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<int> ReciprocalEstimateRefinementSteps(
71 "x86-recip-refinement-steps", cl::init(1),
72 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
73 "result of the hardware reciprocal estimate instruction."),
76 // Forward declarations.
77 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
80 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
81 SelectionDAG &DAG, SDLoc dl,
82 unsigned vectorWidth) {
83 assert((vectorWidth == 128 || vectorWidth == 256) &&
84 "Unsupported vector width");
85 EVT VT = Vec.getValueType();
86 EVT ElVT = VT.getVectorElementType();
87 unsigned Factor = VT.getSizeInBits()/vectorWidth;
88 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
89 VT.getVectorNumElements()/Factor);
91 // Extract from UNDEF is UNDEF.
92 if (Vec.getOpcode() == ISD::UNDEF)
93 return DAG.getUNDEF(ResultVT);
95 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
96 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
98 // This is the index of the first element of the vectorWidth-bit chunk
100 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
103 // If the input is a buildvector just emit a smaller one.
104 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
105 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
106 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
109 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
110 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
113 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
114 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
115 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
116 /// instructions or a simple subregister reference. Idx is an index in the
117 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
118 /// lowering EXTRACT_VECTOR_ELT operations easier.
119 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
120 SelectionDAG &DAG, SDLoc dl) {
121 assert((Vec.getValueType().is256BitVector() ||
122 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
123 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
126 /// Generate a DAG to grab 256-bits from a 512-bit vector.
127 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
128 SelectionDAG &DAG, SDLoc dl) {
129 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
130 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
133 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
134 unsigned IdxVal, SelectionDAG &DAG,
135 SDLoc dl, unsigned vectorWidth) {
136 assert((vectorWidth == 128 || vectorWidth == 256) &&
137 "Unsupported vector width");
138 // Inserting UNDEF is Result
139 if (Vec.getOpcode() == ISD::UNDEF)
141 EVT VT = Vec.getValueType();
142 EVT ElVT = VT.getVectorElementType();
143 EVT ResultVT = Result.getValueType();
145 // Insert the relevant vectorWidth bits.
146 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
148 // This is the index of the first element of the vectorWidth-bit chunk
150 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
153 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
154 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
157 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
158 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
159 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
160 /// simple superregister reference. Idx is an index in the 128 bits
161 /// we want. It need not be aligned to a 128-bit boundary. That makes
162 /// lowering INSERT_VECTOR_ELT operations easier.
163 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
164 SelectionDAG &DAG,SDLoc dl) {
165 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
166 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
169 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
170 SelectionDAG &DAG, SDLoc dl) {
171 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
172 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
175 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
176 /// instructions. This is used because creating CONCAT_VECTOR nodes of
177 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
178 /// large BUILD_VECTORS.
179 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
180 unsigned NumElems, SelectionDAG &DAG,
182 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
183 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
186 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
187 unsigned NumElems, SelectionDAG &DAG,
189 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
190 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
193 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
194 const X86Subtarget &STI)
195 : TargetLowering(TM), Subtarget(&STI) {
196 X86ScalarSSEf64 = Subtarget->hasSSE2();
197 X86ScalarSSEf32 = Subtarget->hasSSE1();
198 TD = getDataLayout();
200 // Set up the TargetLowering object.
201 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
203 // X86 is weird. It always uses i8 for shift amounts and setcc results.
204 setBooleanContents(ZeroOrOneBooleanContent);
205 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
206 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
208 // For 64-bit, since we have so many registers, use the ILP scheduler.
209 // For 32-bit, use the register pressure specific scheduling.
210 // For Atom, always use ILP scheduling.
211 if (Subtarget->isAtom())
212 setSchedulingPreference(Sched::ILP);
213 else if (Subtarget->is64Bit())
214 setSchedulingPreference(Sched::ILP);
216 setSchedulingPreference(Sched::RegPressure);
217 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
218 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
220 // Bypass expensive divides on Atom when compiling with O2.
221 if (TM.getOptLevel() >= CodeGenOpt::Default) {
222 if (Subtarget->hasSlowDivide32())
223 addBypassSlowDiv(32, 8);
224 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
225 addBypassSlowDiv(64, 16);
228 if (Subtarget->isTargetKnownWindowsMSVC()) {
229 // Setup Windows compiler runtime calls.
230 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
231 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
232 setLibcallName(RTLIB::SREM_I64, "_allrem");
233 setLibcallName(RTLIB::UREM_I64, "_aullrem");
234 setLibcallName(RTLIB::MUL_I64, "_allmul");
235 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
236 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
237 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
238 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
239 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
241 // The _ftol2 runtime function has an unusual calling conv, which
242 // is modeled by a special pseudo-instruction.
243 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
244 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
245 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
246 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
249 if (Subtarget->isTargetDarwin()) {
250 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
251 setUseUnderscoreSetJmp(false);
252 setUseUnderscoreLongJmp(false);
253 } else if (Subtarget->isTargetWindowsGNU()) {
254 // MS runtime is weird: it exports _setjmp, but longjmp!
255 setUseUnderscoreSetJmp(true);
256 setUseUnderscoreLongJmp(false);
258 setUseUnderscoreSetJmp(true);
259 setUseUnderscoreLongJmp(true);
262 // Set up the register classes.
263 addRegisterClass(MVT::i8, &X86::GR8RegClass);
264 addRegisterClass(MVT::i16, &X86::GR16RegClass);
265 addRegisterClass(MVT::i32, &X86::GR32RegClass);
266 if (Subtarget->is64Bit())
267 addRegisterClass(MVT::i64, &X86::GR64RegClass);
269 for (MVT VT : MVT::integer_valuetypes())
270 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
272 // We don't accept any truncstore of integer registers.
273 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
274 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
275 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
276 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
277 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
278 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
280 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
282 // SETOEQ and SETUNE require checking two conditions.
283 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
284 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
285 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
286 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
287 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
288 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
290 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
292 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
293 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
294 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
298 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
299 } else if (!TM.Options.UseSoftFloat) {
300 // We have an algorithm for SSE2->double, and we turn this into a
301 // 64-bit FILD followed by conditional FADD for other targets.
302 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
303 // We have an algorithm for SSE2, and we turn this into a 64-bit
304 // FILD for other targets.
305 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
308 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
310 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
311 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
313 if (!TM.Options.UseSoftFloat) {
314 // SSE has no i16 to fp conversion, only i32
315 if (X86ScalarSSEf32) {
316 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
317 // f32 and f64 cases are Legal, f80 case is not
318 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
320 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
321 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
324 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
325 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
328 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
329 // are Legal, f80 is custom lowered.
330 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
331 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
333 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
335 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
336 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
338 if (X86ScalarSSEf32) {
339 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
340 // f32 and f64 cases are Legal, f80 case is not
341 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
343 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
344 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
347 // Handle FP_TO_UINT by promoting the destination to a larger signed
349 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
350 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
351 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
353 if (Subtarget->is64Bit()) {
354 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
355 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
356 } else if (!TM.Options.UseSoftFloat) {
357 // Since AVX is a superset of SSE3, only check for SSE here.
358 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
359 // Expand FP_TO_UINT into a select.
360 // FIXME: We would like to use a Custom expander here eventually to do
361 // the optimal thing for SSE vs. the default expansion in the legalizer.
362 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
364 // With SSE3 we can use fisttpll to convert to a signed i64; without
365 // SSE, we're stuck with a fistpll.
366 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
369 if (isTargetFTOL()) {
370 // Use the _ftol2 runtime function, which has a pseudo-instruction
371 // to handle its weird calling convention.
372 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
375 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
376 if (!X86ScalarSSEf64) {
377 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
378 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
379 if (Subtarget->is64Bit()) {
380 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
381 // Without SSE, i64->f64 goes through memory.
382 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
386 // Scalar integer divide and remainder are lowered to use operations that
387 // produce two results, to match the available instructions. This exposes
388 // the two-result form to trivial CSE, which is able to combine x/y and x%y
389 // into a single instruction.
391 // Scalar integer multiply-high is also lowered to use two-result
392 // operations, to match the available instructions. However, plain multiply
393 // (low) operations are left as Legal, as there are single-result
394 // instructions for this in x86. Using the two-result multiply instructions
395 // when both high and low results are needed must be arranged by dagcombine.
396 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
398 setOperationAction(ISD::MULHS, VT, Expand);
399 setOperationAction(ISD::MULHU, VT, Expand);
400 setOperationAction(ISD::SDIV, VT, Expand);
401 setOperationAction(ISD::UDIV, VT, Expand);
402 setOperationAction(ISD::SREM, VT, Expand);
403 setOperationAction(ISD::UREM, VT, Expand);
405 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
406 setOperationAction(ISD::ADDC, VT, Custom);
407 setOperationAction(ISD::ADDE, VT, Custom);
408 setOperationAction(ISD::SUBC, VT, Custom);
409 setOperationAction(ISD::SUBE, VT, Custom);
412 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
413 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
414 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
415 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
416 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
417 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
418 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
419 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
420 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
421 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
422 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
423 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
424 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
425 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
426 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
427 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
430 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
431 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
432 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
433 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
434 setOperationAction(ISD::FREM , MVT::f32 , Expand);
435 setOperationAction(ISD::FREM , MVT::f64 , Expand);
436 setOperationAction(ISD::FREM , MVT::f80 , Expand);
437 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
439 // Promote the i8 variants and force them on up to i32 which has a shorter
441 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
442 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
443 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
444 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
445 if (Subtarget->hasBMI()) {
446 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
447 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
448 if (Subtarget->is64Bit())
449 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
451 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
452 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
453 if (Subtarget->is64Bit())
454 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
457 if (Subtarget->hasLZCNT()) {
458 // When promoting the i8 variants, force them to i32 for a shorter
460 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
461 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
462 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
463 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
464 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
465 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
466 if (Subtarget->is64Bit())
467 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
469 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
470 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
471 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
472 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
473 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
474 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
475 if (Subtarget->is64Bit()) {
476 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
477 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
481 // Special handling for half-precision floating point conversions.
482 // If we don't have F16C support, then lower half float conversions
483 // into library calls.
484 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
485 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
486 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
489 // There's never any support for operations beyond MVT::f32.
490 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
491 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
492 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
493 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
495 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
496 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
497 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
498 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
499 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
500 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
502 if (Subtarget->hasPOPCNT()) {
503 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
505 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
506 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
507 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
508 if (Subtarget->is64Bit())
509 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
512 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
514 if (!Subtarget->hasMOVBE())
515 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
517 // These should be promoted to a larger select which is supported.
518 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
519 // X86 wants to expand cmov itself.
520 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
521 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
522 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
523 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
524 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
525 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
526 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
527 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
528 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
529 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
530 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
531 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
532 if (Subtarget->is64Bit()) {
533 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
534 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
536 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
537 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
538 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
539 // support continuation, user-level threading, and etc.. As a result, no
540 // other SjLj exception interfaces are implemented and please don't build
541 // your own exception handling based on them.
542 // LLVM/Clang supports zero-cost DWARF exception handling.
