1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/ADT/VariadicFunction.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 static cl::opt<bool> ExperimentalVectorShuffleLowering(
71 "x86-experimental-vector-shuffle-lowering", cl::init(true),
72 cl::desc("Enable an experimental vector shuffle lowering code path."),
75 static cl::opt<bool> ExperimentalVectorShuffleLegality(
76 "x86-experimental-vector-shuffle-legality", cl::init(false),
77 cl::desc("Enable experimental shuffle legality based on the experimental "
78 "shuffle lowering. Should only be used with the experimental "
82 static cl::opt<int> ReciprocalEstimateRefinementSteps(
83 "x86-recip-refinement-steps", cl::init(1),
84 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
85 "result of the hardware reciprocal estimate instruction."),
88 // Forward declarations.
89 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
92 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
93 SelectionDAG &DAG, SDLoc dl,
94 unsigned vectorWidth) {
95 assert((vectorWidth == 128 || vectorWidth == 256) &&
96 "Unsupported vector width");
97 EVT VT = Vec.getValueType();
98 EVT ElVT = VT.getVectorElementType();
99 unsigned Factor = VT.getSizeInBits()/vectorWidth;
100 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
101 VT.getVectorNumElements()/Factor);
103 // Extract from UNDEF is UNDEF.
104 if (Vec.getOpcode() == ISD::UNDEF)
105 return DAG.getUNDEF(ResultVT);
107 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
108 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
110 // This is the index of the first element of the vectorWidth-bit chunk
112 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
115 // If the input is a buildvector just emit a smaller one.
116 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
117 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
118 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
121 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
122 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
125 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
126 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
127 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
128 /// instructions or a simple subregister reference. Idx is an index in the
129 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
130 /// lowering EXTRACT_VECTOR_ELT operations easier.
131 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
132 SelectionDAG &DAG, SDLoc dl) {
133 assert((Vec.getValueType().is256BitVector() ||
134 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
135 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
138 /// Generate a DAG to grab 256-bits from a 512-bit vector.
139 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
140 SelectionDAG &DAG, SDLoc dl) {
141 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
142 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
145 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
146 unsigned IdxVal, SelectionDAG &DAG,
147 SDLoc dl, unsigned vectorWidth) {
148 assert((vectorWidth == 128 || vectorWidth == 256) &&
149 "Unsupported vector width");
150 // Inserting UNDEF is Result
151 if (Vec.getOpcode() == ISD::UNDEF)
153 EVT VT = Vec.getValueType();
154 EVT ElVT = VT.getVectorElementType();
155 EVT ResultVT = Result.getValueType();
157 // Insert the relevant vectorWidth bits.
158 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
160 // This is the index of the first element of the vectorWidth-bit chunk
162 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
165 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
166 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
169 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
170 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
171 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
172 /// simple superregister reference. Idx is an index in the 128 bits
173 /// we want. It need not be aligned to a 128-bit boundary. That makes
174 /// lowering INSERT_VECTOR_ELT operations easier.
175 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
176 SelectionDAG &DAG,SDLoc dl) {
177 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
178 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
181 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
182 SelectionDAG &DAG, SDLoc dl) {
183 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
184 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
187 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
188 /// instructions. This is used because creating CONCAT_VECTOR nodes of
189 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
190 /// large BUILD_VECTORS.
191 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
192 unsigned NumElems, SelectionDAG &DAG,
194 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
195 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
198 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
199 unsigned NumElems, SelectionDAG &DAG,
201 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
202 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
205 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
206 const X86Subtarget &STI)
207 : TargetLowering(TM), Subtarget(&STI) {
208 X86ScalarSSEf64 = Subtarget->hasSSE2();
209 X86ScalarSSEf32 = Subtarget->hasSSE1();
210 TD = getDataLayout();
212 // Set up the TargetLowering object.
213 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
215 // X86 is weird. It always uses i8 for shift amounts and setcc results.
216 setBooleanContents(ZeroOrOneBooleanContent);
217 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
218 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
220 // For 64-bit, since we have so many registers, use the ILP scheduler.
221 // For 32-bit, use the register pressure specific scheduling.
222 // For Atom, always use ILP scheduling.
223 if (Subtarget->isAtom())
224 setSchedulingPreference(Sched::ILP);
225 else if (Subtarget->is64Bit())
226 setSchedulingPreference(Sched::ILP);
228 setSchedulingPreference(Sched::RegPressure);
229 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
230 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
232 // Bypass expensive divides on Atom when compiling with O2.
233 if (TM.getOptLevel() >= CodeGenOpt::Default) {
234 if (Subtarget->hasSlowDivide32())
235 addBypassSlowDiv(32, 8);
236 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
237 addBypassSlowDiv(64, 16);
240 if (Subtarget->isTargetKnownWindowsMSVC()) {
241 // Setup Windows compiler runtime calls.
242 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
243 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
244 setLibcallName(RTLIB::SREM_I64, "_allrem");
245 setLibcallName(RTLIB::UREM_I64, "_aullrem");
246 setLibcallName(RTLIB::MUL_I64, "_allmul");
247 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
248 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
249 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
250 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
251 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
253 // The _ftol2 runtime function has an unusual calling conv, which
254 // is modeled by a special pseudo-instruction.
255 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
256 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
257 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
258 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
261 if (Subtarget->isTargetDarwin()) {
262 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
263 setUseUnderscoreSetJmp(false);
264 setUseUnderscoreLongJmp(false);
265 } else if (Subtarget->isTargetWindowsGNU()) {
266 // MS runtime is weird: it exports _setjmp, but longjmp!
267 setUseUnderscoreSetJmp(true);
268 setUseUnderscoreLongJmp(false);
270 setUseUnderscoreSetJmp(true);
271 setUseUnderscoreLongJmp(true);
274 // Set up the register classes.
275 addRegisterClass(MVT::i8, &X86::GR8RegClass);
276 addRegisterClass(MVT::i16, &X86::GR16RegClass);
277 addRegisterClass(MVT::i32, &X86::GR32RegClass);
278 if (Subtarget->is64Bit())
279 addRegisterClass(MVT::i64, &X86::GR64RegClass);
281 for (MVT VT : MVT::integer_valuetypes())
282 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
284 // We don't accept any truncstore of integer registers.
285 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
286 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
287 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
288 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
289 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
290 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
292 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
294 // SETOEQ and SETUNE require checking two conditions.
295 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
296 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
297 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
298 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
299 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
300 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
302 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
304 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
305 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
306 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
308 if (Subtarget->is64Bit()) {
309 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
310 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
311 } else if (!TM.Options.UseSoftFloat) {
312 // We have an algorithm for SSE2->double, and we turn this into a
313 // 64-bit FILD followed by conditional FADD for other targets.
314 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
315 // We have an algorithm for SSE2, and we turn this into a 64-bit
316 // FILD for other targets.
317 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
320 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
322 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
323 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
325 if (!TM.Options.UseSoftFloat) {
326 // SSE has no i16 to fp conversion, only i32
327 if (X86ScalarSSEf32) {
328 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
329 // f32 and f64 cases are Legal, f80 case is not
330 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
332 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
333 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
336 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
337 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
340 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
341 // are Legal, f80 is custom lowered.
342 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
343 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
345 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
347 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
348 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
350 if (X86ScalarSSEf32) {
351 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
352 // f32 and f64 cases are Legal, f80 case is not
353 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
355 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
356 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
359 // Handle FP_TO_UINT by promoting the destination to a larger signed
361 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
362 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
363 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
365 if (Subtarget->is64Bit()) {
366 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
367 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
368 } else if (!TM.Options.UseSoftFloat) {
369 // Since AVX is a superset of SSE3, only check for SSE here.
370 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
371 // Expand FP_TO_UINT into a select.
372 // FIXME: We would like to use a Custom expander here eventually to do
373 // the optimal thing for SSE vs. the default expansion in the legalizer.
374 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
376 // With SSE3 we can use fisttpll to convert to a signed i64; without
377 // SSE, we're stuck with a fistpll.
378 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
381 if (isTargetFTOL()) {
382 // Use the _ftol2 runtime function, which has a pseudo-instruction
383 // to handle its weird calling convention.
384 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
387 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
388 if (!X86ScalarSSEf64) {
389 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
390 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
391 if (Subtarget->is64Bit()) {
392 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
393 // Without SSE, i64->f64 goes through memory.
394 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
398 // Scalar integer divide and remainder are lowered to use operations that
399 // produce two results, to match the available instructions. This exposes
400 // the two-result form to trivial CSE, which is able to combine x/y and x%y
401 // into a single instruction.
403 // Scalar integer multiply-high is also lowered to use two-result
404 // operations, to match the available instructions. However, plain multiply
405 // (low) operations are left as Legal, as there are single-result
406 // instructions for this in x86. Using the two-result multiply instructions
407 // when both high and low results are needed must be arranged by dagcombine.
408 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
410 setOperationAction(ISD::MULHS, VT, Expand);
411 setOperationAction(ISD::MULHU, VT, Expand);
412 setOperationAction(ISD::SDIV, VT, Expand);
413 setOperationAction(ISD::UDIV, VT, Expand);
414 setOperationAction(ISD::SREM, VT, Expand);
415 setOperationAction(ISD::UREM, VT, Expand);
417 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
418 setOperationAction(ISD::ADDC, VT, Custom);
419 setOperationAction(ISD::ADDE, VT, Custom);
420 setOperationAction(ISD::SUBC, VT, Custom);
421 setOperationAction(ISD::SUBE, VT, Custom);
424 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
425 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
426 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
427 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
428 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
429 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
430 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
431 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
432 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
433 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
434 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
435 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
436 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
437 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
438 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
439 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
440 if (Subtarget->is64Bit())
441 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
442 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
443 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
444 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
445 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
446 setOperationAction(ISD::FREM , MVT::f32 , Expand);
447 setOperationAction(ISD::FREM , MVT::f64 , Expand);
448 setOperationAction(ISD::FREM , MVT::f80 , Expand);
449 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
451 // Promote the i8 variants and force them on up to i32 which has a shorter
453 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
454 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
455 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
456 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
457 if (Subtarget->hasBMI()) {
458 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
459 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
460 if (Subtarget->is64Bit())
461 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
463 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
464 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
465 if (Subtarget->is64Bit())
466 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
469 if (Subtarget->hasLZCNT()) {
470 // When promoting the i8 variants, force them to i32 for a shorter
472 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
473 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
474 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
475 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
476 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
477 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
478 if (Subtarget->is64Bit())
479 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
481 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
482 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
483 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
484 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
485 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
486 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
487 if (Subtarget->is64Bit()) {
488 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
489 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
493 // Special handling for half-precision floating point conversions.
494 // If we don't have F16C support, then lower half float conversions
495 // into library calls.
496 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
497 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
498 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
501 // There's never any support for operations beyond MVT::f32.
502 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
503 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
504 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
505 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
507 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
508 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
509 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
510 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
511 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
512 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
514 if (Subtarget->hasPOPCNT()) {
515 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
517 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
518 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
519 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
520 if (Subtarget->is64Bit())
521 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
524 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
526 if (!Subtarget->hasMOVBE())
527 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
529 // These should be promoted to a larger select which is supported.
530 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
531 // X86 wants to expand cmov itself.
532 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
533 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
534 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
535 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
536 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
537 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
538 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
539 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
540 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
541 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
542 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
543 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
544 if (Subtarget->is64Bit()) {
545 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
546 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
548 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
549 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
550 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
551 // support continuation, user-level threading, and etc.. As a result, no
552 // other SjLj exception interfaces are implemented and please don't build
553 // your own exception handling based on them.
554 // LLVM/Clang supports zero-cost DWARF exception handling.
555 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
556 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
559 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
560 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
561 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
562 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
563 if (Subtarget->is64Bit())
564 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
565 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
566 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
567 if (Subtarget->is64Bit()) {
568 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
569 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
570 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
571 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
572 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
574 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
575 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
576 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
577 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
578 if (Subtarget->is64Bit()) {
579 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
580 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
581 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
584 if (Subtarget->hasSSE1())
585 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
587 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
589 // Expand certain atomics
590 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
592 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
593 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
594 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
597 if (Subtarget->hasCmpxchg16b()) {
598 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
601 // FIXME - use subtarget debug flags
602 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
603 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
604 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
607 if (Subtarget->is64Bit()) {
608 setExceptionPointerRegister(X86::RAX);
609 setExceptionSelectorRegister(X86::RDX);
611 setExceptionPointerRegister(X86::EAX);
612 setExceptionSelectorRegister(X86::EDX);
614 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
615 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
617 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
618 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
620 setOperationAction(ISD::TRAP, MVT::Other, Legal);
621 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
623 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
624 setOperationAction(ISD::VASTART , MVT::Other, Custom);
625 setOperationAction(ISD::VAEND , MVT::Other, Expand);
626 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
627 // TargetInfo::X86_64ABIBuiltinVaList
628 setOperationAction(ISD::VAARG , MVT::Other, Custom);
629 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
631 // TargetInfo::CharPtrBuiltinVaList
632 setOperationAction(ISD::VAARG , MVT::Other, Expand);
633 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
636 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
637 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
639 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
641 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
642 // f32 and f64 use SSE.
643 // Set up the FP register classes.
644 addRegisterClass(MVT::f32, &X86::FR32RegClass);
645 addRegisterClass(MVT::f64, &X86::FR64RegClass);
647 // Use ANDPD to simulate FABS.
648 setOperationAction(ISD::FABS , MVT::f64, Custom);
649 setOperationAction(ISD::FABS , MVT::f32, Custom);
651 // Use XORP to simulate FNEG.
652 setOperationAction(ISD::FNEG , MVT::f64, Custom);
653 setOperationAction(ISD::FNEG , MVT::f32, Custom);
655 // Use ANDPD and ORPD to simulate FCOPYSIGN.
656 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
657 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
659 // Lower this to FGETSIGNx86 plus an AND.
660 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
661 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
663 // We don't support sin/cos/fmod
664 setOperationAction(ISD::FSIN , MVT::f64, Expand);
665 setOperationAction(ISD::FCOS , MVT::f64, Expand);
666 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
667 setOperationAction(ISD::FSIN , MVT::f32, Expand);
668 setOperationAction(ISD::FCOS , MVT::f32, Expand);
669 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
671 // Expand FP immediates into loads from the stack, except for the special
673 addLegalFPImmediate(APFloat(+0.0)); // xorpd
674 addLegalFPImmediate(APFloat(+0.0f)); // xorps
675 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
676 // Use SSE for f32, x87 for f64.
677 // Set up the FP register classes.
678 addRegisterClass(MVT::f32, &X86::FR32RegClass);
679 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
681 // Use ANDPS to simulate FABS.
682 setOperationAction(ISD::FABS , MVT::f32, Custom);
684 // Use XORP to simulate FNEG.
685 setOperationAction(ISD::FNEG , MVT::f32, Custom);
687 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
689 // Use ANDPS and ORPS to simulate FCOPYSIGN.
690 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
691 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
693 // We don't support sin/cos/fmod
694 setOperationAction(ISD::FSIN , MVT::f32, Expand);
695 setOperationAction(ISD::FCOS , MVT::f32, Expand);
696 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
698 // Special cases we handle for FP constants.
699 addLegalFPImmediate(APFloat(+0.0f)); // xorps
700 addLegalFPImmediate(APFloat(+0.0)); // FLD0
701 addLegalFPImmediate(APFloat(+1.0)); // FLD1
702 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
703 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
705 if (!TM.Options.UnsafeFPMath) {
706 setOperationAction(ISD::FSIN , MVT::f64, Expand);
707 setOperationAction(ISD::FCOS , MVT::f64, Expand);
708 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
710 } else if (!TM.Options.UseSoftFloat) {
711 // f32 and f64 in x87.
712 // Set up the FP register classes.
713 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
714 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
716 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
717 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
718 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
719 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
721 if (!TM.Options.UnsafeFPMath) {
722 setOperationAction(ISD::FSIN , MVT::f64, Expand);
723 setOperationAction(ISD::FSIN , MVT::f32, Expand);
724 setOperationAction(ISD::FCOS , MVT::f64, Expand);
725 setOperationAction(ISD::FCOS , MVT::f32, Expand);
726 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
727 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
729 addLegalFPImmediate(APFloat(+0.0)); // FLD0
730 addLegalFPImmediate(APFloat(+1.0)); // FLD1
731 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
732 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
733 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
734 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
735 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
736 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
739 // We don't support FMA.
740 setOperationAction(ISD::FMA, MVT::f64, Expand);
741 setOperationAction(ISD::FMA, MVT::f32, Expand);
743 // Long double always uses X87.
744 if (!TM.Options.UseSoftFloat) {
745 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
746 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
747 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
749 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
750 addLegalFPImmediate(TmpFlt); // FLD0
752 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
755 APFloat TmpFlt2(+1.0);
756 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
758 addLegalFPImmediate(TmpFlt2); // FLD1
759 TmpFlt2.changeSign();
760 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
763 if (!TM.Options.UnsafeFPMath) {
764 setOperationAction(ISD::FSIN , MVT::f80, Expand);
765 setOperationAction(ISD::FCOS , MVT::f80, Expand);
766 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
769 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
770 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
771 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
772 setOperationAction(ISD::FRINT, MVT::f80, Expand);
773 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
774 setOperationAction(ISD::FMA, MVT::f80, Expand);
777 // Always use a library call for pow.
778 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
779 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
780 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
782 setOperationAction(ISD::FLOG, MVT::f80, Expand);
783 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
784 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
785 setOperationAction(ISD::FEXP, MVT::f80, Expand);
786 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
787 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
788 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
790 // First set operation action for all vector types to either promote
791 // (for widening) or expand (for scalarization). Then we will selectively
792 // turn on ones that can be effectively codegen'd.
793 for (MVT VT : MVT::vector_valuetypes()) {
794 setOperationAction(ISD::ADD , VT, Expand);
795 setOperationAction(ISD::SUB , VT, Expand);
796 setOperationAction(ISD::FADD, VT, Expand);
797 setOperationAction(ISD::FNEG, VT, Expand);
798 setOperationAction(ISD::FSUB, VT, Expand);
799 setOperationAction(ISD::MUL , VT, Expand);
800 setOperationAction(ISD::FMUL, VT, Expand);
801 setOperationAction(ISD::SDIV, VT, Expand);
802 setOperationAction(ISD::UDIV, VT, Expand);
803 setOperationAction(ISD::FDIV, VT, Expand);
804 setOperationAction(ISD::SREM, VT, Expand);
805 setOperationAction(ISD::UREM, VT, Expand);
806 setOperationAction(ISD::LOAD, VT, Expand);
807 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
808 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
809 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
810 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
811 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
812 setOperationAction(ISD::FABS, VT, Expand);
813 setOperationAction(ISD::FSIN, VT, Expand);
814 setOperationAction(ISD::FSINCOS, VT, Expand);
815 setOperationAction(ISD::FCOS, VT, Expand);
816 setOperationAction(ISD::FSINCOS, VT, Expand);
817 setOperationAction(ISD::FREM, VT, Expand);
818 setOperationAction(ISD::FMA, VT, Expand);
819 setOperationAction(ISD::FPOWI, VT, Expand);
820 setOperationAction(ISD::FSQRT, VT, Expand);
821 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
822 setOperationAction(ISD::FFLOOR, VT, Expand);
823 setOperationAction(ISD::FCEIL, VT, Expand);
824 setOperationAction(ISD::FTRUNC, VT, Expand);
825 setOperationAction(ISD::FRINT, VT, Expand);
826 setOperationAction(ISD::FNEARBYINT, VT, Expand);
827 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
828 setOperationAction(ISD::MULHS, VT, Expand);
829 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
830 setOperationAction(ISD::MULHU, VT, Expand);
831 setOperationAction(ISD::SDIVREM, VT, Expand);
832 setOperationAction(ISD::UDIVREM, VT, Expand);
833 setOperationAction(ISD::FPOW, VT, Expand);
834 setOperationAction(ISD::CTPOP, VT, Expand);
835 setOperationAction(ISD::CTTZ, VT, Expand);
836 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
837 setOperationAction(ISD::CTLZ, VT, Expand);
838 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
839 setOperationAction(ISD::SHL, VT, Expand);
840 setOperationAction(ISD::SRA, VT, Expand);
841 setOperationAction(ISD::SRL, VT, Expand);
842 setOperationAction(ISD::ROTL, VT, Expand);
843 setOperationAction(ISD::ROTR, VT, Expand);
844 setOperationAction(ISD::BSWAP, VT, Expand);
845 setOperationAction(ISD::SETCC, VT, Expand);
846 setOperationAction(ISD::FLOG, VT, Expand);
847 setOperationAction(ISD::FLOG2, VT, Expand);
848 setOperationAction(ISD::FLOG10, VT, Expand);
849 setOperationAction(ISD::FEXP, VT, Expand);
850 setOperationAction(ISD::FEXP2, VT, Expand);
851 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
852 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
853 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
854 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
855 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
856 setOperationAction(ISD::TRUNCATE, VT, Expand);
857 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
858 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
859 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
860 setOperationAction(ISD::VSELECT, VT, Expand);
861 setOperationAction(ISD::SELECT_CC, VT, Expand);
862 for (MVT InnerVT : MVT::vector_valuetypes()) {
863 setTruncStoreAction(InnerVT, VT, Expand);
865 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
866 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
868 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
869 // types, we have to deal with them whether we ask for Expansion or not.
870 // Setting Expand causes its own optimisation problems though, so leave
872 if (VT.getVectorElementType() == MVT::i1)
873 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
877 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
878 // with -msoft-float, disable use of MMX as well.
879 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
880 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
881 // No operations on x86mmx supported, everything uses intrinsics.
884 // MMX-sized vectors (other than x86mmx) are expected to be expanded
885 // into smaller operations.
886 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
887 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
888 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
889 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
890 setOperationAction(ISD::AND, MVT::v8i8, Expand);
891 setOperationAction(ISD::AND, MVT::v4i16, Expand);
892 setOperationAction(ISD::AND, MVT::v2i32, Expand);
893 setOperationAction(ISD::AND, MVT::v1i64, Expand);
894 setOperationAction(ISD::OR, MVT::v8i8, Expand);
895 setOperationAction(ISD::OR, MVT::v4i16, Expand);
896 setOperationAction(ISD::OR, MVT::v2i32, Expand);
897 setOperationAction(ISD::OR, MVT::v1i64, Expand);
898 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
899 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
900 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
901 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
902 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
903 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
904 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
905 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
907 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
908 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
909 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
910 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
911 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
912 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
913 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
914 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
916 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
917 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
919 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
920 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
921 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
922 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
923 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
924 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
925 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
926 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
927 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
928 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
930 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
931 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
934 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
935 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
937 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
938 // registers cannot be used even for integer operations.
939 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
940 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
941 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
942 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
944 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
945 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
946 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
947 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
948 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
949 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
950 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
951 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
952 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
953 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
954 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
955 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
956 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
957 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
958 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
959 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
960 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
961 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
962 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
963 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
964 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
965 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
967 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
968 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
969 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
970 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
972 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
973 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
974 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
975 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
976 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
978 // Only provide customized ctpop vector bit twiddling for vector types we
979 // know to perform better than using the popcnt instructions on each vector
980 // element. If popcnt isn't supported, always provide the custom version.
981 if (!Subtarget->hasPOPCNT()) {
982 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
983 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
986 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
987 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
988 MVT VT = (MVT::SimpleValueType)i;
989 // Do not attempt to custom lower non-power-of-2 vectors
990 if (!isPowerOf2_32(VT.getVectorNumElements()))
992 // Do not attempt to custom lower non-128-bit vectors
993 if (!VT.is128BitVector())
995 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
996 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
997 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1000 // We support custom legalizing of sext and anyext loads for specific
1001 // memory vector types which we can load as a scalar (or sequence of
1002 // scalars) and extend in-register to a legal 128-bit vector type. For sext
1003 // loads these must work with a single scalar load.
1004 for (MVT VT : MVT::integer_vector_valuetypes()) {
1005 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
1006 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
1007 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
1008 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
1009 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
1010 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
1011 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
1012 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
1013 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
1016 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1017 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1018 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1019 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1020 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1021 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1023 if (Subtarget->is64Bit()) {
1024 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1025 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1028 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1029 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1030 MVT VT = (MVT::SimpleValueType)i;
1032 // Do not attempt to promote non-128-bit vectors
1033 if (!VT.is128BitVector())
1036 setOperationAction(ISD::AND, VT, Promote);
1037 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1038 setOperationAction(ISD::OR, VT, Promote);
1039 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1040 setOperationAction(ISD::XOR, VT, Promote);
1041 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1042 setOperationAction(ISD::LOAD, VT, Promote);
1043 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1044 setOperationAction(ISD::SELECT, VT, Promote);
1045 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1048 // Custom lower v2i64 and v2f64 selects.
1049 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1050 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1051 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1052 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1054 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1055 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1057 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1058 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1059 // As there is no 64-bit GPR available, we need build a special custom
1060 // sequence to convert from v2i32 to v2f32.
1061 if (!Subtarget->is64Bit())
1062 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1064 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1065 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1067 for (MVT VT : MVT::fp_vector_valuetypes())
1068 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
1070 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1071 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1072 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1075 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1076 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1077 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1078 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1079 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1080 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1081 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1082 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1083 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1084 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1085 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1087 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1088 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1089 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1090 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1091 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1092 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1093 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1094 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1095 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1096 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1098 // FIXME: Do we need to handle scalar-to-vector here?
1099 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1101 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1102 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1103 setOperationAction(ISD::VSELECT, MVT::v4i32, Custom);
1104 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
1105 setOperationAction(ISD::VSELECT, MVT::v8i16, Custom);
1106 // There is no BLENDI for byte vectors. We don't need to custom lower
1107 // some vselects for now.
1108 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1110 // SSE41 brings specific instructions for doing vector sign extend even in
1111 // cases where we don't have SRA.
1112 for (MVT VT : MVT::integer_vector_valuetypes()) {
1113 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
1114 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
1115 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
1118 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
1119 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1120 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1121 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1122 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1123 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1124 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1126 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1127 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1128 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1129 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1130 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1131 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1133 // i8 and i16 vectors are custom because the source register and source
1134 // source memory operand types are not the same width. f32 vectors are
1135 // custom since the immediate controlling the insert encodes additional
1137 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1138 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1139 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1140 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1142 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1143 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1144 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1145 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1147 // FIXME: these should be Legal, but that's only for the case where
1148 // the index is constant. For now custom expand to deal with that.
1149 if (Subtarget->is64Bit()) {
1150 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1151 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1155 if (Subtarget->hasSSE2()) {
1156 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1157 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1159 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1160 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1162 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1163 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1165 // In the customized shift lowering, the legal cases in AVX2 will be
1167 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1168 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1170 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1171 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1173 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1176 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1177 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1178 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1179 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1180 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1181 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1182 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1184 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1185 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1186 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1188 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1189 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1190 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1191 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1192 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1193 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1194 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1195 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1196 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1197 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1198 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1199 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1201 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1202 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1203 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1204 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1205 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1206 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1207 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1208 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1209 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1210 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1211 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1212 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1214 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1215 // even though v8i16 is a legal type.
1216 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1217 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1218 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1220 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1221 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1222 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1224 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1225 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1227 for (MVT VT : MVT::fp_vector_valuetypes())
1228 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1230 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1231 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1233 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1234 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1236 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1237 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1239 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1240 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1241 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1242 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1244 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1245 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1246 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1248 setOperationAction(ISD::VSELECT, MVT::v4f64, Custom);
1249 setOperationAction(ISD::VSELECT, MVT::v4i64, Custom);
1250 setOperationAction(ISD::VSELECT, MVT::v8i32, Custom);
1251 setOperationAction(ISD::VSELECT, MVT::v8f32, Custom);
1253 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1254 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1255 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1256 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1257 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1258 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1259 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1260 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1261 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1262 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1263 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1264 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1266 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1267 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1268 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1269 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1270 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1271 setOperationAction(ISD::FMA, MVT::f32, Legal);
1272 setOperationAction(ISD::FMA, MVT::f64, Legal);
1275 if (Subtarget->hasInt256()) {
1276 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1277 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1278 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1279 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1281 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1282 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1283 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1284 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1286 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1287 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1288 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1289 // Don't lower v32i8 because there is no 128-bit byte mul
1291 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1292 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1293 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1294 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1296 setOperationAction(ISD::VSELECT, MVT::v16i16, Custom);
1297 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1299 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1300 // when we have a 256bit-wide blend with immediate.
1301 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1303 // Only provide customized ctpop vector bit twiddling for vector types we
1304 // know to perform better than using the popcnt instructions on each
1305 // vector element. If popcnt isn't supported, always provide the custom
1307 if (!Subtarget->hasPOPCNT())
1308 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1310 // Custom CTPOP always performs better on natively supported v8i32
1311 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1313 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1314 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1315 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1316 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1317 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1318 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1319 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1321 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1322 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1323 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1324 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1325 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1326 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1328 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1329 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1330 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1331 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1333 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1334 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1335 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1336 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1338 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1339 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1340 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1341 // Don't lower v32i8 because there is no 128-bit byte mul
1344 // In the customized shift lowering, the legal cases in AVX2 will be
1346 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1347 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1349 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1350 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1352 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1354 // Custom lower several nodes for 256-bit types.
1355 for (MVT VT : MVT::vector_valuetypes()) {
1356 if (VT.getScalarSizeInBits() >= 32) {
1357 setOperationAction(ISD::MLOAD, VT, Legal);
1358 setOperationAction(ISD::MSTORE, VT, Legal);
1360 // Extract subvector is special because the value type
1361 // (result) is 128-bit but the source is 256-bit wide.
1362 if (VT.is128BitVector()) {
1363 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1365 // Do not attempt to custom lower other non-256-bit vectors
1366 if (!VT.is256BitVector())
1369 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1370 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1371 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1372 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1373 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1374 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1375 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1378 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1379 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1380 MVT VT = (MVT::SimpleValueType)i;
1382 // Do not attempt to promote non-256-bit vectors
1383 if (!VT.is256BitVector())
1386 setOperationAction(ISD::AND, VT, Promote);
1387 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1388 setOperationAction(ISD::OR, VT, Promote);
1389 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1390 setOperationAction(ISD::XOR, VT, Promote);
1391 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1392 setOperationAction(ISD::LOAD, VT, Promote);
1393 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1394 setOperationAction(ISD::SELECT, VT, Promote);
1395 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1399 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1400 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1401 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1402 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1403 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1405 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1406 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1407 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1409 for (MVT VT : MVT::fp_vector_valuetypes())
1410 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1412 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1413 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1414 setOperationAction(ISD::XOR, MVT::i1, Legal);
1415 setOperationAction(ISD::OR, MVT::i1, Legal);
1416 setOperationAction(ISD::AND, MVT::i1, Legal);
1417 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1418 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1419 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1420 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1421 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1423 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1424 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1425 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1426 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1427 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1428 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1430 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1431 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1432 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1433 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1434 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1435 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1436 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1437 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1439 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1440 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1441 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1442 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1443 if (Subtarget->is64Bit()) {
1444 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1445 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1446 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1447 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1449 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1450 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1451 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1452 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1453 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1454 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1455 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1456 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1457 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1458 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1459 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1460 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1461 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1462 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1464 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1465 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1466 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1467 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1468 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1469 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1470 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1471 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1472 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1473 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1474 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1475 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1476 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1478 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1479 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1480 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1481 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1482 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1483 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1485 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1486 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1488 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1490 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1491 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1492 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1493 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1494 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1495 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1496 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1497 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1498 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1500 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1501 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1503 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1504 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1506 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1508 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1509 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1511 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1512 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1514 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1515 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1517 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1518 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1519 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1520 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1521 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1522 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1524 if (Subtarget->hasCDI()) {
1525 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1526 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1529 // Custom lower several nodes.
1530 for (MVT VT : MVT::vector_valuetypes()) {
1531 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1532 // Extract subvector is special because the value type
1533 // (result) is 256/128-bit but the source is 512-bit wide.
1534 if (VT.is128BitVector() || VT.is256BitVector()) {
1535 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1537 if (VT.getVectorElementType() == MVT::i1)
1538 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1540 // Do not attempt to custom lower other non-512-bit vectors
1541 if (!VT.is512BitVector())
1544 if ( EltSize >= 32) {
1545 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1546 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1547 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1548 setOperationAction(ISD::VSELECT, VT, Legal);
1549 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1550 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1551 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1552 setOperationAction(ISD::MLOAD, VT, Legal);
1553 setOperationAction(ISD::MSTORE, VT, Legal);
1556 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1557 MVT VT = (MVT::SimpleValueType)i;
1559 // Do not attempt to promote non-512-bit vectors.
1560 if (!VT.is512BitVector())
1563 setOperationAction(ISD::SELECT, VT, Promote);
1564 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1568 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1569 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1570 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1572 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1573 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1575 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1576 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1577 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1578 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1579 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1580 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1581 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1582 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1583 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1585 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1586 const MVT VT = (MVT::SimpleValueType)i;
1588 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1590 // Do not attempt to promote non-512-bit vectors.
1591 if (!VT.is512BitVector())
1595 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1596 setOperationAction(ISD::VSELECT, VT, Legal);
1601 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1602 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1603 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1605 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1606 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1607 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal);
1609 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1610 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1611 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1612 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1613 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1614 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1617 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1618 // of this type with custom code.
1619 for (MVT VT : MVT::vector_valuetypes())
1620 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
1622 // We want to custom lower some of our intrinsics.
1623 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1624 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1625 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1626 if (!Subtarget->is64Bit())
1627 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1629 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1630 // handle type legalization for these operations here.
1632 // FIXME: We really should do custom legalization for addition and
1633 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1634 // than generic legalization for 64-bit multiplication-with-overflow, though.
1635 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1636 // Add/Sub/Mul with overflow operations are custom lowered.
1638 setOperationAction(ISD::SADDO, VT, Custom);
1639 setOperationAction(ISD::UADDO, VT, Custom);
1640 setOperationAction(ISD::SSUBO, VT, Custom);
1641 setOperationAction(ISD::USUBO, VT, Custom);
1642 setOperationAction(ISD::SMULO, VT, Custom);
1643 setOperationAction(ISD::UMULO, VT, Custom);
1647 if (!Subtarget->is64Bit()) {
1648 // These libcalls are not available in 32-bit.
1649 setLibcallName(RTLIB::SHL_I128, nullptr);
1650 setLibcallName(RTLIB::SRL_I128, nullptr);
1651 setLibcallName(RTLIB::SRA_I128, nullptr);
1654 // Combine sin / cos into one node or libcall if possible.
1655 if (Subtarget->hasSinCos()) {
1656 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1657 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1658 if (Subtarget->isTargetDarwin()) {
1659 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1660 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1661 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1662 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1666 if (Subtarget->isTargetWin64()) {
1667 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1668 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1669 setOperationAction(ISD::SREM, MVT::i128, Custom);
1670 setOperationAction(ISD::UREM, MVT::i128, Custom);
1671 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1672 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1675 // We have target-specific dag combine patterns for the following nodes:
1676 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1677 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1678 setTargetDAGCombine(ISD::VSELECT);
1679 setTargetDAGCombine(ISD::SELECT);
1680 setTargetDAGCombine(ISD::SHL);
1681 setTargetDAGCombine(ISD::SRA);
1682 setTargetDAGCombine(ISD::SRL);
1683 setTargetDAGCombine(ISD::OR);
1684 setTargetDAGCombine(ISD::AND);
1685 setTargetDAGCombine(ISD::ADD);
1686 setTargetDAGCombine(ISD::FADD);
1687 setTargetDAGCombine(ISD::FSUB);
1688 setTargetDAGCombine(ISD::FMA);
1689 setTargetDAGCombine(ISD::SUB);
1690 setTargetDAGCombine(ISD::LOAD);
1691 setTargetDAGCombine(ISD::MLOAD);
1692 setTargetDAGCombine(ISD::STORE);
1693 setTargetDAGCombine(ISD::MSTORE);
1694 setTargetDAGCombine(ISD::ZERO_EXTEND);
1695 setTargetDAGCombine(ISD::ANY_EXTEND);
1696 setTargetDAGCombine(ISD::SIGN_EXTEND);
1697 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1698 setTargetDAGCombine(ISD::TRUNCATE);
1699 setTargetDAGCombine(ISD::SINT_TO_FP);
1700 setTargetDAGCombine(ISD::SETCC);
1701 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1702 setTargetDAGCombine(ISD::BUILD_VECTOR);
1703 setTargetDAGCombine(ISD::MUL);
1704 setTargetDAGCombine(ISD::XOR);
1706 computeRegisterProperties();
1708 // On Darwin, -Os means optimize for size without hurting performance,
1709 // do not reduce the limit.
1710 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1711 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1712 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1713 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1714 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1715 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1716 setPrefLoopAlignment(4); // 2^4 bytes.
1718 // Predictable cmov don't hurt on atom because it's in-order.
1719 PredictableSelectIsExpensive = !Subtarget->isAtom();
1720 EnableExtLdPromotion = true;
1721 setPrefFunctionAlignment(4); // 2^4 bytes.
1723 verifyIntrinsicTables();
1726 // This has so far only been implemented for 64-bit MachO.
1727 bool X86TargetLowering::useLoadStackGuardNode() const {
1728 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1731 TargetLoweringBase::LegalizeTypeAction
1732 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1733 if (ExperimentalVectorWideningLegalization &&
1734 VT.getVectorNumElements() != 1 &&
1735 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1736 return TypeWidenVector;
1738 return TargetLoweringBase::getPreferredVectorAction(VT);
1741 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1743 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1745 const unsigned NumElts = VT.getVectorNumElements();
1746 const EVT EltVT = VT.getVectorElementType();
1747 if (VT.is512BitVector()) {
1748 if (Subtarget->hasAVX512())
1749 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1750 EltVT == MVT::f32 || EltVT == MVT::f64)
1752 case 8: return MVT::v8i1;
1753 case 16: return MVT::v16i1;
1755 if (Subtarget->hasBWI())
1756 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1758 case 32: return MVT::v32i1;
1759 case 64: return MVT::v64i1;
1763 if (VT.is256BitVector() || VT.is128BitVector()) {
1764 if (Subtarget->hasVLX())
1765 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1766 EltVT == MVT::f32 || EltVT == MVT::f64)
1768 case 2: return MVT::v2i1;
1769 case 4: return MVT::v4i1;
1770 case 8: return MVT::v8i1;
1772 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1773 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1775 case 8: return MVT::v8i1;
1776 case 16: return MVT::v16i1;
1777 case 32: return MVT::v32i1;
1781 return VT.changeVectorElementTypeToInteger();
1784 /// Helper for getByValTypeAlignment to determine
1785 /// the desired ByVal argument alignment.
1786 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1789 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1790 if (VTy->getBitWidth() == 128)
1792 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1793 unsigned EltAlign = 0;
1794 getMaxByValAlign(ATy->getElementType(), EltAlign);
1795 if (EltAlign > MaxAlign)
1796 MaxAlign = EltAlign;
1797 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1798 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1799 unsigned EltAlign = 0;
1800 getMaxByValAlign(STy->getElementType(i), EltAlign);
1801 if (EltAlign > MaxAlign)
1802 MaxAlign = EltAlign;
1809 /// Return the desired alignment for ByVal aggregate
1810 /// function arguments in the caller parameter area. For X86, aggregates
1811 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1812 /// are at 4-byte boundaries.
1813 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1814 if (Subtarget->is64Bit()) {
1815 // Max of 8 and alignment of type.
1816 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1823 if (Subtarget->hasSSE1())
1824 getMaxByValAlign(Ty, Align);
1828 /// Returns the target specific optimal type for load
1829 /// and store operations as a result of memset, memcpy, and memmove
1830 /// lowering. If DstAlign is zero that means it's safe to destination
1831 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1832 /// means there isn't a need to check it against alignment requirement,
1833 /// probably because the source does not need to be loaded. If 'IsMemset' is
1834 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1835 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1836 /// source is constant so it does not need to be loaded.
1837 /// It returns EVT::Other if the type should be determined using generic
1838 /// target-independent logic.
1840 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1841 unsigned DstAlign, unsigned SrcAlign,
1842 bool IsMemset, bool ZeroMemset,
1844 MachineFunction &MF) const {
1845 const Function *F = MF.getFunction();
1846 if ((!IsMemset || ZeroMemset) &&
1847 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1848 Attribute::NoImplicitFloat)) {
1850 (Subtarget->isUnalignedMemAccessFast() ||
1851 ((DstAlign == 0 || DstAlign >= 16) &&
1852 (SrcAlign == 0 || SrcAlign >= 16)))) {
1854 if (Subtarget->hasInt256())
1856 if (Subtarget->hasFp256())
1859 if (Subtarget->hasSSE2())
1861 if (Subtarget->hasSSE1())
1863 } else if (!MemcpyStrSrc && Size >= 8 &&
1864 !Subtarget->is64Bit() &&
1865 Subtarget->hasSSE2()) {
1866 // Do not use f64 to lower memcpy if source is string constant. It's
1867 // better to use i32 to avoid the loads.
1871 if (Subtarget->is64Bit() && Size >= 8)
1876 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1878 return X86ScalarSSEf32;
1879 else if (VT == MVT::f64)
1880 return X86ScalarSSEf64;
1885 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1890 *Fast = Subtarget->isUnalignedMemAccessFast();
1894 /// Return the entry encoding for a jump table in the
1895 /// current function. The returned value is a member of the
1896 /// MachineJumpTableInfo::JTEntryKind enum.
1897 unsigned X86TargetLowering::getJumpTableEncoding() const {
1898 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1900 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1901 Subtarget->isPICStyleGOT())
1902 return MachineJumpTableInfo::EK_Custom32;
1904 // Otherwise, use the normal jump table encoding heuristics.
1905 return TargetLowering::getJumpTableEncoding();
1909 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1910 const MachineBasicBlock *MBB,
1911 unsigned uid,MCContext &Ctx) const{
1912 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1913 Subtarget->isPICStyleGOT());
1914 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1916 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1917 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1920 /// Returns relocation base for the given PIC jumptable.
1921 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1922 SelectionDAG &DAG) const {
1923 if (!Subtarget->is64Bit())
1924 // This doesn't have SDLoc associated with it, but is not really the
1925 // same as a Register.
1926 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1930 /// This returns the relocation base for the given PIC jumptable,
1931 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1932 const MCExpr *X86TargetLowering::
1933 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1934 MCContext &Ctx) const {
1935 // X86-64 uses RIP relative addressing based on the jump table label.
1936 if (Subtarget->isPICStyleRIPRel())
1937 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1939 // Otherwise, the reference is relative to the PIC base.
1940 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1943 // FIXME: Why this routine is here? Move to RegInfo!
1944 std::pair<const TargetRegisterClass*, uint8_t>
1945 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1946 const TargetRegisterClass *RRC = nullptr;
1948 switch (VT.SimpleTy) {
1950 return TargetLowering::findRepresentativeClass(VT);
1951 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1952 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1955 RRC = &X86::VR64RegClass;
1957 case MVT::f32: case MVT::f64:
1958 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1959 case MVT::v4f32: case MVT::v2f64:
1960 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1962 RRC = &X86::VR128RegClass;
1965 return std::make_pair(RRC, Cost);
1968 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1969 unsigned &Offset) const {
1970 if (!Subtarget->isTargetLinux())
1973 if (Subtarget->is64Bit()) {
1974 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1976 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1988 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1989 unsigned DestAS) const {
1990 assert(SrcAS != DestAS && "Expected different address spaces!");
1992 return SrcAS < 256 && DestAS < 256;
1995 //===----------------------------------------------------------------------===//
1996 // Return Value Calling Convention Implementation
1997 //===----------------------------------------------------------------------===//
1999 #include "X86GenCallingConv.inc"
2002 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2003 MachineFunction &MF, bool isVarArg,
2004 const SmallVectorImpl<ISD::OutputArg> &Outs,
2005 LLVMContext &Context) const {
2006 SmallVector<CCValAssign, 16> RVLocs;
2007 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2008 return CCInfo.CheckReturn(Outs, RetCC_X86);
2011 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2012 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2017 X86TargetLowering::LowerReturn(SDValue Chain,
2018 CallingConv::ID CallConv, bool isVarArg,
2019 const SmallVectorImpl<ISD::OutputArg> &Outs,
2020 const SmallVectorImpl<SDValue> &OutVals,
2021 SDLoc dl, SelectionDAG &DAG) const {
2022 MachineFunction &MF = DAG.getMachineFunction();
2023 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2025 SmallVector<CCValAssign, 16> RVLocs;
2026 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2027 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2030 SmallVector<SDValue, 6> RetOps;
2031 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2032 // Operand #1 = Bytes To Pop
2033 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
2036 // Copy the result values into the output registers.
2037 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2038 CCValAssign &VA = RVLocs[i];
2039 assert(VA.isRegLoc() && "Can only return in registers!");
2040 SDValue ValToCopy = OutVals[i];
2041 EVT ValVT = ValToCopy.getValueType();
2043 // Promote values to the appropriate types.
2044 if (VA.getLocInfo() == CCValAssign::SExt)
2045 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2046 else if (VA.getLocInfo() == CCValAssign::ZExt)
2047 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2048 else if (VA.getLocInfo() == CCValAssign::AExt)
2049 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2050 else if (VA.getLocInfo() == CCValAssign::BCvt)
2051 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2053 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2054 "Unexpected FP-extend for return value.");
2056 // If this is x86-64, and we disabled SSE, we can't return FP values,
2057 // or SSE or MMX vectors.
2058 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2059 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2060 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2061 report_fatal_error("SSE register return with SSE disabled");
2063 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2064 // llvm-gcc has never done it right and no one has noticed, so this
2065 // should be OK for now.
2066 if (ValVT == MVT::f64 &&
2067 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2068 report_fatal_error("SSE2 register return with SSE2 disabled");
2070 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2071 // the RET instruction and handled by the FP Stackifier.
2072 if (VA.getLocReg() == X86::FP0 ||
2073 VA.getLocReg() == X86::FP1) {
2074 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2075 // change the value to the FP stack register class.
2076 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2077 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2078 RetOps.push_back(ValToCopy);
2079 // Don't emit a copytoreg.
2083 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2084 // which is returned in RAX / RDX.
2085 if (Subtarget->is64Bit()) {
2086 if (ValVT == MVT::x86mmx) {
2087 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2088 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2089 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2091 // If we don't have SSE2 available, convert to v4f32 so the generated
2092 // register is legal.
2093 if (!Subtarget->hasSSE2())
2094 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2099 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2100 Flag = Chain.getValue(1);
2101 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2104 // The x86-64 ABIs require that for returning structs by value we copy
2105 // the sret argument into %rax/%eax (depending on ABI) for the return.
2106 // Win32 requires us to put the sret argument to %eax as well.
2107 // We saved the argument into a virtual register in the entry block,
2108 // so now we copy the value out and into %rax/%eax.
2109 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
2110 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
2111 MachineFunction &MF = DAG.getMachineFunction();
2112 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2113 unsigned Reg = FuncInfo->getSRetReturnReg();
2115 "SRetReturnReg should have been set in LowerFormalArguments().");
2116 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2119 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2120 X86::RAX : X86::EAX;
2121 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2122 Flag = Chain.getValue(1);
2124 // RAX/EAX now acts like a return value.
2125 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2128 RetOps[0] = Chain; // Update chain.
2130 // Add the flag if we have it.
2132 RetOps.push_back(Flag);
2134 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2137 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2138 if (N->getNumValues() != 1)
2140 if (!N->hasNUsesOfValue(1, 0))
2143 SDValue TCChain = Chain;
2144 SDNode *Copy = *N->use_begin();
2145 if (Copy->getOpcode() == ISD::CopyToReg) {
2146 // If the copy has a glue operand, we conservatively assume it isn't safe to
2147 // perform a tail call.
2148 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2150 TCChain = Copy->getOperand(0);
2151 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2154 bool HasRet = false;
2155 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2157 if (UI->getOpcode() != X86ISD::RET_FLAG)
2159 // If we are returning more than one value, we can definitely
2160 // not make a tail call see PR19530
2161 if (UI->getNumOperands() > 4)
2163 if (UI->getNumOperands() == 4 &&
2164 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2177 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2178 ISD::NodeType ExtendKind) const {
2180 // TODO: Is this also valid on 32-bit?
2181 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2182 ReturnMVT = MVT::i8;
2184 ReturnMVT = MVT::i32;
2186 EVT MinVT = getRegisterType(Context, ReturnMVT);
2187 return VT.bitsLT(MinVT) ? MinVT : VT;
2190 /// Lower the result values of a call into the
2191 /// appropriate copies out of appropriate physical registers.
2194 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2195 CallingConv::ID CallConv, bool isVarArg,
2196 const SmallVectorImpl<ISD::InputArg> &Ins,
2197 SDLoc dl, SelectionDAG &DAG,
2198 SmallVectorImpl<SDValue> &InVals) const {
2200 // Assign locations to each value returned by this call.
2201 SmallVector<CCValAssign, 16> RVLocs;
2202 bool Is64Bit = Subtarget->is64Bit();
2203 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2205 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2207 // Copy all of the result registers out of their specified physreg.
2208 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2209 CCValAssign &VA = RVLocs[i];
2210 EVT CopyVT = VA.getValVT();
2212 // If this is x86-64, and we disabled SSE, we can't return FP values
2213 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2214 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2215 report_fatal_error("SSE register return with SSE disabled");
2218 // If we prefer to use the value in xmm registers, copy it out as f80 and
2219 // use a truncate to move it from fp stack reg to xmm reg.
2220 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2221 isScalarFPTypeInSSEReg(VA.getValVT()))
2224 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2225 CopyVT, InFlag).getValue(1);
2226 SDValue Val = Chain.getValue(0);
2228 if (CopyVT != VA.getValVT())
2229 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2230 // This truncation won't change the value.
2231 DAG.getIntPtrConstant(1));
2233 InFlag = Chain.getValue(2);
2234 InVals.push_back(Val);
2240 //===----------------------------------------------------------------------===//
2241 // C & StdCall & Fast Calling Convention implementation
2242 //===----------------------------------------------------------------------===//
2243 // StdCall calling convention seems to be standard for many Windows' API
2244 // routines and around. It differs from C calling convention just a little:
2245 // callee should clean up the stack, not caller. Symbols should be also
2246 // decorated in some fancy way :) It doesn't support any vector arguments.
2247 // For info on fast calling convention see Fast Calling Convention (tail call)
2248 // implementation LowerX86_32FastCCCallTo.
2250 /// CallIsStructReturn - Determines whether a call uses struct return
2252 enum StructReturnType {
2257 static StructReturnType
2258 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2260 return NotStructReturn;
2262 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2263 if (!Flags.isSRet())
2264 return NotStructReturn;
2265 if (Flags.isInReg())
2266 return RegStructReturn;
2267 return StackStructReturn;
2270 /// Determines whether a function uses struct return semantics.
2271 static StructReturnType
2272 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2274 return NotStructReturn;
2276 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2277 if (!Flags.isSRet())
2278 return NotStructReturn;
2279 if (Flags.isInReg())
2280 return RegStructReturn;
2281 return StackStructReturn;
2284 /// Make a copy of an aggregate at address specified by "Src" to address
2285 /// "Dst" with size and alignment information specified by the specific
2286 /// parameter attribute. The copy will be passed as a byval function parameter.
2288 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2289 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2291 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2293 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2294 /*isVolatile*/false, /*AlwaysInline=*/true,
2295 MachinePointerInfo(), MachinePointerInfo());
2298 /// Return true if the calling convention is one that
2299 /// supports tail call optimization.
2300 static bool IsTailCallConvention(CallingConv::ID CC) {
2301 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2302 CC == CallingConv::HiPE);
2305 /// \brief Return true if the calling convention is a C calling convention.
2306 static bool IsCCallConvention(CallingConv::ID CC) {
2307 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2308 CC == CallingConv::X86_64_SysV);
2311 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2312 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2316 CallingConv::ID CalleeCC = CS.getCallingConv();
2317 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2323 /// Return true if the function is being made into
2324 /// a tailcall target by changing its ABI.
2325 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2326 bool GuaranteedTailCallOpt) {
2327 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2331 X86TargetLowering::LowerMemArgument(SDValue Chain,
2332 CallingConv::ID CallConv,
2333 const SmallVectorImpl<ISD::InputArg> &Ins,
2334 SDLoc dl, SelectionDAG &DAG,
2335 const CCValAssign &VA,
2336 MachineFrameInfo *MFI,
2338 // Create the nodes corresponding to a load from this parameter slot.
2339 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2340 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2341 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2342 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2345 // If value is passed by pointer we have address passed instead of the value
2347 if (VA.getLocInfo() == CCValAssign::Indirect)
2348 ValVT = VA.getLocVT();
2350 ValVT = VA.getValVT();
2352 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2353 // changed with more analysis.
2354 // In case of tail call optimization mark all arguments mutable. Since they
2355 // could be overwritten by lowering of arguments in case of a tail call.
2356 if (Flags.isByVal()) {
2357 unsigned Bytes = Flags.getByValSize();
2358 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2359 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2360 return DAG.getFrameIndex(FI, getPointerTy());
2362 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2363 VA.getLocMemOffset(), isImmutable);
2364 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2365 return DAG.getLoad(ValVT, dl, Chain, FIN,
2366 MachinePointerInfo::getFixedStack(FI),
2367 false, false, false, 0);
2371 // FIXME: Get this from tablegen.
2372 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2373 const X86Subtarget *Subtarget) {
2374 assert(Subtarget->is64Bit());
2376 if (Subtarget->isCallingConvWin64(CallConv)) {
2377 static const MCPhysReg GPR64ArgRegsWin64[] = {
2378 X86::RCX, X86::RDX, X86::R8, X86::R9
2380 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2383 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2384 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2386 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2389 // FIXME: Get this from tablegen.
2390 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2391 CallingConv::ID CallConv,
2392 const X86Subtarget *Subtarget) {
2393 assert(Subtarget->is64Bit());
2394 if (Subtarget->isCallingConvWin64(CallConv)) {
2395 // The XMM registers which might contain var arg parameters are shadowed
2396 // in their paired GPR. So we only need to save the GPR to their home
2398 // TODO: __vectorcall will change this.
2402 const Function *Fn = MF.getFunction();
2403 bool NoImplicitFloatOps = Fn->getAttributes().
2404 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
2405 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2406 "SSE register cannot be used when SSE is disabled!");
2407 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2408 !Subtarget->hasSSE1())
2409 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2413 static const MCPhysReg XMMArgRegs64Bit[] = {
2414 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2415 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2417 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2421 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2422 CallingConv::ID CallConv,
2424 const SmallVectorImpl<ISD::InputArg> &Ins,
2427 SmallVectorImpl<SDValue> &InVals)
2429 MachineFunction &MF = DAG.getMachineFunction();
2430 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2432 const Function* Fn = MF.getFunction();
2433 if (Fn->hasExternalLinkage() &&
2434 Subtarget->isTargetCygMing() &&
2435 Fn->getName() == "main")
2436 FuncInfo->setForceFramePointer(true);
2438 MachineFrameInfo *MFI = MF.getFrameInfo();
2439 bool Is64Bit = Subtarget->is64Bit();
2440 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2442 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2443 "Var args not supported with calling convention fastcc, ghc or hipe");
2445 // Assign locations to all of the incoming arguments.
2446 SmallVector<CCValAssign, 16> ArgLocs;
2447 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2449 // Allocate shadow area for Win64
2451 CCInfo.AllocateStack(32, 8);
2453 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2455 unsigned LastVal = ~0U;
2457 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2458 CCValAssign &VA = ArgLocs[i];
2459 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2461 assert(VA.getValNo() != LastVal &&
2462 "Don't support value assigned to multiple locs yet");
2464 LastVal = VA.getValNo();
2466 if (VA.isRegLoc()) {
2467 EVT RegVT = VA.getLocVT();
2468 const TargetRegisterClass *RC;
2469 if (RegVT == MVT::i32)
2470 RC = &X86::GR32RegClass;
2471 else if (Is64Bit && RegVT == MVT::i64)
2472 RC = &X86::GR64RegClass;
2473 else if (RegVT == MVT::f32)
2474 RC = &X86::FR32RegClass;
2475 else if (RegVT == MVT::f64)
2476 RC = &X86::FR64RegClass;
2477 else if (RegVT.is512BitVector())
2478 RC = &X86::VR512RegClass;
2479 else if (RegVT.is256BitVector())
2480 RC = &X86::VR256RegClass;
2481 else if (RegVT.is128BitVector())
2482 RC = &X86::VR128RegClass;
2483 else if (RegVT == MVT::x86mmx)
2484 RC = &X86::VR64RegClass;
2485 else if (RegVT == MVT::i1)
2486 RC = &X86::VK1RegClass;
2487 else if (RegVT == MVT::v8i1)
2488 RC = &X86::VK8RegClass;
2489 else if (RegVT == MVT::v16i1)
2490 RC = &X86::VK16RegClass;
2491 else if (RegVT == MVT::v32i1)
2492 RC = &X86::VK32RegClass;
2493 else if (RegVT == MVT::v64i1)
2494 RC = &X86::VK64RegClass;
2496 llvm_unreachable("Unknown argument type!");
2498 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2499 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2501 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2502 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2504 if (VA.getLocInfo() == CCValAssign::SExt)
2505 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2506 DAG.getValueType(VA.getValVT()));
2507 else if (VA.getLocInfo() == CCValAssign::ZExt)
2508 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2509 DAG.getValueType(VA.getValVT()));
2510 else if (VA.getLocInfo() == CCValAssign::BCvt)
2511 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2513 if (VA.isExtInLoc()) {
2514 // Handle MMX values passed in XMM regs.
2515 if (RegVT.isVector())
2516 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2518 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2521 assert(VA.isMemLoc());
2522 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2525 // If value is passed via pointer - do a load.
2526 if (VA.getLocInfo() == CCValAssign::Indirect)
2527 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2528 MachinePointerInfo(), false, false, false, 0);
2530 InVals.push_back(ArgValue);
2533 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2534 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2535 // The x86-64 ABIs require that for returning structs by value we copy
2536 // the sret argument into %rax/%eax (depending on ABI) for the return.
2537 // Win32 requires us to put the sret argument to %eax as well.
2538 // Save the argument into a virtual register so that we can access it
2539 // from the return points.
2540 if (Ins[i].Flags.isSRet()) {
2541 unsigned Reg = FuncInfo->getSRetReturnReg();
2543 MVT PtrTy = getPointerTy();
2544 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2545 FuncInfo->setSRetReturnReg(Reg);
2547 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2548 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2554 unsigned StackSize = CCInfo.getNextStackOffset();
2555 // Align stack specially for tail calls.
2556 if (FuncIsMadeTailCallSafe(CallConv,
2557 MF.getTarget().Options.GuaranteedTailCallOpt))
2558 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2560 // If the function takes variable number of arguments, make a frame index for
2561 // the start of the first vararg value... for expansion of llvm.va_start. We
2562 // can skip this if there are no va_start calls.
2563 if (MFI->hasVAStart() &&
2564 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2565 CallConv != CallingConv::X86_ThisCall))) {
2566 FuncInfo->setVarArgsFrameIndex(
2567 MFI->CreateFixedObject(1, StackSize, true));
2570 // Figure out if XMM registers are in use.
2571 assert(!(MF.getTarget().Options.UseSoftFloat &&
2572 Fn->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2573 Attribute::NoImplicitFloat)) &&
2574 "SSE register cannot be used when SSE is disabled!");
2576 // 64-bit calling conventions support varargs and register parameters, so we
2577 // have to do extra work to spill them in the prologue.
2578 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2579 // Find the first unallocated argument registers.
2580 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2581 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2582 unsigned NumIntRegs =
2583 CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size());
2584 unsigned NumXMMRegs =
2585 CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size());
2586 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2587 "SSE register cannot be used when SSE is disabled!");
2589 // Gather all the live in physical registers.
2590 SmallVector<SDValue, 6> LiveGPRs;
2591 SmallVector<SDValue, 8> LiveXMMRegs;
2593 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2594 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2596 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2598 if (!ArgXMMs.empty()) {
2599 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2600 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2601 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2602 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2603 LiveXMMRegs.push_back(
2604 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2609 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2610 // Get to the caller-allocated home save location. Add 8 to account
2611 // for the return address.
2612 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2613 FuncInfo->setRegSaveFrameIndex(
2614 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2615 // Fixup to set vararg frame on shadow area (4 x i64).
2617 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2619 // For X86-64, if there are vararg parameters that are passed via
2620 // registers, then we must store them to their spots on the stack so
2621 // they may be loaded by deferencing the result of va_next.
2622 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2623 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2624 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2625 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2628 // Store the integer parameter registers.
2629 SmallVector<SDValue, 8> MemOps;
2630 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2632 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2633 for (SDValue Val : LiveGPRs) {
2634 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2635 DAG.getIntPtrConstant(Offset));
2637 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2638 MachinePointerInfo::getFixedStack(
2639 FuncInfo->getRegSaveFrameIndex(), Offset),
2641 MemOps.push_back(Store);
2645 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2646 // Now store the XMM (fp + vector) parameter registers.
2647 SmallVector<SDValue, 12> SaveXMMOps;
2648 SaveXMMOps.push_back(Chain);
2649 SaveXMMOps.push_back(ALVal);
2650 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2651 FuncInfo->getRegSaveFrameIndex()));
2652 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2653 FuncInfo->getVarArgsFPOffset()));
2654 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2656 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2657 MVT::Other, SaveXMMOps));
2660 if (!MemOps.empty())
2661 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2664 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2665 // Find the largest legal vector type.
2666 MVT VecVT = MVT::Other;
2667 // FIXME: Only some x86_32 calling conventions support AVX512.
2668 if (Subtarget->hasAVX512() &&
2669 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2670 CallConv == CallingConv::Intel_OCL_BI)))
2671 VecVT = MVT::v16f32;
2672 else if (Subtarget->hasAVX())
2674 else if (Subtarget->hasSSE2())
2677 // We forward some GPRs and some vector types.
2678 SmallVector<MVT, 2> RegParmTypes;
2679 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2680 RegParmTypes.push_back(IntVT);
2681 if (VecVT != MVT::Other)
2682 RegParmTypes.push_back(VecVT);
2684 // Compute the set of forwarded registers. The rest are scratch.
2685 SmallVectorImpl<ForwardedRegister> &Forwards =
2686 FuncInfo->getForwardedMustTailRegParms();
2687 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2689 // Conservatively forward AL on x86_64, since it might be used for varargs.
2690 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2691 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2692 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2695 // Copy all forwards from physical to virtual registers.
2696 for (ForwardedRegister &F : Forwards) {
2697 // FIXME: Can we use a less constrained schedule?
2698 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2699 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2700 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2704 // Some CCs need callee pop.
2705 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2706 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2707 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2709 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2710 // If this is an sret function, the return should pop the hidden pointer.
2711 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2712 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2713 argsAreStructReturn(Ins) == StackStructReturn)
2714 FuncInfo->setBytesToPopOnReturn(4);
2718 // RegSaveFrameIndex is X86-64 only.
2719 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2720 if (CallConv == CallingConv::X86_FastCall ||
2721 CallConv == CallingConv::X86_ThisCall)
2722 // fastcc functions can't have varargs.
2723 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2726 FuncInfo->setArgumentStackSize(StackSize);
2732 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2733 SDValue StackPtr, SDValue Arg,
2734 SDLoc dl, SelectionDAG &DAG,
2735 const CCValAssign &VA,
2736 ISD::ArgFlagsTy Flags) const {
2737 unsigned LocMemOffset = VA.getLocMemOffset();
2738 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2739 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2740 if (Flags.isByVal())
2741 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2743 return DAG.getStore(Chain, dl, Arg, PtrOff,
2744 MachinePointerInfo::getStack(LocMemOffset),
2748 /// Emit a load of return address if tail call
2749 /// optimization is performed and it is required.
2751 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2752 SDValue &OutRetAddr, SDValue Chain,
2753 bool IsTailCall, bool Is64Bit,
2754 int FPDiff, SDLoc dl) const {
2755 // Adjust the Return address stack slot.
2756 EVT VT = getPointerTy();
2757 OutRetAddr = getReturnAddressFrameIndex(DAG);
2759 // Load the "old" Return address.
2760 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2761 false, false, false, 0);
2762 return SDValue(OutRetAddr.getNode(), 1);
2765 /// Emit a store of the return address if tail call
2766 /// optimization is performed and it is required (FPDiff!=0).
2767 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2768 SDValue Chain, SDValue RetAddrFrIdx,
2769 EVT PtrVT, unsigned SlotSize,
2770 int FPDiff, SDLoc dl) {
2771 // Store the return address to the appropriate stack slot.
2772 if (!FPDiff) return Chain;
2773 // Calculate the new stack slot for the return address.
2774 int NewReturnAddrFI =
2775 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2777 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2778 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2779 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2785 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2786 SmallVectorImpl<SDValue> &InVals) const {
2787 SelectionDAG &DAG = CLI.DAG;
2789 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2790 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2791 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2792 SDValue Chain = CLI.Chain;
2793 SDValue Callee = CLI.Callee;
2794 CallingConv::ID CallConv = CLI.CallConv;
2795 bool &isTailCall = CLI.IsTailCall;
2796 bool isVarArg = CLI.IsVarArg;
2798 MachineFunction &MF = DAG.getMachineFunction();
2799 bool Is64Bit = Subtarget->is64Bit();
2800 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2801 StructReturnType SR = callIsStructReturn(Outs);
2802 bool IsSibcall = false;
2803 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2805 if (MF.getTarget().Options.DisableTailCalls)
2808 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2810 // Force this to be a tail call. The verifier rules are enough to ensure
2811 // that we can lower this successfully without moving the return address
2814 } else if (isTailCall) {
2815 // Check if it's really possible to do a tail call.
2816 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2817 isVarArg, SR != NotStructReturn,
2818 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2819 Outs, OutVals, Ins, DAG);
2821 // Sibcalls are automatically detected tailcalls which do not require
2823 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2830 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2831 "Var args not supported with calling convention fastcc, ghc or hipe");
2833 // Analyze operands of the call, assigning locations to each operand.
2834 SmallVector<CCValAssign, 16> ArgLocs;
2835 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2837 // Allocate shadow area for Win64
2839 CCInfo.AllocateStack(32, 8);
2841 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2843 // Get a count of how many bytes are to be pushed on the stack.
2844 unsigned NumBytes = CCInfo.getNextStackOffset();
2846 // This is a sibcall. The memory operands are available in caller's
2847 // own caller's stack.
2849 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2850 IsTailCallConvention(CallConv))
2851 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2854 if (isTailCall && !IsSibcall && !IsMustTail) {
2855 // Lower arguments at fp - stackoffset + fpdiff.
2856 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2858 FPDiff = NumBytesCallerPushed - NumBytes;
2860 // Set the delta of movement of the returnaddr stackslot.
2861 // But only set if delta is greater than previous delta.
2862 if (FPDiff < X86Info->getTCReturnAddrDelta())
2863 X86Info->setTCReturnAddrDelta(FPDiff);
2866 unsigned NumBytesToPush = NumBytes;
2867 unsigned NumBytesToPop = NumBytes;
2869 // If we have an inalloca argument, all stack space has already been allocated
2870 // for us and be right at the top of the stack. We don't support multiple
2871 // arguments passed in memory when using inalloca.
2872 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2874 if (!ArgLocs.back().isMemLoc())
2875 report_fatal_error("cannot use inalloca attribute on a register "
2877 if (ArgLocs.back().getLocMemOffset() != 0)
2878 report_fatal_error("any parameter with the inalloca attribute must be "
2879 "the only memory argument");
2883 Chain = DAG.getCALLSEQ_START(
2884 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2886 SDValue RetAddrFrIdx;
2887 // Load return address for tail calls.
2888 if (isTailCall && FPDiff)
2889 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2890 Is64Bit, FPDiff, dl);
2892 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2893 SmallVector<SDValue, 8> MemOpChains;
2896 // Walk the register/memloc assignments, inserting copies/loads. In the case
2897 // of tail call optimization arguments are handle later.
2898 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2899 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2900 // Skip inalloca arguments, they have already been written.
2901 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2902 if (Flags.isInAlloca())
2905 CCValAssign &VA = ArgLocs[i];
2906 EVT RegVT = VA.getLocVT();
2907 SDValue Arg = OutVals[i];
2908 bool isByVal = Flags.isByVal();
2910 // Promote the value if needed.
2911 switch (VA.getLocInfo()) {
2912 default: llvm_unreachable("Unknown loc info!");
2913 case CCValAssign::Full: break;
2914 case CCValAssign::SExt:
2915 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2917 case CCValAssign::ZExt:
2918 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2920 case CCValAssign::AExt:
2921 if (RegVT.is128BitVector()) {
2922 // Special case: passing MMX values in XMM registers.
2923 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2924 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2925 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2927 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2929 case CCValAssign::BCvt:
2930 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2932 case CCValAssign::Indirect: {
2933 // Store the argument.
2934 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2935 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2936 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2937 MachinePointerInfo::getFixedStack(FI),
2944 if (VA.isRegLoc()) {
2945 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2946 if (isVarArg && IsWin64) {
2947 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2948 // shadow reg if callee is a varargs function.
2949 unsigned ShadowReg = 0;
2950 switch (VA.getLocReg()) {
2951 case X86::XMM0: ShadowReg = X86::RCX; break;
2952 case X86::XMM1: ShadowReg = X86::RDX; break;
2953 case X86::XMM2: ShadowReg = X86::R8; break;
2954 case X86::XMM3: ShadowReg = X86::R9; break;
2957 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2959 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2960 assert(VA.isMemLoc());
2961 if (!StackPtr.getNode())
2962 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2964 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2965 dl, DAG, VA, Flags));
2969 if (!MemOpChains.empty())
2970 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2972 if (Subtarget->isPICStyleGOT()) {
2973 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2976 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2977 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2979 // If we are tail calling and generating PIC/GOT style code load the
2980 // address of the callee into ECX. The value in ecx is used as target of
2981 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2982 // for tail calls on PIC/GOT architectures. Normally we would just put the
2983 // address of GOT into ebx and then call target@PLT. But for tail calls
2984 // ebx would be restored (since ebx is callee saved) before jumping to the
2987 // Note: The actual moving to ECX is done further down.
2988 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2989 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2990 !G->getGlobal()->hasProtectedVisibility())
2991 Callee = LowerGlobalAddress(Callee, DAG);
2992 else if (isa<ExternalSymbolSDNode>(Callee))
2993 Callee = LowerExternalSymbol(Callee, DAG);
2997 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2998 // From AMD64 ABI document:
2999 // For calls that may call functions that use varargs or stdargs
3000 // (prototype-less calls or calls to functions containing ellipsis (...) in
3001 // the declaration) %al is used as hidden argument to specify the number
3002 // of SSE registers used. The contents of %al do not need to match exactly
3003 // the number of registers, but must be an ubound on the number of SSE
3004 // registers used and is in the range 0 - 8 inclusive.
3006 // Count the number of XMM registers allocated.
3007 static const MCPhysReg XMMArgRegs[] = {
3008 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3009 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3011 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
3012 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3013 && "SSE registers cannot be used when SSE is disabled");
3015 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3016 DAG.getConstant(NumXMMRegs, MVT::i8)));
3019 if (isVarArg && IsMustTail) {
3020 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3021 for (const auto &F : Forwards) {
3022 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3023 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3027 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3028 // don't need this because the eligibility check rejects calls that require
3029 // shuffling arguments passed in memory.
3030 if (!IsSibcall && isTailCall) {
3031 // Force all the incoming stack arguments to be loaded from the stack
3032 // before any new outgoing arguments are stored to the stack, because the
3033 // outgoing stack slots may alias the incoming argument stack slots, and
3034 // the alias isn't otherwise explicit. This is slightly more conservative
3035 // than necessary, because it means that each store effectively depends
3036 // on every argument instead of just those arguments it would clobber.
3037 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3039 SmallVector<SDValue, 8> MemOpChains2;
3042 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3043 CCValAssign &VA = ArgLocs[i];
3046 assert(VA.isMemLoc());
3047 SDValue Arg = OutVals[i];
3048 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3049 // Skip inalloca arguments. They don't require any work.
3050 if (Flags.isInAlloca())
3052 // Create frame index.
3053 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3054 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3055 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3056 FIN = DAG.getFrameIndex(FI, getPointerTy());
3058 if (Flags.isByVal()) {
3059 // Copy relative to framepointer.
3060 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3061 if (!StackPtr.getNode())
3062 StackPtr = DAG.getCopyFromReg(Chain, dl,
3063 RegInfo->getStackRegister(),
3065 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3067 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3071 // Store relative to framepointer.
3072 MemOpChains2.push_back(
3073 DAG.getStore(ArgChain, dl, Arg, FIN,
3074 MachinePointerInfo::getFixedStack(FI),
3079 if (!MemOpChains2.empty())
3080 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3082 // Store the return address to the appropriate stack slot.
3083 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3084 getPointerTy(), RegInfo->getSlotSize(),
3088 // Build a sequence of copy-to-reg nodes chained together with token chain
3089 // and flag operands which copy the outgoing args into registers.
3091 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3092 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3093 RegsToPass[i].second, InFlag);
3094 InFlag = Chain.getValue(1);
3097 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3098 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3099 // In the 64-bit large code model, we have to make all calls
3100 // through a register, since the call instruction's 32-bit
3101 // pc-relative offset may not be large enough to hold the whole
3103 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3104 // If the callee is a GlobalAddress node (quite common, every direct call
3105 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3107 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3109 // We should use extra load for direct calls to dllimported functions in
3111 const GlobalValue *GV = G->getGlobal();
3112 if (!GV->hasDLLImportStorageClass()) {
3113 unsigned char OpFlags = 0;
3114 bool ExtraLoad = false;
3115 unsigned WrapperKind = ISD::DELETED_NODE;
3117 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3118 // external symbols most go through the PLT in PIC mode. If the symbol
3119 // has hidden or protected visibility, or if it is static or local, then
3120 // we don't need to use the PLT - we can directly call it.
3121 if (Subtarget->isTargetELF() &&
3122 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3123 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3124 OpFlags = X86II::MO_PLT;
3125 } else if (Subtarget->isPICStyleStubAny() &&
3126 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3127 (!Subtarget->getTargetTriple().isMacOSX() ||
3128 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3129 // PC-relative references to external symbols should go through $stub,
3130 // unless we're building with the leopard linker or later, which
3131 // automatically synthesizes these stubs.
3132 OpFlags = X86II::MO_DARWIN_STUB;
3133 } else if (Subtarget->isPICStyleRIPRel() &&
3134 isa<Function>(GV) &&
3135 cast<Function>(GV)->getAttributes().
3136 hasAttribute(AttributeSet::FunctionIndex,
3137 Attribute::NonLazyBind)) {
3138 // If the function is marked as non-lazy, generate an indirect call
3139 // which loads from the GOT directly. This avoids runtime overhead
3140 // at the cost of eager binding (and one extra byte of encoding).
3141 OpFlags = X86II::MO_GOTPCREL;
3142 WrapperKind = X86ISD::WrapperRIP;
3146 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3147 G->getOffset(), OpFlags);
3149 // Add a wrapper if needed.
3150 if (WrapperKind != ISD::DELETED_NODE)
3151 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3152 // Add extra indirection if needed.
3154 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3155 MachinePointerInfo::getGOT(),
3156 false, false, false, 0);
3158 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3159 unsigned char OpFlags = 0;
3161 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3162 // external symbols should go through the PLT.
3163 if (Subtarget->isTargetELF() &&
3164 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3165 OpFlags = X86II::MO_PLT;
3166 } else if (Subtarget->isPICStyleStubAny() &&
3167 (!Subtarget->getTargetTriple().isMacOSX() ||
3168 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3169 // PC-relative references to external symbols should go through $stub,
3170 // unless we're building with the leopard linker or later, which
3171 // automatically synthesizes these stubs.
3172 OpFlags = X86II::MO_DARWIN_STUB;
3175 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3177 } else if (Subtarget->isTarget64BitILP32() &&
3178 Callee->getValueType(0) == MVT::i32) {
3179 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3180 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3183 // Returns a chain & a flag for retval copy to use.
3184 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3185 SmallVector<SDValue, 8> Ops;
3187 if (!IsSibcall && isTailCall) {
3188 Chain = DAG.getCALLSEQ_END(Chain,
3189 DAG.getIntPtrConstant(NumBytesToPop, true),
3190 DAG.getIntPtrConstant(0, true), InFlag, dl);
3191 InFlag = Chain.getValue(1);
3194 Ops.push_back(Chain);
3195 Ops.push_back(Callee);
3198 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3200 // Add argument registers to the end of the list so that they are known live
3202 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3203 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3204 RegsToPass[i].second.getValueType()));
3206 // Add a register mask operand representing the call-preserved registers.
3207 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3208 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3209 assert(Mask && "Missing call preserved mask for calling convention");
3210 Ops.push_back(DAG.getRegisterMask(Mask));
3212 if (InFlag.getNode())
3213 Ops.push_back(InFlag);
3217 //// If this is the first return lowered for this function, add the regs
3218 //// to the liveout set for the function.
3219 // This isn't right, although it's probably harmless on x86; liveouts
3220 // should be computed from returns not tail calls. Consider a void
3221 // function making a tail call to a function returning int.
3222 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3225 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3226 InFlag = Chain.getValue(1);
3228 // Create the CALLSEQ_END node.
3229 unsigned NumBytesForCalleeToPop;
3230 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3231 DAG.getTarget().Options.GuaranteedTailCallOpt))
3232 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3233 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3234 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3235 SR == StackStructReturn)
3236 // If this is a call to a struct-return function, the callee
3237 // pops the hidden struct pointer, so we have to push it back.
3238 // This is common for Darwin/X86, Linux & Mingw32 targets.
3239 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3240 NumBytesForCalleeToPop = 4;
3242 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3244 // Returns a flag for retval copy to use.
3246 Chain = DAG.getCALLSEQ_END(Chain,
3247 DAG.getIntPtrConstant(NumBytesToPop, true),
3248 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3251 InFlag = Chain.getValue(1);
3254 // Handle result values, copying them out of physregs into vregs that we
3256 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3257 Ins, dl, DAG, InVals);
3260 //===----------------------------------------------------------------------===//
3261 // Fast Calling Convention (tail call) implementation
3262 //===----------------------------------------------------------------------===//
3264 // Like std call, callee cleans arguments, convention except that ECX is
3265 // reserved for storing the tail called function address. Only 2 registers are
3266 // free for argument passing (inreg). Tail call optimization is performed
3268 // * tailcallopt is enabled
3269 // * caller/callee are fastcc
3270 // On X86_64 architecture with GOT-style position independent code only local
3271 // (within module) calls are supported at the moment.
3272 // To keep the stack aligned according to platform abi the function
3273 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3274 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3275 // If a tail called function callee has more arguments than the caller the
3276 // caller needs to make sure that there is room to move the RETADDR to. This is
3277 // achieved by reserving an area the size of the argument delta right after the
3278 // original RETADDR, but before the saved framepointer or the spilled registers
3279 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3291 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3292 /// for a 16 byte align requirement.
3294 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3295 SelectionDAG& DAG) const {
3296 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3297 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3298 unsigned StackAlignment = TFI.getStackAlignment();
3299 uint64_t AlignMask = StackAlignment - 1;
3300 int64_t Offset = StackSize;
3301 unsigned SlotSize = RegInfo->getSlotSize();
3302 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3303 // Number smaller than 12 so just add the difference.
3304 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3306 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3307 Offset = ((~AlignMask) & Offset) + StackAlignment +
3308 (StackAlignment-SlotSize);
3313 /// MatchingStackOffset - Return true if the given stack call argument is
3314 /// already available in the same position (relatively) of the caller's
3315 /// incoming argument stack.
3317 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3318 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3319 const X86InstrInfo *TII) {
3320 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3322 if (Arg.getOpcode() == ISD::CopyFromReg) {
3323 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3324 if (!TargetRegisterInfo::isVirtualRegister(VR))
3326 MachineInstr *Def = MRI->getVRegDef(VR);
3329 if (!Flags.isByVal()) {
3330 if (!TII->isLoadFromStackSlot(Def, FI))
3333 unsigned Opcode = Def->getOpcode();
3334 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3335 Opcode == X86::LEA64_32r) &&
3336 Def->getOperand(1).isFI()) {
3337 FI = Def->getOperand(1).getIndex();
3338 Bytes = Flags.getByValSize();
3342 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3343 if (Flags.isByVal())
3344 // ByVal argument is passed in as a pointer but it's now being
3345 // dereferenced. e.g.
3346 // define @foo(%struct.X* %A) {
3347 // tail call @bar(%struct.X* byval %A)
3350 SDValue Ptr = Ld->getBasePtr();
3351 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3354 FI = FINode->getIndex();
3355 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3356 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3357 FI = FINode->getIndex();
3358 Bytes = Flags.getByValSize();
3362 assert(FI != INT_MAX);
3363 if (!MFI->isFixedObjectIndex(FI))
3365 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3368 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3369 /// for tail call optimization. Targets which want to do tail call
3370 /// optimization should implement this function.
3372 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3373 CallingConv::ID CalleeCC,
3375 bool isCalleeStructRet,
3376 bool isCallerStructRet,
3378 const SmallVectorImpl<ISD::OutputArg> &Outs,
3379 const SmallVectorImpl<SDValue> &OutVals,
3380 const SmallVectorImpl<ISD::InputArg> &Ins,
3381 SelectionDAG &DAG) const {
3382 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3385 // If -tailcallopt is specified, make fastcc functions tail-callable.
3386 const MachineFunction &MF = DAG.getMachineFunction();
3387 const Function *CallerF = MF.getFunction();
3389 // If the function return type is x86_fp80 and the callee return type is not,
3390 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3391 // perform a tailcall optimization here.
3392 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3395 CallingConv::ID CallerCC = CallerF->getCallingConv();
3396 bool CCMatch = CallerCC == CalleeCC;
3397 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3398 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3400 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3401 if (IsTailCallConvention(CalleeCC) && CCMatch)
3406 // Look for obvious safe cases to perform tail call optimization that do not
3407 // require ABI changes. This is what gcc calls sibcall.
3409 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3410 // emit a special epilogue.
3411 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3412 if (RegInfo->needsStackRealignment(MF))
3415 // Also avoid sibcall optimization if either caller or callee uses struct
3416 // return semantics.
3417 if (isCalleeStructRet || isCallerStructRet)
3420 // An stdcall/thiscall caller is expected to clean up its arguments; the
3421 // callee isn't going to do that.
3422 // FIXME: this is more restrictive than needed. We could produce a tailcall
3423 // when the stack adjustment matches. For example, with a thiscall that takes
3424 // only one argument.
3425 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3426 CallerCC == CallingConv::X86_ThisCall))
3429 // Do not sibcall optimize vararg calls unless all arguments are passed via
3431 if (isVarArg && !Outs.empty()) {
3433 // Optimizing for varargs on Win64 is unlikely to be safe without
3434 // additional testing.
3435 if (IsCalleeWin64 || IsCallerWin64)
3438 SmallVector<CCValAssign, 16> ArgLocs;
3439 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3442 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3443 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3444 if (!ArgLocs[i].isRegLoc())
3448 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3449 // stack. Therefore, if it's not used by the call it is not safe to optimize
3450 // this into a sibcall.
3451 bool Unused = false;
3452 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3459 SmallVector<CCValAssign, 16> RVLocs;
3460 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3462 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3463 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3464 CCValAssign &VA = RVLocs[i];
3465 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3470 // If the calling conventions do not match, then we'd better make sure the
3471 // results are returned in the same way as what the caller expects.
3473 SmallVector<CCValAssign, 16> RVLocs1;
3474 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3476 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3478 SmallVector<CCValAssign, 16> RVLocs2;
3479 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3481 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3483 if (RVLocs1.size() != RVLocs2.size())
3485 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3486 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3488 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3490 if (RVLocs1[i].isRegLoc()) {
3491 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3494 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3500 // If the callee takes no arguments then go on to check the results of the
3502 if (!Outs.empty()) {
3503 // Check if stack adjustment is needed. For now, do not do this if any
3504 // argument is passed on the stack.