543 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
544 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
547 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
548 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
549 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
550 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
551 if (Subtarget->is64Bit())
552 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
553 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
554 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
555 if (Subtarget->is64Bit()) {
556 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
557 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
558 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
559 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
560 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
562 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
563 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
564 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
565 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
566 if (Subtarget->is64Bit()) {
567 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
568 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
569 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
572 if (Subtarget->hasSSE1())
573 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
575 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
577 // Expand certain atomics
578 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
580 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
581 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
582 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
585 if (Subtarget->hasCmpxchg16b()) {
586 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
589 // FIXME - use subtarget debug flags
590 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
591 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
592 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
595 if (Subtarget->is64Bit()) {
596 setExceptionPointerRegister(X86::RAX);
597 setExceptionSelectorRegister(X86::RDX);
599 setExceptionPointerRegister(X86::EAX);
600 setExceptionSelectorRegister(X86::EDX);
602 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
603 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
605 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
606 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
608 setOperationAction(ISD::TRAP, MVT::Other, Legal);
609 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
611 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
612 setOperationAction(ISD::VASTART , MVT::Other, Custom);
613 setOperationAction(ISD::VAEND , MVT::Other, Expand);
614 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
615 // TargetInfo::X86_64ABIBuiltinVaList
616 setOperationAction(ISD::VAARG , MVT::Other, Custom);
617 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
619 // TargetInfo::CharPtrBuiltinVaList
620 setOperationAction(ISD::VAARG , MVT::Other, Expand);
621 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
624 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
625 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
627 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
629 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
630 // f32 and f64 use SSE.
631 // Set up the FP register classes.
632 addRegisterClass(MVT::f32, &X86::FR32RegClass);
633 addRegisterClass(MVT::f64, &X86::FR64RegClass);
635 // Use ANDPD to simulate FABS.
636 setOperationAction(ISD::FABS , MVT::f64, Custom);
637 setOperationAction(ISD::FABS , MVT::f32, Custom);
639 // Use XORP to simulate FNEG.
640 setOperationAction(ISD::FNEG , MVT::f64, Custom);
641 setOperationAction(ISD::FNEG , MVT::f32, Custom);
643 // Use ANDPD and ORPD to simulate FCOPYSIGN.
644 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
645 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
647 // Lower this to FGETSIGNx86 plus an AND.
648 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
649 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
651 // We don't support sin/cos/fmod
652 setOperationAction(ISD::FSIN , MVT::f64, Expand);
653 setOperationAction(ISD::FCOS , MVT::f64, Expand);
654 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
655 setOperationAction(ISD::FSIN , MVT::f32, Expand);
656 setOperationAction(ISD::FCOS , MVT::f32, Expand);
657 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
659 // Expand FP immediates into loads from the stack, except for the special
661 addLegalFPImmediate(APFloat(+0.0)); // xorpd
662 addLegalFPImmediate(APFloat(+0.0f)); // xorps
663 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
664 // Use SSE for f32, x87 for f64.
665 // Set up the FP register classes.
666 addRegisterClass(MVT::f32, &X86::FR32RegClass);
667 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
669 // Use ANDPS to simulate FABS.
670 setOperationAction(ISD::FABS , MVT::f32, Custom);
672 // Use XORP to simulate FNEG.
673 setOperationAction(ISD::FNEG , MVT::f32, Custom);
675 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
677 // Use ANDPS and ORPS to simulate FCOPYSIGN.
678 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
679 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
681 // We don't support sin/cos/fmod
682 setOperationAction(ISD::FSIN , MVT::f32, Expand);
683 setOperationAction(ISD::FCOS , MVT::f32, Expand);
684 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
686 // Special cases we handle for FP constants.
687 addLegalFPImmediate(APFloat(+0.0f)); // xorps
688 addLegalFPImmediate(APFloat(+0.0)); // FLD0
689 addLegalFPImmediate(APFloat(+1.0)); // FLD1
690 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
691 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
693 if (!TM.Options.UnsafeFPMath) {
694 setOperationAction(ISD::FSIN , MVT::f64, Expand);
695 setOperationAction(ISD::FCOS , MVT::f64, Expand);
696 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
698 } else if (!TM.Options.UseSoftFloat) {
699 // f32 and f64 in x87.
700 // Set up the FP register classes.
701 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
702 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
704 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
705 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
706 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
707 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
709 if (!TM.Options.UnsafeFPMath) {
710 setOperationAction(ISD::FSIN , MVT::f64, Expand);
711 setOperationAction(ISD::FSIN , MVT::f32, Expand);
712 setOperationAction(ISD::FCOS , MVT::f64, Expand);
713 setOperationAction(ISD::FCOS , MVT::f32, Expand);
714 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
715 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
717 addLegalFPImmediate(APFloat(+0.0)); // FLD0
718 addLegalFPImmediate(APFloat(+1.0)); // FLD1
719 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
720 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
721 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
722 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
723 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
724 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
727 // We don't support FMA.
728 setOperationAction(ISD::FMA, MVT::f64, Expand);
729 setOperationAction(ISD::FMA, MVT::f32, Expand);
731 // Long double always uses X87.
732 if (!TM.Options.UseSoftFloat) {
733 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
734 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
735 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
737 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
738 addLegalFPImmediate(TmpFlt); // FLD0
740 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
743 APFloat TmpFlt2(+1.0);
744 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
746 addLegalFPImmediate(TmpFlt2); // FLD1
747 TmpFlt2.changeSign();
748 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
751 if (!TM.Options.UnsafeFPMath) {
752 setOperationAction(ISD::FSIN , MVT::f80, Expand);
753 setOperationAction(ISD::FCOS , MVT::f80, Expand);
754 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
757 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
758 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
759 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
760 setOperationAction(ISD::FRINT, MVT::f80, Expand);
761 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
762 setOperationAction(ISD::FMA, MVT::f80, Expand);
765 // Always use a library call for pow.
766 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
767 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
768 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
770 setOperationAction(ISD::FLOG, MVT::f80, Expand);
771 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
772 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
773 setOperationAction(ISD::FEXP, MVT::f80, Expand);
774 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
775 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
776 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
778 // First set operation action for all vector types to either promote
779 // (for widening) or expand (for scalarization). Then we will selectively
780 // turn on ones that can be effectively codegen'd.
781 for (MVT VT : MVT::vector_valuetypes()) {
782 setOperationAction(ISD::ADD , VT, Expand);
783 setOperationAction(ISD::SUB , VT, Expand);
784 setOperationAction(ISD::FADD, VT, Expand);
785 setOperationAction(ISD::FNEG, VT, Expand);
786 setOperationAction(ISD::FSUB, VT, Expand);
787 setOperationAction(ISD::MUL , VT, Expand);
788 setOperationAction(ISD::FMUL, VT, Expand);
789 setOperationAction(ISD::SDIV, VT, Expand);
790 setOperationAction(ISD::UDIV, VT, Expand);
791 setOperationAction(ISD::FDIV, VT, Expand);
792 setOperationAction(ISD::SREM, VT, Expand);
793 setOperationAction(ISD::UREM, VT, Expand);
794 setOperationAction(ISD::LOAD, VT, Expand);
795 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
797 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
798 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
799 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
800 setOperationAction(ISD::FABS, VT, Expand);
801 setOperationAction(ISD::FSIN, VT, Expand);
802 setOperationAction(ISD::FSINCOS, VT, Expand);
803 setOperationAction(ISD::FCOS, VT, Expand);
804 setOperationAction(ISD::FSINCOS, VT, Expand);
805 setOperationAction(ISD::FREM, VT, Expand);
806 setOperationAction(ISD::FMA, VT, Expand);
807 setOperationAction(ISD::FPOWI, VT, Expand);
808 setOperationAction(ISD::FSQRT, VT, Expand);
809 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
810 setOperationAction(ISD::FFLOOR, VT, Expand);
811 setOperationAction(ISD::FCEIL, VT, Expand);
812 setOperationAction(ISD::FTRUNC, VT, Expand);
813 setOperationAction(ISD::FRINT, VT, Expand);
814 setOperationAction(ISD::FNEARBYINT, VT, Expand);
815 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
816 setOperationAction(ISD::MULHS, VT, Expand);
817 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
818 setOperationAction(ISD::MULHU, VT, Expand);
819 setOperationAction(ISD::SDIVREM, VT, Expand);
820 setOperationAction(ISD::UDIVREM, VT, Expand);
821 setOperationAction(ISD::FPOW, VT, Expand);
822 setOperationAction(ISD::CTPOP, VT, Expand);
823 setOperationAction(ISD::CTTZ, VT, Expand);
824 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
825 setOperationAction(ISD::CTLZ, VT, Expand);
826 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
827 setOperationAction(ISD::SHL, VT, Expand);
828 setOperationAction(ISD::SRA, VT, Expand);
829 setOperationAction(ISD::SRL, VT, Expand);
830 setOperationAction(ISD::ROTL, VT, Expand);
831 setOperationAction(ISD::ROTR, VT, Expand);
832 setOperationAction(ISD::BSWAP, VT, Expand);
833 setOperationAction(ISD::SETCC, VT, Expand);
834 setOperationAction(ISD::FLOG, VT, Expand);
835 setOperationAction(ISD::FLOG2, VT, Expand);
836 setOperationAction(ISD::FLOG10, VT, Expand);
837 setOperationAction(ISD::FEXP, VT, Expand);
838 setOperationAction(ISD::FEXP2, VT, Expand);
839 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
840 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
841 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
842 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
843 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
844 setOperationAction(ISD::TRUNCATE, VT, Expand);
845 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
846 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
847 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
848 setOperationAction(ISD::VSELECT, VT, Expand);
849 setOperationAction(ISD::SELECT_CC, VT, Expand);
850 for (MVT InnerVT : MVT::vector_valuetypes()) {
851 setTruncStoreAction(InnerVT, VT, Expand);
853 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
854 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
856 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
857 // types, we have to deal with them whether we ask for Expansion or not.
858 // Setting Expand causes its own optimisation problems though, so leave
860 if (VT.getVectorElementType() == MVT::i1)
861 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
865 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
866 // with -msoft-float, disable use of MMX as well.
867 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
868 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
869 // No operations on x86mmx supported, everything uses intrinsics.
872 // MMX-sized vectors (other than x86mmx) are expected to be expanded
873 // into smaller operations.
874 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
875 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
876 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
877 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
878 setOperationAction(ISD::AND, MVT::v8i8, Expand);
879 setOperationAction(ISD::AND, MVT::v4i16, Expand);
880 setOperationAction(ISD::AND, MVT::v2i32, Expand);
881 setOperationAction(ISD::AND, MVT::v1i64, Expand);
882 setOperationAction(ISD::OR, MVT::v8i8, Expand);
883 setOperationAction(ISD::OR, MVT::v4i16, Expand);
884 setOperationAction(ISD::OR, MVT::v2i32, Expand);
885 setOperationAction(ISD::OR, MVT::v1i64, Expand);
886 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
887 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
888 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
889 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
890 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
891 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
892 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
893 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
895 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
896 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
897 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
898 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
899 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
900 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
901 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
902 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
904 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
905 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
907 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
908 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
909 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
910 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
911 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
912 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
913 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
914 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
917 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
918 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
919 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
920 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
923 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
924 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
926 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
927 // registers cannot be used even for integer operations.
928 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
929 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
930 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
931 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
933 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
934 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
935 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
936 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
937 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
938 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
939 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
940 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
941 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
942 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
943 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
944 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
945 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
946 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
947 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
948 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
949 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
950 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
951 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
952 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
953 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
954 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
956 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
957 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
958 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
959 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
961 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
962 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
963 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
964 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
965 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
967 // Only provide customized ctpop vector bit twiddling for vector types we
968 // know to perform better than using the popcnt instructions on each vector
969 // element. If popcnt isn't supported, always provide the custom version.
970 if (!Subtarget->hasPOPCNT()) {
971 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
972 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
975 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
976 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
977 MVT VT = (MVT::SimpleValueType)i;
978 // Do not attempt to custom lower non-power-of-2 vectors
979 if (!isPowerOf2_32(VT.getVectorNumElements()))
981 // Do not attempt to custom lower non-128-bit vectors
982 if (!VT.is128BitVector())
984 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
985 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
986 setOperationAction(ISD::VSELECT, VT, Custom);
987 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
990 // We support custom legalizing of sext and anyext loads for specific
991 // memory vector types which we can load as a scalar (or sequence of
992 // scalars) and extend in-register to a legal 128-bit vector type. For sext
993 // loads these must work with a single scalar load.
994 for (MVT VT : MVT::integer_vector_valuetypes()) {
995 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
996 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
997 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
998 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
999 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
1000 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
1001 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
1002 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
1003 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
1006 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1007 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1008 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1009 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1010 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1011 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1013 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1015 if (Subtarget->is64Bit()) {
1016 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1020 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1021 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1022 MVT VT = (MVT::SimpleValueType)i;
1024 // Do not attempt to promote non-128-bit vectors
1025 if (!VT.is128BitVector())
1028 setOperationAction(ISD::AND, VT, Promote);
1029 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1030 setOperationAction(ISD::OR, VT, Promote);
1031 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1032 setOperationAction(ISD::XOR, VT, Promote);
1033 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1034 setOperationAction(ISD::LOAD, VT, Promote);
1035 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1036 setOperationAction(ISD::SELECT, VT, Promote);
1037 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1040 // Custom lower v2i64 and v2f64 selects.
1041 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1042 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1043 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1044 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1046 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1047 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1049 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1050 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1051 // As there is no 64-bit GPR available, we need build a special custom
1052 // sequence to convert from v2i32 to v2f32.
1053 if (!Subtarget->is64Bit())
1054 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1056 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1057 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1059 for (MVT VT : MVT::fp_vector_valuetypes())
1060 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
1062 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1063 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1064 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1067 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1068 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1069 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1070 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1071 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1072 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1073 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1074 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1075 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1076 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1077 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1079 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1080 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1081 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1082 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1083 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1084 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1085 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1086 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1087 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1088 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1090 // FIXME: Do we need to handle scalar-to-vector here?
1091 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1093 // We directly match byte blends in the backend as they match the VSELECT
1095 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1097 // SSE41 brings specific instructions for doing vector sign extend even in
1098 // cases where we don't have SRA.
1099 for (MVT VT : MVT::integer_vector_valuetypes()) {
1100 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
1101 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
1102 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
1105 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
1106 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1107 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1108 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1109 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1110 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1111 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1113 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1114 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1115 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1116 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1117 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1118 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1120 // i8 and i16 vectors are custom because the source register and source
1121 // source memory operand types are not the same width. f32 vectors are
1122 // custom since the immediate controlling the insert encodes additional
1124 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1125 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1126 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1127 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1129 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1130 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1131 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1132 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1134 // FIXME: these should be Legal, but that's only for the case where
1135 // the index is constant. For now custom expand to deal with that.
1136 if (Subtarget->is64Bit()) {
1137 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1138 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1142 if (Subtarget->hasSSE2()) {
1143 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1144 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1146 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1147 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1149 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1150 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1152 // In the customized shift lowering, the legal cases in AVX2 will be
1154 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1155 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1157 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1158 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1160 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1163 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1164 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1165 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1166 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1167 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1168 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1169 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1171 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1172 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1173 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1175 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1176 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1177 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1178 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1179 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1180 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1181 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1182 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1183 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1184 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1185 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1186 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1188 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1189 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1190 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1191 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1192 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1193 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1194 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1195 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1196 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1197 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1198 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1199 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1201 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1202 // even though v8i16 is a legal type.
1203 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1204 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1205 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1207 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1208 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1209 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1211 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1212 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1214 for (MVT VT : MVT::fp_vector_valuetypes())
1215 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1217 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1218 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1220 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1221 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1223 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1224 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1226 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1227 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1228 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1229 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1232 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1233 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1235 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1236 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1237 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1238 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1239 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1240 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1241 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1242 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1243 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1244 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1245 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1246 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1248 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1249 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1250 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1251 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1252 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1253 setOperationAction(ISD::FMA, MVT::f32, Legal);
1254 setOperationAction(ISD::FMA, MVT::f64, Legal);
1257 if (Subtarget->hasInt256()) {
1258 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1259 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1260 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1261 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1263 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1264 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1265 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1266 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1268 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1269 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1270 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1271 // Don't lower v32i8 because there is no 128-bit byte mul
1273 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1274 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1275 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1276 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1278 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1279 // when we have a 256bit-wide blend with immediate.
1280 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1282 // Only provide customized ctpop vector bit twiddling for vector types we
1283 // know to perform better than using the popcnt instructions on each
1284 // vector element. If popcnt isn't supported, always provide the custom
1286 if (!Subtarget->hasPOPCNT())
1287 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1289 // Custom CTPOP always performs better on natively supported v8i32
1290 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1292 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1293 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1294 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1295 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1296 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1297 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1298 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1300 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1301 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1302 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1303 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1304 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1305 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1307 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1308 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1309 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1310 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1312 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1313 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1314 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1315 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1317 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1318 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1319 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1320 // Don't lower v32i8 because there is no 128-bit byte mul
1323 // In the customized shift lowering, the legal cases in AVX2 will be
1325 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1326 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1328 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1329 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1331 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1333 // Custom lower several nodes for 256-bit types.
1334 for (MVT VT : MVT::vector_valuetypes()) {
1335 if (VT.getScalarSizeInBits() >= 32) {
1336 setOperationAction(ISD::MLOAD, VT, Legal);
1337 setOperationAction(ISD::MSTORE, VT, Legal);
1339 // Extract subvector is special because the value type
1340 // (result) is 128-bit but the source is 256-bit wide.
1341 if (VT.is128BitVector()) {
1342 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1344 // Do not attempt to custom lower other non-256-bit vectors
1345 if (!VT.is256BitVector())
1348 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1349 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1350 setOperationAction(ISD::VSELECT, VT, Custom);
1351 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1352 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1353 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1354 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1355 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1358 if (Subtarget->hasInt256())
1359 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1362 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1363 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1364 MVT VT = (MVT::SimpleValueType)i;
1366 // Do not attempt to promote non-256-bit vectors
1367 if (!VT.is256BitVector())
1370 setOperationAction(ISD::AND, VT, Promote);
1371 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1372 setOperationAction(ISD::OR, VT, Promote);
1373 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1374 setOperationAction(ISD::XOR, VT, Promote);
1375 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1376 setOperationAction(ISD::LOAD, VT, Promote);
1377 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1378 setOperationAction(ISD::SELECT, VT, Promote);
1379 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1383 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1384 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1385 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1386 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1387 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1389 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1390 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1391 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1393 for (MVT VT : MVT::fp_vector_valuetypes())
1394 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1396 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1397 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1398 setOperationAction(ISD::XOR, MVT::i1, Legal);
1399 setOperationAction(ISD::OR, MVT::i1, Legal);
1400 setOperationAction(ISD::AND, MVT::i1, Legal);
1401 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1402 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1403 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1404 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1405 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1407 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1408 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1409 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1410 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1411 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1412 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1414 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1415 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1416 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1417 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1418 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1419 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1420 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1421 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1423 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1424 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1425 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1426 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1427 if (Subtarget->is64Bit()) {
1428 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1429 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1430 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1431 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1433 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1434 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1435 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1438 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1439 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1440 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1441 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1442 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1443 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1444 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1445 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1446 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1448 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1449 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1450 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1451 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1452 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1453 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1454 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1455 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1456 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1457 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1458 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1459 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1460 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1462 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1463 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1464 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1465 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1466 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1467 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1468 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1469 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1470 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1471 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1474 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1475 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1476 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1477 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1478 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1480 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1481 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1483 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1485 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1486 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1487 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1488 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1489 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1490 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1491 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1492 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1493 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1495 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1496 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1498 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1499 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1501 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1503 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1504 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1506 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1507 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1509 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1510 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1512 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1513 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1514 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1515 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1516 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1517 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1519 if (Subtarget->hasCDI()) {
1520 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1521 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1524 // Custom lower several nodes.
1525 for (MVT VT : MVT::vector_valuetypes()) {
1526 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1527 // Extract subvector is special because the value type
1528 // (result) is 256/128-bit but the source is 512-bit wide.
1529 if (VT.is128BitVector() || VT.is256BitVector()) {
1530 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1532 if (VT.getVectorElementType() == MVT::i1)
1533 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1535 // Do not attempt to custom lower other non-512-bit vectors
1536 if (!VT.is512BitVector())
1539 if ( EltSize >= 32) {
1540 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1541 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1542 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1543 setOperationAction(ISD::VSELECT, VT, Legal);
1544 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1545 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1546 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1547 setOperationAction(ISD::MLOAD, VT, Legal);
1548 setOperationAction(ISD::MSTORE, VT, Legal);
1551 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1552 MVT VT = (MVT::SimpleValueType)i;
1554 // Do not attempt to promote non-512-bit vectors.
1555 if (!VT.is512BitVector())
1558 setOperationAction(ISD::SELECT, VT, Promote);
1559 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1563 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1564 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1565 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1567 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1568 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1570 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1571 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1572 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1573 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1574 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1575 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1576 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1577 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1578 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1580 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1581 const MVT VT = (MVT::SimpleValueType)i;
1583 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1585 // Do not attempt to promote non-512-bit vectors.
1586 if (!VT.is512BitVector())
1590 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1591 setOperationAction(ISD::VSELECT, VT, Legal);
1596 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1597 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1598 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1600 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1601 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1602 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal);
1604 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1605 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1606 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1607 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1608 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1609 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1612 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1613 // of this type with custom code.
1614 for (MVT VT : MVT::vector_valuetypes())
1615 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
1617 // We want to custom lower some of our intrinsics.
1618 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1619 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1620 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1621 if (!Subtarget->is64Bit())
1622 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1624 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1625 // handle type legalization for these operations here.
1627 // FIXME: We really should do custom legalization for addition and
1628 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1629 // than generic legalization for 64-bit multiplication-with-overflow, though.
1630 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1631 // Add/Sub/Mul with overflow operations are custom lowered.