3505 SmallVector<CCValAssign, 16> ArgLocs;
3506 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3509 // Allocate shadow area for Win64
3511 CCInfo.AllocateStack(32, 8);
3513 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3514 if (CCInfo.getNextStackOffset()) {
3515 MachineFunction &MF = DAG.getMachineFunction();
3516 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3519 // Check if the arguments are already laid out in the right way as
3520 // the caller's fixed stack objects.
3521 MachineFrameInfo *MFI = MF.getFrameInfo();
3522 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3523 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3524 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3525 CCValAssign &VA = ArgLocs[i];
3526 SDValue Arg = OutVals[i];
3527 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3528 if (VA.getLocInfo() == CCValAssign::Indirect)
3530 if (!VA.isRegLoc()) {
3531 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3538 // If the tailcall address may be in a register, then make sure it's
3539 // possible to register allocate for it. In 32-bit, the call address can
3540 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3541 // callee-saved registers are restored. These happen to be the same
3542 // registers used to pass 'inreg' arguments so watch out for those.
3543 if (!Subtarget->is64Bit() &&
3544 ((!isa<GlobalAddressSDNode>(Callee) &&
3545 !isa<ExternalSymbolSDNode>(Callee)) ||
3546 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3547 unsigned NumInRegs = 0;
3548 // In PIC we need an extra register to formulate the address computation
3550 unsigned MaxInRegs =
3551 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3553 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3554 CCValAssign &VA = ArgLocs[i];
3557 unsigned Reg = VA.getLocReg();
3560 case X86::EAX: case X86::EDX: case X86::ECX:
3561 if (++NumInRegs == MaxInRegs)
3573 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3574 const TargetLibraryInfo *libInfo) const {
3575 return X86::createFastISel(funcInfo, libInfo);
3578 //===----------------------------------------------------------------------===//
3579 // Other Lowering Hooks
3580 //===----------------------------------------------------------------------===//
3582 static bool MayFoldLoad(SDValue Op) {
3583 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3586 static bool MayFoldIntoStore(SDValue Op) {
3587 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3590 static bool isTargetShuffle(unsigned Opcode) {
3592 default: return false;
3593 case X86ISD::BLENDI:
3594 case X86ISD::PSHUFB:
3595 case X86ISD::PSHUFD:
3596 case X86ISD::PSHUFHW:
3597 case X86ISD::PSHUFLW:
3599 case X86ISD::PALIGNR:
3600 case X86ISD::MOVLHPS:
3601 case X86ISD::MOVLHPD:
3602 case X86ISD::MOVHLPS:
3603 case X86ISD::MOVLPS:
3604 case X86ISD::MOVLPD:
3605 case X86ISD::MOVSHDUP:
3606 case X86ISD::MOVSLDUP:
3607 case X86ISD::MOVDDUP:
3610 case X86ISD::UNPCKL:
3611 case X86ISD::UNPCKH:
3612 case X86ISD::VPERMILPI:
3613 case X86ISD::VPERM2X128:
3614 case X86ISD::VPERMI:
3619 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3620 SDValue V1, SelectionDAG &DAG) {
3622 default: llvm_unreachable("Unknown x86 shuffle node");
3623 case X86ISD::MOVSHDUP:
3624 case X86ISD::MOVSLDUP:
3625 case X86ISD::MOVDDUP:
3626 return DAG.getNode(Opc, dl, VT, V1);
3630 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3631 SDValue V1, unsigned TargetMask,
3632 SelectionDAG &DAG) {
3634 default: llvm_unreachable("Unknown x86 shuffle node");
3635 case X86ISD::PSHUFD:
3636 case X86ISD::PSHUFHW:
3637 case X86ISD::PSHUFLW:
3638 case X86ISD::VPERMILPI:
3639 case X86ISD::VPERMI:
3640 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3644 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3645 SDValue V1, SDValue V2, unsigned TargetMask,
3646 SelectionDAG &DAG) {
3648 default: llvm_unreachable("Unknown x86 shuffle node");
3649 case X86ISD::PALIGNR:
3650 case X86ISD::VALIGN:
3652 case X86ISD::VPERM2X128:
3653 return DAG.getNode(Opc, dl, VT, V1, V2,
3654 DAG.getConstant(TargetMask, MVT::i8));
3658 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3659 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3661 default: llvm_unreachable("Unknown x86 shuffle node");
3662 case X86ISD::MOVLHPS:
3663 case X86ISD::MOVLHPD:
3664 case X86ISD::MOVHLPS:
3665 case X86ISD::MOVLPS:
3666 case X86ISD::MOVLPD:
3669 case X86ISD::UNPCKL:
3670 case X86ISD::UNPCKH:
3671 return DAG.getNode(Opc, dl, VT, V1, V2);
3675 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3676 MachineFunction &MF = DAG.getMachineFunction();
3677 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3678 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3679 int ReturnAddrIndex = FuncInfo->getRAIndex();
3681 if (ReturnAddrIndex == 0) {
3682 // Set up a frame object for the return address.
3683 unsigned SlotSize = RegInfo->getSlotSize();
3684 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3687 FuncInfo->setRAIndex(ReturnAddrIndex);
3690 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3693 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3694 bool hasSymbolicDisplacement) {
3695 // Offset should fit into 32 bit immediate field.
3696 if (!isInt<32>(Offset))
3699 // If we don't have a symbolic displacement - we don't have any extra
3701 if (!hasSymbolicDisplacement)
3704 // FIXME: Some tweaks might be needed for medium code model.
3705 if (M != CodeModel::Small && M != CodeModel::Kernel)
3708 // For small code model we assume that latest object is 16MB before end of 31
3709 // bits boundary. We may also accept pretty large negative constants knowing
3710 // that all objects are in the positive half of address space.
3711 if (M == CodeModel::Small && Offset < 16*1024*1024)
3714 // For kernel code model we know that all object resist in the negative half
3715 // of 32bits address space. We may not accept negative offsets, since they may
3716 // be just off and we may accept pretty large positive ones.
3717 if (M == CodeModel::Kernel && Offset >= 0)
3723 /// isCalleePop - Determines whether the callee is required to pop its
3724 /// own arguments. Callee pop is necessary to support tail calls.
3725 bool X86::isCalleePop(CallingConv::ID CallingConv,
3726 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3727 switch (CallingConv) {
3730 case CallingConv::X86_StdCall:
3731 case CallingConv::X86_FastCall:
3732 case CallingConv::X86_ThisCall:
3734 case CallingConv::Fast:
3735 case CallingConv::GHC:
3736 case CallingConv::HiPE:
3743 /// \brief Return true if the condition is an unsigned comparison operation.
3744 static bool isX86CCUnsigned(unsigned X86CC) {
3746 default: llvm_unreachable("Invalid integer condition!");
3747 case X86::COND_E: return true;
3748 case X86::COND_G: return false;
3749 case X86::COND_GE: return false;
3750 case X86::COND_L: return false;
3751 case X86::COND_LE: return false;
3752 case X86::COND_NE: return true;
3753 case X86::COND_B: return true;
3754 case X86::COND_A: return true;
3755 case X86::COND_BE: return true;
3756 case X86::COND_AE: return true;
3758 llvm_unreachable("covered switch fell through?!");
3761 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3762 /// specific condition code, returning the condition code and the LHS/RHS of the
3763 /// comparison to make.
3764 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3765 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3767 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3768 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3769 // X > -1 -> X == 0, jump !sign.
3770 RHS = DAG.getConstant(0, RHS.getValueType());
3771 return X86::COND_NS;
3773 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3774 // X < 0 -> X == 0, jump on sign.
3777 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3779 RHS = DAG.getConstant(0, RHS.getValueType());
3780 return X86::COND_LE;
3784 switch (SetCCOpcode) {
3785 default: llvm_unreachable("Invalid integer condition!");
3786 case ISD::SETEQ: return X86::COND_E;
3787 case ISD::SETGT: return X86::COND_G;
3788 case ISD::SETGE: return X86::COND_GE;
3789 case ISD::SETLT: return X86::COND_L;
3790 case ISD::SETLE: return X86::COND_LE;
3791 case ISD::SETNE: return X86::COND_NE;
3792 case ISD::SETULT: return X86::COND_B;
3793 case ISD::SETUGT: return X86::COND_A;
3794 case ISD::SETULE: return X86::COND_BE;
3795 case ISD::SETUGE: return X86::COND_AE;
3799 // First determine if it is required or is profitable to flip the operands.
3801 // If LHS is a foldable load, but RHS is not, flip the condition.
3802 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3803 !ISD::isNON_EXTLoad(RHS.getNode())) {
3804 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3805 std::swap(LHS, RHS);
3808 switch (SetCCOpcode) {
3814 std::swap(LHS, RHS);
3818 // On a floating point condition, the flags are set as follows:
3820 // 0 | 0 | 0 | X > Y
3821 // 0 | 0 | 1 | X < Y
3822 // 1 | 0 | 0 | X == Y
3823 // 1 | 1 | 1 | unordered
3824 switch (SetCCOpcode) {
3825 default: llvm_unreachable("Condcode should be pre-legalized away");
3827 case ISD::SETEQ: return X86::COND_E;
3828 case ISD::SETOLT: // flipped
3830 case ISD::SETGT: return X86::COND_A;
3831 case ISD::SETOLE: // flipped
3833 case ISD::SETGE: return X86::COND_AE;
3834 case ISD::SETUGT: // flipped
3836 case ISD::SETLT: return X86::COND_B;
3837 case ISD::SETUGE: // flipped
3839 case ISD::SETLE: return X86::COND_BE;
3841 case ISD::SETNE: return X86::COND_NE;
3842 case ISD::SETUO: return X86::COND_P;
3843 case ISD::SETO: return X86::COND_NP;
3845 case ISD::SETUNE: return X86::COND_INVALID;
3849 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3850 /// code. Current x86 isa includes the following FP cmov instructions:
3851 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3852 static bool hasFPCMov(unsigned X86CC) {
3868 /// isFPImmLegal - Returns true if the target can instruction select the
3869 /// specified FP immediate natively. If false, the legalizer will
3870 /// materialize the FP immediate as a load from a constant pool.
3871 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3872 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3873 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3879 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3880 ISD::LoadExtType ExtTy,
3882 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3883 // relocation target a movq or addq instruction: don't let the load shrink.
3884 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3885 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3886 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3887 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3891 /// \brief Returns true if it is beneficial to convert a load of a constant
3892 /// to just the constant itself.
3893 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3895 assert(Ty->isIntegerTy());
3897 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3898 if (BitSize == 0 || BitSize > 64)
3903 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3904 unsigned Index) const {
3905 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3908 return (Index == 0 || Index == ResVT.getVectorNumElements());
3911 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3912 // Speculate cttz only if we can directly use TZCNT.
3913 return Subtarget->hasBMI();
3916 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3917 // Speculate ctlz only if we can directly use LZCNT.
3918 return Subtarget->hasLZCNT();
3921 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3922 /// the specified range (L, H].
3923 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3924 return (Val < 0) || (Val >= Low && Val < Hi);
3927 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3928 /// specified value.
3929 static bool isUndefOrEqual(int Val, int CmpVal) {
3930 return (Val < 0 || Val == CmpVal);
3933 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3934 /// from position Pos and ending in Pos+Size, falls within the specified
3935 /// sequential range (Low, Low+Size]. or is undef.
3936 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3937 unsigned Pos, unsigned Size, int Low) {
3938 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3939 if (!isUndefOrEqual(Mask[i], Low))
3944 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3945 /// is suitable for input to PSHUFD. That is, it doesn't reference the other
3946 /// operand - by default will match for first operand.
3947 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT,
3948 bool TestSecondOperand = false) {
3949 if (VT != MVT::v4f32 && VT != MVT::v4i32 &&
3950 VT != MVT::v2f64 && VT != MVT::v2i64)
3953 unsigned NumElems = VT.getVectorNumElements();
3954 unsigned Lo = TestSecondOperand ? NumElems : 0;
3955 unsigned Hi = Lo + NumElems;
3957 for (unsigned i = 0; i < NumElems; ++i)
3958 if (!isUndefOrInRange(Mask[i], (int)Lo, (int)Hi))
3964 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3965 /// is suitable for input to PSHUFHW.
3966 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3967 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3970 // Lower quadword copied in order or undef.
3971 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3974 // Upper quadword shuffled.
3975 for (unsigned i = 4; i != 8; ++i)
3976 if (!isUndefOrInRange(Mask[i], 4, 8))
3979 if (VT == MVT::v16i16) {
3980 // Lower quadword copied in order or undef.
3981 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3984 // Upper quadword shuffled.
3985 for (unsigned i = 12; i != 16; ++i)
3986 if (!isUndefOrInRange(Mask[i], 12, 16))
3993 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3994 /// is suitable for input to PSHUFLW.
3995 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3996 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3999 // Upper quadword copied in order.
4000 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
4003 // Lower quadword shuffled.
4004 for (unsigned i = 0; i != 4; ++i)
4005 if (!isUndefOrInRange(Mask[i], 0, 4))
4008 if (VT == MVT::v16i16) {
4009 // Upper quadword copied in order.
4010 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
4013 // Lower quadword shuffled.
4014 for (unsigned i = 8; i != 12; ++i)
4015 if (!isUndefOrInRange(Mask[i], 8, 12))
4022 /// \brief Return true if the mask specifies a shuffle of elements that is
4023 /// suitable for input to intralane (palignr) or interlane (valign) vector
4025 static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) {
4026 unsigned NumElts = VT.getVectorNumElements();
4027 unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128;
4028 unsigned NumLaneElts = NumElts/NumLanes;
4030 // Do not handle 64-bit element shuffles with palignr.
4031 if (NumLaneElts == 2)
4034 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
4036 for (i = 0; i != NumLaneElts; ++i) {
4041 // Lane is all undef, go to next lane
4042 if (i == NumLaneElts)
4045 int Start = Mask[i+l];
4047 // Make sure its in this lane in one of the sources
4048 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
4049 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
4052 // If not lane 0, then we must match lane 0
4053 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
4056 // Correct second source to be contiguous with first source
4057 if (Start >= (int)NumElts)
4058 Start -= NumElts - NumLaneElts;
4060 // Make sure we're shifting in the right direction.
4061 if (Start <= (int)(i+l))
4066 // Check the rest of the elements to see if they are consecutive.
4067 for (++i; i != NumLaneElts; ++i) {
4068 int Idx = Mask[i+l];
4070 // Make sure its in this lane
4071 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
4072 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
4075 // If not lane 0, then we must match lane 0
4076 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
4079 if (Idx >= (int)NumElts)
4080 Idx -= NumElts - NumLaneElts;
4082 if (!isUndefOrEqual(Idx, Start+i))
4091 /// \brief Return true if the node specifies a shuffle of elements that is
4092 /// suitable for input to PALIGNR.
4093 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
4094 const X86Subtarget *Subtarget) {
4095 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
4096 (VT.is256BitVector() && !Subtarget->hasInt256()) ||
4097 VT.is512BitVector())
4098 // FIXME: Add AVX512BW.
4101 return isAlignrMask(Mask, VT, false);
4104 /// \brief Return true if the node specifies a shuffle of elements that is
4105 /// suitable for input to VALIGN.
4106 static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT,
4107 const X86Subtarget *Subtarget) {
4108 // FIXME: Add AVX512VL.
4109 if (!VT.is512BitVector() || !Subtarget->hasAVX512())
4111 return isAlignrMask(Mask, VT, true);
4114 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4115 /// the two vector operands have swapped position.
4116 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
4117 unsigned NumElems) {
4118 for (unsigned i = 0; i != NumElems; ++i) {
4122 else if (idx < (int)NumElems)
4123 Mask[i] = idx + NumElems;
4125 Mask[i] = idx - NumElems;
4129 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
4130 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
4131 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
4132 /// reverse of what x86 shuffles want.
4133 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
4135 unsigned NumElems = VT.getVectorNumElements();
4136 unsigned NumLanes = VT.getSizeInBits()/128;
4137 unsigned NumLaneElems = NumElems/NumLanes;
4139 if (NumLaneElems != 2 && NumLaneElems != 4)
4142 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4143 bool symetricMaskRequired =
4144 (VT.getSizeInBits() >= 256) && (EltSize == 32);
4146 // VSHUFPSY divides the resulting vector into 4 chunks.
4147 // The sources are also splitted into 4 chunks, and each destination
4148 // chunk must come from a different source chunk.
4150 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
4151 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
4153 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
4154 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
4156 // VSHUFPDY divides the resulting vector into 4 chunks.
4157 // The sources are also splitted into 4 chunks, and each destination
4158 // chunk must come from a different source chunk.
4160 // SRC1 => X3 X2 X1 X0
4161 // SRC2 => Y3 Y2 Y1 Y0
4163 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
4165 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
4166 unsigned HalfLaneElems = NumLaneElems/2;
4167 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
4168 for (unsigned i = 0; i != NumLaneElems; ++i) {
4169 int Idx = Mask[i+l];
4170 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
4171 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
4173 // For VSHUFPSY, the mask of the second half must be the same as the
4174 // first but with the appropriate offsets. This works in the same way as
4175 // VPERMILPS works with masks.
4176 if (!symetricMaskRequired || Idx < 0)
4178 if (MaskVal[i] < 0) {
4179 MaskVal[i] = Idx - l;
4182 if ((signed)(Idx - l) != MaskVal[i])
4190 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
4191 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
4192 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
4193 if (!VT.is128BitVector())
4196 unsigned NumElems = VT.getVectorNumElements();
4201 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
4202 return isUndefOrEqual(Mask[0], 6) &&
4203 isUndefOrEqual(Mask[1], 7) &&
4204 isUndefOrEqual(Mask[2], 2) &&
4205 isUndefOrEqual(Mask[3], 3);
4208 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
4209 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
4211 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
4212 if (!VT.is128BitVector())
4215 unsigned NumElems = VT.getVectorNumElements();
4220 return isUndefOrEqual(Mask[0], 2) &&
4221 isUndefOrEqual(Mask[1], 3) &&
4222 isUndefOrEqual(Mask[2], 2) &&
4223 isUndefOrEqual(Mask[3], 3);
4226 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
4227 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
4228 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
4229 if (!VT.is128BitVector())
4232 unsigned NumElems = VT.getVectorNumElements();
4234 if (NumElems != 2 && NumElems != 4)
4237 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4238 if (!isUndefOrEqual(Mask[i], i + NumElems))
4241 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4242 if (!isUndefOrEqual(Mask[i], i))
4248 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
4249 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
4250 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
4251 if (!VT.is128BitVector())
4254 unsigned NumElems = VT.getVectorNumElements();
4256 if (NumElems != 2 && NumElems != 4)
4259 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4260 if (!isUndefOrEqual(Mask[i], i))
4263 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4264 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
4270 /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
4271 /// specifies a shuffle of elements that is suitable for input to INSERTPS.
4272 /// i. e: If all but one element come from the same vector.
4273 static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
4274 // TODO: Deal with AVX's VINSERTPS
4275 if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
4278 unsigned CorrectPosV1 = 0;
4279 unsigned CorrectPosV2 = 0;
4280 for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
4281 if (Mask[i] == -1) {
4289 else if (Mask[i] == i + 4)
4293 if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
4294 // We have 3 elements (undefs count as elements from any vector) from one
4295 // vector, and one from another.
4302 // Some special combinations that can be optimized.
4305 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
4306 SelectionDAG &DAG) {
4307 MVT VT = SVOp->getSimpleValueType(0);
4310 if (VT != MVT::v8i32 && VT != MVT::v8f32)
4313 ArrayRef<int> Mask = SVOp->getMask();
4315 // These are the special masks that may be optimized.
4316 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
4317 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
4318 bool MatchEvenMask = true;
4319 bool MatchOddMask = true;
4320 for (int i=0; i<8; ++i) {
4321 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
4322 MatchEvenMask = false;
4323 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
4324 MatchOddMask = false;
4327 if (!MatchEvenMask && !MatchOddMask)
4330 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
4332 SDValue Op0 = SVOp->getOperand(0);
4333 SDValue Op1 = SVOp->getOperand(1);
4335 if (MatchEvenMask) {
4336 // Shift the second operand right to 32 bits.
4337 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
4338 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
4340 // Shift the first operand left to 32 bits.
4341 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
4342 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
4344 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
4345 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
4348 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
4349 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
4350 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
4351 bool HasInt256, bool V2IsSplat = false) {
4353 assert(VT.getSizeInBits() >= 128 &&
4354 "Unsupported vector type for unpckl");
4356 unsigned NumElts = VT.getVectorNumElements();
4357 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4358 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4361 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4362 "Unsupported vector type for unpckh");
4364 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4365 unsigned NumLanes = VT.getSizeInBits()/128;
4366 unsigned NumLaneElts = NumElts/NumLanes;
4368 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4369 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4370 int BitI = Mask[l+i];
4371 int BitI1 = Mask[l+i+1];
4372 if (!isUndefOrEqual(BitI, j))
4375 if (!isUndefOrEqual(BitI1, NumElts))
4378 if (!isUndefOrEqual(BitI1, j + NumElts))
4387 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
4388 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
4389 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
4390 bool HasInt256, bool V2IsSplat = false) {
4391 assert(VT.getSizeInBits() >= 128 &&
4392 "Unsupported vector type for unpckh");
4394 unsigned NumElts = VT.getVectorNumElements();
4395 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4396 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4399 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4400 "Unsupported vector type for unpckh");
4402 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4403 unsigned NumLanes = VT.getSizeInBits()/128;
4404 unsigned NumLaneElts = NumElts/NumLanes;
4406 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4407 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4408 int BitI = Mask[l+i];
4409 int BitI1 = Mask[l+i+1];
4410 if (!isUndefOrEqual(BitI, j))
4413 if (isUndefOrEqual(BitI1, NumElts))
4416 if (!isUndefOrEqual(BitI1, j+NumElts))
4424 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
4425 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
4427 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4428 unsigned NumElts = VT.getVectorNumElements();
4429 bool Is256BitVec = VT.is256BitVector();
4431 if (VT.is512BitVector())
4433 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4434 "Unsupported vector type for unpckh");
4436 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
4437 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4440 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
4441 // FIXME: Need a better way to get rid of this, there's no latency difference
4442 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
4443 // the former later. We should also remove the "_undef" special mask.
4444 if (NumElts == 4 && Is256BitVec)
4447 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4448 // independently on 128-bit lanes.
4449 unsigned NumLanes = VT.getSizeInBits()/128;
4450 unsigned NumLaneElts = NumElts/NumLanes;
4452 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4453 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4454 int BitI = Mask[l+i];
4455 int BitI1 = Mask[l+i+1];
4457 if (!isUndefOrEqual(BitI, j))
4459 if (!isUndefOrEqual(BitI1, j))
4467 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4468 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4470 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4471 unsigned NumElts = VT.getVectorNumElements();
4473 if (VT.is512BitVector())
4476 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4477 "Unsupported vector type for unpckh");
4479 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4480 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4483 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4484 // independently on 128-bit lanes.
4485 unsigned NumLanes = VT.getSizeInBits()/128;
4486 unsigned NumLaneElts = NumElts/NumLanes;
4488 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4489 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4490 int BitI = Mask[l+i];
4491 int BitI1 = Mask[l+i+1];
4492 if (!isUndefOrEqual(BitI, j))
4494 if (!isUndefOrEqual(BitI1, j))
4501 // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
4502 // (src1[0], src0[1]), manipulation with 256-bit sub-vectors
4503 static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
4504 if (!VT.is512BitVector())
4507 unsigned NumElts = VT.getVectorNumElements();
4508 unsigned HalfSize = NumElts/2;
4509 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
4510 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
4515 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
4516 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
4524 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4525 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4526 /// MOVSD, and MOVD, i.e. setting the lowest element.
4527 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4528 if (VT.getVectorElementType().getSizeInBits() < 32)
4530 if (!VT.is128BitVector())
4533 unsigned NumElts = VT.getVectorNumElements();
4535 if (!isUndefOrEqual(Mask[0], NumElts))
4538 for (unsigned i = 1; i != NumElts; ++i)
4539 if (!isUndefOrEqual(Mask[i], i))
4545 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4546 /// as permutations between 128-bit chunks or halves. As an example: this
4548 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4549 /// The first half comes from the second half of V1 and the second half from the
4550 /// the second half of V2.
4551 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4552 if (!HasFp256 || !VT.is256BitVector())
4555 // The shuffle result is divided into half A and half B. In total the two
4556 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4557 // B must come from C, D, E or F.
4558 unsigned HalfSize = VT.getVectorNumElements()/2;
4559 bool MatchA = false, MatchB = false;
4561 // Check if A comes from one of C, D, E, F.
4562 for (unsigned Half = 0; Half != 4; ++Half) {
4563 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4569 // Check if B comes from one of C, D, E, F.
4570 for (unsigned Half = 0; Half != 4; ++Half) {
4571 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4577 return MatchA && MatchB;
4580 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4581 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4582 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4583 MVT VT = SVOp->getSimpleValueType(0);
4585 unsigned HalfSize = VT.getVectorNumElements()/2;
4587 unsigned FstHalf = 0, SndHalf = 0;
4588 for (unsigned i = 0; i < HalfSize; ++i) {
4589 if (SVOp->getMaskElt(i) > 0) {
4590 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4594 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4595 if (SVOp->getMaskElt(i) > 0) {
4596 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4601 return (FstHalf | (SndHalf << 4));
4604 // Symetric in-lane mask. Each lane has 4 elements (for imm8)
4605 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4606 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4610 unsigned NumElts = VT.getVectorNumElements();
4612 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4613 for (unsigned i = 0; i != NumElts; ++i) {
4616 Imm8 |= Mask[i] << (i*2);
4621 unsigned LaneSize = 4;
4622 SmallVector<int, 4> MaskVal(LaneSize, -1);
4624 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4625 for (unsigned i = 0; i != LaneSize; ++i) {
4626 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4630 if (MaskVal[i] < 0) {
4631 MaskVal[i] = Mask[i+l] - l;
4632 Imm8 |= MaskVal[i] << (i*2);
4635 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4642 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4643 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4644 /// Note that VPERMIL mask matching is different depending whether theunderlying
4645 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4646 /// to the same elements of the low, but to the higher half of the source.
4647 /// In VPERMILPD the two lanes could be shuffled independently of each other
4648 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4649 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4650 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4651 if (VT.getSizeInBits() < 256 || EltSize < 32)
4653 bool symetricMaskRequired = (EltSize == 32);
4654 unsigned NumElts = VT.getVectorNumElements();
4656 unsigned NumLanes = VT.getSizeInBits()/128;
4657 unsigned LaneSize = NumElts/NumLanes;
4658 // 2 or 4 elements in one lane
4660 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4661 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4662 for (unsigned i = 0; i != LaneSize; ++i) {
4663 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4665 if (symetricMaskRequired) {
4666 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4667 ExpectedMaskVal[i] = Mask[i+l] - l;
4670 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4678 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4679 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4680 /// element of vector 2 and the other elements to come from vector 1 in order.
4681 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4682 bool V2IsSplat = false, bool V2IsUndef = false) {
4683 if (!VT.is128BitVector())
4686 unsigned NumOps = VT.getVectorNumElements();
4687 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4690 if (!isUndefOrEqual(Mask[0], 0))
4693 for (unsigned i = 1; i != NumOps; ++i)
4694 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4695 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4696 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4702 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4703 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4704 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4705 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4706 const X86Subtarget *Subtarget) {
4707 if (!Subtarget->hasSSE3())
4710 unsigned NumElems = VT.getVectorNumElements();
4712 if ((VT.is128BitVector() && NumElems != 4) ||
4713 (VT.is256BitVector() && NumElems != 8) ||
4714 (VT.is512BitVector() && NumElems != 16))
4717 // "i+1" is the value the indexed mask element must have
4718 for (unsigned i = 0; i != NumElems; i += 2)
4719 if (!isUndefOrEqual(Mask[i], i+1) ||
4720 !isUndefOrEqual(Mask[i+1], i+1))
4726 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4727 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4728 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4729 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4730 const X86Subtarget *Subtarget) {
4731 if (!Subtarget->hasSSE3())
4734 unsigned NumElems = VT.getVectorNumElements();
4736 if ((VT.is128BitVector() && NumElems != 4) ||
4737 (VT.is256BitVector() && NumElems != 8) ||
4738 (VT.is512BitVector() && NumElems != 16))
4741 // "i" is the value the indexed mask element must have
4742 for (unsigned i = 0; i != NumElems; i += 2)
4743 if (!isUndefOrEqual(Mask[i], i) ||
4744 !isUndefOrEqual(Mask[i+1], i))
4750 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4751 /// specifies a shuffle of elements that is suitable for input to 256-bit
4752 /// version of MOVDDUP.
4753 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4754 if (!HasFp256 || !VT.is256BitVector())
4757 unsigned NumElts = VT.getVectorNumElements();
4761 for (unsigned i = 0; i != NumElts/2; ++i)
4762 if (!isUndefOrEqual(Mask[i], 0))
4764 for (unsigned i = NumElts/2; i != NumElts; ++i)
4765 if (!isUndefOrEqual(Mask[i], NumElts/2))
4770 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4771 /// specifies a shuffle of elements that is suitable for input to 128-bit
4772 /// version of MOVDDUP.
4773 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4774 if (!VT.is128BitVector())
4777 unsigned e = VT.getVectorNumElements() / 2;
4778 for (unsigned i = 0; i != e; ++i)
4779 if (!isUndefOrEqual(Mask[i], i))
4781 for (unsigned i = 0; i != e; ++i)
4782 if (!isUndefOrEqual(Mask[e+i], i))
4787 /// isVEXTRACTIndex - Return true if the specified
4788 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4789 /// suitable for instruction that extract 128 or 256 bit vectors
4790 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4791 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4792 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4795 // The index should be aligned on a vecWidth-bit boundary.
4797 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4799 MVT VT = N->getSimpleValueType(0);
4800 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4801 bool Result = (Index * ElSize) % vecWidth == 0;
4806 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4807 /// operand specifies a subvector insert that is suitable for input to
4808 /// insertion of 128 or 256-bit subvectors
4809 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4810 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4811 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4813 // The index should be aligned on a vecWidth-bit boundary.
4815 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4817 MVT VT = N->getSimpleValueType(0);
4818 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4819 bool Result = (Index * ElSize) % vecWidth == 0;
4824 bool X86::isVINSERT128Index(SDNode *N) {
4825 return isVINSERTIndex(N, 128);
4828 bool X86::isVINSERT256Index(SDNode *N) {
4829 return isVINSERTIndex(N, 256);
4832 bool X86::isVEXTRACT128Index(SDNode *N) {
4833 return isVEXTRACTIndex(N, 128);
4836 bool X86::isVEXTRACT256Index(SDNode *N) {
4837 return isVEXTRACTIndex(N, 256);
4840 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4841 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4842 /// Handles 128-bit and 256-bit.
4843 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4844 MVT VT = N->getSimpleValueType(0);
4846 assert((VT.getSizeInBits() >= 128) &&
4847 "Unsupported vector type for PSHUF/SHUFP");
4849 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4850 // independently on 128-bit lanes.
4851 unsigned NumElts = VT.getVectorNumElements();
4852 unsigned NumLanes = VT.getSizeInBits()/128;
4853 unsigned NumLaneElts = NumElts/NumLanes;
4855 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4856 "Only supports 2, 4 or 8 elements per lane");
4858 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4860 for (unsigned i = 0; i != NumElts; ++i) {
4861 int Elt = N->getMaskElt(i);
4862 if (Elt < 0) continue;
4863 Elt &= NumLaneElts - 1;
4864 unsigned ShAmt = (i << Shift) % 8;
4865 Mask |= Elt << ShAmt;
4871 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4872 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4873 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4874 MVT VT = N->getSimpleValueType(0);
4876 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4877 "Unsupported vector type for PSHUFHW");
4879 unsigned NumElts = VT.getVectorNumElements();
4882 for (unsigned l = 0; l != NumElts; l += 8) {
4883 // 8 nodes per lane, but we only care about the last 4.
4884 for (unsigned i = 0; i < 4; ++i) {
4885 int Elt = N->getMaskElt(l+i+4);
4886 if (Elt < 0) continue;
4887 Elt &= 0x3; // only 2-bits.
4888 Mask |= Elt << (i * 2);
4895 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4896 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4897 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4898 MVT VT = N->getSimpleValueType(0);
4900 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4901 "Unsupported vector type for PSHUFHW");
4903 unsigned NumElts = VT.getVectorNumElements();
4906 for (unsigned l = 0; l != NumElts; l += 8) {
4907 // 8 nodes per lane, but we only care about the first 4.
4908 for (unsigned i = 0; i < 4; ++i) {
4909 int Elt = N->getMaskElt(l+i);
4910 if (Elt < 0) continue;
4911 Elt &= 0x3; // only 2-bits
4912 Mask |= Elt << (i * 2);
4919 /// \brief Return the appropriate immediate to shuffle the specified
4920 /// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with
4921 /// VALIGN (if Interlane is true) instructions.
4922 static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp,
4924 MVT VT = SVOp->getSimpleValueType(0);
4925 unsigned EltSize = InterLane ? 1 :
4926 VT.getVectorElementType().getSizeInBits() >> 3;
4928 unsigned NumElts = VT.getVectorNumElements();
4929 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4930 unsigned NumLaneElts = NumElts/NumLanes;
4934 for (i = 0; i != NumElts; ++i) {
4935 Val = SVOp->getMaskElt(i);
4939 if (Val >= (int)NumElts)
4940 Val -= NumElts - NumLaneElts;
4942 assert(Val - i > 0 && "PALIGNR imm should be positive");
4943 return (Val - i) * EltSize;
4946 /// \brief Return the appropriate immediate to shuffle the specified
4947 /// VECTOR_SHUFFLE mask with the PALIGNR instruction.
4948 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4949 return getShuffleAlignrImmediate(SVOp, false);
4952 /// \brief Return the appropriate immediate to shuffle the specified
4953 /// VECTOR_SHUFFLE mask with the VALIGN instruction.
4954 static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) {
4955 return getShuffleAlignrImmediate(SVOp, true);
4959 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4960 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4961 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4962 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4965 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4967 MVT VecVT = N->getOperand(0).getSimpleValueType();
4968 MVT ElVT = VecVT.getVectorElementType();
4970 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4971 return Index / NumElemsPerChunk;
4974 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4975 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4976 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4977 llvm_unreachable("Illegal insert subvector for VINSERT");
4980 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4982 MVT VecVT = N->getSimpleValueType(0);
4983 MVT ElVT = VecVT.getVectorElementType();
4985 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4986 return Index / NumElemsPerChunk;
4989 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4990 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4991 /// and VINSERTI128 instructions.
4992 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4993 return getExtractVEXTRACTImmediate(N, 128);
4996 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4997 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4998 /// and VINSERTI64x4 instructions.
4999 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
5000 return getExtractVEXTRACTImmediate(N, 256);
5003 /// getInsertVINSERT128Immediate - Return the appropriate immediate
5004 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
5005 /// and VINSERTI128 instructions.
5006 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
5007 return getInsertVINSERTImmediate(N, 128);
5010 /// getInsertVINSERT256Immediate - Return the appropriate immediate
5011 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
5012 /// and VINSERTI64x4 instructions.
5013 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
5014 return getInsertVINSERTImmediate(N, 256);
5017 /// isZero - Returns true if Elt is a constant integer zero
5018 static bool isZero(SDValue V) {
5019 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
5020 return C && C->isNullValue();
5023 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
5025 bool X86::isZeroNode(SDValue Elt) {
5028 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
5029 return CFP->getValueAPF().isPosZero();
5033 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
5034 /// match movhlps. The lower half elements should come from upper half of
5035 /// V1 (and in order), and the upper half elements should come from the upper
5036 /// half of V2 (and in order).
5037 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
5038 if (!VT.is128BitVector())
5040 if (VT.getVectorNumElements() != 4)
5042 for (unsigned i = 0, e = 2; i != e; ++i)
5043 if (!isUndefOrEqual(Mask[i], i+2))
5045 for (unsigned i = 2; i != 4; ++i)
5046 if (!isUndefOrEqual(Mask[i], i+4))
5051 /// isScalarLoadToVector - Returns true if the node is a scalar load that
5052 /// is promoted to a vector. It also returns the LoadSDNode by reference if
5054 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
5055 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
5057 N = N->getOperand(0).getNode();
5058 if (!ISD::isNON_EXTLoad(N))
5061 *LD = cast<LoadSDNode>(N);
5065 // Test whether the given value is a vector value which will be legalized
5067 static bool WillBeConstantPoolLoad(SDNode *N) {
5068 if (N->getOpcode() != ISD::BUILD_VECTOR)
5071 // Check for any non-constant elements.
5072 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5073 switch (N->getOperand(i).getNode()->getOpcode()) {
5075 case ISD::ConstantFP:
5082 // Vectors of all-zeros and all-ones are materialized with special
5083 // instructions rather than being loaded.
5084 return !ISD::isBuildVectorAllZeros(N) &&
5085 !ISD::isBuildVectorAllOnes(N);
5088 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
5089 /// match movlp{s|d}. The lower half elements should come from lower half of
5090 /// V1 (and in order), and the upper half elements should come from the upper
5091 /// half of V2 (and in order). And since V1 will become the source of the
5092 /// MOVLP, it must be either a vector load or a scalar load to vector.
5093 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
5094 ArrayRef<int> Mask, MVT VT) {
5095 if (!VT.is128BitVector())
5098 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
5100 // Is V2 is a vector load, don't do this transformation. We will try to use
5101 // load folding shufps op.
5102 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
5105 unsigned NumElems = VT.getVectorNumElements();
5107 if (NumElems != 2 && NumElems != 4)
5109 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
5110 if (!isUndefOrEqual(Mask[i], i))
5112 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
5113 if (!isUndefOrEqual(Mask[i], i+NumElems))
5118 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
5119 /// to an zero vector.
5120 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
5121 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
5122 SDValue V1 = N->getOperand(0);
5123 SDValue V2 = N->getOperand(1);
5124 unsigned NumElems = N->getValueType(0).getVectorNumElements();
5125 for (unsigned i = 0; i != NumElems; ++i) {
5126 int Idx = N->getMaskElt(i);
5127 if (Idx >= (int)NumElems) {
5128 unsigned Opc = V2.getOpcode();
5129 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
5131 if (Opc != ISD::BUILD_VECTOR ||
5132 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
5134 } else if (Idx >= 0) {
5135 unsigned Opc = V1.getOpcode();
5136 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
5138 if (Opc != ISD::BUILD_VECTOR ||
5139 !X86::isZeroNode(V1.getOperand(Idx)))
5146 /// getZeroVector - Returns a vector of specified type with all zero elements.
5148 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
5149 SelectionDAG &DAG, SDLoc dl) {
5150 assert(VT.isVector() && "Expected a vector type");
5152 // Always build SSE zero vectors as <4 x i32> bitcasted
5153 // to their dest type. This ensures they get CSE'd.
5155 if (VT.is128BitVector()) { // SSE
5156 if (Subtarget->hasSSE2()) { // SSE2
5157 SDValue Cst = DAG.getConstant(0, MVT::i32);
5158 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5160 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5161 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
5163 } else if (VT.is256BitVector()) { // AVX
5164 if (Subtarget->hasInt256()) { // AVX2
5165 SDValue Cst = DAG.getConstant(0, MVT::i32);
5166 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5167 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5169 // 256-bit logic and arithmetic instructions in AVX are all
5170 // floating-point, no support for integer ops. Emit fp zeroed vectors.
5171 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5172 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5173 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
5175 } else if (VT.is512BitVector()) { // AVX-512
5176 SDValue Cst = DAG.getConstant(0, MVT::i32);
5177 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5178 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5179 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
5180 } else if (VT.getScalarType() == MVT::i1) {
5181 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
5182 SDValue Cst = DAG.getConstant(0, MVT::i1);
5183 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5184 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5186 llvm_unreachable("Unexpected vector type");
5188 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5191 /// getOnesVector - Returns a vector of specified type with all bits set.
5192 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
5193 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
5194 /// Then bitcast to their original type, ensuring they get CSE'd.
5195 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
5197 assert(VT.isVector() && "Expected a vector type");
5199 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
5201 if (VT.is256BitVector()) {
5202 if (HasInt256) { // AVX2
5203 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5204 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5206 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5207 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
5209 } else if (VT.is128BitVector()) {
5210 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5212 llvm_unreachable("Unexpected vector type");
5214 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5217 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
5218 /// that point to V2 points to its first element.
5219 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
5220 for (unsigned i = 0; i != NumElems; ++i) {
5221 if (Mask[i] > (int)NumElems) {
5227 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
5228 /// operation of specified width.
5229 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
5231 unsigned NumElems = VT.getVectorNumElements();
5232 SmallVector<int, 8> Mask;
5233 Mask.push_back(NumElems);
5234 for (unsigned i = 1; i != NumElems; ++i)
5236 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5239 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
5240 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5242 unsigned NumElems = VT.getVectorNumElements();
5243 SmallVector<int, 8> Mask;
5244 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
5246 Mask.push_back(i + NumElems);
5248 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5251 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
5252 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5254 unsigned NumElems = VT.getVectorNumElements();
5255 SmallVector<int, 8> Mask;
5256 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
5257 Mask.push_back(i + Half);
5258 Mask.push_back(i + NumElems + Half);
5260 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5263 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
5264 // a generic shuffle instruction because the target has no such instructions.
5265 // Generate shuffles which repeat i16 and i8 several times until they can be
5266 // represented by v4f32 and then be manipulated by target suported shuffles.
5267 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
5268 MVT VT = V.getSimpleValueType();
5269 int NumElems = VT.getVectorNumElements();
5272 while (NumElems > 4) {
5273 if (EltNo < NumElems/2) {
5274 V = getUnpackl(DAG, dl, VT, V, V);
5276 V = getUnpackh(DAG, dl, VT, V, V);
5277 EltNo -= NumElems/2;
5284 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
5285 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
5286 MVT VT = V.getSimpleValueType();
5289 if (VT.is128BitVector()) {
5290 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
5291 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
5292 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
5294 } else if (VT.is256BitVector()) {
5295 // To use VPERMILPS to splat scalars, the second half of indicies must
5296 // refer to the higher part, which is a duplication of the lower one,
5297 // because VPERMILPS can only handle in-lane permutations.
5298 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
5299 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
5301 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
5302 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
5305 llvm_unreachable("Vector size not supported");
5307 return DAG.getNode(ISD::BITCAST, dl, VT, V);
5310 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
5311 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
5312 MVT SrcVT = SV->getSimpleValueType(0);
5313 SDValue V1 = SV->getOperand(0);
5316 int EltNo = SV->getSplatIndex();
5317 int NumElems = SrcVT.getVectorNumElements();
5318 bool Is256BitVec = SrcVT.is256BitVector();
5320 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
5321 "Unknown how to promote splat for type");
5323 // Extract the 128-bit part containing the splat element and update
5324 // the splat element index when it refers to the higher register.
5326 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
5327 if (EltNo >= NumElems/2)
5328 EltNo -= NumElems/2;
5331 // All i16 and i8 vector types can't be used directly by a generic shuffle
5332 // instruction because the target has no such instruction. Generate shuffles
5333 // which repeat i16 and i8 several times until they fit in i32, and then can
5334 // be manipulated by target suported shuffles.
5335 MVT EltVT = SrcVT.getVectorElementType();
5336 if (EltVT == MVT::i8 || EltVT == MVT::i16)
5337 V1 = PromoteSplati8i16(V1, DAG, EltNo);
5339 // Recreate the 256-bit vector and place the same 128-bit vector
5340 // into the low and high part. This is necessary because we want
5341 // to use VPERM* to shuffle the vectors
5343 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
5346 return getLegalSplat(DAG, V1, EltNo);
5349 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
5350 /// vector of zero or undef vector. This produces a shuffle where the low
5351 /// element of V2 is swizzled into the zero/undef vector, landing at element
5352 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
5353 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
5355 const X86Subtarget *Subtarget,
5356 SelectionDAG &DAG) {
5357 MVT VT = V2.getSimpleValueType();
5359 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
5360 unsigned NumElems = VT.getVectorNumElements();
5361 SmallVector<int, 16> MaskVec;
5362 for (unsigned i = 0; i != NumElems; ++i)
5363 // If this is the insertion idx, put the low elt of V2 here.
5364 MaskVec.push_back(i == Idx ? NumElems : i);
5365 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
5368 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
5369 /// target specific opcode. Returns true if the Mask could be calculated. Sets
5370 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
5371 /// shuffles which use a single input multiple times, and in those cases it will
5372 /// adjust the mask to only have indices within that single input.
5373 static bool getTargetShuffleMask(SDNode *N, MVT VT,
5374 SmallVectorImpl<int> &Mask, bool &IsUnary) {
5375 unsigned NumElems = VT.getVectorNumElements();
5379 bool IsFakeUnary = false;
5380 switch(N->getOpcode()) {
5381 case X86ISD::BLENDI:
5382 ImmN = N->getOperand(N->getNumOperands()-1);
5383 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5386 ImmN = N->getOperand(N->getNumOperands()-1);
5387 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5388 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5390 case X86ISD::UNPCKH:
5391 DecodeUNPCKHMask(VT, Mask);
5392 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5394 case X86ISD::UNPCKL:
5395 DecodeUNPCKLMask(VT, Mask);
5396 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5398 case X86ISD::MOVHLPS:
5399 DecodeMOVHLPSMask(NumElems, Mask);
5400 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5402 case X86ISD::MOVLHPS:
5403 DecodeMOVLHPSMask(NumElems, Mask);
5404 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5406 case X86ISD::PALIGNR:
5407 ImmN = N->getOperand(N->getNumOperands()-1);
5408 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5410 case X86ISD::PSHUFD:
5411 case X86ISD::VPERMILPI:
5412 ImmN = N->getOperand(N->getNumOperands()-1);
5413 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5416 case X86ISD::PSHUFHW:
5417 ImmN = N->getOperand(N->getNumOperands()-1);
5418 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5421 case X86ISD::PSHUFLW:
5422 ImmN = N->getOperand(N->getNumOperands()-1);
5423 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5426 case X86ISD::PSHUFB: {
5428 SDValue MaskNode = N->getOperand(1);
5429 while (MaskNode->getOpcode() == ISD::BITCAST)
5430 MaskNode = MaskNode->getOperand(0);
5432 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
5433 // If we have a build-vector, then things are easy.
5434 EVT VT = MaskNode.getValueType();
5435 assert(VT.isVector() &&
5436 "Can't produce a non-vector with a build_vector!");
5437 if (!VT.isInteger())
5440 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
5442 SmallVector<uint64_t, 32> RawMask;
5443 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
5444 SDValue Op = MaskNode->getOperand(i);
5445 if (Op->getOpcode() == ISD::UNDEF) {
5446 RawMask.push_back((uint64_t)SM_SentinelUndef);
5449 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
5452 APInt MaskElement = CN->getAPIntValue();
5454 // We now have to decode the element which could be any integer size and
5455 // extract each byte of it.
5456 for (int j = 0; j < NumBytesPerElement; ++j) {
5457 // Note that this is x86 and so always little endian: the low byte is
5458 // the first byte of the mask.
5459 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
5460 MaskElement = MaskElement.lshr(8);
5463 DecodePSHUFBMask(RawMask, Mask);
5467 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
5471 SDValue Ptr = MaskLoad->getBasePtr();
5472 if (Ptr->getOpcode() == X86ISD::Wrapper)
5473 Ptr = Ptr->getOperand(0);
5475 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
5476 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5479 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5480 DecodePSHUFBMask(C, Mask);
5486 case X86ISD::VPERMI:
5487 ImmN = N->getOperand(N->getNumOperands()-1);
5488 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5493 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
5495 case X86ISD::VPERM2X128:
5496 ImmN = N->getOperand(N->getNumOperands()-1);
5497 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5498 if (Mask.empty()) return false;
5500 case X86ISD::MOVSLDUP:
5501 DecodeMOVSLDUPMask(VT, Mask);
5504 case X86ISD::MOVSHDUP:
5505 DecodeMOVSHDUPMask(VT, Mask);
5508 case X86ISD::MOVDDUP:
5509 DecodeMOVDDUPMask(VT, Mask);
5512 case X86ISD::MOVLHPD:
5513 case X86ISD::MOVLPD:
5514 case X86ISD::MOVLPS:
5515 // Not yet implemented
5517 default: llvm_unreachable("unknown target shuffle node");
5520 // If we have a fake unary shuffle, the shuffle mask is spread across two
5521 // inputs that are actually the same node. Re-map the mask to always point
5522 // into the first input.
5525 if (M >= (int)Mask.size())
5531 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5532 /// element of the result of the vector shuffle.
5533 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5536 return SDValue(); // Limit search depth.
5538 SDValue V = SDValue(N, 0);
5539 EVT VT = V.getValueType();
5540 unsigned Opcode = V.getOpcode();
5542 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5543 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5544 int Elt = SV->getMaskElt(Index);
5547 return DAG.getUNDEF(VT.getVectorElementType());
5549 unsigned NumElems = VT.getVectorNumElements();
5550 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5551 : SV->getOperand(1);
5552 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5555 // Recurse into target specific vector shuffles to find scalars.
5556 if (isTargetShuffle(Opcode)) {
5557 MVT ShufVT = V.getSimpleValueType();
5558 unsigned NumElems = ShufVT.getVectorNumElements();
5559 SmallVector<int, 16> ShuffleMask;
5562 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5565 int Elt = ShuffleMask[Index];
5567 return DAG.getUNDEF(ShufVT.getVectorElementType());
5569 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5571 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5575 // Actual nodes that may contain scalar elements
5576 if (Opcode == ISD::BITCAST) {
5577 V = V.getOperand(0);
5578 EVT SrcVT = V.getValueType();
5579 unsigned NumElems = VT.getVectorNumElements();
5581 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5585 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5586 return (Index == 0) ? V.getOperand(0)
5587 : DAG.getUNDEF(VT.getVectorElementType());
5589 if (V.getOpcode() == ISD::BUILD_VECTOR)
5590 return V.getOperand(Index);
5595 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5596 /// shuffle operation which come from a consecutively from a zero. The
5597 /// search can start in two different directions, from left or right.
5598 /// We count undefs as zeros until PreferredNum is reached.
5599 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5600 unsigned NumElems, bool ZerosFromLeft,
5602 unsigned PreferredNum = -1U) {
5603 unsigned NumZeros = 0;
5604 for (unsigned i = 0; i != NumElems; ++i) {
5605 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5606 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5610 if (X86::isZeroNode(Elt))
5612 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5613 NumZeros = std::min(NumZeros + 1, PreferredNum);
5621 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5622 /// correspond consecutively to elements from one of the vector operands,
5623 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5625 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5626 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5627 unsigned NumElems, unsigned &OpNum) {
5628 bool SeenV1 = false;
5629 bool SeenV2 = false;
5631 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5632 int Idx = SVOp->getMaskElt(i);
5633 // Ignore undef indicies
5637 if (Idx < (int)NumElems)
5642 // Only accept consecutive elements from the same vector
5643 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5647 OpNum = SeenV1 ? 0 : 1;
5651 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5652 /// logical left shift of a vector.
5653 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5654 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5656 SVOp->getSimpleValueType(0).getVectorNumElements();
5657 unsigned NumZeros = getNumOfConsecutiveZeros(
5658 SVOp, NumElems, false /* check zeros from right */, DAG,
5659 SVOp->getMaskElt(0));
5665 // Considering the elements in the mask that are not consecutive zeros,
5666 // check if they consecutively come from only one of the source vectors.
5668 // V1 = {X, A, B, C} 0
5670 // vector_shuffle V1, V2 <1, 2, 3, X>
5672 if (!isShuffleMaskConsecutive(SVOp,
5673 0, // Mask Start Index
5674 NumElems-NumZeros, // Mask End Index(exclusive)
5675 NumZeros, // Where to start looking in the src vector
5676 NumElems, // Number of elements in vector
5677 OpSrc)) // Which source operand ?
5682 ShVal = SVOp->getOperand(OpSrc);
5686 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5687 /// logical left shift of a vector.
5688 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5689 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5691 SVOp->getSimpleValueType(0).getVectorNumElements();
5692 unsigned NumZeros = getNumOfConsecutiveZeros(
5693 SVOp, NumElems, true /* check zeros from left */, DAG,
5694 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5700 // Considering the elements in the mask that are not consecutive zeros,
5701 // check if they consecutively come from only one of the source vectors.
5703 // 0 { A, B, X, X } = V2
5705 // vector_shuffle V1, V2 <X, X, 4, 5>
5707 if (!isShuffleMaskConsecutive(SVOp,
5708 NumZeros, // Mask Start Index
5709 NumElems, // Mask End Index(exclusive)
5710 0, // Where to start looking in the src vector
5711 NumElems, // Number of elements in vector
5712 OpSrc)) // Which source operand ?
5717 ShVal = SVOp->getOperand(OpSrc);
5721 /// isVectorShift - Returns true if the shuffle can be implemented as a
5722 /// logical left or right shift of a vector.
5723 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5724 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5725 // Although the logic below support any bitwidth size, there are no
5726 // shift instructions which handle more than 128-bit vectors.
5727 if (!SVOp->getSimpleValueType(0).is128BitVector())
5730 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5731 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5737 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5739 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5740 unsigned NumNonZero, unsigned NumZero,
5742 const X86Subtarget* Subtarget,
5743 const TargetLowering &TLI) {
5750 for (unsigned i = 0; i < 16; ++i) {
5751 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5752 if (ThisIsNonZero && First) {
5754 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5756 V = DAG.getUNDEF(MVT::v8i16);
5761 SDValue ThisElt, LastElt;
5762 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5763 if (LastIsNonZero) {
5764 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5765 MVT::i16, Op.getOperand(i-1));
5767 if (ThisIsNonZero) {
5768 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5769 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5770 ThisElt, DAG.getConstant(8, MVT::i8));
5772 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5776 if (ThisElt.getNode())
5777 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5778 DAG.getIntPtrConstant(i/2));
5782 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5785 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5787 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5788 unsigned NumNonZero, unsigned NumZero,
5790 const X86Subtarget* Subtarget,
5791 const TargetLowering &TLI) {
5798 for (unsigned i = 0; i < 8; ++i) {
5799 bool isNonZero = (NonZeros & (1 << i)) != 0;
5803 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5805 V = DAG.getUNDEF(MVT::v8i16);
5808 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5809 MVT::v8i16, V, Op.getOperand(i),
5810 DAG.getIntPtrConstant(i));
5817 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
5818 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5819 const X86Subtarget *Subtarget,
5820 const TargetLowering &TLI) {
5821 // Find all zeroable elements.
5823 for (int i=0; i < 4; ++i) {
5824 SDValue Elt = Op->getOperand(i);
5825 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5827 assert(std::count_if(&Zeroable[0], &Zeroable[4],
5828 [](bool M) { return !M; }) > 1 &&
5829 "We expect at least two non-zero elements!");
5831 // We only know how to deal with build_vector nodes where elements are either
5832 // zeroable or extract_vector_elt with constant index.
5833 SDValue FirstNonZero;
5834 unsigned FirstNonZeroIdx;
5835 for (unsigned i=0; i < 4; ++i) {
5838 SDValue Elt = Op->getOperand(i);
5839 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5840 !isa<ConstantSDNode>(Elt.getOperand(1)))
5842 // Make sure that this node is extracting from a 128-bit vector.
5843 MVT VT = Elt.getOperand(0).getSimpleValueType();
5844 if (!VT.is128BitVector())
5846 if (!FirstNonZero.getNode()) {
5848 FirstNonZeroIdx = i;
5852 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5853 SDValue V1 = FirstNonZero.getOperand(0);
5854 MVT VT = V1.getSimpleValueType();
5856 // See if this build_vector can be lowered as a blend with zero.
5858 unsigned EltMaskIdx, EltIdx;
5860 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5861 if (Zeroable[EltIdx]) {
5862 // The zero vector will be on the right hand side.
5863 Mask[EltIdx] = EltIdx+4;
5867 Elt = Op->getOperand(EltIdx);
5868 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5869 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5870 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5872 Mask[EltIdx] = EltIdx;
5876 // Let the shuffle legalizer deal with blend operations.
5877 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5878 if (V1.getSimpleValueType() != VT)
5879 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5880 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5883 // See if we can lower this build_vector to a INSERTPS.
5884 if (!Subtarget->hasSSE41())
5887 SDValue V2 = Elt.getOperand(0);
5888 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5891 bool CanFold = true;
5892 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5896 SDValue Current = Op->getOperand(i);
5897 SDValue SrcVector = Current->getOperand(0);
5900 CanFold = SrcVector == V1 &&
5901 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5907 assert(V1.getNode() && "Expected at least two non-zero elements!");
5908 if (V1.getSimpleValueType() != MVT::v4f32)
5909 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5910 if (V2.getSimpleValueType() != MVT::v4f32)
5911 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5913 // Ok, we can emit an INSERTPS instruction.
5915 for (int i = 0; i < 4; ++i)
5919 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5920 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5921 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
5922 DAG.getIntPtrConstant(InsertPSMask));
5923 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
5926 /// Return a vector logical shift node.
5927 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5928 unsigned NumBits, SelectionDAG &DAG,
5929 const TargetLowering &TLI, SDLoc dl) {
5930 assert(VT.is128BitVector() && "Unknown type for VShift");
5931 MVT ShVT = MVT::v2i64;
5932 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5933 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5934 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
5935 SDValue ShiftVal = DAG.getConstant(NumBits, ScalarShiftTy);
5936 return DAG.getNode(ISD::BITCAST, dl, VT,
5937 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5941 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5943 // Check if the scalar load can be widened into a vector load. And if
5944 // the address is "base + cst" see if the cst can be "absorbed" into
5945 // the shuffle mask.
5946 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5947 SDValue Ptr = LD->getBasePtr();
5948 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5950 EVT PVT = LD->getValueType(0);
5951 if (PVT != MVT::i32 && PVT != MVT::f32)
5956 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5957 FI = FINode->getIndex();
5959 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5960 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5961 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5962 Offset = Ptr.getConstantOperandVal(1);
5963 Ptr = Ptr.getOperand(0);
5968 // FIXME: 256-bit vector instructions don't require a strict alignment,
5969 // improve this code to support it better.
5970 unsigned RequiredAlign = VT.getSizeInBits()/8;
5971 SDValue Chain = LD->getChain();
5972 // Make sure the stack object alignment is at least 16 or 32.
5973 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5974 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5975 if (MFI->isFixedObjectIndex(FI)) {
5976 // Can't change the alignment. FIXME: It's possible to compute
5977 // the exact stack offset and reference FI + adjust offset instead.
5978 // If someone *really* cares about this. That's the way to implement it.
5981 MFI->setObjectAlignment(FI, RequiredAlign);
5985 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5986 // Ptr + (Offset & ~15).
5989 if ((Offset % RequiredAlign) & 3)
5991 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5993 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5994 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
5996 int EltNo = (Offset - StartOffset) >> 2;
5997 unsigned NumElems = VT.getVectorNumElements();
5999 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
6000 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
6001 LD->getPointerInfo().getWithOffset(StartOffset),
6002 false, false, false, 0);
6004 SmallVector<int, 8> Mask;
6005 for (unsigned i = 0; i != NumElems; ++i)
6006 Mask.push_back(EltNo);
6008 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
6014 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
6015 /// elements can be replaced by a single large load which has the same value as
6016 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
6018 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
6020 /// FIXME: we'd also like to handle the case where the last elements are zero
6021 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
6022 /// There's even a handy isZeroNode for that purpose.
6023 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
6024 SDLoc &DL, SelectionDAG &DAG,
6025 bool isAfterLegalize) {
6026 unsigned NumElems = Elts.size();
6028 LoadSDNode *LDBase = nullptr;
6029 unsigned LastLoadedElt = -1U;
6031 // For each element in the initializer, see if we've found a load or an undef.
6032 // If we don't find an initial load element, or later load elements are
6033 // non-consecutive, bail out.
6034 for (unsigned i = 0; i < NumElems; ++i) {
6035 SDValue Elt = Elts[i];
6036 // Look through a bitcast.
6037 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
6038 Elt = Elt.getOperand(0);
6039 if (!Elt.getNode() ||
6040 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6043 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
6045 LDBase = cast<LoadSDNode>(Elt.getNode());
6049 if (Elt.getOpcode() == ISD::UNDEF)
6052 LoadSDNode *LD = cast<LoadSDNode>(Elt);
6053 EVT LdVT = Elt.getValueType();
6054 // Each loaded element must be the correct fractional portion of the
6055 // requested vector load.
6056 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
6058 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
6063 // If we have found an entire vector of loads and undefs, then return a large
6064 // load of the entire vector width starting at the base pointer. If we found
6065 // consecutive loads for the low half, generate a vzext_load node.
6066 if (LastLoadedElt == NumElems - 1) {
6067 assert(LDBase && "Did not find base load for merging consecutive loads");
6068 EVT EltVT = LDBase->getValueType(0);
6069 // Ensure that the input vector size for the merged loads matches the
6070 // cumulative size of the input elements.
6071 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
6074 if (isAfterLegalize &&
6075 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
6078 SDValue NewLd = SDValue();
6080 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
6081 LDBase->getPointerInfo(), LDBase->isVolatile(),
6082 LDBase->isNonTemporal(), LDBase->isInvariant(),
6083 LDBase->getAlignment());
6085 if (LDBase->hasAnyUseOfValue(1)) {
6086 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6088 SDValue(NewLd.getNode(), 1));
6089 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6090 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6091 SDValue(NewLd.getNode(), 1));
6097 //TODO: The code below fires only for for loading the low v2i32 / v2f32
6098 //of a v4i32 / v4f32. It's probably worth generalizing.
6099 EVT EltVT = VT.getVectorElementType();
6100 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
6101 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
6102 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
6103 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
6105 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
6106 LDBase->getPointerInfo(),
6107 LDBase->getAlignment(),
6108 false/*isVolatile*/, true/*ReadMem*/,
6111 // Make sure the newly-created LOAD is in the same position as LDBase in
6112 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
6113 // update uses of LDBase's output chain to use the TokenFactor.
6114 if (LDBase->hasAnyUseOfValue(1)) {
6115 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6116 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
6117 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6118 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6119 SDValue(ResNode.getNode(), 1));
6122 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
6127 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
6128 /// to generate a splat value for the following cases:
6129 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
6130 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
6131 /// a scalar load, or a constant.
6132 /// The VBROADCAST node is returned when a pattern is found,
6133 /// or SDValue() otherwise.
6134 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
6135 SelectionDAG &DAG) {
6136 // VBROADCAST requires AVX.
6137 // TODO: Splats could be generated for non-AVX CPUs using SSE
6138 // instructions, but there's less potential gain for only 128-bit vectors.
6139 if (!Subtarget->hasAVX())
6142 MVT VT = Op.getSimpleValueType();
6145 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
6146 "Unsupported vector type for broadcast.");
6151 switch (Op.getOpcode()) {
6153 // Unknown pattern found.
6156 case ISD::BUILD_VECTOR: {
6157 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
6158 BitVector UndefElements;
6159 SDValue Splat = BVOp->getSplatValue(&UndefElements);
6161 // We need a splat of a single value to use broadcast, and it doesn't
6162 // make any sense if the value is only in one element of the vector.
6163 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
6167 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6168 Ld.getOpcode() == ISD::ConstantFP);
6170 // Make sure that all of the users of a non-constant load are from the
6171 // BUILD_VECTOR node.
6172 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
6177 case ISD::VECTOR_SHUFFLE: {
6178 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6180 // Shuffles must have a splat mask where the first element is
6182 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
6185 SDValue Sc = Op.getOperand(0);
6186 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
6187 Sc.getOpcode() != ISD::BUILD_VECTOR) {
6189 if (!Subtarget->hasInt256())
6192 // Use the register form of the broadcast instruction available on AVX2.
6193 if (VT.getSizeInBits() >= 256)
6194 Sc = Extract128BitVector(Sc, 0, DAG, dl);
6195 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
6198 Ld = Sc.getOperand(0);
6199 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6200 Ld.getOpcode() == ISD::ConstantFP);
6202 // The scalar_to_vector node and the suspected
6203 // load node must have exactly one user.
6204 // Constants may have multiple users.
6206 // AVX-512 has register version of the broadcast
6207 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
6208 Ld.getValueType().getSizeInBits() >= 32;
6209 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
6216 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
6217 bool IsGE256 = (VT.getSizeInBits() >= 256);
6219 // When optimizing for size, generate up to 5 extra bytes for a broadcast
6220 // instruction to save 8 or more bytes of constant pool data.
6221 // TODO: If multiple splats are generated to load the same constant,
6222 // it may be detrimental to overall size. There needs to be a way to detect
6223 // that condition to know if this is truly a size win.
6224 const Function *F = DAG.getMachineFunction().getFunction();
6225 bool OptForSize = F->getAttributes().
6226 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
6228 // Handle broadcasting a single constant scalar from the constant pool
6230 // On Sandybridge (no AVX2), it is still better to load a constant vector
6231 // from the constant pool and not to broadcast it from a scalar.
6232 // But override that restriction when optimizing for size.
6233 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
6234 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
6235 EVT CVT = Ld.getValueType();
6236 assert(!CVT.isVector() && "Must not broadcast a vector type");
6238 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
6239 // For size optimization, also splat v2f64 and v2i64, and for size opt
6240 // with AVX2, also splat i8 and i16.
6241 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
6242 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6243 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
6244 const Constant *C = nullptr;
6245 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
6246 C = CI->getConstantIntValue();
6247 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
6248 C = CF->getConstantFPValue();
6250 assert(C && "Invalid constant type");
6252 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6253 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
6254 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
6255 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
6256 MachinePointerInfo::getConstantPool(),
6257 false, false, false, Alignment);
6259 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6263 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
6265 // Handle AVX2 in-register broadcasts.
6266 if (!IsLoad && Subtarget->hasInt256() &&
6267 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
6268 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6270 // The scalar source must be a normal load.
6274 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6275 (Subtarget->hasVLX() && ScalarSize == 64))
6276 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6278 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
6279 // double since there is no vbroadcastsd xmm
6280 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
6281 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
6282 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6285 // Unsupported broadcast.
6289 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
6290 /// underlying vector and index.
6292 /// Modifies \p ExtractedFromVec to the real vector and returns the real
6294 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
6296 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
6297 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
6300 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
6302 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
6304 // (extract_vector_elt (vector_shuffle<2,u,u,u>
6305 // (extract_subvector (v8f32 %vreg0), Constant<4>),
6308 // In this case the vector is the extract_subvector expression and the index
6309 // is 2, as specified by the shuffle.
6310 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
6311 SDValue ShuffleVec = SVOp->getOperand(0);
6312 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
6313 assert(ShuffleVecVT.getVectorElementType() ==
6314 ExtractedFromVec.getSimpleValueType().getVectorElementType());
6316 int ShuffleIdx = SVOp->getMaskElt(Idx);
6317 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
6318 ExtractedFromVec = ShuffleVec;
6324 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
6325 MVT VT = Op.getSimpleValueType();
6327 // Skip if insert_vec_elt is not supported.
6328 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6329 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
6333 unsigned NumElems = Op.getNumOperands();
6337 SmallVector<unsigned, 4> InsertIndices;
6338 SmallVector<int, 8> Mask(NumElems, -1);
6340 for (unsigned i = 0; i != NumElems; ++i) {
6341 unsigned Opc = Op.getOperand(i).getOpcode();
6343 if (Opc == ISD::UNDEF)
6346 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
6347 // Quit if more than 1 elements need inserting.
6348 if (InsertIndices.size() > 1)
6351 InsertIndices.push_back(i);
6355 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
6356 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
6357 // Quit if non-constant index.
6358 if (!isa<ConstantSDNode>(ExtIdx))
6360 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
6362 // Quit if extracted from vector of different type.
6363 if (ExtractedFromVec.getValueType() != VT)
6366 if (!VecIn1.getNode())
6367 VecIn1 = ExtractedFromVec;
6368 else if (VecIn1 != ExtractedFromVec) {
6369 if (!VecIn2.getNode())
6370 VecIn2 = ExtractedFromVec;
6371 else if (VecIn2 != ExtractedFromVec)
6372 // Quit if more than 2 vectors to shuffle
6376 if (ExtractedFromVec == VecIn1)
6378 else if (ExtractedFromVec == VecIn2)
6379 Mask[i] = Idx + NumElems;
6382 if (!VecIn1.getNode())
6385 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
6386 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
6387 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
6388 unsigned Idx = InsertIndices[i];
6389 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
6390 DAG.getIntPtrConstant(Idx));
6396 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
6398 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
6400 MVT VT = Op.getSimpleValueType();
6401 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
6402 "Unexpected type in LowerBUILD_VECTORvXi1!");
6405 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6406 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
6407 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6408 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6411 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
6412 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
6413 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6414 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6417 bool AllContants = true;
6418 uint64_t Immediate = 0;
6419 int NonConstIdx = -1;
6420 bool IsSplat = true;
6421 unsigned NumNonConsts = 0;
6422 unsigned NumConsts = 0;
6423 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
6424 SDValue In = Op.getOperand(idx);
6425 if (In.getOpcode() == ISD::UNDEF)
6427 if (!isa<ConstantSDNode>(In)) {
6428 AllContants = false;
6433 if (cast<ConstantSDNode>(In)->getZExtValue())
6434 Immediate |= (1ULL << idx);
6436 if (In != Op.getOperand(0))
6441 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
6442 DAG.getConstant(Immediate, MVT::i16));
6443 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
6444 DAG.getIntPtrConstant(0));
6447 if (NumNonConsts == 1 && NonConstIdx != 0) {
6450 SDValue VecAsImm = DAG.getConstant(Immediate,
6451 MVT::getIntegerVT(VT.getSizeInBits()));
6452 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
6455 DstVec = DAG.getUNDEF(VT);
6456 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
6457 Op.getOperand(NonConstIdx),
6458 DAG.getIntPtrConstant(NonConstIdx));
6460 if (!IsSplat && (NonConstIdx != 0))
6461 llvm_unreachable("Unsupported BUILD_VECTOR operation");
6462 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
6465 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6466 DAG.getConstant(-1, SelectVT),
6467 DAG.getConstant(0, SelectVT));
6469 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6470 DAG.getConstant((Immediate | 1), SelectVT),
6471 DAG.getConstant(Immediate, SelectVT));
6472 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
6475 /// \brief Return true if \p N implements a horizontal binop and return the
6476 /// operands for the horizontal binop into V0 and V1.
6478 /// This is a helper function of PerformBUILD_VECTORCombine.
6479 /// This function checks that the build_vector \p N in input implements a
6480 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
6481 /// operation to match.
6482 /// For example, if \p Opcode is equal to ISD::ADD, then this function
6483 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
6484 /// is equal to ISD::SUB, then this function checks if this is a horizontal
6487 /// This function only analyzes elements of \p N whose indices are
6488 /// in range [BaseIdx, LastIdx).
6489 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
6491 unsigned BaseIdx, unsigned LastIdx,
6492 SDValue &V0, SDValue &V1) {
6493 EVT VT = N->getValueType(0);
6495 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
6496 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
6497 "Invalid Vector in input!");
6499 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
6500 bool CanFold = true;
6501 unsigned ExpectedVExtractIdx = BaseIdx;
6502 unsigned NumElts = LastIdx - BaseIdx;
6503 V0 = DAG.getUNDEF(VT);
6504 V1 = DAG.getUNDEF(VT);
6506 // Check if N implements a horizontal binop.
6507 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
6508 SDValue Op = N->getOperand(i + BaseIdx);
6511 if (Op->getOpcode() == ISD::UNDEF) {
6512 // Update the expected vector extract index.
6513 if (i * 2 == NumElts)
6514 ExpectedVExtractIdx = BaseIdx;
6515 ExpectedVExtractIdx += 2;
6519 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
6524 SDValue Op0 = Op.getOperand(0);
6525 SDValue Op1 = Op.getOperand(1);
6527 // Try to match the following pattern:
6528 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
6529 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6530 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6531 Op0.getOperand(0) == Op1.getOperand(0) &&
6532 isa<ConstantSDNode>(Op0.getOperand(1)) &&
6533 isa<ConstantSDNode>(Op1.getOperand(1)));
6537 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6538 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
6540 if (i * 2 < NumElts) {
6541 if (V0.getOpcode() == ISD::UNDEF)
6542 V0 = Op0.getOperand(0);
6544 if (V1.getOpcode() == ISD::UNDEF)
6545 V1 = Op0.getOperand(0);
6546 if (i * 2 == NumElts)
6547 ExpectedVExtractIdx = BaseIdx;
6550 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
6551 if (I0 == ExpectedVExtractIdx)
6552 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
6553 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
6554 // Try to match the following dag sequence:
6555 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
6556 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
6560 ExpectedVExtractIdx += 2;
6566 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
6567 /// a concat_vector.
6569 /// This is a helper function of PerformBUILD_VECTORCombine.
6570 /// This function expects two 256-bit vectors called V0 and V1.
6571 /// At first, each vector is split into two separate 128-bit vectors.
6572 /// Then, the resulting 128-bit vectors are used to implement two
6573 /// horizontal binary operations.
6575 /// The kind of horizontal binary operation is defined by \p X86Opcode.
6577 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
6578 /// the two new horizontal binop.
6579 /// When Mode is set, the first horizontal binop dag node would take as input
6580 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
6581 /// horizontal binop dag node would take as input the lower 128-bit of V1
6582 /// and the upper 128-bit of V1.
6584 /// HADD V0_LO, V0_HI
6585 /// HADD V1_LO, V1_HI
6587 /// Otherwise, the first horizontal binop dag node takes as input the lower
6588 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
6589 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
6591 /// HADD V0_LO, V1_LO
6592 /// HADD V0_HI, V1_HI
6594 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
6595 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6596 /// the upper 128-bits of the result.
6597 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6598 SDLoc DL, SelectionDAG &DAG,
6599 unsigned X86Opcode, bool Mode,
6600 bool isUndefLO, bool isUndefHI) {
6601 EVT VT = V0.getValueType();
6602 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6603 "Invalid nodes in input!");
6605 unsigned NumElts = VT.getVectorNumElements();
6606 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6607 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6608 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6609 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6610 EVT NewVT = V0_LO.getValueType();
6612 SDValue LO = DAG.getUNDEF(NewVT);
6613 SDValue HI = DAG.getUNDEF(NewVT);
6616 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6617 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6618 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6619 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6620 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6622 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6623 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6624 V1_LO->getOpcode() != ISD::UNDEF))
6625 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6627 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6628 V1_HI->getOpcode() != ISD::UNDEF))
6629 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6632 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6635 /// \brief Try to fold a build_vector that performs an 'addsub' into the
6636 /// sequence of 'vadd + vsub + blendi'.
6637 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
6638 const X86Subtarget *Subtarget) {
6640 EVT VT = BV->getValueType(0);
6641 unsigned NumElts = VT.getVectorNumElements();
6642 SDValue InVec0 = DAG.getUNDEF(VT);
6643 SDValue InVec1 = DAG.getUNDEF(VT);
6645 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6646 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6648 // Odd-numbered elements in the input build vector are obtained from
6649 // adding two integer/float elements.
6650 // Even-numbered elements in the input build vector are obtained from
6651 // subtracting two integer/float elements.
6652 unsigned ExpectedOpcode = ISD::FSUB;
6653 unsigned NextExpectedOpcode = ISD::FADD;
6654 bool AddFound = false;
6655 bool SubFound = false;
6657 for (unsigned i = 0, e = NumElts; i != e; i++) {
6658 SDValue Op = BV->getOperand(i);
6660 // Skip 'undef' values.
6661 unsigned Opcode = Op.getOpcode();
6662 if (Opcode == ISD::UNDEF) {
6663 std::swap(ExpectedOpcode, NextExpectedOpcode);
6667 // Early exit if we found an unexpected opcode.
6668 if (Opcode != ExpectedOpcode)
6671 SDValue Op0 = Op.getOperand(0);
6672 SDValue Op1 = Op.getOperand(1);
6674 // Try to match the following pattern:
6675 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6676 // Early exit if we cannot match that sequence.
6677 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6678 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6679 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6680 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6681 Op0.getOperand(1) != Op1.getOperand(1))
6684 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6688 // We found a valid add/sub node. Update the information accordingly.
6694 // Update InVec0 and InVec1.
6695 if (InVec0.getOpcode() == ISD::UNDEF)
6696 InVec0 = Op0.getOperand(0);
6697 if (InVec1.getOpcode() == ISD::UNDEF)
6698 InVec1 = Op1.getOperand(0);
6700 // Make sure that operands in input to each add/sub node always
6701 // come from a same pair of vectors.
6702 if (InVec0 != Op0.getOperand(0)) {
6703 if (ExpectedOpcode == ISD::FSUB)
6706 // FADD is commutable. Try to commute the operands
6707 // and then test again.
6708 std::swap(Op0, Op1);
6709 if (InVec0 != Op0.getOperand(0))
6713 if (InVec1 != Op1.getOperand(0))
6716 // Update the pair of expected opcodes.
6717 std::swap(ExpectedOpcode, NextExpectedOpcode);
6720 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6721 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6722 InVec1.getOpcode() != ISD::UNDEF)
6723 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6728 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
6729 const X86Subtarget *Subtarget) {
6731 EVT VT = N->getValueType(0);
6732 unsigned NumElts = VT.getVectorNumElements();
6733 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
6734 SDValue InVec0, InVec1;
6736 // Try to match an ADDSUB.
6737 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
6738 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
6739 SDValue Value = matchAddSub(BV, DAG, Subtarget);
6740 if (Value.getNode())
6744 // Try to match horizontal ADD/SUB.
6745 unsigned NumUndefsLO = 0;
6746 unsigned NumUndefsHI = 0;
6747 unsigned Half = NumElts/2;
6749 // Count the number of UNDEF operands in the build_vector in input.
6750 for (unsigned i = 0, e = Half; i != e; ++i)
6751 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6754 for (unsigned i = Half, e = NumElts; i != e; ++i)
6755 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6758 // Early exit if this is either a build_vector of all UNDEFs or all the
6759 // operands but one are UNDEF.
6760 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6763 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6764 // Try to match an SSE3 float HADD/HSUB.
6765 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6766 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6768 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6769 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6770 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6771 // Try to match an SSSE3 integer HADD/HSUB.
6772 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6773 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6775 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6776 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6779 if (!Subtarget->hasAVX())
6782 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6783 // Try to match an AVX horizontal add/sub of packed single/double
6784 // precision floating point values from 256-bit vectors.
6785 SDValue InVec2, InVec3;
6786 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6787 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6788 ((InVec0.getOpcode() == ISD::UNDEF ||
6789 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6790 ((InVec1.getOpcode() == ISD::UNDEF ||
6791 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6792 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6794 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6795 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6796 ((InVec0.getOpcode() == ISD::UNDEF ||
6797 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6798 ((InVec1.getOpcode() == ISD::UNDEF ||
6799 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6800 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6801 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6802 // Try to match an AVX2 horizontal add/sub of signed integers.
6803 SDValue InVec2, InVec3;
6805 bool CanFold = true;
6807 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6808 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6809 ((InVec0.getOpcode() == ISD::UNDEF ||
6810 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6811 ((InVec1.getOpcode() == ISD::UNDEF ||
6812 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6813 X86Opcode = X86ISD::HADD;
6814 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6815 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6816 ((InVec0.getOpcode() == ISD::UNDEF ||
6817 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6818 ((InVec1.getOpcode() == ISD::UNDEF ||
6819 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6820 X86Opcode = X86ISD::HSUB;
6825 // Fold this build_vector into a single horizontal add/sub.
6826 // Do this only if the target has AVX2.
6827 if (Subtarget->hasAVX2())
6828 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6830 // Do not try to expand this build_vector into a pair of horizontal
6831 // add/sub if we can emit a pair of scalar add/sub.
6832 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6835 // Convert this build_vector into a pair of horizontal binop followed by
6837 bool isUndefLO = NumUndefsLO == Half;
6838 bool isUndefHI = NumUndefsHI == Half;
6839 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6840 isUndefLO, isUndefHI);
6844 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6845 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6847 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6848 X86Opcode = X86ISD::HADD;
6849 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6850 X86Opcode = X86ISD::HSUB;
6851 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6852 X86Opcode = X86ISD::FHADD;
6853 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6854 X86Opcode = X86ISD::FHSUB;
6858 // Don't try to expand this build_vector into a pair of horizontal add/sub
6859 // if we can simply emit a pair of scalar add/sub.
6860 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6863 // Convert this build_vector into two horizontal add/sub followed by
6865 bool isUndefLO = NumUndefsLO == Half;
6866 bool isUndefHI = NumUndefsHI == Half;
6867 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6868 isUndefLO, isUndefHI);
6875 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6878 MVT VT = Op.getSimpleValueType();
6879 MVT ExtVT = VT.getVectorElementType();
6880 unsigned NumElems = Op.getNumOperands();
6882 // Generate vectors for predicate vectors.
6883 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6884 return LowerBUILD_VECTORvXi1(Op, DAG);
6886 // Vectors containing all zeros can be matched by pxor and xorps later
6887 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6888 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6889 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6890 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6893 return getZeroVector(VT, Subtarget, DAG, dl);
6896 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6897 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6898 // vpcmpeqd on 256-bit vectors.
6899 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6900 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6903 if (!VT.is512BitVector())
6904 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
6907 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
6908 if (Broadcast.getNode())
6911 unsigned EVTBits = ExtVT.getSizeInBits();
6913 unsigned NumZero = 0;
6914 unsigned NumNonZero = 0;
6915 unsigned NonZeros = 0;
6916 bool IsAllConstants = true;
6917 SmallSet<SDValue, 8> Values;
6918 for (unsigned i = 0; i < NumElems; ++i) {
6919 SDValue Elt = Op.getOperand(i);
6920 if (Elt.getOpcode() == ISD::UNDEF)
6923 if (Elt.getOpcode() != ISD::Constant &&
6924 Elt.getOpcode() != ISD::ConstantFP)
6925 IsAllConstants = false;
6926 if (X86::isZeroNode(Elt))
6929 NonZeros |= (1 << i);
6934 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6935 if (NumNonZero == 0)
6936 return DAG.getUNDEF(VT);
6938 // Special case for single non-zero, non-undef, element.
6939 if (NumNonZero == 1) {
6940 unsigned Idx = countTrailingZeros(NonZeros);
6941 SDValue Item = Op.getOperand(Idx);
6943 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6944 // the value are obviously zero, truncate the value to i32 and do the
6945 // insertion that way. Only do this if the value is non-constant or if the
6946 // value is a constant being inserted into element 0. It is cheaper to do
6947 // a constant pool load than it is to do a movd + shuffle.
6948 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6949 (!IsAllConstants || Idx == 0)) {
6950 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6952 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6953 EVT VecVT = MVT::v4i32;
6954 unsigned VecElts = 4;
6956 // Truncate the value (which may itself be a constant) to i32, and
6957 // convert it to a vector with movd (S2V+shuffle to zero extend).
6958 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6959 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6961 // If using the new shuffle lowering, just directly insert this.
6962 if (ExperimentalVectorShuffleLowering)
6964 ISD::BITCAST, dl, VT,
6965 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
6967 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6969 // Now we have our 32-bit value zero extended in the low element of
6970 // a vector. If Idx != 0, swizzle it into place.
6972 SmallVector<int, 4> Mask;
6973 Mask.push_back(Idx);
6974 for (unsigned i = 1; i != VecElts; ++i)
6976 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
6979 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6983 // If we have a constant or non-constant insertion into the low element of
6984 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6985 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6986 // depending on what the source datatype is.
6989 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6991 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6992 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6993 if (VT.is256BitVector() || VT.is512BitVector()) {
6994 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6995 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6996 Item, DAG.getIntPtrConstant(0));
6998 assert(VT.is128BitVector() && "Expected an SSE value type!");
6999 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
7000 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
7001 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
7004 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
7005 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
7006 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
7007 if (VT.is256BitVector()) {
7008 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
7009 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
7011 assert(VT.is128BitVector() && "Expected an SSE value type!");
7012 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
7014 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
7018 // Is it a vector logical left shift?
7019 if (NumElems == 2 && Idx == 1 &&
7020 X86::isZeroNode(Op.getOperand(0)) &&
7021 !X86::isZeroNode(Op.getOperand(1))) {
7022 unsigned NumBits = VT.getSizeInBits();
7023 return getVShift(true, VT,
7024 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7025 VT, Op.getOperand(1)),
7026 NumBits/2, DAG, *this, dl);
7029 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
7032 // Otherwise, if this is a vector with i32 or f32 elements, and the element
7033 // is a non-constant being inserted into an element other than the low one,
7034 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
7035 // movd/movss) to move this into the low element, then shuffle it into
7037 if (EVTBits == 32) {
7038 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
7040 // If using the new shuffle lowering, just directly insert this.
7041 if (ExperimentalVectorShuffleLowering)
7042 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
7044 // Turn it into a shuffle of zero and zero-extended scalar to vector.
7045 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
7046 SmallVector<int, 8> MaskVec;
7047 for (unsigned i = 0; i != NumElems; ++i)
7048 MaskVec.push_back(i == Idx ? 0 : 1);
7049 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
7053 // Splat is obviously ok. Let legalizer expand it to a shuffle.
7054 if (Values.size() == 1) {
7055 if (EVTBits == 32) {
7056 // Instead of a shuffle like this:
7057 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
7058 // Check if it's possible to issue this instead.
7059 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
7060 unsigned Idx = countTrailingZeros(NonZeros);
7061 SDValue Item = Op.getOperand(Idx);
7062 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
7063 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
7068 // A vector full of immediates; various special cases are already
7069 // handled, so this is best done with a single constant-pool load.
7073 // For AVX-length vectors, see if we can use a vector load to get all of the
7074 // elements, otherwise build the individual 128-bit pieces and use
7075 // shuffles to put them in place.
7076 if (VT.is256BitVector() || VT.is512BitVector()) {
7077 SmallVector<SDValue, 64> V;
7078 for (unsigned i = 0; i != NumElems; ++i)
7079 V.push_back(Op.getOperand(i));
7081 // Check for a build vector of consecutive loads.
7082 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
7085 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
7087 // Build both the lower and upper subvector.
7088 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7089 makeArrayRef(&V[0], NumElems/2));
7090 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7091 makeArrayRef(&V[NumElems / 2], NumElems/2));
7093 // Recreate the wider vector with the lower and upper part.
7094 if (VT.is256BitVector())
7095 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7096 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7099 // Let legalizer expand 2-wide build_vectors.
7100 if (EVTBits == 64) {
7101 if (NumNonZero == 1) {
7102 // One half is zero or undef.
7103 unsigned Idx = countTrailingZeros(NonZeros);
7104 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
7105 Op.getOperand(Idx));
7106 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
7111 // If element VT is < 32 bits, convert it to inserts into a zero vector.
7112 if (EVTBits == 8 && NumElems == 16) {
7113 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
7115 if (V.getNode()) return V;
7118 if (EVTBits == 16 && NumElems == 8) {
7119 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
7121 if (V.getNode()) return V;
7124 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
7125 if (EVTBits == 32 && NumElems == 4) {
7126 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
7131 // If element VT is == 32 bits, turn it into a number of shuffles.
7132 SmallVector<SDValue, 8> V(NumElems);
7133 if (NumElems == 4 && NumZero > 0) {
7134 for (unsigned i = 0; i < 4; ++i) {
7135 bool isZero = !(NonZeros & (1 << i));
7137 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
7139 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7142 for (unsigned i = 0; i < 2; ++i) {
7143 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
7146 V[i] = V[i*2]; // Must be a zero vector.
7149 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
7152 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
7155 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
7160 bool Reverse1 = (NonZeros & 0x3) == 2;
7161 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
7165 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
7166 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
7168 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
7171 if (Values.size() > 1 && VT.is128BitVector()) {
7172 // Check for a build vector of consecutive loads.
7173 for (unsigned i = 0; i < NumElems; ++i)
7174 V[i] = Op.getOperand(i);
7176 // Check for elements which are consecutive loads.
7177 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
7181 // Check for a build vector from mostly shuffle plus few inserting.
7182 SDValue Sh = buildFromShuffleMostly(Op, DAG);
7186 // For SSE 4.1, use insertps to put the high elements into the low element.
7187 if (Subtarget->hasSSE41()) {
7189 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
7190 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
7192 Result = DAG.getUNDEF(VT);
7194 for (unsigned i = 1; i < NumElems; ++i) {
7195 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
7196 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
7197 Op.getOperand(i), DAG.getIntPtrConstant(i));
7202 // Otherwise, expand into a number of unpckl*, start by extending each of
7203 // our (non-undef) elements to the full vector width with the element in the
7204 // bottom slot of the vector (which generates no code for SSE).
7205 for (unsigned i = 0; i < NumElems; ++i) {
7206 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
7207 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7209 V[i] = DAG.getUNDEF(VT);
7212 // Next, we iteratively mix elements, e.g. for v4f32:
7213 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
7214 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
7215 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
7216 unsigned EltStride = NumElems >> 1;
7217 while (EltStride != 0) {
7218 for (unsigned i = 0; i < EltStride; ++i) {
7219 // If V[i+EltStride] is undef and this is the first round of mixing,
7220 // then it is safe to just drop this shuffle: V[i] is already in the
7221 // right place, the one element (since it's the first round) being
7222 // inserted as undef can be dropped. This isn't safe for successive
7223 // rounds because they will permute elements within both vectors.
7224 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
7225 EltStride == NumElems/2)
7228 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
7237 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
7238 // to create 256-bit vectors from two other 128-bit ones.
7239 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7241 MVT ResVT = Op.getSimpleValueType();
7243 assert((ResVT.is256BitVector() ||
7244 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
7246 SDValue V1 = Op.getOperand(0);
7247 SDValue V2 = Op.getOperand(1);
7248 unsigned NumElems = ResVT.getVectorNumElements();
7249 if(ResVT.is256BitVector())
7250 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7252 if (Op.getNumOperands() == 4) {
7253 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
7254 ResVT.getVectorNumElements()/2);
7255 SDValue V3 = Op.getOperand(2);
7256 SDValue V4 = Op.getOperand(3);
7257 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
7258 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
7260 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7263 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7264 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
7265 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
7266 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
7267 Op.getNumOperands() == 4)));
7269 // AVX can use the vinsertf128 instruction to create 256-bit vectors
7270 // from two other 128-bit ones.
7272 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
7273 return LowerAVXCONCAT_VECTORS(Op, DAG);
7277 //===----------------------------------------------------------------------===//
7278 // Vector shuffle lowering
7280 // This is an experimental code path for lowering vector shuffles on x86. It is
7281 // designed to handle arbitrary vector shuffles and blends, gracefully
7282 // degrading performance as necessary. It works hard to recognize idiomatic
7283 // shuffles and lower them to optimal instruction patterns without leaving
7284 // a framework that allows reasonably efficient handling of all vector shuffle
7286 //===----------------------------------------------------------------------===//
7288 /// \brief Tiny helper function to identify a no-op mask.
7290 /// This is a somewhat boring predicate function. It checks whether the mask
7291 /// array input, which is assumed to be a single-input shuffle mask of the kind
7292 /// used by the X86 shuffle instructions (not a fully general
7293 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
7294 /// in-place shuffle are 'no-op's.
7295 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
7296 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7297 if (Mask[i] != -1 && Mask[i] != i)
7302 /// \brief Helper function to classify a mask as a single-input mask.
7304 /// This isn't a generic single-input test because in the vector shuffle
7305 /// lowering we canonicalize single inputs to be the first input operand. This
7306 /// means we can more quickly test for a single input by only checking whether
7307 /// an input from the second operand exists. We also assume that the size of
7308 /// mask corresponds to the size of the input vectors which isn't true in the
7309 /// fully general case.
7310 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
7312 if (M >= (int)Mask.size())
7317 /// \brief Test whether there are elements crossing 128-bit lanes in this
7320 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
7321 /// and we routinely test for these.
7322 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
7323 int LaneSize = 128 / VT.getScalarSizeInBits();
7324 int Size = Mask.size();
7325 for (int i = 0; i < Size; ++i)
7326 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
7331 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
7333 /// This checks a shuffle mask to see if it is performing the same
7334 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
7335 /// that it is also not lane-crossing. It may however involve a blend from the
7336 /// same lane of a second vector.
7338 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
7339 /// non-trivial to compute in the face of undef lanes. The representation is
7340 /// *not* suitable for use with existing 128-bit shuffles as it will contain
7341 /// entries from both V1 and V2 inputs to the wider mask.
7343 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
7344 SmallVectorImpl<int> &RepeatedMask) {
7345 int LaneSize = 128 / VT.getScalarSizeInBits();
7346 RepeatedMask.resize(LaneSize, -1);
7347 int Size = Mask.size();
7348 for (int i = 0; i < Size; ++i) {
7351 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
7352 // This entry crosses lanes, so there is no way to model this shuffle.
7355 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
7356 if (RepeatedMask[i % LaneSize] == -1)
7357 // This is the first non-undef entry in this slot of a 128-bit lane.
7358 RepeatedMask[i % LaneSize] =
7359 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
7360 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
7361 // Found a mismatch with the repeated mask.
7367 // Hide this symbol with an anonymous namespace instead of 'static' so that MSVC
7368 // 2013 will allow us to use it as a non-type template parameter.
7371 /// \brief Implementation of the \c isShuffleEquivalent variadic functor.
7373 /// See its documentation for details.
7374 bool isShuffleEquivalentImpl(ArrayRef<int> Mask, ArrayRef<const int *> Args) {
7375 if (Mask.size() != Args.size())
7377 for (int i = 0, e = Mask.size(); i < e; ++i) {
7378 assert(*Args[i] >= 0 && "Arguments must be positive integers!");
7379 if (Mask[i] != -1 && Mask[i] != *Args[i])
7387 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
7390 /// This is a fast way to test a shuffle mask against a fixed pattern:
7392 /// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... }
7394 /// It returns true if the mask is exactly as wide as the argument list, and
7395 /// each element of the mask is either -1 (signifying undef) or the value given
7396 /// in the argument.
7397 static const VariadicFunction1<
7398 bool, ArrayRef<int>, int, isShuffleEquivalentImpl> isShuffleEquivalent = {};
7400 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
7402 /// This helper function produces an 8-bit shuffle immediate corresponding to
7403 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
7404 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
7407 /// NB: We rely heavily on "undef" masks preserving the input lane.
7408 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
7409 SelectionDAG &DAG) {
7410 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
7411 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
7412 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
7413 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
7414 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
7417 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
7418 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
7419 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
7420 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
7421 return DAG.getConstant(Imm, MVT::i8);
7424 /// \brief Try to emit a blend instruction for a shuffle.
7426 /// This doesn't do any checks for the availability of instructions for blending
7427 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7428 /// be matched in the backend with the type given. What it does check for is
7429 /// that the shuffle mask is in fact a blend.
7430 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7431 SDValue V2, ArrayRef<int> Mask,
7432 const X86Subtarget *Subtarget,
7433 SelectionDAG &DAG) {
7435 unsigned BlendMask = 0;
7436 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7437 if (Mask[i] >= Size) {
7438 if (Mask[i] != i + Size)
7439 return SDValue(); // Shuffled V2 input!
7440 BlendMask |= 1u << i;
7443 if (Mask[i] >= 0 && Mask[i] != i)
7444 return SDValue(); // Shuffled V1 input!
7446 switch (VT.SimpleTy) {
7451 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7452 DAG.getConstant(BlendMask, MVT::i8));
7456 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7460 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7461 // that instruction.
7462 if (Subtarget->hasAVX2()) {
7463 // Scale the blend by the number of 32-bit dwords per element.
7464 int Scale = VT.getScalarSizeInBits() / 32;
7466 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7467 if (Mask[i] >= Size)
7468 for (int j = 0; j < Scale; ++j)
7469 BlendMask |= 1u << (i * Scale + j);
7471 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7472 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
7473 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
7474 return DAG.getNode(ISD::BITCAST, DL, VT,
7475 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7476 DAG.getConstant(BlendMask, MVT::i8)));
7480 // For integer shuffles we need to expand the mask and cast the inputs to
7481 // v8i16s prior to blending.
7482 int Scale = 8 / VT.getVectorNumElements();
7484 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7485 if (Mask[i] >= Size)
7486 for (int j = 0; j < Scale; ++j)
7487 BlendMask |= 1u << (i * Scale + j);
7489 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
7490 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
7491 return DAG.getNode(ISD::BITCAST, DL, VT,
7492 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7493 DAG.getConstant(BlendMask, MVT::i8)));
7497 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7498 SmallVector<int, 8> RepeatedMask;
7499 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7500 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7501 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7503 for (int i = 0; i < 8; ++i)
7504 if (RepeatedMask[i] >= 16)
7505 BlendMask |= 1u << i;
7506 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7507 DAG.getConstant(BlendMask, MVT::i8));
7512 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7513 // Scale the blend by the number of bytes per element.
7514 int Scale = VT.getScalarSizeInBits() / 8;
7515 assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!");
7517 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7518 // mix of LLVM's code generator and the x86 backend. We tell the code
7519 // generator that boolean values in the elements of an x86 vector register
7520 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7521 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7522 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7523 // of the element (the remaining are ignored) and 0 in that high bit would
7524 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7525 // the LLVM model for boolean values in vector elements gets the relevant
7526 // bit set, it is set backwards and over constrained relative to x86's
7528 SDValue VSELECTMask[32];
7529 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7530 for (int j = 0; j < Scale; ++j)
7531 VSELECTMask[Scale * i + j] =
7532 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7533 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8);
7535 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1);
7536 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2);
7538 ISD::BITCAST, DL, VT,
7539 DAG.getNode(ISD::VSELECT, DL, MVT::v32i8,
7540 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, VSELECTMask),
7545 llvm_unreachable("Not a supported integer vector type!");
7549 /// \brief Generic routine to lower a shuffle and blend as a decomposed set of
7550 /// unblended shuffles followed by an unshuffled blend.
7552 /// This matches the extremely common pattern for handling combined
7553 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7555 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7559 SelectionDAG &DAG) {
7560 // Shuffle the input elements into the desired positions in V1 and V2 and
7561 // blend them together.
7562 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7563 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7564 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7565 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7566 if (Mask[i] >= 0 && Mask[i] < Size) {
7567 V1Mask[i] = Mask[i];
7569 } else if (Mask[i] >= Size) {
7570 V2Mask[i] = Mask[i] - Size;
7571 BlendMask[i] = i + Size;
7574 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7575 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7576 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7579 /// \brief Try to lower a vector shuffle as a byte rotation.
7581 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7582 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7583 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7584 /// try to generically lower a vector shuffle through such an pattern. It
7585 /// does not check for the profitability of lowering either as PALIGNR or
7586 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7587 /// This matches shuffle vectors that look like:
7589 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7591 /// Essentially it concatenates V1 and V2, shifts right by some number of
7592 /// elements, and takes the low elements as the result. Note that while this is
7593 /// specified as a *right shift* because x86 is little-endian, it is a *left
7594 /// rotate* of the vector lanes.
7596 /// Note that this only handles 128-bit vector widths currently.
7597 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7600 const X86Subtarget *Subtarget,
7601 SelectionDAG &DAG) {
7602 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7604 // We need to detect various ways of spelling a rotation:
7605 // [11, 12, 13, 14, 15, 0, 1, 2]
7606 // [-1, 12, 13, 14, -1, -1, 1, -1]
7607 // [-1, -1, -1, -1, -1, -1, 1, 2]
7608 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7609 // [-1, 4, 5, 6, -1, -1, 9, -1]
7610 // [-1, 4, 5, 6, -1, -1, -1, -1]
7613 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7616 assert(Mask[i] >= 0 && "Only -1 is a valid negative mask element!");
7618 // Based on the mod-Size value of this mask element determine where
7619 // a rotated vector would have started.
7620 int StartIdx = i - (Mask[i] % Size);
7622 // The identity rotation isn't interesting, stop.
7625 // If we found the tail of a vector the rotation must be the missing
7626 // front. If we found the head of a vector, it must be how much of the head.
7627 int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
7630 Rotation = CandidateRotation;
7631 else if (Rotation != CandidateRotation)
7632 // The rotations don't match, so we can't match this mask.
7635 // Compute which value this mask is pointing at.
7636 SDValue MaskV = Mask[i] < Size ? V1 : V2;
7638 // Compute which of the two target values this index should be assigned to.
7639 // This reflects whether the high elements are remaining or the low elements
7641 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7643 // Either set up this value if we've not encountered it before, or check
7644 // that it remains consistent.
7647 else if (TargetV != MaskV)
7648 // This may be a rotation, but it pulls from the inputs in some
7649 // unsupported interleaving.
7653 // Check that we successfully analyzed the mask, and normalize the results.
7654 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7655 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7661 assert(VT.getSizeInBits() == 128 &&
7662 "Rotate-based lowering only supports 128-bit lowering!");
7663 assert(Mask.size() <= 16 &&
7664 "Can shuffle at most 16 bytes in a 128-bit vector!");
7666 // The actual rotate instruction rotates bytes, so we need to scale the
7667 // rotation based on how many bytes are in the vector.
7668 int Scale = 16 / Mask.size();
7670 // SSSE3 targets can use the palignr instruction
7671 if (Subtarget->hasSSSE3()) {
7672 // Cast the inputs to v16i8 to match PALIGNR.
7673 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Lo);
7674 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Hi);
7676 return DAG.getNode(ISD::BITCAST, DL, VT,
7677 DAG.getNode(X86ISD::PALIGNR, DL, MVT::v16i8, Hi, Lo,
7678 DAG.getConstant(Rotation * Scale, MVT::i8)));
7681 // Default SSE2 implementation
7682 int LoByteShift = 16 - Rotation * Scale;
7683 int HiByteShift = Rotation * Scale;
7685 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7686 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
7687 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
7689 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7690 DAG.getConstant(8 * LoByteShift, MVT::i8));
7691 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7692 DAG.getConstant(8 * HiByteShift, MVT::i8));
7693 return DAG.getNode(ISD::BITCAST, DL, VT,
7694 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7697 /// \brief Compute whether each element of a shuffle is zeroable.
7699 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
7700 /// Either it is an undef element in the shuffle mask, the element of the input
7701 /// referenced is undef, or the element of the input referenced is known to be
7702 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
7703 /// as many lanes with this technique as possible to simplify the remaining
7705 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
7706 SDValue V1, SDValue V2) {
7707 SmallBitVector Zeroable(Mask.size(), false);
7709 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7710 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7712 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7714 // Handle the easy cases.
7715 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
7720 // If this is an index into a build_vector node, dig out the input value and
7722 SDValue V = M < Size ? V1 : V2;
7723 if (V.getOpcode() != ISD::BUILD_VECTOR)
7726 SDValue Input = V.getOperand(M % Size);
7727 // The UNDEF opcode check really should be dead code here, but not quite
7728 // worth asserting on (it isn't invalid, just unexpected).
7729 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
7736 /// \brief Try to emit a bitmask instruction for a shuffle.
7738 /// This handles cases where we can model a blend exactly as a bitmask due to
7739 /// one of the inputs being zeroable.
7740 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
7741 SDValue V2, ArrayRef<int> Mask,
7742 SelectionDAG &DAG) {
7743 MVT EltVT = VT.getScalarType();
7744 int NumEltBits = EltVT.getSizeInBits();
7745 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
7746 SDValue Zero = DAG.getConstant(0, IntEltVT);
7747 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
7748 if (EltVT.isFloatingPoint()) {
7749 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
7750 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
7752 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
7753 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7755 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7758 if (Mask[i] % Size != i)
7759 return SDValue(); // Not a blend.
7761 V = Mask[i] < Size ? V1 : V2;
7762 else if (V != (Mask[i] < Size ? V1 : V2))
7763 return SDValue(); // Can only let one input through the mask.
7765 VMaskOps[i] = AllOnes;
7768 return SDValue(); // No non-zeroable elements!
7770 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
7771 V = DAG.getNode(VT.isFloatingPoint() ? X86ISD::FAND : ISD::AND, DL, VT, V,
7776 /// \brief Try to lower a vector shuffle as a byte shift (shifts in zeros).
7778 /// Attempts to match a shuffle mask against the PSRLDQ and PSLLDQ SSE2
7779 /// byte-shift instructions. The mask must consist of a shifted sequential
7780 /// shuffle from one of the input vectors and zeroable elements for the
7781 /// remaining 'shifted in' elements.
7783 /// Note that this only handles 128-bit vector widths currently.
7784 static SDValue lowerVectorShuffleAsByteShift(SDLoc DL, MVT VT, SDValue V1,
7785 SDValue V2, ArrayRef<int> Mask,
7786 SelectionDAG &DAG) {
7787 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7789 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7791 int Size = Mask.size();
7792 int Scale = 16 / Size;
7794 for (int Shift = 1; Shift < Size; Shift++) {
7795 int ByteShift = Shift * Scale;
7797 // PSRLDQ : (little-endian) right byte shift
7798 // [ 5, 6, 7, zz, zz, zz, zz, zz]
7799 // [ -1, 5, 6, 7, zz, zz, zz, zz]
7800 // [ 1, 2, -1, -1, -1, -1, zz, zz]
7801 bool ZeroableRight = true;
7802 for (int i = Size - Shift; i < Size; i++) {
7803 ZeroableRight &= Zeroable[i];
7806 if (ZeroableRight) {
7807 bool ValidShiftRight1 =
7808 isSequentialOrUndefInRange(Mask, 0, Size - Shift, Shift);
7809 bool ValidShiftRight2 =
7810 isSequentialOrUndefInRange(Mask, 0, Size - Shift, Size + Shift);
7812 if (ValidShiftRight1 || ValidShiftRight2) {
7813 // Cast the inputs to v2i64 to match PSRLDQ.
7814 SDValue &TargetV = ValidShiftRight1 ? V1 : V2;
7815 SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
7816 SDValue Shifted = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, V,
7817 DAG.getConstant(ByteShift * 8, MVT::i8));
7818 return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
7822 // PSLLDQ : (little-endian) left byte shift
7823 // [ zz, 0, 1, 2, 3, 4, 5, 6]
7824 // [ zz, zz, -1, -1, 2, 3, 4, -1]
7825 // [ zz, zz, zz, zz, zz, zz, -1, 1]
7826 bool ZeroableLeft = true;
7827 for (int i = 0; i < Shift; i++) {
7828 ZeroableLeft &= Zeroable[i];
7832 bool ValidShiftLeft1 =
7833 isSequentialOrUndefInRange(Mask, Shift, Size - Shift, 0);
7834 bool ValidShiftLeft2 =
7835 isSequentialOrUndefInRange(Mask, Shift, Size - Shift, Size);
7837 if (ValidShiftLeft1 || ValidShiftLeft2) {
7838 // Cast the inputs to v2i64 to match PSLLDQ.
7839 SDValue &TargetV = ValidShiftLeft1 ? V1 : V2;
7840 SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
7841 SDValue Shifted = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, V,
7842 DAG.getConstant(ByteShift * 8, MVT::i8));
7843 return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
7851 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7853 /// Attempts to match a shuffle mask against the PSRL(W/D/Q) and PSLL(W/D/Q)
7854 /// SSE2 and AVX2 logical bit-shift instructions. The function matches
7855 /// elements from one of the input vectors shuffled to the left or right
7856 /// with zeroable elements 'shifted in'.
7857 static SDValue lowerVectorShuffleAsBitShift(SDLoc DL, MVT VT, SDValue V1,
7858 SDValue V2, ArrayRef<int> Mask,
7859 SelectionDAG &DAG) {
7860 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7862 int Size = Mask.size();
7863 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7865 // PSRL : (little-endian) right bit shift.
7868 // PSHL : (little-endian) left bit shift.
7870 // [ -1, 4, zz, -1 ]
7871 auto MatchBitShift = [&](int Shift, int Scale) -> SDValue {
7872 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7873 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7874 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7875 "Illegal integer vector type");
7877 bool MatchLeft = true, MatchRight = true;
7878 for (int i = 0; i != Size; i += Scale) {
7879 for (int j = 0; j != Shift; j++) {
7880 MatchLeft &= Zeroable[i + j];
7882 for (int j = Scale - Shift; j != Scale; j++) {
7883 MatchRight &= Zeroable[i + j];
7886 if (!(MatchLeft || MatchRight))
7889 bool MatchV1 = true, MatchV2 = true;
7890 for (int i = 0; i != Size; i += Scale) {
7891 unsigned Pos = MatchLeft ? i + Shift : i;
7892 unsigned Low = MatchLeft ? i : i + Shift;
7893 unsigned Len = Scale - Shift;
7894 MatchV1 &= isSequentialOrUndefInRange(Mask, Pos, Len, Low);
7895 MatchV2 &= isSequentialOrUndefInRange(Mask, Pos, Len, Low + Size);
7897 if (!(MatchV1 || MatchV2))
7900 // Cast the inputs to ShiftVT to match VSRLI/VSHLI and back again.
7901 unsigned OpCode = MatchLeft ? X86ISD::VSHLI : X86ISD::VSRLI;
7902 int ShiftAmt = Shift * VT.getScalarSizeInBits();
7903 SDValue V = MatchV1 ? V1 : V2;
7904 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
7905 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
7906 return DAG.getNode(ISD::BITCAST, DL, VT, V);
7909 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7910 // keep doubling the size of the integer elements up to that. We can
7911 // then shift the elements of the integer vector by whole multiples of
7912 // their width within the elements of the larger integer vector. Test each
7913 // multiple to see if we can find a match with the moved element indices
7914 // and that the shifted in elements are all zeroable.
7915 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 64; Scale *= 2)
7916 for (int Shift = 1; Shift != Scale; Shift++)
7917 if (SDValue BitShift = MatchBitShift(Shift, Scale))
7924 /// \brief Lower a vector shuffle as a zero or any extension.
7926 /// Given a specific number of elements, element bit width, and extension
7927 /// stride, produce either a zero or any extension based on the available
7928 /// features of the subtarget.
7929 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7930 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7931 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7932 assert(Scale > 1 && "Need a scale to extend.");
7933 int NumElements = VT.getVectorNumElements();
7934 int EltBits = VT.getScalarSizeInBits();
7935 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7936 "Only 8, 16, and 32 bit elements can be extended.");
7937 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7939 // Found a valid zext mask! Try various lowering strategies based on the
7940 // input type and available ISA extensions.
7941 if (Subtarget->hasSSE41()) {
7942 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7943 NumElements / Scale);
7944 return DAG.getNode(ISD::BITCAST, DL, VT,
7945 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7948 // For any extends we can cheat for larger element sizes and use shuffle
7949 // instructions that can fold with a load and/or copy.
7950 if (AnyExt && EltBits == 32) {
7951 int PSHUFDMask[4] = {0, -1, 1, -1};
7953 ISD::BITCAST, DL, VT,
7954 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7955 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7956 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7958 if (AnyExt && EltBits == 16 && Scale > 2) {
7959 int PSHUFDMask[4] = {0, -1, 0, -1};
7960 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7961 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7962 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
7963 int PSHUFHWMask[4] = {1, -1, -1, -1};
7965 ISD::BITCAST, DL, VT,
7966 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7967 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
7968 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
7971 // If this would require more than 2 unpack instructions to expand, use
7972 // pshufb when available. We can only use more than 2 unpack instructions
7973 // when zero extending i8 elements which also makes it easier to use pshufb.
7974 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7975 assert(NumElements == 16 && "Unexpected byte vector width!");
7976 SDValue PSHUFBMask[16];
7977 for (int i = 0; i < 16; ++i)
7979 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
7980 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
7981 return DAG.getNode(ISD::BITCAST, DL, VT,
7982 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7983 DAG.getNode(ISD::BUILD_VECTOR, DL,
7984 MVT::v16i8, PSHUFBMask)));
7987 // Otherwise emit a sequence of unpacks.
7989 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7990 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7991 : getZeroVector(InputVT, Subtarget, DAG, DL);
7992 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
7993 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7997 } while (Scale > 1);
7998 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
8001 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
8003 /// This routine will try to do everything in its power to cleverly lower
8004 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
8005 /// check for the profitability of this lowering, it tries to aggressively
8006 /// match this pattern. It will use all of the micro-architectural details it
8007 /// can to emit an efficient lowering. It handles both blends with all-zero
8008 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
8009 /// masking out later).
8011 /// The reason we have dedicated lowering for zext-style shuffles is that they
8012 /// are both incredibly common and often quite performance sensitive.
8013 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
8014 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
8015 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8016 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8018 int Bits = VT.getSizeInBits();
8019 int NumElements = VT.getVectorNumElements();
8020 assert(VT.getScalarSizeInBits() <= 32 &&
8021 "Exceeds 32-bit integer zero extension limit");
8022 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
8024 // Define a helper function to check a particular ext-scale and lower to it if
8026 auto Lower = [&](int Scale) -> SDValue {
8029 for (int i = 0; i < NumElements; ++i) {
8031 continue; // Valid anywhere but doesn't tell us anything.
8032 if (i % Scale != 0) {
8033 // Each of the extended elements need to be zeroable.
8037 // We no longer are in the anyext case.
8042 // Each of the base elements needs to be consecutive indices into the
8043 // same input vector.
8044 SDValue V = Mask[i] < NumElements ? V1 : V2;
8047 else if (InputV != V)
8048 return SDValue(); // Flip-flopping inputs.
8050 if (Mask[i] % NumElements != i / Scale)
8051 return SDValue(); // Non-consecutive strided elements.
8054 // If we fail to find an input, we have a zero-shuffle which should always
8055 // have already been handled.
8056 // FIXME: Maybe handle this here in case during blending we end up with one?
8060 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
8061 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
8064 // The widest scale possible for extending is to a 64-bit integer.
8065 assert(Bits % 64 == 0 &&
8066 "The number of bits in a vector must be divisible by 64 on x86!");
8067 int NumExtElements = Bits / 64;
8069 // Each iteration, try extending the elements half as much, but into twice as
8071 for (; NumExtElements < NumElements; NumExtElements *= 2) {
8072 assert(NumElements % NumExtElements == 0 &&
8073 "The input vector size must be divisible by the extended size.");
8074 if (SDValue V = Lower(NumElements / NumExtElements))
8078 // General extends failed, but 128-bit vectors may be able to use MOVQ.
8082 // Returns one of the source operands if the shuffle can be reduced to a
8083 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
8084 auto CanZExtLowHalf = [&]() {
8085 for (int i = NumElements / 2; i != NumElements; i++)
8088 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
8090 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
8095 if (SDValue V = CanZExtLowHalf()) {
8096 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
8097 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
8098 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8101 // No viable ext lowering found.
8105 /// \brief Try to get a scalar value for a specific element of a vector.
8107 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
8108 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
8109 SelectionDAG &DAG) {
8110 MVT VT = V.getSimpleValueType();
8111 MVT EltVT = VT.getVectorElementType();
8112 while (V.getOpcode() == ISD::BITCAST)
8113 V = V.getOperand(0);
8114 // If the bitcasts shift the element size, we can't extract an equivalent
8116 MVT NewVT = V.getSimpleValueType();
8117 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
8120 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8121 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
8122 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
8127 /// \brief Helper to test for a load that can be folded with x86 shuffles.
8129 /// This is particularly important because the set of instructions varies
8130 /// significantly based on whether the operand is a load or not.
8131 static bool isShuffleFoldableLoad(SDValue V) {
8132 while (V.getOpcode() == ISD::BITCAST)
8133 V = V.getOperand(0);
8135 return ISD::isNON_EXTLoad(V.getNode());
8138 /// \brief Try to lower insertion of a single element into a zero vector.
8140 /// This is a common pattern that we have especially efficient patterns to lower
8141 /// across all subtarget feature sets.
8142 static SDValue lowerVectorShuffleAsElementInsertion(
8143 MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
8144 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8145 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8147 MVT EltVT = VT.getVectorElementType();
8149 int V2Index = std::find_if(Mask.begin(), Mask.end(),
8150 [&Mask](int M) { return M >= (int)Mask.size(); }) -
8152 bool IsV1Zeroable = true;
8153 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8154 if (i != V2Index && !Zeroable[i]) {
8155 IsV1Zeroable = false;
8159 // Check for a single input from a SCALAR_TO_VECTOR node.
8160 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
8161 // all the smarts here sunk into that routine. However, the current
8162 // lowering of BUILD_VECTOR makes that nearly impossible until the old
8163 // vector shuffle lowering is dead.
8164 if (SDValue V2S = getScalarValueForVectorElement(
8165 V2, Mask[V2Index] - Mask.size(), DAG)) {
8166 // We need to zext the scalar if it is smaller than an i32.
8167 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
8168 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
8169 // Using zext to expand a narrow element won't work for non-zero
8174 // Zero-extend directly to i32.
8176 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
8178 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
8179 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
8180 EltVT == MVT::i16) {
8181 // Either not inserting from the low element of the input or the input
8182 // element size is too small to use VZEXT_MOVL to clear the high bits.
8186 if (!IsV1Zeroable) {
8187 // If V1 can't be treated as a zero vector we have fewer options to lower
8188 // this. We can't support integer vectors or non-zero targets cheaply, and
8189 // the V1 elements can't be permuted in any way.
8190 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8191 if (!VT.isFloatingPoint() || V2Index != 0)
8193 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8194 V1Mask[V2Index] = -1;
8195 if (!isNoopShuffleMask(V1Mask))
8197 // This is essentially a special case blend operation, but if we have
8198 // general purpose blend operations, they are always faster. Bail and let
8199 // the rest of the lowering handle these as blends.
8200 if (Subtarget->hasSSE41())
8203 // Otherwise, use MOVSD or MOVSS.
8204 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8205 "Only two types of floating point element types to handle!");
8206 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8210 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8212 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8215 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8216 // the desired position. Otherwise it is more efficient to do a vector
8217 // shift left. We know that we can do a vector shift left because all
8218 // the inputs are zero.
8219 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8220 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8221 V2Shuffle[V2Index] = 0;
8222 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8224 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
8226 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8228 V2Index * EltVT.getSizeInBits(),
8229 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
8230 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8236 /// \brief Try to lower broadcast of a single element.
8238 /// For convenience, this code also bundles all of the subtarget feature set
8239 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8240 /// a convenient way to factor it out.
8241 static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V,
8243 const X86Subtarget *Subtarget,
8244 SelectionDAG &DAG) {
8245 if (!Subtarget->hasAVX())
8247 if (VT.isInteger() && !Subtarget->hasAVX2())
8250 // Check that the mask is a broadcast.
8251 int BroadcastIdx = -1;
8253 if (M >= 0 && BroadcastIdx == -1)
8255 else if (M >= 0 && M != BroadcastIdx)
8258 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8259 "a sorted mask where the broadcast "
8262 // Go up the chain of (vector) values to try and find a scalar load that
8263 // we can combine with the broadcast.
8265 switch (V.getOpcode()) {
8266 case ISD::CONCAT_VECTORS: {
8267 int OperandSize = Mask.size() / V.getNumOperands();
8268 V = V.getOperand(BroadcastIdx / OperandSize);
8269 BroadcastIdx %= OperandSize;
8273 case ISD::INSERT_SUBVECTOR: {
8274 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8275 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8279 int BeginIdx = (int)ConstantIdx->getZExtValue();
8281 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
8282 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8283 BroadcastIdx -= BeginIdx;
8294 // Check if this is a broadcast of a scalar. We special case lowering
8295 // for scalars so that we can more effectively fold with loads.
8296 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8297 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8298 V = V.getOperand(BroadcastIdx);
8300 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
8302 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8304 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8305 // We can't broadcast from a vector register w/o AVX2, and we can only
8306 // broadcast from the zero-element of a vector register.
8310 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8313 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8314 // INSERTPS when the V1 elements are already in the correct locations
8315 // because otherwise we can just always use two SHUFPS instructions which
8316 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8317 // perform INSERTPS if a single V1 element is out of place and all V2
8318 // elements are zeroable.
8319 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8321 SelectionDAG &DAG) {
8322 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8323 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8324 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8325 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8327 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8330 int V1DstIndex = -1;
8331 int V2DstIndex = -1;
8332 bool V1UsedInPlace = false;
8334 for (int i = 0; i < 4; i++) {
8335 // Synthesize a zero mask from the zeroable elements (includes undefs).
8341 // Flag if we use any V1 inputs in place.
8343 V1UsedInPlace = true;
8347 // We can only insert a single non-zeroable element.
8348 if (V1DstIndex != -1 || V2DstIndex != -1)
8352 // V1 input out of place for insertion.
8355 // V2 input for insertion.
8360 // Don't bother if we have no (non-zeroable) element for insertion.
8361 if (V1DstIndex == -1 && V2DstIndex == -1)
8364 // Determine element insertion src/dst indices. The src index is from the
8365 // start of the inserted vector, not the start of the concatenated vector.
8366 unsigned V2SrcIndex = 0;
8367 if (V1DstIndex != -1) {
8368 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8369 // and don't use the original V2 at all.
8370 V2SrcIndex = Mask[V1DstIndex];
8371 V2DstIndex = V1DstIndex;
8374 V2SrcIndex = Mask[V2DstIndex] - 4;
8377 // If no V1 inputs are used in place, then the result is created only from
8378 // the zero mask and the V2 insertion - so remove V1 dependency.
8380 V1 = DAG.getUNDEF(MVT::v4f32);
8382 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8383 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8385 // Insert the V2 element into the desired position.
8387 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8388 DAG.getConstant(InsertPSMask, MVT::i8));
8391 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8393 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8394 /// support for floating point shuffles but not integer shuffles. These
8395 /// instructions will incur a domain crossing penalty on some chips though so
8396 /// it is better to avoid lowering through this for integer vectors where
8398 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8399 const X86Subtarget *Subtarget,
8400 SelectionDAG &DAG) {
8402 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8403 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8404 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8405 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8406 ArrayRef<int> Mask = SVOp->getMask();
8407 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8409 if (isSingleInputShuffleMask(Mask)) {
8410 // Use low duplicate instructions for masks that match their pattern.
8411 if (Subtarget->hasSSE3())
8412 if (isShuffleEquivalent(Mask, 0, 0))
8413 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8415 // Straight shuffle of a single input vector. Simulate this by using the
8416 // single input as both of the "inputs" to this instruction..
8417 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8419 if (Subtarget->hasAVX()) {
8420 // If we have AVX, we can use VPERMILPS which will allow folding a load
8421 // into the shuffle.
8422 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8423 DAG.getConstant(SHUFPDMask, MVT::i8));
8426 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
8427 DAG.getConstant(SHUFPDMask, MVT::i8));
8429 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8430 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8432 // Use dedicated unpack instructions for masks that match their pattern.
8433 if (isShuffleEquivalent(Mask, 0, 2))
8434 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8435 if (isShuffleEquivalent(Mask, 1, 3))
8436 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8438 // If we have a single input, insert that into V1 if we can do so cheaply.
8439 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8440 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8441 MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
8443 // Try inverting the insertion since for v2 masks it is easy to do and we
8444 // can't reliably sort the mask one way or the other.
8445 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8446 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8447 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8448 MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG))
8452 // Try to use one of the special instruction patterns to handle two common
8453 // blend patterns if a zero-blend above didn't work.
8454 if (isShuffleEquivalent(Mask, 0, 3) || isShuffleEquivalent(Mask, 1, 3))
8455 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8456 // We can either use a special instruction to load over the low double or
8457 // to move just the low double.
8459 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8461 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8463 if (Subtarget->hasSSE41())
8464 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8468 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8469 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
8470 DAG.getConstant(SHUFPDMask, MVT::i8));
8473 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8475 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8476 /// the integer unit to minimize domain crossing penalties. However, for blends
8477 /// it falls back to the floating point shuffle operation with appropriate bit
8479 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8480 const X86Subtarget *Subtarget,
8481 SelectionDAG &DAG) {
8483 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8484 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8485 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8486 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8487 ArrayRef<int> Mask = SVOp->getMask();
8488 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8490 if (isSingleInputShuffleMask(Mask)) {
8491 // Check for being able to broadcast a single element.
8492 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1,
8493 Mask, Subtarget, DAG))
8496 // Straight shuffle of a single input vector. For everything from SSE2
8497 // onward this has a single fast instruction with no scary immediates.
8498 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8499 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
8500 int WidenedMask[4] = {
8501 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8502 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8504 ISD::BITCAST, DL, MVT::v2i64,
8505 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
8506 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
8509 // Try to use byte shift instructions.
8510 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8511 DL, MVT::v2i64, V1, V2, Mask, DAG))
8514 // If we have a single input from V2 insert that into V1 if we can do so
8516 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8517 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8518 MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
8520 // Try inverting the insertion since for v2 masks it is easy to do and we
8521 // can't reliably sort the mask one way or the other.
8522 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8523 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8524 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8525 MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG))
8529 // Use dedicated unpack instructions for masks that match their pattern.
8530 if (isShuffleEquivalent(Mask, 0, 2))
8531 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8532 if (isShuffleEquivalent(Mask, 1, 3))
8533 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8535 if (Subtarget->hasSSE41())
8536 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8540 // Try to use byte rotation instructions.
8541 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8542 if (Subtarget->hasSSSE3())
8543 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8544 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8547 // We implement this with SHUFPD which is pretty lame because it will likely
8548 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8549 // However, all the alternatives are still more cycles and newer chips don't
8550 // have this problem. It would be really nice if x86 had better shuffles here.
8551 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
8552 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
8553 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
8554 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8557 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8559 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8560 /// It makes no assumptions about whether this is the *best* lowering, it simply
8562 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8563 ArrayRef<int> Mask, SDValue V1,
8564 SDValue V2, SelectionDAG &DAG) {
8565 SDValue LowV = V1, HighV = V2;
8566 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8569 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8571 if (NumV2Elements == 1) {
8573 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8576 // Compute the index adjacent to V2Index and in the same half by toggling
8578 int V2AdjIndex = V2Index ^ 1;
8580 if (Mask[V2AdjIndex] == -1) {
8581 // Handles all the cases where we have a single V2 element and an undef.
8582 // This will only ever happen in the high lanes because we commute the
8583 // vector otherwise.
8585 std::swap(LowV, HighV);
8586 NewMask[V2Index] -= 4;
8588 // Handle the case where the V2 element ends up adjacent to a V1 element.
8589 // To make this work, blend them together as the first step.
8590 int V1Index = V2AdjIndex;
8591 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8592 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8593 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8595 // Now proceed to reconstruct the final blend as we have the necessary
8596 // high or low half formed.
8603 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8604 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8606 } else if (NumV2Elements == 2) {
8607 if (Mask[0] < 4 && Mask[1] < 4) {
8608 // Handle the easy case where we have V1 in the low lanes and V2 in the
8612 } else if (Mask[2] < 4 && Mask[3] < 4) {
8613 // We also handle the reversed case because this utility may get called
8614 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8615 // arrange things in the right direction.
8621 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8622 // trying to place elements directly, just blend them and set up the final
8623 // shuffle to place them.
8625 // The first two blend mask elements are for V1, the second two are for
8627 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8628 Mask[2] < 4 ? Mask[2] : Mask[3],
8629 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8630 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8631 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8632 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8634 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8637 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8638 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8639 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8640 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8643 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8644 getV4X86ShuffleImm8ForMask(NewMask, DAG));
8647 /// \brief Lower 4-lane 32-bit floating point shuffles.
8649 /// Uses instructions exclusively from the floating point unit to minimize
8650 /// domain crossing penalties, as these are sufficient to implement all v4f32
8652 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8653 const X86Subtarget *Subtarget,
8654 SelectionDAG &DAG) {
8656 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8657 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8658 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8659 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8660 ArrayRef<int> Mask = SVOp->getMask();
8661 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8664 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8666 if (NumV2Elements == 0) {
8667 // Check for being able to broadcast a single element.
8668 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1,
8669 Mask, Subtarget, DAG))
8672 // Use even/odd duplicate instructions for masks that match their pattern.
8673 if (Subtarget->hasSSE3()) {
8674 if (isShuffleEquivalent(Mask, 0, 0, 2, 2))
8675 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8676 if (isShuffleEquivalent(Mask, 1, 1, 3, 3))
8677 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8680 if (Subtarget->hasAVX()) {
8681 // If we have AVX, we can use VPERMILPS which will allow folding a load
8682 // into the shuffle.
8683 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8684 getV4X86ShuffleImm8ForMask(Mask, DAG));
8687 // Otherwise, use a straight shuffle of a single input vector. We pass the
8688 // input vector to both operands to simulate this with a SHUFPS.
8689 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8690 getV4X86ShuffleImm8ForMask(Mask, DAG));
8693 // Use dedicated unpack instructions for masks that match their pattern.
8694 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8695 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8696 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8697 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8699 // There are special ways we can lower some single-element blends. However, we
8700 // have custom ways we can lower more complex single-element blends below that
8701 // we defer to if both this and BLENDPS fail to match, so restrict this to
8702 // when the V2 input is targeting element 0 of the mask -- that is the fast
8704 if (NumV2Elements == 1 && Mask[0] >= 4)
8705 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
8706 Mask, Subtarget, DAG))
8709 if (Subtarget->hasSSE41()) {
8710 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8714 // Use INSERTPS if we can complete the shuffle efficiently.
8715 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8719 // Otherwise fall back to a SHUFPS lowering strategy.
8720 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8723 /// \brief Lower 4-lane i32 vector shuffles.
8725 /// We try to handle these with integer-domain shuffles where we can, but for
8726 /// blends we use the floating point domain blend instructions.
8727 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8728 const X86Subtarget *Subtarget,
8729 SelectionDAG &DAG) {
8731 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8732 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8733 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8734 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8735 ArrayRef<int> Mask = SVOp->getMask();
8736 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8738 // Whenever we can lower this as a zext, that instruction is strictly faster
8739 // than any alternative. It also allows us to fold memory operands into the
8740 // shuffle in many cases.
8741 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8742 Mask, Subtarget, DAG))
8746 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8748 if (NumV2Elements == 0) {
8749 // Check for being able to broadcast a single element.
8750 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1,
8751 Mask, Subtarget, DAG))
8754 // Straight shuffle of a single input vector. For everything from SSE2
8755 // onward this has a single fast instruction with no scary immediates.
8756 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8757 // but we aren't actually going to use the UNPCK instruction because doing
8758 // so prevents folding a load into this instruction or making a copy.
8759 const int UnpackLoMask[] = {0, 0, 1, 1};
8760 const int UnpackHiMask[] = {2, 2, 3, 3};
8761 if (isShuffleEquivalent(Mask, 0, 0, 1, 1))
8762 Mask = UnpackLoMask;
8763 else if (isShuffleEquivalent(Mask, 2, 2, 3, 3))
8764 Mask = UnpackHiMask;
8766 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8767 getV4X86ShuffleImm8ForMask(Mask, DAG));
8770 // Try to use bit shift instructions.
8771 if (SDValue Shift = lowerVectorShuffleAsBitShift(
8772 DL, MVT::v4i32, V1, V2, Mask, DAG))
8775 // Try to use byte shift instructions.
8776 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8777 DL, MVT::v4i32, V1, V2, Mask, DAG))
8780 // There are special ways we can lower some single-element blends.
8781 if (NumV2Elements == 1)
8782 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
8783 Mask, Subtarget, DAG))
8786 if (Subtarget->hasSSE41())
8787 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8791 if (SDValue Masked =
8792 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8795 // Use dedicated unpack instructions for masks that match their pattern.
8796 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8797 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8798 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8799 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8801 // Try to use byte rotation instructions.
8802 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8803 if (Subtarget->hasSSSE3())
8804 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8805 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8808 // We implement this with SHUFPS because it can blend from two vectors.
8809 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8810 // up the inputs, bypassing domain shift penalties that we would encur if we
8811 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8813 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
8814 DAG.getVectorShuffle(
8816 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
8817 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
8820 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8821 /// shuffle lowering, and the most complex part.
8823 /// The lowering strategy is to try to form pairs of input lanes which are
8824 /// targeted at the same half of the final vector, and then use a dword shuffle
8825 /// to place them onto the right half, and finally unpack the paired lanes into
8826 /// their final position.
8828 /// The exact breakdown of how to form these dword pairs and align them on the
8829 /// correct sides is really tricky. See the comments within the function for
8830 /// more of the details.
8831 static SDValue lowerV8I16SingleInputVectorShuffle(
8832 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
8833 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8834 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
8835 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8836 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8838 SmallVector<int, 4> LoInputs;
8839 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8840 [](int M) { return M >= 0; });
8841 std::sort(LoInputs.begin(), LoInputs.end());
8842 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8843 SmallVector<int, 4> HiInputs;
8844 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8845 [](int M) { return M >= 0; });
8846 std::sort(HiInputs.begin(), HiInputs.end());
8847 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8849 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8850 int NumHToL = LoInputs.size() - NumLToL;
8852 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8853 int NumHToH = HiInputs.size() - NumLToH;
8854 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8855 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8856 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8857 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8859 // Check for being able to broadcast a single element.
8860 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V,
8861 Mask, Subtarget, DAG))
8864 // Try to use bit shift instructions.
8865 if (SDValue Shift = lowerVectorShuffleAsBitShift(
8866 DL, MVT::v8i16, V, V, Mask, DAG))
8869 // Try to use byte shift instructions.
8870 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8871 DL, MVT::v8i16, V, V, Mask, DAG))
8874 // Use dedicated unpack instructions for masks that match their pattern.
8875 if (isShuffleEquivalent(Mask, 0, 0, 1, 1, 2, 2, 3, 3))
8876 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
8877 if (isShuffleEquivalent(Mask, 4, 4, 5, 5, 6, 6, 7, 7))
8878 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
8880 // Try to use byte rotation instructions.
8881 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8882 DL, MVT::v8i16, V, V, Mask, Subtarget, DAG))
8885 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8886 // such inputs we can swap two of the dwords across the half mark and end up
8887 // with <=2 inputs to each half in each half. Once there, we can fall through
8888 // to the generic code below. For example:
8890 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8891 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8893 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8894 // and an existing 2-into-2 on the other half. In this case we may have to
8895 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8896 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8897 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8898 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8899 // half than the one we target for fixing) will be fixed when we re-enter this
8900 // path. We will also combine away any sequence of PSHUFD instructions that
8901 // result into a single instruction. Here is an example of the tricky case:
8903 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8904 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8906 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8908 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8909 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8911 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8912 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8914 // The result is fine to be handled by the generic logic.
8915 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8916 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8917 int AOffset, int BOffset) {
8918 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8919 "Must call this with A having 3 or 1 inputs from the A half.");
8920 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8921 "Must call this with B having 1 or 3 inputs from the B half.");
8922 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8923 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8925 // Compute the index of dword with only one word among the three inputs in
8926 // a half by taking the sum of the half with three inputs and subtracting
8927 // the sum of the actual three inputs. The difference is the remaining
8930 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8931 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8932 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8933 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8934 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8935 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8936 int TripleNonInputIdx =
8937 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8938 TripleDWord = TripleNonInputIdx / 2;
8940 // We use xor with one to compute the adjacent DWord to whichever one the
8942 OneInputDWord = (OneInput / 2) ^ 1;
8944 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8945 // and BToA inputs. If there is also such a problem with the BToB and AToB
8946 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8947 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8948 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8949 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8950 // Compute how many inputs will be flipped by swapping these DWords. We
8952 // to balance this to ensure we don't form a 3-1 shuffle in the other
8954 int NumFlippedAToBInputs =
8955 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8956 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8957 int NumFlippedBToBInputs =
8958 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8959 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8960 if ((NumFlippedAToBInputs == 1 &&
8961 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8962 (NumFlippedBToBInputs == 1 &&
8963 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8964 // We choose whether to fix the A half or B half based on whether that
8965 // half has zero flipped inputs. At zero, we may not be able to fix it
8966 // with that half. We also bias towards fixing the B half because that
8967 // will more commonly be the high half, and we have to bias one way.
8968 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8969 ArrayRef<int> Inputs) {
8970 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8971 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8972 PinnedIdx ^ 1) != Inputs.end();
8973 // Determine whether the free index is in the flipped dword or the
8974 // unflipped dword based on where the pinned index is. We use this bit
8975 // in an xor to conditionally select the adjacent dword.
8976 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8977 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8978 FixFreeIdx) != Inputs.end();
8979 if (IsFixIdxInput == IsFixFreeIdxInput)
8981 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8982 FixFreeIdx) != Inputs.end();
8983 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8984 "We need to be changing the number of flipped inputs!");
8985 int PSHUFHalfMask[] = {0, 1, 2, 3};
8986 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8987 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8989 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
8992 if (M != -1 && M == FixIdx)
8994 else if (M != -1 && M == FixFreeIdx)
8997 if (NumFlippedBToBInputs != 0) {
8999 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9000 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
9002 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
9004 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
9005 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
9010 int PSHUFDMask[] = {0, 1, 2, 3};
9011 PSHUFDMask[ADWord] = BDWord;
9012 PSHUFDMask[BDWord] = ADWord;
9013 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9014 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
9015 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
9016 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9018 // Adjust the mask to match the new locations of A and B.
9020 if (M != -1 && M/2 == ADWord)
9021 M = 2 * BDWord + M % 2;
9022 else if (M != -1 && M/2 == BDWord)
9023 M = 2 * ADWord + M % 2;
9025 // Recurse back into this routine to re-compute state now that this isn't
9026 // a 3 and 1 problem.
9027 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
9030 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
9031 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
9032 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
9033 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
9035 // At this point there are at most two inputs to the low and high halves from
9036 // each half. That means the inputs can always be grouped into dwords and
9037 // those dwords can then be moved to the correct half with a dword shuffle.
9038 // We use at most one low and one high word shuffle to collect these paired
9039 // inputs into dwords, and finally a dword shuffle to place them.
9040 int PSHUFLMask[4] = {-1, -1, -1, -1};
9041 int PSHUFHMask[4] = {-1, -1, -1, -1};
9042 int PSHUFDMask[4] = {-1, -1, -1, -1};
9044 // First fix the masks for all the inputs that are staying in their
9045 // original halves. This will then dictate the targets of the cross-half
9047 auto fixInPlaceInputs =
9048 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
9049 MutableArrayRef<int> SourceHalfMask,
9050 MutableArrayRef<int> HalfMask, int HalfOffset) {
9051 if (InPlaceInputs.empty())
9053 if (InPlaceInputs.size() == 1) {
9054 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9055 InPlaceInputs[0] - HalfOffset;
9056 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
9059 if (IncomingInputs.empty()) {
9060 // Just fix all of the in place inputs.
9061 for (int Input : InPlaceInputs) {
9062 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
9063 PSHUFDMask[Input / 2] = Input / 2;
9068 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
9069 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
9070 InPlaceInputs[0] - HalfOffset;
9071 // Put the second input next to the first so that they are packed into
9072 // a dword. We find the adjacent index by toggling the low bit.
9073 int AdjIndex = InPlaceInputs[0] ^ 1;
9074 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
9075 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
9076 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
9078 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
9079 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
9081 // Now gather the cross-half inputs and place them into a free dword of
9082 // their target half.
9083 // FIXME: This operation could almost certainly be simplified dramatically to
9084 // look more like the 3-1 fixing operation.
9085 auto moveInputsToRightHalf = [&PSHUFDMask](
9086 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
9087 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
9088 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
9090 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
9091 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
9093 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
9095 int LowWord = Word & ~1;
9096 int HighWord = Word | 1;
9097 return isWordClobbered(SourceHalfMask, LowWord) ||
9098 isWordClobbered(SourceHalfMask, HighWord);
9101 if (IncomingInputs.empty())
9104 if (ExistingInputs.empty()) {
9105 // Map any dwords with inputs from them into the right half.
9106 for (int Input : IncomingInputs) {
9107 // If the source half mask maps over the inputs, turn those into
9108 // swaps and use the swapped lane.
9109 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
9110 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
9111 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
9112 Input - SourceOffset;
9113 // We have to swap the uses in our half mask in one sweep.
9114 for (int &M : HalfMask)
9115 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
9117 else if (M == Input)
9118 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9120 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
9121 Input - SourceOffset &&
9122 "Previous placement doesn't match!");
9124 // Note that this correctly re-maps both when we do a swap and when
9125 // we observe the other side of the swap above. We rely on that to
9126 // avoid swapping the members of the input list directly.
9127 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
9130 // Map the input's dword into the correct half.
9131 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
9132 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
9134 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
9136 "Previous placement doesn't match!");
9139 // And just directly shift any other-half mask elements to be same-half
9140 // as we will have mirrored the dword containing the element into the
9141 // same position within that half.
9142 for (int &M : HalfMask)
9143 if (M >= SourceOffset && M < SourceOffset + 4) {
9144 M = M - SourceOffset + DestOffset;
9145 assert(M >= 0 && "This should never wrap below zero!");
9150 // Ensure we have the input in a viable dword of its current half. This
9151 // is particularly tricky because the original position may be clobbered
9152 // by inputs being moved and *staying* in that half.
9153 if (IncomingInputs.size() == 1) {
9154 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9155 int InputFixed = std::find(std::begin(SourceHalfMask),
9156 std::end(SourceHalfMask), -1) -
9157 std::begin(SourceHalfMask) + SourceOffset;
9158 SourceHalfMask[InputFixed - SourceOffset] =
9159 IncomingInputs[0] - SourceOffset;
9160 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
9162 IncomingInputs[0] = InputFixed;
9164 } else if (IncomingInputs.size() == 2) {
9165 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
9166 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
9167 // We have two non-adjacent or clobbered inputs we need to extract from
9168 // the source half. To do this, we need to map them into some adjacent
9169 // dword slot in the source mask.
9170 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9171 IncomingInputs[1] - SourceOffset};
9173 // If there is a free slot in the source half mask adjacent to one of
9174 // the inputs, place the other input in it. We use (Index XOR 1) to
9175 // compute an adjacent index.
9176 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9177 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9178 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9179 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9180 InputsFixed[1] = InputsFixed[0] ^ 1;
9181 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9182 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9183 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9184 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9185 InputsFixed[0] = InputsFixed[1] ^ 1;
9186 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9187 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9188 // The two inputs are in the same DWord but it is clobbered and the
9189 // adjacent DWord isn't used at all. Move both inputs to the free
9191 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9192 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9193 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9194 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9196 // The only way we hit this point is if there is no clobbering
9197 // (because there are no off-half inputs to this half) and there is no
9198 // free slot adjacent to one of the inputs. In this case, we have to
9199 // swap an input with a non-input.
9200 for (int i = 0; i < 4; ++i)
9201 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9202 "We can't handle any clobbers here!");
9203 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9204 "Cannot have adjacent inputs here!");
9206 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9207 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9209 // We also have to update the final source mask in this case because
9210 // it may need to undo the above swap.
9211 for (int &M : FinalSourceHalfMask)
9212 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9213 M = InputsFixed[1] + SourceOffset;
9214 else if (M == InputsFixed[1] + SourceOffset)
9215 M = (InputsFixed[0] ^ 1) + SourceOffset;
9217 InputsFixed[1] = InputsFixed[0] ^ 1;
9220 // Point everything at the fixed inputs.
9221 for (int &M : HalfMask)
9222 if (M == IncomingInputs[0])
9223 M = InputsFixed[0] + SourceOffset;
9224 else if (M == IncomingInputs[1])
9225 M = InputsFixed[1] + SourceOffset;
9227 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9228 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9231 llvm_unreachable("Unhandled input size!");
9234 // Now hoist the DWord down to the right half.
9235 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9236 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9237 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9238 for (int &M : HalfMask)
9239 for (int Input : IncomingInputs)
9241 M = FreeDWord * 2 + Input % 2;
9243 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9244 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9245 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9246 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9248 // Now enact all the shuffles we've computed to move the inputs into their
9250 if (!isNoopShuffleMask(PSHUFLMask))
9251 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9252 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
9253 if (!isNoopShuffleMask(PSHUFHMask))
9254 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9255 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
9256 if (!isNoopShuffleMask(PSHUFDMask))
9257 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9258 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
9259 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
9260 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9262 // At this point, each half should contain all its inputs, and we can then
9263 // just shuffle them into their final position.
9264 assert(std::count_if(LoMask.begin(), LoMask.end(),
9265 [](int M) { return M >= 4; }) == 0 &&
9266 "Failed to lift all the high half inputs to the low mask!");
9267 assert(std::count_if(HiMask.begin(), HiMask.end(),
9268 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9269 "Failed to lift all the low half inputs to the high mask!");
9271 // Do a half shuffle for the low mask.
9272 if (!isNoopShuffleMask(LoMask))
9273 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9274 getV4X86ShuffleImm8ForMask(LoMask, DAG));
9276 // Do a half shuffle with the high mask after shifting its values down.
9277 for (int &M : HiMask)
9280 if (!isNoopShuffleMask(HiMask))
9281 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9282 getV4X86ShuffleImm8ForMask(HiMask, DAG));
9287 /// \brief Detect whether the mask pattern should be lowered through
9290 /// This essentially tests whether viewing the mask as an interleaving of two
9291 /// sub-sequences reduces the cross-input traffic of a blend operation. If so,
9292 /// lowering it through interleaving is a significantly better strategy.
9293 static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
9294 int NumEvenInputs[2] = {0, 0};
9295 int NumOddInputs[2] = {0, 0};
9296 int NumLoInputs[2] = {0, 0};
9297 int NumHiInputs[2] = {0, 0};
9298 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
9302 int InputIdx = Mask[i] >= Size;
9305 ++NumLoInputs[InputIdx];
9307 ++NumHiInputs[InputIdx];
9310 ++NumEvenInputs[InputIdx];
9312 ++NumOddInputs[InputIdx];
9315 // The minimum number of cross-input results for both the interleaved and
9316 // split cases. If interleaving results in fewer cross-input results, return
9318 int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
9319 NumEvenInputs[0] + NumOddInputs[1]);
9320 int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
9321 NumLoInputs[0] + NumHiInputs[1]);
9322 return InterleavedCrosses < SplitCrosses;
9325 /// \brief Blend two v8i16 vectors using a naive unpack strategy.
9327 /// This strategy only works when the inputs from each vector fit into a single
9328 /// half of that vector, and generally there are not so many inputs as to leave
9329 /// the in-place shuffles required highly constrained (and thus expensive). It
9330 /// shifts all the inputs into a single side of both input vectors and then
9331 /// uses an unpack to interleave these inputs in a single vector. At that
9332 /// point, we will fall back on the generic single input shuffle lowering.
9333 static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
9335 MutableArrayRef<int> Mask,
9336 const X86Subtarget *Subtarget,
9337 SelectionDAG &DAG) {
9338 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9339 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9340 SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
9341 for (int i = 0; i < 8; ++i)
9342 if (Mask[i] >= 0 && Mask[i] < 4)
9343 LoV1Inputs.push_back(i);
9344 else if (Mask[i] >= 4 && Mask[i] < 8)
9345 HiV1Inputs.push_back(i);
9346 else if (Mask[i] >= 8 && Mask[i] < 12)
9347 LoV2Inputs.push_back(i);
9348 else if (Mask[i] >= 12)
9349 HiV2Inputs.push_back(i);
9351 int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
9352 int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
9355 assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
9356 assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
9357 assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
9359 bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
9360 HiV1Inputs.size() + HiV2Inputs.size();
9362 auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
9363 ArrayRef<int> HiInputs, bool MoveToLo,
9365 ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
9366 ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
9367 if (BadInputs.empty())
9370 int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9371 int MoveOffset = MoveToLo ? 0 : 4;
9373 if (GoodInputs.empty()) {
9374 for (int BadInput : BadInputs) {
9375 MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
9376 Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
9379 if (GoodInputs.size() == 2) {
9380 // If the low inputs are spread across two dwords, pack them into
9382 MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
9383 MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
9384 Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
9385 Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
9387 // Otherwise pin the good inputs.
9388 for (int GoodInput : GoodInputs)
9389 MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
9392 if (BadInputs.size() == 2) {
9393 // If we have two bad inputs then there may be either one or two good
9394 // inputs fixed in place. Find a fixed input, and then find the *other*
9395 // two adjacent indices by using modular arithmetic.
9397 std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
9398 [](int M) { return M >= 0; }) -
9399 std::begin(MoveMask);
9401 ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset;
9402 assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
9403 assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
9404 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9405 MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
9406 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9407 Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
9409 assert(BadInputs.size() == 1 && "All sizes handled");
9410 int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset,
9411 std::end(MoveMask), -1) -
9412 std::begin(MoveMask);
9413 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9414 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9418 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
9421 V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
9423 V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
9426 // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
9427 // cross-half traffic in the final shuffle.
9429 // Munge the mask to be a single-input mask after the unpack merges the
9433 M = 2 * (M % 4) + (M / 8);
9435 return DAG.getVectorShuffle(
9436 MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
9437 DL, MVT::v8i16, V1, V2),
9438 DAG.getUNDEF(MVT::v8i16), Mask);
9441 /// \brief Generic lowering of 8-lane i16 shuffles.
9443 /// This handles both single-input shuffles and combined shuffle/blends with
9444 /// two inputs. The single input shuffles are immediately delegated to
9445 /// a dedicated lowering routine.
9447 /// The blends are lowered in one of three fundamental ways. If there are few
9448 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9449 /// of the input is significantly cheaper when lowered as an interleaving of
9450 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9451 /// halves of the inputs separately (making them have relatively few inputs)
9452 /// and then concatenate them.
9453 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9454 const X86Subtarget *Subtarget,
9455 SelectionDAG &DAG) {
9457 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9458 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9459 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9460 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9461 ArrayRef<int> OrigMask = SVOp->getMask();
9462 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9463 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9464 MutableArrayRef<int> Mask(MaskStorage);
9466 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9468 // Whenever we can lower this as a zext, that instruction is strictly faster
9469 // than any alternative.
9470 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9471 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9474 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9475 auto isV2 = [](int M) { return M >= 8; };
9477 int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
9478 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9480 if (NumV2Inputs == 0)
9481 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
9483 assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
9484 "to be V1-input shuffles.");
9486 // Try to use bit shift instructions.
9487 if (SDValue Shift = lowerVectorShuffleAsBitShift(
9488 DL, MVT::v8i16, V1, V2, Mask, DAG))
9491 // Try to use byte shift instructions.
9492 if (SDValue Shift = lowerVectorShuffleAsByteShift(
9493 DL, MVT::v8i16, V1, V2, Mask, DAG))
9496 // There are special ways we can lower some single-element blends.
9497 if (NumV2Inputs == 1)
9498 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
9499 Mask, Subtarget, DAG))
9502 if (Subtarget->hasSSE41())
9503 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9507 if (SDValue Masked =
9508 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9511 // Use dedicated unpack instructions for masks that match their pattern.
9512 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 2, 10, 3, 11))
9513 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9514 if (isShuffleEquivalent(Mask, 4, 12, 5, 13, 6, 14, 7, 15))
9515 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9517 // Try to use byte rotation instructions.
9518 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9519 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9522 if (NumV1Inputs + NumV2Inputs <= 4)
9523 return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
9525 // Check whether an interleaving lowering is likely to be more efficient.
9526 // This isn't perfect but it is a strong heuristic that tends to work well on
9527 // the kinds of shuffles that show up in practice.
9529 // FIXME: Handle 1x, 2x, and 4x interleaving.
9530 if (shouldLowerAsInterleaving(Mask)) {
9531 // FIXME: Figure out whether we should pack these into the low or high
9534 int EMask[8], OMask[8];
9535 for (int i = 0; i < 4; ++i) {
9536 EMask[i] = Mask[2*i];
9537 OMask[i] = Mask[2*i + 1];
9542 SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
9543 SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
9545 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
9548 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9549 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9551 for (int i = 0; i < 4; ++i) {
9552 LoBlendMask[i] = Mask[i];
9553 HiBlendMask[i] = Mask[i + 4];
9556 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
9557 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
9558 LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
9559 HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
9561 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9562 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
9565 /// \brief Check whether a compaction lowering can be done by dropping even
9566 /// elements and compute how many times even elements must be dropped.
9568 /// This handles shuffles which take every Nth element where N is a power of
9569 /// two. Example shuffle masks:
9571 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9572 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9573 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9574 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9575 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9576 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9578 /// Any of these lanes can of course be undef.
9580 /// This routine only supports N <= 3.
9581 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9584 /// \returns N above, or the number of times even elements must be dropped if
9585 /// there is such a number. Otherwise returns zero.
9586 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9587 // Figure out whether we're looping over two inputs or just one.
9588 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9590 // The modulus for the shuffle vector entries is based on whether this is
9591 // a single input or not.
9592 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9593 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9594 "We should only be called with masks with a power-of-2 size!");
9596 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9598 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9599 // and 2^3 simultaneously. This is because we may have ambiguity with
9600 // partially undef inputs.
9601 bool ViableForN[3] = {true, true, true};
9603 for (int i = 0, e = Mask.size(); i < e; ++i) {
9604 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9609 bool IsAnyViable = false;
9610 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9611 if (ViableForN[j]) {
9614 // The shuffle mask must be equal to (i * 2^N) % M.
9615 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9618 ViableForN[j] = false;
9620 // Early exit if we exhaust the possible powers of two.
9625 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9629 // Return 0 as there is no viable power of two.
9633 /// \brief Generic lowering of v16i8 shuffles.
9635 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9636 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9637 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9638 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9640 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9641 const X86Subtarget *Subtarget,
9642 SelectionDAG &DAG) {
9644 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9645 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9646 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9647 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9648 ArrayRef<int> OrigMask = SVOp->getMask();
9649 assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9651 // Try to use bit shift instructions.
9652 if (SDValue Shift = lowerVectorShuffleAsBitShift(
9653 DL, MVT::v16i8, V1, V2, OrigMask, DAG))
9656 // Try to use byte shift instructions.
9657 if (SDValue Shift = lowerVectorShuffleAsByteShift(
9658 DL, MVT::v16i8, V1, V2, OrigMask, DAG))
9661 // Try to use byte rotation instructions.
9662 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9663 DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
9666 // Try to use a zext lowering.
9667 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9668 DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
9671 int MaskStorage[16] = {
9672 OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9673 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7],
9674 OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11],
9675 OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
9676 MutableArrayRef<int> Mask(MaskStorage);
9677 MutableArrayRef<int> LoMask = Mask.slice(0, 8);
9678 MutableArrayRef<int> HiMask = Mask.slice(8, 8);
9681 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9683 // For single-input shuffles, there are some nicer lowering tricks we can use.
9684 if (NumV2Elements == 0) {
9685 // Check for being able to broadcast a single element.
9686 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1,
9687 Mask, Subtarget, DAG))
9690 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9691 // Notably, this handles splat and partial-splat shuffles more efficiently.
9692 // However, it only makes sense if the pre-duplication shuffle simplifies
9693 // things significantly. Currently, this means we need to be able to
9694 // express the pre-duplication shuffle as an i16 shuffle.
9696 // FIXME: We should check for other patterns which can be widened into an
9697 // i16 shuffle as well.
9698 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9699 for (int i = 0; i < 16; i += 2)
9700 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9705 auto tryToWidenViaDuplication = [&]() -> SDValue {
9706 if (!canWidenViaDuplication(Mask))
9708 SmallVector<int, 4> LoInputs;
9709 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9710 [](int M) { return M >= 0 && M < 8; });
9711 std::sort(LoInputs.begin(), LoInputs.end());
9712 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9714 SmallVector<int, 4> HiInputs;
9715 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9716 [](int M) { return M >= 8; });
9717 std::sort(HiInputs.begin(), HiInputs.end());
9718 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9721 bool TargetLo = LoInputs.size() >= HiInputs.size();
9722 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9723 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9725 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9726 SmallDenseMap<int, int, 8> LaneMap;
9727 for (int I : InPlaceInputs) {
9728 PreDupI16Shuffle[I/2] = I/2;
9731 int j = TargetLo ? 0 : 4, je = j + 4;
9732 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9733 // Check if j is already a shuffle of this input. This happens when
9734 // there are two adjacent bytes after we move the low one.
9735 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9736 // If we haven't yet mapped the input, search for a slot into which
9738 while (j < je && PreDupI16Shuffle[j] != -1)
9742 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9745 // Map this input with the i16 shuffle.
9746 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9749 // Update the lane map based on the mapping we ended up with.
9750 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9753 ISD::BITCAST, DL, MVT::v16i8,
9754 DAG.getVectorShuffle(MVT::v8i16, DL,
9755 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9756 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9758 // Unpack the bytes to form the i16s that will be shuffled into place.
9759 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9760 MVT::v16i8, V1, V1);
9762 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9763 for (int i = 0; i < 16; ++i)
9764 if (Mask[i] != -1) {
9765 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9766 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9767 if (PostDupI16Shuffle[i / 2] == -1)
9768 PostDupI16Shuffle[i / 2] = MappedMask;
9770 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9771 "Conflicting entrties in the original shuffle!");
9774 ISD::BITCAST, DL, MVT::v16i8,
9775 DAG.getVectorShuffle(MVT::v8i16, DL,
9776 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9777 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9779 if (SDValue V = tryToWidenViaDuplication())
9783 // Check whether an interleaving lowering is likely to be more efficient.
9784 // This isn't perfect but it is a strong heuristic that tends to work well on
9785 // the kinds of shuffles that show up in practice.
9787 // FIXME: We need to handle other interleaving widths (i16, i32, ...).
9788 if (shouldLowerAsInterleaving(Mask)) {
9789 int NumLoHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
9790 return (M >= 0 && M < 8) || (M >= 16 && M < 24);
9792 int NumHiHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
9793 return (M >= 8 && M < 16) || M >= 24;
9795 int EMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
9796 -1, -1, -1, -1, -1, -1, -1, -1};
9797 int OMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
9798 -1, -1, -1, -1, -1, -1, -1, -1};
9799 bool UnpackLo = NumLoHalf >= NumHiHalf;
9800 MutableArrayRef<int> TargetEMask(UnpackLo ? EMask : EMask + 8, 8);
9801 MutableArrayRef<int> TargetOMask(UnpackLo ? OMask : OMask + 8, 8);
9802 for (int i = 0; i < 8; ++i) {
9803 TargetEMask[i] = Mask[2 * i];
9804 TargetOMask[i] = Mask[2 * i + 1];
9807 SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
9808 SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
9810 return DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9811 MVT::v16i8, Evens, Odds);
9814 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9815 // with PSHUFB. It is important to do this before we attempt to generate any
9816 // blends but after all of the single-input lowerings. If the single input
9817 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9818 // want to preserve that and we can DAG combine any longer sequences into
9819 // a PSHUFB in the end. But once we start blending from multiple inputs,
9820 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9821 // and there are *very* few patterns that would actually be faster than the
9822 // PSHUFB approach because of its ability to zero lanes.
9824 // FIXME: The only exceptions to the above are blends which are exact
9825 // interleavings with direct instructions supporting them. We currently don't
9826 // handle those well here.
9827 if (Subtarget->hasSSSE3()) {
9830 bool V1InUse = false;
9831 bool V2InUse = false;
9832 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9834 for (int i = 0; i < 16; ++i) {
9835 if (Mask[i] == -1) {
9836 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9838 const int ZeroMask = 0x80;
9839 int V1Idx = (Mask[i] < 16 ? Mask[i] : ZeroMask);
9840 int V2Idx = (Mask[i] < 16 ? ZeroMask : Mask[i] - 16);
9842 V1Idx = V2Idx = ZeroMask;
9843 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
9844 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
9845 V1InUse |= (ZeroMask != V1Idx);
9846 V2InUse |= (ZeroMask != V2Idx);
9851 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1,
9852 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9854 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2,
9855 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9857 // If we need shuffled inputs from both, blend the two.
9858 if (V1InUse && V2InUse)
9859 return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9861 return V1; // Single inputs are easy.
9863 return V2; // Single inputs are easy.
9864 // Shuffling to a zeroable vector.
9865 return getZeroVector(MVT::v16i8, Subtarget, DAG, DL);
9868 // There are special ways we can lower some single-element blends.
9869 if (NumV2Elements == 1)
9870 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2,
9871 Mask, Subtarget, DAG))
9874 // Check whether a compaction lowering can be done. This handles shuffles
9875 // which take every Nth element for some even N. See the helper function for
9878 // We special case these as they can be particularly efficiently handled with
9879 // the PACKUSB instruction on x86 and they show up in common patterns of
9880 // rearranging bytes to truncate wide elements.
9881 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9882 // NumEvenDrops is the power of two stride of the elements. Another way of
9883 // thinking about it is that we need to drop the even elements this many
9884 // times to get the original input.
9885 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9887 // First we need to zero all the dropped bytes.
9888 assert(NumEvenDrops <= 3 &&
9889 "No support for dropping even elements more than 3 times.");
9890 // We use the mask type to pick which bytes are preserved based on how many
9891 // elements are dropped.
9892 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9893 SDValue ByteClearMask =
9894 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
9895 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
9896 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9898 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9900 // Now pack things back together.
9901 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
9902 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
9903 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9904 for (int i = 1; i < NumEvenDrops; ++i) {
9905 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
9906 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9912 int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9913 int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9914 int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9915 int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9917 auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
9918 MutableArrayRef<int> V1HalfBlendMask,
9919 MutableArrayRef<int> V2HalfBlendMask) {
9920 for (int i = 0; i < 8; ++i)
9921 if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
9922 V1HalfBlendMask[i] = HalfMask[i];
9924 } else if (HalfMask[i] >= 16) {
9925 V2HalfBlendMask[i] = HalfMask[i] - 16;
9926 HalfMask[i] = i + 8;
9929 buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
9930 buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
9932 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9934 auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
9935 MutableArrayRef<int> HiBlendMask) {
9937 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9938 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9940 if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
9941 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9942 std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
9943 [](int M) { return M >= 0 && M % 2 == 1; })) {
9944 // Use a mask to drop the high bytes.
9945 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
9946 V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
9947 DAG.getConstant(0x00FF, MVT::v8i16));
9949 // This will be a single vector shuffle instead of a blend so nuke V2.
9950 V2 = DAG.getUNDEF(MVT::v8i16);
9952 // Squash the masks to point directly into V1.
9953 for (int &M : LoBlendMask)
9956 for (int &M : HiBlendMask)
9960 // Otherwise just unpack the low half of V into V1 and the high half into
9961 // V2 so that we can blend them as i16s.
9962 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9963 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9964 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9965 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9968 SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
9969 SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
9970 return std::make_pair(BlendedLo, BlendedHi);
9972 SDValue V1Lo, V1Hi, V2Lo, V2Hi;
9973 std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
9974 std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
9976 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
9977 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
9979 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9982 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9984 /// This routine breaks down the specific type of 128-bit shuffle and
9985 /// dispatches to the lowering routines accordingly.
9986 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9987 MVT VT, const X86Subtarget *Subtarget,
9988 SelectionDAG &DAG) {
9989 switch (VT.SimpleTy) {
9991 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9993 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9995 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9997 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9999 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10001 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10004 llvm_unreachable("Unimplemented!");
10008 /// \brief Helper function to test whether a shuffle mask could be
10009 /// simplified by widening the elements being shuffled.
10011 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
10012 /// leaves it in an unspecified state.
10014 /// NOTE: This must handle normal vector shuffle masks and *target* vector
10015 /// shuffle masks. The latter have the special property of a '-2' representing
10016 /// a zero-ed lane of a vector.
10017 static bool canWidenShuffleElements(ArrayRef<int> Mask,
10018 SmallVectorImpl<int> &WidenedMask) {
10019 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
10020 // If both elements are undef, its trivial.
10021 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
10022 WidenedMask.push_back(SM_SentinelUndef);
10026 // Check for an undef mask and a mask value properly aligned to fit with
10027 // a pair of values. If we find such a case, use the non-undef mask's value.
10028 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
10029 WidenedMask.push_back(Mask[i + 1] / 2);
10032 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
10033 WidenedMask.push_back(Mask[i] / 2);
10037 // When zeroing, we need to spread the zeroing across both lanes to widen.
10038 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
10039 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
10040 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
10041 WidenedMask.push_back(SM_SentinelZero);
10047 // Finally check if the two mask values are adjacent and aligned with
10049 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
10050 WidenedMask.push_back(Mask[i] / 2);
10054 // Otherwise we can't safely widen the elements used in this shuffle.
10057 assert(WidenedMask.size() == Mask.size() / 2 &&
10058 "Incorrect size of mask after widening the elements!");
10063 /// \brief Generic routine to split ector shuffle into half-sized shuffles.
10065 /// This routine just extracts two subvectors, shuffles them independently, and
10066 /// then concatenates them back together. This should work effectively with all
10067 /// AVX vector shuffle types.
10068 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10069 SDValue V2, ArrayRef<int> Mask,
10070 SelectionDAG &DAG) {
10071 assert(VT.getSizeInBits() >= 256 &&
10072 "Only for 256-bit or wider vector shuffles!");
10073 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
10074 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
10076 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
10077 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
10079 int NumElements = VT.getVectorNumElements();
10080 int SplitNumElements = NumElements / 2;
10081 MVT ScalarVT = VT.getScalarType();
10082 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
10084 SDValue LoV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
10085 DAG.getIntPtrConstant(0));
10086 SDValue HiV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
10087 DAG.getIntPtrConstant(SplitNumElements));
10088 SDValue LoV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
10089 DAG.getIntPtrConstant(0));
10090 SDValue HiV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
10091 DAG.getIntPtrConstant(SplitNumElements));
10093 // Now create two 4-way blends of these half-width vectors.
10094 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
10095 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
10096 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
10097 for (int i = 0; i < SplitNumElements; ++i) {
10098 int M = HalfMask[i];
10099 if (M >= NumElements) {
10100 if (M >= NumElements + SplitNumElements)
10104 V2BlendMask.push_back(M - NumElements);
10105 V1BlendMask.push_back(-1);
10106 BlendMask.push_back(SplitNumElements + i);
10107 } else if (M >= 0) {
10108 if (M >= SplitNumElements)
10112 V2BlendMask.push_back(-1);
10113 V1BlendMask.push_back(M);
10114 BlendMask.push_back(i);
10116 V2BlendMask.push_back(-1);
10117 V1BlendMask.push_back(-1);
10118 BlendMask.push_back(-1);
10122 // Because the lowering happens after all combining takes place, we need to
10123 // manually combine these blend masks as much as possible so that we create
10124 // a minimal number of high-level vector shuffle nodes.
10126 // First try just blending the halves of V1 or V2.
10127 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
10128 return DAG.getUNDEF(SplitVT);
10129 if (!UseLoV2 && !UseHiV2)
10130 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10131 if (!UseLoV1 && !UseHiV1)
10132 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10134 SDValue V1Blend, V2Blend;
10135 if (UseLoV1 && UseHiV1) {
10137 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
10139 // We only use half of V1 so map the usage down into the final blend mask.
10140 V1Blend = UseLoV1 ? LoV1 : HiV1;
10141 for (int i = 0; i < SplitNumElements; ++i)
10142 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
10143 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
10145 if (UseLoV2 && UseHiV2) {
10147 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
10149 // We only use half of V2 so map the usage down into the final blend mask.
10150 V2Blend = UseLoV2 ? LoV2 : HiV2;
10151 for (int i = 0; i < SplitNumElements; ++i)
10152 if (BlendMask[i] >= SplitNumElements)
10153 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
10155 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
10157 SDValue Lo = HalfBlend(LoMask);
10158 SDValue Hi = HalfBlend(HiMask);
10159 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
10162 /// \brief Either split a vector in halves or decompose the shuffles and the
10165 /// This is provided as a good fallback for many lowerings of non-single-input
10166 /// shuffles with more than one 128-bit lane. In those cases, we want to select
10167 /// between splitting the shuffle into 128-bit components and stitching those
10168 /// back together vs. extracting the single-input shuffles and blending those
10170 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
10171 SDValue V2, ArrayRef<int> Mask,
10172 SelectionDAG &DAG) {
10173 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
10174 "lower single-input shuffles as it "
10175 "could then recurse on itself.");
10176 int Size = Mask.size();
10178 // If this can be modeled as a broadcast of two elements followed by a blend,
10179 // prefer that lowering. This is especially important because broadcasts can
10180 // often fold with memory operands.