1633 setOperationAction(ISD::SADDO, VT, Custom);
1634 setOperationAction(ISD::UADDO, VT, Custom);
1635 setOperationAction(ISD::SSUBO, VT, Custom);
1636 setOperationAction(ISD::USUBO, VT, Custom);
1637 setOperationAction(ISD::SMULO, VT, Custom);
1638 setOperationAction(ISD::UMULO, VT, Custom);
1642 if (!Subtarget->is64Bit()) {
1643 // These libcalls are not available in 32-bit.
1644 setLibcallName(RTLIB::SHL_I128, nullptr);
1645 setLibcallName(RTLIB::SRL_I128, nullptr);
1646 setLibcallName(RTLIB::SRA_I128, nullptr);
1649 // Combine sin / cos into one node or libcall if possible.
1650 if (Subtarget->hasSinCos()) {
1651 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1652 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1653 if (Subtarget->isTargetDarwin()) {
1654 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1655 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1656 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1657 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1661 if (Subtarget->isTargetWin64()) {
1662 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1663 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1664 setOperationAction(ISD::SREM, MVT::i128, Custom);
1665 setOperationAction(ISD::UREM, MVT::i128, Custom);
1666 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1667 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1670 // We have target-specific dag combine patterns for the following nodes:
1671 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1672 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1673 setTargetDAGCombine(ISD::BITCAST);
1674 setTargetDAGCombine(ISD::VSELECT);
1675 setTargetDAGCombine(ISD::SELECT);
1676 setTargetDAGCombine(ISD::SHL);
1677 setTargetDAGCombine(ISD::SRA);
1678 setTargetDAGCombine(ISD::SRL);
1679 setTargetDAGCombine(ISD::OR);
1680 setTargetDAGCombine(ISD::AND);
1681 setTargetDAGCombine(ISD::ADD);
1682 setTargetDAGCombine(ISD::FADD);
1683 setTargetDAGCombine(ISD::FSUB);
1684 setTargetDAGCombine(ISD::FMA);
1685 setTargetDAGCombine(ISD::SUB);
1686 setTargetDAGCombine(ISD::LOAD);
1687 setTargetDAGCombine(ISD::MLOAD);
1688 setTargetDAGCombine(ISD::STORE);
1689 setTargetDAGCombine(ISD::MSTORE);
1690 setTargetDAGCombine(ISD::ZERO_EXTEND);
1691 setTargetDAGCombine(ISD::ANY_EXTEND);
1692 setTargetDAGCombine(ISD::SIGN_EXTEND);
1693 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1694 setTargetDAGCombine(ISD::TRUNCATE);
1695 setTargetDAGCombine(ISD::SINT_TO_FP);
1696 setTargetDAGCombine(ISD::SETCC);
1697 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1698 setTargetDAGCombine(ISD::BUILD_VECTOR);
1699 setTargetDAGCombine(ISD::MUL);
1700 setTargetDAGCombine(ISD::XOR);
1702 computeRegisterProperties();
1704 // On Darwin, -Os means optimize for size without hurting performance,
1705 // do not reduce the limit.
1706 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1707 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1708 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1709 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1710 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1711 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1712 setPrefLoopAlignment(4); // 2^4 bytes.
1714 // Predictable cmov don't hurt on atom because it's in-order.
1715 PredictableSelectIsExpensive = !Subtarget->isAtom();
1716 EnableExtLdPromotion = true;
1717 setPrefFunctionAlignment(4); // 2^4 bytes.
1719 verifyIntrinsicTables();
1722 // This has so far only been implemented for 64-bit MachO.
1723 bool X86TargetLowering::useLoadStackGuardNode() const {
1724 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1727 TargetLoweringBase::LegalizeTypeAction
1728 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1729 if (ExperimentalVectorWideningLegalization &&
1730 VT.getVectorNumElements() != 1 &&
1731 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1732 return TypeWidenVector;
1734 return TargetLoweringBase::getPreferredVectorAction(VT);
1737 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1739 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1741 const unsigned NumElts = VT.getVectorNumElements();
1742 const EVT EltVT = VT.getVectorElementType();
1743 if (VT.is512BitVector()) {
1744 if (Subtarget->hasAVX512())
1745 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1746 EltVT == MVT::f32 || EltVT == MVT::f64)
1748 case 8: return MVT::v8i1;
1749 case 16: return MVT::v16i1;
1751 if (Subtarget->hasBWI())
1752 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1754 case 32: return MVT::v32i1;
1755 case 64: return MVT::v64i1;
1759 if (VT.is256BitVector() || VT.is128BitVector()) {
1760 if (Subtarget->hasVLX())
1761 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1762 EltVT == MVT::f32 || EltVT == MVT::f64)
1764 case 2: return MVT::v2i1;
1765 case 4: return MVT::v4i1;
1766 case 8: return MVT::v8i1;
1768 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1769 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1771 case 8: return MVT::v8i1;
1772 case 16: return MVT::v16i1;
1773 case 32: return MVT::v32i1;
1777 return VT.changeVectorElementTypeToInteger();
1780 /// Helper for getByValTypeAlignment to determine
1781 /// the desired ByVal argument alignment.
1782 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1785 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1786 if (VTy->getBitWidth() == 128)
1788 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1789 unsigned EltAlign = 0;
1790 getMaxByValAlign(ATy->getElementType(), EltAlign);
1791 if (EltAlign > MaxAlign)
1792 MaxAlign = EltAlign;
1793 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1794 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1795 unsigned EltAlign = 0;
1796 getMaxByValAlign(STy->getElementType(i), EltAlign);
1797 if (EltAlign > MaxAlign)
1798 MaxAlign = EltAlign;
1805 /// Return the desired alignment for ByVal aggregate
1806 /// function arguments in the caller parameter area. For X86, aggregates
1807 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1808 /// are at 4-byte boundaries.
1809 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1810 if (Subtarget->is64Bit()) {
1811 // Max of 8 and alignment of type.
1812 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1819 if (Subtarget->hasSSE1())
1820 getMaxByValAlign(Ty, Align);
1824 /// Returns the target specific optimal type for load
1825 /// and store operations as a result of memset, memcpy, and memmove
1826 /// lowering. If DstAlign is zero that means it's safe to destination
1827 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1828 /// means there isn't a need to check it against alignment requirement,
1829 /// probably because the source does not need to be loaded. If 'IsMemset' is
1830 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1831 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1832 /// source is constant so it does not need to be loaded.
1833 /// It returns EVT::Other if the type should be determined using generic
1834 /// target-independent logic.
1836 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1837 unsigned DstAlign, unsigned SrcAlign,
1838 bool IsMemset, bool ZeroMemset,
1840 MachineFunction &MF) const {
1841 const Function *F = MF.getFunction();
1842 if ((!IsMemset || ZeroMemset) &&
1843 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1845 (Subtarget->isUnalignedMemAccessFast() ||
1846 ((DstAlign == 0 || DstAlign >= 16) &&
1847 (SrcAlign == 0 || SrcAlign >= 16)))) {
1849 if (Subtarget->hasInt256())
1851 if (Subtarget->hasFp256())
1854 if (Subtarget->hasSSE2())
1856 if (Subtarget->hasSSE1())
1858 } else if (!MemcpyStrSrc && Size >= 8 &&
1859 !Subtarget->is64Bit() &&
1860 Subtarget->hasSSE2()) {
1861 // Do not use f64 to lower memcpy if source is string constant. It's
1862 // better to use i32 to avoid the loads.
1866 if (Subtarget->is64Bit() && Size >= 8)
1871 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1873 return X86ScalarSSEf32;
1874 else if (VT == MVT::f64)
1875 return X86ScalarSSEf64;
1880 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1885 *Fast = Subtarget->isUnalignedMemAccessFast();
1889 /// Return the entry encoding for a jump table in the
1890 /// current function. The returned value is a member of the
1891 /// MachineJumpTableInfo::JTEntryKind enum.
1892 unsigned X86TargetLowering::getJumpTableEncoding() const {
1893 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1895 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1896 Subtarget->isPICStyleGOT())
1897 return MachineJumpTableInfo::EK_Custom32;
1899 // Otherwise, use the normal jump table encoding heuristics.
1900 return TargetLowering::getJumpTableEncoding();
1904 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1905 const MachineBasicBlock *MBB,
1906 unsigned uid,MCContext &Ctx) const{
1907 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1908 Subtarget->isPICStyleGOT());
1909 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1911 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1912 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1915 /// Returns relocation base for the given PIC jumptable.
1916 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1917 SelectionDAG &DAG) const {
1918 if (!Subtarget->is64Bit())
1919 // This doesn't have SDLoc associated with it, but is not really the
1920 // same as a Register.
1921 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1925 /// This returns the relocation base for the given PIC jumptable,
1926 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1927 const MCExpr *X86TargetLowering::
1928 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1929 MCContext &Ctx) const {
1930 // X86-64 uses RIP relative addressing based on the jump table label.
1931 if (Subtarget->isPICStyleRIPRel())
1932 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1934 // Otherwise, the reference is relative to the PIC base.
1935 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1938 // FIXME: Why this routine is here? Move to RegInfo!
1939 std::pair<const TargetRegisterClass*, uint8_t>
1940 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1941 const TargetRegisterClass *RRC = nullptr;
1943 switch (VT.SimpleTy) {
1945 return TargetLowering::findRepresentativeClass(VT);
1946 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1947 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1950 RRC = &X86::VR64RegClass;
1952 case MVT::f32: case MVT::f64:
1953 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1954 case MVT::v4f32: case MVT::v2f64:
1955 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1957 RRC = &X86::VR128RegClass;
1960 return std::make_pair(RRC, Cost);
1963 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1964 unsigned &Offset) const {
1965 if (!Subtarget->isTargetLinux())
1968 if (Subtarget->is64Bit()) {
1969 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1971 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1983 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1984 unsigned DestAS) const {
1985 assert(SrcAS != DestAS && "Expected different address spaces!");
1987 return SrcAS < 256 && DestAS < 256;
1990 //===----------------------------------------------------------------------===//
1991 // Return Value Calling Convention Implementation
1992 //===----------------------------------------------------------------------===//
1994 #include "X86GenCallingConv.inc"
1997 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1998 MachineFunction &MF, bool isVarArg,
1999 const SmallVectorImpl<ISD::OutputArg> &Outs,
2000 LLVMContext &Context) const {
2001 SmallVector<CCValAssign, 16> RVLocs;
2002 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2003 return CCInfo.CheckReturn(Outs, RetCC_X86);
2006 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2007 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2012 X86TargetLowering::LowerReturn(SDValue Chain,
2013 CallingConv::ID CallConv, bool isVarArg,
2014 const SmallVectorImpl<ISD::OutputArg> &Outs,
2015 const SmallVectorImpl<SDValue> &OutVals,
2016 SDLoc dl, SelectionDAG &DAG) const {
2017 MachineFunction &MF = DAG.getMachineFunction();
2018 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2020 SmallVector<CCValAssign, 16> RVLocs;
2021 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2022 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2025 SmallVector<SDValue, 6> RetOps;
2026 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2027 // Operand #1 = Bytes To Pop
2028 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
2031 // Copy the result values into the output registers.
2032 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2033 CCValAssign &VA = RVLocs[i];
2034 assert(VA.isRegLoc() && "Can only return in registers!");
2035 SDValue ValToCopy = OutVals[i];
2036 EVT ValVT = ValToCopy.getValueType();
2038 // Promote values to the appropriate types.