10181 auto DoBothBroadcast = [&] {
10182 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
10185 if (V2BroadcastIdx == -1)
10186 V2BroadcastIdx = M - Size;
10187 else if (M - Size != V2BroadcastIdx)
10189 } else if (M >= 0) {
10190 if (V1BroadcastIdx == -1)
10191 V1BroadcastIdx = M;
10192 else if (M != V1BroadcastIdx)
10197 if (DoBothBroadcast())
10198 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10201 // If the inputs all stem from a single 128-bit lane of each input, then we
10202 // split them rather than blending because the split will decompose to
10203 // unusually few instructions.
10204 int LaneCount = VT.getSizeInBits() / 128;
10205 int LaneSize = Size / LaneCount;
10206 SmallBitVector LaneInputs[2];
10207 LaneInputs[0].resize(LaneCount, false);
10208 LaneInputs[1].resize(LaneCount, false);
10209 for (int i = 0; i < Size; ++i)
10211 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10212 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10213 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10215 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10216 // that the decomposed single-input shuffles don't end up here.
10217 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10220 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10221 /// a permutation and blend of those lanes.
10223 /// This essentially blends the out-of-lane inputs to each lane into the lane
10224 /// from a permuted copy of the vector. This lowering strategy results in four
10225 /// instructions in the worst case for a single-input cross lane shuffle which
10226 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10227 /// of. Special cases for each particular shuffle pattern should be handled
10228 /// prior to trying this lowering.
10229 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10230 SDValue V1, SDValue V2,
10231 ArrayRef<int> Mask,
10232 SelectionDAG &DAG) {
10233 // FIXME: This should probably be generalized for 512-bit vectors as well.
10234 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
10235 int LaneSize = Mask.size() / 2;
10237 // If there are only inputs from one 128-bit lane, splitting will in fact be
10238 // less expensive. The flags track wether the given lane contains an element
10239 // that crosses to another lane.
10240 bool LaneCrossing[2] = {false, false};
10241 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10242 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10243 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10244 if (!LaneCrossing[0] || !LaneCrossing[1])
10245 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10247 if (isSingleInputShuffleMask(Mask)) {
10248 SmallVector<int, 32> FlippedBlendMask;
10249 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10250 FlippedBlendMask.push_back(
10251 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10253 : Mask[i] % LaneSize +
10254 (i / LaneSize) * LaneSize + Size));
10256 // Flip the vector, and blend the results which should now be in-lane. The
10257 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10258 // 5 for the high source. The value 3 selects the high half of source 2 and
10259 // the value 2 selects the low half of source 2. We only use source 2 to
10260 // allow folding it into a memory operand.
10261 unsigned PERMMask = 3 | 2 << 4;
10262 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10263 V1, DAG.getConstant(PERMMask, MVT::i8));
10264 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10267 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10268 // will be handled by the above logic and a blend of the results, much like
10269 // other patterns in AVX.
10270 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10273 /// \brief Handle lowering 2-lane 128-bit shuffles.
10274 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10275 SDValue V2, ArrayRef<int> Mask,
10276 const X86Subtarget *Subtarget,
10277 SelectionDAG &DAG) {
10278 // Blends are faster and handle all the non-lane-crossing cases.
10279 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10283 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10284 VT.getVectorNumElements() / 2);
10285 // Check for patterns which can be matched with a single insert of a 128-bit
10287 if (isShuffleEquivalent(Mask, 0, 1, 0, 1) ||
10288 isShuffleEquivalent(Mask, 0, 1, 4, 5)) {
10289 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10290 DAG.getIntPtrConstant(0));
10291 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10292 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
10293 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10295 if (isShuffleEquivalent(Mask, 0, 1, 6, 7)) {
10296 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10297 DAG.getIntPtrConstant(0));
10298 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
10299 DAG.getIntPtrConstant(2));
10300 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10303 // Otherwise form a 128-bit permutation.
10304 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
10305 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
10306 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10307 DAG.getConstant(PermMask, MVT::i8));
10310 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10311 /// shuffling each lane.
10313 /// This will only succeed when the result of fixing the 128-bit lanes results
10314 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10315 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10316 /// the lane crosses early and then use simpler shuffles within each lane.
10318 /// FIXME: It might be worthwhile at some point to support this without
10319 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10320 /// in x86 only floating point has interesting non-repeating shuffles, and even
10321 /// those are still *marginally* more expensive.
10322 static SDValue lowerVectorShuffleByMerging128BitLanes(
10323 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10324 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10325 assert(!isSingleInputShuffleMask(Mask) &&
10326 "This is only useful with multiple inputs.");
10328 int Size = Mask.size();
10329 int LaneSize = 128 / VT.getScalarSizeInBits();
10330 int NumLanes = Size / LaneSize;
10331 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10333 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10334 // check whether the in-128-bit lane shuffles share a repeating pattern.
10335 SmallVector<int, 4> Lanes;
10336 Lanes.resize(NumLanes, -1);
10337 SmallVector<int, 4> InLaneMask;
10338 InLaneMask.resize(LaneSize, -1);
10339 for (int i = 0; i < Size; ++i) {
10343 int j = i / LaneSize;
10345 if (Lanes[j] < 0) {
10346 // First entry we've seen for this lane.
10347 Lanes[j] = Mask[i] / LaneSize;
10348 } else if (Lanes[j] != Mask[i] / LaneSize) {
10349 // This doesn't match the lane selected previously!
10353 // Check that within each lane we have a consistent shuffle mask.
10354 int k = i % LaneSize;
10355 if (InLaneMask[k] < 0) {
10356 InLaneMask[k] = Mask[i] % LaneSize;
10357 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10358 // This doesn't fit a repeating in-lane mask.
10363 // First shuffle the lanes into place.
10364 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10365 VT.getSizeInBits() / 64);
10366 SmallVector<int, 8> LaneMask;
10367 LaneMask.resize(NumLanes * 2, -1);
10368 for (int i = 0; i < NumLanes; ++i)
10369 if (Lanes[i] >= 0) {
10370 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10371 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10374 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
10375 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
10376 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10378 // Cast it back to the type we actually want.
10379 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
10381 // Now do a simple shuffle that isn't lane crossing.
10382 SmallVector<int, 8> NewMask;
10383 NewMask.resize(Size, -1);
10384 for (int i = 0; i < Size; ++i)
10386 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10387 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10388 "Must not introduce lane crosses at this point!");
10390 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10393 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10396 /// This returns true if the elements from a particular input are already in the
10397 /// slot required by the given mask and require no permutation.
10398 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10399 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10400 int Size = Mask.size();
10401 for (int i = 0; i < Size; ++i)
10402 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10408 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10410 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10411 /// isn't available.
10412 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10413 const X86Subtarget *Subtarget,
10414 SelectionDAG &DAG) {
10416 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10417 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10418 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10419 ArrayRef<int> Mask = SVOp->getMask();
10420 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10422 SmallVector<int, 4> WidenedMask;
10423 if (canWidenShuffleElements(Mask, WidenedMask))
10424 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10427 if (isSingleInputShuffleMask(Mask)) {
10428 // Check for being able to broadcast a single element.
10429 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1,
10430 Mask, Subtarget, DAG))
10433 // Use low duplicate instructions for masks that match their pattern.
10434 if (isShuffleEquivalent(Mask, 0, 0, 2, 2))
10435 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10437 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10438 // Non-half-crossing single input shuffles can be lowerid with an
10439 // interleaved permutation.
10440 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10441 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10442 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10443 DAG.getConstant(VPERMILPMask, MVT::i8));
10446 // With AVX2 we have direct support for this permutation.
10447 if (Subtarget->hasAVX2())
10448 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10449 getV4X86ShuffleImm8ForMask(Mask, DAG));
10451 // Otherwise, fall back.
10452 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10456 // X86 has dedicated unpack instructions that can handle specific blend
10457 // operations: UNPCKH and UNPCKL.
10458 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
10459 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10460 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
10461 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10463 // If we have a single input to the zero element, insert that into V1 if we
10464 // can do so cheaply.
10465 int NumV2Elements =
10466 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
10467 if (NumV2Elements == 1 && Mask[0] >= 4)
10468 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10469 MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
10472 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10476 // Check if the blend happens to exactly fit that of SHUFPD.
10477 if ((Mask[0] == -1 || Mask[0] < 2) &&
10478 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
10479 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
10480 (Mask[3] == -1 || Mask[3] >= 6)) {
10481 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
10482 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
10483 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
10484 DAG.getConstant(SHUFPDMask, MVT::i8));
10486 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
10487 (Mask[1] == -1 || Mask[1] < 2) &&
10488 (Mask[2] == -1 || Mask[2] >= 6) &&
10489 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
10490 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
10491 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
10492 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
10493 DAG.getConstant(SHUFPDMask, MVT::i8));
10496 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10497 // shuffle. However, if we have AVX2 and either inputs are already in place,
10498 // we will be able to shuffle even across lanes the other input in a single
10499 // instruction so skip this pattern.
10500 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10501 isShuffleMaskInputInPlace(1, Mask))))
10502 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10503 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10506 // If we have AVX2 then we always want to lower with a blend because an v4 we
10507 // can fully permute the elements.
10508 if (Subtarget->hasAVX2())
10509 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10512 // Otherwise fall back on generic lowering.
10513 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10516 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10518 /// This routine is only called when we have AVX2 and thus a reasonable
10519 /// instruction set for v4i64 shuffling..
10520 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10521 const X86Subtarget *Subtarget,
10522 SelectionDAG &DAG) {
10524 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10525 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10526 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10527 ArrayRef<int> Mask = SVOp->getMask();
10528 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10529 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10531 SmallVector<int, 4> WidenedMask;
10532 if (canWidenShuffleElements(Mask, WidenedMask))
10533 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10536 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10540 // Check for being able to broadcast a single element.
10541 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1,
10542 Mask, Subtarget, DAG))
10545 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10546 // use lower latency instructions that will operate on both 128-bit lanes.
10547 SmallVector<int, 2> RepeatedMask;
10548 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10549 if (isSingleInputShuffleMask(Mask)) {
10550 int PSHUFDMask[] = {-1, -1, -1, -1};
10551 for (int i = 0; i < 2; ++i)
10552 if (RepeatedMask[i] >= 0) {
10553 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10554 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10556 return DAG.getNode(
10557 ISD::BITCAST, DL, MVT::v4i64,
10558 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10559 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
10560 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
10563 // Use dedicated unpack instructions for masks that match their pattern.
10564 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
10565 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10566 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
10567 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10570 // AVX2 provides a direct instruction for permuting a single input across
10572 if (isSingleInputShuffleMask(Mask))
10573 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10574 getV4X86ShuffleImm8ForMask(Mask, DAG));
10576 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10577 // shuffle. However, if we have AVX2 and either inputs are already in place,
10578 // we will be able to shuffle even across lanes the other input in a single
10579 // instruction so skip this pattern.
10580 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10581 isShuffleMaskInputInPlace(1, Mask))))
10582 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10583 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10586 // Otherwise fall back on generic blend lowering.
10587 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10591 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10593 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10594 /// isn't available.
10595 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10596 const X86Subtarget *Subtarget,
10597 SelectionDAG &DAG) {
10599 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10600 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10601 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10602 ArrayRef<int> Mask = SVOp->getMask();
10603 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10605 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10609 // Check for being able to broadcast a single element.
10610 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1,
10611 Mask, Subtarget, DAG))
10614 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10615 // options to efficiently lower the shuffle.
10616 SmallVector<int, 4> RepeatedMask;
10617 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10618 assert(RepeatedMask.size() == 4 &&
10619 "Repeated masks must be half the mask width!");
10621 // Use even/odd duplicate instructions for masks that match their pattern.
10622 if (isShuffleEquivalent(Mask, 0, 0, 2, 2, 4, 4, 6, 6))
10623 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10624 if (isShuffleEquivalent(Mask, 1, 1, 3, 3, 5, 5, 7, 7))
10625 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10627 if (isSingleInputShuffleMask(Mask))
10628 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10629 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10631 // Use dedicated unpack instructions for masks that match their pattern.
10632 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10633 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10634 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10635 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10637 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10638 // have already handled any direct blends. We also need to squash the
10639 // repeated mask into a simulated v4f32 mask.
10640 for (int i = 0; i < 4; ++i)
10641 if (RepeatedMask[i] >= 8)
10642 RepeatedMask[i] -= 4;
10643 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10646 // If we have a single input shuffle with different shuffle patterns in the
10647 // two 128-bit lanes use the variable mask to VPERMILPS.
10648 if (isSingleInputShuffleMask(Mask)) {
10649 SDValue VPermMask[8];
10650 for (int i = 0; i < 8; ++i)
10651 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10652 : DAG.getConstant(Mask[i], MVT::i32);
10653 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10654 return DAG.getNode(
10655 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10656 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10658 if (Subtarget->hasAVX2())
10659 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
10660 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
10661 DAG.getNode(ISD::BUILD_VECTOR, DL,
10662 MVT::v8i32, VPermMask)),
10665 // Otherwise, fall back.
10666 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10670 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10672 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10673 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10676 // If we have AVX2 then we always want to lower with a blend because at v8 we
10677 // can fully permute the elements.
10678 if (Subtarget->hasAVX2())
10679 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10682 // Otherwise fall back on generic lowering.
10683 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10686 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10688 /// This routine is only called when we have AVX2 and thus a reasonable
10689 /// instruction set for v8i32 shuffling..
10690 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10691 const X86Subtarget *Subtarget,
10692 SelectionDAG &DAG) {
10694 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10695 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10696 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10697 ArrayRef<int> Mask = SVOp->getMask();
10698 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10699 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10701 // Whenever we can lower this as a zext, that instruction is strictly faster
10702 // than any alternative. It also allows us to fold memory operands into the
10703 // shuffle in many cases.
10704 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10705 Mask, Subtarget, DAG))
10708 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10712 // Check for being able to broadcast a single element.
10713 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1,
10714 Mask, Subtarget, DAG))
10717 // If the shuffle mask is repeated in each 128-bit lane we can use more
10718 // efficient instructions that mirror the shuffles across the two 128-bit
10720 SmallVector<int, 4> RepeatedMask;
10721 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10722 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10723 if (isSingleInputShuffleMask(Mask))
10724 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10725 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10727 // Use dedicated unpack instructions for masks that match their pattern.
10728 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10729 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10730 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10731 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10734 // If the shuffle patterns aren't repeated but it is a single input, directly
10735 // generate a cross-lane VPERMD instruction.
10736 if (isSingleInputShuffleMask(Mask)) {
10737 SDValue VPermMask[8];
10738 for (int i = 0; i < 8; ++i)
10739 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10740 : DAG.getConstant(Mask[i], MVT::i32);
10741 return DAG.getNode(
10742 X86ISD::VPERMV, DL, MVT::v8i32,
10743 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10746 // Try to use bit shift instructions.
10747 if (SDValue Shift = lowerVectorShuffleAsBitShift(
10748 DL, MVT::v8i32, V1, V2, Mask, DAG))
10751 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10753 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10754 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10757 // Otherwise fall back on generic blend lowering.
10758 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10762 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10764 /// This routine is only called when we have AVX2 and thus a reasonable
10765 /// instruction set for v16i16 shuffling..
10766 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10767 const X86Subtarget *Subtarget,
10768 SelectionDAG &DAG) {
10770 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10771 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10772 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10773 ArrayRef<int> Mask = SVOp->getMask();
10774 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10775 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10777 // Whenever we can lower this as a zext, that instruction is strictly faster
10778 // than any alternative. It also allows us to fold memory operands into the
10779 // shuffle in many cases.
10780 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10781 Mask, Subtarget, DAG))
10784 // Check for being able to broadcast a single element.
10785 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1,
10786 Mask, Subtarget, DAG))
10789 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10793 // Use dedicated unpack instructions for masks that match their pattern.
10794 if (isShuffleEquivalent(Mask,
10795 // First 128-bit lane:
10796 0, 16, 1, 17, 2, 18, 3, 19,
10797 // Second 128-bit lane:
10798 8, 24, 9, 25, 10, 26, 11, 27))
10799 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10800 if (isShuffleEquivalent(Mask,
10801 // First 128-bit lane:
10802 4, 20, 5, 21, 6, 22, 7, 23,
10803 // Second 128-bit lane:
10804 12, 28, 13, 29, 14, 30, 15, 31))
10805 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10807 if (isSingleInputShuffleMask(Mask)) {
10808 // There are no generalized cross-lane shuffle operations available on i16
10810 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10811 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10814 SDValue PSHUFBMask[32];
10815 for (int i = 0; i < 16; ++i) {
10816 if (Mask[i] == -1) {
10817 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10821 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10822 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10823 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
10824 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
10826 return DAG.getNode(
10827 ISD::BITCAST, DL, MVT::v16i16,
10829 X86ISD::PSHUFB, DL, MVT::v32i8,
10830 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
10831 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
10834 // Try to use bit shift instructions.
10835 if (SDValue Shift = lowerVectorShuffleAsBitShift(
10836 DL, MVT::v16i16, V1, V2, Mask, DAG))
10839 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10841 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10842 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10845 // Otherwise fall back on generic lowering.
10846 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10849 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10851 /// This routine is only called when we have AVX2 and thus a reasonable
10852 /// instruction set for v32i8 shuffling..
10853 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10854 const X86Subtarget *Subtarget,
10855 SelectionDAG &DAG) {
10857 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10858 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10859 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10860 ArrayRef<int> Mask = SVOp->getMask();
10861 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10862 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10864 // Whenever we can lower this as a zext, that instruction is strictly faster
10865 // than any alternative. It also allows us to fold memory operands into the
10866 // shuffle in many cases.
10867 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10868 Mask, Subtarget, DAG))
10871 // Check for being able to broadcast a single element.
10872 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1,
10873 Mask, Subtarget, DAG))
10876 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10880 // Use dedicated unpack instructions for masks that match their pattern.
10881 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10883 if (isShuffleEquivalent(
10885 // First 128-bit lane:
10886 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10887 // Second 128-bit lane:
10888 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55))
10889 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10890 if (isShuffleEquivalent(
10892 // First 128-bit lane:
10893 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10894 // Second 128-bit lane:
10895 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63))
10896 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10898 if (isSingleInputShuffleMask(Mask)) {
10899 // There are no generalized cross-lane shuffle operations available on i8
10901 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10902 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10905 SDValue PSHUFBMask[32];
10906 for (int i = 0; i < 32; ++i)
10909 ? DAG.getUNDEF(MVT::i8)
10910 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
10912 return DAG.getNode(
10913 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10914 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10917 // Try to use bit shift instructions.
10918 if (SDValue Shift = lowerVectorShuffleAsBitShift(
10919 DL, MVT::v32i8, V1, V2, Mask, DAG))
10922 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10924 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10925 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10928 // Otherwise fall back on generic lowering.
10929 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10932 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10934 /// This routine either breaks down the specific type of a 256-bit x86 vector
10935 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10936 /// together based on the available instructions.
10937 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10938 MVT VT, const X86Subtarget *Subtarget,
10939 SelectionDAG &DAG) {
10941 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10942 ArrayRef<int> Mask = SVOp->getMask();
10944 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10945 // check for those subtargets here and avoid much of the subtarget querying in
10946 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10947 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10948 // floating point types there eventually, just immediately cast everything to
10949 // a float and operate entirely in that domain.
10950 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10951 int ElementBits = VT.getScalarSizeInBits();
10952 if (ElementBits < 32)
10953 // No floating point type available, decompose into 128-bit vectors.
10954 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10956 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10957 VT.getVectorNumElements());
10958 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
10959 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
10960 return DAG.getNode(ISD::BITCAST, DL, VT,
10961 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10964 switch (VT.SimpleTy) {
10966 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10968 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10970 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10972 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10974 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10976 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10979 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10983 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10984 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10985 const X86Subtarget *Subtarget,
10986 SelectionDAG &DAG) {
10988 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10989 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10990 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10991 ArrayRef<int> Mask = SVOp->getMask();
10992 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10994 // X86 has dedicated unpack instructions that can handle specific blend
10995 // operations: UNPCKH and UNPCKL.
10996 if (isShuffleEquivalent(Mask, 0, 8, 2, 10, 4, 12, 6, 14))
10997 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10998 if (isShuffleEquivalent(Mask, 1, 9, 3, 11, 5, 13, 7, 15))
10999 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
11001 // FIXME: Implement direct support for this type!
11002 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
11005 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
11006 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11007 const X86Subtarget *Subtarget,
11008 SelectionDAG &DAG) {
11010 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11011 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
11012 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11013 ArrayRef<int> Mask = SVOp->getMask();
11014 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11016 // Use dedicated unpack instructions for masks that match their pattern.
11017 if (isShuffleEquivalent(Mask,
11018 0, 16, 1, 17, 4, 20, 5, 21,
11019 8, 24, 9, 25, 12, 28, 13, 29))
11020 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
11021 if (isShuffleEquivalent(Mask,
11022 2, 18, 3, 19, 6, 22, 7, 23,
11023 10, 26, 11, 27, 14, 30, 15, 31))
11024 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
11026 // FIXME: Implement direct support for this type!
11027 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
11030 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
11031 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11032 const X86Subtarget *Subtarget,
11033 SelectionDAG &DAG) {
11035 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11036 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
11037 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11038 ArrayRef<int> Mask = SVOp->getMask();
11039 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
11041 // X86 has dedicated unpack instructions that can handle specific blend
11042 // operations: UNPCKH and UNPCKL.
11043 if (isShuffleEquivalent(Mask, 0, 8, 2, 10, 4, 12, 6, 14))
11044 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
11045 if (isShuffleEquivalent(Mask, 1, 9, 3, 11, 5, 13, 7, 15))
11046 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
11048 // FIXME: Implement direct support for this type!
11049 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
11052 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
11053 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11054 const X86Subtarget *Subtarget,
11055 SelectionDAG &DAG) {
11057 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11058 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
11059 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11060 ArrayRef<int> Mask = SVOp->getMask();
11061 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
11063 // Use dedicated unpack instructions for masks that match their pattern.
11064 if (isShuffleEquivalent(Mask,
11065 0, 16, 1, 17, 4, 20, 5, 21,
11066 8, 24, 9, 25, 12, 28, 13, 29))
11067 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
11068 if (isShuffleEquivalent(Mask,
11069 2, 18, 3, 19, 6, 22, 7, 23,
11070 10, 26, 11, 27, 14, 30, 15, 31))
11071 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
11073 // FIXME: Implement direct support for this type!
11074 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
11077 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
11078 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11079 const X86Subtarget *Subtarget,
11080 SelectionDAG &DAG) {
11082 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11083 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
11084 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11085 ArrayRef<int> Mask = SVOp->getMask();
11086 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
11087 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
11089 // FIXME: Implement direct support for this type!
11090 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
11093 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
11094 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11095 const X86Subtarget *Subtarget,
11096 SelectionDAG &DAG) {
11098 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11099 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
11100 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11101 ArrayRef<int> Mask = SVOp->getMask();
11102 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
11103 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
11105 // FIXME: Implement direct support for this type!
11106 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
11109 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
11111 /// This routine either breaks down the specific type of a 512-bit x86 vector
11112 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
11113 /// together based on the available instructions.
11114 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11115 MVT VT, const X86Subtarget *Subtarget,
11116 SelectionDAG &DAG) {
11118 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11119 ArrayRef<int> Mask = SVOp->getMask();
11120 assert(Subtarget->hasAVX512() &&
11121 "Cannot lower 512-bit vectors w/ basic ISA!");
11123 // Check for being able to broadcast a single element.
11124 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
11125 Mask, Subtarget, DAG))
11128 // Dispatch to each element type for lowering. If we don't have supprot for
11129 // specific element type shuffles at 512 bits, immediately split them and
11130 // lower them. Each lowering routine of a given type is allowed to assume that
11131 // the requisite ISA extensions for that element type are available.
11132 switch (VT.SimpleTy) {
11134 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11136 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11138 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11140 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11142 if (Subtarget->hasBWI())
11143 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11146 if (Subtarget->hasBWI())
11147 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11151 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11154 // Otherwise fall back on splitting.
11155 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11158 /// \brief Top-level lowering for x86 vector shuffles.
11160 /// This handles decomposition, canonicalization, and lowering of all x86
11161 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11162 /// above in helper routines. The canonicalization attempts to widen shuffles
11163 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11164 /// s.t. only one of the two inputs needs to be tested, etc.
11165 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11166 SelectionDAG &DAG) {
11167 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11168 ArrayRef<int> Mask = SVOp->getMask();
11169 SDValue V1 = Op.getOperand(0);
11170 SDValue V2 = Op.getOperand(1);
11171 MVT VT = Op.getSimpleValueType();
11172 int NumElements = VT.getVectorNumElements();
11175 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
11177 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11178 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11179 if (V1IsUndef && V2IsUndef)
11180 return DAG.getUNDEF(VT);
11182 // When we create a shuffle node we put the UNDEF node to second operand,
11183 // but in some cases the first operand may be transformed to UNDEF.
11184 // In this case we should just commute the node.
11186 return DAG.getCommutedVectorShuffle(*SVOp);
11188 // Check for non-undef masks pointing at an undef vector and make the masks
11189 // undef as well. This makes it easier to match the shuffle based solely on
11193 if (M >= NumElements) {
11194 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11195 for (int &M : NewMask)
11196 if (M >= NumElements)
11198 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11201 // Try to collapse shuffles into using a vector type with fewer elements but
11202 // wider element types. We cap this to not form integers or floating point
11203 // elements wider than 64 bits, but it might be interesting to form i128
11204 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11205 SmallVector<int, 16> WidenedMask;
11206 if (VT.getScalarSizeInBits() < 64 &&
11207 canWidenShuffleElements(Mask, WidenedMask)) {
11208 MVT NewEltVT = VT.isFloatingPoint()
11209 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11210 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11211 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11212 // Make sure that the new vector type is legal. For example, v2f64 isn't
11214 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11215 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
11216 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
11217 return DAG.getNode(ISD::BITCAST, dl, VT,
11218 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11222 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11223 for (int M : SVOp->getMask())
11225 ++NumUndefElements;
11226 else if (M < NumElements)
11231 // Commute the shuffle as needed such that more elements come from V1 than
11232 // V2. This allows us to match the shuffle pattern strictly on how many
11233 // elements come from V1 without handling the symmetric cases.
11234 if (NumV2Elements > NumV1Elements)
11235 return DAG.getCommutedVectorShuffle(*SVOp);
11237 // When the number of V1 and V2 elements are the same, try to minimize the
11238 // number of uses of V2 in the low half of the vector. When that is tied,
11239 // ensure that the sum of indices for V1 is equal to or lower than the sum
11240 // indices for V2. When those are equal, try to ensure that the number of odd
11241 // indices for V1 is lower than the number of odd indices for V2.
11242 if (NumV1Elements == NumV2Elements) {
11243 int LowV1Elements = 0, LowV2Elements = 0;
11244 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11245 if (M >= NumElements)
11249 if (LowV2Elements > LowV1Elements) {
11250 return DAG.getCommutedVectorShuffle(*SVOp);
11251 } else if (LowV2Elements == LowV1Elements) {
11252 int SumV1Indices = 0, SumV2Indices = 0;
11253 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11254 if (SVOp->getMask()[i] >= NumElements)
11256 else if (SVOp->getMask()[i] >= 0)
11258 if (SumV2Indices < SumV1Indices) {
11259 return DAG.getCommutedVectorShuffle(*SVOp);
11260 } else if (SumV2Indices == SumV1Indices) {
11261 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11262 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11263 if (SVOp->getMask()[i] >= NumElements)
11264 NumV2OddIndices += i % 2;
11265 else if (SVOp->getMask()[i] >= 0)
11266 NumV1OddIndices += i % 2;
11267 if (NumV2OddIndices < NumV1OddIndices)
11268 return DAG.getCommutedVectorShuffle(*SVOp);
11273 // For each vector width, delegate to a specialized lowering routine.
11274 if (VT.getSizeInBits() == 128)
11275 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11277 if (VT.getSizeInBits() == 256)
11278 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11280 // Force AVX-512 vectors to be scalarized for now.
11281 // FIXME: Implement AVX-512 support!
11282 if (VT.getSizeInBits() == 512)
11283 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11285 llvm_unreachable("Unimplemented!");
11289 //===----------------------------------------------------------------------===//
11290 // Legacy vector shuffle lowering
11292 // This code is the legacy code handling vector shuffles until the above
11293 // replaces its functionality and performance.
11294 //===----------------------------------------------------------------------===//
11296 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
11297 bool hasInt256, unsigned *MaskOut = nullptr) {
11298 MVT EltVT = VT.getVectorElementType();
11300 // There is no blend with immediate in AVX-512.
11301 if (VT.is512BitVector())
11304 if (!hasSSE41 || EltVT == MVT::i8)
11306 if (!hasInt256 && VT == MVT::v16i16)
11309 unsigned MaskValue = 0;
11310 unsigned NumElems = VT.getVectorNumElements();
11311 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11312 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11313 unsigned NumElemsInLane = NumElems / NumLanes;
11315 // Blend for v16i16 should be symetric for the both lanes.
11316 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11318 int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
11319 int EltIdx = MaskVals[i];
11321 if ((EltIdx < 0 || EltIdx == (int)i) &&
11322 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
11325 if (((unsigned)EltIdx == (i + NumElems)) &&
11326 (SndLaneEltIdx < 0 ||
11327 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
11328 MaskValue |= (1 << i);
11334 *MaskOut = MaskValue;
11338 // Try to lower a shuffle node into a simple blend instruction.
11339 // This function assumes isBlendMask returns true for this
11340 // SuffleVectorSDNode
11341 static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
11342 unsigned MaskValue,
11343 const X86Subtarget *Subtarget,
11344 SelectionDAG &DAG) {
11345 MVT VT = SVOp->getSimpleValueType(0);
11346 MVT EltVT = VT.getVectorElementType();
11347 assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
11348 Subtarget->hasInt256() && "Trying to lower a "
11349 "VECTOR_SHUFFLE to a Blend but "
11350 "with the wrong mask"));
11351 SDValue V1 = SVOp->getOperand(0);
11352 SDValue V2 = SVOp->getOperand(1);
11354 unsigned NumElems = VT.getVectorNumElements();
11356 // Convert i32 vectors to floating point if it is not AVX2.
11357 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
11359 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
11360 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
11362 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
11363 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
11366 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
11367 DAG.getConstant(MaskValue, MVT::i32));
11368 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
11371 /// In vector type \p VT, return true if the element at index \p InputIdx
11372 /// falls on a different 128-bit lane than \p OutputIdx.
11373 static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
11374 unsigned OutputIdx) {
11375 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
11376 return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
11379 /// Generate a PSHUFB if possible. Selects elements from \p V1 according to
11380 /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
11381 /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
11382 /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
11384 static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
11385 SelectionDAG &DAG) {
11386 MVT VT = V1.getSimpleValueType();
11387 assert(VT.is128BitVector() || VT.is256BitVector());
11389 MVT EltVT = VT.getVectorElementType();
11390 unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
11391 unsigned NumElts = VT.getVectorNumElements();
11393 SmallVector<SDValue, 32> PshufbMask;
11394 for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
11395 int InputIdx = MaskVals[OutputIdx];
11396 unsigned InputByteIdx;
11398 if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
11399 InputByteIdx = 0x80;
11401 // Cross lane is not allowed.
11402 if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
11404 InputByteIdx = InputIdx * EltSizeInBytes;
11405 // Index is an byte offset within the 128-bit lane.
11406 InputByteIdx &= 0xf;
11409 for (unsigned j = 0; j < EltSizeInBytes; ++j) {
11410 PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
11411 if (InputByteIdx != 0x80)
11416 MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
11418 V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
11419 return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
11420 DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
11423 // v8i16 shuffles - Prefer shuffles in the following order:
11424 // 1. [all] pshuflw, pshufhw, optional move
11425 // 2. [ssse3] 1 x pshufb
11426 // 3. [ssse3] 2 x pshufb + 1 x por
11427 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
11429 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
11430 SelectionDAG &DAG) {
11431 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11432 SDValue V1 = SVOp->getOperand(0);
11433 SDValue V2 = SVOp->getOperand(1);
11435 SmallVector<int, 8> MaskVals;
11437 // Determine if more than 1 of the words in each of the low and high quadwords
11438 // of the result come from the same quadword of one of the two inputs. Undef
11439 // mask values count as coming from any quadword, for better codegen.
11441 // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
11442 // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
11443 unsigned LoQuad[] = { 0, 0, 0, 0 };
11444 unsigned HiQuad[] = { 0, 0, 0, 0 };
11445 // Indices of quads used.
11446 std::bitset<4> InputQuads;
11447 for (unsigned i = 0; i < 8; ++i) {
11448 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
11449 int EltIdx = SVOp->getMaskElt(i);
11450 MaskVals.push_back(EltIdx);
11458 ++Quad[EltIdx / 4];
11459 InputQuads.set(EltIdx / 4);
11462 int BestLoQuad = -1;
11463 unsigned MaxQuad = 1;
11464 for (unsigned i = 0; i < 4; ++i) {
11465 if (LoQuad[i] > MaxQuad) {
11467 MaxQuad = LoQuad[i];
11471 int BestHiQuad = -1;
11473 for (unsigned i = 0; i < 4; ++i) {
11474 if (HiQuad[i] > MaxQuad) {
11476 MaxQuad = HiQuad[i];
11480 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
11481 // of the two input vectors, shuffle them into one input vector so only a
11482 // single pshufb instruction is necessary. If there are more than 2 input
11483 // quads, disable the next transformation since it does not help SSSE3.
11484 bool V1Used = InputQuads[0] || InputQuads[1];
11485 bool V2Used = InputQuads[2] || InputQuads[3];
11486 if (Subtarget->hasSSSE3()) {
11487 if (InputQuads.count() == 2 && V1Used && V2Used) {
11488 BestLoQuad = InputQuads[0] ? 0 : 1;
11489 BestHiQuad = InputQuads[2] ? 2 : 3;
11491 if (InputQuads.count() > 2) {
11497 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
11498 // the shuffle mask. If a quad is scored as -1, that means that it contains
11499 // words from all 4 input quadwords.
11501 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
11503 BestLoQuad < 0 ? 0 : BestLoQuad,
11504 BestHiQuad < 0 ? 1 : BestHiQuad
11506 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
11507 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
11508 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
11509 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
11511 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
11512 // source words for the shuffle, to aid later transformations.
11513 bool AllWordsInNewV = true;
11514 bool InOrder[2] = { true, true };
11515 for (unsigned i = 0; i != 8; ++i) {
11516 int idx = MaskVals[i];
11518 InOrder[i/4] = false;
11519 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
11521 AllWordsInNewV = false;
11525 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
11526 if (AllWordsInNewV) {
11527 for (int i = 0; i != 8; ++i) {
11528 int idx = MaskVals[i];
11531 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
11532 if ((idx != i) && idx < 4)
11534 if ((idx != i) && idx > 3)
11543 // If we've eliminated the use of V2, and the new mask is a pshuflw or
11544 // pshufhw, that's as cheap as it gets. Return the new shuffle.
11545 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
11546 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
11547 unsigned TargetMask = 0;
11548 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
11549 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
11550 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11551 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
11552 getShufflePSHUFLWImmediate(SVOp);
11553 V1 = NewV.getOperand(0);
11554 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
11558 // Promote splats to a larger type which usually leads to more efficient code.
11559 // FIXME: Is this true if pshufb is available?
11560 if (SVOp->isSplat())
11561 return PromoteSplat(SVOp, DAG);
11563 // If we have SSSE3, and all words of the result are from 1 input vector,
11564 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
11565 // is present, fall back to case 4.
11566 if (Subtarget->hasSSSE3()) {
11567 SmallVector<SDValue,16> pshufbMask;
11569 // If we have elements from both input vectors, set the high bit of the
11570 // shuffle mask element to zero out elements that come from V2 in the V1
11571 // mask, and elements that come from V1 in the V2 mask, so that the two
11572 // results can be OR'd together.
11573 bool TwoInputs = V1Used && V2Used;
11574 V1 = getPSHUFB(MaskVals, V1, dl, DAG);
11576 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11578 // Calculate the shuffle mask for the second input, shuffle it, and
11579 // OR it with the first shuffled input.
11580 CommuteVectorShuffleMask(MaskVals, 8);
11581 V2 = getPSHUFB(MaskVals, V2, dl, DAG);
11582 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
11583 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11586 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
11587 // and update MaskVals with new element order.
11588 std::bitset<8> InOrder;
11589 if (BestLoQuad >= 0) {
11590 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
11591 for (int i = 0; i != 4; ++i) {
11592 int idx = MaskVals[i];
11595 } else if ((idx / 4) == BestLoQuad) {
11596 MaskV[i] = idx & 3;
11600 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11603 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11604 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11605 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
11606 NewV.getOperand(0),
11607 getShufflePSHUFLWImmediate(SVOp), DAG);
11611 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
11612 // and update MaskVals with the new element order.
11613 if (BestHiQuad >= 0) {
11614 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
11615 for (unsigned i = 4; i != 8; ++i) {
11616 int idx = MaskVals[i];
11619 } else if ((idx / 4) == BestHiQuad) {
11620 MaskV[i] = (idx & 3) + 4;
11624 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11627 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11628 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11629 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
11630 NewV.getOperand(0),
11631 getShufflePSHUFHWImmediate(SVOp), DAG);
11635 // In case BestHi & BestLo were both -1, which means each quadword has a word
11636 // from each of the four input quadwords, calculate the InOrder bitvector now
11637 // before falling through to the insert/extract cleanup.
11638 if (BestLoQuad == -1 && BestHiQuad == -1) {
11640 for (int i = 0; i != 8; ++i)
11641 if (MaskVals[i] < 0 || MaskVals[i] == i)
11645 // The other elements are put in the right place using pextrw and pinsrw.
11646 for (unsigned i = 0; i != 8; ++i) {
11649 int EltIdx = MaskVals[i];
11652 SDValue ExtOp = (EltIdx < 8) ?
11653 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
11654 DAG.getIntPtrConstant(EltIdx)) :
11655 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
11656 DAG.getIntPtrConstant(EltIdx - 8));
11657 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
11658 DAG.getIntPtrConstant(i));
11663 /// \brief v16i16 shuffles
11665 /// FIXME: We only support generation of a single pshufb currently. We can
11666 /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
11667 /// well (e.g 2 x pshufb + 1 x por).
11669 LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
11670 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11671 SDValue V1 = SVOp->getOperand(0);
11672 SDValue V2 = SVOp->getOperand(1);
11675 if (V2.getOpcode() != ISD::UNDEF)
11678 SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
11679 return getPSHUFB(MaskVals, V1, dl, DAG);
11682 // v16i8 shuffles - Prefer shuffles in the following order:
11683 // 1. [ssse3] 1 x pshufb
11684 // 2. [ssse3] 2 x pshufb + 1 x por
11685 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
11686 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
11687 const X86Subtarget* Subtarget,
11688 SelectionDAG &DAG) {
11689 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11690 SDValue V1 = SVOp->getOperand(0);
11691 SDValue V2 = SVOp->getOperand(1);
11693 ArrayRef<int> MaskVals = SVOp->getMask();
11695 // Promote splats to a larger type which usually leads to more efficient code.
11696 // FIXME: Is this true if pshufb is available?
11697 if (SVOp->isSplat())
11698 return PromoteSplat(SVOp, DAG);
11700 // If we have SSSE3, case 1 is generated when all result bytes come from
11701 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
11702 // present, fall back to case 3.
11704 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
11705 if (Subtarget->hasSSSE3()) {
11706 SmallVector<SDValue,16> pshufbMask;
11708 // If all result elements are from one input vector, then only translate
11709 // undef mask values to 0x80 (zero out result) in the pshufb mask.
11711 // Otherwise, we have elements from both input vectors, and must zero out
11712 // elements that come from V2 in the first mask, and V1 in the second mask
11713 // so that we can OR them together.
11714 for (unsigned i = 0; i != 16; ++i) {
11715 int EltIdx = MaskVals[i];
11716 if (EltIdx < 0 || EltIdx >= 16)
11718 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
11720 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
11721 DAG.getNode(ISD::BUILD_VECTOR, dl,
11722 MVT::v16i8, pshufbMask));
11724 // As PSHUFB will zero elements with negative indices, it's safe to ignore
11725 // the 2nd operand if it's undefined or zero.
11726 if (V2.getOpcode() == ISD::UNDEF ||
11727 ISD::isBuildVectorAllZeros(V2.getNode()))
11730 // Calculate the shuffle mask for the second input, shuffle it, and
11731 // OR it with the first shuffled input.
11732 pshufbMask.clear();
11733 for (unsigned i = 0; i != 16; ++i) {
11734 int EltIdx = MaskVals[i];
11735 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
11736 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
11738 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
11739 DAG.getNode(ISD::BUILD_VECTOR, dl,
11740 MVT::v16i8, pshufbMask));
11741 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
11744 // No SSSE3 - Calculate in place words and then fix all out of place words
11745 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
11746 // the 16 different words that comprise the two doublequadword input vectors.
11747 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11748 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
11750 for (int i = 0; i != 8; ++i) {
11751 int Elt0 = MaskVals[i*2];
11752 int Elt1 = MaskVals[i*2+1];
11754 // This word of the result is all undef, skip it.
11755 if (Elt0 < 0 && Elt1 < 0)
11758 // This word of the result is already in the correct place, skip it.
11759 if ((Elt0 == i*2) && (Elt1 == i*2+1))
11762 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
11763 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
11766 // If Elt0 and Elt1 are defined, are consecutive, and can be load
11767 // using a single extract together, load it and store it.
11768 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
11769 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
11770 DAG.getIntPtrConstant(Elt1 / 2));
11771 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
11772 DAG.getIntPtrConstant(i));
11776 // If Elt1 is defined, extract it from the appropriate source. If the
11777 // source byte is not also odd, shift the extracted word left 8 bits
11778 // otherwise clear the bottom 8 bits if we need to do an or.
11780 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
11781 DAG.getIntPtrConstant(Elt1 / 2));
11782 if ((Elt1 & 1) == 0)
11783 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
11785 TLI.getShiftAmountTy(InsElt.getValueType())));
11786 else if (Elt0 >= 0)
11787 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
11788 DAG.getConstant(0xFF00, MVT::i16));
11790 // If Elt0 is defined, extract it from the appropriate source. If the
11791 // source byte is not also even, shift the extracted word right 8 bits. If
11792 // Elt1 was also defined, OR the extracted values together before
11793 // inserting them in the result.
11795 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
11796 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
11797 if ((Elt0 & 1) != 0)
11798 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
11800 TLI.getShiftAmountTy(InsElt0.getValueType())));
11801 else if (Elt1 >= 0)
11802 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
11803 DAG.getConstant(0x00FF, MVT::i16));
11804 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
11807 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
11808 DAG.getIntPtrConstant(i));
11810 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
11813 // v32i8 shuffles - Translate to VPSHUFB if possible.
11815 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
11816 const X86Subtarget *Subtarget,
11817 SelectionDAG &DAG) {
11818 MVT VT = SVOp->getSimpleValueType(0);
11819 SDValue V1 = SVOp->getOperand(0);
11820 SDValue V2 = SVOp->getOperand(1);
11822 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
11824 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11825 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
11826 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
11828 // VPSHUFB may be generated if
11829 // (1) one of input vector is undefined or zeroinitializer.
11830 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
11831 // And (2) the mask indexes don't cross the 128-bit lane.
11832 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
11833 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
11836 if (V1IsAllZero && !V2IsAllZero) {
11837 CommuteVectorShuffleMask(MaskVals, 32);
11840 return getPSHUFB(MaskVals, V1, dl, DAG);
11843 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
11844 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
11845 /// done when every pair / quad of shuffle mask elements point to elements in
11846 /// the right sequence. e.g.
11847 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
11849 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
11850 SelectionDAG &DAG) {
11851 MVT VT = SVOp->getSimpleValueType(0);
11853 unsigned NumElems = VT.getVectorNumElements();
11856 switch (VT.SimpleTy) {
11857 default: llvm_unreachable("Unexpected!");
11860 return SDValue(SVOp, 0);
11861 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
11862 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
11863 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
11864 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
11865 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
11866 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
11869 SmallVector<int, 8> MaskVec;
11870 for (unsigned i = 0; i != NumElems; i += Scale) {
11872 for (unsigned j = 0; j != Scale; ++j) {
11873 int EltIdx = SVOp->getMaskElt(i+j);
11877 StartIdx = (EltIdx / Scale);
11878 if (EltIdx != (int)(StartIdx*Scale + j))
11881 MaskVec.push_back(StartIdx);
11884 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
11885 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
11886 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
11889 /// getVZextMovL - Return a zero-extending vector move low node.
11891 static SDValue getVZextMovL(MVT VT, MVT OpVT,
11892 SDValue SrcOp, SelectionDAG &DAG,
11893 const X86Subtarget *Subtarget, SDLoc dl) {
11894 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
11895 LoadSDNode *LD = nullptr;
11896 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
11897 LD = dyn_cast<LoadSDNode>(SrcOp);
11899 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
11901 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
11902 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
11903 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11904 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
11905 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
11907 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
11908 return DAG.getNode(ISD::BITCAST, dl, VT,
11909 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
11910 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11912 SrcOp.getOperand(0)
11918 return DAG.getNode(ISD::BITCAST, dl, VT,
11919 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
11920 DAG.getNode(ISD::BITCAST, dl,
11924 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
11925 /// which could not be matched by any known target speficic shuffle
11927 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
11929 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
11930 if (NewOp.getNode())
11933 MVT VT = SVOp->getSimpleValueType(0);
11935 unsigned NumElems = VT.getVectorNumElements();
11936 unsigned NumLaneElems = NumElems / 2;
11939 MVT EltVT = VT.getVectorElementType();
11940 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
11943 SmallVector<int, 16> Mask;
11944 for (unsigned l = 0; l < 2; ++l) {
11945 // Build a shuffle mask for the output, discovering on the fly which
11946 // input vectors to use as shuffle operands (recorded in InputUsed).
11947 // If building a suitable shuffle vector proves too hard, then bail
11948 // out with UseBuildVector set.
11949 bool UseBuildVector = false;
11950 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
11951 unsigned LaneStart = l * NumLaneElems;
11952 for (unsigned i = 0; i != NumLaneElems; ++i) {
11953 // The mask element. This indexes into the input.
11954 int Idx = SVOp->getMaskElt(i+LaneStart);
11956 // the mask element does not index into any input vector.
11957 Mask.push_back(-1);
11961 // The input vector this mask element indexes into.
11962 int Input = Idx / NumLaneElems;
11964 // Turn the index into an offset from the start of the input vector.
11965 Idx -= Input * NumLaneElems;
11967 // Find or create a shuffle vector operand to hold this input.
11969 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
11970 if (InputUsed[OpNo] == Input)
11971 // This input vector is already an operand.
11973 if (InputUsed[OpNo] < 0) {
11974 // Create a new operand for this input vector.
11975 InputUsed[OpNo] = Input;
11980 if (OpNo >= array_lengthof(InputUsed)) {
11981 // More than two input vectors used! Give up on trying to create a
11982 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
11983 UseBuildVector = true;
11987 // Add the mask index for the new shuffle vector.
11988 Mask.push_back(Idx + OpNo * NumLaneElems);
11991 if (UseBuildVector) {
11992 SmallVector<SDValue, 16> SVOps;
11993 for (unsigned i = 0; i != NumLaneElems; ++i) {
11994 // The mask element. This indexes into the input.
11995 int Idx = SVOp->getMaskElt(i+LaneStart);
11997 SVOps.push_back(DAG.getUNDEF(EltVT));
12001 // The input vector this mask element indexes into.
12002 int Input = Idx / NumElems;
12004 // Turn the index into an offset from the start of the input vector.
12005 Idx -= Input * NumElems;
12007 // Extract the vector element by hand.
12008 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
12009 SVOp->getOperand(Input),
12010 DAG.getIntPtrConstant(Idx)));
12013 // Construct the output using a BUILD_VECTOR.
12014 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
12015 } else if (InputUsed[0] < 0) {
12016 // No input vectors were used! The result is undefined.
12017 Output[l] = DAG.getUNDEF(NVT);
12019 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
12020 (InputUsed[0] % 2) * NumLaneElems,
12022 // If only one input was used, use an undefined vector for the other.
12023 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
12024 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
12025 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
12026 // At least one input vector was used. Create a new shuffle vector.
12027 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
12033 // Concatenate the result back
12034 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
12037 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
12038 /// 4 elements, and match them with several different shuffle types.
12040 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
12041 SDValue V1 = SVOp->getOperand(0);
12042 SDValue V2 = SVOp->getOperand(1);
12044 MVT VT = SVOp->getSimpleValueType(0);
12046 assert(VT.is128BitVector() && "Unsupported vector size");
12048 std::pair<int, int> Locs[4];
12049 int Mask1[] = { -1, -1, -1, -1 };
12050 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
12052 unsigned NumHi = 0;
12053 unsigned NumLo = 0;
12054 for (unsigned i = 0; i != 4; ++i) {
12055 int Idx = PermMask[i];
12057 Locs[i] = std::make_pair(-1, -1);
12059 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
12061 Locs[i] = std::make_pair(0, NumLo);
12062 Mask1[NumLo] = Idx;
12065 Locs[i] = std::make_pair(1, NumHi);
12067 Mask1[2+NumHi] = Idx;
12073 if (NumLo <= 2 && NumHi <= 2) {
12074 // If no more than two elements come from either vector. This can be
12075 // implemented with two shuffles. First shuffle gather the elements.
12076 // The second shuffle, which takes the first shuffle as both of its
12077 // vector operands, put the elements into the right order.
12078 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12080 int Mask2[] = { -1, -1, -1, -1 };
12082 for (unsigned i = 0; i != 4; ++i)
12083 if (Locs[i].first != -1) {
12084 unsigned Idx = (i < 2) ? 0 : 4;
12085 Idx += Locs[i].first * 2 + Locs[i].second;
12089 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
12092 if (NumLo == 3 || NumHi == 3) {
12093 // Otherwise, we must have three elements from one vector, call it X, and
12094 // one element from the other, call it Y. First, use a shufps to build an
12095 // intermediate vector with the one element from Y and the element from X
12096 // that will be in the same half in the final destination (the indexes don't
12097 // matter). Then, use a shufps to build the final vector, taking the half
12098 // containing the element from Y from the intermediate, and the other half
12101 // Normalize it so the 3 elements come from V1.
12102 CommuteVectorShuffleMask(PermMask, 4);
12106 // Find the element from V2.
12108 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
12109 int Val = PermMask[HiIndex];
12116 Mask1[0] = PermMask[HiIndex];
12118 Mask1[2] = PermMask[HiIndex^1];
12120 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12122 if (HiIndex >= 2) {
12123 Mask1[0] = PermMask[0];
12124 Mask1[1] = PermMask[1];
12125 Mask1[2] = HiIndex & 1 ? 6 : 4;
12126 Mask1[3] = HiIndex & 1 ? 4 : 6;
12127 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
12130 Mask1[0] = HiIndex & 1 ? 2 : 0;
12131 Mask1[1] = HiIndex & 1 ? 0 : 2;
12132 Mask1[2] = PermMask[2];
12133 Mask1[3] = PermMask[3];
12138 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
12141 // Break it into (shuffle shuffle_hi, shuffle_lo).
12142 int LoMask[] = { -1, -1, -1, -1 };
12143 int HiMask[] = { -1, -1, -1, -1 };
12145 int *MaskPtr = LoMask;
12146 unsigned MaskIdx = 0;
12147 unsigned LoIdx = 0;
12148 unsigned HiIdx = 2;
12149 for (unsigned i = 0; i != 4; ++i) {
12156 int Idx = PermMask[i];
12158 Locs[i] = std::make_pair(-1, -1);
12159 } else if (Idx < 4) {
12160 Locs[i] = std::make_pair(MaskIdx, LoIdx);
12161 MaskPtr[LoIdx] = Idx;
12164 Locs[i] = std::make_pair(MaskIdx, HiIdx);
12165 MaskPtr[HiIdx] = Idx;
12170 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
12171 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
12172 int MaskOps[] = { -1, -1, -1, -1 };
12173 for (unsigned i = 0; i != 4; ++i)
12174 if (Locs[i].first != -1)
12175 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
12176 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
12179 static bool MayFoldVectorLoad(SDValue V) {
12180 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
12181 V = V.getOperand(0);
12183 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
12184 V = V.getOperand(0);
12185 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
12186 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
12187 // BUILD_VECTOR (load), undef
12188 V = V.getOperand(0);
12190 return MayFoldLoad(V);
12194 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
12195 MVT VT = Op.getSimpleValueType();
12197 // Canonizalize to v2f64.
12198 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
12199 return DAG.getNode(ISD::BITCAST, dl, VT,
12200 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
12205 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
12207 SDValue V1 = Op.getOperand(0);
12208 SDValue V2 = Op.getOperand(1);
12209 MVT VT = Op.getSimpleValueType();
12211 assert(VT != MVT::v2i64 && "unsupported shuffle type");
12213 if (HasSSE2 && VT == MVT::v2f64)
12214 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
12216 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
12217 return DAG.getNode(ISD::BITCAST, dl, VT,
12218 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
12219 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
12220 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
12224 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
12225 SDValue V1 = Op.getOperand(0);
12226 SDValue V2 = Op.getOperand(1);
12227 MVT VT = Op.getSimpleValueType();
12229 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
12230 "unsupported shuffle type");
12232 if (V2.getOpcode() == ISD::UNDEF)
12236 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
12240 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
12241 SDValue V1 = Op.getOperand(0);
12242 SDValue V2 = Op.getOperand(1);
12243 MVT VT = Op.getSimpleValueType();
12244 unsigned NumElems = VT.getVectorNumElements();
12246 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
12247 // operand of these instructions is only memory, so check if there's a
12248 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
12250 bool CanFoldLoad = false;
12252 // Trivial case, when V2 comes from a load.
12253 if (MayFoldVectorLoad(V2))
12254 CanFoldLoad = true;
12256 // When V1 is a load, it can be folded later into a store in isel, example:
12257 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
12259 // (MOVLPSmr addr:$src1, VR128:$src2)
12260 // So, recognize this potential and also use MOVLPS or MOVLPD
12261 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
12262 CanFoldLoad = true;
12264 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12266 if (HasSSE2 && NumElems == 2)
12267 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
12270 // If we don't care about the second element, proceed to use movss.
12271 if (SVOp->getMaskElt(1) != -1)
12272 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
12275 // movl and movlp will both match v2i64, but v2i64 is never matched by
12276 // movl earlier because we make it strict to avoid messing with the movlp load
12277 // folding logic (see the code above getMOVLP call). Match it here then,
12278 // this is horrible, but will stay like this until we move all shuffle
12279 // matching to x86 specific nodes. Note that for the 1st condition all
12280 // types are matched with movsd.
12282 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
12283 // as to remove this logic from here, as much as possible
12284 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
12285 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12286 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12289 assert(VT != MVT::v4i32 && "unsupported shuffle type");
12291 // Invert the operand order and use SHUFPS to match it.
12292 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
12293 getShuffleSHUFImmediate(SVOp), DAG);
12296 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
12297 SelectionDAG &DAG) {
12299 MVT VT = Load->getSimpleValueType(0);
12300 MVT EVT = VT.getVectorElementType();
12301 SDValue Addr = Load->getOperand(1);
12302 SDValue NewAddr = DAG.getNode(
12303 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
12304 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
12307 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
12308 DAG.getMachineFunction().getMachineMemOperand(
12309 Load->getMemOperand(), 0, EVT.getStoreSize()));
12313 // It is only safe to call this function if isINSERTPSMask is true for
12314 // this shufflevector mask.
12315 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
12316 SelectionDAG &DAG) {
12317 // Generate an insertps instruction when inserting an f32 from memory onto a
12318 // v4f32 or when copying a member from one v4f32 to another.
12319 // We also use it for transferring i32 from one register to another,
12320 // since it simply copies the same bits.
12321 // If we're transferring an i32 from memory to a specific element in a
12322 // register, we output a generic DAG that will match the PINSRD
12324 MVT VT = SVOp->getSimpleValueType(0);
12325 MVT EVT = VT.getVectorElementType();
12326 SDValue V1 = SVOp->getOperand(0);
12327 SDValue V2 = SVOp->getOperand(1);
12328 auto Mask = SVOp->getMask();
12329 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
12330 "unsupported vector type for insertps/pinsrd");
12332 auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
12333 auto FromV2Predicate = [](const int &i) { return i >= 4; };
12334 int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
12338 unsigned DestIndex;
12342 DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
12345 // If we have 1 element from each vector, we have to check if we're
12346 // changing V1's element's place. If so, we're done. Otherwise, we
12347 // should assume we're changing V2's element's place and behave
12349 int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
12350 assert(DestIndex <= INT32_MAX && "truncated destination index");
12351 if (FromV1 == FromV2 &&
12352 static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
12356 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12359 assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
12360 "More than one element from V1 and from V2, or no elements from one "
12361 "of the vectors. This case should not have returned true from "
12366 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12369 // Get an index into the source vector in the range [0,4) (the mask is
12370 // in the range [0,8) because it can address V1 and V2)
12371 unsigned SrcIndex = Mask[DestIndex] % 4;
12372 if (MayFoldLoad(From)) {
12373 // Trivial case, when From comes from a load and is only used by the
12374 // shuffle. Make it use insertps from the vector that we need from that
12377 NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
12378 if (!NewLoad.getNode())
12381 if (EVT == MVT::f32) {
12382 // Create this as a scalar to vector to match the instruction pattern.
12383 SDValue LoadScalarToVector =
12384 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
12385 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
12386 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
12388 } else { // EVT == MVT::i32
12389 // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
12390 // instruction, to match the PINSRD instruction, which loads an i32 to a
12391 // certain vector element.
12392 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
12393 DAG.getConstant(DestIndex, MVT::i32));
12397 // Vector-element-to-vector
12398 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
12399 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
12402 // Reduce a vector shuffle to zext.
12403 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
12404 SelectionDAG &DAG) {
12405 // PMOVZX is only available from SSE41.
12406 if (!Subtarget->hasSSE41())
12409 MVT VT = Op.getSimpleValueType();
12411 // Only AVX2 support 256-bit vector integer extending.
12412 if (!Subtarget->hasInt256() && VT.is256BitVector())
12415 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12417 SDValue V1 = Op.getOperand(0);
12418 SDValue V2 = Op.getOperand(1);
12419 unsigned NumElems = VT.getVectorNumElements();
12421 // Extending is an unary operation and the element type of the source vector
12422 // won't be equal to or larger than i64.
12423 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
12424 VT.getVectorElementType() == MVT::i64)
12427 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
12428 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
12429 while ((1U << Shift) < NumElems) {
12430 if (SVOp->getMaskElt(1U << Shift) == 1)
12433 // The maximal ratio is 8, i.e. from i8 to i64.
12438 // Check the shuffle mask.
12439 unsigned Mask = (1U << Shift) - 1;
12440 for (unsigned i = 0; i != NumElems; ++i) {
12441 int EltIdx = SVOp->getMaskElt(i);
12442 if ((i & Mask) != 0 && EltIdx != -1)
12444 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
12448 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
12449 MVT NeVT = MVT::getIntegerVT(NBits);
12450 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
12452 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
12455 return DAG.getNode(ISD::BITCAST, DL, VT,
12456 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
12459 static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
12460 SelectionDAG &DAG) {
12461 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12462 MVT VT = Op.getSimpleValueType();
12464 SDValue V1 = Op.getOperand(0);
12465 SDValue V2 = Op.getOperand(1);
12467 if (isZeroShuffle(SVOp))
12468 return getZeroVector(VT, Subtarget, DAG, dl);
12470 // Handle splat operations
12471 if (SVOp->isSplat()) {
12472 // Use vbroadcast whenever the splat comes from a foldable load
12473 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
12474 if (Broadcast.getNode())
12478 // Check integer expanding shuffles.
12479 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
12480 if (NewOp.getNode())
12483 // If the shuffle can be profitably rewritten as a narrower shuffle, then
12485 if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
12486 VT == MVT::v32i8) {
12487 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12488 if (NewOp.getNode())
12489 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
12490 } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
12491 // FIXME: Figure out a cleaner way to do this.
12492 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
12493 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12494 if (NewOp.getNode()) {
12495 MVT NewVT = NewOp.getSimpleValueType();
12496 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
12497 NewVT, true, false))
12498 return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
12501 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
12502 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12503 if (NewOp.getNode()) {
12504 MVT NewVT = NewOp.getSimpleValueType();
12505 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
12506 return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
12515 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
12516 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12517 SDValue V1 = Op.getOperand(0);
12518 SDValue V2 = Op.getOperand(1);
12519 MVT VT = Op.getSimpleValueType();
12521 unsigned NumElems = VT.getVectorNumElements();
12522 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
12523 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
12524 bool V1IsSplat = false;
12525 bool V2IsSplat = false;
12526 bool HasSSE2 = Subtarget->hasSSE2();
12527 bool HasFp256 = Subtarget->hasFp256();
12528 bool HasInt256 = Subtarget->hasInt256();
12529 MachineFunction &MF = DAG.getMachineFunction();
12530 bool OptForSize = MF.getFunction()->getAttributes().
12531 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
12533 // Check if we should use the experimental vector shuffle lowering. If so,
12534 // delegate completely to that code path.
12535 if (ExperimentalVectorShuffleLowering)
12536 return lowerVectorShuffle(Op, Subtarget, DAG);
12538 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
12540 if (V1IsUndef && V2IsUndef)
12541 return DAG.getUNDEF(VT);
12543 // When we create a shuffle node we put the UNDEF node to second operand,
12544 // but in some cases the first operand may be transformed to UNDEF.
12545 // In this case we should just commute the node.
12547 return DAG.getCommutedVectorShuffle(*SVOp);
12549 // Vector shuffle lowering takes 3 steps:
12551 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
12552 // narrowing and commutation of operands should be handled.
12553 // 2) Matching of shuffles with known shuffle masks to x86 target specific
12555 // 3) Rewriting of unmatched masks into new generic shuffle operations,
12556 // so the shuffle can be broken into other shuffles and the legalizer can
12557 // try the lowering again.
12559 // The general idea is that no vector_shuffle operation should be left to
12560 // be matched during isel, all of them must be converted to a target specific
12563 // Normalize the input vectors. Here splats, zeroed vectors, profitable
12564 // narrowing and commutation of operands should be handled. The actual code
12565 // doesn't include all of those, work in progress...
12566 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
12567 if (NewOp.getNode())
12570 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
12572 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
12573 // unpckh_undef). Only use pshufd if speed is more important than size.
12574 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
12575 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12576 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
12577 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12579 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
12580 V2IsUndef && MayFoldVectorLoad(V1))
12581 return getMOVDDup(Op, dl, V1, DAG);
12583 if (isMOVHLPS_v_undef_Mask(M, VT))
12584 return getMOVHighToLow(Op, dl, DAG);
12586 // Use to match splats
12587 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
12588 (VT == MVT::v2f64 || VT == MVT::v2i64))
12589 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12591 if (isPSHUFDMask(M, VT)) {
12592 // The actual implementation will match the mask in the if above and then
12593 // during isel it can match several different instructions, not only pshufd
12594 // as its name says, sad but true, emulate the behavior for now...
12595 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
12596 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
12598 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
12600 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
12601 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
12603 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
12604 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask,
12607 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
12611 if (isPALIGNRMask(M, VT, Subtarget))
12612 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
12613 getShufflePALIGNRImmediate(SVOp),
12616 if (isVALIGNMask(M, VT, Subtarget))
12617 return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2,
12618 getShuffleVALIGNImmediate(SVOp),
12621 // Check if this can be converted into a logical shift.
12622 bool isLeft = false;
12623 unsigned ShAmt = 0;
12625 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
12626 if (isShift && ShVal.hasOneUse()) {
12627 // If the shifted value has multiple uses, it may be cheaper to use
12628 // v_set0 + movlhps or movhlps, etc.
12629 MVT EltVT = VT.getVectorElementType();
12630 ShAmt *= EltVT.getSizeInBits();
12631 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12634 if (isMOVLMask(M, VT)) {
12635 if (ISD::isBuildVectorAllZeros(V1.getNode()))
12636 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
12637 if (!isMOVLPMask(M, VT)) {
12638 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
12639 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12641 if (VT == MVT::v4i32 || VT == MVT::v4f32)
12642 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12646 // FIXME: fold these into legal mask.
12647 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
12648 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
12650 if (isMOVHLPSMask(M, VT))
12651 return getMOVHighToLow(Op, dl, DAG);
12653 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
12654 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
12656 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
12657 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
12659 if (isMOVLPMask(M, VT))
12660 return getMOVLP(Op, dl, DAG, HasSSE2);
12662 if (ShouldXformToMOVHLPS(M, VT) ||
12663 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
12664 return DAG.getCommutedVectorShuffle(*SVOp);
12667 // No better options. Use a vshldq / vsrldq.
12668 MVT EltVT = VT.getVectorElementType();
12669 ShAmt *= EltVT.getSizeInBits();
12670 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12673 bool Commuted = false;
12674 // FIXME: This should also accept a bitcast of a splat? Be careful, not
12675 // 1,1,1,1 -> v8i16 though.
12676 BitVector UndefElements;
12677 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
12678 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12680 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
12681 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12684 // Canonicalize the splat or undef, if present, to be on the RHS.
12685 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
12686 CommuteVectorShuffleMask(M, NumElems);
12688 std::swap(V1IsSplat, V2IsSplat);
12692 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
12693 // Shuffling low element of v1 into undef, just return v1.
12696 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
12697 // the instruction selector will not match, so get a canonical MOVL with
12698 // swapped operands to undo the commute.
12699 return getMOVL(DAG, dl, VT, V2, V1);
12702 if (isUNPCKLMask(M, VT, HasInt256))
12703 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12705 if (isUNPCKHMask(M, VT, HasInt256))
12706 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12709 // Normalize mask so all entries that point to V2 points to its first
12710 // element then try to match unpck{h|l} again. If match, return a
12711 // new vector_shuffle with the corrected mask.p
12712 SmallVector<int, 8> NewMask(M.begin(), M.end());
12713 NormalizeMask(NewMask, NumElems);
12714 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
12715 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12716 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
12717 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12721 // Commute is back and try unpck* again.
12722 // FIXME: this seems wrong.
12723 CommuteVectorShuffleMask(M, NumElems);
12725 std::swap(V1IsSplat, V2IsSplat);
12727 if (isUNPCKLMask(M, VT, HasInt256))
12728 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12730 if (isUNPCKHMask(M, VT, HasInt256))
12731 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12734 // Normalize the node to match x86 shuffle ops if needed
12735 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
12736 return DAG.getCommutedVectorShuffle(*SVOp);
12738 // The checks below are all present in isShuffleMaskLegal, but they are
12739 // inlined here right now to enable us to directly emit target specific
12740 // nodes, and remove one by one until they don't return Op anymore.
12742 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
12743 SVOp->getSplatIndex() == 0 && V2IsUndef) {
12744 if (VT == MVT::v2f64 || VT == MVT::v2i64)
12745 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12748 if (isPSHUFHWMask(M, VT, HasInt256))
12749 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
12750 getShufflePSHUFHWImmediate(SVOp),
12753 if (isPSHUFLWMask(M, VT, HasInt256))
12754 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
12755 getShufflePSHUFLWImmediate(SVOp),
12758 unsigned MaskValue;
12759 if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(),
12761 return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
12763 if (isSHUFPMask(M, VT))
12764 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
12765 getShuffleSHUFImmediate(SVOp), DAG);
12767 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
12768 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12769 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
12770 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12772 //===--------------------------------------------------------------------===//
12773 // Generate target specific nodes for 128 or 256-bit shuffles only
12774 // supported in the AVX instruction set.
12777 // Handle VMOVDDUPY permutations
12778 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
12779 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
12781 // Handle VPERMILPS/D* permutations
12782 if (isVPERMILPMask(M, VT)) {
12783 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
12784 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
12785 getShuffleSHUFImmediate(SVOp), DAG);
12786 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1,
12787 getShuffleSHUFImmediate(SVOp), DAG);
12791 if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
12792 return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
12793 Idx*(NumElems/2), DAG, dl);
12795 // Handle VPERM2F128/VPERM2I128 permutations
12796 if (isVPERM2X128Mask(M, VT, HasFp256))
12797 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
12798 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
12800 if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
12801 return getINSERTPS(SVOp, dl, DAG);
12804 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
12805 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
12807 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
12808 VT.is512BitVector()) {
12809 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
12810 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
12811 SmallVector<SDValue, 16> permclMask;
12812 for (unsigned i = 0; i != NumElems; ++i) {
12813 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
12816 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
12818 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
12819 return DAG.getNode(X86ISD::VPERMV, dl, VT,
12820 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
12821 return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
12822 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
12825 //===--------------------------------------------------------------------===//
12826 // Since no target specific shuffle was selected for this generic one,
12827 // lower it into other known shuffles. FIXME: this isn't true yet, but
12828 // this is the plan.
12831 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
12832 if (VT == MVT::v8i16) {
12833 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
12834 if (NewOp.getNode())
12838 if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
12839 SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
12840 if (NewOp.getNode())
12844 if (VT == MVT::v16i8) {
12845 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
12846 if (NewOp.getNode())
12850 if (VT == MVT::v32i8) {
12851 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
12852 if (NewOp.getNode())
12856 // Handle all 128-bit wide vectors with 4 elements, and match them with
12857 // several different shuffle types.
12858 if (NumElems == 4 && VT.is128BitVector())
12859 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
12861 // Handle general 256-bit shuffles
12862 if (VT.is256BitVector())
12863 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
12868 // This function assumes its argument is a BUILD_VECTOR of constants or
12869 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
12871 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
12872 unsigned &MaskValue) {
12874 unsigned NumElems = BuildVector->getNumOperands();
12875 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
12876 unsigned NumLanes = (NumElems - 1) / 8 + 1;
12877 unsigned NumElemsInLane = NumElems / NumLanes;
12879 // Blend for v16i16 should be symetric for the both lanes.
12880 for (unsigned i = 0; i < NumElemsInLane; ++i) {
12881 SDValue EltCond = BuildVector->getOperand(i);
12882 SDValue SndLaneEltCond =
12883 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
12885 int Lane1Cond = -1, Lane2Cond = -1;
12886 if (isa<ConstantSDNode>(EltCond))
12887 Lane1Cond = !isZero(EltCond);
12888 if (isa<ConstantSDNode>(SndLaneEltCond))
12889 Lane2Cond = !isZero(SndLaneEltCond);
12891 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
12892 // Lane1Cond != 0, means we want the first argument.
12893 // Lane1Cond == 0, means we want the second argument.
12894 // The encoding of this argument is 0 for the first argument, 1
12895 // for the second. Therefore, invert the condition.
12896 MaskValue |= !Lane1Cond << i;
12897 else if (Lane1Cond < 0)
12898 MaskValue |= !Lane2Cond << i;
12905 /// \brief Try to lower a VSELECT instruction to an immediate-controlled blend
12907 static SDValue lowerVSELECTtoBLENDI(SDValue Op, const X86Subtarget *Subtarget,
12908 SelectionDAG &DAG) {
12909 SDValue Cond = Op.getOperand(0);
12910 SDValue LHS = Op.getOperand(1);
12911 SDValue RHS = Op.getOperand(2);
12913 MVT VT = Op.getSimpleValueType();
12914 MVT EltVT = VT.getVectorElementType();
12915 unsigned NumElems = VT.getVectorNumElements();
12917 // There is no blend with immediate in AVX-512.
12918 if (VT.is512BitVector())
12921 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
12923 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
12926 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
12929 // Check the mask for BLEND and build the value.
12930 unsigned MaskValue = 0;
12931 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
12934 // Convert i32 vectors to floating point if it is not AVX2.
12935 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
12937 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
12938 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
12940 LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS);
12941 RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS);
12944 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS,
12945 DAG.getConstant(MaskValue, MVT::i32));
12946 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
12949 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
12950 // A vselect where all conditions and data are constants can be optimized into
12951 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
12952 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
12953 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
12954 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
12957 SDValue BlendOp = lowerVSELECTtoBLENDI(Op, Subtarget, DAG);
12958 if (BlendOp.getNode())
12961 // Some types for vselect were previously set to Expand, not Legal or
12962 // Custom. Return an empty SDValue so we fall-through to Expand, after
12963 // the Custom lowering phase.
12964 MVT VT = Op.getSimpleValueType();
12965 switch (VT.SimpleTy) {
12970 if (Subtarget->hasBWI() && Subtarget->hasVLX())
12975 // We couldn't create a "Blend with immediate" node.
12976 // This node should still be legal, but we'll have to emit a blendv*
12981 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
12982 MVT VT = Op.getSimpleValueType();
12985 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
12988 if (VT.getSizeInBits() == 8) {
12989 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
12990 Op.getOperand(0), Op.getOperand(1));
12991 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
12992 DAG.getValueType(VT));
12993 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
12996 if (VT.getSizeInBits() == 16) {
12997 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12998 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
13000 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
13001 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13002 DAG.getNode(ISD::BITCAST, dl,
13005 Op.getOperand(1)));
13006 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
13007 Op.getOperand(0), Op.getOperand(1));
13008 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
13009 DAG.getValueType(VT));
13010 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
13013 if (VT == MVT::f32) {
13014 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
13015 // the result back to FR32 register. It's only worth matching if the
13016 // result has a single use which is a store or a bitcast to i32. And in
13017 // the case of a store, it's not worth it if the index is a constant 0,
13018 // because a MOVSSmr can be used instead, which is smaller and faster.
13019 if (!Op.hasOneUse())
13021 SDNode *User = *Op.getNode()->use_begin();
13022 if ((User->getOpcode() != ISD::STORE ||
13023 (isa<ConstantSDNode>(Op.getOperand(1)) &&
13024 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
13025 (User->getOpcode() != ISD::BITCAST ||
13026 User->getValueType(0) != MVT::i32))
13028 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13029 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
13032 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
13035 if (VT == MVT::i32 || VT == MVT::i64) {
13036 // ExtractPS/pextrq works with constant index.
13037 if (isa<ConstantSDNode>(Op.getOperand(1)))
13043 /// Extract one bit from mask vector, like v16i1 or v8i1.
13044 /// AVX-512 feature.
13046 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
13047 SDValue Vec = Op.getOperand(0);
13049 MVT VecVT = Vec.getSimpleValueType();
13050 SDValue Idx = Op.getOperand(1);
13051 MVT EltVT = Op.getSimpleValueType();
13053 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
13054 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
13055 "Unexpected vector type in ExtractBitFromMaskVector");
13057 // variable index can't be handled in mask registers,
13058 // extend vector to VR512
13059 if (!isa<ConstantSDNode>(Idx)) {
13060 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
13061 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
13062 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
13063 ExtVT.getVectorElementType(), Ext, Idx);
13064 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
13067 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13068 const TargetRegisterClass* rc = getRegClassFor(VecVT);
13069 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
13070 rc = getRegClassFor(MVT::v16i1);
13071 unsigned MaxSift = rc->getSize()*8 - 1;
13072 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
13073 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
13074 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
13075 DAG.getConstant(MaxSift, MVT::i8));
13076 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
13077 DAG.getIntPtrConstant(0));
13081 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
13082 SelectionDAG &DAG) const {
13084 SDValue Vec = Op.getOperand(0);
13085 MVT VecVT = Vec.getSimpleValueType();
13086 SDValue Idx = Op.getOperand(1);
13088 if (Op.getSimpleValueType() == MVT::i1)
13089 return ExtractBitFromMaskVector(Op, DAG);
13091 if (!isa<ConstantSDNode>(Idx)) {
13092 if (VecVT.is512BitVector() ||
13093 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
13094 VecVT.getVectorElementType().getSizeInBits() == 32)) {
13097 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
13098 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
13099 MaskEltVT.getSizeInBits());
13101 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
13102 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
13103 getZeroVector(MaskVT, Subtarget, DAG, dl),
13104 Idx, DAG.getConstant(0, getPointerTy()));
13105 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
13106 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
13107 Perm, DAG.getConstant(0, getPointerTy()));
13112 // If this is a 256-bit vector result, first extract the 128-bit vector and
13113 // then extract the element from the 128-bit vector.
13114 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
13116 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13117 // Get the 128-bit vector.
13118 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
13119 MVT EltVT = VecVT.getVectorElementType();
13121 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
13123 //if (IdxVal >= NumElems/2)
13124 // IdxVal -= NumElems/2;
13125 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
13126 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
13127 DAG.getConstant(IdxVal, MVT::i32));
13130 assert(VecVT.is128BitVector() && "Unexpected vector length");
13132 if (Subtarget->hasSSE41()) {
13133 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
13138 MVT VT = Op.getSimpleValueType();
13139 // TODO: handle v16i8.
13140 if (VT.getSizeInBits() == 16) {
13141 SDValue Vec = Op.getOperand(0);
13142 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13144 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
13145 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
13146 DAG.getNode(ISD::BITCAST, dl,
13148 Op.getOperand(1)));
13149 // Transform it so it match pextrw which produces a 32-bit result.
13150 MVT EltVT = MVT::i32;
13151 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
13152 Op.getOperand(0), Op.getOperand(1));
13153 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
13154 DAG.getValueType(VT));
13155 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
13158 if (VT.getSizeInBits() == 32) {
13159 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13163 // SHUFPS the element to the lowest double word, then movss.
13164 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
13165 MVT VVT = Op.getOperand(0).getSimpleValueType();
13166 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
13167 DAG.getUNDEF(VVT), Mask);
13168 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
13169 DAG.getIntPtrConstant(0));
13172 if (VT.getSizeInBits() == 64) {
13173 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
13174 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
13175 // to match extract_elt for f64.
13176 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
13180 // UNPCKHPD the element to the lowest double word, then movsd.
13181 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
13182 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
13183 int Mask[2] = { 1, -1 };
13184 MVT VVT = Op.getOperand(0).getSimpleValueType();
13185 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
13186 DAG.getUNDEF(VVT), Mask);
13187 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
13188 DAG.getIntPtrConstant(0));
13194 /// Insert one bit to mask vector, like v16i1 or v8i1.
13195 /// AVX-512 feature.
13197 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
13199 SDValue Vec = Op.getOperand(0);
13200 SDValue Elt = Op.getOperand(1);
13201 SDValue Idx = Op.getOperand(2);
13202 MVT VecVT = Vec.getSimpleValueType();
13204 if (!isa<ConstantSDNode>(Idx)) {
13205 // Non constant index. Extend source and destination,
13206 // insert element and then truncate the result.
13207 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
13208 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
13209 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
13210 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
13211 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
13212 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
13215 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13216 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
13217 if (Vec.getOpcode() == ISD::UNDEF)
13218 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
13219 DAG.getConstant(IdxVal, MVT::i8));
13220 const TargetRegisterClass* rc = getRegClassFor(VecVT);
13221 unsigned MaxSift = rc->getSize()*8 - 1;
13222 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
13223 DAG.getConstant(MaxSift, MVT::i8));
13224 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
13225 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
13226 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
13229 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
13230 SelectionDAG &DAG) const {
13231 MVT VT = Op.getSimpleValueType();
13232 MVT EltVT = VT.getVectorElementType();
13234 if (EltVT == MVT::i1)
13235 return InsertBitToMaskVector(Op, DAG);
13238 SDValue N0 = Op.getOperand(0);
13239 SDValue N1 = Op.getOperand(1);
13240 SDValue N2 = Op.getOperand(2);
13241 if (!isa<ConstantSDNode>(N2))
13243 auto *N2C = cast<ConstantSDNode>(N2);
13244 unsigned IdxVal = N2C->getZExtValue();
13246 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
13247 // into that, and then insert the subvector back into the result.
13248 if (VT.is256BitVector() || VT.is512BitVector()) {
13249 // Get the desired 128-bit vector half.
13250 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
13252 // Insert the element into the desired half.
13253 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
13254 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
13256 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
13257 DAG.getConstant(IdxIn128, MVT::i32));
13259 // Insert the changed part back to the 256-bit vector
13260 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
13262 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
13264 if (Subtarget->hasSSE41()) {
13265 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
13267 if (VT == MVT::v8i16) {
13268 Opc = X86ISD::PINSRW;
13270 assert(VT == MVT::v16i8);
13271 Opc = X86ISD::PINSRB;
13274 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
13276 if (N1.getValueType() != MVT::i32)
13277 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13278 if (N2.getValueType() != MVT::i32)
13279 N2 = DAG.getIntPtrConstant(IdxVal);
13280 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
13283 if (EltVT == MVT::f32) {
13284 // Bits [7:6] of the constant are the source select. This will always be
13285 // zero here. The DAG Combiner may combine an extract_elt index into
13287 // bits. For example (insert (extract, 3), 2) could be matched by
13289 // the '3' into bits [7:6] of X86ISD::INSERTPS.
13290 // Bits [5:4] of the constant are the destination select. This is the
13291 // value of the incoming immediate.
13292 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
13293 // combine either bitwise AND or insert of float 0.0 to set these bits.
13294 N2 = DAG.getIntPtrConstant(IdxVal << 4);
13295 // Create this as a scalar to vector..
13296 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
13297 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
13300 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
13301 // PINSR* works with constant index.
13306 if (EltVT == MVT::i8)
13309 if (EltVT.getSizeInBits() == 16) {
13310 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
13311 // as its second argument.
13312 if (N1.getValueType() != MVT::i32)
13313 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13314 if (N2.getValueType() != MVT::i32)
13315 N2 = DAG.getIntPtrConstant(IdxVal);
13316 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
13321 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
13323 MVT OpVT = Op.getSimpleValueType();
13325 // If this is a 256-bit vector result, first insert into a 128-bit
13326 // vector and then insert into the 256-bit vector.
13327 if (!OpVT.is128BitVector()) {
13328 // Insert into a 128-bit vector.
13329 unsigned SizeFactor = OpVT.getSizeInBits()/128;
13330 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
13331 OpVT.getVectorNumElements() / SizeFactor);
13333 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
13335 // Insert the 128-bit vector.
13336 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
13339 if (OpVT == MVT::v1i64 &&
13340 Op.getOperand(0).getValueType() == MVT::i64)
13341 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
13343 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
13344 assert(OpVT.is128BitVector() && "Expected an SSE type!");
13345 return DAG.getNode(ISD::BITCAST, dl, OpVT,
13346 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
13349 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
13350 // a simple subregister reference or explicit instructions to grab
13351 // upper bits of a vector.
13352 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13353 SelectionDAG &DAG) {
13355 SDValue In = Op.getOperand(0);
13356 SDValue Idx = Op.getOperand(1);
13357 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13358 MVT ResVT = Op.getSimpleValueType();
13359 MVT InVT = In.getSimpleValueType();
13361 if (Subtarget->hasFp256()) {
13362 if (ResVT.is128BitVector() &&
13363 (InVT.is256BitVector() || InVT.is512BitVector()) &&
13364 isa<ConstantSDNode>(Idx)) {
13365 return Extract128BitVector(In, IdxVal, DAG, dl);
13367 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
13368 isa<ConstantSDNode>(Idx)) {
13369 return Extract256BitVector(In, IdxVal, DAG, dl);
13375 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
13376 // simple superregister reference or explicit instructions to insert
13377 // the upper bits of a vector.
13378 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13379 SelectionDAG &DAG) {
13380 if (!Subtarget->hasAVX())
13384 SDValue Vec = Op.getOperand(0);
13385 SDValue SubVec = Op.getOperand(1);
13386 SDValue Idx = Op.getOperand(2);
13388 if (!isa<ConstantSDNode>(Idx))
13391 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13392 MVT OpVT = Op.getSimpleValueType();
13393 MVT SubVecVT = SubVec.getSimpleValueType();
13395 // Fold two 16-byte subvector loads into one 32-byte load:
13396 // (insert_subvector (insert_subvector undef, (load addr), 0),
13397 // (load addr + 16), Elts/2)
13399 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
13400 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
13401 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
13402 !Subtarget->isUnalignedMem32Slow()) {
13403 SDValue SubVec2 = Vec.getOperand(1);
13404 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
13405 if (Idx2->getZExtValue() == 0) {
13406 SDValue Ops[] = { SubVec2, SubVec };
13407 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
13414 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
13415 SubVecVT.is128BitVector())
13416 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
13418 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
13419 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
13424 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
13425 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
13426 // one of the above mentioned nodes. It has to be wrapped because otherwise
13427 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
13428 // be used to form addressing mode. These wrapped nodes will be selected
13431 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
13432 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
13434 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13435 // global base reg.
13436 unsigned char OpFlag = 0;
13437 unsigned WrapperKind = X86ISD::Wrapper;
13438 CodeModel::Model M = DAG.getTarget().getCodeModel();
13440 if (Subtarget->isPICStyleRIPRel() &&
13441 (M == CodeModel::Small || M == CodeModel::Kernel))
13442 WrapperKind = X86ISD::WrapperRIP;
13443 else if (Subtarget->isPICStyleGOT())
13444 OpFlag = X86II::MO_GOTOFF;
13445 else if (Subtarget->isPICStyleStubPIC())
13446 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13448 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
13449 CP->getAlignment(),
13450 CP->getOffset(), OpFlag);
13452 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13453 // With PIC, the address is actually $g + Offset.
13455 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13456 DAG.getNode(X86ISD::GlobalBaseReg,
13457 SDLoc(), getPointerTy()),
13464 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
13465 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
13467 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13468 // global base reg.
13469 unsigned char OpFlag = 0;
13470 unsigned WrapperKind = X86ISD::Wrapper;
13471 CodeModel::Model M = DAG.getTarget().getCodeModel();
13473 if (Subtarget->isPICStyleRIPRel() &&
13474 (M == CodeModel::Small || M == CodeModel::Kernel))
13475 WrapperKind = X86ISD::WrapperRIP;
13476 else if (Subtarget->isPICStyleGOT())
13477 OpFlag = X86II::MO_GOTOFF;
13478 else if (Subtarget->isPICStyleStubPIC())
13479 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13481 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
13484 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13486 // With PIC, the address is actually $g + Offset.
13488 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13489 DAG.getNode(X86ISD::GlobalBaseReg,
13490 SDLoc(), getPointerTy()),
13497 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
13498 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
13500 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13501 // global base reg.
13502 unsigned char OpFlag = 0;
13503 unsigned WrapperKind = X86ISD::Wrapper;
13504 CodeModel::Model M = DAG.getTarget().getCodeModel();
13506 if (Subtarget->isPICStyleRIPRel() &&
13507 (M == CodeModel::Small || M == CodeModel::Kernel)) {
13508 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
13509 OpFlag = X86II::MO_GOTPCREL;
13510 WrapperKind = X86ISD::WrapperRIP;
13511 } else if (Subtarget->isPICStyleGOT()) {
13512 OpFlag = X86II::MO_GOT;
13513 } else if (Subtarget->isPICStyleStubPIC()) {
13514 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
13515 } else if (Subtarget->isPICStyleStubNoDynamic()) {
13516 OpFlag = X86II::MO_DARWIN_NONLAZY;
13519 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
13522 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13524 // With PIC, the address is actually $g + Offset.
13525 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
13526 !Subtarget->is64Bit()) {
13527 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13528 DAG.getNode(X86ISD::GlobalBaseReg,
13529 SDLoc(), getPointerTy()),
13533 // For symbols that require a load from a stub to get the address, emit the
13535 if (isGlobalStubReference(OpFlag))
13536 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
13537 MachinePointerInfo::getGOT(), false, false, false, 0);
13543 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
13544 // Create the TargetBlockAddressAddress node.
13545 unsigned char OpFlags =
13546 Subtarget->ClassifyBlockAddressReference();
13547 CodeModel::Model M = DAG.getTarget().getCodeModel();
13548 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
13549 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
13551 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
13554 if (Subtarget->isPICStyleRIPRel() &&
13555 (M == CodeModel::Small || M == CodeModel::Kernel))
13556 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13558 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13560 // With PIC, the address is actually $g + Offset.
13561 if (isGlobalRelativeToPICBase(OpFlags)) {
13562 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13563 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13571 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
13572 int64_t Offset, SelectionDAG &DAG) const {
13573 // Create the TargetGlobalAddress node, folding in the constant
13574 // offset if it is legal.
13575 unsigned char OpFlags =
13576 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
13577 CodeModel::Model M = DAG.getTarget().getCodeModel();
13579 if (OpFlags == X86II::MO_NO_FLAG &&
13580 X86::isOffsetSuitableForCodeModel(Offset, M)) {
13581 // A direct static reference to a global.
13582 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
13585 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
13588 if (Subtarget->isPICStyleRIPRel() &&
13589 (M == CodeModel::Small || M == CodeModel::Kernel))
13590 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13592 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13594 // With PIC, the address is actually $g + Offset.
13595 if (isGlobalRelativeToPICBase(OpFlags)) {
13596 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13597 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13601 // For globals that require a load from a stub to get the address, emit the
13603 if (isGlobalStubReference(OpFlags))
13604 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
13605 MachinePointerInfo::getGOT(), false, false, false, 0);
13607 // If there was a non-zero offset that we didn't fold, create an explicit
13608 // addition for it.
13610 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
13611 DAG.getConstant(Offset, getPointerTy()));
13617 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
13618 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
13619 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
13620 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
13624 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
13625 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
13626 unsigned char OperandFlags, bool LocalDynamic = false) {
13627 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13628 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13630 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13631 GA->getValueType(0),
13635 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
13639 SDValue Ops[] = { Chain, TGA, *InFlag };
13640 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13642 SDValue Ops[] = { Chain, TGA };
13643 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13646 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
13647 MFI->setAdjustsStack(true);
13648 MFI->setHasCalls(true);
13650 SDValue Flag = Chain.getValue(1);
13651 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
13654 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
13656 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13659 SDLoc dl(GA); // ? function entry point might be better
13660 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13661 DAG.getNode(X86ISD::GlobalBaseReg,
13662 SDLoc(), PtrVT), InFlag);
13663 InFlag = Chain.getValue(1);
13665 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
13668 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
13670 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13672 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
13673 X86::RAX, X86II::MO_TLSGD);
13676 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
13682 // Get the start address of the TLS block for this module.
13683 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
13684 .getInfo<X86MachineFunctionInfo>();
13685 MFI->incNumLocalDynamicTLSAccesses();
13689 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
13690 X86II::MO_TLSLD, /*LocalDynamic=*/true);
13693 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13694 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
13695 InFlag = Chain.getValue(1);
13696 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
13697 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
13700 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
13704 unsigned char OperandFlags = X86II::MO_DTPOFF;
13705 unsigned WrapperKind = X86ISD::Wrapper;
13706 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13707 GA->getValueType(0),
13708 GA->getOffset(), OperandFlags);
13709 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
13711 // Add x@dtpoff with the base.
13712 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
13715 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
13716 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13717 const EVT PtrVT, TLSModel::Model model,
13718 bool is64Bit, bool isPIC) {
13721 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
13722 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
13723 is64Bit ? 257 : 256));
13725 SDValue ThreadPointer =
13726 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
13727 MachinePointerInfo(Ptr), false, false, false, 0);
13729 unsigned char OperandFlags = 0;
13730 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
13732 unsigned WrapperKind = X86ISD::Wrapper;
13733 if (model == TLSModel::LocalExec) {
13734 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
13735 } else if (model == TLSModel::InitialExec) {
13737 OperandFlags = X86II::MO_GOTTPOFF;
13738 WrapperKind = X86ISD::WrapperRIP;
13740 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
13743 llvm_unreachable("Unexpected model");
13746 // emit "addl x@ntpoff,%eax" (local exec)
13747 // or "addl x@indntpoff,%eax" (initial exec)
13748 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
13750 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
13751 GA->getOffset(), OperandFlags);
13752 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
13754 if (model == TLSModel::InitialExec) {
13755 if (isPIC && !is64Bit) {
13756 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
13757 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
13761 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
13762 MachinePointerInfo::getGOT(), false, false, false, 0);
13765 // The address of the thread local variable is the add of the thread
13766 // pointer with the offset of the variable.
13767 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
13771 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
13773 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
13774 const GlobalValue *GV = GA->getGlobal();
13776 if (Subtarget->isTargetELF()) {
13777 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
13780 case TLSModel::GeneralDynamic:
13781 if (Subtarget->is64Bit())
13782 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
13783 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
13784 case TLSModel::LocalDynamic:
13785 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
13786 Subtarget->is64Bit());
13787 case TLSModel::InitialExec:
13788 case TLSModel::LocalExec:
13789 return LowerToTLSExecModel(
13790 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
13791 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
13793 llvm_unreachable("Unknown TLS model.");
13796 if (Subtarget->isTargetDarwin()) {
13797 // Darwin only has one model of TLS. Lower to that.
13798 unsigned char OpFlag = 0;
13799 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
13800 X86ISD::WrapperRIP : X86ISD::Wrapper;
13802 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13803 // global base reg.
13804 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
13805 !Subtarget->is64Bit();
13807 OpFlag = X86II::MO_TLVP_PIC_BASE;
13809 OpFlag = X86II::MO_TLVP;
13811 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
13812 GA->getValueType(0),
13813 GA->getOffset(), OpFlag);
13814 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13816 // With PIC32, the address is actually $g + Offset.
13818 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13819 DAG.getNode(X86ISD::GlobalBaseReg,
13820 SDLoc(), getPointerTy()),
13823 // Lowering the machine isd will make sure everything is in the right
13825 SDValue Chain = DAG.getEntryNode();
13826 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13827 SDValue Args[] = { Chain, Offset };
13828 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
13830 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
13831 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13832 MFI->setAdjustsStack(true);
13834 // And our return value (tls address) is in the standard call return value
13836 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
13837 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
13838 Chain.getValue(1));
13841 if (Subtarget->isTargetKnownWindowsMSVC() ||
13842 Subtarget->isTargetWindowsGNU()) {
13843 // Just use the implicit TLS architecture
13844 // Need to generate someting similar to:
13845 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
13847 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
13848 // mov rcx, qword [rdx+rcx*8]
13849 // mov eax, .tls$:tlsvar
13850 // [rax+rcx] contains the address
13851 // Windows 64bit: gs:0x58
13852 // Windows 32bit: fs:__tls_array
13855 SDValue Chain = DAG.getEntryNode();
13857 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
13858 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
13859 // use its literal value of 0x2C.
13860 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
13861 ? Type::getInt8PtrTy(*DAG.getContext(),
13863 : Type::getInt32PtrTy(*DAG.getContext(),
13867 Subtarget->is64Bit()
13868 ? DAG.getIntPtrConstant(0x58)
13869 : (Subtarget->isTargetWindowsGNU()
13870 ? DAG.getIntPtrConstant(0x2C)
13871 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
13873 SDValue ThreadPointer =
13874 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
13875 MachinePointerInfo(Ptr), false, false, false, 0);
13877 // Load the _tls_index variable
13878 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
13879 if (Subtarget->is64Bit())
13880 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
13881 IDX, MachinePointerInfo(), MVT::i32,
13882 false, false, false, 0);
13884 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
13885 false, false, false, 0);
13887 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
13889 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
13891 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
13892 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
13893 false, false, false, 0);
13895 // Get the offset of start of .tls section
13896 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13897 GA->getValueType(0),
13898 GA->getOffset(), X86II::MO_SECREL);
13899 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
13901 // The address of the thread local variable is the add of the thread
13902 // pointer with the offset of the variable.
13903 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
13906 llvm_unreachable("TLS not implemented for this target.");
13909 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
13910 /// and take a 2 x i32 value to shift plus a shift amount.
13911 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
13912 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
13913 MVT VT = Op.getSimpleValueType();
13914 unsigned VTBits = VT.getSizeInBits();
13916 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
13917 SDValue ShOpLo = Op.getOperand(0);
13918 SDValue ShOpHi = Op.getOperand(1);
13919 SDValue ShAmt = Op.getOperand(2);
13920 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
13921 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
13923 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
13924 DAG.getConstant(VTBits - 1, MVT::i8));
13925 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
13926 DAG.getConstant(VTBits - 1, MVT::i8))
13927 : DAG.getConstant(0, VT);
13929 SDValue Tmp2, Tmp3;
13930 if (Op.getOpcode() == ISD::SHL_PARTS) {
13931 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
13932 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
13934 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
13935 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
13938 // If the shift amount is larger or equal than the width of a part we can't
13939 // rely on the results of shld/shrd. Insert a test and select the appropriate
13940 // values for large shift amounts.
13941 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
13942 DAG.getConstant(VTBits, MVT::i8));
13943 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13944 AndNode, DAG.getConstant(0, MVT::i8));
13947 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13948 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
13949 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
13951 if (Op.getOpcode() == ISD::SHL_PARTS) {
13952 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
13953 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
13955 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
13956 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
13959 SDValue Ops[2] = { Lo, Hi };
13960 return DAG.getMergeValues(Ops, dl);
13963 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
13964 SelectionDAG &DAG) const {
13965 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
13968 if (SrcVT.isVector()) {
13969 if (SrcVT.getVectorElementType() == MVT::i1) {
13970 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
13971 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
13972 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
13973 Op.getOperand(0)));
13978 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
13979 "Unknown SINT_TO_FP to lower!");
13981 // These are really Legal; return the operand so the caller accepts it as
13983 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
13985 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
13986 Subtarget->is64Bit()) {
13990 unsigned Size = SrcVT.getSizeInBits()/8;
13991 MachineFunction &MF = DAG.getMachineFunction();
13992 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
13993 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
13994 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13996 MachinePointerInfo::getFixedStack(SSFI),
13998 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
14001 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
14003 SelectionDAG &DAG) const {
14007 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
14009 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
14011 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
14013 unsigned ByteSize = SrcVT.getSizeInBits()/8;
14015 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
14016 MachineMemOperand *MMO;
14018 int SSFI = FI->getIndex();
14020 DAG.getMachineFunction()
14021 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14022 MachineMemOperand::MOLoad, ByteSize, ByteSize);
14024 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
14025 StackSlot = StackSlot.getOperand(1);
14027 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
14028 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
14030 Tys, Ops, SrcVT, MMO);
14033 Chain = Result.getValue(1);
14034 SDValue InFlag = Result.getValue(2);
14036 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
14037 // shouldn't be necessary except that RFP cannot be live across
14038 // multiple blocks. When stackifier is fixed, they can be uncoupled.
14039 MachineFunction &MF = DAG.getMachineFunction();
14040 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
14041 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
14042 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14043 Tys = DAG.getVTList(MVT::Other);
14045 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
14047 MachineMemOperand *MMO =
14048 DAG.getMachineFunction()
14049 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14050 MachineMemOperand::MOStore, SSFISize, SSFISize);
14052 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
14053 Ops, Op.getValueType(), MMO);
14054 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
14055 MachinePointerInfo::getFixedStack(SSFI),
14056 false, false, false, 0);
14062 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
14063 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
14064 SelectionDAG &DAG) const {
14065 // This algorithm is not obvious. Here it is what we're trying to output:
14068 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
14069 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
14071 haddpd %xmm0, %xmm0
14073 pshufd $0x4e, %xmm0, %xmm1
14079 LLVMContext *Context = DAG.getContext();
14081 // Build some magic constants.
14082 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
14083 Constant *C0 = ConstantDataVector::get(*Context, CV0);
14084 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
14086 SmallVector<Constant*,2> CV1;
14088 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
14089 APInt(64, 0x4330000000000000ULL))));
14091 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
14092 APInt(64, 0x4530000000000000ULL))));
14093 Constant *C1 = ConstantVector::get(CV1);
14094 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
14096 // Load the 64-bit value into an XMM register.
14097 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
14099 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
14100 MachinePointerInfo::getConstantPool(),
14101 false, false, false, 16);
14102 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
14103 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
14106 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
14107 MachinePointerInfo::getConstantPool(),
14108 false, false, false, 16);
14109 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
14110 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
14113 if (Subtarget->hasSSE3()) {
14114 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
14115 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
14117 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
14118 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
14120 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
14121 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
14125 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
14126 DAG.getIntPtrConstant(0));
14129 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
14130 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
14131 SelectionDAG &DAG) const {
14133 // FP constant to bias correct the final result.
14134 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
14137 // Load the 32-bit value into an XMM register.
14138 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
14141 // Zero out the upper parts of the register.
14142 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
14144 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
14145 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
14146 DAG.getIntPtrConstant(0));
14148 // Or the load with the bias.
14149 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
14150 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
14151 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
14152 MVT::v2f64, Load)),
14153 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
14154 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
14155 MVT::v2f64, Bias)));
14156 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
14157 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
14158 DAG.getIntPtrConstant(0));
14160 // Subtract the bias.
14161 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
14163 // Handle final rounding.
14164 EVT DestVT = Op.getValueType();
14166 if (DestVT.bitsLT(MVT::f64))
14167 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
14168 DAG.getIntPtrConstant(0));
14169 if (DestVT.bitsGT(MVT::f64))
14170 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
14172 // Handle final rounding.
14176 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
14177 const X86Subtarget &Subtarget) {
14178 // The algorithm is the following:
14179 // #ifdef __SSE4_1__
14180 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
14181 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
14182 // (uint4) 0x53000000, 0xaa);
14184 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
14185 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
14187 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
14188 // return (float4) lo + fhi;
14191 SDValue V = Op->getOperand(0);
14192 EVT VecIntVT = V.getValueType();
14193 bool Is128 = VecIntVT == MVT::v4i32;
14194 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
14195 // If we convert to something else than the supported type, e.g., to v4f64,
14197 if (VecFloatVT != Op->getValueType(0))
14200 unsigned NumElts = VecIntVT.getVectorNumElements();
14201 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
14202 "Unsupported custom type");
14203 assert(NumElts <= 8 && "The size of the constant array must be fixed");
14205 // In the #idef/#else code, we have in common:
14206 // - The vector of constants:
14212 // Create the splat vector for 0x4b000000.
14213 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
14214 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
14215 CstLow, CstLow, CstLow, CstLow};
14216 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14217 makeArrayRef(&CstLowArray[0], NumElts));
14218 // Create the splat vector for 0x53000000.
14219 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
14220 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
14221 CstHigh, CstHigh, CstHigh, CstHigh};
14222 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14223 makeArrayRef(&CstHighArray[0], NumElts));
14225 // Create the right shift.
14226 SDValue CstShift = DAG.getConstant(16, MVT::i32);
14227 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
14228 CstShift, CstShift, CstShift, CstShift};
14229 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
14230 makeArrayRef(&CstShiftArray[0], NumElts));
14231 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
14234 if (Subtarget.hasSSE41()) {
14235 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
14236 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
14237 SDValue VecCstLowBitcast =
14238 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
14239 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
14240 // Low will be bitcasted right away, so do not bother bitcasting back to its
14242 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
14243 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
14244 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
14245 // (uint4) 0x53000000, 0xaa);
14246 SDValue VecCstHighBitcast =
14247 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
14248 SDValue VecShiftBitcast =
14249 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
14250 // High will be bitcasted right away, so do not bother bitcasting back to
14251 // its original type.
14252 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
14253 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
14255 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
14256 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
14257 CstMask, CstMask, CstMask);
14258 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
14259 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
14260 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
14262 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
14263 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
14266 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
14267 SDValue CstFAdd = DAG.getConstantFP(
14268 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
14269 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
14270 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
14271 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
14272 makeArrayRef(&CstFAddArray[0], NumElts));
14274 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
14275 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
14277 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
14278 // return (float4) lo + fhi;
14279 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
14280 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
14283 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
14284 SelectionDAG &DAG) const {
14285 SDValue N0 = Op.getOperand(0);
14286 MVT SVT = N0.getSimpleValueType();
14289 switch (SVT.SimpleTy) {
14291 llvm_unreachable("Custom UINT_TO_FP is not supported!");
14296 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
14297 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
14298 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
14302 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
14304 llvm_unreachable(nullptr);
14307 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
14308 SelectionDAG &DAG) const {
14309 SDValue N0 = Op.getOperand(0);
14312 if (Op.getValueType().isVector())
14313 return lowerUINT_TO_FP_vec(Op, DAG);
14315 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
14316 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
14317 // the optimization here.
14318 if (DAG.SignBitIsZero(N0))
14319 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
14321 MVT SrcVT = N0.getSimpleValueType();
14322 MVT DstVT = Op.getSimpleValueType();
14323 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
14324 return LowerUINT_TO_FP_i64(Op, DAG);
14325 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
14326 return LowerUINT_TO_FP_i32(Op, DAG);
14327 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
14330 // Make a 64-bit buffer, and use it to build an FILD.
14331 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
14332 if (SrcVT == MVT::i32) {
14333 SDValue WordOff = DAG.getConstant(4, getPointerTy());
14334 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
14335 getPointerTy(), StackSlot, WordOff);
14336 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14337 StackSlot, MachinePointerInfo(),
14339 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
14340 OffsetSlot, MachinePointerInfo(),
14342 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
14346 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
14347 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14348 StackSlot, MachinePointerInfo(),
14350 // For i64 source, we need to add the appropriate power of 2 if the input
14351 // was negative. This is the same as the optimization in
14352 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
14353 // we must be careful to do the computation in x87 extended precision, not
14354 // in SSE. (The generic code can't know it's OK to do this, or how to.)
14355 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
14356 MachineMemOperand *MMO =
14357 DAG.getMachineFunction()
14358 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14359 MachineMemOperand::MOLoad, 8, 8);
14361 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
14362 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
14363 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
14366 APInt FF(32, 0x5F800000ULL);
14368 // Check whether the sign bit is set.
14369 SDValue SignSet = DAG.getSetCC(dl,
14370 getSetCCResultType(*DAG.getContext(), MVT::i64),
14371 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
14374 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
14375 SDValue FudgePtr = DAG.getConstantPool(
14376 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
14379 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
14380 SDValue Zero = DAG.getIntPtrConstant(0);
14381 SDValue Four = DAG.getIntPtrConstant(4);
14382 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
14384 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
14386 // Load the value out, extending it from f32 to f80.
14387 // FIXME: Avoid the extend by constructing the right constant pool?
14388 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
14389 FudgePtr, MachinePointerInfo::getConstantPool(),
14390 MVT::f32, false, false, false, 4);
14391 // Extend everything to 80 bits to force it to be done on x87.
14392 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
14393 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
14396 std::pair<SDValue,SDValue>
14397 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
14398 bool IsSigned, bool IsReplace) const {
14401 EVT DstTy = Op.getValueType();
14403 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
14404 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
14408 assert(DstTy.getSimpleVT() <= MVT::i64 &&
14409 DstTy.getSimpleVT() >= MVT::i16 &&
14410 "Unknown FP_TO_INT to lower!");
14412 // These are really Legal.
14413 if (DstTy == MVT::i32 &&
14414 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14415 return std::make_pair(SDValue(), SDValue());
14416 if (Subtarget->is64Bit() &&
14417 DstTy == MVT::i64 &&
14418 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14419 return std::make_pair(SDValue(), SDValue());
14421 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
14422 // stack slot, or into the FTOL runtime function.
14423 MachineFunction &MF = DAG.getMachineFunction();
14424 unsigned MemSize = DstTy.getSizeInBits()/8;
14425 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14426 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14429 if (!IsSigned && isIntegerTypeFTOL(DstTy))
14430 Opc = X86ISD::WIN_FTOL;
14432 switch (DstTy.getSimpleVT().SimpleTy) {
14433 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
14434 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
14435 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
14436 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
14439 SDValue Chain = DAG.getEntryNode();
14440 SDValue Value = Op.getOperand(0);
14441 EVT TheVT = Op.getOperand(0).getValueType();
14442 // FIXME This causes a redundant load/store if the SSE-class value is already
14443 // in memory, such as if it is on the callstack.
14444 if (isScalarFPTypeInSSEReg(TheVT)) {
14445 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
14446 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
14447 MachinePointerInfo::getFixedStack(SSFI),
14449 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
14451 Chain, StackSlot, DAG.getValueType(TheVT)
14454 MachineMemOperand *MMO =
14455 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14456 MachineMemOperand::MOLoad, MemSize, MemSize);
14457 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
14458 Chain = Value.getValue(1);
14459 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14460 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14463 MachineMemOperand *MMO =
14464 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14465 MachineMemOperand::MOStore, MemSize, MemSize);
14467 if (Opc != X86ISD::WIN_FTOL) {
14468 // Build the FP_TO_INT*_IN_MEM
14469 SDValue Ops[] = { Chain, Value, StackSlot };
14470 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
14472 return std::make_pair(FIST, StackSlot);
14474 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
14475 DAG.getVTList(MVT::Other, MVT::Glue),
14477 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
14478 MVT::i32, ftol.getValue(1));
14479 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
14480 MVT::i32, eax.getValue(2));
14481 SDValue Ops[] = { eax, edx };
14482 SDValue pair = IsReplace
14483 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
14484 : DAG.getMergeValues(Ops, DL);
14485 return std::make_pair(pair, SDValue());
14489 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
14490 const X86Subtarget *Subtarget) {
14491 MVT VT = Op->getSimpleValueType(0);
14492 SDValue In = Op->getOperand(0);
14493 MVT InVT = In.getSimpleValueType();
14496 // Optimize vectors in AVX mode:
14499 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
14500 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
14501 // Concat upper and lower parts.
14504 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
14505 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
14506 // Concat upper and lower parts.
14509 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
14510 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
14511 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
14514 if (Subtarget->hasInt256())
14515 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
14517 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
14518 SDValue Undef = DAG.getUNDEF(InVT);
14519 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
14520 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14521 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14523 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
14524 VT.getVectorNumElements()/2);
14526 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
14527 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
14529 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14532 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
14533 SelectionDAG &DAG) {
14534 MVT VT = Op->getSimpleValueType(0);
14535 SDValue In = Op->getOperand(0);
14536 MVT InVT = In.getSimpleValueType();
14538 unsigned int NumElts = VT.getVectorNumElements();
14539 if (NumElts != 8 && NumElts != 16)
14542 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
14543 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
14545 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
14546 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14547 // Now we have only mask extension
14548 assert(InVT.getVectorElementType() == MVT::i1);
14549 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
14550 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14551 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
14552 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14553 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14554 MachinePointerInfo::getConstantPool(),
14555 false, false, false, Alignment);
14557 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
14558 if (VT.is512BitVector())
14560 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
14563 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14564 SelectionDAG &DAG) {
14565 if (Subtarget->hasFp256()) {
14566 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14574 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14575 SelectionDAG &DAG) {
14577 MVT VT = Op.getSimpleValueType();
14578 SDValue In = Op.getOperand(0);
14579 MVT SVT = In.getSimpleValueType();
14581 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
14582 return LowerZERO_EXTEND_AVX512(Op, DAG);
14584 if (Subtarget->hasFp256()) {
14585 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14590 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
14591 VT.getVectorNumElements() != SVT.getVectorNumElements());
14595 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
14597 MVT VT = Op.getSimpleValueType();
14598 SDValue In = Op.getOperand(0);
14599 MVT InVT = In.getSimpleValueType();
14601 if (VT == MVT::i1) {
14602 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
14603 "Invalid scalar TRUNCATE operation");
14604 if (InVT.getSizeInBits() >= 32)
14606 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
14607 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
14609 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
14610 "Invalid TRUNCATE operation");
14612 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
14613 if (VT.getVectorElementType().getSizeInBits() >=8)
14614 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
14616 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14617 unsigned NumElts = InVT.getVectorNumElements();
14618 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
14619 if (InVT.getSizeInBits() < 512) {
14620 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
14621 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
14625 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
14626 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14627 SDValue CP = DAG.getConstantPool(C, getPointerTy());
14628 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14629 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14630 MachinePointerInfo::getConstantPool(),
14631 false, false, false, Alignment);
14632 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
14633 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
14634 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
14637 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
14638 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
14639 if (Subtarget->hasInt256()) {
14640 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
14641 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
14642 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
14644 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
14645 DAG.getIntPtrConstant(0));
14648 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14649 DAG.getIntPtrConstant(0));
14650 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14651 DAG.getIntPtrConstant(2));
14652 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14653 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14654 static const int ShufMask[] = {0, 2, 4, 6};
14655 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
14658 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
14659 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
14660 if (Subtarget->hasInt256()) {
14661 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
14663 SmallVector<SDValue,32> pshufbMask;
14664 for (unsigned i = 0; i < 2; ++i) {
14665 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
14666 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
14667 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
14668 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
14669 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
14670 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
14671 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
14672 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
14673 for (unsigned j = 0; j < 8; ++j)
14674 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
14676 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
14677 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
14678 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
14680 static const int ShufMask[] = {0, 2, -1, -1};
14681 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
14683 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14684 DAG.getIntPtrConstant(0));
14685 return DAG.getNode(ISD::BITCAST, DL, VT, In);
14688 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14689 DAG.getIntPtrConstant(0));
14691 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14692 DAG.getIntPtrConstant(4));
14694 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
14695 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
14697 // The PSHUFB mask:
14698 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
14699 -1, -1, -1, -1, -1, -1, -1, -1};
14701 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
14702 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
14703 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
14705 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14706 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14708 // The MOVLHPS Mask:
14709 static const int ShufMask2[] = {0, 1, 4, 5};
14710 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
14711 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
14714 // Handle truncation of V256 to V128 using shuffles.
14715 if (!VT.is128BitVector() || !InVT.is256BitVector())
14718 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
14720 unsigned NumElems = VT.getVectorNumElements();
14721 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
14723 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
14724 // Prepare truncation shuffle mask
14725 for (unsigned i = 0; i != NumElems; ++i)
14726 MaskVec[i] = i * 2;
14727 SDValue V = DAG.getVectorShuffle(NVT, DL,
14728 DAG.getNode(ISD::BITCAST, DL, NVT, In),
14729 DAG.getUNDEF(NVT), &MaskVec[0]);
14730 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
14731 DAG.getIntPtrConstant(0));
14734 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
14735 SelectionDAG &DAG) const {
14736 assert(!Op.getSimpleValueType().isVector());
14738 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
14739 /*IsSigned=*/ true, /*IsReplace=*/ false);
14740 SDValue FIST = Vals.first, StackSlot = Vals.second;
14741 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
14742 if (!FIST.getNode()) return Op;
14744 if (StackSlot.getNode())
14745 // Load the result.
14746 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
14747 FIST, StackSlot, MachinePointerInfo(),
14748 false, false, false, 0);
14750 // The node is the result.
14754 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
14755 SelectionDAG &DAG) const {
14756 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
14757 /*IsSigned=*/ false, /*IsReplace=*/ false);
14758 SDValue FIST = Vals.first, StackSlot = Vals.second;
14759 assert(FIST.getNode() && "Unexpected failure");
14761 if (StackSlot.getNode())
14762 // Load the result.
14763 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
14764 FIST, StackSlot, MachinePointerInfo(),
14765 false, false, false, 0);
14767 // The node is the result.
14771 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
14773 MVT VT = Op.getSimpleValueType();
14774 SDValue In = Op.getOperand(0);
14775 MVT SVT = In.getSimpleValueType();
14777 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
14779 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
14780 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
14781 In, DAG.getUNDEF(SVT)));
14784 /// The only differences between FABS and FNEG are the mask and the logic op.
14785 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
14786 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
14787 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
14788 "Wrong opcode for lowering FABS or FNEG.");
14790 bool IsFABS = (Op.getOpcode() == ISD::FABS);
14792 // If this is a FABS and it has an FNEG user, bail out to fold the combination
14793 // into an FNABS. We'll lower the FABS after that if it is still in use.
14795 for (SDNode *User : Op->uses())
14796 if (User->getOpcode() == ISD::FNEG)
14799 SDValue Op0 = Op.getOperand(0);
14800 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
14803 MVT VT = Op.getSimpleValueType();
14804 // Assume scalar op for initialization; update for vector if needed.
14805 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
14806 // generate a 16-byte vector constant and logic op even for the scalar case.
14807 // Using a 16-byte mask allows folding the load of the mask with
14808 // the logic op, so it can save (~4 bytes) on code size.
14810 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
14811 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
14812 // decide if we should generate a 16-byte constant mask when we only need 4 or
14813 // 8 bytes for the scalar case.
14814 if (VT.isVector()) {
14815 EltVT = VT.getVectorElementType();
14816 NumElts = VT.getVectorNumElements();
14819 unsigned EltBits = EltVT.getSizeInBits();
14820 LLVMContext *Context = DAG.getContext();
14821 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
14823 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
14824 Constant *C = ConstantInt::get(*Context, MaskElt);
14825 C = ConstantVector::getSplat(NumElts, C);
14826 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14827 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
14828 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
14829 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
14830 MachinePointerInfo::getConstantPool(),
14831 false, false, false, Alignment);
14833 if (VT.isVector()) {
14834 // For a vector, cast operands to a vector type, perform the logic op,
14835 // and cast the result back to the original value type.
14836 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
14837 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
14838 SDValue Operand = IsFNABS ?
14839 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
14840 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
14841 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
14842 return DAG.getNode(ISD::BITCAST, dl, VT,
14843 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
14846 // If not vector, then scalar.
14847 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
14848 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
14849 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
14852 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
14853 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14854 LLVMContext *Context = DAG.getContext();
14855 SDValue Op0 = Op.getOperand(0);
14856 SDValue Op1 = Op.getOperand(1);
14858 MVT VT = Op.getSimpleValueType();
14859 MVT SrcVT = Op1.getSimpleValueType();
14861 // If second operand is smaller, extend it first.
14862 if (SrcVT.bitsLT(VT)) {
14863 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
14866 // And if it is bigger, shrink it first.
14867 if (SrcVT.bitsGT(VT)) {
14868 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
14872 // At this point the operands and the result should have the same
14873 // type, and that won't be f80 since that is not custom lowered.
14875 const fltSemantics &Sem =
14876 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
14877 const unsigned SizeInBits = VT.getSizeInBits();
14879 SmallVector<Constant *, 4> CV(
14880 VT == MVT::f64 ? 2 : 4,
14881 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
14883 // First, clear all bits but the sign bit from the second operand (sign).
14884 CV[0] = ConstantFP::get(*Context,
14885 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
14886 Constant *C = ConstantVector::get(CV);
14887 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
14888 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
14889 MachinePointerInfo::getConstantPool(),
14890 false, false, false, 16);
14891 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
14893 // Next, clear the sign bit from the first operand (magnitude).
14894 // If it's a constant, we can clear it here.
14895 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
14896 APFloat APF = Op0CN->getValueAPF();
14897 // If the magnitude is a positive zero, the sign bit alone is enough.
14898 if (APF.isPosZero())
14901 CV[0] = ConstantFP::get(*Context, APF);
14903 CV[0] = ConstantFP::get(
14905 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
14907 C = ConstantVector::get(CV);
14908 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
14909 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
14910 MachinePointerInfo::getConstantPool(),
14911 false, false, false, 16);
14912 // If the magnitude operand wasn't a constant, we need to AND out the sign.
14913 if (!isa<ConstantFPSDNode>(Op0))
14914 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
14916 // OR the magnitude value with the sign bit.
14917 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
14920 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
14921 SDValue N0 = Op.getOperand(0);
14923 MVT VT = Op.getSimpleValueType();
14925 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
14926 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
14927 DAG.getConstant(1, VT));
14928 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
14931 // Check whether an OR'd tree is PTEST-able.
14932 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
14933 SelectionDAG &DAG) {
14934 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
14936 if (!Subtarget->hasSSE41())
14939 if (!Op->hasOneUse())
14942 SDNode *N = Op.getNode();
14945 SmallVector<SDValue, 8> Opnds;
14946 DenseMap<SDValue, unsigned> VecInMap;
14947 SmallVector<SDValue, 8> VecIns;
14948 EVT VT = MVT::Other;
14950 // Recognize a special case where a vector is casted into wide integer to
14952 Opnds.push_back(N->getOperand(0));
14953 Opnds.push_back(N->getOperand(1));
14955 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
14956 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
14957 // BFS traverse all OR'd operands.
14958 if (I->getOpcode() == ISD::OR) {
14959 Opnds.push_back(I->getOperand(0));
14960 Opnds.push_back(I->getOperand(1));
14961 // Re-evaluate the number of nodes to be traversed.
14962 e += 2; // 2 more nodes (LHS and RHS) are pushed.
14966 // Quit if a non-EXTRACT_VECTOR_ELT
14967 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
14970 // Quit if without a constant index.
14971 SDValue Idx = I->getOperand(1);
14972 if (!isa<ConstantSDNode>(Idx))
14975 SDValue ExtractedFromVec = I->getOperand(0);
14976 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
14977 if (M == VecInMap.end()) {
14978 VT = ExtractedFromVec.getValueType();
14979 // Quit if not 128/256-bit vector.
14980 if (!VT.is128BitVector() && !VT.is256BitVector())
14982 // Quit if not the same type.
14983 if (VecInMap.begin() != VecInMap.end() &&
14984 VT != VecInMap.begin()->first.getValueType())
14986 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
14987 VecIns.push_back(ExtractedFromVec);
14989 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
14992 assert((VT.is128BitVector() || VT.is256BitVector()) &&
14993 "Not extracted from 128-/256-bit vector.");
14995 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
14997 for (DenseMap<SDValue, unsigned>::const_iterator
14998 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
14999 // Quit if not all elements are used.
15000 if (I->second != FullMask)
15004 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
15006 // Cast all vectors into TestVT for PTEST.
15007 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
15008 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
15010 // If more than one full vectors are evaluated, OR them first before PTEST.
15011 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
15012 // Each iteration will OR 2 nodes and append the result until there is only
15013 // 1 node left, i.e. the final OR'd value of all vectors.
15014 SDValue LHS = VecIns[Slot];
15015 SDValue RHS = VecIns[Slot + 1];
15016 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
15019 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
15020 VecIns.back(), VecIns.back());
15023 /// \brief return true if \c Op has a use that doesn't just read flags.
15024 static bool hasNonFlagsUse(SDValue Op) {
15025 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
15027 SDNode *User = *UI;
15028 unsigned UOpNo = UI.getOperandNo();
15029 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
15030 // Look pass truncate.
15031 UOpNo = User->use_begin().getOperandNo();
15032 User = *User->use_begin();
15035 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
15036 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
15042 /// Emit nodes that will be selected as "test Op0,Op0", or something
15044 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
15045 SelectionDAG &DAG) const {
15046 if (Op.getValueType() == MVT::i1)
15047 // KORTEST instruction should be selected
15048 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
15049 DAG.getConstant(0, Op.getValueType()));
15051 // CF and OF aren't always set the way we want. Determine which
15052 // of these we need.
15053 bool NeedCF = false;
15054 bool NeedOF = false;
15057 case X86::COND_A: case X86::COND_AE:
15058 case X86::COND_B: case X86::COND_BE:
15061 case X86::COND_G: case X86::COND_GE:
15062 case X86::COND_L: case X86::COND_LE:
15063 case X86::COND_O: case X86::COND_NO: {
15064 // Check if we really need to set the
15065 // Overflow flag. If NoSignedWrap is present
15066 // that is not actually needed.
15067 switch (Op->getOpcode()) {
15072 const BinaryWithFlagsSDNode *BinNode =
15073 cast<BinaryWithFlagsSDNode>(Op.getNode());
15074 if (BinNode->hasNoSignedWrap())
15084 // See if we can use the EFLAGS value from the operand instead of
15085 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
15086 // we prove that the arithmetic won't overflow, we can't use OF or CF.
15087 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
15088 // Emit a CMP with 0, which is the TEST pattern.
15089 //if (Op.getValueType() == MVT::i1)
15090 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
15091 // DAG.getConstant(0, MVT::i1));
15092 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
15093 DAG.getConstant(0, Op.getValueType()));
15095 unsigned Opcode = 0;
15096 unsigned NumOperands = 0;
15098 // Truncate operations may prevent the merge of the SETCC instruction
15099 // and the arithmetic instruction before it. Attempt to truncate the operands
15100 // of the arithmetic instruction and use a reduced bit-width instruction.
15101 bool NeedTruncation = false;
15102 SDValue ArithOp = Op;
15103 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
15104 SDValue Arith = Op->getOperand(0);
15105 // Both the trunc and the arithmetic op need to have one user each.
15106 if (Arith->hasOneUse())
15107 switch (Arith.getOpcode()) {
15114 NeedTruncation = true;
15120 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
15121 // which may be the result of a CAST. We use the variable 'Op', which is the
15122 // non-casted variable when we check for possible users.
15123 switch (ArithOp.getOpcode()) {
15125 // Due to an isel shortcoming, be conservative if this add is likely to be
15126 // selected as part of a load-modify-store instruction. When the root node
15127 // in a match is a store, isel doesn't know how to remap non-chain non-flag
15128 // uses of other nodes in the match, such as the ADD in this case. This
15129 // leads to the ADD being left around and reselected, with the result being
15130 // two adds in the output. Alas, even if none our users are stores, that
15131 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
15132 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
15133 // climbing the DAG back to the root, and it doesn't seem to be worth the
15135 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
15136 UE = Op.getNode()->use_end(); UI != UE; ++UI)
15137 if (UI->getOpcode() != ISD::CopyToReg &&
15138 UI->getOpcode() != ISD::SETCC &&
15139 UI->getOpcode() != ISD::STORE)
15142 if (ConstantSDNode *C =
15143 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
15144 // An add of one will be selected as an INC.
15145 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
15146 Opcode = X86ISD::INC;
15151 // An add of negative one (subtract of one) will be selected as a DEC.
15152 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
15153 Opcode = X86ISD::DEC;
15159 // Otherwise use a regular EFLAGS-setting add.
15160 Opcode = X86ISD::ADD;
15165 // If we have a constant logical shift that's only used in a comparison
15166 // against zero turn it into an equivalent AND. This allows turning it into
15167 // a TEST instruction later.
15168 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
15169 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
15170 EVT VT = Op.getValueType();
15171 unsigned BitWidth = VT.getSizeInBits();
15172 unsigned ShAmt = Op->getConstantOperandVal(1);
15173 if (ShAmt >= BitWidth) // Avoid undefined shifts.
15175 APInt Mask = ArithOp.getOpcode() == ISD::SRL
15176 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
15177 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
15178 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
15180 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
15181 DAG.getConstant(Mask, VT));
15182 DAG.ReplaceAllUsesWith(Op, New);
15188 // If the primary and result isn't used, don't bother using X86ISD::AND,
15189 // because a TEST instruction will be better.
15190 if (!hasNonFlagsUse(Op))
15196 // Due to the ISEL shortcoming noted above, be conservative if this op is
15197 // likely to be selected as part of a load-modify-store instruction.
15198 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
15199 UE = Op.getNode()->use_end(); UI != UE; ++UI)
15200 if (UI->getOpcode() == ISD::STORE)
15203 // Otherwise use a regular EFLAGS-setting instruction.
15204 switch (ArithOp.getOpcode()) {
15205 default: llvm_unreachable("unexpected operator!");
15206 case ISD::SUB: Opcode = X86ISD::SUB; break;
15207 case ISD::XOR: Opcode = X86ISD::XOR; break;
15208 case ISD::AND: Opcode = X86ISD::AND; break;
15210 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
15211 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
15212 if (EFLAGS.getNode())
15215 Opcode = X86ISD::OR;
15229 return SDValue(Op.getNode(), 1);
15235 // If we found that truncation is beneficial, perform the truncation and
15237 if (NeedTruncation) {
15238 EVT VT = Op.getValueType();
15239 SDValue WideVal = Op->getOperand(0);
15240 EVT WideVT = WideVal.getValueType();
15241 unsigned ConvertedOp = 0;
15242 // Use a target machine opcode to prevent further DAGCombine
15243 // optimizations that may separate the arithmetic operations
15244 // from the setcc node.
15245 switch (WideVal.getOpcode()) {
15247 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
15248 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
15249 case ISD::AND: ConvertedOp = X86ISD::AND; break;
15250 case ISD::OR: ConvertedOp = X86ISD::OR; break;
15251 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
15255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15256 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
15257 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
15258 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
15259 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
15265 // Emit a CMP with 0, which is the TEST pattern.
15266 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
15267 DAG.getConstant(0, Op.getValueType()));
15269 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15270 SmallVector<SDValue, 4> Ops;
15271 for (unsigned i = 0; i != NumOperands; ++i)
15272 Ops.push_back(Op.getOperand(i));
15274 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
15275 DAG.ReplaceAllUsesWith(Op, New);
15276 return SDValue(New.getNode(), 1);
15279 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
15281 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
15282 SDLoc dl, SelectionDAG &DAG) const {
15283 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
15284 if (C->getAPIntValue() == 0)
15285 return EmitTest(Op0, X86CC, dl, DAG);
15287 if (Op0.getValueType() == MVT::i1)
15288 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
15291 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
15292 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
15293 // Do the comparison at i32 if it's smaller, besides the Atom case.
15294 // This avoids subregister aliasing issues. Keep the smaller reference
15295 // if we're optimizing for size, however, as that'll allow better folding
15296 // of memory operations.
15297 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
15298 !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
15299 AttributeSet::FunctionIndex, Attribute::MinSize) &&
15300 !Subtarget->isAtom()) {
15301 unsigned ExtendOp =
15302 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
15303 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
15304 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
15306 // Use SUB instead of CMP to enable CSE between SUB and CMP.
15307 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
15308 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
15310 return SDValue(Sub.getNode(), 1);
15312 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
15315 /// Convert a comparison if required by the subtarget.
15316 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
15317 SelectionDAG &DAG) const {
15318 // If the subtarget does not support the FUCOMI instruction, floating-point
15319 // comparisons have to be converted.
15320 if (Subtarget->hasCMov() ||
15321 Cmp.getOpcode() != X86ISD::CMP ||
15322 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
15323 !Cmp.getOperand(1).getValueType().isFloatingPoint())
15326 // The instruction selector will select an FUCOM instruction instead of
15327 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
15328 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
15329 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
15331 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
15332 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
15333 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
15334 DAG.getConstant(8, MVT::i8));
15335 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
15336 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
15339 /// The minimum architected relative accuracy is 2^-12. We need one
15340 /// Newton-Raphson step to have a good float result (24 bits of precision).
15341 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
15342 DAGCombinerInfo &DCI,
15343 unsigned &RefinementSteps,
15344 bool &UseOneConstNR) const {
15345 // FIXME: We should use instruction latency models to calculate the cost of
15346 // each potential sequence, but this is very hard to do reliably because
15347 // at least Intel's Core* chips have variable timing based on the number of
15348 // significant digits in the divisor and/or sqrt operand.
15349 if (!Subtarget->useSqrtEst())
15352 EVT VT = Op.getValueType();
15354 // SSE1 has rsqrtss and rsqrtps.
15355 // TODO: Add support for AVX512 (v16f32).
15356 // It is likely not profitable to do this for f64 because a double-precision
15357 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
15358 // instructions: convert to single, rsqrtss, convert back to double, refine
15359 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
15360 // along with FMA, this could be a throughput win.
15361 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15362 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15363 RefinementSteps = 1;
15364 UseOneConstNR = false;
15365 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
15370 /// The minimum architected relative accuracy is 2^-12. We need one
15371 /// Newton-Raphson step to have a good float result (24 bits of precision).
15372 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
15373 DAGCombinerInfo &DCI,
15374 unsigned &RefinementSteps) const {
15375 // FIXME: We should use instruction latency models to calculate the cost of
15376 // each potential sequence, but this is very hard to do reliably because
15377 // at least Intel's Core* chips have variable timing based on the number of
15378 // significant digits in the divisor.
15379 if (!Subtarget->useReciprocalEst())
15382 EVT VT = Op.getValueType();
15384 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
15385 // TODO: Add support for AVX512 (v16f32).
15386 // It is likely not profitable to do this for f64 because a double-precision
15387 // reciprocal estimate with refinement on x86 prior to FMA requires
15388 // 15 instructions: convert to single, rcpss, convert back to double, refine
15389 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
15390 // along with FMA, this could be a throughput win.
15391 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15392 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15393 RefinementSteps = ReciprocalEstimateRefinementSteps;
15394 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
15399 static bool isAllOnes(SDValue V) {
15400 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
15401 return C && C->isAllOnesValue();
15404 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
15405 /// if it's possible.
15406 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
15407 SDLoc dl, SelectionDAG &DAG) const {
15408 SDValue Op0 = And.getOperand(0);
15409 SDValue Op1 = And.getOperand(1);
15410 if (Op0.getOpcode() == ISD::TRUNCATE)
15411 Op0 = Op0.getOperand(0);
15412 if (Op1.getOpcode() == ISD::TRUNCATE)
15413 Op1 = Op1.getOperand(0);
15416 if (Op1.getOpcode() == ISD::SHL)
15417 std::swap(Op0, Op1);
15418 if (Op0.getOpcode() == ISD::SHL) {
15419 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
15420 if (And00C->getZExtValue() == 1) {
15421 // If we looked past a truncate, check that it's only truncating away
15423 unsigned BitWidth = Op0.getValueSizeInBits();
15424 unsigned AndBitWidth = And.getValueSizeInBits();
15425 if (BitWidth > AndBitWidth) {
15427 DAG.computeKnownBits(Op0, Zeros, Ones);
15428 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
15432 RHS = Op0.getOperand(1);
15434 } else if (Op1.getOpcode() == ISD::Constant) {
15435 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
15436 uint64_t AndRHSVal = AndRHS->getZExtValue();
15437 SDValue AndLHS = Op0;
15439 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
15440 LHS = AndLHS.getOperand(0);
15441 RHS = AndLHS.getOperand(1);
15444 // Use BT if the immediate can't be encoded in a TEST instruction.
15445 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
15447 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
15451 if (LHS.getNode()) {
15452 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
15453 // instruction. Since the shift amount is in-range-or-undefined, we know
15454 // that doing a bittest on the i32 value is ok. We extend to i32 because
15455 // the encoding for the i16 version is larger than the i32 version.
15456 // Also promote i16 to i32 for performance / code size reason.
15457 if (LHS.getValueType() == MVT::i8 ||
15458 LHS.getValueType() == MVT::i16)
15459 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
15461 // If the operand types disagree, extend the shift amount to match. Since
15462 // BT ignores high bits (like shifts) we can use anyextend.
15463 if (LHS.getValueType() != RHS.getValueType())
15464 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
15466 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
15467 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
15468 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15469 DAG.getConstant(Cond, MVT::i8), BT);
15475 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
15477 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
15482 // SSE Condition code mapping:
15491 switch (SetCCOpcode) {
15492 default: llvm_unreachable("Unexpected SETCC condition");
15494 case ISD::SETEQ: SSECC = 0; break;
15496 case ISD::SETGT: Swap = true; // Fallthrough
15498 case ISD::SETOLT: SSECC = 1; break;
15500 case ISD::SETGE: Swap = true; // Fallthrough
15502 case ISD::SETOLE: SSECC = 2; break;
15503 case ISD::SETUO: SSECC = 3; break;
15505 case ISD::SETNE: SSECC = 4; break;
15506 case ISD::SETULE: Swap = true; // Fallthrough
15507 case ISD::SETUGE: SSECC = 5; break;
15508 case ISD::SETULT: Swap = true; // Fallthrough
15509 case ISD::SETUGT: SSECC = 6; break;
15510 case ISD::SETO: SSECC = 7; break;
15512 case ISD::SETONE: SSECC = 8; break;
15515 std::swap(Op0, Op1);
15520 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
15521 // ones, and then concatenate the result back.
15522 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
15523 MVT VT = Op.getSimpleValueType();
15525 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
15526 "Unsupported value type for operation");
15528 unsigned NumElems = VT.getVectorNumElements();
15530 SDValue CC = Op.getOperand(2);
15532 // Extract the LHS vectors
15533 SDValue LHS = Op.getOperand(0);
15534 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15535 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15537 // Extract the RHS vectors
15538 SDValue RHS = Op.getOperand(1);
15539 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15540 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15542 // Issue the operation on the smaller types and concatenate the result back
15543 MVT EltVT = VT.getVectorElementType();
15544 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15545 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15546 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
15547 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
15550 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
15551 const X86Subtarget *Subtarget) {
15552 SDValue Op0 = Op.getOperand(0);
15553 SDValue Op1 = Op.getOperand(1);
15554 SDValue CC = Op.getOperand(2);
15555 MVT VT = Op.getSimpleValueType();
15558 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
15559 Op.getValueType().getScalarType() == MVT::i1 &&
15560 "Cannot set masked compare for this operation");
15562 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15564 bool Unsigned = false;
15567 switch (SetCCOpcode) {
15568 default: llvm_unreachable("Unexpected SETCC condition");
15569 case ISD::SETNE: SSECC = 4; break;
15570 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
15571 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
15572 case ISD::SETLT: Swap = true; //fall-through
15573 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
15574 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
15575 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
15576 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
15577 case ISD::SETULE: Unsigned = true; //fall-through
15578 case ISD::SETLE: SSECC = 2; break;
15582 std::swap(Op0, Op1);
15584 return DAG.getNode(Opc, dl, VT, Op0, Op1);
15585 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
15586 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15587 DAG.getConstant(SSECC, MVT::i8));
15590 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
15591 /// operand \p Op1. If non-trivial (for example because it's not constant)
15592 /// return an empty value.
15593 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
15595 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
15599 MVT VT = Op1.getSimpleValueType();
15600 MVT EVT = VT.getVectorElementType();
15601 unsigned n = VT.getVectorNumElements();
15602 SmallVector<SDValue, 8> ULTOp1;
15604 for (unsigned i = 0; i < n; ++i) {
15605 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
15606 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
15609 // Avoid underflow.
15610 APInt Val = Elt->getAPIntValue();
15614 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
15617 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
15620 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
15621 SelectionDAG &DAG) {
15622 SDValue Op0 = Op.getOperand(0);
15623 SDValue Op1 = Op.getOperand(1);
15624 SDValue CC = Op.getOperand(2);
15625 MVT VT = Op.getSimpleValueType();
15626 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15627 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
15632 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
15633 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
15636 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
15637 unsigned Opc = X86ISD::CMPP;
15638 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
15639 assert(VT.getVectorNumElements() <= 16);
15640 Opc = X86ISD::CMPM;
15642 // In the two special cases we can't handle, emit two comparisons.
15645 unsigned CombineOpc;
15646 if (SetCCOpcode == ISD::SETUEQ) {
15647 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
15649 assert(SetCCOpcode == ISD::SETONE);
15650 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
15653 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15654 DAG.getConstant(CC0, MVT::i8));
15655 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15656 DAG.getConstant(CC1, MVT::i8));
15657 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
15659 // Handle all other FP comparisons here.
15660 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15661 DAG.getConstant(SSECC, MVT::i8));
15664 // Break 256-bit integer vector compare into smaller ones.
15665 if (VT.is256BitVector() && !Subtarget->hasInt256())
15666 return Lower256IntVSETCC(Op, DAG);
15668 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
15669 EVT OpVT = Op1.getValueType();
15670 if (Subtarget->hasAVX512()) {
15671 if (Op1.getValueType().is512BitVector() ||
15672 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
15673 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
15674 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
15676 // In AVX-512 architecture setcc returns mask with i1 elements,
15677 // But there is no compare instruction for i8 and i16 elements in KNL.
15678 // We are not talking about 512-bit operands in this case, these
15679 // types are illegal.
15681 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
15682 OpVT.getVectorElementType().getSizeInBits() >= 8))
15683 return DAG.getNode(ISD::TRUNCATE, dl, VT,
15684 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
15687 // We are handling one of the integer comparisons here. Since SSE only has
15688 // GT and EQ comparisons for integer, swapping operands and multiple
15689 // operations may be required for some comparisons.
15691 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
15692 bool Subus = false;
15694 switch (SetCCOpcode) {
15695 default: llvm_unreachable("Unexpected SETCC condition");
15696 case ISD::SETNE: Invert = true;
15697 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
15698 case ISD::SETLT: Swap = true;
15699 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
15700 case ISD::SETGE: Swap = true;
15701 case ISD::SETLE: Opc = X86ISD::PCMPGT;
15702 Invert = true; break;
15703 case ISD::SETULT: Swap = true;
15704 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
15705 FlipSigns = true; break;
15706 case ISD::SETUGE: Swap = true;
15707 case ISD::SETULE: Opc = X86ISD::PCMPGT;
15708 FlipSigns = true; Invert = true; break;
15711 // Special case: Use min/max operations for SETULE/SETUGE
15712 MVT VET = VT.getVectorElementType();
15714 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
15715 || (Subtarget->hasSSE2() && (VET == MVT::i8));
15718 switch (SetCCOpcode) {
15720 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
15721 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
15724 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
15727 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
15728 if (!MinMax && hasSubus) {
15729 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
15731 // t = psubus Op0, Op1
15732 // pcmpeq t, <0..0>
15733 switch (SetCCOpcode) {
15735 case ISD::SETULT: {
15736 // If the comparison is against a constant we can turn this into a
15737 // setule. With psubus, setule does not require a swap. This is
15738 // beneficial because the constant in the register is no longer
15739 // destructed as the destination so it can be hoisted out of a loop.
15740 // Only do this pre-AVX since vpcmp* is no longer destructive.
15741 if (Subtarget->hasAVX())
15743 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
15744 if (ULEOp1.getNode()) {
15746 Subus = true; Invert = false; Swap = false;
15750 // Psubus is better than flip-sign because it requires no inversion.
15751 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
15752 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
15756 Opc = X86ISD::SUBUS;
15762 std::swap(Op0, Op1);
15764 // Check that the operation in question is available (most are plain SSE2,
15765 // but PCMPGTQ and PCMPEQQ have different requirements).
15766 if (VT == MVT::v2i64) {
15767 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
15768 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
15770 // First cast everything to the right type.
15771 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
15772 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
15774 // Since SSE has no unsigned integer comparisons, we need to flip the sign
15775 // bits of the inputs before performing those operations. The lower
15776 // compare is always unsigned.
15779 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
15781 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
15782 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
15783 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
15784 Sign, Zero, Sign, Zero);
15786 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
15787 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
15789 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
15790 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
15791 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
15793 // Create masks for only the low parts/high parts of the 64 bit integers.
15794 static const int MaskHi[] = { 1, 1, 3, 3 };
15795 static const int MaskLo[] = { 0, 0, 2, 2 };
15796 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
15797 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
15798 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
15800 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
15801 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
15804 Result = DAG.getNOT(dl, Result, MVT::v4i32);
15806 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
15809 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
15810 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
15811 // pcmpeqd + pshufd + pand.
15812 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
15814 // First cast everything to the right type.
15815 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
15816 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
15819 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
15821 // Make sure the lower and upper halves are both all-ones.
15822 static const int Mask[] = { 1, 0, 3, 2 };
15823 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
15824 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
15827 Result = DAG.getNOT(dl, Result, MVT::v4i32);
15829 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
15833 // Since SSE has no unsigned integer comparisons, we need to flip the sign
15834 // bits of the inputs before performing those operations.
15836 EVT EltVT = VT.getVectorElementType();
15837 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
15838 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
15839 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
15842 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
15844 // If the logical-not of the result is required, perform that now.
15846 Result = DAG.getNOT(dl, Result, VT);
15849 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
15852 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
15853 getZeroVector(VT, Subtarget, DAG, dl));
15858 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
15860 MVT VT = Op.getSimpleValueType();
15862 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
15864 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
15865 && "SetCC type must be 8-bit or 1-bit integer");
15866 SDValue Op0 = Op.getOperand(0);
15867 SDValue Op1 = Op.getOperand(1);
15869 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
15871 // Optimize to BT if possible.
15872 // Lower (X & (1 << N)) == 0 to BT(X, N).
15873 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
15874 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
15875 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
15876 Op1.getOpcode() == ISD::Constant &&
15877 cast<ConstantSDNode>(Op1)->isNullValue() &&
15878 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15879 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
15880 if (NewSetCC.getNode()) {
15882 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
15887 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
15889 if (Op1.getOpcode() == ISD::Constant &&
15890 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
15891 cast<ConstantSDNode>(Op1)->isNullValue()) &&
15892 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15894 // If the input is a setcc, then reuse the input setcc or use a new one with
15895 // the inverted condition.
15896 if (Op0.getOpcode() == X86ISD::SETCC) {
15897 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
15898 bool Invert = (CC == ISD::SETNE) ^
15899 cast<ConstantSDNode>(Op1)->isNullValue();
15903 CCode = X86::GetOppositeBranchCondition(CCode);
15904 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15905 DAG.getConstant(CCode, MVT::i8),
15906 Op0.getOperand(1));
15908 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
15912 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
15913 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
15914 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15916 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
15917 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
15920 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
15921 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
15922 if (X86CC == X86::COND_INVALID)
15925 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
15926 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
15927 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15928 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
15930 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
15934 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
15935 static bool isX86LogicalCmp(SDValue Op) {
15936 unsigned Opc = Op.getNode()->getOpcode();
15937 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
15938 Opc == X86ISD::SAHF)
15940 if (Op.getResNo() == 1 &&
15941 (Opc == X86ISD::ADD ||
15942 Opc == X86ISD::SUB ||
15943 Opc == X86ISD::ADC ||
15944 Opc == X86ISD::SBB ||
15945 Opc == X86ISD::SMUL ||
15946 Opc == X86ISD::UMUL ||
15947 Opc == X86ISD::INC ||
15948 Opc == X86ISD::DEC ||
15949 Opc == X86ISD::OR ||
15950 Opc == X86ISD::XOR ||
15951 Opc == X86ISD::AND))
15954 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
15960 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
15961 if (V.getOpcode() != ISD::TRUNCATE)
15964 SDValue VOp0 = V.getOperand(0);
15965 unsigned InBits = VOp0.getValueSizeInBits();
15966 unsigned Bits = V.getValueSizeInBits();
15967 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
15970 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
15971 bool addTest = true;
15972 SDValue Cond = Op.getOperand(0);
15973 SDValue Op1 = Op.getOperand(1);
15974 SDValue Op2 = Op.getOperand(2);
15976 EVT VT = Op1.getValueType();
15979 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
15980 // are available. Otherwise fp cmovs get lowered into a less efficient branch
15981 // sequence later on.
15982 if (Cond.getOpcode() == ISD::SETCC &&
15983 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
15984 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
15985 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
15986 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
15987 int SSECC = translateX86FSETCC(
15988 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
15991 if (Subtarget->hasAVX512()) {
15992 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
15993 DAG.getConstant(SSECC, MVT::i8));
15994 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
15996 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
15997 DAG.getConstant(SSECC, MVT::i8));
15998 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
15999 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
16000 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
16004 if (Cond.getOpcode() == ISD::SETCC) {
16005 SDValue NewCond = LowerSETCC(Cond, DAG);
16006 if (NewCond.getNode())
16010 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
16011 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
16012 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
16013 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
16014 if (Cond.getOpcode() == X86ISD::SETCC &&
16015 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
16016 isZero(Cond.getOperand(1).getOperand(1))) {
16017 SDValue Cmp = Cond.getOperand(1);
16019 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
16021 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
16022 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
16023 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
16025 SDValue CmpOp0 = Cmp.getOperand(0);
16026 // Apply further optimizations for special cases
16027 // (select (x != 0), -1, 0) -> neg & sbb
16028 // (select (x == 0), 0, -1) -> neg & sbb
16029 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
16030 if (YC->isNullValue() &&
16031 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
16032 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
16033 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
16034 DAG.getConstant(0, CmpOp0.getValueType()),
16036 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16037 DAG.getConstant(X86::COND_B, MVT::i8),
16038 SDValue(Neg.getNode(), 1));
16042 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
16043 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
16044 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16046 SDValue Res = // Res = 0 or -1.
16047 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16048 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
16050 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
16051 Res = DAG.getNOT(DL, Res, Res.getValueType());
16053 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
16054 if (!N2C || !N2C->isNullValue())
16055 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
16060 // Look past (and (setcc_carry (cmp ...)), 1).
16061 if (Cond.getOpcode() == ISD::AND &&
16062 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
16063 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
16064 if (C && C->getAPIntValue() == 1)
16065 Cond = Cond.getOperand(0);
16068 // If condition flag is set by a X86ISD::CMP, then use it as the condition
16069 // setting operand in place of the X86ISD::SETCC.
16070 unsigned CondOpcode = Cond.getOpcode();
16071 if (CondOpcode == X86ISD::SETCC ||
16072 CondOpcode == X86ISD::SETCC_CARRY) {
16073 CC = Cond.getOperand(0);
16075 SDValue Cmp = Cond.getOperand(1);
16076 unsigned Opc = Cmp.getOpcode();
16077 MVT VT = Op.getSimpleValueType();
16079 bool IllegalFPCMov = false;
16080 if (VT.isFloatingPoint() && !VT.isVector() &&
16081 !isScalarFPTypeInSSEReg(VT)) // FPStack?
16082 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
16084 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
16085 Opc == X86ISD::BT) { // FIXME
16089 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
16090 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
16091 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
16092 Cond.getOperand(0).getValueType() != MVT::i8)) {
16093 SDValue LHS = Cond.getOperand(0);
16094 SDValue RHS = Cond.getOperand(1);
16095 unsigned X86Opcode;
16098 switch (CondOpcode) {
16099 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
16100 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
16101 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
16102 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
16103 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
16104 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
16105 default: llvm_unreachable("unexpected overflowing operator");
16107 if (CondOpcode == ISD::UMULO)
16108 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
16111 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
16113 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
16115 if (CondOpcode == ISD::UMULO)
16116 Cond = X86Op.getValue(2);
16118 Cond = X86Op.getValue(1);
16120 CC = DAG.getConstant(X86Cond, MVT::i8);
16125 // Look pass the truncate if the high bits are known zero.
16126 if (isTruncWithZeroHighBitsInput(Cond, DAG))
16127 Cond = Cond.getOperand(0);
16129 // We know the result of AND is compared against zero. Try to match
16131 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
16132 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
16133 if (NewSetCC.getNode()) {
16134 CC = NewSetCC.getOperand(0);
16135 Cond = NewSetCC.getOperand(1);
16142 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16143 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
16146 // a < b ? -1 : 0 -> RES = ~setcc_carry
16147 // a < b ? 0 : -1 -> RES = setcc_carry
16148 // a >= b ? -1 : 0 -> RES = setcc_carry
16149 // a >= b ? 0 : -1 -> RES = ~setcc_carry
16150 if (Cond.getOpcode() == X86ISD::SUB) {
16151 Cond = ConvertCmpIfNecessary(Cond, DAG);
16152 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
16154 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
16155 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
16156 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
16157 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
16158 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
16159 return DAG.getNOT(DL, Res, Res.getValueType());
16164 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
16165 // widen the cmov and push the truncate through. This avoids introducing a new
16166 // branch during isel and doesn't add any extensions.
16167 if (Op.getValueType() == MVT::i8 &&
16168 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
16169 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
16170 if (T1.getValueType() == T2.getValueType() &&
16171 // Blacklist CopyFromReg to avoid partial register stalls.
16172 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
16173 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
16174 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
16175 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
16179 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
16180 // condition is true.
16181 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
16182 SDValue Ops[] = { Op2, Op1, CC, Cond };
16183 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
16186 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
16187 SelectionDAG &DAG) {
16188 MVT VT = Op->getSimpleValueType(0);
16189 SDValue In = Op->getOperand(0);
16190 MVT InVT = In.getSimpleValueType();
16191 MVT VTElt = VT.getVectorElementType();
16192 MVT InVTElt = InVT.getVectorElementType();
16196 if ((InVTElt == MVT::i1) &&
16197 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
16198 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
16200 ((Subtarget->hasBWI() && VT.is512BitVector() &&
16201 VTElt.getSizeInBits() <= 16)) ||
16203 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
16204 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
16206 ((Subtarget->hasDQI() && VT.is512BitVector() &&
16207 VTElt.getSizeInBits() >= 32))))
16208 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16210 unsigned int NumElts = VT.getVectorNumElements();
16212 if (NumElts != 8 && NumElts != 16)
16215 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
16216 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
16217 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
16218 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16222 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
16224 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
16225 Constant *C = ConstantInt::get(*DAG.getContext(),
16226 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
16228 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
16229 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
16230 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
16231 MachinePointerInfo::getConstantPool(),
16232 false, false, false, Alignment);
16233 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
16234 if (VT.is512BitVector())
16236 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
16239 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
16240 SelectionDAG &DAG) {
16241 MVT VT = Op->getSimpleValueType(0);
16242 SDValue In = Op->getOperand(0);
16243 MVT InVT = In.getSimpleValueType();
16246 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
16247 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
16249 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
16250 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
16251 (VT != MVT::v16i16 || InVT != MVT::v16i8))
16254 if (Subtarget->hasInt256())
16255 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
16257 // Optimize vectors in AVX mode
16258 // Sign extend v8i16 to v8i32 and
16261 // Divide input vector into two parts
16262 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
16263 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
16264 // concat the vectors to original VT
16266 unsigned NumElems = InVT.getVectorNumElements();
16267 SDValue Undef = DAG.getUNDEF(InVT);
16269 SmallVector<int,8> ShufMask1(NumElems, -1);
16270 for (unsigned i = 0; i != NumElems/2; ++i)
16273 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
16275 SmallVector<int,8> ShufMask2(NumElems, -1);
16276 for (unsigned i = 0; i != NumElems/2; ++i)
16277 ShufMask2[i] = i + NumElems/2;
16279 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
16281 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
16282 VT.getVectorNumElements()/2);
16284 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
16285 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
16287 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
16290 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
16291 // may emit an illegal shuffle but the expansion is still better than scalar
16292 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
16293 // we'll emit a shuffle and a arithmetic shift.
16294 // TODO: It is possible to support ZExt by zeroing the undef values during
16295 // the shuffle phase or after the shuffle.
16296 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
16297 SelectionDAG &DAG) {
16298 MVT RegVT = Op.getSimpleValueType();
16299 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
16300 assert(RegVT.isInteger() &&
16301 "We only custom lower integer vector sext loads.");
16303 // Nothing useful we can do without SSE2 shuffles.
16304 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
16306 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
16308 EVT MemVT = Ld->getMemoryVT();
16309 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16310 unsigned RegSz = RegVT.getSizeInBits();
16312 ISD::LoadExtType Ext = Ld->getExtensionType();
16314 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
16315 && "Only anyext and sext are currently implemented.");
16316 assert(MemVT != RegVT && "Cannot extend to the same type");
16317 assert(MemVT.isVector() && "Must load a vector from memory");
16319 unsigned NumElems = RegVT.getVectorNumElements();
16320 unsigned MemSz = MemVT.getSizeInBits();
16321 assert(RegSz > MemSz && "Register size must be greater than the mem size");
16323 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
16324 // The only way in which we have a legal 256-bit vector result but not the
16325 // integer 256-bit operations needed to directly lower a sextload is if we
16326 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
16327 // a 128-bit vector and a normal sign_extend to 256-bits that should get
16328 // correctly legalized. We do this late to allow the canonical form of
16329 // sextload to persist throughout the rest of the DAG combiner -- it wants
16330 // to fold together any extensions it can, and so will fuse a sign_extend
16331 // of an sextload into a sextload targeting a wider value.
16333 if (MemSz == 128) {
16334 // Just switch this to a normal load.
16335 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
16336 "it must be a legal 128-bit vector "
16338 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
16339 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
16340 Ld->isInvariant(), Ld->getAlignment());
16342 assert(MemSz < 128 &&
16343 "Can't extend a type wider than 128 bits to a 256 bit vector!");
16344 // Do an sext load to a 128-bit vector type. We want to use the same
16345 // number of elements, but elements half as wide. This will end up being
16346 // recursively lowered by this routine, but will succeed as we definitely
16347 // have all the necessary features if we're using AVX1.
16349 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
16350 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
16352 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
16353 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
16354 Ld->isNonTemporal(), Ld->isInvariant(),
16355 Ld->getAlignment());
16358 // Replace chain users with the new chain.
16359 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
16360 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
16362 // Finally, do a normal sign-extend to the desired register.
16363 return DAG.getSExtOrTrunc(Load, dl, RegVT);
16366 // All sizes must be a power of two.
16367 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
16368 "Non-power-of-two elements are not custom lowered!");
16370 // Attempt to load the original value using scalar loads.
16371 // Find the largest scalar type that divides the total loaded size.
16372 MVT SclrLoadTy = MVT::i8;
16373 for (MVT Tp : MVT::integer_valuetypes()) {
16374 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
16379 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
16380 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
16382 SclrLoadTy = MVT::f64;
16384 // Calculate the number of scalar loads that we need to perform
16385 // in order to load our vector from memory.
16386 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
16388 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
16389 "Can only lower sext loads with a single scalar load!");
16391 unsigned loadRegZize = RegSz;
16392 if (Ext == ISD::SEXTLOAD && RegSz == 256)
16395 // Represent our vector as a sequence of elements which are the
16396 // largest scalar that we can load.
16397 EVT LoadUnitVecVT = EVT::getVectorVT(
16398 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
16400 // Represent the data using the same element type that is stored in
16401 // memory. In practice, we ''widen'' MemVT.
16403 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
16404 loadRegZize / MemVT.getScalarType().getSizeInBits());
16406 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
16407 "Invalid vector type");
16409 // We can't shuffle using an illegal type.
16410 assert(TLI.isTypeLegal(WideVecVT) &&
16411 "We only lower types that form legal widened vector types");
16413 SmallVector<SDValue, 8> Chains;
16414 SDValue Ptr = Ld->getBasePtr();
16415 SDValue Increment =
16416 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
16417 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
16419 for (unsigned i = 0; i < NumLoads; ++i) {
16420 // Perform a single load.
16421 SDValue ScalarLoad =
16422 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
16423 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
16424 Ld->getAlignment());
16425 Chains.push_back(ScalarLoad.getValue(1));
16426 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
16427 // another round of DAGCombining.
16429 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
16431 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
16432 ScalarLoad, DAG.getIntPtrConstant(i));
16434 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16437 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
16439 // Bitcast the loaded value to a vector of the original element type, in
16440 // the size of the target vector type.
16441 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
16442 unsigned SizeRatio = RegSz / MemSz;
16444 if (Ext == ISD::SEXTLOAD) {
16445 // If we have SSE4.1, we can directly emit a VSEXT node.
16446 if (Subtarget->hasSSE41()) {
16447 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
16448 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16452 // Otherwise we'll shuffle the small elements in the high bits of the
16453 // larger type and perform an arithmetic shift. If the shift is not legal
16454 // it's better to scalarize.
16455 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
16456 "We can't implement a sext load without an arithmetic right shift!");
16458 // Redistribute the loaded elements into the different locations.
16459 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16460 for (unsigned i = 0; i != NumElems; ++i)
16461 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
16463 SDValue Shuff = DAG.getVectorShuffle(
16464 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16466 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16468 // Build the arithmetic shift.
16469 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
16470 MemVT.getVectorElementType().getSizeInBits();
16472 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
16474 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16478 // Redistribute the loaded elements into the different locations.
16479 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16480 for (unsigned i = 0; i != NumElems; ++i)
16481 ShuffleVec[i * SizeRatio] = i;
16483 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
16484 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16486 // Bitcast to the requested type.
16487 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16488 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16492 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
16493 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
16494 // from the AND / OR.
16495 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
16496 Opc = Op.getOpcode();
16497 if (Opc != ISD::OR && Opc != ISD::AND)
16499 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16500 Op.getOperand(0).hasOneUse() &&
16501 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
16502 Op.getOperand(1).hasOneUse());
16505 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
16506 // 1 and that the SETCC node has a single use.
16507 static bool isXor1OfSetCC(SDValue Op) {
16508 if (Op.getOpcode() != ISD::XOR)
16510 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
16511 if (N1C && N1C->getAPIntValue() == 1) {
16512 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16513 Op.getOperand(0).hasOneUse();
16518 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
16519 bool addTest = true;
16520 SDValue Chain = Op.getOperand(0);
16521 SDValue Cond = Op.getOperand(1);
16522 SDValue Dest = Op.getOperand(2);
16525 bool Inverted = false;
16527 if (Cond.getOpcode() == ISD::SETCC) {
16528 // Check for setcc([su]{add,sub,mul}o == 0).
16529 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
16530 isa<ConstantSDNode>(Cond.getOperand(1)) &&
16531 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
16532 Cond.getOperand(0).getResNo() == 1 &&
16533 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
16534 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
16535 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
16536 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
16537 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
16538 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
16540 Cond = Cond.getOperand(0);
16542 SDValue NewCond = LowerSETCC(Cond, DAG);
16543 if (NewCond.getNode())
16548 // FIXME: LowerXALUO doesn't handle these!!
16549 else if (Cond.getOpcode() == X86ISD::ADD ||
16550 Cond.getOpcode() == X86ISD::SUB ||
16551 Cond.getOpcode() == X86ISD::SMUL ||
16552 Cond.getOpcode() == X86ISD::UMUL)
16553 Cond = LowerXALUO(Cond, DAG);
16556 // Look pass (and (setcc_carry (cmp ...)), 1).
16557 if (Cond.getOpcode() == ISD::AND &&
16558 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
16559 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
16560 if (C && C->getAPIntValue() == 1)
16561 Cond = Cond.getOperand(0);
16564 // If condition flag is set by a X86ISD::CMP, then use it as the condition
16565 // setting operand in place of the X86ISD::SETCC.
16566 unsigned CondOpcode = Cond.getOpcode();
16567 if (CondOpcode == X86ISD::SETCC ||
16568 CondOpcode == X86ISD::SETCC_CARRY) {
16569 CC = Cond.getOperand(0);
16571 SDValue Cmp = Cond.getOperand(1);
16572 unsigned Opc = Cmp.getOpcode();
16573 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
16574 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
16578 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
16582 // These can only come from an arithmetic instruction with overflow,
16583 // e.g. SADDO, UADDO.
16584 Cond = Cond.getNode()->getOperand(1);
16590 CondOpcode = Cond.getOpcode();
16591 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
16592 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
16593 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
16594 Cond.getOperand(0).getValueType() != MVT::i8)) {
16595 SDValue LHS = Cond.getOperand(0);
16596 SDValue RHS = Cond.getOperand(1);
16597 unsigned X86Opcode;
16600 // Keep this in sync with LowerXALUO, otherwise we might create redundant
16601 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
16603 switch (CondOpcode) {
16604 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
16606 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16608 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
16611 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
16612 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
16614 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16616 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
16619 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
16620 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
16621 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
16622 default: llvm_unreachable("unexpected overflowing operator");
16625 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
16626 if (CondOpcode == ISD::UMULO)
16627 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
16630 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
16632 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
16634 if (CondOpcode == ISD::UMULO)
16635 Cond = X86Op.getValue(2);
16637 Cond = X86Op.getValue(1);
16639 CC = DAG.getConstant(X86Cond, MVT::i8);
16643 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
16644 SDValue Cmp = Cond.getOperand(0).getOperand(1);
16645 if (CondOpc == ISD::OR) {
16646 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
16647 // two branches instead of an explicit OR instruction with a
16649 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16650 isX86LogicalCmp(Cmp)) {
16651 CC = Cond.getOperand(0).getOperand(0);
16652 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16653 Chain, Dest, CC, Cmp);
16654 CC = Cond.getOperand(1).getOperand(0);
16658 } else { // ISD::AND
16659 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
16660 // two branches instead of an explicit AND instruction with a
16661 // separate test. However, we only do this if this block doesn't
16662 // have a fall-through edge, because this requires an explicit
16663 // jmp when the condition is false.
16664 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16665 isX86LogicalCmp(Cmp) &&
16666 Op.getNode()->hasOneUse()) {
16667 X86::CondCode CCode =
16668 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16669 CCode = X86::GetOppositeBranchCondition(CCode);
16670 CC = DAG.getConstant(CCode, MVT::i8);
16671 SDNode *User = *Op.getNode()->use_begin();
16672 // Look for an unconditional branch following this conditional branch.
16673 // We need this because we need to reverse the successors in order
16674 // to implement FCMP_OEQ.
16675 if (User->getOpcode() == ISD::BR) {
16676 SDValue FalseBB = User->getOperand(1);
16678 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16679 assert(NewBR == User);
16683 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16684 Chain, Dest, CC, Cmp);
16685 X86::CondCode CCode =
16686 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
16687 CCode = X86::GetOppositeBranchCondition(CCode);
16688 CC = DAG.getConstant(CCode, MVT::i8);
16694 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
16695 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
16696 // It should be transformed during dag combiner except when the condition
16697 // is set by a arithmetics with overflow node.
16698 X86::CondCode CCode =
16699 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16700 CCode = X86::GetOppositeBranchCondition(CCode);
16701 CC = DAG.getConstant(CCode, MVT::i8);
16702 Cond = Cond.getOperand(0).getOperand(1);
16704 } else if (Cond.getOpcode() == ISD::SETCC &&
16705 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
16706 // For FCMP_OEQ, we can emit
16707 // two branches instead of an explicit AND instruction with a
16708 // separate test. However, we only do this if this block doesn't
16709 // have a fall-through edge, because this requires an explicit
16710 // jmp when the condition is false.
16711 if (Op.getNode()->hasOneUse()) {
16712 SDNode *User = *Op.getNode()->use_begin();
16713 // Look for an unconditional branch following this conditional branch.
16714 // We need this because we need to reverse the successors in order
16715 // to implement FCMP_OEQ.
16716 if (User->getOpcode() == ISD::BR) {
16717 SDValue FalseBB = User->getOperand(1);
16719 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16720 assert(NewBR == User);
16724 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
16725 Cond.getOperand(0), Cond.getOperand(1));
16726 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16727 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16728 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16729 Chain, Dest, CC, Cmp);
16730 CC = DAG.getConstant(X86::COND_P, MVT::i8);
16735 } else if (Cond.getOpcode() == ISD::SETCC &&
16736 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
16737 // For FCMP_UNE, we can emit
16738 // two branches instead of an explicit AND instruction with a
16739 // separate test. However, we only do this if this block doesn't
16740 // have a fall-through edge, because this requires an explicit
16741 // jmp when the condition is false.
16742 if (Op.getNode()->hasOneUse()) {
16743 SDNode *User = *Op.getNode()->use_begin();
16744 // Look for an unconditional branch following this conditional branch.
16745 // We need this because we need to reverse the successors in order
16746 // to implement FCMP_UNE.
16747 if (User->getOpcode() == ISD::BR) {
16748 SDValue FalseBB = User->getOperand(1);
16750 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16751 assert(NewBR == User);
16754 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
16755 Cond.getOperand(0), Cond.getOperand(1));
16756 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16757 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16758 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16759 Chain, Dest, CC, Cmp);
16760 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
16770 // Look pass the truncate if the high bits are known zero.
16771 if (isTruncWithZeroHighBitsInput(Cond, DAG))
16772 Cond = Cond.getOperand(0);
16774 // We know the result of AND is compared against zero. Try to match
16776 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
16777 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
16778 if (NewSetCC.getNode()) {
16779 CC = NewSetCC.getOperand(0);
16780 Cond = NewSetCC.getOperand(1);
16787 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
16788 CC = DAG.getConstant(X86Cond, MVT::i8);
16789 Cond = EmitTest(Cond, X86Cond, dl, DAG);
16791 Cond = ConvertCmpIfNecessary(Cond, DAG);
16792 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16793 Chain, Dest, CC, Cond);
16796 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
16797 // Calls to _alloca are needed to probe the stack when allocating more than 4k
16798 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
16799 // that the guard pages used by the OS virtual memory manager are allocated in
16800 // correct sequence.
16802 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
16803 SelectionDAG &DAG) const {
16804 MachineFunction &MF = DAG.getMachineFunction();
16805 bool SplitStack = MF.shouldSplitStack();
16806 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
16811 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16812 SDNode* Node = Op.getNode();
16814 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
16815 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
16816 " not tell us which reg is the stack pointer!");
16817 EVT VT = Node->getValueType(0);
16818 SDValue Tmp1 = SDValue(Node, 0);
16819 SDValue Tmp2 = SDValue(Node, 1);
16820 SDValue Tmp3 = Node->getOperand(2);
16821 SDValue Chain = Tmp1.getOperand(0);
16823 // Chain the dynamic stack allocation so that it doesn't modify the stack
16824 // pointer when other instructions are using the stack.