2039 if (VA.getLocInfo() == CCValAssign::SExt)
2040 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2041 else if (VA.getLocInfo() == CCValAssign::ZExt)
2042 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2043 else if (VA.getLocInfo() == CCValAssign::AExt)
2044 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2045 else if (VA.getLocInfo() == CCValAssign::BCvt)
2046 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2048 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2049 "Unexpected FP-extend for return value.");
2051 // If this is x86-64, and we disabled SSE, we can't return FP values,
2052 // or SSE or MMX vectors.
2053 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2054 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2055 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2056 report_fatal_error("SSE register return with SSE disabled");
2058 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2059 // llvm-gcc has never done it right and no one has noticed, so this
2060 // should be OK for now.
2061 if (ValVT == MVT::f64 &&
2062 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2063 report_fatal_error("SSE2 register return with SSE2 disabled");
2065 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2066 // the RET instruction and handled by the FP Stackifier.
2067 if (VA.getLocReg() == X86::FP0 ||
2068 VA.getLocReg() == X86::FP1) {
2069 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2070 // change the value to the FP stack register class.
2071 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2072 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2073 RetOps.push_back(ValToCopy);
2074 // Don't emit a copytoreg.
2078 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2079 // which is returned in RAX / RDX.
2080 if (Subtarget->is64Bit()) {
2081 if (ValVT == MVT::x86mmx) {
2082 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2083 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2084 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2086 // If we don't have SSE2 available, convert to v4f32 so the generated
2087 // register is legal.
2088 if (!Subtarget->hasSSE2())
2089 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2094 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2095 Flag = Chain.getValue(1);
2096 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2099 // The x86-64 ABIs require that for returning structs by value we copy
2100 // the sret argument into %rax/%eax (depending on ABI) for the return.
2101 // Win32 requires us to put the sret argument to %eax as well.
2102 // We saved the argument into a virtual register in the entry block,
2103 // so now we copy the value out and into %rax/%eax.
2105 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2106 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2107 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2108 // either case FuncInfo->setSRetReturnReg() will have been called.
2109 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2110 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
2111 "No need for an sret register");
2112 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2115 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2116 X86::RAX : X86::EAX;
2117 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2118 Flag = Chain.getValue(1);
2120 // RAX/EAX now acts like a return value.
2121 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2124 RetOps[0] = Chain; // Update chain.
2126 // Add the flag if we have it.
2128 RetOps.push_back(Flag);
2130 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2133 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2134 if (N->getNumValues() != 1)
2136 if (!N->hasNUsesOfValue(1, 0))
2139 SDValue TCChain = Chain;
2140 SDNode *Copy = *N->use_begin();
2141 if (Copy->getOpcode() == ISD::CopyToReg) {
2142 // If the copy has a glue operand, we conservatively assume it isn't safe to
2143 // perform a tail call.
2144 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2146 TCChain = Copy->getOperand(0);
2147 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2150 bool HasRet = false;
2151 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2153 if (UI->getOpcode() != X86ISD::RET_FLAG)
2155 // If we are returning more than one value, we can definitely
2156 // not make a tail call see PR19530
2157 if (UI->getNumOperands() > 4)
2159 if (UI->getNumOperands() == 4 &&
2160 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2173 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2174 ISD::NodeType ExtendKind) const {
2176 // TODO: Is this also valid on 32-bit?
2177 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2178 ReturnMVT = MVT::i8;
2180 ReturnMVT = MVT::i32;
2182 EVT MinVT = getRegisterType(Context, ReturnMVT);
2183 return VT.bitsLT(MinVT) ? MinVT : VT;
2186 /// Lower the result values of a call into the
2187 /// appropriate copies out of appropriate physical registers.
2190 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2191 CallingConv::ID CallConv, bool isVarArg,
2192 const SmallVectorImpl<ISD::InputArg> &Ins,
2193 SDLoc dl, SelectionDAG &DAG,
2194 SmallVectorImpl<SDValue> &InVals) const {
2196 // Assign locations to each value returned by this call.
2197 SmallVector<CCValAssign, 16> RVLocs;
2198 bool Is64Bit = Subtarget->is64Bit();
2199 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2201 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2203 // Copy all of the result registers out of their specified physreg.
2204 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2205 CCValAssign &VA = RVLocs[i];
2206 EVT CopyVT = VA.getValVT();
2208 // If this is x86-64, and we disabled SSE, we can't return FP values
2209 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2210 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2211 report_fatal_error("SSE register return with SSE disabled");
2214 // If we prefer to use the value in xmm registers, copy it out as f80 and
2215 // use a truncate to move it from fp stack reg to xmm reg.
2216 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2217 isScalarFPTypeInSSEReg(VA.getValVT()))
2220 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2221 CopyVT, InFlag).getValue(1);
2222 SDValue Val = Chain.getValue(0);
2224 if (CopyVT != VA.getValVT())
2225 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2226 // This truncation won't change the value.
2227 DAG.getIntPtrConstant(1));
2229 InFlag = Chain.getValue(2);
2230 InVals.push_back(Val);
2236 //===----------------------------------------------------------------------===//
2237 // C & StdCall & Fast Calling Convention implementation
2238 //===----------------------------------------------------------------------===//
2239 // StdCall calling convention seems to be standard for many Windows' API
2240 // routines and around. It differs from C calling convention just a little:
2241 // callee should clean up the stack, not caller. Symbols should be also
2242 // decorated in some fancy way :) It doesn't support any vector arguments.
2243 // For info on fast calling convention see Fast Calling Convention (tail call)
2244 // implementation LowerX86_32FastCCCallTo.
2246 /// CallIsStructReturn - Determines whether a call uses struct return
2248 enum StructReturnType {
2253 static StructReturnType
2254 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2256 return NotStructReturn;
2258 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2259 if (!Flags.isSRet())
2260 return NotStructReturn;
2261 if (Flags.isInReg())
2262 return RegStructReturn;
2263 return StackStructReturn;
2266 /// Determines whether a function uses struct return semantics.
2267 static StructReturnType
2268 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2270 return NotStructReturn;
2272 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2273 if (!Flags.isSRet())
2274 return NotStructReturn;
2275 if (Flags.isInReg())
2276 return RegStructReturn;
2277 return StackStructReturn;
2280 /// Make a copy of an aggregate at address specified by "Src" to address
2281 /// "Dst" with size and alignment information specified by the specific
2282 /// parameter attribute. The copy will be passed as a byval function parameter.
2284 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2285 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2287 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2289 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2290 /*isVolatile*/false, /*AlwaysInline=*/true,
2291 MachinePointerInfo(), MachinePointerInfo());
2294 /// Return true if the calling convention is one that
2295 /// supports tail call optimization.
2296 static bool IsTailCallConvention(CallingConv::ID CC) {
2297 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2298 CC == CallingConv::HiPE);
2301 /// \brief Return true if the calling convention is a C calling convention.
2302 static bool IsCCallConvention(CallingConv::ID CC) {
2303 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2304 CC == CallingConv::X86_64_SysV);
2307 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2308 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2312 CallingConv::ID CalleeCC = CS.getCallingConv();
2313 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2319 /// Return true if the function is being made into
2320 /// a tailcall target by changing its ABI.
2321 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2322 bool GuaranteedTailCallOpt) {
2323 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2327 X86TargetLowering::LowerMemArgument(SDValue Chain,
2328 CallingConv::ID CallConv,
2329 const SmallVectorImpl<ISD::InputArg> &Ins,
2330 SDLoc dl, SelectionDAG &DAG,
2331 const CCValAssign &VA,
2332 MachineFrameInfo *MFI,
2334 // Create the nodes corresponding to a load from this parameter slot.
2335 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2336 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2337 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2338 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2341 // If value is passed by pointer we have address passed instead of the value
2343 if (VA.getLocInfo() == CCValAssign::Indirect)
2344 ValVT = VA.getLocVT();
2346 ValVT = VA.getValVT();
2348 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2349 // changed with more analysis.
2350 // In case of tail call optimization mark all arguments mutable. Since they
2351 // could be overwritten by lowering of arguments in case of a tail call.
2352 if (Flags.isByVal()) {
2353 unsigned Bytes = Flags.getByValSize();
2354 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2355 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2356 return DAG.getFrameIndex(FI, getPointerTy());
2358 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2359 VA.getLocMemOffset(), isImmutable);
2360 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2361 return DAG.getLoad(ValVT, dl, Chain, FIN,
2362 MachinePointerInfo::getFixedStack(FI),
2363 false, false, false, 0);
2367 // FIXME: Get this from tablegen.
2368 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2369 const X86Subtarget *Subtarget) {
2370 assert(Subtarget->is64Bit());
2372 if (Subtarget->isCallingConvWin64(CallConv)) {
2373 static const MCPhysReg GPR64ArgRegsWin64[] = {
2374 X86::RCX, X86::RDX, X86::R8, X86::R9
2376 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2379 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2380 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2382 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2385 // FIXME: Get this from tablegen.
2386 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2387 CallingConv::ID CallConv,
2388 const X86Subtarget *Subtarget) {
2389 assert(Subtarget->is64Bit());
2390 if (Subtarget->isCallingConvWin64(CallConv)) {
2391 // The XMM registers which might contain var arg parameters are shadowed
2392 // in their paired GPR. So we only need to save the GPR to their home
2394 // TODO: __vectorcall will change this.
2398 const Function *Fn = MF.getFunction();
2399 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2400 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2401 "SSE register cannot be used when SSE is disabled!");
2402 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2403 !Subtarget->hasSSE1())
2404 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2408 static const MCPhysReg XMMArgRegs64Bit[] = {
2409 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2410 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2412 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2416 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2417 CallingConv::ID CallConv,
2419 const SmallVectorImpl<ISD::InputArg> &Ins,
2422 SmallVectorImpl<SDValue> &InVals)
2424 MachineFunction &MF = DAG.getMachineFunction();
2425 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2427 const Function* Fn = MF.getFunction();
2428 if (Fn->hasExternalLinkage() &&
2429 Subtarget->isTargetCygMing() &&
2430 Fn->getName() == "main")
2431 FuncInfo->setForceFramePointer(true);
2433 MachineFrameInfo *MFI = MF.getFrameInfo();
2434 bool Is64Bit = Subtarget->is64Bit();
2435 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2437 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2438 "Var args not supported with calling convention fastcc, ghc or hipe");
2440 // Assign locations to all of the incoming arguments.