16825 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
16828 SDValue Size = Tmp2.getOperand(1);
16829 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
16830 Chain = SP.getValue(1);
16831 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
16832 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16833 unsigned StackAlign = TFI.getStackAlignment();
16834 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
16835 if (Align > StackAlign)
16836 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
16837 DAG.getConstant(-(uint64_t)Align, VT));
16838 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
16840 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
16841 DAG.getIntPtrConstant(0, true), SDValue(),
16844 SDValue Ops[2] = { Tmp1, Tmp2 };
16845 return DAG.getMergeValues(Ops, dl);
16849 SDValue Chain = Op.getOperand(0);
16850 SDValue Size = Op.getOperand(1);
16851 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
16852 EVT VT = Op.getNode()->getValueType(0);
16854 bool Is64Bit = Subtarget->is64Bit();
16855 EVT SPTy = getPointerTy();
16858 MachineRegisterInfo &MRI = MF.getRegInfo();
16861 // The 64 bit implementation of segmented stacks needs to clobber both r10
16862 // r11. This makes it impossible to use it along with nested parameters.
16863 const Function *F = MF.getFunction();
16865 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
16867 if (I->hasNestAttr())
16868 report_fatal_error("Cannot use segmented stacks with functions that "
16869 "have nested arguments.");
16872 const TargetRegisterClass *AddrRegClass =
16873 getRegClassFor(getPointerTy());
16874 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
16875 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
16876 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
16877 DAG.getRegister(Vreg, SPTy));
16878 SDValue Ops1[2] = { Value, Chain };
16879 return DAG.getMergeValues(Ops1, dl);
16882 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
16884 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
16885 Flag = Chain.getValue(1);
16886 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
16888 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
16890 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16891 unsigned SPReg = RegInfo->getStackRegister();
16892 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
16893 Chain = SP.getValue(1);
16896 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
16897 DAG.getConstant(-(uint64_t)Align, VT));
16898 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
16901 SDValue Ops1[2] = { SP, Chain };
16902 return DAG.getMergeValues(Ops1, dl);
16906 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
16907 MachineFunction &MF = DAG.getMachineFunction();
16908 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16910 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16913 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
16914 // vastart just stores the address of the VarArgsFrameIndex slot into the
16915 // memory location argument.
16916 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
16918 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
16919 MachinePointerInfo(SV), false, false, 0);
16923 // gp_offset (0 - 6 * 8)
16924 // fp_offset (48 - 48 + 8 * 16)
16925 // overflow_arg_area (point to parameters coming in memory).
16927 SmallVector<SDValue, 8> MemOps;
16928 SDValue FIN = Op.getOperand(1);
16930 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
16931 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
16933 FIN, MachinePointerInfo(SV), false, false, 0);
16934 MemOps.push_back(Store);
16937 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16938 FIN, DAG.getIntPtrConstant(4));
16939 Store = DAG.getStore(Op.getOperand(0), DL,
16940 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
16942 FIN, MachinePointerInfo(SV, 4), false, false, 0);
16943 MemOps.push_back(Store);
16945 // Store ptr to overflow_arg_area
16946 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16947 FIN, DAG.getIntPtrConstant(4));
16948 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
16950 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
16951 MachinePointerInfo(SV, 8),
16953 MemOps.push_back(Store);
16955 // Store ptr to reg_save_area.
16956 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16957 FIN, DAG.getIntPtrConstant(8));
16958 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
16960 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
16961 MachinePointerInfo(SV, 16), false, false, 0);
16962 MemOps.push_back(Store);
16963 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
16966 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
16967 assert(Subtarget->is64Bit() &&
16968 "LowerVAARG only handles 64-bit va_arg!");
16969 assert((Subtarget->isTargetLinux() ||
16970 Subtarget->isTargetDarwin()) &&
16971 "Unhandled target in LowerVAARG");
16972 assert(Op.getNode()->getNumOperands() == 4);
16973 SDValue Chain = Op.getOperand(0);
16974 SDValue SrcPtr = Op.getOperand(1);
16975 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16976 unsigned Align = Op.getConstantOperandVal(3);
16979 EVT ArgVT = Op.getNode()->getValueType(0);
16980 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16981 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
16984 // Decide which area this value should be read from.
16985 // TODO: Implement the AMD64 ABI in its entirety. This simple
16986 // selection mechanism works only for the basic types.
16987 if (ArgVT == MVT::f80) {
16988 llvm_unreachable("va_arg for f80 not yet implemented");
16989 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
16990 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
16991 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
16992 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
16994 llvm_unreachable("Unhandled argument type in LowerVAARG");
16997 if (ArgMode == 2) {
16998 // Sanity Check: Make sure using fp_offset makes sense.
16999 assert(!DAG.getTarget().Options.UseSoftFloat &&
17000 !(DAG.getMachineFunction()
17001 .getFunction()->getAttributes()
17002 .hasAttribute(AttributeSet::FunctionIndex,
17003 Attribute::NoImplicitFloat)) &&
17004 Subtarget->hasSSE1());
17007 // Insert VAARG_64 node into the DAG
17008 // VAARG_64 returns two values: Variable Argument Address, Chain
17009 SmallVector<SDValue, 11> InstOps;
17010 InstOps.push_back(Chain);
17011 InstOps.push_back(SrcPtr);
17012 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
17013 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
17014 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
17015 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
17016 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
17017 VTs, InstOps, MVT::i64,
17018 MachinePointerInfo(SV),
17020 /*Volatile=*/false,
17022 /*WriteMem=*/true);
17023 Chain = VAARG.getValue(1);
17025 // Load the next argument and return it
17026 return DAG.getLoad(ArgVT, dl,
17029 MachinePointerInfo(),
17030 false, false, false, 0);
17033 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
17034 SelectionDAG &DAG) {
17035 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
17036 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
17037 SDValue Chain = Op.getOperand(0);
17038 SDValue DstPtr = Op.getOperand(1);
17039 SDValue SrcPtr = Op.getOperand(2);
17040 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
17041 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17044 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
17045 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
17047 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
17050 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
17051 // amount is a constant. Takes immediate version of shift as input.
17052 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
17053 SDValue SrcOp, uint64_t ShiftAmt,
17054 SelectionDAG &DAG) {
17055 MVT ElementType = VT.getVectorElementType();
17057 // Fold this packed shift into its first operand if ShiftAmt is 0.
17061 // Check for ShiftAmt >= element width
17062 if (ShiftAmt >= ElementType.getSizeInBits()) {
17063 if (Opc == X86ISD::VSRAI)
17064 ShiftAmt = ElementType.getSizeInBits() - 1;
17066 return DAG.getConstant(0, VT);
17069 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
17070 && "Unknown target vector shift-by-constant node");
17072 // Fold this packed vector shift into a build vector if SrcOp is a
17073 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
17074 if (VT == SrcOp.getSimpleValueType() &&
17075 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
17076 SmallVector<SDValue, 8> Elts;
17077 unsigned NumElts = SrcOp->getNumOperands();
17078 ConstantSDNode *ND;
17081 default: llvm_unreachable(nullptr);
17082 case X86ISD::VSHLI:
17083 for (unsigned i=0; i!=NumElts; ++i) {
17084 SDValue CurrentOp = SrcOp->getOperand(i);
17085 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17086 Elts.push_back(CurrentOp);
17089 ND = cast<ConstantSDNode>(CurrentOp);
17090 const APInt &C = ND->getAPIntValue();
17091 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
17094 case X86ISD::VSRLI:
17095 for (unsigned i=0; i!=NumElts; ++i) {
17096 SDValue CurrentOp = SrcOp->getOperand(i);
17097 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17098 Elts.push_back(CurrentOp);
17101 ND = cast<ConstantSDNode>(CurrentOp);
17102 const APInt &C = ND->getAPIntValue();
17103 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
17106 case X86ISD::VSRAI:
17107 for (unsigned i=0; i!=NumElts; ++i) {
17108 SDValue CurrentOp = SrcOp->getOperand(i);
17109 if (CurrentOp->getOpcode() == ISD::UNDEF) {
17110 Elts.push_back(CurrentOp);
17113 ND = cast<ConstantSDNode>(CurrentOp);
17114 const APInt &C = ND->getAPIntValue();
17115 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
17120 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17123 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
17126 // getTargetVShiftNode - Handle vector element shifts where the shift amount
17127 // may or may not be a constant. Takes immediate version of shift as input.
17128 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
17129 SDValue SrcOp, SDValue ShAmt,
17130 SelectionDAG &DAG) {
17131 MVT SVT = ShAmt.getSimpleValueType();
17132 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
17134 // Catch shift-by-constant.
17135 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
17136 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
17137 CShAmt->getZExtValue(), DAG);
17139 // Change opcode to non-immediate version
17141 default: llvm_unreachable("Unknown target vector shift node");
17142 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
17143 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
17144 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
17147 const X86Subtarget &Subtarget =
17148 static_cast<const X86Subtarget &>(DAG.getSubtarget());
17149 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
17150 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
17151 // Let the shuffle legalizer expand this shift amount node.
17152 SDValue Op0 = ShAmt.getOperand(0);
17153 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
17154 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
17156 // Need to build a vector containing shift amount.
17157 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
17158 SmallVector<SDValue, 4> ShOps;
17159 ShOps.push_back(ShAmt);
17160 if (SVT == MVT::i32) {
17161 ShOps.push_back(DAG.getConstant(0, SVT));
17162 ShOps.push_back(DAG.getUNDEF(SVT));
17164 ShOps.push_back(DAG.getUNDEF(SVT));
17166 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
17167 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
17170 // The return type has to be a 128-bit type with the same element
17171 // type as the input type.
17172 MVT EltVT = VT.getVectorElementType();
17173 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
17175 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
17176 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
17179 /// \brief Return (and \p Op, \p Mask) for compare instructions or
17180 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
17181 /// necessary casting for \p Mask when lowering masking intrinsics.
17182 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
17183 SDValue PreservedSrc,
17184 const X86Subtarget *Subtarget,
17185 SelectionDAG &DAG) {
17186 EVT VT = Op.getValueType();
17187 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
17188 MVT::i1, VT.getVectorNumElements());
17189 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17190 Mask.getValueType().getSizeInBits());
17193 assert(MaskVT.isSimple() && "invalid mask type");
17195 if (isAllOnes(Mask))
17198 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17199 // are extracted by EXTRACT_SUBVECTOR.
17200 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17201 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17202 DAG.getIntPtrConstant(0));
17204 switch (Op.getOpcode()) {
17206 case X86ISD::PCMPEQM:
17207 case X86ISD::PCMPGTM:
17209 case X86ISD::CMPMU:
17210 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
17212 if (PreservedSrc.getOpcode() == ISD::UNDEF)
17213 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
17214 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
17217 /// \brief Creates an SDNode for a predicated scalar operation.
17218 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
17219 /// The mask is comming as MVT::i8 and it should be truncated
17220 /// to MVT::i1 while lowering masking intrinsics.
17221 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
17222 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
17223 /// a scalar instruction.
17224 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
17225 SDValue PreservedSrc,
17226 const X86Subtarget *Subtarget,
17227 SelectionDAG &DAG) {
17228 if (isAllOnes(Mask))
17231 EVT VT = Op.getValueType();
17233 // The mask should be of type MVT::i1
17234 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
17236 if (PreservedSrc.getOpcode() == ISD::UNDEF)
17237 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
17238 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
17241 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17242 SelectionDAG &DAG) {
17244 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17245 EVT VT = Op.getValueType();
17246 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
17248 switch(IntrData->Type) {
17249 case INTR_TYPE_1OP:
17250 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
17251 case INTR_TYPE_2OP:
17252 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17254 case INTR_TYPE_3OP:
17255 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17256 Op.getOperand(2), Op.getOperand(3));
17257 case INTR_TYPE_1OP_MASK_RM: {
17258 SDValue Src = Op.getOperand(1);
17259 SDValue Src0 = Op.getOperand(2);
17260 SDValue Mask = Op.getOperand(3);
17261 SDValue RoundingMode = Op.getOperand(4);
17262 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
17264 Mask, Src0, Subtarget, DAG);
17266 case INTR_TYPE_SCALAR_MASK_RM: {
17267 SDValue Src1 = Op.getOperand(1);
17268 SDValue Src2 = Op.getOperand(2);
17269 SDValue Src0 = Op.getOperand(3);
17270 SDValue Mask = Op.getOperand(4);
17271 SDValue RoundingMode = Op.getOperand(5);
17272 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
17274 Mask, Src0, Subtarget, DAG);
17276 case INTR_TYPE_2OP_MASK: {
17277 SDValue Mask = Op.getOperand(4);
17278 SDValue PassThru = Op.getOperand(3);
17279 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
17280 if (IntrWithRoundingModeOpcode != 0) {
17281 unsigned Round = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
17282 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
17283 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
17284 dl, Op.getValueType(),
17285 Op.getOperand(1), Op.getOperand(2),
17286 Op.getOperand(3), Op.getOperand(5)),
17287 Mask, PassThru, Subtarget, DAG);
17290 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
17293 Mask, PassThru, Subtarget, DAG);
17295 case FMA_OP_MASK: {
17296 SDValue Src1 = Op.getOperand(1);
17297 SDValue Src2 = Op.getOperand(2);
17298 SDValue Src3 = Op.getOperand(3);
17299 SDValue Mask = Op.getOperand(4);
17300 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
17301 if (IntrWithRoundingModeOpcode != 0) {
17302 SDValue Rnd = Op.getOperand(5);
17303 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
17304 X86::STATIC_ROUNDING::CUR_DIRECTION)
17305 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
17306 dl, Op.getValueType(),
17307 Src1, Src2, Src3, Rnd),
17308 Mask, Src1, Subtarget, DAG);
17310 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
17311 dl, Op.getValueType(),
17313 Mask, Src1, Subtarget, DAG);
17316 case CMP_MASK_CC: {
17317 // Comparison intrinsics with masks.
17318 // Example of transformation:
17319 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
17320 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
17322 // (v8i1 (insert_subvector undef,
17323 // (v2i1 (and (PCMPEQM %a, %b),
17324 // (extract_subvector
17325 // (v8i1 (bitcast %mask)), 0))), 0))))
17326 EVT VT = Op.getOperand(1).getValueType();
17327 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17328 VT.getVectorNumElements());
17329 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
17330 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17331 Mask.getValueType().getSizeInBits());
17333 if (IntrData->Type == CMP_MASK_CC) {
17334 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17335 Op.getOperand(2), Op.getOperand(3));
17337 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
17338 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17341 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
17342 DAG.getTargetConstant(0, MaskVT),
17344 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
17345 DAG.getUNDEF(BitcastVT), CmpMask,
17346 DAG.getIntPtrConstant(0));
17347 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
17349 case COMI: { // Comparison intrinsics
17350 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
17351 SDValue LHS = Op.getOperand(1);
17352 SDValue RHS = Op.getOperand(2);
17353 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
17354 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
17355 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
17356 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17357 DAG.getConstant(X86CC, MVT::i8), Cond);
17358 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17361 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
17362 Op.getOperand(1), Op.getOperand(2), DAG);
17364 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
17365 Op.getSimpleValueType(),
17367 Op.getOperand(2), DAG),
17368 Op.getOperand(4), Op.getOperand(3), Subtarget,
17370 case COMPRESS_EXPAND_IN_REG: {
17371 SDValue Mask = Op.getOperand(3);
17372 SDValue DataToCompress = Op.getOperand(1);
17373 SDValue PassThru = Op.getOperand(2);
17374 if (isAllOnes(Mask)) // return data as is
17375 return Op.getOperand(1);
17376 EVT VT = Op.getValueType();
17377 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17378 VT.getVectorNumElements());
17379 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17380 Mask.getValueType().getSizeInBits());
17382 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17383 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17384 DAG.getIntPtrConstant(0));
17386 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
17390 SDValue Mask = Op.getOperand(3);
17391 EVT VT = Op.getValueType();
17392 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17393 VT.getVectorNumElements());
17394 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17395 Mask.getValueType().getSizeInBits());
17397 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17398 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17399 DAG.getIntPtrConstant(0));
17400 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
17409 default: return SDValue(); // Don't custom lower most intrinsics.
17411 case Intrinsic::x86_avx512_mask_valign_q_512:
17412 case Intrinsic::x86_avx512_mask_valign_d_512:
17413 // Vector source operands are swapped.
17414 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
17415 Op.getValueType(), Op.getOperand(2),
17418 Op.getOperand(5), Op.getOperand(4),
17421 // ptest and testp intrinsics. The intrinsic these come from are designed to
17422 // return an integer value, not just an instruction so lower it to the ptest
17423 // or testp pattern and a setcc for the result.
17424 case Intrinsic::x86_sse41_ptestz:
17425 case Intrinsic::x86_sse41_ptestc:
17426 case Intrinsic::x86_sse41_ptestnzc:
17427 case Intrinsic::x86_avx_ptestz_256:
17428 case Intrinsic::x86_avx_ptestc_256:
17429 case Intrinsic::x86_avx_ptestnzc_256:
17430 case Intrinsic::x86_avx_vtestz_ps:
17431 case Intrinsic::x86_avx_vtestc_ps:
17432 case Intrinsic::x86_avx_vtestnzc_ps:
17433 case Intrinsic::x86_avx_vtestz_pd:
17434 case Intrinsic::x86_avx_vtestc_pd:
17435 case Intrinsic::x86_avx_vtestnzc_pd:
17436 case Intrinsic::x86_avx_vtestz_ps_256:
17437 case Intrinsic::x86_avx_vtestc_ps_256:
17438 case Intrinsic::x86_avx_vtestnzc_ps_256:
17439 case Intrinsic::x86_avx_vtestz_pd_256:
17440 case Intrinsic::x86_avx_vtestc_pd_256:
17441 case Intrinsic::x86_avx_vtestnzc_pd_256: {
17442 bool IsTestPacked = false;
17445 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
17446 case Intrinsic::x86_avx_vtestz_ps:
17447 case Intrinsic::x86_avx_vtestz_pd:
17448 case Intrinsic::x86_avx_vtestz_ps_256:
17449 case Intrinsic::x86_avx_vtestz_pd_256:
17450 IsTestPacked = true; // Fallthrough
17451 case Intrinsic::x86_sse41_ptestz:
17452 case Intrinsic::x86_avx_ptestz_256:
17454 X86CC = X86::COND_E;
17456 case Intrinsic::x86_avx_vtestc_ps:
17457 case Intrinsic::x86_avx_vtestc_pd:
17458 case Intrinsic::x86_avx_vtestc_ps_256:
17459 case Intrinsic::x86_avx_vtestc_pd_256:
17460 IsTestPacked = true; // Fallthrough
17461 case Intrinsic::x86_sse41_ptestc:
17462 case Intrinsic::x86_avx_ptestc_256:
17464 X86CC = X86::COND_B;
17466 case Intrinsic::x86_avx_vtestnzc_ps:
17467 case Intrinsic::x86_avx_vtestnzc_pd:
17468 case Intrinsic::x86_avx_vtestnzc_ps_256:
17469 case Intrinsic::x86_avx_vtestnzc_pd_256:
17470 IsTestPacked = true; // Fallthrough
17471 case Intrinsic::x86_sse41_ptestnzc:
17472 case Intrinsic::x86_avx_ptestnzc_256:
17474 X86CC = X86::COND_A;
17478 SDValue LHS = Op.getOperand(1);
17479 SDValue RHS = Op.getOperand(2);
17480 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
17481 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
17482 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17483 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
17484 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17486 case Intrinsic::x86_avx512_kortestz_w:
17487 case Intrinsic::x86_avx512_kortestc_w: {
17488 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
17489 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
17490 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
17491 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17492 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
17493 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17494 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17497 case Intrinsic::x86_sse42_pcmpistria128:
17498 case Intrinsic::x86_sse42_pcmpestria128:
17499 case Intrinsic::x86_sse42_pcmpistric128:
17500 case Intrinsic::x86_sse42_pcmpestric128:
17501 case Intrinsic::x86_sse42_pcmpistrio128:
17502 case Intrinsic::x86_sse42_pcmpestrio128:
17503 case Intrinsic::x86_sse42_pcmpistris128:
17504 case Intrinsic::x86_sse42_pcmpestris128:
17505 case Intrinsic::x86_sse42_pcmpistriz128:
17506 case Intrinsic::x86_sse42_pcmpestriz128: {
17510 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17511 case Intrinsic::x86_sse42_pcmpistria128:
17512 Opcode = X86ISD::PCMPISTRI;
17513 X86CC = X86::COND_A;
17515 case Intrinsic::x86_sse42_pcmpestria128:
17516 Opcode = X86ISD::PCMPESTRI;
17517 X86CC = X86::COND_A;
17519 case Intrinsic::x86_sse42_pcmpistric128:
17520 Opcode = X86ISD::PCMPISTRI;
17521 X86CC = X86::COND_B;
17523 case Intrinsic::x86_sse42_pcmpestric128:
17524 Opcode = X86ISD::PCMPESTRI;
17525 X86CC = X86::COND_B;
17527 case Intrinsic::x86_sse42_pcmpistrio128:
17528 Opcode = X86ISD::PCMPISTRI;
17529 X86CC = X86::COND_O;
17531 case Intrinsic::x86_sse42_pcmpestrio128:
17532 Opcode = X86ISD::PCMPESTRI;
17533 X86CC = X86::COND_O;
17535 case Intrinsic::x86_sse42_pcmpistris128:
17536 Opcode = X86ISD::PCMPISTRI;
17537 X86CC = X86::COND_S;
17539 case Intrinsic::x86_sse42_pcmpestris128:
17540 Opcode = X86ISD::PCMPESTRI;
17541 X86CC = X86::COND_S;
17543 case Intrinsic::x86_sse42_pcmpistriz128:
17544 Opcode = X86ISD::PCMPISTRI;
17545 X86CC = X86::COND_E;
17547 case Intrinsic::x86_sse42_pcmpestriz128:
17548 Opcode = X86ISD::PCMPESTRI;
17549 X86CC = X86::COND_E;
17552 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17553 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17554 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17555 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17556 DAG.getConstant(X86CC, MVT::i8),
17557 SDValue(PCMP.getNode(), 1));
17558 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17561 case Intrinsic::x86_sse42_pcmpistri128:
17562 case Intrinsic::x86_sse42_pcmpestri128: {
17564 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17565 Opcode = X86ISD::PCMPISTRI;
17567 Opcode = X86ISD::PCMPESTRI;
17569 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17570 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17571 return DAG.getNode(Opcode, dl, VTs, NewOps);
17576 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17577 SDValue Src, SDValue Mask, SDValue Base,
17578 SDValue Index, SDValue ScaleOp, SDValue Chain,
17579 const X86Subtarget * Subtarget) {
17581 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17582 assert(C && "Invalid scale type");
17583 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17584 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17585 Index.getSimpleValueType().getVectorNumElements());
17587 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17589 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17591 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17592 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17593 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17594 SDValue Segment = DAG.getRegister(0, MVT::i32);
17595 if (Src.getOpcode() == ISD::UNDEF)
17596 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
17597 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17598 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17599 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17600 return DAG.getMergeValues(RetOps, dl);
17603 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17604 SDValue Src, SDValue Mask, SDValue Base,
17605 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17607 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17608 assert(C && "Invalid scale type");
17609 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17610 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17611 SDValue Segment = DAG.getRegister(0, MVT::i32);
17612 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17613 Index.getSimpleValueType().getVectorNumElements());
17615 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17617 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17619 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17620 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17621 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17622 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17623 return SDValue(Res, 1);
17626 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17627 SDValue Mask, SDValue Base, SDValue Index,
17628 SDValue ScaleOp, SDValue Chain) {
17630 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17631 assert(C && "Invalid scale type");
17632 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17633 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17634 SDValue Segment = DAG.getRegister(0, MVT::i32);
17636 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17638 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17640 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17642 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17643 //SDVTList VTs = DAG.getVTList(MVT::Other);
17644 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17645 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17646 return SDValue(Res, 0);
17649 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17650 // read performance monitor counters (x86_rdpmc).
17651 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17652 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17653 SmallVectorImpl<SDValue> &Results) {
17654 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17655 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17658 // The ECX register is used to select the index of the performance counter
17660 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17662 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17664 // Reads the content of a 64-bit performance counter and returns it in the
17665 // registers EDX:EAX.
17666 if (Subtarget->is64Bit()) {
17667 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17668 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17671 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17672 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17675 Chain = HI.getValue(1);
17677 if (Subtarget->is64Bit()) {
17678 // The EAX register is loaded with the low-order 32 bits. The EDX register
17679 // is loaded with the supported high-order bits of the counter.
17680 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17681 DAG.getConstant(32, MVT::i8));
17682 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17683 Results.push_back(Chain);
17687 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17688 SDValue Ops[] = { LO, HI };
17689 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17690 Results.push_back(Pair);
17691 Results.push_back(Chain);
17694 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17695 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17696 // also used to custom lower READCYCLECOUNTER nodes.
17697 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17698 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17699 SmallVectorImpl<SDValue> &Results) {
17700 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17701 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17704 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17705 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17706 // and the EAX register is loaded with the low-order 32 bits.
17707 if (Subtarget->is64Bit()) {
17708 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17709 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17712 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17713 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17716 SDValue Chain = HI.getValue(1);
17718 if (Opcode == X86ISD::RDTSCP_DAG) {
17719 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17721 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17722 // the ECX register. Add 'ecx' explicitly to the chain.
17723 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17725 // Explicitly store the content of ECX at the location passed in input
17726 // to the 'rdtscp' intrinsic.
17727 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17728 MachinePointerInfo(), false, false, 0);
17731 if (Subtarget->is64Bit()) {
17732 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17733 // the EAX register is loaded with the low-order 32 bits.
17734 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17735 DAG.getConstant(32, MVT::i8));
17736 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17737 Results.push_back(Chain);
17741 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17742 SDValue Ops[] = { LO, HI };
17743 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17744 Results.push_back(Pair);
17745 Results.push_back(Chain);
17748 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17749 SelectionDAG &DAG) {
17750 SmallVector<SDValue, 2> Results;
17752 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17754 return DAG.getMergeValues(Results, DL);
17758 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17759 SelectionDAG &DAG) {
17760 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17762 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17767 switch(IntrData->Type) {
17769 llvm_unreachable("Unknown Intrinsic Type");
17773 // Emit the node with the right value type.
17774 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17775 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17777 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17778 // Otherwise return the value from Rand, which is always 0, casted to i32.
17779 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17780 DAG.getConstant(1, Op->getValueType(1)),
17781 DAG.getConstant(X86::COND_B, MVT::i32),
17782 SDValue(Result.getNode(), 1) };
17783 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17784 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17787 // Return { result, isValid, chain }.
17788 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17789 SDValue(Result.getNode(), 2));
17792 //gather(v1, mask, index, base, scale);
17793 SDValue Chain = Op.getOperand(0);
17794 SDValue Src = Op.getOperand(2);
17795 SDValue Base = Op.getOperand(3);
17796 SDValue Index = Op.getOperand(4);
17797 SDValue Mask = Op.getOperand(5);
17798 SDValue Scale = Op.getOperand(6);
17799 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
17803 //scatter(base, mask, index, v1, scale);
17804 SDValue Chain = Op.getOperand(0);
17805 SDValue Base = Op.getOperand(2);
17806 SDValue Mask = Op.getOperand(3);
17807 SDValue Index = Op.getOperand(4);
17808 SDValue Src = Op.getOperand(5);
17809 SDValue Scale = Op.getOperand(6);
17810 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
17813 SDValue Hint = Op.getOperand(6);
17815 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
17816 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
17817 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
17818 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17819 SDValue Chain = Op.getOperand(0);
17820 SDValue Mask = Op.getOperand(2);
17821 SDValue Index = Op.getOperand(3);
17822 SDValue Base = Op.getOperand(4);
17823 SDValue Scale = Op.getOperand(5);
17824 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17826 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17828 SmallVector<SDValue, 2> Results;
17829 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
17830 return DAG.getMergeValues(Results, dl);
17832 // Read Performance Monitoring Counters.
17834 SmallVector<SDValue, 2> Results;
17835 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17836 return DAG.getMergeValues(Results, dl);
17838 // XTEST intrinsics.
17840 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17841 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17842 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17843 DAG.getConstant(X86::COND_NE, MVT::i8),
17845 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17846 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17847 Ret, SDValue(InTrans.getNode(), 1));
17851 SmallVector<SDValue, 2> Results;
17852 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17853 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17854 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17855 DAG.getConstant(-1, MVT::i8));
17856 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17857 Op.getOperand(4), GenCF.getValue(1));
17858 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17859 Op.getOperand(5), MachinePointerInfo(),
17861 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17862 DAG.getConstant(X86::COND_B, MVT::i8),
17864 Results.push_back(SetCC);
17865 Results.push_back(Store);
17866 return DAG.getMergeValues(Results, dl);
17868 case COMPRESS_TO_MEM: {
17870 SDValue Mask = Op.getOperand(4);
17871 SDValue DataToCompress = Op.getOperand(3);
17872 SDValue Addr = Op.getOperand(2);
17873 SDValue Chain = Op.getOperand(0);
17875 if (isAllOnes(Mask)) // return just a store
17876 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17877 MachinePointerInfo(), false, false, 0);
17879 EVT VT = DataToCompress.getValueType();
17880 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17881 VT.getVectorNumElements());
17882 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17883 Mask.getValueType().getSizeInBits());
17884 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17885 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17886 DAG.getIntPtrConstant(0));
17888 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
17889 DataToCompress, DAG.getUNDEF(VT));
17890 return DAG.getStore(Chain, dl, Compressed, Addr,
17891 MachinePointerInfo(), false, false, 0);
17893 case EXPAND_FROM_MEM: {
17895 SDValue Mask = Op.getOperand(4);
17896 SDValue PathThru = Op.getOperand(3);
17897 SDValue Addr = Op.getOperand(2);
17898 SDValue Chain = Op.getOperand(0);
17899 EVT VT = Op.getValueType();
17901 if (isAllOnes(Mask)) // return just a load
17902 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17904 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17905 VT.getVectorNumElements());
17906 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17907 Mask.getValueType().getSizeInBits());
17908 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17909 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17910 DAG.getIntPtrConstant(0));
17912 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17913 false, false, false, 0);
17915 SmallVector<SDValue, 2> Results;
17916 Results.push_back(DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand,
17918 Results.push_back(Chain);
17919 return DAG.getMergeValues(Results, dl);
17924 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17925 SelectionDAG &DAG) const {
17926 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17927 MFI->setReturnAddressIsTaken(true);
17929 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17932 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17934 EVT PtrVT = getPointerTy();
17937 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17938 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17939 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
17940 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17941 DAG.getNode(ISD::ADD, dl, PtrVT,
17942 FrameAddr, Offset),
17943 MachinePointerInfo(), false, false, false, 0);
17946 // Just load the return address.
17947 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17948 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17949 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17952 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17953 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17954 MFI->setFrameAddressIsTaken(true);
17956 EVT VT = Op.getValueType();
17957 SDLoc dl(Op); // FIXME probably not meaningful
17958 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17959 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17960 unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(
17961 DAG.getMachineFunction());
17962 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17963 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17964 "Invalid Frame Register!");
17965 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17967 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17968 MachinePointerInfo(),
17969 false, false, false, 0);
17973 // FIXME? Maybe this could be a TableGen attribute on some registers and
17974 // this table could be generated automatically from RegInfo.
17975 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
17977 unsigned Reg = StringSwitch<unsigned>(RegName)
17978 .Case("esp", X86::ESP)
17979 .Case("rsp", X86::RSP)
17983 report_fatal_error("Invalid register name global variable");
17986 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17987 SelectionDAG &DAG) const {
17988 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17989 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
17992 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17993 SDValue Chain = Op.getOperand(0);
17994 SDValue Offset = Op.getOperand(1);
17995 SDValue Handler = Op.getOperand(2);
17998 EVT PtrVT = getPointerTy();
17999 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18000 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
18001 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
18002 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
18003 "Invalid Frame Register!");
18004 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
18005 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
18007 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
18008 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
18009 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
18010 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
18012 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
18014 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
18015 DAG.getRegister(StoreAddrReg, PtrVT));
18018 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
18019 SelectionDAG &DAG) const {
18021 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
18022 DAG.getVTList(MVT::i32, MVT::Other),
18023 Op.getOperand(0), Op.getOperand(1));
18026 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
18027 SelectionDAG &DAG) const {
18029 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
18030 Op.getOperand(0), Op.getOperand(1));
18033 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
18034 return Op.getOperand(0);
18037 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
18038 SelectionDAG &DAG) const {
18039 SDValue Root = Op.getOperand(0);
18040 SDValue Trmp = Op.getOperand(1); // trampoline
18041 SDValue FPtr = Op.getOperand(2); // nested function
18042 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
18045 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
18046 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18048 if (Subtarget->is64Bit()) {
18049 SDValue OutChains[6];
18051 // Large code-model.
18052 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
18053 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
18055 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
18056 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
18058 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
18060 // Load the pointer to the nested function into R11.
18061 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
18062 SDValue Addr = Trmp;
18063 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18064 Addr, MachinePointerInfo(TrmpAddr),
18067 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18068 DAG.getConstant(2, MVT::i64));
18069 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
18070 MachinePointerInfo(TrmpAddr, 2),
18073 // Load the 'nest' parameter value into R10.
18074 // R10 is specified in X86CallingConv.td
18075 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
18076 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18077 DAG.getConstant(10, MVT::i64));
18078 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18079 Addr, MachinePointerInfo(TrmpAddr, 10),
18082 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18083 DAG.getConstant(12, MVT::i64));
18084 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
18085 MachinePointerInfo(TrmpAddr, 12),
18088 // Jump to the nested function.
18089 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
18090 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18091 DAG.getConstant(20, MVT::i64));
18092 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
18093 Addr, MachinePointerInfo(TrmpAddr, 20),
18096 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
18097 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
18098 DAG.getConstant(22, MVT::i64));
18099 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
18100 MachinePointerInfo(TrmpAddr, 22),
18103 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
18105 const Function *Func =
18106 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
18107 CallingConv::ID CC = Func->getCallingConv();
18112 llvm_unreachable("Unsupported calling convention");
18113 case CallingConv::C:
18114 case CallingConv::X86_StdCall: {
18115 // Pass 'nest' parameter in ECX.
18116 // Must be kept in sync with X86CallingConv.td
18117 NestReg = X86::ECX;
18119 // Check that ECX wasn't needed by an 'inreg' parameter.
18120 FunctionType *FTy = Func->getFunctionType();
18121 const AttributeSet &Attrs = Func->getAttributes();
18123 if (!Attrs.isEmpty() && !Func->isVarArg()) {
18124 unsigned InRegCount = 0;
18127 for (FunctionType::param_iterator I = FTy->param_begin(),
18128 E = FTy->param_end(); I != E; ++I, ++Idx)
18129 if (Attrs.hasAttribute(Idx, Attribute::InReg))
18130 // FIXME: should only count parameters that are lowered to integers.
18131 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
18133 if (InRegCount > 2) {
18134 report_fatal_error("Nest register in use - reduce number of inreg"
18140 case CallingConv::X86_FastCall:
18141 case CallingConv::X86_ThisCall:
18142 case CallingConv::Fast:
18143 // Pass 'nest' parameter in EAX.
18144 // Must be kept in sync with X86CallingConv.td
18145 NestReg = X86::EAX;
18149 SDValue OutChains[4];
18150 SDValue Addr, Disp;
18152 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18153 DAG.getConstant(10, MVT::i32));
18154 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
18156 // This is storing the opcode for MOV32ri.
18157 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
18158 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
18159 OutChains[0] = DAG.getStore(Root, dl,
18160 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
18161 Trmp, MachinePointerInfo(TrmpAddr),
18164 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18165 DAG.getConstant(1, MVT::i32));
18166 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
18167 MachinePointerInfo(TrmpAddr, 1),
18170 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
18171 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18172 DAG.getConstant(5, MVT::i32));
18173 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
18174 MachinePointerInfo(TrmpAddr, 5),
18177 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
18178 DAG.getConstant(6, MVT::i32));
18179 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
18180 MachinePointerInfo(TrmpAddr, 6),
18183 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
18187 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
18188 SelectionDAG &DAG) const {
18190 The rounding mode is in bits 11:10 of FPSR, and has the following
18192 00 Round to nearest
18197 FLT_ROUNDS, on the other hand, expects the following:
18204 To perform the conversion, we do:
18205 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
18208 MachineFunction &MF = DAG.getMachineFunction();
18209 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
18210 unsigned StackAlignment = TFI.getStackAlignment();
18211 MVT VT = Op.getSimpleValueType();
18214 // Save FP Control Word to stack slot
18215 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
18216 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
18218 MachineMemOperand *MMO =
18219 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
18220 MachineMemOperand::MOStore, 2, 2);
18222 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
18223 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
18224 DAG.getVTList(MVT::Other),
18225 Ops, MVT::i16, MMO);
18227 // Load FP Control Word from stack slot
18228 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
18229 MachinePointerInfo(), false, false, false, 0);
18231 // Transform as necessary
18233 DAG.getNode(ISD::SRL, DL, MVT::i16,
18234 DAG.getNode(ISD::AND, DL, MVT::i16,
18235 CWD, DAG.getConstant(0x800, MVT::i16)),
18236 DAG.getConstant(11, MVT::i8));
18238 DAG.getNode(ISD::SRL, DL, MVT::i16,
18239 DAG.getNode(ISD::AND, DL, MVT::i16,
18240 CWD, DAG.getConstant(0x400, MVT::i16)),
18241 DAG.getConstant(9, MVT::i8));
18244 DAG.getNode(ISD::AND, DL, MVT::i16,
18245 DAG.getNode(ISD::ADD, DL, MVT::i16,
18246 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
18247 DAG.getConstant(1, MVT::i16)),
18248 DAG.getConstant(3, MVT::i16));
18250 return DAG.getNode((VT.getSizeInBits() < 16 ?
18251 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
18254 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
18255 MVT VT = Op.getSimpleValueType();
18257 unsigned NumBits = VT.getSizeInBits();
18260 Op = Op.getOperand(0);
18261 if (VT == MVT::i8) {
18262 // Zero extend to i32 since there is not an i8 bsr.
18264 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18267 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
18268 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18269 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18271 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18274 DAG.getConstant(NumBits+NumBits-1, OpVT),
18275 DAG.getConstant(X86::COND_E, MVT::i8),
18278 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18280 // Finally xor with NumBits-1.
18281 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18284 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18288 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
18289 MVT VT = Op.getSimpleValueType();
18291 unsigned NumBits = VT.getSizeInBits();
18294 Op = Op.getOperand(0);
18295 if (VT == MVT::i8) {
18296 // Zero extend to i32 since there is not an i8 bsr.
18298 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18301 // Issue a bsr (scan bits in reverse).
18302 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18303 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18305 // And xor with NumBits-1.
18306 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18309 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18313 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18314 MVT VT = Op.getSimpleValueType();
18315 unsigned NumBits = VT.getSizeInBits();
18317 Op = Op.getOperand(0);
18319 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18320 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18321 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
18323 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18326 DAG.getConstant(NumBits, VT),
18327 DAG.getConstant(X86::COND_E, MVT::i8),
18330 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18333 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18334 // ones, and then concatenate the result back.
18335 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18336 MVT VT = Op.getSimpleValueType();
18338 assert(VT.is256BitVector() && VT.isInteger() &&
18339 "Unsupported value type for operation");
18341 unsigned NumElems = VT.getVectorNumElements();
18344 // Extract the LHS vectors
18345 SDValue LHS = Op.getOperand(0);
18346 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18347 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18349 // Extract the RHS vectors
18350 SDValue RHS = Op.getOperand(1);
18351 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18352 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18354 MVT EltVT = VT.getVectorElementType();
18355 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18357 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18358 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18359 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18362 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18363 assert(Op.getSimpleValueType().is256BitVector() &&
18364 Op.getSimpleValueType().isInteger() &&
18365 "Only handle AVX 256-bit vector integer operation");
18366 return Lower256IntArith(Op, DAG);
18369 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18370 assert(Op.getSimpleValueType().is256BitVector() &&
18371 Op.getSimpleValueType().isInteger() &&
18372 "Only handle AVX 256-bit vector integer operation");
18373 return Lower256IntArith(Op, DAG);
18376 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18377 SelectionDAG &DAG) {
18379 MVT VT = Op.getSimpleValueType();
18381 // Decompose 256-bit ops into smaller 128-bit ops.
18382 if (VT.is256BitVector() && !Subtarget->hasInt256())
18383 return Lower256IntArith(Op, DAG);
18385 SDValue A = Op.getOperand(0);
18386 SDValue B = Op.getOperand(1);
18388 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18389 if (VT == MVT::v4i32) {
18390 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18391 "Should not custom lower when pmuldq is available!");
18393 // Extract the odd parts.
18394 static const int UnpackMask[] = { 1, -1, 3, -1 };
18395 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18396 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18398 // Multiply the even parts.
18399 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18400 // Now multiply odd parts.
18401 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18403 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
18404 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
18406 // Merge the two vectors back together with a shuffle. This expands into 2
18408 static const int ShufMask[] = { 0, 4, 2, 6 };
18409 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18412 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18413 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18415 // Ahi = psrlqi(a, 32);
18416 // Bhi = psrlqi(b, 32);
18418 // AloBlo = pmuludq(a, b);
18419 // AloBhi = pmuludq(a, Bhi);
18420 // AhiBlo = pmuludq(Ahi, b);
18422 // AloBhi = psllqi(AloBhi, 32);
18423 // AhiBlo = psllqi(AhiBlo, 32);
18424 // return AloBlo + AloBhi + AhiBlo;
18426 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18427 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18429 // Bit cast to 32-bit vectors for MULUDQ
18430 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18431 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18432 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
18433 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
18434 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
18435 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
18437 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18438 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18439 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18441 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18442 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18444 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18445 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18448 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18449 assert(Subtarget->isTargetWin64() && "Unexpected target");
18450 EVT VT = Op.getValueType();
18451 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18452 "Unexpected return type for lowering");
18456 switch (Op->getOpcode()) {
18457 default: llvm_unreachable("Unexpected request for libcall!");
18458 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18459 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18460 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18461 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18462 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18463 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18467 SDValue InChain = DAG.getEntryNode();
18469 TargetLowering::ArgListTy Args;
18470 TargetLowering::ArgListEntry Entry;
18471 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18472 EVT ArgVT = Op->getOperand(i).getValueType();
18473 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18474 "Unexpected argument type for lowering");
18475 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18476 Entry.Node = StackPtr;
18477 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18479 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18480 Entry.Ty = PointerType::get(ArgTy,0);
18481 Entry.isSExt = false;
18482 Entry.isZExt = false;
18483 Args.push_back(Entry);
18486 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18489 TargetLowering::CallLoweringInfo CLI(DAG);
18490 CLI.setDebugLoc(dl).setChain(InChain)
18491 .setCallee(getLibcallCallingConv(LC),
18492 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18493 Callee, std::move(Args), 0)
18494 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18496 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18497 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
18500 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18501 SelectionDAG &DAG) {
18502 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18503 EVT VT = Op0.getValueType();
18506 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18507 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18509 // PMULxD operations multiply each even value (starting at 0) of LHS with
18510 // the related value of RHS and produce a widen result.
18511 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18512 // => <2 x i64> <ae|cg>
18514 // In other word, to have all the results, we need to perform two PMULxD:
18515 // 1. one with the even values.
18516 // 2. one with the odd values.
18517 // To achieve #2, with need to place the odd values at an even position.
18519 // Place the odd value at an even position (basically, shift all values 1
18520 // step to the left):
18521 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18522 // <a|b|c|d> => <b|undef|d|undef>
18523 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18524 // <e|f|g|h> => <f|undef|h|undef>
18525 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18527 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18529 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18530 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18532 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18533 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18534 // => <2 x i64> <ae|cg>
18535 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
18536 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18537 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18538 // => <2 x i64> <bf|dh>
18539 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
18540 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18542 // Shuffle it back into the right order.
18543 SDValue Highs, Lows;
18544 if (VT == MVT::v8i32) {
18545 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18546 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18547 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18548 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18550 const int HighMask[] = {1, 5, 3, 7};
18551 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18552 const int LowMask[] = {0, 4, 2, 6};
18553 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18556 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18557 // unsigned multiply.
18558 if (IsSigned && !Subtarget->hasSSE41()) {
18560 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
18561 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18562 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18563 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18564 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18566 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18567 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18570 // The first result of MUL_LOHI is actually the low value, followed by the
18572 SDValue Ops[] = {Lows, Highs};
18573 return DAG.getMergeValues(Ops, dl);
18576 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18577 const X86Subtarget *Subtarget) {
18578 MVT VT = Op.getSimpleValueType();
18580 SDValue R = Op.getOperand(0);
18581 SDValue Amt = Op.getOperand(1);
18583 // Optimize shl/srl/sra with constant shift amount.
18584 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18585 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18586 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18588 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
18589 (Subtarget->hasInt256() &&
18590 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
18591 (Subtarget->hasAVX512() &&
18592 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
18593 if (Op.getOpcode() == ISD::SHL)
18594 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
18596 if (Op.getOpcode() == ISD::SRL)
18597 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
18599 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
18600 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
18604 if (VT == MVT::v16i8) {
18605 if (Op.getOpcode() == ISD::SHL) {
18606 // Make a large shift.
18607 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18608 MVT::v8i16, R, ShiftAmt,
18610 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18611 // Zero out the rightmost bits.
18612 SmallVector<SDValue, 16> V(16,
18613 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18615 return DAG.getNode(ISD::AND, dl, VT, SHL,
18616 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18618 if (Op.getOpcode() == ISD::SRL) {
18619 // Make a large shift.
18620 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18621 MVT::v8i16, R, ShiftAmt,
18623 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18624 // Zero out the leftmost bits.
18625 SmallVector<SDValue, 16> V(16,
18626 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18628 return DAG.getNode(ISD::AND, dl, VT, SRL,
18629 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18631 if (Op.getOpcode() == ISD::SRA) {
18632 if (ShiftAmt == 7) {
18633 // R s>> 7 === R s< 0
18634 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18635 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18638 // R s>> a === ((R u>> a) ^ m) - m
18639 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18640 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
18642 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18643 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18644 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18647 llvm_unreachable("Unknown shift opcode.");
18650 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
18651 if (Op.getOpcode() == ISD::SHL) {
18652 // Make a large shift.
18653 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18654 MVT::v16i16, R, ShiftAmt,
18656 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18657 // Zero out the rightmost bits.
18658 SmallVector<SDValue, 32> V(32,
18659 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18661 return DAG.getNode(ISD::AND, dl, VT, SHL,
18662 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18664 if (Op.getOpcode() == ISD::SRL) {
18665 // Make a large shift.
18666 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18667 MVT::v16i16, R, ShiftAmt,
18669 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18670 // Zero out the leftmost bits.
18671 SmallVector<SDValue, 32> V(32,
18672 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18674 return DAG.getNode(ISD::AND, dl, VT, SRL,
18675 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18677 if (Op.getOpcode() == ISD::SRA) {
18678 if (ShiftAmt == 7) {
18679 // R s>> 7 === R s< 0
18680 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18681 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18684 // R s>> a === ((R u>> a) ^ m) - m
18685 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18686 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
18688 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18689 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18690 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18693 llvm_unreachable("Unknown shift opcode.");
18698 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18699 if (!Subtarget->is64Bit() &&
18700 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18701 Amt.getOpcode() == ISD::BITCAST &&
18702 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18703 Amt = Amt.getOperand(0);
18704 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18705 VT.getVectorNumElements();
18706 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18707 uint64_t ShiftAmt = 0;
18708 for (unsigned i = 0; i != Ratio; ++i) {
18709 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
18713 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18715 // Check remaining shift amounts.
18716 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18717 uint64_t ShAmt = 0;
18718 for (unsigned j = 0; j != Ratio; ++j) {
18719 ConstantSDNode *C =
18720 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18724 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18726 if (ShAmt != ShiftAmt)
18729 switch (Op.getOpcode()) {
18731 llvm_unreachable("Unknown shift opcode!");
18733 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
18736 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
18739 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
18747 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18748 const X86Subtarget* Subtarget) {
18749 MVT VT = Op.getSimpleValueType();
18751 SDValue R = Op.getOperand(0);
18752 SDValue Amt = Op.getOperand(1);
18754 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
18755 VT == MVT::v4i32 || VT == MVT::v8i16 ||
18756 (Subtarget->hasInt256() &&
18757 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
18758 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
18759 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
18761 EVT EltVT = VT.getVectorElementType();
18763 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18764 // Check if this build_vector node is doing a splat.
18765 // If so, then set BaseShAmt equal to the splat value.
18766 BaseShAmt = BV->getSplatValue();
18767 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18768 BaseShAmt = SDValue();
18770 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18771 Amt = Amt.getOperand(0);
18773 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18774 if (SVN && SVN->isSplat()) {
18775 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18776 SDValue InVec = Amt.getOperand(0);
18777 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18778 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18779 "Unexpected shuffle index found!");
18780 BaseShAmt = InVec.getOperand(SplatIdx);
18781 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18782 if (ConstantSDNode *C =
18783 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18784 if (C->getZExtValue() == SplatIdx)
18785 BaseShAmt = InVec.getOperand(1);
18790 // Avoid introducing an extract element from a shuffle.
18791 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18792 DAG.getIntPtrConstant(SplatIdx));
18796 if (BaseShAmt.getNode()) {
18797 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18798 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18799 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18800 else if (EltVT.bitsLT(MVT::i32))
18801 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18803 switch (Op.getOpcode()) {
18805 llvm_unreachable("Unknown shift opcode!");
18807 switch (VT.SimpleTy) {
18808 default: return SDValue();
18817 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
18820 switch (VT.SimpleTy) {
18821 default: return SDValue();
18828 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
18831 switch (VT.SimpleTy) {
18832 default: return SDValue();
18841 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
18847 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18848 if (!Subtarget->is64Bit() &&
18849 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
18850 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
18851 Amt.getOpcode() == ISD::BITCAST &&
18852 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18853 Amt = Amt.getOperand(0);
18854 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18855 VT.getVectorNumElements();
18856 std::vector<SDValue> Vals(Ratio);
18857 for (unsigned i = 0; i != Ratio; ++i)
18858 Vals[i] = Amt.getOperand(i);
18859 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18860 for (unsigned j = 0; j != Ratio; ++j)
18861 if (Vals[j] != Amt.getOperand(i + j))
18864 switch (Op.getOpcode()) {
18866 llvm_unreachable("Unknown shift opcode!");
18868 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
18870 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
18872 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
18879 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18880 SelectionDAG &DAG) {
18881 MVT VT = Op.getSimpleValueType();
18883 SDValue R = Op.getOperand(0);
18884 SDValue Amt = Op.getOperand(1);
18887 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18888 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18890 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
18894 V = LowerScalarVariableShift(Op, DAG, Subtarget);
18898 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
18900 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
18901 if (Subtarget->hasInt256()) {
18902 if (Op.getOpcode() == ISD::SRL &&
18903 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18904 VT == MVT::v4i64 || VT == MVT::v8i32))
18906 if (Op.getOpcode() == ISD::SHL &&
18907 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18908 VT == MVT::v4i64 || VT == MVT::v8i32))
18910 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
18914 // If possible, lower this packed shift into a vector multiply instead of
18915 // expanding it into a sequence of scalar shifts.
18916 // Do this only if the vector shift count is a constant build_vector.
18917 if (Op.getOpcode() == ISD::SHL &&
18918 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18919 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18920 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18921 SmallVector<SDValue, 8> Elts;
18922 EVT SVT = VT.getScalarType();
18923 unsigned SVTBits = SVT.getSizeInBits();
18924 const APInt &One = APInt(SVTBits, 1);
18925 unsigned NumElems = VT.getVectorNumElements();
18927 for (unsigned i=0; i !=NumElems; ++i) {
18928 SDValue Op = Amt->getOperand(i);
18929 if (Op->getOpcode() == ISD::UNDEF) {
18930 Elts.push_back(Op);
18934 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18935 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18936 uint64_t ShAmt = C.getZExtValue();
18937 if (ShAmt >= SVTBits) {
18938 Elts.push_back(DAG.getUNDEF(SVT));
18941 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
18943 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18944 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18947 // Lower SHL with variable shift amount.
18948 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18949 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
18951 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
18952 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
18953 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18954 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18957 // If possible, lower this shift as a sequence of two shifts by
18958 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18960 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18962 // Could be rewritten as:
18963 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18965 // The advantage is that the two shifts from the example would be
18966 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18967 // the vector shift into four scalar shifts plus four pairs of vector
18969 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18970 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18971 unsigned TargetOpcode = X86ISD::MOVSS;
18972 bool CanBeSimplified;
18973 // The splat value for the first packed shift (the 'X' from the example).
18974 SDValue Amt1 = Amt->getOperand(0);
18975 // The splat value for the second packed shift (the 'Y' from the example).
18976 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18977 Amt->getOperand(2);
18979 // See if it is possible to replace this node with a sequence of
18980 // two shifts followed by a MOVSS/MOVSD
18981 if (VT == MVT::v4i32) {
18982 // Check if it is legal to use a MOVSS.
18983 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18984 Amt2 == Amt->getOperand(3);
18985 if (!CanBeSimplified) {
18986 // Otherwise, check if we can still simplify this node using a MOVSD.
18987 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18988 Amt->getOperand(2) == Amt->getOperand(3);
18989 TargetOpcode = X86ISD::MOVSD;
18990 Amt2 = Amt->getOperand(2);
18993 // Do similar checks for the case where the machine value type
18995 CanBeSimplified = Amt1 == Amt->getOperand(1);
18996 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18997 CanBeSimplified = Amt2 == Amt->getOperand(i);
18999 if (!CanBeSimplified) {
19000 TargetOpcode = X86ISD::MOVSD;
19001 CanBeSimplified = true;
19002 Amt2 = Amt->getOperand(4);
19003 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
19004 CanBeSimplified = Amt1 == Amt->getOperand(i);
19005 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
19006 CanBeSimplified = Amt2 == Amt->getOperand(j);
19010 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
19011 isa<ConstantSDNode>(Amt2)) {
19012 // Replace this node with two shifts followed by a MOVSS/MOVSD.
19013 EVT CastVT = MVT::v4i32;
19015 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
19016 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
19018 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
19019 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
19020 if (TargetOpcode == X86ISD::MOVSD)
19021 CastVT = MVT::v2i64;
19022 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
19023 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
19024 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
19026 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
19030 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
19031 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
19034 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
19035 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
19037 // Turn 'a' into a mask suitable for VSELECT
19038 SDValue VSelM = DAG.getConstant(0x80, VT);
19039 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19040 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19042 SDValue CM1 = DAG.getConstant(0x0f, VT);
19043 SDValue CM2 = DAG.getConstant(0x3f, VT);
19045 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
19046 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
19047 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
19048 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
19049 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
19052 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
19053 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19054 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19056 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
19057 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
19058 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
19059 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
19060 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
19063 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
19064 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
19065 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
19067 // return VSELECT(r, r+r, a);
19068 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
19069 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
19073 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
19074 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
19075 // solution better.
19076 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
19077 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
19079 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
19080 R = DAG.getNode(ExtOpc, dl, NewVT, R);
19081 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
19082 return DAG.getNode(ISD::TRUNCATE, dl, VT,
19083 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
19086 // Decompose 256-bit shifts into smaller 128-bit shifts.
19087 if (VT.is256BitVector()) {
19088 unsigned NumElems = VT.getVectorNumElements();
19089 MVT EltVT = VT.getVectorElementType();
19090 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19092 // Extract the two vectors
19093 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
19094 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
19096 // Recreate the shift amount vectors
19097 SDValue Amt1, Amt2;
19098 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
19099 // Constant shift amount
19100 SmallVector<SDValue, 4> Amt1Csts;
19101 SmallVector<SDValue, 4> Amt2Csts;
19102 for (unsigned i = 0; i != NumElems/2; ++i)
19103 Amt1Csts.push_back(Amt->getOperand(i));
19104 for (unsigned i = NumElems/2; i != NumElems; ++i)
19105 Amt2Csts.push_back(Amt->getOperand(i));
19107 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
19108 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
19110 // Variable shift amount
19111 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
19112 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
19115 // Issue new vector shifts for the smaller types
19116 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
19117 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
19119 // Concatenate the result back
19120 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
19126 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
19127 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
19128 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
19129 // looks for this combo and may remove the "setcc" instruction if the "setcc"
19130 // has only one use.
19131 SDNode *N = Op.getNode();
19132 SDValue LHS = N->getOperand(0);
19133 SDValue RHS = N->getOperand(1);
19134 unsigned BaseOp = 0;
19137 switch (Op.getOpcode()) {
19138 default: llvm_unreachable("Unknown ovf instruction!");
19140 // A subtract of one will be selected as a INC. Note that INC doesn't
19141 // set CF, so we can't do this for UADDO.
19142 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
19144 BaseOp = X86ISD::INC;
19145 Cond = X86::COND_O;
19148 BaseOp = X86ISD::ADD;
19149 Cond = X86::COND_O;
19152 BaseOp = X86ISD::ADD;
19153 Cond = X86::COND_B;
19156 // A subtract of one will be selected as a DEC. Note that DEC doesn't
19157 // set CF, so we can't do this for USUBO.
19158 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
19160 BaseOp = X86ISD::DEC;
19161 Cond = X86::COND_O;
19164 BaseOp = X86ISD::SUB;
19165 Cond = X86::COND_O;
19168 BaseOp = X86ISD::SUB;
19169 Cond = X86::COND_B;
19172 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
19173 Cond = X86::COND_O;
19175 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
19176 if (N->getValueType(0) == MVT::i8) {
19177 BaseOp = X86ISD::UMUL8;
19178 Cond = X86::COND_O;
19181 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
19183 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
19186 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
19187 DAG.getConstant(X86::COND_O, MVT::i32),
19188 SDValue(Sum.getNode(), 2));
19190 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19194 // Also sets EFLAGS.
19195 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
19196 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
19199 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
19200 DAG.getConstant(Cond, MVT::i32),
19201 SDValue(Sum.getNode(), 1));
19203 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
19206 // Sign extension of the low part of vector elements. This may be used either
19207 // when sign extend instructions are not available or if the vector element
19208 // sizes already match the sign-extended size. If the vector elements are in
19209 // their pre-extended size and sign extend instructions are available, that will
19210 // be handled by LowerSIGN_EXTEND.
19211 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
19212 SelectionDAG &DAG) const {
19214 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
19215 MVT VT = Op.getSimpleValueType();
19217 if (!Subtarget->hasSSE2() || !VT.isVector())
19220 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
19221 ExtraVT.getScalarType().getSizeInBits();
19223 switch (VT.SimpleTy) {
19224 default: return SDValue();
19227 if (!Subtarget->hasFp256())
19229 if (!Subtarget->hasInt256()) {
19230 // needs to be split
19231 unsigned NumElems = VT.getVectorNumElements();
19233 // Extract the LHS vectors
19234 SDValue LHS = Op.getOperand(0);
19235 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
19236 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
19238 MVT EltVT = VT.getVectorElementType();
19239 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19241 EVT ExtraEltVT = ExtraVT.getVectorElementType();
19242 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
19243 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
19245 SDValue Extra = DAG.getValueType(ExtraVT);
19247 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
19248 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
19250 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
19255 SDValue Op0 = Op.getOperand(0);
19257 // This is a sign extension of some low part of vector elements without
19258 // changing the size of the vector elements themselves:
19259 // Shift-Left + Shift-Right-Algebraic.
19260 SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
19262 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
19268 /// Returns true if the operand type is exactly twice the native width, and
19269 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19270 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19271 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19272 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
19273 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19276 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19277 else if (OpWidth == 128)
19278 return Subtarget->hasCmpxchg16b();
19283 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19284 return needsCmpXchgNb(SI->getValueOperand()->getType());
19287 // Note: this turns large loads into lock cmpxchg8b/16b.
19288 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19289 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19290 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19291 return needsCmpXchgNb(PTy->getElementType());
19294 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19295 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19296 const Type *MemType = AI->getType();
19298 // If the operand is too big, we must see if cmpxchg8/16b is available
19299 // and default to library calls otherwise.
19300 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19301 return needsCmpXchgNb(MemType);
19303 AtomicRMWInst::BinOp Op = AI->getOperation();
19306 llvm_unreachable("Unknown atomic operation");
19307 case AtomicRMWInst::Xchg:
19308 case AtomicRMWInst::Add:
19309 case AtomicRMWInst::Sub:
19310 // It's better to use xadd, xsub or xchg for these in all cases.
19312 case AtomicRMWInst::Or:
19313 case AtomicRMWInst::And:
19314 case AtomicRMWInst::Xor:
19315 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19316 // prefix to a normal instruction for these operations.
19317 return !AI->use_empty();
19318 case AtomicRMWInst::Nand:
19319 case AtomicRMWInst::Max:
19320 case AtomicRMWInst::Min:
19321 case AtomicRMWInst::UMax:
19322 case AtomicRMWInst::UMin:
19323 // These always require a non-trivial set of data operations on x86. We must
19324 // use a cmpxchg loop.
19329 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19330 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19331 // no-sse2). There isn't any reason to disable it if the target processor
19333 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19337 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19338 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19339 const Type *MemType = AI->getType();
19340 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19341 // there is no benefit in turning such RMWs into loads, and it is actually
19342 // harmful as it introduces a mfence.
19343 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19346 auto Builder = IRBuilder<>(AI);
19347 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19348 auto SynchScope = AI->getSynchScope();
19349 // We must restrict the ordering to avoid generating loads with Release or
19350 // ReleaseAcquire orderings.
19351 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19352 auto Ptr = AI->getPointerOperand();
19354 // Before the load we need a fence. Here is an example lifted from
19355 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19358 // x.store(1, relaxed);
19359 // r1 = y.fetch_add(0, release);
19361 // y.fetch_add(42, acquire);
19362 // r2 = x.load(relaxed);
19363 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19364 // lowered to just a load without a fence. A mfence flushes the store buffer,
19365 // making the optimization clearly correct.
19366 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19367 // otherwise, we might be able to be more agressive on relaxed idempotent
19368 // rmw. In practice, they do not look useful, so we don't try to be
19369 // especially clever.
19370 if (SynchScope == SingleThread) {
19371 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19372 // the IR level, so we must wrap it in an intrinsic.
19374 } else if (hasMFENCE(*Subtarget)) {
19375 Function *MFence = llvm::Intrinsic::getDeclaration(M,
19376 Intrinsic::x86_sse2_mfence);
19377 Builder.CreateCall(MFence);
19379 // FIXME: it might make sense to use a locked operation here but on a
19380 // different cache-line to prevent cache-line bouncing. In practice it
19381 // is probably a small win, and x86 processors without mfence are rare
19382 // enough that we do not bother.
19386 // Finally we can emit the atomic load.
19387 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19388 AI->getType()->getPrimitiveSizeInBits());
19389 Loaded->setAtomic(Order, SynchScope);
19390 AI->replaceAllUsesWith(Loaded);
19391 AI->eraseFromParent();
19395 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19396 SelectionDAG &DAG) {
19398 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19399 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19400 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19401 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19403 // The only fence that needs an instruction is a sequentially-consistent
19404 // cross-thread fence.
19405 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19406 if (hasMFENCE(*Subtarget))
19407 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19409 SDValue Chain = Op.getOperand(0);
19410 SDValue Zero = DAG.getConstant(0, MVT::i32);
19412 DAG.getRegister(X86::ESP, MVT::i32), // Base
19413 DAG.getTargetConstant(1, MVT::i8), // Scale
19414 DAG.getRegister(0, MVT::i32), // Index
19415 DAG.getTargetConstant(0, MVT::i32), // Disp
19416 DAG.getRegister(0, MVT::i32), // Segment.
19420 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19421 return SDValue(Res, 0);
19424 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19425 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19428 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19429 SelectionDAG &DAG) {
19430 MVT T = Op.getSimpleValueType();
19434 switch(T.SimpleTy) {
19435 default: llvm_unreachable("Invalid value type!");
19436 case MVT::i8: Reg = X86::AL; size = 1; break;
19437 case MVT::i16: Reg = X86::AX; size = 2; break;
19438 case MVT::i32: Reg = X86::EAX; size = 4; break;
19440 assert(Subtarget->is64Bit() && "Node not type legal!");
19441 Reg = X86::RAX; size = 8;
19444 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19445 Op.getOperand(2), SDValue());
19446 SDValue Ops[] = { cpIn.getValue(0),
19449 DAG.getTargetConstant(size, MVT::i8),
19450 cpIn.getValue(1) };
19451 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19452 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19453 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19457 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19458 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19459 MVT::i32, cpOut.getValue(2));
19460 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19461 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
19463 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19464 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19465 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19469 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19470 SelectionDAG &DAG) {
19471 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19472 MVT DstVT = Op.getSimpleValueType();
19474 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19475 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19476 if (DstVT != MVT::f64)
19477 // This conversion needs to be expanded.
19480 SDValue InVec = Op->getOperand(0);
19482 unsigned NumElts = SrcVT.getVectorNumElements();
19483 EVT SVT = SrcVT.getVectorElementType();
19485 // Widen the vector in input in the case of MVT::v2i32.
19486 // Example: from MVT::v2i32 to MVT::v4i32.
19487 SmallVector<SDValue, 16> Elts;
19488 for (unsigned i = 0, e = NumElts; i != e; ++i)
19489 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19490 DAG.getIntPtrConstant(i)));
19492 // Explicitly mark the extra elements as Undef.
19493 SDValue Undef = DAG.getUNDEF(SVT);
19494 for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i)
19495 Elts.push_back(Undef);
19497 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19498 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19499 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
19500 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19501 DAG.getIntPtrConstant(0));
19504 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19505 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19506 assert((DstVT == MVT::i64 ||
19507 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19508 "Unexpected custom BITCAST");
19509 // i64 <=> MMX conversions are Legal.
19510 if (SrcVT==MVT::i64 && DstVT.isVector())
19512 if (DstVT==MVT::i64 && SrcVT.isVector())
19514 // MMX <=> MMX conversions are Legal.
19515 if (SrcVT.isVector() && DstVT.isVector())
19517 // All other conversions need to be expanded.
19521 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19522 SelectionDAG &DAG) {
19523 SDNode *Node = Op.getNode();
19526 Op = Op.getOperand(0);
19527 EVT VT = Op.getValueType();
19528 assert((VT.is128BitVector() || VT.is256BitVector()) &&
19529 "CTPOP lowering only implemented for 128/256-bit wide vector types");
19531 unsigned NumElts = VT.getVectorNumElements();
19532 EVT EltVT = VT.getVectorElementType();
19533 unsigned Len = EltVT.getSizeInBits();
19535 // This is the vectorized version of the "best" algorithm from
19536 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19537 // with a minor tweak to use a series of adds + shifts instead of vector
19538 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
19540 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
19541 // v8i32 => Always profitable
19543 // FIXME: There a couple of possible improvements:
19545 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
19546 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
19548 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
19549 "CTPOP not implemented for this vector element type.");
19551 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
19552 // extra legalization.
19553 bool NeedsBitcast = EltVT == MVT::i32;
19554 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
19556 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
19557 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
19558 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
19560 // v = v - ((v >> 1) & 0x55555555...)
19561 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
19562 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
19563 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
19565 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19567 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
19568 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
19570 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
19572 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
19573 if (VT != And.getValueType())
19574 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19575 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
19577 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19578 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
19579 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
19580 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
19581 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
19583 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
19584 if (NeedsBitcast) {
19585 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19586 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
19587 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
19590 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
19591 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
19592 if (VT != AndRHS.getValueType()) {
19593 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
19594 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
19596 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
19598 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19599 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
19600 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
19601 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
19602 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19604 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
19605 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
19606 if (NeedsBitcast) {
19607 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19608 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
19610 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
19611 if (VT != And.getValueType())
19612 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19614 // The algorithm mentioned above uses:
19615 // v = (v * 0x01010101...) >> (Len - 8)
19617 // Change it to use vector adds + vector shifts which yield faster results on
19618 // Haswell than using vector integer multiplication.
19620 // For i32 elements:
19621 // v = v + (v >> 8)
19622 // v = v + (v >> 16)
19624 // For i64 elements:
19625 // v = v + (v >> 8)
19626 // v = v + (v >> 16)
19627 // v = v + (v >> 32)
19630 SmallVector<SDValue, 8> Csts;
19631 for (unsigned i = 8; i <= Len/2; i *= 2) {
19632 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
19633 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
19634 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
19635 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19639 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
19640 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
19641 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
19642 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
19643 if (NeedsBitcast) {
19644 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19645 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
19647 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
19648 if (VT != And.getValueType())
19649 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19654 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19655 SDNode *Node = Op.getNode();
19657 EVT T = Node->getValueType(0);
19658 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19659 DAG.getConstant(0, T), Node->getOperand(2));
19660 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19661 cast<AtomicSDNode>(Node)->getMemoryVT(),
19662 Node->getOperand(0),
19663 Node->getOperand(1), negOp,
19664 cast<AtomicSDNode>(Node)->getMemOperand(),
19665 cast<AtomicSDNode>(Node)->getOrdering(),
19666 cast<AtomicSDNode>(Node)->getSynchScope());
19669 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19670 SDNode *Node = Op.getNode();
19672 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19674 // Convert seq_cst store -> xchg
19675 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19676 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19677 // (The only way to get a 16-byte store is cmpxchg16b)
19678 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19679 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19680 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19681 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19682 cast<AtomicSDNode>(Node)->getMemoryVT(),
19683 Node->getOperand(0),
19684 Node->getOperand(1), Node->getOperand(2),
19685 cast<AtomicSDNode>(Node)->getMemOperand(),
19686 cast<AtomicSDNode>(Node)->getOrdering(),
19687 cast<AtomicSDNode>(Node)->getSynchScope());
19688 return Swap.getValue(1);
19690 // Other atomic stores have a simple pattern.
19694 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19695 EVT VT = Op.getNode()->getSimpleValueType(0);
19697 // Let legalize expand this if it isn't a legal type yet.
19698 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19701 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19704 bool ExtraOp = false;
19705 switch (Op.getOpcode()) {
19706 default: llvm_unreachable("Invalid code");
19707 case ISD::ADDC: Opc = X86ISD::ADD; break;
19708 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19709 case ISD::SUBC: Opc = X86ISD::SUB; break;
19710 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19714 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19716 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19717 Op.getOperand(1), Op.getOperand(2));
19720 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19721 SelectionDAG &DAG) {
19722 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19724 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19725 // which returns the values as { float, float } (in XMM0) or
19726 // { double, double } (which is returned in XMM0, XMM1).
19728 SDValue Arg = Op.getOperand(0);
19729 EVT ArgVT = Arg.getValueType();
19730 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19732 TargetLowering::ArgListTy Args;
19733 TargetLowering::ArgListEntry Entry;
19737 Entry.isSExt = false;
19738 Entry.isZExt = false;
19739 Args.push_back(Entry);
19741 bool isF64 = ArgVT == MVT::f64;
19742 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19743 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19744 // the results are returned via SRet in memory.
19745 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19746 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19747 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
19749 Type *RetTy = isF64
19750 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19751 : (Type*)VectorType::get(ArgTy, 4);
19753 TargetLowering::CallLoweringInfo CLI(DAG);
19754 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19755 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19757 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19760 // Returned in xmm0 and xmm1.
19761 return CallResult.first;
19763 // Returned in bits 0:31 and 32:64 xmm0.
19764 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19765 CallResult.first, DAG.getIntPtrConstant(0));
19766 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19767 CallResult.first, DAG.getIntPtrConstant(1));
19768 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19769 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19772 /// LowerOperation - Provide custom lowering hooks for some operations.
19774 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19775 switch (Op.getOpcode()) {
19776 default: llvm_unreachable("Should not custom lower this!");
19777 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
19778 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19779 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19780 return LowerCMP_SWAP(Op, Subtarget, DAG);
19781 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19782 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19783 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19784 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19785 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
19786 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
19787 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19788 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19789 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19790 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19791 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19792 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19793 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19794 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19795 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19796 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19797 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19798 case ISD::SHL_PARTS:
19799 case ISD::SRA_PARTS:
19800 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19801 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19802 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19803 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19804 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19805 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19806 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19807 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19808 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19809 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19810 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19812 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19813 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19814 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19815 case ISD::SETCC: return LowerSETCC(Op, DAG);
19816 case ISD::SELECT: return LowerSELECT(Op, DAG);
19817 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19818 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19819 case ISD::VASTART: return LowerVASTART(Op, DAG);
19820 case ISD::VAARG: return LowerVAARG(Op, DAG);
19821 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19822 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19823 case ISD::INTRINSIC_VOID:
19824 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19825 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19826 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19827 case ISD::FRAME_TO_ARGS_OFFSET:
19828 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19829 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19830 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19831 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19832 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19833 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19834 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19835 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19836 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
19837 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
19838 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
19839 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19840 case ISD::UMUL_LOHI:
19841 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19844 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19850 case ISD::UMULO: return LowerXALUO(Op, DAG);
19851 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19852 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19856 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19857 case ISD::ADD: return LowerADD(Op, DAG);
19858 case ISD::SUB: return LowerSUB(Op, DAG);
19859 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19863 /// ReplaceNodeResults - Replace a node with an illegal result type
19864 /// with a new node built out of custom code.
19865 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19866 SmallVectorImpl<SDValue>&Results,
19867 SelectionDAG &DAG) const {
19869 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19870 switch (N->getOpcode()) {
19872 llvm_unreachable("Do not know how to custom type legalize this operation!");
19873 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19874 case X86ISD::FMINC:
19876 case X86ISD::FMAXC:
19877 case X86ISD::FMAX: {
19878 EVT VT = N->getValueType(0);
19879 if (VT != MVT::v2f32)
19880 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19881 SDValue UNDEF = DAG.getUNDEF(VT);
19882 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19883 N->getOperand(0), UNDEF);
19884 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19885 N->getOperand(1), UNDEF);
19886 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19889 case ISD::SIGN_EXTEND_INREG:
19894 // We don't want to expand or promote these.
19901 case ISD::UDIVREM: {
19902 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19903 Results.push_back(V);
19906 case ISD::FP_TO_SINT:
19907 case ISD::FP_TO_UINT: {
19908 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19910 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
19913 std::pair<SDValue,SDValue> Vals =
19914 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19915 SDValue FIST = Vals.first, StackSlot = Vals.second;
19916 if (FIST.getNode()) {
19917 EVT VT = N->getValueType(0);
19918 // Return a load from the stack slot.
19919 if (StackSlot.getNode())
19920 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19921 MachinePointerInfo(),
19922 false, false, false, 0));
19924 Results.push_back(FIST);
19928 case ISD::UINT_TO_FP: {
19929 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19930 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19931 N->getValueType(0) != MVT::v2f32)
19933 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19935 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
19937 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19938 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19939 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
19940 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
19941 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19942 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19945 case ISD::FP_ROUND: {
19946 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19948 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19949 Results.push_back(V);
19952 case ISD::INTRINSIC_W_CHAIN: {
19953 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19955 default : llvm_unreachable("Do not know how to custom type "
19956 "legalize this intrinsic operation!");
19957 case Intrinsic::x86_rdtsc:
19958 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19960 case Intrinsic::x86_rdtscp:
19961 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19963 case Intrinsic::x86_rdpmc:
19964 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19967 case ISD::READCYCLECOUNTER: {
19968 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19971 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19972 EVT T = N->getValueType(0);
19973 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19974 bool Regs64bit = T == MVT::i128;
19975 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19976 SDValue cpInL, cpInH;
19977 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19978 DAG.getConstant(0, HalfT));
19979 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19980 DAG.getConstant(1, HalfT));
19981 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19982 Regs64bit ? X86::RAX : X86::EAX,
19984 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19985 Regs64bit ? X86::RDX : X86::EDX,
19986 cpInH, cpInL.getValue(1));
19987 SDValue swapInL, swapInH;
19988 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19989 DAG.getConstant(0, HalfT));
19990 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19991 DAG.getConstant(1, HalfT));
19992 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19993 Regs64bit ? X86::RBX : X86::EBX,
19994 swapInL, cpInH.getValue(1));
19995 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19996 Regs64bit ? X86::RCX : X86::ECX,
19997 swapInH, swapInL.getValue(1));
19998 SDValue Ops[] = { swapInH.getValue(0),
20000 swapInH.getValue(1) };
20001 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
20002 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
20003 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
20004 X86ISD::LCMPXCHG8_DAG;
20005 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
20006 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
20007 Regs64bit ? X86::RAX : X86::EAX,
20008 HalfT, Result.getValue(1));
20009 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
20010 Regs64bit ? X86::RDX : X86::EDX,
20011 HalfT, cpOutL.getValue(2));
20012 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
20014 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
20015 MVT::i32, cpOutH.getValue(2));
20017 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
20018 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
20019 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
20021 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
20022 Results.push_back(Success);
20023 Results.push_back(EFLAGS.getValue(1));
20026 case ISD::ATOMIC_SWAP:
20027 case ISD::ATOMIC_LOAD_ADD:
20028 case ISD::ATOMIC_LOAD_SUB:
20029 case ISD::ATOMIC_LOAD_AND:
20030 case ISD::ATOMIC_LOAD_OR:
20031 case ISD::ATOMIC_LOAD_XOR:
20032 case ISD::ATOMIC_LOAD_NAND:
20033 case ISD::ATOMIC_LOAD_MIN:
20034 case ISD::ATOMIC_LOAD_MAX:
20035 case ISD::ATOMIC_LOAD_UMIN:
20036 case ISD::ATOMIC_LOAD_UMAX:
20037 case ISD::ATOMIC_LOAD: {
20038 // Delegate to generic TypeLegalization. Situations we can really handle
20039 // should have already been dealt with by AtomicExpandPass.cpp.
20042 case ISD::BITCAST: {
20043 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
20044 EVT DstVT = N->getValueType(0);
20045 EVT SrcVT = N->getOperand(0)->getValueType(0);
20047 if (SrcVT != MVT::f64 ||
20048 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
20051 unsigned NumElts = DstVT.getVectorNumElements();
20052 EVT SVT = DstVT.getVectorElementType();
20053 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
20054 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
20055 MVT::v2f64, N->getOperand(0));
20056 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
20058 if (ExperimentalVectorWideningLegalization) {
20059 // If we are legalizing vectors by widening, we already have the desired
20060 // legal vector type, just return it.
20061 Results.push_back(ToVecInt);
20065 SmallVector<SDValue, 8> Elts;
20066 for (unsigned i = 0, e = NumElts; i != e; ++i)
20067 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
20068 ToVecInt, DAG.getIntPtrConstant(i)));
20070 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
20075 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
20077 default: return nullptr;
20078 case X86ISD::BSF: return "X86ISD::BSF";
20079 case X86ISD::BSR: return "X86ISD::BSR";
20080 case X86ISD::SHLD: return "X86ISD::SHLD";
20081 case X86ISD::SHRD: return "X86ISD::SHRD";
20082 case X86ISD::FAND: return "X86ISD::FAND";
20083 case X86ISD::FANDN: return "X86ISD::FANDN";
20084 case X86ISD::FOR: return "X86ISD::FOR";
20085 case X86ISD::FXOR: return "X86ISD::FXOR";
20086 case X86ISD::FSRL: return "X86ISD::FSRL";
20087 case X86ISD::FILD: return "X86ISD::FILD";
20088 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
20089 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
20090 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
20091 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
20092 case X86ISD::FLD: return "X86ISD::FLD";
20093 case X86ISD::FST: return "X86ISD::FST";
20094 case X86ISD::CALL: return "X86ISD::CALL";
20095 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20096 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20097 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20098 case X86ISD::BT: return "X86ISD::BT";
20099 case X86ISD::CMP: return "X86ISD::CMP";
20100 case X86ISD::COMI: return "X86ISD::COMI";
20101 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20102 case X86ISD::CMPM: return "X86ISD::CMPM";
20103 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20104 case X86ISD::SETCC: return "X86ISD::SETCC";
20105 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20106 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20107 case X86ISD::CMOV: return "X86ISD::CMOV";
20108 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20109 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20110 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20111 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20112 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20113 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20114 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20115 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20116 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20117 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20118 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20119 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20120 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20121 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20122 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20123 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20124 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20125 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20126 case X86ISD::HADD: return "X86ISD::HADD";
20127 case X86ISD::HSUB: return "X86ISD::HSUB";
20128 case X86ISD::FHADD: return "X86ISD::FHADD";
20129 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20130 case X86ISD::UMAX: return "X86ISD::UMAX";
20131 case X86ISD::UMIN: return "X86ISD::UMIN";
20132 case X86ISD::SMAX: return "X86ISD::SMAX";
20133 case X86ISD::SMIN: return "X86ISD::SMIN";
20134 case X86ISD::FMAX: return "X86ISD::FMAX";
20135 case X86ISD::FMIN: return "X86ISD::FMIN";
20136 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20137 case X86ISD::FMINC: return "X86ISD::FMINC";
20138 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20139 case X86ISD::FRCP: return "X86ISD::FRCP";
20140 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20141 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20142 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20143 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20144 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20145 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20146 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20147 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20148 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20149 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20150 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20151 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20152 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20153 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20154 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20155 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20156 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20157 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
20158 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20159 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20160 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20161 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20162 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20163 case X86ISD::VSHL: return "X86ISD::VSHL";
20164 case X86ISD::VSRL: return "X86ISD::VSRL";
20165 case X86ISD::VSRA: return "X86ISD::VSRA";
20166 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20167 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20168 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20169 case X86ISD::CMPP: return "X86ISD::CMPP";
20170 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20171 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20172 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20173 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20174 case X86ISD::ADD: return "X86ISD::ADD";
20175 case X86ISD::SUB: return "X86ISD::SUB";
20176 case X86ISD::ADC: return "X86ISD::ADC";
20177 case X86ISD::SBB: return "X86ISD::SBB";
20178 case X86ISD::SMUL: return "X86ISD::SMUL";
20179 case X86ISD::UMUL: return "X86ISD::UMUL";
20180 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20181 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20182 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20183 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20184 case X86ISD::INC: return "X86ISD::INC";
20185 case X86ISD::DEC: return "X86ISD::DEC";
20186 case X86ISD::OR: return "X86ISD::OR";
20187 case X86ISD::XOR: return "X86ISD::XOR";
20188 case X86ISD::AND: return "X86ISD::AND";
20189 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20190 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20191 case X86ISD::PTEST: return "X86ISD::PTEST";
20192 case X86ISD::TESTP: return "X86ISD::TESTP";
20193 case X86ISD::TESTM: return "X86ISD::TESTM";
20194 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20195 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20196 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20197 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20198 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20199 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20200 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20201 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20202 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20203 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20204 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20205 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20206 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20207 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20208 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20209 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20210 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20211 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20212 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20213 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20214 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20215 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20216 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20217 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
20218 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20219 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20220 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20221 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20222 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20223 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20224 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20225 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20226 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20227 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20228 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20229 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20230 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20231 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20232 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
20233 case X86ISD::SAHF: return "X86ISD::SAHF";
20234 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20235 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20236 case X86ISD::FMADD: return "X86ISD::FMADD";
20237 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20238 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20239 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20240 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20241 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20242 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20243 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20244 case X86ISD::XTEST: return "X86ISD::XTEST";
20245 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20246 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20247 case X86ISD::SELECT: return "X86ISD::SELECT";
20248 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20249 case X86ISD::RCP28: return "X86ISD::RCP28";
20250 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20254 // isLegalAddressingMode - Return true if the addressing mode represented
20255 // by AM is legal for this target, for a load/store of the specified type.
20256 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
20258 // X86 supports extremely general addressing modes.
20259 CodeModel::Model M = getTargetMachine().getCodeModel();
20260 Reloc::Model R = getTargetMachine().getRelocationModel();
20262 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20263 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20268 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20270 // If a reference to this global requires an extra load, we can't fold it.
20271 if (isGlobalStubReference(GVFlags))
20274 // If BaseGV requires a register for the PIC base, we cannot also have a
20275 // BaseReg specified.
20276 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20279 // If lower 4G is not available, then we must use rip-relative addressing.
20280 if ((M != CodeModel::Small || R != Reloc::Static) &&
20281 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20285 switch (AM.Scale) {
20291 // These scales always work.
20296 // These scales are formed with basereg+scalereg. Only accept if there is
20301 default: // Other stuff never works.
20308 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20309 unsigned Bits = Ty->getScalarSizeInBits();
20311 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20312 // particularly cheaper than those without.
20316 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20317 // variable shifts just as cheap as scalar ones.
20318 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20321 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20322 // fully general vector.
20326 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20327 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20329 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20330 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20331 return NumBits1 > NumBits2;
20334 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20335 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20338 if (!isTypeLegal(EVT::getEVT(Ty1)))
20341 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20343 // Assuming the caller doesn't have a zeroext or signext return parameter,
20344 // truncation all the way down to i1 is valid.
20348 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20349 return isInt<32>(Imm);
20352 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20353 // Can also use sub to handle negated immediates.
20354 return isInt<32>(Imm);
20357 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20358 if (!VT1.isInteger() || !VT2.isInteger())
20360 unsigned NumBits1 = VT1.getSizeInBits();
20361 unsigned NumBits2 = VT2.getSizeInBits();
20362 return NumBits1 > NumBits2;
20365 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20366 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20367 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20370 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20371 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20372 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20375 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20376 EVT VT1 = Val.getValueType();
20377 if (isZExtFree(VT1, VT2))
20380 if (Val.getOpcode() != ISD::LOAD)
20383 if (!VT1.isSimple() || !VT1.isInteger() ||
20384 !VT2.isSimple() || !VT2.isInteger())
20387 switch (VT1.getSimpleVT().SimpleTy) {
20392 // X86 has 8, 16, and 32-bit zero-extending loads.
20400 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20401 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
20404 VT = VT.getScalarType();
20406 if (!VT.isSimple())
20409 switch (VT.getSimpleVT().SimpleTy) {
20420 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20421 // i16 instructions are longer (0x66 prefix) and potentially slower.
20422 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20425 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20426 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20427 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20428 /// are assumed to be legal.
20430 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20432 if (!VT.isSimple())
20435 MVT SVT = VT.getSimpleVT();
20437 // Very little shuffling can be done for 64-bit vectors right now.
20438 if (VT.getSizeInBits() == 64)
20441 // This is an experimental legality test that is tailored to match the
20442 // legality test of the experimental lowering more closely. They are gated
20443 // separately to ease testing of performance differences.
20444 if (ExperimentalVectorShuffleLegality)
20445 // We only care that the types being shuffled are legal. The lowering can
20446 // handle any possible shuffle mask that results.
20447 return isTypeLegal(SVT);
20449 // If this is a single-input shuffle with no 128 bit lane crossings we can
20450 // lower it into pshufb.
20451 if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
20452 (SVT.is256BitVector() && Subtarget->hasInt256())) {
20453 bool isLegal = true;
20454 for (unsigned I = 0, E = M.size(); I != E; ++I) {
20455 if (M[I] >= (int)SVT.getVectorNumElements() ||
20456 ShuffleCrosses128bitLane(SVT, I, M[I])) {
20465 // FIXME: blends, shifts.
20466 return (SVT.getVectorNumElements() == 2 ||
20467 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
20468 isMOVLMask(M, SVT) ||
20469 isCommutedMOVLMask(M, SVT) ||
20470 isMOVHLPSMask(M, SVT) ||
20471 isSHUFPMask(M, SVT) ||
20472 isSHUFPMask(M, SVT, /* Commuted */ true) ||
20473 isPSHUFDMask(M, SVT) ||
20474 isPSHUFDMask(M, SVT, /* SecondOperand */ true) ||
20475 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
20476 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
20477 isPALIGNRMask(M, SVT, Subtarget) ||
20478 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
20479 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
20480 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20481 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20482 isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()) ||
20483 (Subtarget->hasSSE41() && isINSERTPSMask(M, SVT)));
20487 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20489 if (!VT.isSimple())
20492 MVT SVT = VT.getSimpleVT();
20494 // This is an experimental legality test that is tailored to match the
20495 // legality test of the experimental lowering more closely. They are gated
20496 // separately to ease testing of performance differences.
20497 if (ExperimentalVectorShuffleLegality)
20498 // The new vector shuffle lowering is very good at managing zero-inputs.
20499 return isShuffleMaskLegal(Mask, VT);
20501 unsigned NumElts = SVT.getVectorNumElements();
20502 // FIXME: This collection of masks seems suspect.
20505 if (NumElts == 4 && SVT.is128BitVector()) {
20506 return (isMOVLMask(Mask, SVT) ||
20507 isCommutedMOVLMask(Mask, SVT, true) ||
20508 isSHUFPMask(Mask, SVT) ||
20509 isSHUFPMask(Mask, SVT, /* Commuted */ true) ||
20510 isBlendMask(Mask, SVT, Subtarget->hasSSE41(),
20511 Subtarget->hasInt256()));
20516 //===----------------------------------------------------------------------===//
20517 // X86 Scheduler Hooks
20518 //===----------------------------------------------------------------------===//
20520 /// Utility function to emit xbegin specifying the start of an RTM region.
20521 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20522 const TargetInstrInfo *TII) {
20523 DebugLoc DL = MI->getDebugLoc();
20525 const BasicBlock *BB = MBB->getBasicBlock();
20526 MachineFunction::iterator I = MBB;
20529 // For the v = xbegin(), we generate
20540 MachineBasicBlock *thisMBB = MBB;
20541 MachineFunction *MF = MBB->getParent();
20542 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20543 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20544 MF->insert(I, mainMBB);
20545 MF->insert(I, sinkMBB);
20547 // Transfer the remainder of BB and its successor edges to sinkMBB.
20548 sinkMBB->splice(sinkMBB->begin(), MBB,
20549 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20550 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20554 // # fallthrough to mainMBB
20555 // # abortion to sinkMBB
20556 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20557 thisMBB->addSuccessor(mainMBB);
20558 thisMBB->addSuccessor(sinkMBB);
20562 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20563 mainMBB->addSuccessor(sinkMBB);
20566 // EAX is live into the sinkMBB
20567 sinkMBB->addLiveIn(X86::EAX);
20568 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20569 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20572 MI->eraseFromParent();
20576 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20577 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20578 // in the .td file.
20579 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20580 const TargetInstrInfo *TII) {
20582 switch (MI->getOpcode()) {
20583 default: llvm_unreachable("illegal opcode!");
20584 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20585 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20586 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20587 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20588 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20589 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20590 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20591 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20594 DebugLoc dl = MI->getDebugLoc();
20595 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20597 unsigned NumArgs = MI->getNumOperands();
20598 for (unsigned i = 1; i < NumArgs; ++i) {
20599 MachineOperand &Op = MI->getOperand(i);
20600 if (!(Op.isReg() && Op.isImplicit()))
20601 MIB.addOperand(Op);
20603 if (MI->hasOneMemOperand())
20604 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20606 BuildMI(*BB, MI, dl,
20607 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20608 .addReg(X86::XMM0);
20610 MI->eraseFromParent();
20614 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20615 // defs in an instruction pattern
20616 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20617 const TargetInstrInfo *TII) {
20619 switch (MI->getOpcode()) {
20620 default: llvm_unreachable("illegal opcode!");
20621 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20622 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20623 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20624 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20625 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20626 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20627 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20628 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20631 DebugLoc dl = MI->getDebugLoc();
20632 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20634 unsigned NumArgs = MI->getNumOperands(); // remove the results
20635 for (unsigned i = 1; i < NumArgs; ++i) {
20636 MachineOperand &Op = MI->getOperand(i);
20637 if (!(Op.isReg() && Op.isImplicit()))
20638 MIB.addOperand(Op);
20640 if (MI->hasOneMemOperand())
20641 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20643 BuildMI(*BB, MI, dl,
20644 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20647 MI->eraseFromParent();
20651 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20652 const X86Subtarget *Subtarget) {
20653 DebugLoc dl = MI->getDebugLoc();
20654 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20655 // Address into RAX/EAX, other two args into ECX, EDX.
20656 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20657 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20658 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20659 for (int i = 0; i < X86::AddrNumOperands; ++i)
20660 MIB.addOperand(MI->getOperand(i));
20662 unsigned ValOps = X86::AddrNumOperands;
20663 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20664 .addReg(MI->getOperand(ValOps).getReg());
20665 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20666 .addReg(MI->getOperand(ValOps+1).getReg());
20668 // The instruction doesn't actually take any operands though.
20669 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20671 MI->eraseFromParent(); // The pseudo is gone now.
20675 MachineBasicBlock *
20676 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20677 MachineBasicBlock *MBB) const {
20678 // Emit va_arg instruction on X86-64.
20680 // Operands to this pseudo-instruction:
20681 // 0 ) Output : destination address (reg)
20682 // 1-5) Input : va_list address (addr, i64mem)
20683 // 6 ) ArgSize : Size (in bytes) of vararg type
20684 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20685 // 8 ) Align : Alignment of type
20686 // 9 ) EFLAGS (implicit-def)
20688 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20689 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
20691 unsigned DestReg = MI->getOperand(0).getReg();
20692 MachineOperand &Base = MI->getOperand(1);
20693 MachineOperand &Scale = MI->getOperand(2);
20694 MachineOperand &Index = MI->getOperand(3);
20695 MachineOperand &Disp = MI->getOperand(4);
20696 MachineOperand &Segment = MI->getOperand(5);
20697 unsigned ArgSize = MI->getOperand(6).getImm();
20698 unsigned ArgMode = MI->getOperand(7).getImm();
20699 unsigned Align = MI->getOperand(8).getImm();
20701 // Memory Reference
20702 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20703 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20704 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20706 // Machine Information
20707 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20708 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20709 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20710 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20711 DebugLoc DL = MI->getDebugLoc();
20713 // struct va_list {
20716 // i64 overflow_area (address)
20717 // i64 reg_save_area (address)
20719 // sizeof(va_list) = 24
20720 // alignment(va_list) = 8
20722 unsigned TotalNumIntRegs = 6;
20723 unsigned TotalNumXMMRegs = 8;
20724 bool UseGPOffset = (ArgMode == 1);
20725 bool UseFPOffset = (ArgMode == 2);
20726 unsigned MaxOffset = TotalNumIntRegs * 8 +
20727 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20729 /* Align ArgSize to a multiple of 8 */
20730 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20731 bool NeedsAlign = (Align > 8);
20733 MachineBasicBlock *thisMBB = MBB;
20734 MachineBasicBlock *overflowMBB;
20735 MachineBasicBlock *offsetMBB;
20736 MachineBasicBlock *endMBB;
20738 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20739 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20740 unsigned OffsetReg = 0;
20742 if (!UseGPOffset && !UseFPOffset) {
20743 // If we only pull from the overflow region, we don't create a branch.
20744 // We don't need to alter control flow.
20745 OffsetDestReg = 0; // unused
20746 OverflowDestReg = DestReg;
20748 offsetMBB = nullptr;
20749 overflowMBB = thisMBB;
20752 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20753 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20754 // If not, pull from overflow_area. (branch to overflowMBB)
20759 // offsetMBB overflowMBB
20764 // Registers for the PHI in endMBB
20765 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20766 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20768 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20769 MachineFunction *MF = MBB->getParent();
20770 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20771 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20772 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20774 MachineFunction::iterator MBBIter = MBB;
20777 // Insert the new basic blocks
20778 MF->insert(MBBIter, offsetMBB);
20779 MF->insert(MBBIter, overflowMBB);
20780 MF->insert(MBBIter, endMBB);
20782 // Transfer the remainder of MBB and its successor edges to endMBB.
20783 endMBB->splice(endMBB->begin(), thisMBB,
20784 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20785 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20787 // Make offsetMBB and overflowMBB successors of thisMBB
20788 thisMBB->addSuccessor(offsetMBB);
20789 thisMBB->addSuccessor(overflowMBB);
20791 // endMBB is a successor of both offsetMBB and overflowMBB
20792 offsetMBB->addSuccessor(endMBB);
20793 overflowMBB->addSuccessor(endMBB);
20795 // Load the offset value into a register
20796 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20797 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20801 .addDisp(Disp, UseFPOffset ? 4 : 0)
20802 .addOperand(Segment)
20803 .setMemRefs(MMOBegin, MMOEnd);
20805 // Check if there is enough room left to pull this argument.
20806 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20808 .addImm(MaxOffset + 8 - ArgSizeA8);
20810 // Branch to "overflowMBB" if offset >= max
20811 // Fall through to "offsetMBB" otherwise
20812 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20813 .addMBB(overflowMBB);
20816 // In offsetMBB, emit code to use the reg_save_area.
20818 assert(OffsetReg != 0);
20820 // Read the reg_save_area address.
20821 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20822 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20827 .addOperand(Segment)
20828 .setMemRefs(MMOBegin, MMOEnd);
20830 // Zero-extend the offset
20831 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20832 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20835 .addImm(X86::sub_32bit);
20837 // Add the offset to the reg_save_area to get the final address.
20838 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20839 .addReg(OffsetReg64)
20840 .addReg(RegSaveReg);
20842 // Compute the offset for the next argument
20843 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20844 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20846 .addImm(UseFPOffset ? 16 : 8);
20848 // Store it back into the va_list.
20849 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20853 .addDisp(Disp, UseFPOffset ? 4 : 0)
20854 .addOperand(Segment)
20855 .addReg(NextOffsetReg)
20856 .setMemRefs(MMOBegin, MMOEnd);
20859 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20864 // Emit code to use overflow area
20867 // Load the overflow_area address into a register.
20868 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20869 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20874 .addOperand(Segment)
20875 .setMemRefs(MMOBegin, MMOEnd);
20877 // If we need to align it, do so. Otherwise, just copy the address
20878 // to OverflowDestReg.
20880 // Align the overflow address
20881 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20882 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20884 // aligned_addr = (addr + (align-1)) & ~(align-1)
20885 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20886 .addReg(OverflowAddrReg)
20889 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20891 .addImm(~(uint64_t)(Align-1));
20893 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20894 .addReg(OverflowAddrReg);
20897 // Compute the next overflow address after this argument.
20898 // (the overflow address should be kept 8-byte aligned)
20899 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20900 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20901 .addReg(OverflowDestReg)
20902 .addImm(ArgSizeA8);
20904 // Store the new overflow address.
20905 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20910 .addOperand(Segment)
20911 .addReg(NextAddrReg)
20912 .setMemRefs(MMOBegin, MMOEnd);
20914 // If we branched, emit the PHI to the front of endMBB.
20916 BuildMI(*endMBB, endMBB->begin(), DL,
20917 TII->get(X86::PHI), DestReg)
20918 .addReg(OffsetDestReg).addMBB(offsetMBB)
20919 .addReg(OverflowDestReg).addMBB(overflowMBB);
20922 // Erase the pseudo instruction
20923 MI->eraseFromParent();
20928 MachineBasicBlock *
20929 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20931 MachineBasicBlock *MBB) const {
20932 // Emit code to save XMM registers to the stack. The ABI says that the
20933 // number of registers to save is given in %al, so it's theoretically
20934 // possible to do an indirect jump trick to avoid saving all of them,
20935 // however this code takes a simpler approach and just executes all
20936 // of the stores if %al is non-zero. It's less code, and it's probably
20937 // easier on the hardware branch predictor, and stores aren't all that
20938 // expensive anyway.
20940 // Create the new basic blocks. One block contains all the XMM stores,
20941 // and one block is the final destination regardless of whether any
20942 // stores were performed.
20943 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20944 MachineFunction *F = MBB->getParent();
20945 MachineFunction::iterator MBBIter = MBB;
20947 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20948 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20949 F->insert(MBBIter, XMMSaveMBB);
20950 F->insert(MBBIter, EndMBB);
20952 // Transfer the remainder of MBB and its successor edges to EndMBB.
20953 EndMBB->splice(EndMBB->begin(), MBB,
20954 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20955 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20957 // The original block will now fall through to the XMM save block.
20958 MBB->addSuccessor(XMMSaveMBB);
20959 // The XMMSaveMBB will fall through to the end block.
20960 XMMSaveMBB->addSuccessor(EndMBB);
20962 // Now add the instructions.
20963 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20964 DebugLoc DL = MI->getDebugLoc();
20966 unsigned CountReg = MI->getOperand(0).getReg();
20967 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20968 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20970 if (!Subtarget->isTargetWin64()) {
20971 // If %al is 0, branch around the XMM save block.
20972 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20973 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20974 MBB->addSuccessor(EndMBB);
20977 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20978 // that was just emitted, but clearly shouldn't be "saved".
20979 assert((MI->getNumOperands() <= 3 ||
20980 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20981 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20982 && "Expected last argument to be EFLAGS");
20983 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20984 // In the XMM save block, save all the XMM argument registers.
20985 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20986 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20987 MachineMemOperand *MMO =
20988 F->getMachineMemOperand(
20989 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
20990 MachineMemOperand::MOStore,
20991 /*Size=*/16, /*Align=*/16);
20992 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20993 .addFrameIndex(RegSaveFrameIndex)
20994 .addImm(/*Scale=*/1)
20995 .addReg(/*IndexReg=*/0)
20996 .addImm(/*Disp=*/Offset)
20997 .addReg(/*Segment=*/0)
20998 .addReg(MI->getOperand(i).getReg())
20999 .addMemOperand(MMO);
21002 MI->eraseFromParent(); // The pseudo instruction is gone now.
21007 // The EFLAGS operand of SelectItr might be missing a kill marker
21008 // because there were multiple uses of EFLAGS, and ISel didn't know
21009 // which to mark. Figure out whether SelectItr should have had a
21010 // kill marker, and set it if it should. Returns the correct kill
21012 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
21013 MachineBasicBlock* BB,
21014 const TargetRegisterInfo* TRI) {
21015 // Scan forward through BB for a use/def of EFLAGS.
21016 MachineBasicBlock::iterator miI(std::next(SelectItr));
21017 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
21018 const MachineInstr& mi = *miI;
21019 if (mi.readsRegister(X86::EFLAGS))
21021 if (mi.definesRegister(X86::EFLAGS))
21022 break; // Should have kill-flag - update below.
21025 // If we hit the end of the block, check whether EFLAGS is live into a
21027 if (miI == BB->end()) {
21028 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
21029 sEnd = BB->succ_end();
21030 sItr != sEnd; ++sItr) {
21031 MachineBasicBlock* succ = *sItr;
21032 if (succ->isLiveIn(X86::EFLAGS))
21037 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
21038 // out. SelectMI should have a kill flag on EFLAGS.
21039 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
21043 MachineBasicBlock *
21044 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
21045 MachineBasicBlock *BB) const {
21046 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21047 DebugLoc DL = MI->getDebugLoc();
21049 // To "insert" a SELECT_CC instruction, we actually have to insert the
21050 // diamond control-flow pattern. The incoming instruction knows the
21051 // destination vreg to set, the condition code register to branch on, the
21052 // true/false values to select between, and a branch opcode to use.
21053 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21054 MachineFunction::iterator It = BB;
21060 // cmpTY ccX, r1, r2
21062 // fallthrough --> copy0MBB
21063 MachineBasicBlock *thisMBB = BB;
21064 MachineFunction *F = BB->getParent();
21065 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21066 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21067 F->insert(It, copy0MBB);
21068 F->insert(It, sinkMBB);
21070 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21071 // live into the sink and copy blocks.
21072 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21073 if (!MI->killsRegister(X86::EFLAGS) &&
21074 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
21075 copy0MBB->addLiveIn(X86::EFLAGS);
21076 sinkMBB->addLiveIn(X86::EFLAGS);
21079 // Transfer the remainder of BB and its successor edges to sinkMBB.
21080 sinkMBB->splice(sinkMBB->begin(), BB,
21081 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21082 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21084 // Add the true and fallthrough blocks as its successors.
21085 BB->addSuccessor(copy0MBB);
21086 BB->addSuccessor(sinkMBB);
21088 // Create the conditional branch instruction.
21090 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
21091 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21094 // %FalseValue = ...
21095 // # fallthrough to sinkMBB
21096 copy0MBB->addSuccessor(sinkMBB);
21099 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21101 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21102 TII->get(X86::PHI), MI->getOperand(0).getReg())
21103 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
21104 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
21106 MI->eraseFromParent(); // The pseudo instruction is gone now.
21110 MachineBasicBlock *
21111 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21112 MachineBasicBlock *BB) const {
21113 MachineFunction *MF = BB->getParent();
21114 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21115 DebugLoc DL = MI->getDebugLoc();
21116 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21118 assert(MF->shouldSplitStack());
21120 const bool Is64Bit = Subtarget->is64Bit();
21121 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21123 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21124 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21127 // ... [Till the alloca]
21128 // If stacklet is not large enough, jump to mallocMBB
21131 // Allocate by subtracting from RSP
21132 // Jump to continueMBB
21135 // Allocate by call to runtime
21139 // [rest of original BB]
21142 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21143 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21144 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21146 MachineRegisterInfo &MRI = MF->getRegInfo();
21147 const TargetRegisterClass *AddrRegClass =
21148 getRegClassFor(getPointerTy());
21150 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21151 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21152 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21153 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21154 sizeVReg = MI->getOperand(1).getReg(),
21155 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21157 MachineFunction::iterator MBBIter = BB;
21160 MF->insert(MBBIter, bumpMBB);
21161 MF->insert(MBBIter, mallocMBB);
21162 MF->insert(MBBIter, continueMBB);
21164 continueMBB->splice(continueMBB->begin(), BB,
21165 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21166 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21168 // Add code to the main basic block to check if the stack limit has been hit,
21169 // and if so, jump to mallocMBB otherwise to bumpMBB.
21170 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21171 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21172 .addReg(tmpSPVReg).addReg(sizeVReg);
21173 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21174 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21175 .addReg(SPLimitVReg);
21176 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21178 // bumpMBB simply decreases the stack pointer, since we know the current
21179 // stacklet has enough space.
21180 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21181 .addReg(SPLimitVReg);
21182 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21183 .addReg(SPLimitVReg);
21184 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21186 // Calls into a routine in libgcc to allocate more space from the heap.
21187 const uint32_t *RegMask =
21188 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
21190 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21192 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21193 .addExternalSymbol("__morestack_allocate_stack_space")
21194 .addRegMask(RegMask)
21195 .addReg(X86::RDI, RegState::Implicit)
21196 .addReg(X86::RAX, RegState::ImplicitDefine);
21197 } else if (Is64Bit) {
21198 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21200 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21201 .addExternalSymbol("__morestack_allocate_stack_space")
21202 .addRegMask(RegMask)
21203 .addReg(X86::EDI, RegState::Implicit)
21204 .addReg(X86::EAX, RegState::ImplicitDefine);
21206 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21208 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21209 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21210 .addExternalSymbol("__morestack_allocate_stack_space")
21211 .addRegMask(RegMask)
21212 .addReg(X86::EAX, RegState::ImplicitDefine);
21216 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21219 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21220 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21221 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21223 // Set up the CFG correctly.
21224 BB->addSuccessor(bumpMBB);
21225 BB->addSuccessor(mallocMBB);
21226 mallocMBB->addSuccessor(continueMBB);
21227 bumpMBB->addSuccessor(continueMBB);
21229 // Take care of the PHI nodes.
21230 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21231 MI->getOperand(0).getReg())
21232 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21233 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21235 // Delete the original pseudo instruction.
21236 MI->eraseFromParent();
21239 return continueMBB;
21242 MachineBasicBlock *
21243 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21244 MachineBasicBlock *BB) const {
21245 DebugLoc DL = MI->getDebugLoc();
21247 assert(!Subtarget->isTargetMachO());
21249 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
21251 MI->eraseFromParent(); // The pseudo instruction is gone now.
21255 MachineBasicBlock *
21256 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21257 MachineBasicBlock *BB) const {
21258 // This is pretty easy. We're taking the value that we received from
21259 // our load from the relocation, sticking it in either RDI (x86-64)
21260 // or EAX and doing an indirect call. The return value will then
21261 // be in the normal return register.
21262 MachineFunction *F = BB->getParent();
21263 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21264 DebugLoc DL = MI->getDebugLoc();
21266 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21267 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21269 // Get a register mask for the lowered call.
21270 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21271 // proper register mask.
21272 const uint32_t *RegMask =
21273 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
21274 if (Subtarget->is64Bit()) {
21275 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21276 TII->get(X86::MOV64rm), X86::RDI)
21278 .addImm(0).addReg(0)
21279 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21280 MI->getOperand(3).getTargetFlags())
21282 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21283 addDirectMem(MIB, X86::RDI);
21284 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21285 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21286 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21287 TII->get(X86::MOV32rm), X86::EAX)
21289 .addImm(0).addReg(0)
21290 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21291 MI->getOperand(3).getTargetFlags())
21293 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21294 addDirectMem(MIB, X86::EAX);
21295 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21297 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21298 TII->get(X86::MOV32rm), X86::EAX)
21299 .addReg(TII->getGlobalBaseReg(F))
21300 .addImm(0).addReg(0)
21301 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21302 MI->getOperand(3).getTargetFlags())
21304 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21305 addDirectMem(MIB, X86::EAX);
21306 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21309 MI->eraseFromParent(); // The pseudo instruction is gone now.
21313 MachineBasicBlock *
21314 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21315 MachineBasicBlock *MBB) const {
21316 DebugLoc DL = MI->getDebugLoc();
21317 MachineFunction *MF = MBB->getParent();
21318 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21319 MachineRegisterInfo &MRI = MF->getRegInfo();
21321 const BasicBlock *BB = MBB->getBasicBlock();
21322 MachineFunction::iterator I = MBB;
21325 // Memory Reference
21326 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21327 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21330 unsigned MemOpndSlot = 0;
21332 unsigned CurOp = 0;
21334 DstReg = MI->getOperand(CurOp++).getReg();
21335 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21336 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21337 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21338 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21340 MemOpndSlot = CurOp;
21342 MVT PVT = getPointerTy();
21343 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21344 "Invalid Pointer Size!");
21346 // For v = setjmp(buf), we generate
21349 // buf[LabelOffset] = restoreMBB
21350 // SjLjSetup restoreMBB
21356 // v = phi(main, restore)
21359 // if base pointer being used, load it from frame
21362 MachineBasicBlock *thisMBB = MBB;
21363 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21364 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21365 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21366 MF->insert(I, mainMBB);
21367 MF->insert(I, sinkMBB);
21368 MF->push_back(restoreMBB);
21370 MachineInstrBuilder MIB;
21372 // Transfer the remainder of BB and its successor edges to sinkMBB.
21373 sinkMBB->splice(sinkMBB->begin(), MBB,
21374 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21375 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21378 unsigned PtrStoreOpc = 0;
21379 unsigned LabelReg = 0;
21380 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21381 Reloc::Model RM = MF->getTarget().getRelocationModel();
21382 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21383 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21385 // Prepare IP either in reg or imm.
21386 if (!UseImmLabel) {
21387 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21388 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21389 LabelReg = MRI.createVirtualRegister(PtrRC);
21390 if (Subtarget->is64Bit()) {
21391 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21395 .addMBB(restoreMBB)
21398 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21399 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21400 .addReg(XII->getGlobalBaseReg(MF))
21403 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21407 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21409 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21410 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21411 if (i == X86::AddrDisp)
21412 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21414 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21417 MIB.addReg(LabelReg);
21419 MIB.addMBB(restoreMBB);
21420 MIB.setMemRefs(MMOBegin, MMOEnd);
21422 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21423 .addMBB(restoreMBB);
21425 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21426 MIB.addRegMask(RegInfo->getNoPreservedMask());
21427 thisMBB->addSuccessor(mainMBB);
21428 thisMBB->addSuccessor(restoreMBB);
21432 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21433 mainMBB->addSuccessor(sinkMBB);
21436 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21437 TII->get(X86::PHI), DstReg)
21438 .addReg(mainDstReg).addMBB(mainMBB)
21439 .addReg(restoreDstReg).addMBB(restoreMBB);
21442 if (RegInfo->hasBasePointer(*MF)) {
21443 const bool Uses64BitFramePtr =
21444 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21445 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21446 X86FI->setRestoreBasePointer(MF);
21447 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21448 unsigned BasePtr = RegInfo->getBaseRegister();
21449 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21450 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21451 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21452 .setMIFlag(MachineInstr::FrameSetup);
21454 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21455 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21456 restoreMBB->addSuccessor(sinkMBB);
21458 MI->eraseFromParent();
21462 MachineBasicBlock *
21463 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21464 MachineBasicBlock *MBB) const {
21465 DebugLoc DL = MI->getDebugLoc();
21466 MachineFunction *MF = MBB->getParent();
21467 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21468 MachineRegisterInfo &MRI = MF->getRegInfo();
21470 // Memory Reference
21471 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21472 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21474 MVT PVT = getPointerTy();
21475 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21476 "Invalid Pointer Size!");
21478 const TargetRegisterClass *RC =
21479 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21480 unsigned Tmp = MRI.createVirtualRegister(RC);
21481 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21482 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21483 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21484 unsigned SP = RegInfo->getStackRegister();
21486 MachineInstrBuilder MIB;
21488 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21489 const int64_t SPOffset = 2 * PVT.getStoreSize();
21491 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21492 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21495 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21496 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21497 MIB.addOperand(MI->getOperand(i));
21498 MIB.setMemRefs(MMOBegin, MMOEnd);
21500 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21501 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21502 if (i == X86::AddrDisp)
21503 MIB.addDisp(MI->getOperand(i), LabelOffset);
21505 MIB.addOperand(MI->getOperand(i));
21507 MIB.setMemRefs(MMOBegin, MMOEnd);
21509 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21510 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21511 if (i == X86::AddrDisp)
21512 MIB.addDisp(MI->getOperand(i), SPOffset);
21514 MIB.addOperand(MI->getOperand(i));
21516 MIB.setMemRefs(MMOBegin, MMOEnd);
21518 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21520 MI->eraseFromParent();
21524 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21525 // accumulator loops. Writing back to the accumulator allows the coalescer
21526 // to remove extra copies in the loop.
21527 MachineBasicBlock *
21528 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21529 MachineBasicBlock *MBB) const {
21530 MachineOperand &AddendOp = MI->getOperand(3);
21532 // Bail out early if the addend isn't a register - we can't switch these.
21533 if (!AddendOp.isReg())
21536 MachineFunction &MF = *MBB->getParent();
21537 MachineRegisterInfo &MRI = MF.getRegInfo();
21539 // Check whether the addend is defined by a PHI:
21540 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21541 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21542 if (!AddendDef.isPHI())
21545 // Look for the following pattern:
21547 // %addend = phi [%entry, 0], [%loop, %result]
21549 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21553 // %addend = phi [%entry, 0], [%loop, %result]
21555 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21557 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21558 assert(AddendDef.getOperand(i).isReg());
21559 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21560 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21561 if (&PHISrcInst == MI) {
21562 // Found a matching instruction.
21563 unsigned NewFMAOpc = 0;
21564 switch (MI->getOpcode()) {
21565 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21566 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21567 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21568 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21569 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21570 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21571 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21572 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21573 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21574 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21575 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21576 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21577 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21578 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21579 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21580 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21581 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21582 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21583 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21584 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21586 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21587 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21588 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21589 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21590 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21591 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21592 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21593 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21594 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21595 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21596 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21597 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21598 default: llvm_unreachable("Unrecognized FMA variant.");
21601 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21602 MachineInstrBuilder MIB =
21603 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21604 .addOperand(MI->getOperand(0))
21605 .addOperand(MI->getOperand(3))
21606 .addOperand(MI->getOperand(2))
21607 .addOperand(MI->getOperand(1));
21608 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21609 MI->eraseFromParent();
21616 MachineBasicBlock *
21617 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21618 MachineBasicBlock *BB) const {
21619 switch (MI->getOpcode()) {
21620 default: llvm_unreachable("Unexpected instr type to insert");
21621 case X86::TAILJMPd64:
21622 case X86::TAILJMPr64:
21623 case X86::TAILJMPm64:
21624 case X86::TAILJMPd64_REX:
21625 case X86::TAILJMPr64_REX:
21626 case X86::TAILJMPm64_REX:
21627 llvm_unreachable("TAILJMP64 would not be touched here.");
21628 case X86::TCRETURNdi64:
21629 case X86::TCRETURNri64:
21630 case X86::TCRETURNmi64:
21632 case X86::WIN_ALLOCA:
21633 return EmitLoweredWinAlloca(MI, BB);
21634 case X86::SEG_ALLOCA_32:
21635 case X86::SEG_ALLOCA_64:
21636 return EmitLoweredSegAlloca(MI, BB);
21637 case X86::TLSCall_32:
21638 case X86::TLSCall_64:
21639 return EmitLoweredTLSCall(MI, BB);
21640 case X86::CMOV_GR8:
21641 case X86::CMOV_FR32:
21642 case X86::CMOV_FR64:
21643 case X86::CMOV_V4F32:
21644 case X86::CMOV_V2F64:
21645 case X86::CMOV_V2I64:
21646 case X86::CMOV_V8F32:
21647 case X86::CMOV_V4F64:
21648 case X86::CMOV_V4I64:
21649 case X86::CMOV_V16F32:
21650 case X86::CMOV_V8F64:
21651 case X86::CMOV_V8I64:
21652 case X86::CMOV_GR16:
21653 case X86::CMOV_GR32:
21654 case X86::CMOV_RFP32:
21655 case X86::CMOV_RFP64:
21656 case X86::CMOV_RFP80:
21657 return EmitLoweredSelect(MI, BB);
21659 case X86::FP32_TO_INT16_IN_MEM:
21660 case X86::FP32_TO_INT32_IN_MEM:
21661 case X86::FP32_TO_INT64_IN_MEM:
21662 case X86::FP64_TO_INT16_IN_MEM:
21663 case X86::FP64_TO_INT32_IN_MEM:
21664 case X86::FP64_TO_INT64_IN_MEM:
21665 case X86::FP80_TO_INT16_IN_MEM:
21666 case X86::FP80_TO_INT32_IN_MEM:
21667 case X86::FP80_TO_INT64_IN_MEM: {
21668 MachineFunction *F = BB->getParent();
21669 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21670 DebugLoc DL = MI->getDebugLoc();
21672 // Change the floating point control register to use "round towards zero"
21673 // mode when truncating to an integer value.
21674 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21675 addFrameReference(BuildMI(*BB, MI, DL,
21676 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21678 // Load the old value of the high byte of the control word...
21680 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21681 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21684 // Set the high part to be round to zero...
21685 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21688 // Reload the modified control word now...
21689 addFrameReference(BuildMI(*BB, MI, DL,
21690 TII->get(X86::FLDCW16m)), CWFrameIdx);
21692 // Restore the memory image of control word to original value
21693 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21696 // Get the X86 opcode to use.
21698 switch (MI->getOpcode()) {
21699 default: llvm_unreachable("illegal opcode!");
21700 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21701 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21702 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21703 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21704 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21705 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21706 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21707 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21708 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21712 MachineOperand &Op = MI->getOperand(0);
21714 AM.BaseType = X86AddressMode::RegBase;
21715 AM.Base.Reg = Op.getReg();
21717 AM.BaseType = X86AddressMode::FrameIndexBase;
21718 AM.Base.FrameIndex = Op.getIndex();
21720 Op = MI->getOperand(1);
21722 AM.Scale = Op.getImm();
21723 Op = MI->getOperand(2);
21725 AM.IndexReg = Op.getImm();
21726 Op = MI->getOperand(3);
21727 if (Op.isGlobal()) {
21728 AM.GV = Op.getGlobal();
21730 AM.Disp = Op.getImm();
21732 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21733 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21735 // Reload the original control word now.
21736 addFrameReference(BuildMI(*BB, MI, DL,
21737 TII->get(X86::FLDCW16m)), CWFrameIdx);
21739 MI->eraseFromParent(); // The pseudo instruction is gone now.
21742 // String/text processing lowering.
21743 case X86::PCMPISTRM128REG:
21744 case X86::VPCMPISTRM128REG:
21745 case X86::PCMPISTRM128MEM:
21746 case X86::VPCMPISTRM128MEM:
21747 case X86::PCMPESTRM128REG:
21748 case X86::VPCMPESTRM128REG:
21749 case X86::PCMPESTRM128MEM:
21750 case X86::VPCMPESTRM128MEM:
21751 assert(Subtarget->hasSSE42() &&
21752 "Target must have SSE4.2 or AVX features enabled");
21753 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21755 // String/text processing lowering.
21756 case X86::PCMPISTRIREG:
21757 case X86::VPCMPISTRIREG:
21758 case X86::PCMPISTRIMEM:
21759 case X86::VPCMPISTRIMEM:
21760 case X86::PCMPESTRIREG:
21761 case X86::VPCMPESTRIREG:
21762 case X86::PCMPESTRIMEM:
21763 case X86::VPCMPESTRIMEM:
21764 assert(Subtarget->hasSSE42() &&
21765 "Target must have SSE4.2 or AVX features enabled");
21766 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21768 // Thread synchronization.
21770 return EmitMonitor(MI, BB, Subtarget);
21774 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21776 case X86::VASTART_SAVE_XMM_REGS:
21777 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21779 case X86::VAARG_64:
21780 return EmitVAARG64WithCustomInserter(MI, BB);
21782 case X86::EH_SjLj_SetJmp32:
21783 case X86::EH_SjLj_SetJmp64:
21784 return emitEHSjLjSetJmp(MI, BB);
21786 case X86::EH_SjLj_LongJmp32:
21787 case X86::EH_SjLj_LongJmp64:
21788 return emitEHSjLjLongJmp(MI, BB);
21790 case TargetOpcode::STATEPOINT:
21791 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21792 // this point in the process. We diverge later.
21793 return emitPatchPoint(MI, BB);
21795 case TargetOpcode::STACKMAP:
21796 case TargetOpcode::PATCHPOINT:
21797 return emitPatchPoint(MI, BB);
21799 case X86::VFMADDPDr213r:
21800 case X86::VFMADDPSr213r:
21801 case X86::VFMADDSDr213r:
21802 case X86::VFMADDSSr213r:
21803 case X86::VFMSUBPDr213r:
21804 case X86::VFMSUBPSr213r:
21805 case X86::VFMSUBSDr213r:
21806 case X86::VFMSUBSSr213r:
21807 case X86::VFNMADDPDr213r:
21808 case X86::VFNMADDPSr213r:
21809 case X86::VFNMADDSDr213r:
21810 case X86::VFNMADDSSr213r:
21811 case X86::VFNMSUBPDr213r:
21812 case X86::VFNMSUBPSr213r:
21813 case X86::VFNMSUBSDr213r:
21814 case X86::VFNMSUBSSr213r:
21815 case X86::VFMADDSUBPDr213r:
21816 case X86::VFMADDSUBPSr213r:
21817 case X86::VFMSUBADDPDr213r:
21818 case X86::VFMSUBADDPSr213r:
21819 case X86::VFMADDPDr213rY:
21820 case X86::VFMADDPSr213rY:
21821 case X86::VFMSUBPDr213rY:
21822 case X86::VFMSUBPSr213rY:
21823 case X86::VFNMADDPDr213rY:
21824 case X86::VFNMADDPSr213rY:
21825 case X86::VFNMSUBPDr213rY:
21826 case X86::VFNMSUBPSr213rY:
21827 case X86::VFMADDSUBPDr213rY:
21828 case X86::VFMADDSUBPSr213rY:
21829 case X86::VFMSUBADDPDr213rY:
21830 case X86::VFMSUBADDPSr213rY:
21831 return emitFMA3Instr(MI, BB);
21835 //===----------------------------------------------------------------------===//
21836 // X86 Optimization Hooks
21837 //===----------------------------------------------------------------------===//
21839 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21842 const SelectionDAG &DAG,
21843 unsigned Depth) const {
21844 unsigned BitWidth = KnownZero.getBitWidth();
21845 unsigned Opc = Op.getOpcode();
21846 assert((Opc >= ISD::BUILTIN_OP_END ||
21847 Opc == ISD::INTRINSIC_WO_CHAIN ||
21848 Opc == ISD::INTRINSIC_W_CHAIN ||
21849 Opc == ISD::INTRINSIC_VOID) &&
21850 "Should use MaskedValueIsZero if you don't know whether Op"
21851 " is a target node!");
21853 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21867 // These nodes' second result is a boolean.
21868 if (Op.getResNo() == 0)
21871 case X86ISD::SETCC:
21872 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21874 case ISD::INTRINSIC_WO_CHAIN: {
21875 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21876 unsigned NumLoBits = 0;
21879 case Intrinsic::x86_sse_movmsk_ps:
21880 case Intrinsic::x86_avx_movmsk_ps_256:
21881 case Intrinsic::x86_sse2_movmsk_pd:
21882 case Intrinsic::x86_avx_movmsk_pd_256:
21883 case Intrinsic::x86_mmx_pmovmskb:
21884 case Intrinsic::x86_sse2_pmovmskb_128:
21885 case Intrinsic::x86_avx2_pmovmskb: {
21886 // High bits of movmskp{s|d}, pmovmskb are known zero.
21888 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21889 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21890 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21891 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21892 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21893 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21894 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21895 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21897 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21906 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21908 const SelectionDAG &,
21909 unsigned Depth) const {
21910 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21911 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21912 return Op.getValueType().getScalarType().getSizeInBits();
21918 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21919 /// node is a GlobalAddress + offset.
21920 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21921 const GlobalValue* &GA,
21922 int64_t &Offset) const {
21923 if (N->getOpcode() == X86ISD::Wrapper) {
21924 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21925 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21926 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21930 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21933 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21934 /// same as extracting the high 128-bit part of 256-bit vector and then
21935 /// inserting the result into the low part of a new 256-bit vector
21936 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21937 EVT VT = SVOp->getValueType(0);
21938 unsigned NumElems = VT.getVectorNumElements();
21940 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21941 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21942 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21943 SVOp->getMaskElt(j) >= 0)
21949 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21950 /// same as extracting the low 128-bit part of 256-bit vector and then
21951 /// inserting the result into the high part of a new 256-bit vector
21952 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21953 EVT VT = SVOp->getValueType(0);
21954 unsigned NumElems = VT.getVectorNumElements();
21956 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21957 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21958 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21959 SVOp->getMaskElt(j) >= 0)
21965 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21966 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21967 TargetLowering::DAGCombinerInfo &DCI,
21968 const X86Subtarget* Subtarget) {
21970 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21971 SDValue V1 = SVOp->getOperand(0);
21972 SDValue V2 = SVOp->getOperand(1);
21973 EVT VT = SVOp->getValueType(0);
21974 unsigned NumElems = VT.getVectorNumElements();
21976 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21977 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21981 // V UNDEF BUILD_VECTOR UNDEF
21983 // CONCAT_VECTOR CONCAT_VECTOR
21986 // RESULT: V + zero extended
21988 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21989 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21990 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21993 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21996 // To match the shuffle mask, the first half of the mask should
21997 // be exactly the first vector, and all the rest a splat with the
21998 // first element of the second one.
21999 for (unsigned i = 0; i != NumElems/2; ++i)
22000 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22001 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22004 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22005 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22006 if (Ld->hasNUsesOfValue(1, 0)) {
22007 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22008 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22010 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22012 Ld->getPointerInfo(),
22013 Ld->getAlignment(),
22014 false/*isVolatile*/, true/*ReadMem*/,
22015 false/*WriteMem*/);
22017 // Make sure the newly-created LOAD is in the same position as Ld in
22018 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22019 // and update uses of Ld's output chain to use the TokenFactor.
22020 if (Ld->hasAnyUseOfValue(1)) {
22021 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22022 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22023 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22024 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22025 SDValue(ResNode.getNode(), 1));
22028 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
22032 // Emit a zeroed vector and insert the desired subvector on its
22034 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22035 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22036 return DCI.CombineTo(N, InsV);
22039 //===--------------------------------------------------------------------===//
22040 // Combine some shuffles into subvector extracts and inserts:
22043 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22044 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22045 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22046 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22047 return DCI.CombineTo(N, InsV);
22050 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22051 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22052 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22053 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22054 return DCI.CombineTo(N, InsV);
22060 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22063 /// This is the leaf of the recursive combinine below. When we have found some
22064 /// chain of single-use x86 shuffle instructions and accumulated the combined
22065 /// shuffle mask represented by them, this will try to pattern match that mask
22066 /// into either a single instruction if there is a special purpose instruction
22067 /// for this operation, or into a PSHUFB instruction which is a fully general
22068 /// instruction but should only be used to replace chains over a certain depth.
22069 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22070 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22071 TargetLowering::DAGCombinerInfo &DCI,
22072 const X86Subtarget *Subtarget) {
22073 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22075 // Find the operand that enters the chain. Note that multiple uses are OK
22076 // here, we're not going to remove the operand we find.
22077 SDValue Input = Op.getOperand(0);
22078 while (Input.getOpcode() == ISD::BITCAST)
22079 Input = Input.getOperand(0);
22081 MVT VT = Input.getSimpleValueType();
22082 MVT RootVT = Root.getSimpleValueType();
22085 // Just remove no-op shuffle masks.
22086 if (Mask.size() == 1) {
22087 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
22092 // Use the float domain if the operand type is a floating point type.
22093 bool FloatDomain = VT.isFloatingPoint();
22095 // For floating point shuffles, we don't have free copies in the shuffle
22096 // instructions or the ability to load as part of the instruction, so
22097 // canonicalize their shuffles to UNPCK or MOV variants.
22099 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22100 // vectors because it can have a load folded into it that UNPCK cannot. This
22101 // doesn't preclude something switching to the shorter encoding post-RA.
22103 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
22104 bool Lo = Mask.equals(0, 0);
22107 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22108 // is no slower than UNPCKLPD but has the option to fold the input operand
22109 // into even an unaligned memory load.
22110 if (Lo && Subtarget->hasSSE3()) {
22111 Shuffle = X86ISD::MOVDDUP;
22112 ShuffleVT = MVT::v2f64;
22114 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22115 // than the UNPCK variants.
22116 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22117 ShuffleVT = MVT::v4f32;
22119 if (Depth == 1 && Root->getOpcode() == Shuffle)
22120 return false; // Nothing to do!
22121 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22122 DCI.AddToWorklist(Op.getNode());
22123 if (Shuffle == X86ISD::MOVDDUP)
22124 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22126 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22127 DCI.AddToWorklist(Op.getNode());
22128 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22132 if (Subtarget->hasSSE3() &&
22133 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
22134 bool Lo = Mask.equals(0, 0, 2, 2);
22135 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22136 MVT ShuffleVT = MVT::v4f32;
22137 if (Depth == 1 && Root->getOpcode() == Shuffle)
22138 return false; // Nothing to do!
22139 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22140 DCI.AddToWorklist(Op.getNode());
22141 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22142 DCI.AddToWorklist(Op.getNode());
22143 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22147 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
22148 bool Lo = Mask.equals(0, 0, 1, 1);
22149 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22150 MVT ShuffleVT = MVT::v4f32;
22151 if (Depth == 1 && Root->getOpcode() == Shuffle)
22152 return false; // Nothing to do!
22153 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22154 DCI.AddToWorklist(Op.getNode());
22155 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22156 DCI.AddToWorklist(Op.getNode());
22157 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22163 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22164 // variants as none of these have single-instruction variants that are
22165 // superior to the UNPCK formulation.
22166 if (!FloatDomain &&
22167 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
22168 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
22169 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
22170 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
22172 bool Lo = Mask[0] == 0;
22173 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22174 if (Depth == 1 && Root->getOpcode() == Shuffle)
22175 return false; // Nothing to do!
22177 switch (Mask.size()) {
22179 ShuffleVT = MVT::v8i16;
22182 ShuffleVT = MVT::v16i8;
22185 llvm_unreachable("Impossible mask size!");
22187 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22188 DCI.AddToWorklist(Op.getNode());
22189 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22190 DCI.AddToWorklist(Op.getNode());
22191 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22196 // Don't try to re-form single instruction chains under any circumstances now
22197 // that we've done encoding canonicalization for them.
22201 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22202 // can replace them with a single PSHUFB instruction profitably. Intel's
22203 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22204 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22205 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22206 SmallVector<SDValue, 16> PSHUFBMask;
22207 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
22208 int Ratio = 16 / Mask.size();
22209 for (unsigned i = 0; i < 16; ++i) {
22210 if (Mask[i / Ratio] == SM_SentinelUndef) {
22211 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22214 int M = Mask[i / Ratio] != SM_SentinelZero
22215 ? Ratio * Mask[i / Ratio] + i % Ratio
22217 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
22219 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
22220 DCI.AddToWorklist(Op.getNode());
22221 SDValue PSHUFBMaskOp =
22222 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
22223 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22224 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
22225 DCI.AddToWorklist(Op.getNode());
22226 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22231 // Failed to find any combines.
22235 /// \brief Fully generic combining of x86 shuffle instructions.
22237 /// This should be the last combine run over the x86 shuffle instructions. Once
22238 /// they have been fully optimized, this will recursively consider all chains
22239 /// of single-use shuffle instructions, build a generic model of the cumulative
22240 /// shuffle operation, and check for simpler instructions which implement this
22241 /// operation. We use this primarily for two purposes:
22243 /// 1) Collapse generic shuffles to specialized single instructions when
22244 /// equivalent. In most cases, this is just an encoding size win, but
22245 /// sometimes we will collapse multiple generic shuffles into a single
22246 /// special-purpose shuffle.
22247 /// 2) Look for sequences of shuffle instructions with 3 or more total
22248 /// instructions, and replace them with the slightly more expensive SSSE3
22249 /// PSHUFB instruction if available. We do this as the last combining step
22250 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22251 /// a suitable short sequence of other instructions. The PHUFB will either
22252 /// use a register or have to read from memory and so is slightly (but only
22253 /// slightly) more expensive than the other shuffle instructions.
22255 /// Because this is inherently a quadratic operation (for each shuffle in
22256 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22257 /// This should never be an issue in practice as the shuffle lowering doesn't
22258 /// produce sequences of more than 8 instructions.
22260 /// FIXME: We will currently miss some cases where the redundant shuffling
22261 /// would simplify under the threshold for PSHUFB formation because of
22262 /// combine-ordering. To fix this, we should do the redundant instruction
22263 /// combining in this recursive walk.
22264 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22265 ArrayRef<int> RootMask,
22266 int Depth, bool HasPSHUFB,
22268 TargetLowering::DAGCombinerInfo &DCI,
22269 const X86Subtarget *Subtarget) {
22270 // Bound the depth of our recursive combine because this is ultimately
22271 // quadratic in nature.
22275 // Directly rip through bitcasts to find the underlying operand.
22276 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22277 Op = Op.getOperand(0);
22279 MVT VT = Op.getSimpleValueType();
22280 if (!VT.isVector())
22281 return false; // Bail if we hit a non-vector.
22282 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
22283 // version should be added.
22284 if (VT.getSizeInBits() != 128)
22287 assert(Root.getSimpleValueType().isVector() &&
22288 "Shuffles operate on vector types!");
22289 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22290 "Can only combine shuffles of the same vector register size.");
22292 if (!isTargetShuffle(Op.getOpcode()))
22294 SmallVector<int, 16> OpMask;
22296 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22297 // We only can combine unary shuffles which we can decode the mask for.
22298 if (!HaveMask || !IsUnary)
22301 assert(VT.getVectorNumElements() == OpMask.size() &&
22302 "Different mask size from vector size!");
22303 assert(((RootMask.size() > OpMask.size() &&
22304 RootMask.size() % OpMask.size() == 0) ||
22305 (OpMask.size() > RootMask.size() &&
22306 OpMask.size() % RootMask.size() == 0) ||
22307 OpMask.size() == RootMask.size()) &&
22308 "The smaller number of elements must divide the larger.");
22309 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22310 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22311 assert(((RootRatio == 1 && OpRatio == 1) ||
22312 (RootRatio == 1) != (OpRatio == 1)) &&
22313 "Must not have a ratio for both incoming and op masks!");
22315 SmallVector<int, 16> Mask;
22316 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22318 // Merge this shuffle operation's mask into our accumulated mask. Note that
22319 // this shuffle's mask will be the first applied to the input, followed by the
22320 // root mask to get us all the way to the root value arrangement. The reason
22321 // for this order is that we are recursing up the operation chain.
22322 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22323 int RootIdx = i / RootRatio;
22324 if (RootMask[RootIdx] < 0) {
22325 // This is a zero or undef lane, we're done.
22326 Mask.push_back(RootMask[RootIdx]);
22330 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22331 int OpIdx = RootMaskedIdx / OpRatio;
22332 if (OpMask[OpIdx] < 0) {
22333 // The incoming lanes are zero or undef, it doesn't matter which ones we
22335 Mask.push_back(OpMask[OpIdx]);
22339 // Ok, we have non-zero lanes, map them through.
22340 Mask.push_back(OpMask[OpIdx] * OpRatio +
22341 RootMaskedIdx % OpRatio);
22344 // See if we can recurse into the operand to combine more things.
22345 switch (Op.getOpcode()) {
22346 case X86ISD::PSHUFB:
22348 case X86ISD::PSHUFD:
22349 case X86ISD::PSHUFHW:
22350 case X86ISD::PSHUFLW:
22351 if (Op.getOperand(0).hasOneUse() &&
22352 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22353 HasPSHUFB, DAG, DCI, Subtarget))
22357 case X86ISD::UNPCKL:
22358 case X86ISD::UNPCKH:
22359 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
22360 // We can't check for single use, we have to check that this shuffle is the only user.
22361 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22362 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22363 HasPSHUFB, DAG, DCI, Subtarget))
22368 // Minor canonicalization of the accumulated shuffle mask to make it easier
22369 // to match below. All this does is detect masks with squential pairs of
22370 // elements, and shrink them to the half-width mask. It does this in a loop
22371 // so it will reduce the size of the mask to the minimal width mask which
22372 // performs an equivalent shuffle.
22373 SmallVector<int, 16> WidenedMask;
22374 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22375 Mask = std::move(WidenedMask);
22376 WidenedMask.clear();
22379 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22383 /// \brief Get the PSHUF-style mask from PSHUF node.
22385 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22386 /// PSHUF-style masks that can be reused with such instructions.
22387 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22388 SmallVector<int, 4> Mask;
22390 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
22394 switch (N.getOpcode()) {
22395 case X86ISD::PSHUFD:
22397 case X86ISD::PSHUFLW:
22400 case X86ISD::PSHUFHW:
22401 Mask.erase(Mask.begin(), Mask.begin() + 4);
22402 for (int &M : Mask)
22406 llvm_unreachable("No valid shuffle instruction found!");
22410 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22412 /// We walk up the chain and look for a combinable shuffle, skipping over
22413 /// shuffles that we could hoist this shuffle's transformation past without
22414 /// altering anything.
22416 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22418 TargetLowering::DAGCombinerInfo &DCI) {
22419 assert(N.getOpcode() == X86ISD::PSHUFD &&
22420 "Called with something other than an x86 128-bit half shuffle!");
22423 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22424 // of the shuffles in the chain so that we can form a fresh chain to replace
22426 SmallVector<SDValue, 8> Chain;
22427 SDValue V = N.getOperand(0);
22428 for (; V.hasOneUse(); V = V.getOperand(0)) {
22429 switch (V.getOpcode()) {
22431 return SDValue(); // Nothing combined!
22434 // Skip bitcasts as we always know the type for the target specific
22438 case X86ISD::PSHUFD:
22439 // Found another dword shuffle.
22442 case X86ISD::PSHUFLW:
22443 // Check that the low words (being shuffled) are the identity in the
22444 // dword shuffle, and the high words are self-contained.
22445 if (Mask[0] != 0 || Mask[1] != 1 ||
22446 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22449 Chain.push_back(V);
22452 case X86ISD::PSHUFHW:
22453 // Check that the high words (being shuffled) are the identity in the
22454 // dword shuffle, and the low words are self-contained.
22455 if (Mask[2] != 2 || Mask[3] != 3 ||
22456 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22459 Chain.push_back(V);
22462 case X86ISD::UNPCKL:
22463 case X86ISD::UNPCKH:
22464 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22465 // shuffle into a preceding word shuffle.
22466 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
22469 // Search for a half-shuffle which we can combine with.
22470 unsigned CombineOp =
22471 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22472 if (V.getOperand(0) != V.getOperand(1) ||
22473 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22475 Chain.push_back(V);
22476 V = V.getOperand(0);
22478 switch (V.getOpcode()) {
22480 return SDValue(); // Nothing to combine.
22482 case X86ISD::PSHUFLW:
22483 case X86ISD::PSHUFHW:
22484 if (V.getOpcode() == CombineOp)
22487 Chain.push_back(V);
22491 V = V.getOperand(0);
22495 } while (V.hasOneUse());
22498 // Break out of the loop if we break out of the switch.
22502 if (!V.hasOneUse())
22503 // We fell out of the loop without finding a viable combining instruction.
22506 // Merge this node's mask and our incoming mask.
22507 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22508 for (int &M : Mask)
22510 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22511 getV4X86ShuffleImm8ForMask(Mask, DAG));
22513 // Rebuild the chain around this new shuffle.
22514 while (!Chain.empty()) {
22515 SDValue W = Chain.pop_back_val();
22517 if (V.getValueType() != W.getOperand(0).getValueType())
22518 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
22520 switch (W.getOpcode()) {
22522 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22524 case X86ISD::UNPCKL:
22525 case X86ISD::UNPCKH:
22526 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22529 case X86ISD::PSHUFD:
22530 case X86ISD::PSHUFLW:
22531 case X86ISD::PSHUFHW:
22532 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22536 if (V.getValueType() != N.getValueType())
22537 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
22539 // Return the new chain to replace N.
22543 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
22545 /// We walk up the chain, skipping shuffles of the other half and looking
22546 /// through shuffles which switch halves trying to find a shuffle of the same
22547 /// pair of dwords.
22548 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22550 TargetLowering::DAGCombinerInfo &DCI) {
22552 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22553 "Called with something other than an x86 128-bit half shuffle!");
22555 unsigned CombineOpcode = N.getOpcode();
22557 // Walk up a single-use chain looking for a combinable shuffle.
22558 SDValue V = N.getOperand(0);
22559 for (; V.hasOneUse(); V = V.getOperand(0)) {
22560 switch (V.getOpcode()) {
22562 return false; // Nothing combined!
22565 // Skip bitcasts as we always know the type for the target specific
22569 case X86ISD::PSHUFLW:
22570 case X86ISD::PSHUFHW:
22571 if (V.getOpcode() == CombineOpcode)
22574 // Other-half shuffles are no-ops.
22577 // Break out of the loop if we break out of the switch.
22581 if (!V.hasOneUse())
22582 // We fell out of the loop without finding a viable combining instruction.
22585 // Combine away the bottom node as its shuffle will be accumulated into
22586 // a preceding shuffle.
22587 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22589 // Record the old value.
22592 // Merge this node's mask and our incoming mask (adjusted to account for all
22593 // the pshufd instructions encountered).
22594 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22595 for (int &M : Mask)
22597 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22598 getV4X86ShuffleImm8ForMask(Mask, DAG));
22600 // Check that the shuffles didn't cancel each other out. If not, we need to
22601 // combine to the new one.
22603 // Replace the combinable shuffle with the combined one, updating all users
22604 // so that we re-evaluate the chain here.
22605 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22610 /// \brief Try to combine x86 target specific shuffles.
22611 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22612 TargetLowering::DAGCombinerInfo &DCI,
22613 const X86Subtarget *Subtarget) {
22615 MVT VT = N.getSimpleValueType();
22616 SmallVector<int, 4> Mask;
22618 switch (N.getOpcode()) {
22619 case X86ISD::PSHUFD:
22620 case X86ISD::PSHUFLW:
22621 case X86ISD::PSHUFHW:
22622 Mask = getPSHUFShuffleMask(N);
22623 assert(Mask.size() == 4);
22629 // Nuke no-op shuffles that show up after combining.
22630 if (isNoopShuffleMask(Mask))
22631 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22633 // Look for simplifications involving one or two shuffle instructions.
22634 SDValue V = N.getOperand(0);
22635 switch (N.getOpcode()) {
22638 case X86ISD::PSHUFLW:
22639 case X86ISD::PSHUFHW:
22640 assert(VT == MVT::v8i16);
22643 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22644 return SDValue(); // We combined away this shuffle, so we're done.
22646 // See if this reduces to a PSHUFD which is no more expensive and can
22647 // combine with more operations. Note that it has to at least flip the
22648 // dwords as otherwise it would have been removed as a no-op.
22649 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
22650 int DMask[] = {0, 1, 2, 3};
22651 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22652 DMask[DOffset + 0] = DOffset + 1;
22653 DMask[DOffset + 1] = DOffset + 0;
22654 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
22655 DCI.AddToWorklist(V.getNode());
22656 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
22657 getV4X86ShuffleImm8ForMask(DMask, DAG));
22658 DCI.AddToWorklist(V.getNode());
22659 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
22662 // Look for shuffle patterns which can be implemented as a single unpack.
22663 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22664 // only works when we have a PSHUFD followed by two half-shuffles.
22665 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22666 (V.getOpcode() == X86ISD::PSHUFLW ||
22667 V.getOpcode() == X86ISD::PSHUFHW) &&
22668 V.getOpcode() != N.getOpcode() &&
22670 SDValue D = V.getOperand(0);
22671 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22672 D = D.getOperand(0);
22673 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22674 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22675 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22676 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22677 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22679 for (int i = 0; i < 4; ++i) {
22680 WordMask[i + NOffset] = Mask[i] + NOffset;
22681 WordMask[i + VOffset] = VMask[i] + VOffset;
22683 // Map the word mask through the DWord mask.
22685 for (int i = 0; i < 8; ++i)
22686 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22687 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
22688 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
22689 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
22690 std::begin(UnpackLoMask)) ||
22691 std::equal(std::begin(MappedMask), std::end(MappedMask),
22692 std::begin(UnpackHiMask))) {
22693 // We can replace all three shuffles with an unpack.
22694 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
22695 DCI.AddToWorklist(V.getNode());
22696 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22698 DL, MVT::v8i16, V, V);
22705 case X86ISD::PSHUFD:
22706 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22715 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22717 /// We combine this directly on the abstract vector shuffle nodes so it is
22718 /// easier to generically match. We also insert dummy vector shuffle nodes for
22719 /// the operands which explicitly discard the lanes which are unused by this
22720 /// operation to try to flow through the rest of the combiner the fact that
22721 /// they're unused.
22722 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22724 EVT VT = N->getValueType(0);
22726 // We only handle target-independent shuffles.
22727 // FIXME: It would be easy and harmless to use the target shuffle mask
22728 // extraction tool to support more.
22729 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22732 auto *SVN = cast<ShuffleVectorSDNode>(N);
22733 ArrayRef<int> Mask = SVN->getMask();
22734 SDValue V1 = N->getOperand(0);
22735 SDValue V2 = N->getOperand(1);
22737 // We require the first shuffle operand to be the SUB node, and the second to
22738 // be the ADD node.
22739 // FIXME: We should support the commuted patterns.
22740 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22743 // If there are other uses of these operations we can't fold them.
22744 if (!V1->hasOneUse() || !V2->hasOneUse())
22747 // Ensure that both operations have the same operands. Note that we can
22748 // commute the FADD operands.
22749 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22750 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22751 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22754 // We're looking for blends between FADD and FSUB nodes. We insist on these
22755 // nodes being lined up in a specific expected pattern.
22756 if (!(isShuffleEquivalent(Mask, 0, 3) ||
22757 isShuffleEquivalent(Mask, 0, 5, 2, 7) ||
22758 isShuffleEquivalent(Mask, 0, 9, 2, 11, 4, 13, 6, 15)))
22761 // Only specific types are legal at this point, assert so we notice if and
22762 // when these change.
22763 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22764 VT == MVT::v4f64) &&
22765 "Unknown vector type encountered!");
22767 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22770 /// PerformShuffleCombine - Performs several different shuffle combines.
22771 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22772 TargetLowering::DAGCombinerInfo &DCI,
22773 const X86Subtarget *Subtarget) {
22775 SDValue N0 = N->getOperand(0);
22776 SDValue N1 = N->getOperand(1);
22777 EVT VT = N->getValueType(0);
22779 // Don't create instructions with illegal types after legalize types has run.
22780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22781 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22784 // If we have legalized the vector types, look for blends of FADD and FSUB
22785 // nodes that we can fuse into an ADDSUB node.
22786 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22787 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22790 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22791 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22792 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22793 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22795 // During Type Legalization, when promoting illegal vector types,
22796 // the backend might introduce new shuffle dag nodes and bitcasts.
22798 // This code performs the following transformation:
22799 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22800 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22802 // We do this only if both the bitcast and the BINOP dag nodes have
22803 // one use. Also, perform this transformation only if the new binary
22804 // operation is legal. This is to avoid introducing dag nodes that
22805 // potentially need to be further expanded (or custom lowered) into a
22806 // less optimal sequence of dag nodes.
22807 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22808 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22809 N0.getOpcode() == ISD::BITCAST) {
22810 SDValue BC0 = N0.getOperand(0);
22811 EVT SVT = BC0.getValueType();
22812 unsigned Opcode = BC0.getOpcode();
22813 unsigned NumElts = VT.getVectorNumElements();
22815 if (BC0.hasOneUse() && SVT.isVector() &&
22816 SVT.getVectorNumElements() * 2 == NumElts &&
22817 TLI.isOperationLegal(Opcode, VT)) {
22818 bool CanFold = false;
22830 unsigned SVTNumElts = SVT.getVectorNumElements();
22831 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22832 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22833 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22834 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22835 CanFold = SVOp->getMaskElt(i) < 0;
22838 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
22839 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
22840 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22841 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22846 // Only handle 128 wide vector from here on.
22847 if (!VT.is128BitVector())
22850 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22851 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22852 // consecutive, non-overlapping, and in the right order.
22853 SmallVector<SDValue, 16> Elts;
22854 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22855 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22857 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
22861 if (isTargetShuffle(N->getOpcode())) {
22863 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22864 if (Shuffle.getNode())
22867 // Try recursively combining arbitrary sequences of x86 shuffle
22868 // instructions into higher-order shuffles. We do this after combining
22869 // specific PSHUF instruction sequences into their minimal form so that we
22870 // can evaluate how many specialized shuffle instructions are involved in
22871 // a particular chain.
22872 SmallVector<int, 1> NonceMask; // Just a placeholder.
22873 NonceMask.push_back(0);
22874 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22875 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22877 return SDValue(); // This routine will use CombineTo to replace N.
22883 /// PerformTruncateCombine - Converts truncate operation to
22884 /// a sequence of vector shuffle operations.
22885 /// It is possible when we truncate 256-bit vector to 128-bit vector
22886 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
22887 TargetLowering::DAGCombinerInfo &DCI,
22888 const X86Subtarget *Subtarget) {
22892 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22893 /// specific shuffle of a load can be folded into a single element load.
22894 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22895 /// shuffles have been custom lowered so we need to handle those here.
22896 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22897 TargetLowering::DAGCombinerInfo &DCI) {
22898 if (DCI.isBeforeLegalizeOps())
22901 SDValue InVec = N->getOperand(0);
22902 SDValue EltNo = N->getOperand(1);
22904 if (!isa<ConstantSDNode>(EltNo))
22907 EVT OriginalVT = InVec.getValueType();
22909 if (InVec.getOpcode() == ISD::BITCAST) {
22910 // Don't duplicate a load with other uses.
22911 if (!InVec.hasOneUse())
22913 EVT BCVT = InVec.getOperand(0).getValueType();
22914 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22916 InVec = InVec.getOperand(0);
22919 EVT CurrentVT = InVec.getValueType();
22921 if (!isTargetShuffle(InVec.getOpcode()))
22924 // Don't duplicate a load with other uses.
22925 if (!InVec.hasOneUse())
22928 SmallVector<int, 16> ShuffleMask;
22930 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22931 ShuffleMask, UnaryShuffle))
22934 // Select the input vector, guarding against out of range extract vector.
22935 unsigned NumElems = CurrentVT.getVectorNumElements();
22936 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22937 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22938 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22939 : InVec.getOperand(1);
22941 // If inputs to shuffle are the same for both ops, then allow 2 uses
22942 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22943 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22945 if (LdNode.getOpcode() == ISD::BITCAST) {
22946 // Don't duplicate a load with other uses.
22947 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22950 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22951 LdNode = LdNode.getOperand(0);
22954 if (!ISD::isNormalLoad(LdNode.getNode()))
22957 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22959 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22962 EVT EltVT = N->getValueType(0);
22963 // If there's a bitcast before the shuffle, check if the load type and
22964 // alignment is valid.
22965 unsigned Align = LN0->getAlignment();
22966 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22967 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
22968 EltVT.getTypeForEVT(*DAG.getContext()));
22970 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22973 // All checks match so transform back to vector_shuffle so that DAG combiner
22974 // can finish the job
22977 // Create shuffle node taking into account the case that its a unary shuffle
22978 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22979 : InVec.getOperand(1);
22980 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22981 InVec.getOperand(0), Shuffle,
22983 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
22984 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22988 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22989 /// generation and convert it from being a bunch of shuffles and extracts
22990 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22991 /// storing the value and loading scalars back, while for x64 we should
22992 /// use 64-bit extracts and shifts.
22993 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22994 TargetLowering::DAGCombinerInfo &DCI) {
22995 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
22996 if (NewOp.getNode())
22999 SDValue InputVector = N->getOperand(0);
23001 // Detect mmx to i32 conversion through a v2i32 elt extract.
23002 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23003 N->getValueType(0) == MVT::i32 &&
23004 InputVector.getValueType() == MVT::v2i32) {
23006 // The bitcast source is a direct mmx result.
23007 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23008 if (MMXSrc.getValueType() == MVT::x86mmx)
23009 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23010 N->getValueType(0),
23011 InputVector.getNode()->getOperand(0));
23013 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23014 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23015 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23016 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
23017 MMXSrcOp.getOpcode() == ISD::BITCAST &&
23018 MMXSrcOp.getValueType() == MVT::v1i64 &&
23019 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23020 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23021 N->getValueType(0),
23022 MMXSrcOp.getOperand(0));
23025 // Only operate on vectors of 4 elements, where the alternative shuffling
23026 // gets to be more expensive.
23027 if (InputVector.getValueType() != MVT::v4i32)
23030 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23031 // single use which is a sign-extend or zero-extend, and all elements are
23033 SmallVector<SDNode *, 4> Uses;
23034 unsigned ExtractedElements = 0;
23035 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23036 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23037 if (UI.getUse().getResNo() != InputVector.getResNo())
23040 SDNode *Extract = *UI;
23041 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23044 if (Extract->getValueType(0) != MVT::i32)
23046 if (!Extract->hasOneUse())
23048 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23049 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23051 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23054 // Record which element was extracted.
23055 ExtractedElements |=
23056 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23058 Uses.push_back(Extract);
23061 // If not all the elements were used, this may not be worthwhile.
23062 if (ExtractedElements != 15)
23065 // Ok, we've now decided to do the transformation.
23066 // If 64-bit shifts are legal, use the extract-shift sequence,
23067 // otherwise bounce the vector off the cache.
23068 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23070 SDLoc dl(InputVector);
23072 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23073 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
23074 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
23075 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23076 DAG.getConstant(0, VecIdxTy));
23077 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23078 DAG.getConstant(1, VecIdxTy));
23080 SDValue ShAmt = DAG.getConstant(32,
23081 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
23082 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23083 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23084 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23085 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23086 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23087 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23089 // Store the value to a temporary stack slot.
23090 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23091 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23092 MachinePointerInfo(), false, false, 0);
23094 EVT ElementType = InputVector.getValueType().getVectorElementType();
23095 unsigned EltSize = ElementType.getSizeInBits() / 8;
23097 // Replace each use (extract) with a load of the appropriate element.
23098 for (unsigned i = 0; i < 4; ++i) {
23099 uint64_t Offset = EltSize * i;
23100 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
23102 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
23103 StackPtr, OffsetVal);
23105 // Load the scalar.
23106 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23107 ScalarAddr, MachinePointerInfo(),
23108 false, false, false, 0);
23113 // Replace the extracts
23114 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23115 UE = Uses.end(); UI != UE; ++UI) {
23116 SDNode *Extract = *UI;
23118 SDValue Idx = Extract->getOperand(1);
23119 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23120 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23123 // The replacement was made in place; don't return anything.
23127 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
23128 static std::pair<unsigned, bool>
23129 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
23130 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
23131 if (!VT.isVector())
23132 return std::make_pair(0, false);
23134 bool NeedSplit = false;
23135 switch (VT.getSimpleVT().SimpleTy) {
23136 default: return std::make_pair(0, false);
23139 if (!Subtarget->hasVLX())
23140 return std::make_pair(0, false);
23144 if (!Subtarget->hasBWI())
23145 return std::make_pair(0, false);
23149 if (!Subtarget->hasAVX512())
23150 return std::make_pair(0, false);
23155 if (!Subtarget->hasAVX2())
23157 if (!Subtarget->hasAVX())
23158 return std::make_pair(0, false);
23163 if (!Subtarget->hasSSE2())
23164 return std::make_pair(0, false);
23167 // SSE2 has only a small subset of the operations.
23168 bool hasUnsigned = Subtarget->hasSSE41() ||
23169 (Subtarget->hasSSE2() && VT == MVT::v16i8);
23170 bool hasSigned = Subtarget->hasSSE41() ||
23171 (Subtarget->hasSSE2() && VT == MVT::v8i16);
23173 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23176 // Check for x CC y ? x : y.
23177 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23178 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23183 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
23186 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
23189 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23192 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23194 // Check for x CC y ? y : x -- a min/max with reversed arms.
23195 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23196 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23201 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
23204 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
23207 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23210 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23214 return std::make_pair(Opc, NeedSplit);
23218 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23219 const X86Subtarget *Subtarget) {
23221 SDValue Cond = N->getOperand(0);
23222 SDValue LHS = N->getOperand(1);
23223 SDValue RHS = N->getOperand(2);
23225 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23226 SDValue CondSrc = Cond->getOperand(0);
23227 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23228 Cond = CondSrc->getOperand(0);
23231 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23234 // A vselect where all conditions and data are constants can be optimized into
23235 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23236 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23237 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23240 unsigned MaskValue = 0;
23241 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23244 MVT VT = N->getSimpleValueType(0);
23245 unsigned NumElems = VT.getVectorNumElements();
23246 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23247 for (unsigned i = 0; i < NumElems; ++i) {
23248 // Be sure we emit undef where we can.
23249 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23250 ShuffleMask[i] = -1;
23252 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23256 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23258 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23261 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23263 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23264 TargetLowering::DAGCombinerInfo &DCI,
23265 const X86Subtarget *Subtarget) {
23267 SDValue Cond = N->getOperand(0);
23268 // Get the LHS/RHS of the select.
23269 SDValue LHS = N->getOperand(1);
23270 SDValue RHS = N->getOperand(2);
23271 EVT VT = LHS.getValueType();
23272 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23274 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23275 // instructions match the semantics of the common C idiom x<y?x:y but not
23276 // x<=y?x:y, because of how they handle negative zero (which can be
23277 // ignored in unsafe-math mode).
23278 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23279 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23280 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23281 (Subtarget->hasSSE2() ||
23282 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23283 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23285 unsigned Opcode = 0;
23286 // Check for x CC y ? x : y.
23287 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23288 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23292 // Converting this to a min would handle NaNs incorrectly, and swapping
23293 // the operands would cause it to handle comparisons between positive
23294 // and negative zero incorrectly.
23295 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23296 if (!DAG.getTarget().Options.UnsafeFPMath &&
23297 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23299 std::swap(LHS, RHS);
23301 Opcode = X86ISD::FMIN;
23304 // Converting this to a min would handle comparisons between positive
23305 // and negative zero incorrectly.
23306 if (!DAG.getTarget().Options.UnsafeFPMath &&
23307 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23309 Opcode = X86ISD::FMIN;
23312 // Converting this to a min would handle both negative zeros and NaNs
23313 // incorrectly, but we can swap the operands to fix both.
23314 std::swap(LHS, RHS);
23318 Opcode = X86ISD::FMIN;
23322 // Converting this to a max would handle comparisons between positive
23323 // and negative zero incorrectly.
23324 if (!DAG.getTarget().Options.UnsafeFPMath &&
23325 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23327 Opcode = X86ISD::FMAX;
23330 // Converting this to a max would handle NaNs incorrectly, and swapping
23331 // the operands would cause it to handle comparisons between positive
23332 // and negative zero incorrectly.
23333 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23334 if (!DAG.getTarget().Options.UnsafeFPMath &&
23335 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23337 std::swap(LHS, RHS);
23339 Opcode = X86ISD::FMAX;
23342 // Converting this to a max would handle both negative zeros and NaNs
23343 // incorrectly, but we can swap the operands to fix both.
23344 std::swap(LHS, RHS);
23348 Opcode = X86ISD::FMAX;
23351 // Check for x CC y ? y : x -- a min/max with reversed arms.
23352 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23353 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23357 // Converting this to a min would handle comparisons between positive
23358 // and negative zero incorrectly, and swapping the operands would
23359 // cause it to handle NaNs incorrectly.
23360 if (!DAG.getTarget().Options.UnsafeFPMath &&
23361 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23362 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23364 std::swap(LHS, RHS);
23366 Opcode = X86ISD::FMIN;
23369 // Converting this to a min would handle NaNs incorrectly.
23370 if (!DAG.getTarget().Options.UnsafeFPMath &&
23371 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23373 Opcode = X86ISD::FMIN;
23376 // Converting this to a min would handle both negative zeros and NaNs
23377 // incorrectly, but we can swap the operands to fix both.
23378 std::swap(LHS, RHS);
23382 Opcode = X86ISD::FMIN;
23386 // Converting this to a max would handle NaNs incorrectly.
23387 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23389 Opcode = X86ISD::FMAX;
23392 // Converting this to a max would handle comparisons between positive
23393 // and negative zero incorrectly, and swapping the operands would
23394 // cause it to handle NaNs incorrectly.
23395 if (!DAG.getTarget().Options.UnsafeFPMath &&
23396 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23397 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23399 std::swap(LHS, RHS);
23401 Opcode = X86ISD::FMAX;
23404 // Converting this to a max would handle both negative zeros and NaNs
23405 // incorrectly, but we can swap the operands to fix both.
23406 std::swap(LHS, RHS);
23410 Opcode = X86ISD::FMAX;
23416 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23419 EVT CondVT = Cond.getValueType();
23420 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23421 CondVT.getVectorElementType() == MVT::i1) {
23422 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23423 // lowering on KNL. In this case we convert it to
23424 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23425 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23426 // Since SKX these selects have a proper lowering.
23427 EVT OpVT = LHS.getValueType();
23428 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23429 (OpVT.getVectorElementType() == MVT::i8 ||
23430 OpVT.getVectorElementType() == MVT::i16) &&
23431 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23432 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23433 DCI.AddToWorklist(Cond.getNode());
23434 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23437 // If this is a select between two integer constants, try to do some
23439 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23440 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23441 // Don't do this for crazy integer types.
23442 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23443 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23444 // so that TrueC (the true value) is larger than FalseC.
23445 bool NeedsCondInvert = false;
23447 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23448 // Efficiently invertible.
23449 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23450 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23451 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23452 NeedsCondInvert = true;
23453 std::swap(TrueC, FalseC);
23456 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23457 if (FalseC->getAPIntValue() == 0 &&
23458 TrueC->getAPIntValue().isPowerOf2()) {
23459 if (NeedsCondInvert) // Invert the condition if needed.
23460 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23461 DAG.getConstant(1, Cond.getValueType()));
23463 // Zero extend the condition if needed.
23464 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23466 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23467 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23468 DAG.getConstant(ShAmt, MVT::i8));
23471 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23472 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23473 if (NeedsCondInvert) // Invert the condition if needed.
23474 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23475 DAG.getConstant(1, Cond.getValueType()));
23477 // Zero extend the condition if needed.
23478 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23479 FalseC->getValueType(0), Cond);
23480 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23481 SDValue(FalseC, 0));
23484 // Optimize cases that will turn into an LEA instruction. This requires
23485 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23486 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23487 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23488 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23490 bool isFastMultiplier = false;
23492 switch ((unsigned char)Diff) {
23494 case 1: // result = add base, cond
23495 case 2: // result = lea base( , cond*2)
23496 case 3: // result = lea base(cond, cond*2)
23497 case 4: // result = lea base( , cond*4)
23498 case 5: // result = lea base(cond, cond*4)
23499 case 8: // result = lea base( , cond*8)
23500 case 9: // result = lea base(cond, cond*8)
23501 isFastMultiplier = true;
23506 if (isFastMultiplier) {
23507 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23508 if (NeedsCondInvert) // Invert the condition if needed.
23509 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23510 DAG.getConstant(1, Cond.getValueType()));
23512 // Zero extend the condition if needed.
23513 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23515 // Scale the condition by the difference.
23517 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23518 DAG.getConstant(Diff, Cond.getValueType()));
23520 // Add the base if non-zero.
23521 if (FalseC->getAPIntValue() != 0)
23522 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23523 SDValue(FalseC, 0));
23530 // Canonicalize max and min:
23531 // (x > y) ? x : y -> (x >= y) ? x : y
23532 // (x < y) ? x : y -> (x <= y) ? x : y
23533 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23534 // the need for an extra compare
23535 // against zero. e.g.
23536 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23538 // testl %edi, %edi
23540 // cmovgl %edi, %eax
23544 // cmovsl %eax, %edi
23545 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23546 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23547 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23548 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23553 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23554 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23555 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23556 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23561 // Early exit check
23562 if (!TLI.isTypeLegal(VT))
23565 // Match VSELECTs into subs with unsigned saturation.
23566 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23567 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23568 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23569 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23570 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23572 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23573 // left side invert the predicate to simplify logic below.
23575 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23577 CC = ISD::getSetCCInverse(CC, true);
23578 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23582 if (Other.getNode() && Other->getNumOperands() == 2 &&
23583 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23584 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23585 SDValue CondRHS = Cond->getOperand(1);
23587 // Look for a general sub with unsigned saturation first.
23588 // x >= y ? x-y : 0 --> subus x, y
23589 // x > y ? x-y : 0 --> subus x, y
23590 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23591 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23592 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23594 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23595 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23596 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23597 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23598 // If the RHS is a constant we have to reverse the const
23599 // canonicalization.
23600 // x > C-1 ? x+-C : 0 --> subus x, C
23601 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23602 CondRHSConst->getAPIntValue() ==
23603 (-OpRHSConst->getAPIntValue() - 1))
23604 return DAG.getNode(
23605 X86ISD::SUBUS, DL, VT, OpLHS,
23606 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
23608 // Another special case: If C was a sign bit, the sub has been
23609 // canonicalized into a xor.
23610 // FIXME: Would it be better to use computeKnownBits to determine
23611 // whether it's safe to decanonicalize the xor?
23612 // x s< 0 ? x^C : 0 --> subus x, C
23613 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23614 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23615 OpRHSConst->getAPIntValue().isSignBit())
23616 // Note that we have to rebuild the RHS constant here to ensure we
23617 // don't rely on particular values of undef lanes.
23618 return DAG.getNode(
23619 X86ISD::SUBUS, DL, VT, OpLHS,
23620 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
23625 // Try to match a min/max vector operation.
23626 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
23627 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
23628 unsigned Opc = ret.first;
23629 bool NeedSplit = ret.second;
23631 if (Opc && NeedSplit) {
23632 unsigned NumElems = VT.getVectorNumElements();
23633 // Extract the LHS vectors
23634 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
23635 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
23637 // Extract the RHS vectors
23638 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
23639 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
23641 // Create min/max for each subvector
23642 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
23643 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
23645 // Merge the result
23646 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
23648 return DAG.getNode(Opc, DL, VT, LHS, RHS);
23651 // Simplify vector selection if condition value type matches vselect
23653 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23654 assert(Cond.getValueType().isVector() &&
23655 "vector select expects a vector selector!");
23657 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23658 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23660 // Try invert the condition if true value is not all 1s and false value
23662 if (!TValIsAllOnes && !FValIsAllZeros &&
23663 // Check if the selector will be produced by CMPP*/PCMP*
23664 Cond.getOpcode() == ISD::SETCC &&
23665 // Check if SETCC has already been promoted
23666 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
23667 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23668 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23670 if (TValIsAllZeros || FValIsAllOnes) {
23671 SDValue CC = Cond.getOperand(2);
23672 ISD::CondCode NewCC =
23673 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23674 Cond.getOperand(0).getValueType().isInteger());
23675 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23676 std::swap(LHS, RHS);
23677 TValIsAllOnes = FValIsAllOnes;
23678 FValIsAllZeros = TValIsAllZeros;
23682 if (TValIsAllOnes || FValIsAllZeros) {
23685 if (TValIsAllOnes && FValIsAllZeros)
23687 else if (TValIsAllOnes)
23688 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
23689 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
23690 else if (FValIsAllZeros)
23691 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23692 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
23694 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
23698 // If we know that this node is legal then we know that it is going to be
23699 // matched by one of the SSE/AVX BLEND instructions. These instructions only
23700 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
23701 // to simplify previous instructions.