2441 SmallVector<CCValAssign, 16> ArgLocs;
2442 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2444 // Allocate shadow area for Win64
2446 CCInfo.AllocateStack(32, 8);
2448 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2450 unsigned LastVal = ~0U;
2452 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2453 CCValAssign &VA = ArgLocs[i];
2454 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2456 assert(VA.getValNo() != LastVal &&
2457 "Don't support value assigned to multiple locs yet");
2459 LastVal = VA.getValNo();
2461 if (VA.isRegLoc()) {
2462 EVT RegVT = VA.getLocVT();
2463 const TargetRegisterClass *RC;
2464 if (RegVT == MVT::i32)
2465 RC = &X86::GR32RegClass;
2466 else if (Is64Bit && RegVT == MVT::i64)
2467 RC = &X86::GR64RegClass;
2468 else if (RegVT == MVT::f32)
2469 RC = &X86::FR32RegClass;
2470 else if (RegVT == MVT::f64)
2471 RC = &X86::FR64RegClass;
2472 else if (RegVT.is512BitVector())
2473 RC = &X86::VR512RegClass;
2474 else if (RegVT.is256BitVector())
2475 RC = &X86::VR256RegClass;
2476 else if (RegVT.is128BitVector())
2477 RC = &X86::VR128RegClass;
2478 else if (RegVT == MVT::x86mmx)
2479 RC = &X86::VR64RegClass;
2480 else if (RegVT == MVT::i1)
2481 RC = &X86::VK1RegClass;
2482 else if (RegVT == MVT::v8i1)
2483 RC = &X86::VK8RegClass;
2484 else if (RegVT == MVT::v16i1)
2485 RC = &X86::VK16RegClass;
2486 else if (RegVT == MVT::v32i1)
2487 RC = &X86::VK32RegClass;
2488 else if (RegVT == MVT::v64i1)
2489 RC = &X86::VK64RegClass;
2491 llvm_unreachable("Unknown argument type!");
2493 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2494 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2496 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2497 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2499 if (VA.getLocInfo() == CCValAssign::SExt)
2500 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2501 DAG.getValueType(VA.getValVT()));
2502 else if (VA.getLocInfo() == CCValAssign::ZExt)
2503 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2504 DAG.getValueType(VA.getValVT()));
2505 else if (VA.getLocInfo() == CCValAssign::BCvt)
2506 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2508 if (VA.isExtInLoc()) {
2509 // Handle MMX values passed in XMM regs.
2510 if (RegVT.isVector())
2511 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2513 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2516 assert(VA.isMemLoc());
2517 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2520 // If value is passed via pointer - do a load.
2521 if (VA.getLocInfo() == CCValAssign::Indirect)
2522 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2523 MachinePointerInfo(), false, false, false, 0);
2525 InVals.push_back(ArgValue);
2528 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2529 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2530 // The x86-64 ABIs require that for returning structs by value we copy
2531 // the sret argument into %rax/%eax (depending on ABI) for the return.
2532 // Win32 requires us to put the sret argument to %eax as well.
2533 // Save the argument into a virtual register so that we can access it
2534 // from the return points.
2535 if (Ins[i].Flags.isSRet()) {
2536 unsigned Reg = FuncInfo->getSRetReturnReg();
2538 MVT PtrTy = getPointerTy();
2539 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2540 FuncInfo->setSRetReturnReg(Reg);
2542 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2543 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2549 unsigned StackSize = CCInfo.getNextStackOffset();
2550 // Align stack specially for tail calls.
2551 if (FuncIsMadeTailCallSafe(CallConv,
2552 MF.getTarget().Options.GuaranteedTailCallOpt))
2553 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2555 // If the function takes variable number of arguments, make a frame index for
2556 // the start of the first vararg value... for expansion of llvm.va_start. We
2557 // can skip this if there are no va_start calls.
2558 if (MFI->hasVAStart() &&
2559 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2560 CallConv != CallingConv::X86_ThisCall))) {
2561 FuncInfo->setVarArgsFrameIndex(
2562 MFI->CreateFixedObject(1, StackSize, true));
2565 // Figure out if XMM registers are in use.
2566 assert(!(MF.getTarget().Options.UseSoftFloat &&
2567 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2568 "SSE register cannot be used when SSE is disabled!");
2570 // 64-bit calling conventions support varargs and register parameters, so we
2571 // have to do extra work to spill them in the prologue.
2572 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2573 // Find the first unallocated argument registers.
2574 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2575 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2576 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2577 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2578 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2579 "SSE register cannot be used when SSE is disabled!");
2581 // Gather all the live in physical registers.
2582 SmallVector<SDValue, 6> LiveGPRs;
2583 SmallVector<SDValue, 8> LiveXMMRegs;
2585 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2586 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2588 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2590 if (!ArgXMMs.empty()) {
2591 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2592 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2593 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2594 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2595 LiveXMMRegs.push_back(
2596 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2601 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2602 // Get to the caller-allocated home save location. Add 8 to account
2603 // for the return address.
2604 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2605 FuncInfo->setRegSaveFrameIndex(
2606 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2607 // Fixup to set vararg frame on shadow area (4 x i64).
2609 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2611 // For X86-64, if there are vararg parameters that are passed via
2612 // registers, then we must store them to their spots on the stack so
2613 // they may be loaded by deferencing the result of va_next.
2614 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2615 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2616 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2617 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2620 // Store the integer parameter registers.
2621 SmallVector<SDValue, 8> MemOps;
2622 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2624 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2625 for (SDValue Val : LiveGPRs) {
2626 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2627 DAG.getIntPtrConstant(Offset));
2629 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2630 MachinePointerInfo::getFixedStack(
2631 FuncInfo->getRegSaveFrameIndex(), Offset),
2633 MemOps.push_back(Store);
2637 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2638 // Now store the XMM (fp + vector) parameter registers.
2639 SmallVector<SDValue, 12> SaveXMMOps;
2640 SaveXMMOps.push_back(Chain);
2641 SaveXMMOps.push_back(ALVal);
2642 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2643 FuncInfo->getRegSaveFrameIndex()));
2644 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2645 FuncInfo->getVarArgsFPOffset()));
2646 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2648 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2649 MVT::Other, SaveXMMOps));
2652 if (!MemOps.empty())
2653 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2656 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2657 // Find the largest legal vector type.
2658 MVT VecVT = MVT::Other;
2659 // FIXME: Only some x86_32 calling conventions support AVX512.
2660 if (Subtarget->hasAVX512() &&
2661 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2662 CallConv == CallingConv::Intel_OCL_BI)))
2663 VecVT = MVT::v16f32;
2664 else if (Subtarget->hasAVX())
2666 else if (Subtarget->hasSSE2())
2669 // We forward some GPRs and some vector types.
2670 SmallVector<MVT, 2> RegParmTypes;
2671 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2672 RegParmTypes.push_back(IntVT);
2673 if (VecVT != MVT::Other)
2674 RegParmTypes.push_back(VecVT);
2676 // Compute the set of forwarded registers. The rest are scratch.
2677 SmallVectorImpl<ForwardedRegister> &Forwards =
2678 FuncInfo->getForwardedMustTailRegParms();
2679 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2681 // Conservatively forward AL on x86_64, since it might be used for varargs.
2682 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2683 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2684 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2687 // Copy all forwards from physical to virtual registers.
2688 for (ForwardedRegister &F : Forwards) {
2689 // FIXME: Can we use a less constrained schedule?
2690 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2691 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2692 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2696 // Some CCs need callee pop.
2697 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2698 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2699 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2701 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2702 // If this is an sret function, the return should pop the hidden pointer.
2703 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2704 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2705 argsAreStructReturn(Ins) == StackStructReturn)
2706 FuncInfo->setBytesToPopOnReturn(4);
2710 // RegSaveFrameIndex is X86-64 only.
2711 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2712 if (CallConv == CallingConv::X86_FastCall ||
2713 CallConv == CallingConv::X86_ThisCall)
2714 // fastcc functions can't have varargs.
2715 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2718 FuncInfo->setArgumentStackSize(StackSize);
2724 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2725 SDValue StackPtr, SDValue Arg,
2726 SDLoc dl, SelectionDAG &DAG,
2727 const CCValAssign &VA,
2728 ISD::ArgFlagsTy Flags) const {
2729 unsigned LocMemOffset = VA.getLocMemOffset();
2730 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2731 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2732 if (Flags.isByVal())
2733 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2735 return DAG.getStore(Chain, dl, Arg, PtrOff,
2736 MachinePointerInfo::getStack(LocMemOffset),
2740 /// Emit a load of return address if tail call
2741 /// optimization is performed and it is required.
2743 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2744 SDValue &OutRetAddr, SDValue Chain,
2745 bool IsTailCall, bool Is64Bit,
2746 int FPDiff, SDLoc dl) const {
2747 // Adjust the Return address stack slot.
2748 EVT VT = getPointerTy();
2749 OutRetAddr = getReturnAddressFrameIndex(DAG);
2751 // Load the "old" Return address.
2752 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2753 false, false, false, 0);
2754 return SDValue(OutRetAddr.getNode(), 1);
2757 /// Emit a store of the return address if tail call
2758 /// optimization is performed and it is required (FPDiff!=0).
2759 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2760 SDValue Chain, SDValue RetAddrFrIdx,
2761 EVT PtrVT, unsigned SlotSize,
2762 int FPDiff, SDLoc dl) {
2763 // Store the return address to the appropriate stack slot.
2764 if (!FPDiff) return Chain;
2765 // Calculate the new stack slot for the return address.
2766 int NewReturnAddrFI =
2767 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2769 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2770 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2771 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2777 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2778 SmallVectorImpl<SDValue> &InVals) const {
2779 SelectionDAG &DAG = CLI.DAG;
2781 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2782 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2783 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2784 SDValue Chain = CLI.Chain;
2785 SDValue Callee = CLI.Callee;
2786 CallingConv::ID CallConv = CLI.CallConv;
2787 bool &isTailCall = CLI.IsTailCall;
2788 bool isVarArg = CLI.IsVarArg;
2790 MachineFunction &MF = DAG.getMachineFunction();
2791 bool Is64Bit = Subtarget->is64Bit();
2792 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2793 StructReturnType SR = callIsStructReturn(Outs);
2794 bool IsSibcall = false;
2795 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2797 if (MF.getTarget().Options.DisableTailCalls)
2800 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2802 // Force this to be a tail call. The verifier rules are enough to ensure
2803 // that we can lower this successfully without moving the return address
2806 } else if (isTailCall) {
2807 // Check if it's really possible to do a tail call.
2808 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2809 isVarArg, SR != NotStructReturn,
2810 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2811 Outs, OutVals, Ins, DAG);
2813 // Sibcalls are automatically detected tailcalls which do not require
2815 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2822 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2823 "Var args not supported with calling convention fastcc, ghc or hipe");
2825 // Analyze operands of the call, assigning locations to each operand.
2826 SmallVector<CCValAssign, 16> ArgLocs;
2827 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2829 // Allocate shadow area for Win64
2831 CCInfo.AllocateStack(32, 8);
2833 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2835 // Get a count of how many bytes are to be pushed on the stack.
2836 unsigned NumBytes = CCInfo.getNextStackOffset();
2838 // This is a sibcall. The memory operands are available in caller's
2839 // own caller's stack.
2841 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2842 IsTailCallConvention(CallConv))
2843 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2846 if (isTailCall && !IsSibcall && !IsMustTail) {
2847 // Lower arguments at fp - stackoffset + fpdiff.
2848 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2850 FPDiff = NumBytesCallerPushed - NumBytes;
2852 // Set the delta of movement of the returnaddr stackslot.
2853 // But only set if delta is greater than previous delta.
2854 if (FPDiff < X86Info->getTCReturnAddrDelta())
2855 X86Info->setTCReturnAddrDelta(FPDiff);
2858 unsigned NumBytesToPush = NumBytes;
2859 unsigned NumBytesToPop = NumBytes;
2861 // If we have an inalloca argument, all stack space has already been allocated
2862 // for us and be right at the top of the stack. We don't support multiple
2863 // arguments passed in memory when using inalloca.
2864 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2866 if (!ArgLocs.back().isMemLoc())
2867 report_fatal_error("cannot use inalloca attribute on a register "
2869 if (ArgLocs.back().getLocMemOffset() != 0)
2870 report_fatal_error("any parameter with the inalloca attribute must be "
2871 "the only memory argument");
2875 Chain = DAG.getCALLSEQ_START(
2876 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2878 SDValue RetAddrFrIdx;
2879 // Load return address for tail calls.
2880 if (isTailCall && FPDiff)
2881 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2882 Is64Bit, FPDiff, dl);
2884 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2885 SmallVector<SDValue, 8> MemOpChains;
2888 // Walk the register/memloc assignments, inserting copies/loads. In the case
2889 // of tail call optimization arguments are handle later.
2890 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2891 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2892 // Skip inalloca arguments, they have already been written.
2893 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2894 if (Flags.isInAlloca())
2897 CCValAssign &VA = ArgLocs[i];
2898 EVT RegVT = VA.getLocVT();
2899 SDValue Arg = OutVals[i];
2900 bool isByVal = Flags.isByVal();
2902 // Promote the value if needed.
2903 switch (VA.getLocInfo()) {
2904 default: llvm_unreachable("Unknown loc info!");
2905 case CCValAssign::Full: break;
2906 case CCValAssign::SExt:
2907 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2909 case CCValAssign::ZExt:
2910 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2912 case CCValAssign::AExt:
2913 if (RegVT.is128BitVector()) {
2914 // Special case: passing MMX values in XMM registers.
2915 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2916 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2917 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2919 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2921 case CCValAssign::BCvt:
2922 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2924 case CCValAssign::Indirect: {
2925 // Store the argument.
2926 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2927 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2928 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2929 MachinePointerInfo::getFixedStack(FI),
2936 if (VA.isRegLoc()) {
2937 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2938 if (isVarArg && IsWin64) {
2939 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2940 // shadow reg if callee is a varargs function.
2941 unsigned ShadowReg = 0;
2942 switch (VA.getLocReg()) {
2943 case X86::XMM0: ShadowReg = X86::RCX; break;
2944 case X86::XMM1: ShadowReg = X86::RDX; break;
2945 case X86::XMM2: ShadowReg = X86::R8; break;
2946 case X86::XMM3: ShadowReg = X86::R9; break;
2949 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2951 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2952 assert(VA.isMemLoc());
2953 if (!StackPtr.getNode())
2954 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2956 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2957 dl, DAG, VA, Flags));
2961 if (!MemOpChains.empty())
2962 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2964 if (Subtarget->isPICStyleGOT()) {
2965 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2968 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2969 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2971 // If we are tail calling and generating PIC/GOT style code load the
2972 // address of the callee into ECX. The value in ecx is used as target of
2973 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2974 // for tail calls on PIC/GOT architectures. Normally we would just put the
2975 // address of GOT into ebx and then call target@PLT. But for tail calls
2976 // ebx would be restored (since ebx is callee saved) before jumping to the
2979 // Note: The actual moving to ECX is done further down.
2980 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2981 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2982 !G->getGlobal()->hasProtectedVisibility())
2983 Callee = LowerGlobalAddress(Callee, DAG);
2984 else if (isa<ExternalSymbolSDNode>(Callee))
2985 Callee = LowerExternalSymbol(Callee, DAG);
2989 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2990 // From AMD64 ABI document:
2991 // For calls that may call functions that use varargs or stdargs
2992 // (prototype-less calls or calls to functions containing ellipsis (...) in
2993 // the declaration) %al is used as hidden argument to specify the number
2994 // of SSE registers used. The contents of %al do not need to match exactly
2995 // the number of registers, but must be an ubound on the number of SSE
2996 // registers used and is in the range 0 - 8 inclusive.
2998 // Count the number of XMM registers allocated.
2999 static const MCPhysReg XMMArgRegs[] = {
3000 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3001 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3003 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3004 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3005 && "SSE registers cannot be used when SSE is disabled");
3007 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3008 DAG.getConstant(NumXMMRegs, MVT::i8)));
3011 if (isVarArg && IsMustTail) {
3012 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3013 for (const auto &F : Forwards) {
3014 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3015 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3019 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3020 // don't need this because the eligibility check rejects calls that require
3021 // shuffling arguments passed in memory.
3022 if (!IsSibcall && isTailCall) {
3023 // Force all the incoming stack arguments to be loaded from the stack
3024 // before any new outgoing arguments are stored to the stack, because the
3025 // outgoing stack slots may alias the incoming argument stack slots, and
3026 // the alias isn't otherwise explicit. This is slightly more conservative
3027 // than necessary, because it means that each store effectively depends
3028 // on every argument instead of just those arguments it would clobber.
3029 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3031 SmallVector<SDValue, 8> MemOpChains2;
3034 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3035 CCValAssign &VA = ArgLocs[i];
3038 assert(VA.isMemLoc());
3039 SDValue Arg = OutVals[i];
3040 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3041 // Skip inalloca arguments. They don't require any work.
3042 if (Flags.isInAlloca())
3044 // Create frame index.
3045 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3046 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3047 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3048 FIN = DAG.getFrameIndex(FI, getPointerTy());
3050 if (Flags.isByVal()) {
3051 // Copy relative to framepointer.
3052 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3053 if (!StackPtr.getNode())
3054 StackPtr = DAG.getCopyFromReg(Chain, dl,
3055 RegInfo->getStackRegister(),
3057 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3059 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3063 // Store relative to framepointer.
3064 MemOpChains2.push_back(
3065 DAG.getStore(ArgChain, dl, Arg, FIN,
3066 MachinePointerInfo::getFixedStack(FI),
3071 if (!MemOpChains2.empty())
3072 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3074 // Store the return address to the appropriate stack slot.
3075 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3076 getPointerTy(), RegInfo->getSlotSize(),
3080 // Build a sequence of copy-to-reg nodes chained together with token chain
3081 // and flag operands which copy the outgoing args into registers.
3083 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3084 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3085 RegsToPass[i].second, InFlag);
3086 InFlag = Chain.getValue(1);
3089 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3090 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3091 // In the 64-bit large code model, we have to make all calls
3092 // through a register, since the call instruction's 32-bit
3093 // pc-relative offset may not be large enough to hold the whole
3095 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3096 // If the callee is a GlobalAddress node (quite common, every direct call
3097 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3099 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3101 // We should use extra load for direct calls to dllimported functions in
3103 const GlobalValue *GV = G->getGlobal();
3104 if (!GV->hasDLLImportStorageClass()) {
3105 unsigned char OpFlags = 0;
3106 bool ExtraLoad = false;
3107 unsigned WrapperKind = ISD::DELETED_NODE;
3109 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3110 // external symbols most go through the PLT in PIC mode. If the symbol
3111 // has hidden or protected visibility, or if it is static or local, then
3112 // we don't need to use the PLT - we can directly call it.
3113 if (Subtarget->isTargetELF() &&
3114 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3115 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3116 OpFlags = X86II::MO_PLT;
3117 } else if (Subtarget->isPICStyleStubAny() &&
3118 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3119 (!Subtarget->getTargetTriple().isMacOSX() ||
3120 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3121 // PC-relative references to external symbols should go through $stub,
3122 // unless we're building with the leopard linker or later, which
3123 // automatically synthesizes these stubs.
3124 OpFlags = X86II::MO_DARWIN_STUB;
3125 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3126 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3127 // If the function is marked as non-lazy, generate an indirect call
3128 // which loads from the GOT directly. This avoids runtime overhead
3129 // at the cost of eager binding (and one extra byte of encoding).
3130 OpFlags = X86II::MO_GOTPCREL;
3131 WrapperKind = X86ISD::WrapperRIP;
3135 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3136 G->getOffset(), OpFlags);
3138 // Add a wrapper if needed.
3139 if (WrapperKind != ISD::DELETED_NODE)
3140 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3141 // Add extra indirection if needed.
3143 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3144 MachinePointerInfo::getGOT(),
3145 false, false, false, 0);
3147 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3148 unsigned char OpFlags = 0;
3150 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3151 // external symbols should go through the PLT.
3152 if (Subtarget->isTargetELF() &&
3153 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3154 OpFlags = X86II::MO_PLT;
3155 } else if (Subtarget->isPICStyleStubAny() &&
3156 (!Subtarget->getTargetTriple().isMacOSX() ||
3157 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3158 // PC-relative references to external symbols should go through $stub,
3159 // unless we're building with the leopard linker or later, which
3160 // automatically synthesizes these stubs.
3161 OpFlags = X86II::MO_DARWIN_STUB;
3164 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3166 } else if (Subtarget->isTarget64BitILP32() &&
3167 Callee->getValueType(0) == MVT::i32) {
3168 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3169 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3172 // Returns a chain & a flag for retval copy to use.
3173 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3174 SmallVector<SDValue, 8> Ops;
3176 if (!IsSibcall && isTailCall) {
3177 Chain = DAG.getCALLSEQ_END(Chain,
3178 DAG.getIntPtrConstant(NumBytesToPop, true),
3179 DAG.getIntPtrConstant(0, true), InFlag, dl);
3180 InFlag = Chain.getValue(1);
3183 Ops.push_back(Chain);
3184 Ops.push_back(Callee);
3187 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3189 // Add argument registers to the end of the list so that they are known live
3191 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3192 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3193 RegsToPass[i].second.getValueType()));
3195 // Add a register mask operand representing the call-preserved registers.
3196 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3197 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3198 assert(Mask && "Missing call preserved mask for calling convention");
3199 Ops.push_back(DAG.getRegisterMask(Mask));
3201 if (InFlag.getNode())
3202 Ops.push_back(InFlag);
3206 //// If this is the first return lowered for this function, add the regs
3207 //// to the liveout set for the function.
3208 // This isn't right, although it's probably harmless on x86; liveouts
3209 // should be computed from returns not tail calls. Consider a void
3210 // function making a tail call to a function returning int.
3211 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3214 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3215 InFlag = Chain.getValue(1);
3217 // Create the CALLSEQ_END node.
3218 unsigned NumBytesForCalleeToPop;
3219 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3220 DAG.getTarget().Options.GuaranteedTailCallOpt))
3221 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3222 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3223 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3224 SR == StackStructReturn)
3225 // If this is a call to a struct-return function, the callee
3226 // pops the hidden struct pointer, so we have to push it back.
3227 // This is common for Darwin/X86, Linux & Mingw32 targets.
3228 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3229 NumBytesForCalleeToPop = 4;
3231 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3233 // Returns a flag for retval copy to use.