23702 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23703 !DCI.isBeforeLegalize() &&
23704 // We explicitly check against v8i16 and v16i16 because, although
23705 // they're marked as Custom, they might only be legal when Cond is a
23706 // build_vector of constants. This will be taken care in a later
23708 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 &&
23709 VT != MVT::v8i16) &&
23710 // Don't optimize vector of constants. Those are handled by
23711 // the generic code and all the bits must be properly set for
23712 // the generic optimizer.
23713 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23714 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23716 // Don't optimize vector selects that map to mask-registers.
23720 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23721 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23723 APInt KnownZero, KnownOne;
23724 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23725 DCI.isBeforeLegalizeOps());
23726 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23727 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23729 // If we changed the computation somewhere in the DAG, this change
23730 // will affect all users of Cond.
23731 // Make sure it is fine and update all the nodes so that we do not
23732 // use the generic VSELECT anymore. Otherwise, we may perform
23733 // wrong optimizations as we messed up with the actual expectation
23734 // for the vector boolean values.
23735 if (Cond != TLO.Old) {
23736 // Check all uses of that condition operand to check whether it will be
23737 // consumed by non-BLEND instructions, which may depend on all bits are
23739 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23741 if (I->getOpcode() != ISD::VSELECT)
23742 // TODO: Add other opcodes eventually lowered into BLEND.
23745 // Update all the users of the condition, before committing the change,
23746 // so that the VSELECT optimizations that expect the correct vector
23747 // boolean value will not be triggered.
23748 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23750 DAG.ReplaceAllUsesOfValueWith(
23752 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23753 Cond, I->getOperand(1), I->getOperand(2)));
23754 DCI.CommitTargetLoweringOpt(TLO);
23757 // At this point, only Cond is changed. Change the condition
23758 // just for N to keep the opportunity to optimize all other
23759 // users their own way.
23760 DAG.ReplaceAllUsesOfValueWith(
23762 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23763 TLO.New, N->getOperand(1), N->getOperand(2)));
23768 // We should generate an X86ISD::BLENDI from a vselect if its argument
23769 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23770 // constants. This specific pattern gets generated when we split a
23771 // selector for a 512 bit vector in a machine without AVX512 (but with
23772 // 256-bit vectors), during legalization:
23774 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23776 // Iff we find this pattern and the build_vectors are built from
23777 // constants, we translate the vselect into a shuffle_vector that we
23778 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23779 if ((N->getOpcode() == ISD::VSELECT ||
23780 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23781 !DCI.isBeforeLegalize()) {
23782 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23783 if (Shuffle.getNode())
23790 // Check whether a boolean test is testing a boolean value generated by
23791 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23794 // Simplify the following patterns:
23795 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23796 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23797 // to (Op EFLAGS Cond)
23799 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23800 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23801 // to (Op EFLAGS !Cond)
23803 // where Op could be BRCOND or CMOV.
23805 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23806 // Quit if not CMP and SUB with its value result used.
23807 if (Cmp.getOpcode() != X86ISD::CMP &&
23808 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23811 // Quit if not used as a boolean value.
23812 if (CC != X86::COND_E && CC != X86::COND_NE)
23815 // Check CMP operands. One of them should be 0 or 1 and the other should be
23816 // an SetCC or extended from it.
23817 SDValue Op1 = Cmp.getOperand(0);
23818 SDValue Op2 = Cmp.getOperand(1);
23821 const ConstantSDNode* C = nullptr;
23822 bool needOppositeCond = (CC == X86::COND_E);
23823 bool checkAgainstTrue = false; // Is it a comparison against 1?
23825 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23827 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23829 else // Quit if all operands are not constants.
23832 if (C->getZExtValue() == 1) {
23833 needOppositeCond = !needOppositeCond;
23834 checkAgainstTrue = true;
23835 } else if (C->getZExtValue() != 0)
23836 // Quit if the constant is neither 0 or 1.
23839 bool truncatedToBoolWithAnd = false;
23840 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23841 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23842 SetCC.getOpcode() == ISD::TRUNCATE ||
23843 SetCC.getOpcode() == ISD::AND) {
23844 if (SetCC.getOpcode() == ISD::AND) {
23846 ConstantSDNode *CS;
23847 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23848 CS->getZExtValue() == 1)
23850 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23851 CS->getZExtValue() == 1)
23855 SetCC = SetCC.getOperand(OpIdx);
23856 truncatedToBoolWithAnd = true;
23858 SetCC = SetCC.getOperand(0);
23861 switch (SetCC.getOpcode()) {
23862 case X86ISD::SETCC_CARRY:
23863 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23864 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23865 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23866 // truncated to i1 using 'and'.
23867 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23869 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23870 "Invalid use of SETCC_CARRY!");
23872 case X86ISD::SETCC:
23873 // Set the condition code or opposite one if necessary.
23874 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23875 if (needOppositeCond)
23876 CC = X86::GetOppositeBranchCondition(CC);
23877 return SetCC.getOperand(1);
23878 case X86ISD::CMOV: {
23879 // Check whether false/true value has canonical one, i.e. 0 or 1.
23880 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23881 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23882 // Quit if true value is not a constant.
23885 // Quit if false value is not a constant.
23887 SDValue Op = SetCC.getOperand(0);
23888 // Skip 'zext' or 'trunc' node.
23889 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23890 Op.getOpcode() == ISD::TRUNCATE)
23891 Op = Op.getOperand(0);
23892 // A special case for rdrand/rdseed, where 0 is set if false cond is
23894 if ((Op.getOpcode() != X86ISD::RDRAND &&
23895 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23898 // Quit if false value is not the constant 0 or 1.
23899 bool FValIsFalse = true;
23900 if (FVal && FVal->getZExtValue() != 0) {
23901 if (FVal->getZExtValue() != 1)
23903 // If FVal is 1, opposite cond is needed.
23904 needOppositeCond = !needOppositeCond;
23905 FValIsFalse = false;
23907 // Quit if TVal is not the constant opposite of FVal.
23908 if (FValIsFalse && TVal->getZExtValue() != 1)
23910 if (!FValIsFalse && TVal->getZExtValue() != 0)
23912 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23913 if (needOppositeCond)
23914 CC = X86::GetOppositeBranchCondition(CC);
23915 return SetCC.getOperand(3);
23922 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23923 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23924 TargetLowering::DAGCombinerInfo &DCI,
23925 const X86Subtarget *Subtarget) {
23928 // If the flag operand isn't dead, don't touch this CMOV.
23929 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23932 SDValue FalseOp = N->getOperand(0);
23933 SDValue TrueOp = N->getOperand(1);
23934 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23935 SDValue Cond = N->getOperand(3);
23937 if (CC == X86::COND_E || CC == X86::COND_NE) {
23938 switch (Cond.getOpcode()) {
23942 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23943 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23944 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23950 Flags = checkBoolTestSetCCCombine(Cond, CC);
23951 if (Flags.getNode() &&
23952 // Extra check as FCMOV only supports a subset of X86 cond.
23953 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23954 SDValue Ops[] = { FalseOp, TrueOp,
23955 DAG.getConstant(CC, MVT::i8), Flags };
23956 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23959 // If this is a select between two integer constants, try to do some
23960 // optimizations. Note that the operands are ordered the opposite of SELECT
23962 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23963 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23964 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23965 // larger than FalseC (the false value).
23966 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23967 CC = X86::GetOppositeBranchCondition(CC);
23968 std::swap(TrueC, FalseC);
23969 std::swap(TrueOp, FalseOp);
23972 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23973 // This is efficient for any integer data type (including i8/i16) and
23975 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23976 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23977 DAG.getConstant(CC, MVT::i8), Cond);
23979 // Zero extend the condition if needed.
23980 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23982 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23983 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23984 DAG.getConstant(ShAmt, MVT::i8));
23985 if (N->getNumValues() == 2) // Dead flag value?
23986 return DCI.CombineTo(N, Cond, SDValue());
23990 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23991 // for any integer data type, including i8/i16.
23992 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23993 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23994 DAG.getConstant(CC, MVT::i8), Cond);
23996 // Zero extend the condition if needed.
23997 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23998 FalseC->getValueType(0), Cond);
23999 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24000 SDValue(FalseC, 0));
24002 if (N->getNumValues() == 2) // Dead flag value?
24003 return DCI.CombineTo(N, Cond, SDValue());
24007 // Optimize cases that will turn into an LEA instruction. This requires
24008 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24009 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24010 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24011 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24013 bool isFastMultiplier = false;
24015 switch ((unsigned char)Diff) {
24017 case 1: // result = add base, cond
24018 case 2: // result = lea base( , cond*2)
24019 case 3: // result = lea base(cond, cond*2)
24020 case 4: // result = lea base( , cond*4)
24021 case 5: // result = lea base(cond, cond*4)
24022 case 8: // result = lea base( , cond*8)
24023 case 9: // result = lea base(cond, cond*8)
24024 isFastMultiplier = true;
24029 if (isFastMultiplier) {
24030 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24031 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24032 DAG.getConstant(CC, MVT::i8), Cond);
24033 // Zero extend the condition if needed.
24034 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24036 // Scale the condition by the difference.
24038 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24039 DAG.getConstant(Diff, Cond.getValueType()));
24041 // Add the base if non-zero.
24042 if (FalseC->getAPIntValue() != 0)
24043 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24044 SDValue(FalseC, 0));
24045 if (N->getNumValues() == 2) // Dead flag value?
24046 return DCI.CombineTo(N, Cond, SDValue());
24053 // Handle these cases:
24054 // (select (x != c), e, c) -> select (x != c), e, x),
24055 // (select (x == c), c, e) -> select (x == c), x, e)
24056 // where the c is an integer constant, and the "select" is the combination
24057 // of CMOV and CMP.
24059 // The rationale for this change is that the conditional-move from a constant
24060 // needs two instructions, however, conditional-move from a register needs
24061 // only one instruction.
24063 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24064 // some instruction-combining opportunities. This opt needs to be
24065 // postponed as late as possible.
24067 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24068 // the DCI.xxxx conditions are provided to postpone the optimization as
24069 // late as possible.
24071 ConstantSDNode *CmpAgainst = nullptr;
24072 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24073 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24074 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24076 if (CC == X86::COND_NE &&
24077 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24078 CC = X86::GetOppositeBranchCondition(CC);
24079 std::swap(TrueOp, FalseOp);
24082 if (CC == X86::COND_E &&
24083 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24084 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24085 DAG.getConstant(CC, MVT::i8), Cond };
24086 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24094 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
24095 const X86Subtarget *Subtarget) {
24096 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
24098 default: return SDValue();
24099 // SSE/AVX/AVX2 blend intrinsics.
24100 case Intrinsic::x86_avx2_pblendvb:
24101 case Intrinsic::x86_avx2_pblendw:
24102 case Intrinsic::x86_avx2_pblendd_128:
24103 case Intrinsic::x86_avx2_pblendd_256:
24104 // Don't try to simplify this intrinsic if we don't have AVX2.
24105 if (!Subtarget->hasAVX2())
24108 case Intrinsic::x86_avx_blend_pd_256:
24109 case Intrinsic::x86_avx_blend_ps_256:
24110 case Intrinsic::x86_avx_blendv_pd_256:
24111 case Intrinsic::x86_avx_blendv_ps_256:
24112 // Don't try to simplify this intrinsic if we don't have AVX.
24113 if (!Subtarget->hasAVX())
24116 case Intrinsic::x86_sse41_pblendw:
24117 case Intrinsic::x86_sse41_blendpd:
24118 case Intrinsic::x86_sse41_blendps:
24119 case Intrinsic::x86_sse41_blendvps:
24120 case Intrinsic::x86_sse41_blendvpd:
24121 case Intrinsic::x86_sse41_pblendvb: {
24122 SDValue Op0 = N->getOperand(1);
24123 SDValue Op1 = N->getOperand(2);
24124 SDValue Mask = N->getOperand(3);
24126 // Don't try to simplify this intrinsic if we don't have SSE4.1.
24127 if (!Subtarget->hasSSE41())
24130 // fold (blend A, A, Mask) -> A
24133 // fold (blend A, B, allZeros) -> A
24134 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
24136 // fold (blend A, B, allOnes) -> B
24137 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
24140 // Simplify the case where the mask is a constant i32 value.
24141 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
24142 if (C->isNullValue())
24144 if (C->isAllOnesValue())
24151 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
24152 case Intrinsic::x86_sse2_psrai_w:
24153 case Intrinsic::x86_sse2_psrai_d:
24154 case Intrinsic::x86_avx2_psrai_w:
24155 case Intrinsic::x86_avx2_psrai_d:
24156 case Intrinsic::x86_sse2_psra_w:
24157 case Intrinsic::x86_sse2_psra_d:
24158 case Intrinsic::x86_avx2_psra_w:
24159 case Intrinsic::x86_avx2_psra_d: {
24160 SDValue Op0 = N->getOperand(1);
24161 SDValue Op1 = N->getOperand(2);
24162 EVT VT = Op0.getValueType();
24163 assert(VT.isVector() && "Expected a vector type!");
24165 if (isa<BuildVectorSDNode>(Op1))
24166 Op1 = Op1.getOperand(0);
24168 if (!isa<ConstantSDNode>(Op1))
24171 EVT SVT = VT.getVectorElementType();
24172 unsigned SVTBits = SVT.getSizeInBits();
24174 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
24175 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
24176 uint64_t ShAmt = C.getZExtValue();
24178 // Don't try to convert this shift into a ISD::SRA if the shift
24179 // count is bigger than or equal to the element size.
24180 if (ShAmt >= SVTBits)
24183 // Trivial case: if the shift count is zero, then fold this
24184 // into the first operand.
24188 // Replace this packed shift intrinsic with a target independent
24190 SDValue Splat = DAG.getConstant(C, VT);
24191 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
24196 /// PerformMulCombine - Optimize a single multiply with constant into two
24197 /// in order to implement it with two cheaper instructions, e.g.
24198 /// LEA + SHL, LEA + LEA.
24199 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24200 TargetLowering::DAGCombinerInfo &DCI) {
24201 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24204 EVT VT = N->getValueType(0);
24205 if (VT != MVT::i64 && VT != MVT::i32)
24208 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24211 uint64_t MulAmt = C->getZExtValue();
24212 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24215 uint64_t MulAmt1 = 0;
24216 uint64_t MulAmt2 = 0;
24217 if ((MulAmt % 9) == 0) {
24219 MulAmt2 = MulAmt / 9;
24220 } else if ((MulAmt % 5) == 0) {
24222 MulAmt2 = MulAmt / 5;
24223 } else if ((MulAmt % 3) == 0) {
24225 MulAmt2 = MulAmt / 3;
24228 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24231 if (isPowerOf2_64(MulAmt2) &&
24232 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24233 // If second multiplifer is pow2, issue it first. We want the multiply by
24234 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24236 std::swap(MulAmt1, MulAmt2);
24239 if (isPowerOf2_64(MulAmt1))
24240 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24241 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
24243 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24244 DAG.getConstant(MulAmt1, VT));
24246 if (isPowerOf2_64(MulAmt2))
24247 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24248 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
24250 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24251 DAG.getConstant(MulAmt2, VT));
24253 // Do not add new nodes to DAG combiner worklist.
24254 DCI.CombineTo(N, NewMul, false);
24259 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24260 SDValue N0 = N->getOperand(0);
24261 SDValue N1 = N->getOperand(1);
24262 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24263 EVT VT = N0.getValueType();
24265 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24266 // since the result of setcc_c is all zero's or all ones.
24267 if (VT.isInteger() && !VT.isVector() &&
24268 N1C && N0.getOpcode() == ISD::AND &&
24269 N0.getOperand(1).getOpcode() == ISD::Constant) {
24270 SDValue N00 = N0.getOperand(0);
24271 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
24272 ((N00.getOpcode() == ISD::ANY_EXTEND ||
24273 N00.getOpcode() == ISD::ZERO_EXTEND) &&
24274 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
24275 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24276 APInt ShAmt = N1C->getAPIntValue();
24277 Mask = Mask.shl(ShAmt);
24279 return DAG.getNode(ISD::AND, SDLoc(N), VT,
24280 N00, DAG.getConstant(Mask, VT));
24284 // Hardware support for vector shifts is sparse which makes us scalarize the
24285 // vector operations in many cases. Also, on sandybridge ADD is faster than
24287 // (shl V, 1) -> add V,V
24288 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24289 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24290 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24291 // We shift all of the values by one. In many cases we do not have
24292 // hardware support for this operation. This is better expressed as an ADD
24294 if (N1SplatC->getZExtValue() == 1)
24295 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24301 /// \brief Returns a vector of 0s if the node in input is a vector logical
24302 /// shift by a constant amount which is known to be bigger than or equal
24303 /// to the vector element size in bits.
24304 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24305 const X86Subtarget *Subtarget) {
24306 EVT VT = N->getValueType(0);
24308 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24309 (!Subtarget->hasInt256() ||
24310 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24313 SDValue Amt = N->getOperand(1);
24315 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24316 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24317 APInt ShiftAmt = AmtSplat->getAPIntValue();
24318 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24320 // SSE2/AVX2 logical shifts always return a vector of 0s
24321 // if the shift amount is bigger than or equal to
24322 // the element size. The constant shift amount will be
24323 // encoded as a 8-bit immediate.
24324 if (ShiftAmt.trunc(8).uge(MaxAmount))
24325 return getZeroVector(VT, Subtarget, DAG, DL);
24331 /// PerformShiftCombine - Combine shifts.
24332 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24333 TargetLowering::DAGCombinerInfo &DCI,
24334 const X86Subtarget *Subtarget) {
24335 if (N->getOpcode() == ISD::SHL) {
24336 SDValue V = PerformSHLCombine(N, DAG);
24337 if (V.getNode()) return V;
24340 if (N->getOpcode() != ISD::SRA) {
24341 // Try to fold this logical shift into a zero vector.
24342 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
24343 if (V.getNode()) return V;
24349 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24350 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24351 // and friends. Likewise for OR -> CMPNEQSS.
24352 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24353 TargetLowering::DAGCombinerInfo &DCI,
24354 const X86Subtarget *Subtarget) {
24357 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24358 // we're requiring SSE2 for both.
24359 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24360 SDValue N0 = N->getOperand(0);
24361 SDValue N1 = N->getOperand(1);
24362 SDValue CMP0 = N0->getOperand(1);
24363 SDValue CMP1 = N1->getOperand(1);
24366 // The SETCCs should both refer to the same CMP.
24367 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24370 SDValue CMP00 = CMP0->getOperand(0);
24371 SDValue CMP01 = CMP0->getOperand(1);
24372 EVT VT = CMP00.getValueType();
24374 if (VT == MVT::f32 || VT == MVT::f64) {
24375 bool ExpectingFlags = false;
24376 // Check for any users that want flags:
24377 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24378 !ExpectingFlags && UI != UE; ++UI)
24379 switch (UI->getOpcode()) {
24384 ExpectingFlags = true;
24386 case ISD::CopyToReg:
24387 case ISD::SIGN_EXTEND:
24388 case ISD::ZERO_EXTEND:
24389 case ISD::ANY_EXTEND:
24393 if (!ExpectingFlags) {
24394 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24395 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24397 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24398 X86::CondCode tmp = cc0;
24403 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24404 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24405 // FIXME: need symbolic constants for these magic numbers.
24406 // See X86ATTInstPrinter.cpp:printSSECC().
24407 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24408 if (Subtarget->hasAVX512()) {
24409 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24410 CMP01, DAG.getConstant(x86cc, MVT::i8));
24411 if (N->getValueType(0) != MVT::i1)
24412 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24416 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24417 CMP00.getValueType(), CMP00, CMP01,
24418 DAG.getConstant(x86cc, MVT::i8));
24420 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24421 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24423 if (is64BitFP && !Subtarget->is64Bit()) {
24424 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24425 // 64-bit integer, since that's not a legal type. Since
24426 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24427 // bits, but can do this little dance to extract the lowest 32 bits
24428 // and work with those going forward.
24429 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24431 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
24433 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24434 Vector32, DAG.getIntPtrConstant(0));
24438 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
24439 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24440 DAG.getConstant(1, IntVT));
24441 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
24442 return OneBitOfTruth;
24450 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24451 /// so it can be folded inside ANDNP.
24452 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24453 EVT VT = N->getValueType(0);
24455 // Match direct AllOnes for 128 and 256-bit vectors
24456 if (ISD::isBuildVectorAllOnes(N))
24459 // Look through a bit convert.
24460 if (N->getOpcode() == ISD::BITCAST)
24461 N = N->getOperand(0).getNode();
24463 // Sometimes the operand may come from a insert_subvector building a 256-bit
24465 if (VT.is256BitVector() &&
24466 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24467 SDValue V1 = N->getOperand(0);
24468 SDValue V2 = N->getOperand(1);
24470 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24471 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24472 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24473 ISD::isBuildVectorAllOnes(V2.getNode()))
24480 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24481 // register. In most cases we actually compare or select YMM-sized registers
24482 // and mixing the two types creates horrible code. This method optimizes
24483 // some of the transition sequences.
24484 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24485 TargetLowering::DAGCombinerInfo &DCI,
24486 const X86Subtarget *Subtarget) {
24487 EVT VT = N->getValueType(0);
24488 if (!VT.is256BitVector())
24491 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24492 N->getOpcode() == ISD::ZERO_EXTEND ||
24493 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24495 SDValue Narrow = N->getOperand(0);
24496 EVT NarrowVT = Narrow->getValueType(0);
24497 if (!NarrowVT.is128BitVector())
24500 if (Narrow->getOpcode() != ISD::XOR &&
24501 Narrow->getOpcode() != ISD::AND &&
24502 Narrow->getOpcode() != ISD::OR)
24505 SDValue N0 = Narrow->getOperand(0);
24506 SDValue N1 = Narrow->getOperand(1);
24509 // The Left side has to be a trunc.
24510 if (N0.getOpcode() != ISD::TRUNCATE)
24513 // The type of the truncated inputs.
24514 EVT WideVT = N0->getOperand(0)->getValueType(0);
24518 // The right side has to be a 'trunc' or a constant vector.
24519 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24520 ConstantSDNode *RHSConstSplat = nullptr;
24521 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24522 RHSConstSplat = RHSBV->getConstantSplatNode();
24523 if (!RHSTrunc && !RHSConstSplat)
24526 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24528 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24531 // Set N0 and N1 to hold the inputs to the new wide operation.
24532 N0 = N0->getOperand(0);
24533 if (RHSConstSplat) {
24534 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24535 SDValue(RHSConstSplat, 0));
24536 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24537 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24538 } else if (RHSTrunc) {
24539 N1 = N1->getOperand(0);
24542 // Generate the wide operation.
24543 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24544 unsigned Opcode = N->getOpcode();
24546 case ISD::ANY_EXTEND:
24548 case ISD::ZERO_EXTEND: {
24549 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24550 APInt Mask = APInt::getAllOnesValue(InBits);
24551 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24552 return DAG.getNode(ISD::AND, DL, VT,
24553 Op, DAG.getConstant(Mask, VT));
24555 case ISD::SIGN_EXTEND:
24556 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24557 Op, DAG.getValueType(NarrowVT));
24559 llvm_unreachable("Unexpected opcode");
24563 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24564 TargetLowering::DAGCombinerInfo &DCI,
24565 const X86Subtarget *Subtarget) {
24566 EVT VT = N->getValueType(0);
24567 if (DCI.isBeforeLegalizeOps())
24570 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
24574 // Create BEXTR instructions
24575 // BEXTR is ((X >> imm) & (2**size-1))
24576 if (VT == MVT::i32 || VT == MVT::i64) {
24577 SDValue N0 = N->getOperand(0);
24578 SDValue N1 = N->getOperand(1);
24581 // Check for BEXTR.
24582 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24583 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24584 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24585 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24586 if (MaskNode && ShiftNode) {
24587 uint64_t Mask = MaskNode->getZExtValue();
24588 uint64_t Shift = ShiftNode->getZExtValue();
24589 if (isMask_64(Mask)) {
24590 uint64_t MaskSize = CountPopulation_64(Mask);
24591 if (Shift + MaskSize <= VT.getSizeInBits())
24592 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24593 DAG.getConstant(Shift | (MaskSize << 8), VT));
24601 // Want to form ANDNP nodes:
24602 // 1) In the hopes of then easily combining them with OR and AND nodes
24603 // to form PBLEND/PSIGN.
24604 // 2) To match ANDN packed intrinsics
24605 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24608 SDValue N0 = N->getOperand(0);
24609 SDValue N1 = N->getOperand(1);
24612 // Check LHS for vnot
24613 if (N0.getOpcode() == ISD::XOR &&
24614 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24615 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24616 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24618 // Check RHS for vnot
24619 if (N1.getOpcode() == ISD::XOR &&
24620 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24621 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24622 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24627 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24628 TargetLowering::DAGCombinerInfo &DCI,
24629 const X86Subtarget *Subtarget) {
24630 if (DCI.isBeforeLegalizeOps())
24633 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
24637 SDValue N0 = N->getOperand(0);
24638 SDValue N1 = N->getOperand(1);
24639 EVT VT = N->getValueType(0);
24641 // look for psign/blend
24642 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24643 if (!Subtarget->hasSSSE3() ||
24644 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24647 // Canonicalize pandn to RHS
24648 if (N0.getOpcode() == X86ISD::ANDNP)
24650 // or (and (m, y), (pandn m, x))
24651 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24652 SDValue Mask = N1.getOperand(0);
24653 SDValue X = N1.getOperand(1);
24655 if (N0.getOperand(0) == Mask)
24656 Y = N0.getOperand(1);
24657 if (N0.getOperand(1) == Mask)
24658 Y = N0.getOperand(0);
24660 // Check to see if the mask appeared in both the AND and ANDNP and
24664 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24665 // Look through mask bitcast.
24666 if (Mask.getOpcode() == ISD::BITCAST)
24667 Mask = Mask.getOperand(0);
24668 if (X.getOpcode() == ISD::BITCAST)
24669 X = X.getOperand(0);
24670 if (Y.getOpcode() == ISD::BITCAST)
24671 Y = Y.getOperand(0);
24673 EVT MaskVT = Mask.getValueType();
24675 // Validate that the Mask operand is a vector sra node.
24676 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24677 // there is no psrai.b
24678 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24679 unsigned SraAmt = ~0;
24680 if (Mask.getOpcode() == ISD::SRA) {
24681 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24682 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24683 SraAmt = AmtConst->getZExtValue();
24684 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24685 SDValue SraC = Mask.getOperand(1);
24686 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24688 if ((SraAmt + 1) != EltBits)
24693 // Now we know we at least have a plendvb with the mask val. See if
24694 // we can form a psignb/w/d.
24695 // psign = x.type == y.type == mask.type && y = sub(0, x);
24696 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24697 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24698 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24699 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24700 "Unsupported VT for PSIGN");
24701 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24702 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
24704 // PBLENDVB only available on SSE 4.1
24705 if (!Subtarget->hasSSE41())
24708 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24710 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
24711 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
24712 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
24713 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24714 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
24718 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24721 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24722 MachineFunction &MF = DAG.getMachineFunction();
24723 bool OptForSize = MF.getFunction()->getAttributes().
24724 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
24726 // SHLD/SHRD instructions have lower register pressure, but on some
24727 // platforms they have higher latency than the equivalent
24728 // series of shifts/or that would otherwise be generated.
24729 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24730 // have higher latencies and we are not optimizing for size.
24731 if (!OptForSize && Subtarget->isSHLDSlow())
24734 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24736 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24738 if (!N0.hasOneUse() || !N1.hasOneUse())
24741 SDValue ShAmt0 = N0.getOperand(1);
24742 if (ShAmt0.getValueType() != MVT::i8)
24744 SDValue ShAmt1 = N1.getOperand(1);
24745 if (ShAmt1.getValueType() != MVT::i8)
24747 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24748 ShAmt0 = ShAmt0.getOperand(0);
24749 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24750 ShAmt1 = ShAmt1.getOperand(0);
24753 unsigned Opc = X86ISD::SHLD;
24754 SDValue Op0 = N0.getOperand(0);
24755 SDValue Op1 = N1.getOperand(0);
24756 if (ShAmt0.getOpcode() == ISD::SUB) {
24757 Opc = X86ISD::SHRD;
24758 std::swap(Op0, Op1);
24759 std::swap(ShAmt0, ShAmt1);
24762 unsigned Bits = VT.getSizeInBits();
24763 if (ShAmt1.getOpcode() == ISD::SUB) {
24764 SDValue Sum = ShAmt1.getOperand(0);
24765 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24766 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24767 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24768 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24769 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24770 return DAG.getNode(Opc, DL, VT,
24772 DAG.getNode(ISD::TRUNCATE, DL,
24775 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24776 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24778 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24779 return DAG.getNode(Opc, DL, VT,
24780 N0.getOperand(0), N1.getOperand(0),
24781 DAG.getNode(ISD::TRUNCATE, DL,
24788 // Generate NEG and CMOV for integer abs.
24789 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24790 EVT VT = N->getValueType(0);
24792 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24793 // 8-bit integer abs to NEG and CMOV.
24794 if (VT.isInteger() && VT.getSizeInBits() == 8)
24797 SDValue N0 = N->getOperand(0);
24798 SDValue N1 = N->getOperand(1);
24801 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24802 // and change it to SUB and CMOV.
24803 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24804 N0.getOpcode() == ISD::ADD &&
24805 N0.getOperand(1) == N1 &&
24806 N1.getOpcode() == ISD::SRA &&
24807 N1.getOperand(0) == N0.getOperand(0))
24808 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24809 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24810 // Generate SUB & CMOV.
24811 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24812 DAG.getConstant(0, VT), N0.getOperand(0));
24814 SDValue Ops[] = { N0.getOperand(0), Neg,
24815 DAG.getConstant(X86::COND_GE, MVT::i8),
24816 SDValue(Neg.getNode(), 1) };
24817 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24822 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
24823 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24824 TargetLowering::DAGCombinerInfo &DCI,
24825 const X86Subtarget *Subtarget) {
24826 if (DCI.isBeforeLegalizeOps())
24829 if (Subtarget->hasCMov()) {
24830 SDValue RV = performIntegerAbsCombine(N, DAG);
24838 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24839 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24840 TargetLowering::DAGCombinerInfo &DCI,
24841 const X86Subtarget *Subtarget) {
24842 LoadSDNode *Ld = cast<LoadSDNode>(N);
24843 EVT RegVT = Ld->getValueType(0);
24844 EVT MemVT = Ld->getMemoryVT();
24846 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24848 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24849 // into two 16-byte operations.
24850 ISD::LoadExtType Ext = Ld->getExtensionType();
24851 unsigned Alignment = Ld->getAlignment();
24852 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
24853 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24854 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
24855 unsigned NumElems = RegVT.getVectorNumElements();
24859 SDValue Ptr = Ld->getBasePtr();
24860 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
24862 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24864 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24865 Ld->getPointerInfo(), Ld->isVolatile(),
24866 Ld->isNonTemporal(), Ld->isInvariant(),
24868 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24869 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24870 Ld->getPointerInfo(), Ld->isVolatile(),
24871 Ld->isNonTemporal(), Ld->isInvariant(),
24872 std::min(16U, Alignment));
24873 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24875 Load2.getValue(1));
24877 SDValue NewVec = DAG.getUNDEF(RegVT);
24878 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24879 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24880 return DCI.CombineTo(N, NewVec, TF, true);
24886 /// PerformMLOADCombine - Resolve extending loads
24887 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24888 TargetLowering::DAGCombinerInfo &DCI,
24889 const X86Subtarget *Subtarget) {
24890 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24891 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24894 EVT VT = Mld->getValueType(0);
24895 unsigned NumElems = VT.getVectorNumElements();
24896 EVT LdVT = Mld->getMemoryVT();
24899 assert(LdVT != VT && "Cannot extend to the same type");
24900 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24901 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24902 // From, To sizes and ElemCount must be pow of two
24903 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24904 "Unexpected size for extending masked load");
24906 unsigned SizeRatio = ToSz / FromSz;
24907 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24909 // Create a type on which we perform the shuffle
24910 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24911 LdVT.getScalarType(), NumElems*SizeRatio);
24912 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24914 // Convert Src0 value
24915 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
24916 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24917 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24918 for (unsigned i = 0; i != NumElems; ++i)
24919 ShuffleVec[i] = i * SizeRatio;
24921 // Can't shuffle using an illegal type.
24922 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24923 && "WideVecVT should be legal");
24924 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24925 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24927 // Prepare the new mask
24929 SDValue Mask = Mld->getMask();
24930 if (Mask.getValueType() == VT) {
24931 // Mask and original value have the same type
24932 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
24933 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24934 for (unsigned i = 0; i != NumElems; ++i)
24935 ShuffleVec[i] = i * SizeRatio;
24936 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24937 ShuffleVec[i] = NumElems*SizeRatio;
24938 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24939 DAG.getConstant(0, WideVecVT),
24943 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24944 unsigned WidenNumElts = NumElems*SizeRatio;
24945 unsigned MaskNumElts = VT.getVectorNumElements();
24946 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24949 unsigned NumConcat = WidenNumElts / MaskNumElts;
24950 SmallVector<SDValue, 16> Ops(NumConcat);
24951 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
24953 for (unsigned i = 1; i != NumConcat; ++i)
24956 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24959 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24960 Mld->getBasePtr(), NewMask, WideSrc0,
24961 Mld->getMemoryVT(), Mld->getMemOperand(),
24963 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24964 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24967 /// PerformMSTORECombine - Resolve truncating stores
24968 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24969 const X86Subtarget *Subtarget) {
24970 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24971 if (!Mst->isTruncatingStore())
24974 EVT VT = Mst->getValue().getValueType();
24975 unsigned NumElems = VT.getVectorNumElements();
24976 EVT StVT = Mst->getMemoryVT();
24979 assert(StVT != VT && "Cannot truncate to the same type");
24980 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24981 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24983 // From, To sizes and ElemCount must be pow of two
24984 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24985 "Unexpected size for truncating masked store");
24986 // We are going to use the original vector elt for storing.
24987 // Accumulated smaller vector elements must be a multiple of the store size.
24988 assert (((NumElems * FromSz) % ToSz) == 0 &&
24989 "Unexpected ratio for truncating masked store");
24991 unsigned SizeRatio = FromSz / ToSz;
24992 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24994 // Create a type on which we perform the shuffle
24995 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24996 StVT.getScalarType(), NumElems*SizeRatio);
24998 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25000 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
25001 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25002 for (unsigned i = 0; i != NumElems; ++i)
25003 ShuffleVec[i] = i * SizeRatio;
25005 // Can't shuffle using an illegal type.
25006 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
25007 && "WideVecVT should be legal");
25009 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25010 DAG.getUNDEF(WideVecVT),
25014 SDValue Mask = Mst->getMask();
25015 if (Mask.getValueType() == VT) {
25016 // Mask and original value have the same type
25017 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
25018 for (unsigned i = 0; i != NumElems; ++i)
25019 ShuffleVec[i] = i * SizeRatio;
25020 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25021 ShuffleVec[i] = NumElems*SizeRatio;
25022 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25023 DAG.getConstant(0, WideVecVT),
25027 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25028 unsigned WidenNumElts = NumElems*SizeRatio;
25029 unsigned MaskNumElts = VT.getVectorNumElements();
25030 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25033 unsigned NumConcat = WidenNumElts / MaskNumElts;
25034 SmallVector<SDValue, 16> Ops(NumConcat);
25035 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
25037 for (unsigned i = 1; i != NumConcat; ++i)
25040 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25043 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25044 NewMask, StVT, Mst->getMemOperand(), false);
25046 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25047 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25048 const X86Subtarget *Subtarget) {
25049 StoreSDNode *St = cast<StoreSDNode>(N);
25050 EVT VT = St->getValue().getValueType();
25051 EVT StVT = St->getMemoryVT();
25053 SDValue StoredVal = St->getOperand(1);
25054 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25056 // If we are saving a concatenation of two XMM registers and 32-byte stores
25057 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25058 unsigned Alignment = St->getAlignment();
25059 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
25060 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
25061 StVT == VT && !IsAligned) {
25062 unsigned NumElems = VT.getVectorNumElements();
25066 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25067 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25069 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
25070 SDValue Ptr0 = St->getBasePtr();
25071 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25073 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25074 St->getPointerInfo(), St->isVolatile(),
25075 St->isNonTemporal(), Alignment);
25076 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25077 St->getPointerInfo(), St->isVolatile(),
25078 St->isNonTemporal(),
25079 std::min(16U, Alignment));
25080 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25083 // Optimize trunc store (of multiple scalars) to shuffle and store.
25084 // First, pack all of the elements in one place. Next, store to memory
25085 // in fewer chunks.
25086 if (St->isTruncatingStore() && VT.isVector()) {
25087 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25088 unsigned NumElems = VT.getVectorNumElements();
25089 assert(StVT != VT && "Cannot truncate to the same type");
25090 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25091 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25093 // From, To sizes and ElemCount must be pow of two
25094 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25095 // We are going to use the original vector elt for storing.
25096 // Accumulated smaller vector elements must be a multiple of the store size.
25097 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25099 unsigned SizeRatio = FromSz / ToSz;
25101 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25103 // Create a type on which we perform the shuffle
25104 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25105 StVT.getScalarType(), NumElems*SizeRatio);
25107 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25109 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
25110 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25111 for (unsigned i = 0; i != NumElems; ++i)
25112 ShuffleVec[i] = i * SizeRatio;
25114 // Can't shuffle using an illegal type.
25115 if (!TLI.isTypeLegal(WideVecVT))
25118 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25119 DAG.getUNDEF(WideVecVT),
25121 // At this point all of the data is stored at the bottom of the
25122 // register. We now need to save it to mem.
25124 // Find the largest store unit
25125 MVT StoreType = MVT::i8;
25126 for (MVT Tp : MVT::integer_valuetypes()) {
25127 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25131 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25132 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25133 (64 <= NumElems * ToSz))
25134 StoreType = MVT::f64;
25136 // Bitcast the original vector into a vector of store-size units
25137 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25138 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25139 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25140 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
25141 SmallVector<SDValue, 8> Chains;
25142 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
25143 TLI.getPointerTy());
25144 SDValue Ptr = St->getBasePtr();
25146 // Perform one or more big stores into memory.
25147 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25148 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25149 StoreType, ShuffWide,
25150 DAG.getIntPtrConstant(i));
25151 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25152 St->getPointerInfo(), St->isVolatile(),
25153 St->isNonTemporal(), St->getAlignment());
25154 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25155 Chains.push_back(Ch);
25158 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25161 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25162 // the FP state in cases where an emms may be missing.
25163 // A preferable solution to the general problem is to figure out the right
25164 // places to insert EMMS. This qualifies as a quick hack.
25166 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25167 if (VT.getSizeInBits() != 64)
25170 const Function *F = DAG.getMachineFunction().getFunction();
25171 bool NoImplicitFloatOps = F->getAttributes().
25172 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
25173 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
25174 && Subtarget->hasSSE2();
25175 if ((VT.isVector() ||
25176 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25177 isa<LoadSDNode>(St->getValue()) &&
25178 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25179 St->getChain().hasOneUse() && !St->isVolatile()) {
25180 SDNode* LdVal = St->getValue().getNode();
25181 LoadSDNode *Ld = nullptr;
25182 int TokenFactorIndex = -1;
25183 SmallVector<SDValue, 8> Ops;
25184 SDNode* ChainVal = St->getChain().getNode();
25185 // Must be a store of a load. We currently handle two cases: the load
25186 // is a direct child, and it's under an intervening TokenFactor. It is
25187 // possible to dig deeper under nested TokenFactors.
25188 if (ChainVal == LdVal)
25189 Ld = cast<LoadSDNode>(St->getChain());
25190 else if (St->getValue().hasOneUse() &&
25191 ChainVal->getOpcode() == ISD::TokenFactor) {
25192 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25193 if (ChainVal->getOperand(i).getNode() == LdVal) {
25194 TokenFactorIndex = i;
25195 Ld = cast<LoadSDNode>(St->getValue());
25197 Ops.push_back(ChainVal->getOperand(i));
25201 if (!Ld || !ISD::isNormalLoad(Ld))
25204 // If this is not the MMX case, i.e. we are just turning i64 load/store
25205 // into f64 load/store, avoid the transformation if there are multiple
25206 // uses of the loaded value.
25207 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25212 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25213 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25215 if (Subtarget->is64Bit() || F64IsLegal) {
25216 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25217 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25218 Ld->getPointerInfo(), Ld->isVolatile(),
25219 Ld->isNonTemporal(), Ld->isInvariant(),
25220 Ld->getAlignment());
25221 SDValue NewChain = NewLd.getValue(1);
25222 if (TokenFactorIndex != -1) {
25223 Ops.push_back(NewChain);
25224 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25226 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25227 St->getPointerInfo(),
25228 St->isVolatile(), St->isNonTemporal(),
25229 St->getAlignment());
25232 // Otherwise, lower to two pairs of 32-bit loads / stores.
25233 SDValue LoAddr = Ld->getBasePtr();
25234 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25235 DAG.getConstant(4, MVT::i32));
25237 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25238 Ld->getPointerInfo(),
25239 Ld->isVolatile(), Ld->isNonTemporal(),
25240 Ld->isInvariant(), Ld->getAlignment());
25241 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25242 Ld->getPointerInfo().getWithOffset(4),
25243 Ld->isVolatile(), Ld->isNonTemporal(),
25245 MinAlign(Ld->getAlignment(), 4));
25247 SDValue NewChain = LoLd.getValue(1);
25248 if (TokenFactorIndex != -1) {
25249 Ops.push_back(LoLd);
25250 Ops.push_back(HiLd);
25251 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25254 LoAddr = St->getBasePtr();
25255 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25256 DAG.getConstant(4, MVT::i32));
25258 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25259 St->getPointerInfo(),
25260 St->isVolatile(), St->isNonTemporal(),
25261 St->getAlignment());
25262 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25263 St->getPointerInfo().getWithOffset(4),
25265 St->isNonTemporal(),
25266 MinAlign(St->getAlignment(), 4));
25267 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25272 /// Return 'true' if this vector operation is "horizontal"
25273 /// and return the operands for the horizontal operation in LHS and RHS. A
25274 /// horizontal operation performs the binary operation on successive elements
25275 /// of its first operand, then on successive elements of its second operand,
25276 /// returning the resulting values in a vector. For example, if
25277 /// A = < float a0, float a1, float a2, float a3 >
25279 /// B = < float b0, float b1, float b2, float b3 >
25280 /// then the result of doing a horizontal operation on A and B is
25281 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25282 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25283 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25284 /// set to A, RHS to B, and the routine returns 'true'.
25285 /// Note that the binary operation should have the property that if one of the
25286 /// operands is UNDEF then the result is UNDEF.
25287 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25288 // Look for the following pattern: if
25289 // A = < float a0, float a1, float a2, float a3 >
25290 // B = < float b0, float b1, float b2, float b3 >
25292 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25293 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25294 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25295 // which is A horizontal-op B.
25297 // At least one of the operands should be a vector shuffle.
25298 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25299 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25302 MVT VT = LHS.getSimpleValueType();
25304 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25305 "Unsupported vector type for horizontal add/sub");
25307 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25308 // operate independently on 128-bit lanes.
25309 unsigned NumElts = VT.getVectorNumElements();
25310 unsigned NumLanes = VT.getSizeInBits()/128;
25311 unsigned NumLaneElts = NumElts / NumLanes;
25312 assert((NumLaneElts % 2 == 0) &&
25313 "Vector type should have an even number of elements in each lane");
25314 unsigned HalfLaneElts = NumLaneElts/2;
25316 // View LHS in the form
25317 // LHS = VECTOR_SHUFFLE A, B, LMask
25318 // If LHS is not a shuffle then pretend it is the shuffle
25319 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25320 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25323 SmallVector<int, 16> LMask(NumElts);
25324 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25325 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25326 A = LHS.getOperand(0);
25327 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25328 B = LHS.getOperand(1);
25329 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25330 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25332 if (LHS.getOpcode() != ISD::UNDEF)
25334 for (unsigned i = 0; i != NumElts; ++i)
25338 // Likewise, view RHS in the form
25339 // RHS = VECTOR_SHUFFLE C, D, RMask
25341 SmallVector<int, 16> RMask(NumElts);
25342 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25343 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25344 C = RHS.getOperand(0);
25345 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25346 D = RHS.getOperand(1);
25347 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25348 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25350 if (RHS.getOpcode() != ISD::UNDEF)
25352 for (unsigned i = 0; i != NumElts; ++i)
25356 // Check that the shuffles are both shuffling the same vectors.
25357 if (!(A == C && B == D) && !(A == D && B == C))
25360 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25361 if (!A.getNode() && !B.getNode())
25364 // If A and B occur in reverse order in RHS, then "swap" them (which means
25365 // rewriting the mask).
25367 CommuteVectorShuffleMask(RMask, NumElts);
25369 // At this point LHS and RHS are equivalent to
25370 // LHS = VECTOR_SHUFFLE A, B, LMask
25371 // RHS = VECTOR_SHUFFLE A, B, RMask
25372 // Check that the masks correspond to performing a horizontal operation.
25373 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25374 for (unsigned i = 0; i != NumLaneElts; ++i) {
25375 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25377 // Ignore any UNDEF components.
25378 if (LIdx < 0 || RIdx < 0 ||
25379 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25380 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25383 // Check that successive elements are being operated on. If not, this is
25384 // not a horizontal operation.
25385 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25386 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25387 if (!(LIdx == Index && RIdx == Index + 1) &&
25388 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25393 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25394 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25398 /// Do target-specific dag combines on floating point adds.
25399 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25400 const X86Subtarget *Subtarget) {
25401 EVT VT = N->getValueType(0);
25402 SDValue LHS = N->getOperand(0);
25403 SDValue RHS = N->getOperand(1);
25405 // Try to synthesize horizontal adds from adds of shuffles.
25406 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25407 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25408 isHorizontalBinOp(LHS, RHS, true))
25409 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25413 /// Do target-specific dag combines on floating point subs.
25414 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25415 const X86Subtarget *Subtarget) {
25416 EVT VT = N->getValueType(0);
25417 SDValue LHS = N->getOperand(0);
25418 SDValue RHS = N->getOperand(1);
25420 // Try to synthesize horizontal subs from subs of shuffles.
25421 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25422 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25423 isHorizontalBinOp(LHS, RHS, false))
25424 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25428 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25429 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
25430 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25431 // F[X]OR(0.0, x) -> x
25432 // F[X]OR(x, 0.0) -> x
25433 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25434 if (C->getValueAPF().isPosZero())
25435 return N->getOperand(1);
25436 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25437 if (C->getValueAPF().isPosZero())
25438 return N->getOperand(0);
25442 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25443 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25444 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25446 // Only perform optimizations if UnsafeMath is used.
25447 if (!DAG.getTarget().Options.UnsafeFPMath)
25450 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25451 // into FMINC and FMAXC, which are Commutative operations.
25452 unsigned NewOp = 0;
25453 switch (N->getOpcode()) {
25454 default: llvm_unreachable("unknown opcode");
25455 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25456 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25459 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25460 N->getOperand(0), N->getOperand(1));
25463 /// Do target-specific dag combines on X86ISD::FAND nodes.
25464 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25465 // FAND(0.0, x) -> 0.0
25466 // FAND(x, 0.0) -> 0.0
25467 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25468 if (C->getValueAPF().isPosZero())
25469 return N->getOperand(0);
25470 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25471 if (C->getValueAPF().isPosZero())
25472 return N->getOperand(1);
25476 /// Do target-specific dag combines on X86ISD::FANDN nodes
25477 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25478 // FANDN(x, 0.0) -> 0.0
25479 // FANDN(0.0, x) -> x
25480 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25481 if (C->getValueAPF().isPosZero())
25482 return N->getOperand(1);
25483 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25484 if (C->getValueAPF().isPosZero())
25485 return N->getOperand(1);
25489 static SDValue PerformBTCombine(SDNode *N,
25491 TargetLowering::DAGCombinerInfo &DCI) {
25492 // BT ignores high bits in the bit index operand.
25493 SDValue Op1 = N->getOperand(1);
25494 if (Op1.hasOneUse()) {
25495 unsigned BitWidth = Op1.getValueSizeInBits();
25496 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25497 APInt KnownZero, KnownOne;
25498 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25499 !DCI.isBeforeLegalizeOps());
25500 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25501 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25502 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25503 DCI.CommitTargetLoweringOpt(TLO);
25508 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25509 SDValue Op = N->getOperand(0);
25510 if (Op.getOpcode() == ISD::BITCAST)
25511 Op = Op.getOperand(0);
25512 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25513 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25514 VT.getVectorElementType().getSizeInBits() ==
25515 OpVT.getVectorElementType().getSizeInBits()) {
25516 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25521 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25522 const X86Subtarget *Subtarget) {
25523 EVT VT = N->getValueType(0);
25524 if (!VT.isVector())
25527 SDValue N0 = N->getOperand(0);
25528 SDValue N1 = N->getOperand(1);
25529 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25532 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25533 // both SSE and AVX2 since there is no sign-extended shift right
25534 // operation on a vector with 64-bit elements.
25535 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25536 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25537 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25538 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25539 SDValue N00 = N0.getOperand(0);
25541 // EXTLOAD has a better solution on AVX2,
25542 // it may be replaced with X86ISD::VSEXT node.
25543 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25544 if (!ISD::isNormalLoad(N00.getNode()))
25547 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25548 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25550 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25556 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25557 TargetLowering::DAGCombinerInfo &DCI,
25558 const X86Subtarget *Subtarget) {
25559 SDValue N0 = N->getOperand(0);
25560 EVT VT = N->getValueType(0);
25562 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25563 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25564 // This exposes the sext to the sdivrem lowering, so that it directly extends
25565 // from AH (which we otherwise need to do contortions to access).
25566 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25567 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
25569 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25570 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
25571 N0.getOperand(0), N0.getOperand(1));
25572 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25573 return R.getValue(1);
25576 if (!DCI.isBeforeLegalizeOps())
25579 if (!Subtarget->hasFp256())
25582 if (VT.isVector() && VT.getSizeInBits() == 256) {
25583 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
25591 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25592 const X86Subtarget* Subtarget) {
25594 EVT VT = N->getValueType(0);
25596 // Let legalize expand this if it isn't a legal type yet.
25597 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25600 EVT ScalarVT = VT.getScalarType();
25601 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25602 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
25605 SDValue A = N->getOperand(0);
25606 SDValue B = N->getOperand(1);
25607 SDValue C = N->getOperand(2);
25609 bool NegA = (A.getOpcode() == ISD::FNEG);
25610 bool NegB = (B.getOpcode() == ISD::FNEG);
25611 bool NegC = (C.getOpcode() == ISD::FNEG);
25613 // Negative multiplication when NegA xor NegB
25614 bool NegMul = (NegA != NegB);
25616 A = A.getOperand(0);
25618 B = B.getOperand(0);
25620 C = C.getOperand(0);
25624 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25626 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25628 return DAG.getNode(Opcode, dl, VT, A, B, C);
25631 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25632 TargetLowering::DAGCombinerInfo &DCI,
25633 const X86Subtarget *Subtarget) {
25634 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25635 // (and (i32 x86isd::setcc_carry), 1)
25636 // This eliminates the zext. This transformation is necessary because
25637 // ISD::SETCC is always legalized to i8.
25639 SDValue N0 = N->getOperand(0);
25640 EVT VT = N->getValueType(0);
25642 if (N0.getOpcode() == ISD::AND &&
25644 N0.getOperand(0).hasOneUse()) {
25645 SDValue N00 = N0.getOperand(0);
25646 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25647 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25648 if (!C || C->getZExtValue() != 1)
25650 return DAG.getNode(ISD::AND, dl, VT,
25651 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25652 N00.getOperand(0), N00.getOperand(1)),
25653 DAG.getConstant(1, VT));
25657 if (N0.getOpcode() == ISD::TRUNCATE &&
25659 N0.getOperand(0).hasOneUse()) {
25660 SDValue N00 = N0.getOperand(0);
25661 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25662 return DAG.getNode(ISD::AND, dl, VT,
25663 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25664 N00.getOperand(0), N00.getOperand(1)),
25665 DAG.getConstant(1, VT));
25668 if (VT.is256BitVector()) {
25669 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
25674 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25675 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25676 // This exposes the zext to the udivrem lowering, so that it directly extends
25677 // from AH (which we otherwise need to do contortions to access).
25678 if (N0.getOpcode() == ISD::UDIVREM &&
25679 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25680 (VT == MVT::i32 || VT == MVT::i64)) {
25681 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25682 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25683 N0.getOperand(0), N0.getOperand(1));
25684 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25685 return R.getValue(1);
25691 // Optimize x == -y --> x+y == 0
25692 // x != -y --> x+y != 0
25693 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25694 const X86Subtarget* Subtarget) {
25695 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25696 SDValue LHS = N->getOperand(0);
25697 SDValue RHS = N->getOperand(1);
25698 EVT VT = N->getValueType(0);
25701 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25702 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25703 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25704 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
25705 LHS.getValueType(), RHS, LHS.getOperand(1));
25706 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
25707 addV, DAG.getConstant(0, addV.getValueType()), CC);
25709 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25710 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25711 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25712 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
25713 RHS.getValueType(), LHS, RHS.getOperand(1));
25714 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
25715 addV, DAG.getConstant(0, addV.getValueType()), CC);
25718 if (VT.getScalarType() == MVT::i1) {
25719 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25720 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25721 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
25722 if (!IsSEXT0 && !IsVZero0)
25724 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
25725 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25726 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25728 if (!IsSEXT1 && !IsVZero1)
25731 if (IsSEXT0 && IsVZero1) {
25732 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
25733 if (CC == ISD::SETEQ)
25734 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25735 return LHS.getOperand(0);
25737 if (IsSEXT1 && IsVZero0) {
25738 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
25739 if (CC == ISD::SETEQ)
25740 return DAG.getNOT(DL, RHS.getOperand(0), VT);
25741 return RHS.getOperand(0);
25748 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25749 const X86Subtarget *Subtarget) {
25751 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25752 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25753 "X86insertps is only defined for v4x32");
25755 SDValue Ld = N->getOperand(1);
25756 if (MayFoldLoad(Ld)) {
25757 // Extract the countS bits from the immediate so we can get the proper
25758 // address when narrowing the vector load to a specific element.
25759 // When the second source op is a memory address, interps doesn't use
25760 // countS and just gets an f32 from that address.
25761 unsigned DestIndex =
25762 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25763 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25767 // Create this as a scalar to vector to match the instruction pattern.
25768 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25769 // countS bits are ignored when loading from memory on insertps, which
25770 // means we don't need to explicitly set them to 0.
25771 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25772 LoadScalarToVector, N->getOperand(2));
25775 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25776 // as "sbb reg,reg", since it can be extended without zext and produces
25777 // an all-ones bit which is more useful than 0/1 in some cases.
25778 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25781 return DAG.getNode(ISD::AND, DL, VT,
25782 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25783 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
25784 DAG.getConstant(1, VT));
25785 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25786 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25787 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25788 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
25791 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25792 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25793 TargetLowering::DAGCombinerInfo &DCI,
25794 const X86Subtarget *Subtarget) {
25796 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25797 SDValue EFLAGS = N->getOperand(1);
25799 if (CC == X86::COND_A) {
25800 // Try to convert COND_A into COND_B in an attempt to facilitate
25801 // materializing "setb reg".
25803 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25804 // cannot take an immediate as its first operand.
25806 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25807 EFLAGS.getValueType().isInteger() &&
25808 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25809 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25810 EFLAGS.getNode()->getVTList(),
25811 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25812 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25813 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25817 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25818 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25820 if (CC == X86::COND_B)
25821 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25825 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
25826 if (Flags.getNode()) {
25827 SDValue Cond = DAG.getConstant(CC, MVT::i8);
25828 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25834 // Optimize branch condition evaluation.
25836 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25837 TargetLowering::DAGCombinerInfo &DCI,
25838 const X86Subtarget *Subtarget) {
25840 SDValue Chain = N->getOperand(0);
25841 SDValue Dest = N->getOperand(1);
25842 SDValue EFLAGS = N->getOperand(3);
25843 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25847 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
25848 if (Flags.getNode()) {
25849 SDValue Cond = DAG.getConstant(CC, MVT::i8);
25850 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25857 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25858 SelectionDAG &DAG) {
25859 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25860 // optimize away operation when it's from a constant.
25862 // The general transformation is:
25863 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25864 // AND(VECTOR_CMP(x,y), constant2)
25865 // constant2 = UNARYOP(constant)
25867 // Early exit if this isn't a vector operation, the operand of the
25868 // unary operation isn't a bitwise AND, or if the sizes of the operations
25869 // aren't the same.
25870 EVT VT = N->getValueType(0);
25871 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25872 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25873 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25876 // Now check that the other operand of the AND is a constant. We could
25877 // make the transformation for non-constant splats as well, but it's unclear
25878 // that would be a benefit as it would not eliminate any operations, just
25879 // perform one more step in scalar code before moving to the vector unit.
25880 if (BuildVectorSDNode *BV =
25881 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25882 // Bail out if the vector isn't a constant.
25883 if (!BV->isConstant())
25886 // Everything checks out. Build up the new and improved node.
25888 EVT IntVT = BV->getValueType(0);
25889 // Create a new constant of the appropriate type for the transformed
25891 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25892 // The AND node needs bitcasts to/from an integer vector type around it.
25893 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
25894 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25895 N->getOperand(0)->getOperand(0), MaskConst);
25896 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
25903 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25904 const X86Subtarget *Subtarget) {
25905 // First try to optimize away the conversion entirely when it's
25906 // conditionally from a constant. Vectors only.
25907 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
25908 if (Res != SDValue())
25911 // Now move on to more general possibilities.
25912 SDValue Op0 = N->getOperand(0);
25913 EVT InVT = Op0->getValueType(0);
25915 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
25916 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
25918 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
25919 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25920 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
25923 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25924 // a 32-bit target where SSE doesn't support i64->FP operations.
25925 if (Op0.getOpcode() == ISD::LOAD) {
25926 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25927 EVT VT = Ld->getValueType(0);
25928 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
25929 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25930 !Subtarget->is64Bit() && VT == MVT::i64) {
25931 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25932 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
25933 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25940 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25941 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25942 X86TargetLowering::DAGCombinerInfo &DCI) {
25943 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25944 // the result is either zero or one (depending on the input carry bit).
25945 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25946 if (X86::isZeroNode(N->getOperand(0)) &&
25947 X86::isZeroNode(N->getOperand(1)) &&
25948 // We don't have a good way to replace an EFLAGS use, so only do this when
25950 SDValue(N, 1).use_empty()) {
25952 EVT VT = N->getValueType(0);
25953 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
25954 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25955 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25956 DAG.getConstant(X86::COND_B,MVT::i8),
25958 DAG.getConstant(1, VT));
25959 return DCI.CombineTo(N, Res1, CarryOut);
25965 // fold (add Y, (sete X, 0)) -> adc 0, Y
25966 // (add Y, (setne X, 0)) -> sbb -1, Y
25967 // (sub (sete X, 0), Y) -> sbb 0, Y
25968 // (sub (setne X, 0), Y) -> adc -1, Y
25969 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25972 // Look through ZExts.
25973 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25974 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25977 SDValue SetCC = Ext.getOperand(0);
25978 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25981 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25982 if (CC != X86::COND_E && CC != X86::COND_NE)
25985 SDValue Cmp = SetCC.getOperand(1);
25986 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25987 !X86::isZeroNode(Cmp.getOperand(1)) ||
25988 !Cmp.getOperand(0).getValueType().isInteger())
25991 SDValue CmpOp0 = Cmp.getOperand(0);
25992 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25993 DAG.getConstant(1, CmpOp0.getValueType()));
25995 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25996 if (CC == X86::COND_NE)
25997 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25998 DL, OtherVal.getValueType(), OtherVal,
25999 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
26000 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26001 DL, OtherVal.getValueType(), OtherVal,
26002 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
26005 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26006 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26007 const X86Subtarget *Subtarget) {
26008 EVT VT = N->getValueType(0);
26009 SDValue Op0 = N->getOperand(0);
26010 SDValue Op1 = N->getOperand(1);
26012 // Try to synthesize horizontal adds from adds of shuffles.
26013 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26014 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26015 isHorizontalBinOp(Op0, Op1, true))
26016 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26018 return OptimizeConditionalInDecrement(N, DAG);
26021 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26022 const X86Subtarget *Subtarget) {
26023 SDValue Op0 = N->getOperand(0);
26024 SDValue Op1 = N->getOperand(1);
26026 // X86 can't encode an immediate LHS of a sub. See if we can push the
26027 // negation into a preceding instruction.
26028 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26029 // If the RHS of the sub is a XOR with one use and a constant, invert the
26030 // immediate. Then add one to the LHS of the sub so we can turn
26031 // X-Y -> X+~Y+1, saving one register.
26032 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26033 isa<ConstantSDNode>(Op1.getOperand(1))) {
26034 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26035 EVT VT = Op0.getValueType();
26036 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26038 DAG.getConstant(~XorC, VT));
26039 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26040 DAG.getConstant(C->getAPIntValue()+1, VT));
26044 // Try to synthesize horizontal adds from adds of shuffles.
26045 EVT VT = N->getValueType(0);
26046 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26047 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26048 isHorizontalBinOp(Op0, Op1, true))
26049 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26051 return OptimizeConditionalInDecrement(N, DAG);
26054 /// performVZEXTCombine - Performs build vector combines
26055 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26056 TargetLowering::DAGCombinerInfo &DCI,
26057 const X86Subtarget *Subtarget) {
26059 MVT VT = N->getSimpleValueType(0);
26060 SDValue Op = N->getOperand(0);
26061 MVT OpVT = Op.getSimpleValueType();
26062 MVT OpEltVT = OpVT.getVectorElementType();
26063 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26065 // (vzext (bitcast (vzext (x)) -> (vzext x)
26067 while (V.getOpcode() == ISD::BITCAST)
26068 V = V.getOperand(0);
26070 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26071 MVT InnerVT = V.getSimpleValueType();
26072 MVT InnerEltVT = InnerVT.getVectorElementType();
26074 // If the element sizes match exactly, we can just do one larger vzext. This
26075 // is always an exact type match as vzext operates on integer types.
26076 if (OpEltVT == InnerEltVT) {
26077 assert(OpVT == InnerVT && "Types must match for vzext!");
26078 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26081 // The only other way we can combine them is if only a single element of the
26082 // inner vzext is used in the input to the outer vzext.
26083 if (InnerEltVT.getSizeInBits() < InputBits)
26086 // In this case, the inner vzext is completely dead because we're going to
26087 // only look at bits inside of the low element. Just do the outer vzext on
26088 // a bitcast of the input to the inner.
26089 return DAG.getNode(X86ISD::VZEXT, DL, VT,
26090 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
26093 // Check if we can bypass extracting and re-inserting an element of an input
26094 // vector. Essentialy:
26095 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26096 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26097 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26098 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26099 SDValue ExtractedV = V.getOperand(0);
26100 SDValue OrigV = ExtractedV.getOperand(0);
26101 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26102 if (ExtractIdx->getZExtValue() == 0) {
26103 MVT OrigVT = OrigV.getSimpleValueType();
26104 // Extract a subvector if necessary...
26105 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26106 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26107 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26108 OrigVT.getVectorNumElements() / Ratio);
26109 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26110 DAG.getIntPtrConstant(0));
26112 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
26113 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26120 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26121 DAGCombinerInfo &DCI) const {
26122 SelectionDAG &DAG = DCI.DAG;
26123 switch (N->getOpcode()) {
26125 case ISD::EXTRACT_VECTOR_ELT:
26126 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26129 case X86ISD::SHRUNKBLEND:
26130 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26131 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26132 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26133 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26134 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26135 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26138 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26139 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26140 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26141 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26142 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26143 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26144 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26145 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26146 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26147 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26148 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26150 case X86ISD::FOR: return PerformFORCombine(N, DAG);
26152 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26153 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26154 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26155 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26156 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26157 case ISD::ANY_EXTEND:
26158 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26159 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26160 case ISD::SIGN_EXTEND_INREG:
26161 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26162 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
26163 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26164 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26165 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26166 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26167 case X86ISD::SHUFP: // Handle all target specific shuffles
26168 case X86ISD::PALIGNR:
26169 case X86ISD::UNPCKH:
26170 case X86ISD::UNPCKL:
26171 case X86ISD::MOVHLPS:
26172 case X86ISD::MOVLHPS:
26173 case X86ISD::PSHUFB:
26174 case X86ISD::PSHUFD:
26175 case X86ISD::PSHUFHW:
26176 case X86ISD::PSHUFLW:
26177 case X86ISD::MOVSS:
26178 case X86ISD::MOVSD:
26179 case X86ISD::VPERMILPI:
26180 case X86ISD::VPERM2X128:
26181 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26182 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26183 case ISD::INTRINSIC_WO_CHAIN:
26184 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
26185 case X86ISD::INSERTPS: {
26186 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26187 return PerformINSERTPSCombine(N, DAG, Subtarget);
26190 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
26196 /// isTypeDesirableForOp - Return true if the target has native support for
26197 /// the specified value type and it is 'desirable' to use the type for the
26198 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26199 /// instruction encodings are longer and some i16 instructions are slow.
26200 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26201 if (!isTypeLegal(VT))
26203 if (VT != MVT::i16)
26210 case ISD::SIGN_EXTEND:
26211 case ISD::ZERO_EXTEND:
26212 case ISD::ANY_EXTEND:
26225 /// IsDesirableToPromoteOp - This method query the target whether it is
26226 /// beneficial for dag combiner to promote the specified node. If true, it
26227 /// should return the desired promotion type by reference.
26228 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26229 EVT VT = Op.getValueType();
26230 if (VT != MVT::i16)
26233 bool Promote = false;
26234 bool Commute = false;
26235 switch (Op.getOpcode()) {
26238 LoadSDNode *LD = cast<LoadSDNode>(Op);
26239 // If the non-extending load has a single use and it's not live out, then it
26240 // might be folded.
26241 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26242 Op.hasOneUse()*/) {
26243 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26244 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26245 // The only case where we'd want to promote LOAD (rather then it being
26246 // promoted as an operand is when it's only use is liveout.
26247 if (UI->getOpcode() != ISD::CopyToReg)
26254 case ISD::SIGN_EXTEND:
26255 case ISD::ZERO_EXTEND:
26256 case ISD::ANY_EXTEND:
26261 SDValue N0 = Op.getOperand(0);
26262 // Look out for (store (shl (load), x)).
26263 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26276 SDValue N0 = Op.getOperand(0);
26277 SDValue N1 = Op.getOperand(1);
26278 if (!Commute && MayFoldLoad(N1))
26280 // Avoid disabling potential load folding opportunities.
26281 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26283 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26293 //===----------------------------------------------------------------------===//
26294 // X86 Inline Assembly Support
26295 //===----------------------------------------------------------------------===//
26298 // Helper to match a string separated by whitespace.
26299 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
26300 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
26302 for (unsigned i = 0, e = args.size(); i != e; ++i) {
26303 StringRef piece(*args[i]);
26304 if (!s.startswith(piece)) // Check if the piece matches.
26307 s = s.substr(piece.size());
26308 StringRef::size_type pos = s.find_first_not_of(" \t");
26309 if (pos == 0) // We matched a prefix.
26317 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
26320 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26322 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26323 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26324 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26325 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26327 if (AsmPieces.size() == 3)
26329 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26336 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26337 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26339 std::string AsmStr = IA->getAsmString();
26341 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26342 if (!Ty || Ty->getBitWidth() % 16 != 0)
26345 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26346 SmallVector<StringRef, 4> AsmPieces;
26347 SplitString(AsmStr, AsmPieces, ";\n");
26349 switch (AsmPieces.size()) {
26350 default: return false;
26352 // FIXME: this should verify that we are targeting a 486 or better. If not,
26353 // we will turn this bswap into something that will be lowered to logical
26354 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26355 // lower so don't worry about this.
26357 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
26358 matchAsm(AsmPieces[0], "bswapl", "$0") ||
26359 matchAsm(AsmPieces[0], "bswapq", "$0") ||
26360 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
26361 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
26362 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
26363 // No need to check constraints, nothing other than the equivalent of
26364 // "=r,0" would be valid here.
26365 return IntrinsicLowering::LowerToByteSwap(CI);
26368 // rorw $$8, ${0:w} --> llvm.bswap.i16
26369 if (CI->getType()->isIntegerTy(16) &&
26370 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26371 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
26372 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
26374 const std::string &ConstraintsStr = IA->getConstraintString();
26375 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26376 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26377 if (clobbersFlagRegisters(AsmPieces))
26378 return IntrinsicLowering::LowerToByteSwap(CI);
26382 if (CI->getType()->isIntegerTy(32) &&
26383 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26384 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
26385 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
26386 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
26388 const std::string &ConstraintsStr = IA->getConstraintString();
26389 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26390 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26391 if (clobbersFlagRegisters(AsmPieces))
26392 return IntrinsicLowering::LowerToByteSwap(CI);
26395 if (CI->getType()->isIntegerTy(64)) {
26396 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26397 if (Constraints.size() >= 2 &&
26398 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26399 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26400 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26401 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
26402 matchAsm(AsmPieces[1], "bswap", "%edx") &&
26403 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
26404 return IntrinsicLowering::LowerToByteSwap(CI);
26412 /// getConstraintType - Given a constraint letter, return the type of
26413 /// constraint it is for this target.
26414 X86TargetLowering::ConstraintType
26415 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
26416 if (Constraint.size() == 1) {
26417 switch (Constraint[0]) {
26428 return C_RegisterClass;
26452 return TargetLowering::getConstraintType(Constraint);
26455 /// Examine constraint type and operand type and determine a weight value.
26456 /// This object must already have been set up with the operand type
26457 /// and the current alternative constraint selected.
26458 TargetLowering::ConstraintWeight
26459 X86TargetLowering::getSingleConstraintMatchWeight(
26460 AsmOperandInfo &info, const char *constraint) const {
26461 ConstraintWeight weight = CW_Invalid;
26462 Value *CallOperandVal = info.CallOperandVal;
26463 // If we don't have a value, we can't do a match,
26464 // but allow it at the lowest weight.
26465 if (!CallOperandVal)
26467 Type *type = CallOperandVal->getType();
26468 // Look at the constraint type.
26469 switch (*constraint) {
26471 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26482 if (CallOperandVal->getType()->isIntegerTy())
26483 weight = CW_SpecificReg;
26488 if (type->isFloatingPointTy())
26489 weight = CW_SpecificReg;
26492 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26493 weight = CW_SpecificReg;
26497 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26498 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26499 weight = CW_Register;
26502 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26503 if (C->getZExtValue() <= 31)
26504 weight = CW_Constant;
26508 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26509 if (C->getZExtValue() <= 63)
26510 weight = CW_Constant;
26514 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26515 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26516 weight = CW_Constant;
26520 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26521 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26522 weight = CW_Constant;
26526 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26527 if (C->getZExtValue() <= 3)
26528 weight = CW_Constant;
26532 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26533 if (C->getZExtValue() <= 0xff)
26534 weight = CW_Constant;
26539 if (dyn_cast<ConstantFP>(CallOperandVal)) {
26540 weight = CW_Constant;
26544 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26545 if ((C->getSExtValue() >= -0x80000000LL) &&
26546 (C->getSExtValue() <= 0x7fffffffLL))
26547 weight = CW_Constant;
26551 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26552 if (C->getZExtValue() <= 0xffffffff)
26553 weight = CW_Constant;
26560 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26561 /// with another that has more specific requirements based on the type of the
26562 /// corresponding operand.
26563 const char *X86TargetLowering::
26564 LowerXConstraint(EVT ConstraintVT) const {
26565 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26566 // 'f' like normal targets.
26567 if (ConstraintVT.isFloatingPoint()) {
26568 if (Subtarget->hasSSE2())
26570 if (Subtarget->hasSSE1())
26574 return TargetLowering::LowerXConstraint(ConstraintVT);
26577 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26578 /// vector. If it is invalid, don't add anything to Ops.
26579 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26580 std::string &Constraint,
26581 std::vector<SDValue>&Ops,
26582 SelectionDAG &DAG) const {
26585 // Only support length 1 constraints for now.
26586 if (Constraint.length() > 1) return;
26588 char ConstraintLetter = Constraint[0];
26589 switch (ConstraintLetter) {
26592 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26593 if (C->getZExtValue() <= 31) {
26594 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26600 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26601 if (C->getZExtValue() <= 63) {
26602 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26609 if (isInt<8>(C->getSExtValue())) {
26610 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26616 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26617 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26618 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26619 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
26625 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26626 if (C->getZExtValue() <= 3) {
26627 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26633 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26634 if (C->getZExtValue() <= 255) {
26635 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26641 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26642 if (C->getZExtValue() <= 127) {
26643 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26649 // 32-bit signed value
26650 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26651 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26652 C->getSExtValue())) {
26653 // Widen to 64 bits here to get it sign extended.
26654 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
26657 // FIXME gcc accepts some relocatable values here too, but only in certain
26658 // memory models; it's complicated.
26663 // 32-bit unsigned value
26664 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26665 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26666 C->getZExtValue())) {
26667 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26671 // FIXME gcc accepts some relocatable values here too, but only in certain
26672 // memory models; it's complicated.
26676 // Literal immediates are always ok.
26677 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26678 // Widen to 64 bits here to get it sign extended.
26679 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
26683 // In any sort of PIC mode addresses need to be computed at runtime by
26684 // adding in a register or some sort of table lookup. These can't
26685 // be used as immediates.
26686 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26689 // If we are in non-pic codegen mode, we allow the address of a global (with
26690 // an optional displacement) to be used with 'i'.
26691 GlobalAddressSDNode *GA = nullptr;
26692 int64_t Offset = 0;
26694 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26696 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26697 Offset += GA->getOffset();
26699 } else if (Op.getOpcode() == ISD::ADD) {
26700 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26701 Offset += C->getZExtValue();
26702 Op = Op.getOperand(0);
26705 } else if (Op.getOpcode() == ISD::SUB) {
26706 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26707 Offset += -C->getZExtValue();
26708 Op = Op.getOperand(0);
26713 // Otherwise, this isn't something we can handle, reject it.
26717 const GlobalValue *GV = GA->getGlobal();
26718 // If we require an extra load to get this address, as in PIC mode, we
26719 // can't accept it.
26720 if (isGlobalStubReference(
26721 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26724 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26725 GA->getValueType(0), Offset);
26730 if (Result.getNode()) {
26731 Ops.push_back(Result);
26734 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26737 std::pair<unsigned, const TargetRegisterClass*>
26738 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
26740 // First, see if this is a constraint that directly corresponds to an LLVM
26742 if (Constraint.size() == 1) {
26743 // GCC Constraint Letters
26744 switch (Constraint[0]) {
26746 // TODO: Slight differences here in allocation order and leaving
26747 // RIP in the class. Do they matter any more here than they do
26748 // in the normal allocation?
26749 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26750 if (Subtarget->is64Bit()) {
26751 if (VT == MVT::i32 || VT == MVT::f32)
26752 return std::make_pair(0U, &X86::GR32RegClass);
26753 if (VT == MVT::i16)
26754 return std::make_pair(0U, &X86::GR16RegClass);
26755 if (VT == MVT::i8 || VT == MVT::i1)
26756 return std::make_pair(0U, &X86::GR8RegClass);
26757 if (VT == MVT::i64 || VT == MVT::f64)
26758 return std::make_pair(0U, &X86::GR64RegClass);
26761 // 32-bit fallthrough
26762 case 'Q': // Q_REGS
26763 if (VT == MVT::i32 || VT == MVT::f32)
26764 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26765 if (VT == MVT::i16)
26766 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26767 if (VT == MVT::i8 || VT == MVT::i1)
26768 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26769 if (VT == MVT::i64)
26770 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26772 case 'r': // GENERAL_REGS
26773 case 'l': // INDEX_REGS
26774 if (VT == MVT::i8 || VT == MVT::i1)
26775 return std::make_pair(0U, &X86::GR8RegClass);
26776 if (VT == MVT::i16)
26777 return std::make_pair(0U, &X86::GR16RegClass);
26778 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26779 return std::make_pair(0U, &X86::GR32RegClass);
26780 return std::make_pair(0U, &X86::GR64RegClass);
26781 case 'R': // LEGACY_REGS
26782 if (VT == MVT::i8 || VT == MVT::i1)
26783 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26784 if (VT == MVT::i16)
26785 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26786 if (VT == MVT::i32 || !Subtarget->is64Bit())
26787 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26788 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26789 case 'f': // FP Stack registers.
26790 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26791 // value to the correct fpstack register class.
26792 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26793 return std::make_pair(0U, &X86::RFP32RegClass);
26794 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26795 return std::make_pair(0U, &X86::RFP64RegClass);
26796 return std::make_pair(0U, &X86::RFP80RegClass);
26797 case 'y': // MMX_REGS if MMX allowed.
26798 if (!Subtarget->hasMMX()) break;
26799 return std::make_pair(0U, &X86::VR64RegClass);
26800 case 'Y': // SSE_REGS if SSE2 allowed
26801 if (!Subtarget->hasSSE2()) break;
26803 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26804 if (!Subtarget->hasSSE1()) break;
26806 switch (VT.SimpleTy) {
26808 // Scalar SSE types.
26811 return std::make_pair(0U, &X86::FR32RegClass);
26814 return std::make_pair(0U, &X86::FR64RegClass);
26822 return std::make_pair(0U, &X86::VR128RegClass);
26830 return std::make_pair(0U, &X86::VR256RegClass);
26835 return std::make_pair(0U, &X86::VR512RegClass);
26841 // Use the default implementation in TargetLowering to convert the register
26842 // constraint into a member of a register class.
26843 std::pair<unsigned, const TargetRegisterClass*> Res;
26844 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
26846 // Not found as a standard register?
26848 // Map st(0) -> st(7) -> ST0
26849 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26850 tolower(Constraint[1]) == 's' &&
26851 tolower(Constraint[2]) == 't' &&
26852 Constraint[3] == '(' &&
26853 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26854 Constraint[5] == ')' &&
26855 Constraint[6] == '}') {
26857 Res.first = X86::FP0+Constraint[4]-'0';
26858 Res.second = &X86::RFP80RegClass;
26862 // GCC allows "st(0)" to be called just plain "st".
26863 if (StringRef("{st}").equals_lower(Constraint)) {
26864 Res.first = X86::FP0;
26865 Res.second = &X86::RFP80RegClass;
26870 if (StringRef("{flags}").equals_lower(Constraint)) {
26871 Res.first = X86::EFLAGS;
26872 Res.second = &X86::CCRRegClass;
26876 // 'A' means EAX + EDX.
26877 if (Constraint == "A") {
26878 Res.first = X86::EAX;
26879 Res.second = &X86::GR32_ADRegClass;
26885 // Otherwise, check to see if this is a register class of the wrong value
26886 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26887 // turn into {ax},{dx}.
26888 if (Res.second->hasType(VT))
26889 return Res; // Correct type already, nothing to do.
26891 // All of the single-register GCC register classes map their values onto
26892 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
26893 // really want an 8-bit or 32-bit register, map to the appropriate register
26894 // class and return the appropriate register.
26895 if (Res.second == &X86::GR16RegClass) {
26896 if (VT == MVT::i8 || VT == MVT::i1) {
26897 unsigned DestReg = 0;
26898 switch (Res.first) {
26900 case X86::AX: DestReg = X86::AL; break;
26901 case X86::DX: DestReg = X86::DL; break;
26902 case X86::CX: DestReg = X86::CL; break;
26903 case X86::BX: DestReg = X86::BL; break;
26906 Res.first = DestReg;
26907 Res.second = &X86::GR8RegClass;
26909 } else if (VT == MVT::i32 || VT == MVT::f32) {
26910 unsigned DestReg = 0;
26911 switch (Res.first) {
26913 case X86::AX: DestReg = X86::EAX; break;
26914 case X86::DX: DestReg = X86::EDX; break;
26915 case X86::CX: DestReg = X86::ECX; break;
26916 case X86::BX: DestReg = X86::EBX; break;
26917 case X86::SI: DestReg = X86::ESI; break;
26918 case X86::DI: DestReg = X86::EDI; break;
26919 case X86::BP: DestReg = X86::EBP; break;
26920 case X86::SP: DestReg = X86::ESP; break;
26923 Res.first = DestReg;
26924 Res.second = &X86::GR32RegClass;
26926 } else if (VT == MVT::i64 || VT == MVT::f64) {
26927 unsigned DestReg = 0;
26928 switch (Res.first) {
26930 case X86::AX: DestReg = X86::RAX; break;
26931 case X86::DX: DestReg = X86::RDX; break;
26932 case X86::CX: DestReg = X86::RCX; break;
26933 case X86::BX: DestReg = X86::RBX; break;
26934 case X86::SI: DestReg = X86::RSI; break;
26935 case X86::DI: DestReg = X86::RDI; break;
26936 case X86::BP: DestReg = X86::RBP; break;
26937 case X86::SP: DestReg = X86::RSP; break;
26940 Res.first = DestReg;
26941 Res.second = &X86::GR64RegClass;
26944 } else if (Res.second == &X86::FR32RegClass ||
26945 Res.second == &X86::FR64RegClass ||
26946 Res.second == &X86::VR128RegClass ||
26947 Res.second == &X86::VR256RegClass ||
26948 Res.second == &X86::FR32XRegClass ||
26949 Res.second == &X86::FR64XRegClass ||
26950 Res.second == &X86::VR128XRegClass ||
26951 Res.second == &X86::VR256XRegClass ||
26952 Res.second == &X86::VR512RegClass) {
26953 // Handle references to XMM physical registers that got mapped into the
26954 // wrong class. This can happen with constraints like {xmm0} where the
26955 // target independent register mapper will just pick the first match it can
26956 // find, ignoring the required type.
26958 if (VT == MVT::f32 || VT == MVT::i32)
26959 Res.second = &X86::FR32RegClass;
26960 else if (VT == MVT::f64 || VT == MVT::i64)
26961 Res.second = &X86::FR64RegClass;
26962 else if (X86::VR128RegClass.hasType(VT))
26963 Res.second = &X86::VR128RegClass;
26964 else if (X86::VR256RegClass.hasType(VT))
26965 Res.second = &X86::VR256RegClass;
26966 else if (X86::VR512RegClass.hasType(VT))
26967 Res.second = &X86::VR512RegClass;
26973 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
26975 // Scaling factors are not free at all.
26976 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26977 // will take 2 allocations in the out of order engine instead of 1
26978 // for plain addressing mode, i.e. inst (reg1).
26980 // vaddps (%rsi,%drx), %ymm0, %ymm1
26981 // Requires two allocations (one for the load, one for the computation)
26983 // vaddps (%rsi), %ymm0, %ymm1
26984 // Requires just 1 allocation, i.e., freeing allocations for other operations
26985 // and having less micro operations to execute.
26987 // For some X86 architectures, this is even worse because for instance for
26988 // stores, the complex addressing mode forces the instruction to use the
26989 // "load" ports instead of the dedicated "store" port.
26990 // E.g., on Haswell:
26991 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26992 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26993 if (isLegalAddressingMode(AM, Ty))
26994 // Scale represents reg2 * scale, thus account for 1
26995 // as soon as we use a second register.
26996 return AM.Scale != 0;
27000 bool X86TargetLowering::isTargetFTOL() const {
27001 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();