3235 Chain = DAG.getCALLSEQ_END(Chain,
3236 DAG.getIntPtrConstant(NumBytesToPop, true),
3237 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3240 InFlag = Chain.getValue(1);
3243 // Handle result values, copying them out of physregs into vregs that we
3245 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3246 Ins, dl, DAG, InVals);
3249 //===----------------------------------------------------------------------===//
3250 // Fast Calling Convention (tail call) implementation
3251 //===----------------------------------------------------------------------===//
3253 // Like std call, callee cleans arguments, convention except that ECX is
3254 // reserved for storing the tail called function address. Only 2 registers are
3255 // free for argument passing (inreg). Tail call optimization is performed
3257 // * tailcallopt is enabled
3258 // * caller/callee are fastcc
3259 // On X86_64 architecture with GOT-style position independent code only local
3260 // (within module) calls are supported at the moment.
3261 // To keep the stack aligned according to platform abi the function
3262 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3263 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3264 // If a tail called function callee has more arguments than the caller the
3265 // caller needs to make sure that there is room to move the RETADDR to. This is
3266 // achieved by reserving an area the size of the argument delta right after the
3267 // original RETADDR, but before the saved framepointer or the spilled registers
3268 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3280 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3281 /// for a 16 byte align requirement.
3283 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3284 SelectionDAG& DAG) const {
3285 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3286 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3287 unsigned StackAlignment = TFI.getStackAlignment();
3288 uint64_t AlignMask = StackAlignment - 1;
3289 int64_t Offset = StackSize;
3290 unsigned SlotSize = RegInfo->getSlotSize();
3291 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3292 // Number smaller than 12 so just add the difference.
3293 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3295 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3296 Offset = ((~AlignMask) & Offset) + StackAlignment +
3297 (StackAlignment-SlotSize);
3302 /// MatchingStackOffset - Return true if the given stack call argument is
3303 /// already available in the same position (relatively) of the caller's
3304 /// incoming argument stack.
3306 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3307 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3308 const X86InstrInfo *TII) {
3309 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3311 if (Arg.getOpcode() == ISD::CopyFromReg) {
3312 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3313 if (!TargetRegisterInfo::isVirtualRegister(VR))
3315 MachineInstr *Def = MRI->getVRegDef(VR);
3318 if (!Flags.isByVal()) {
3319 if (!TII->isLoadFromStackSlot(Def, FI))
3322 unsigned Opcode = Def->getOpcode();
3323 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3324 Opcode == X86::LEA64_32r) &&
3325 Def->getOperand(1).isFI()) {
3326 FI = Def->getOperand(1).getIndex();
3327 Bytes = Flags.getByValSize();
3331 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3332 if (Flags.isByVal())
3333 // ByVal argument is passed in as a pointer but it's now being
3334 // dereferenced. e.g.
3335 // define @foo(%struct.X* %A) {
3336 // tail call @bar(%struct.X* byval %A)
3339 SDValue Ptr = Ld->getBasePtr();
3340 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3343 FI = FINode->getIndex();
3344 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3345 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3346 FI = FINode->getIndex();
3347 Bytes = Flags.getByValSize();
3351 assert(FI != INT_MAX);
3352 if (!MFI->isFixedObjectIndex(FI))
3354 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3357 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3358 /// for tail call optimization. Targets which want to do tail call
3359 /// optimization should implement this function.
3361 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3362 CallingConv::ID CalleeCC,
3364 bool isCalleeStructRet,
3365 bool isCallerStructRet,
3367 const SmallVectorImpl<ISD::OutputArg> &Outs,
3368 const SmallVectorImpl<SDValue> &OutVals,
3369 const SmallVectorImpl<ISD::InputArg> &Ins,
3370 SelectionDAG &DAG) const {
3371 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3374 // If -tailcallopt is specified, make fastcc functions tail-callable.
3375 const MachineFunction &MF = DAG.getMachineFunction();
3376 const Function *CallerF = MF.getFunction();
3378 // If the function return type is x86_fp80 and the callee return type is not,
3379 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3380 // perform a tailcall optimization here.
3381 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3384 CallingConv::ID CallerCC = CallerF->getCallingConv();
3385 bool CCMatch = CallerCC == CalleeCC;
3386 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3387 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3389 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3390 if (IsTailCallConvention(CalleeCC) && CCMatch)
3395 // Look for obvious safe cases to perform tail call optimization that do not
3396 // require ABI changes. This is what gcc calls sibcall.
3398 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3399 // emit a special epilogue.
3400 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3401 if (RegInfo->needsStackRealignment(MF))
3404 // Also avoid sibcall optimization if either caller or callee uses struct
3405 // return semantics.
3406 if (isCalleeStructRet || isCallerStructRet)
3409 // An stdcall/thiscall caller is expected to clean up its arguments; the
3410 // callee isn't going to do that.
3411 // FIXME: this is more restrictive than needed. We could produce a tailcall
3412 // when the stack adjustment matches. For example, with a thiscall that takes
3413 // only one argument.
3414 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3415 CallerCC == CallingConv::X86_ThisCall))
3418 // Do not sibcall optimize vararg calls unless all arguments are passed via
3420 if (isVarArg && !Outs.empty()) {
3422 // Optimizing for varargs on Win64 is unlikely to be safe without
3423 // additional testing.
3424 if (IsCalleeWin64 || IsCallerWin64)
3427 SmallVector<CCValAssign, 16> ArgLocs;
3428 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3431 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3432 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3433 if (!ArgLocs[i].isRegLoc())
3437 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3438 // stack. Therefore, if it's not used by the call it is not safe to optimize
3439 // this into a sibcall.
3440 bool Unused = false;
3441 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3448 SmallVector<CCValAssign, 16> RVLocs;
3449 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3451 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3452 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3453 CCValAssign &VA = RVLocs[i];
3454 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3459 // If the calling conventions do not match, then we'd better make sure the
3460 // results are returned in the same way as what the caller expects.
3462 SmallVector<CCValAssign, 16> RVLocs1;
3463 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3465 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3467 SmallVector<CCValAssign, 16> RVLocs2;
3468 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3470 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3472 if (RVLocs1.size() != RVLocs2.size())
3474 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3475 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3477 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3479 if (RVLocs1[i].isRegLoc()) {
3480 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3483 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3489 // If the callee takes no arguments then go on to check the results of the
3491 if (!Outs.empty()) {
3492 // Check if stack adjustment is needed. For now, do not do this if any
3493 // argument is passed on the stack.
3494 SmallVector<CCValAssign, 16> ArgLocs;
3495 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3498 // Allocate shadow area for Win64
3500 CCInfo.AllocateStack(32, 8);
3502 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3503 if (CCInfo.getNextStackOffset()) {
3504 MachineFunction &MF = DAG.getMachineFunction();
3505 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3508 // Check if the arguments are already laid out in the right way as
3509 // the caller's fixed stack objects.
3510 MachineFrameInfo *MFI = MF.getFrameInfo();
3511 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3512 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3513 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3514 CCValAssign &VA = ArgLocs[i];
3515 SDValue Arg = OutVals[i];
3516 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3517 if (VA.getLocInfo() == CCValAssign::Indirect)
3519 if (!VA.isRegLoc()) {
3520 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3527 // If the tailcall address may be in a register, then make sure it's
3528 // possible to register allocate for it. In 32-bit, the call address can
3529 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3530 // callee-saved registers are restored. These happen to be the same
3531 // registers used to pass 'inreg' arguments so watch out for those.
3532 if (!Subtarget->is64Bit() &&
3533 ((!isa<GlobalAddressSDNode>(Callee) &&
3534 !isa<ExternalSymbolSDNode>(Callee)) ||
3535 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3536 unsigned NumInRegs = 0;
3537 // In PIC we need an extra register to formulate the address computation
3539 unsigned MaxInRegs =
3540 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3542 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3543 CCValAssign &VA = ArgLocs[i];
3546 unsigned Reg = VA.getLocReg();
3549 case X86::EAX: case X86::EDX: case X86::ECX:
3550 if (++NumInRegs == MaxInRegs)
3562 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3563 const TargetLibraryInfo *libInfo) const {
3564 return X86::createFastISel(funcInfo, libInfo);
3567 //===----------------------------------------------------------------------===//
3568 // Other Lowering Hooks
3569 //===----------------------------------------------------------------------===//
3571 static bool MayFoldLoad(SDValue Op) {
3572 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3575 static bool MayFoldIntoStore(SDValue Op) {
3576 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3579 static bool isTargetShuffle(unsigned Opcode) {
3581 default: return false;
3582 case X86ISD::BLENDI:
3583 case X86ISD::PSHUFB:
3584 case X86ISD::PSHUFD:
3585 case X86ISD::PSHUFHW:
3586 case X86ISD::PSHUFLW:
3588 case X86ISD::PALIGNR:
3589 case X86ISD::MOVLHPS:
3590 case X86ISD::MOVLHPD:
3591 case X86ISD::MOVHLPS:
3592 case X86ISD::MOVLPS:
3593 case X86ISD::MOVLPD:
3594 case X86ISD::MOVSHDUP:
3595 case X86ISD::MOVSLDUP:
3596 case X86ISD::MOVDDUP:
3599 case X86ISD::UNPCKL:
3600 case X86ISD::UNPCKH:
3601 case X86ISD::VPERMILPI:
3602 case X86ISD::VPERM2X128:
3603 case X86ISD::VPERMI:
3608 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3609 SDValue V1, unsigned TargetMask,
3610 SelectionDAG &DAG) {
3612 default: llvm_unreachable("Unknown x86 shuffle node");
3613 case X86ISD::PSHUFD:
3614 case X86ISD::PSHUFHW:
3615 case X86ISD::PSHUFLW:
3616 case X86ISD::VPERMILPI:
3617 case X86ISD::VPERMI:
3618 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3622 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3623 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3625 default: llvm_unreachable("Unknown x86 shuffle node");
3626 case X86ISD::MOVLHPS:
3627 case X86ISD::MOVLHPD:
3628 case X86ISD::MOVHLPS:
3629 case X86ISD::MOVLPS:
3630 case X86ISD::MOVLPD:
3633 case X86ISD::UNPCKL:
3634 case X86ISD::UNPCKH:
3635 return DAG.getNode(Opc, dl, VT, V1, V2);
3639 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3640 MachineFunction &MF = DAG.getMachineFunction();
3641 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3642 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3643 int ReturnAddrIndex = FuncInfo->getRAIndex();
3645 if (ReturnAddrIndex == 0) {
3646 // Set up a frame object for the return address.
3647 unsigned SlotSize = RegInfo->getSlotSize();
3648 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3651 FuncInfo->setRAIndex(ReturnAddrIndex);
3654 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3657 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3658 bool hasSymbolicDisplacement) {
3659 // Offset should fit into 32 bit immediate field.
3660 if (!isInt<32>(Offset))
3663 // If we don't have a symbolic displacement - we don't have any extra
3665 if (!hasSymbolicDisplacement)
3668 // FIXME: Some tweaks might be needed for medium code model.
3669 if (M != CodeModel::Small && M != CodeModel::Kernel)
3672 // For small code model we assume that latest object is 16MB before end of 31
3673 // bits boundary. We may also accept pretty large negative constants knowing
3674 // that all objects are in the positive half of address space.
3675 if (M == CodeModel::Small && Offset < 16*1024*1024)
3678 // For kernel code model we know that all object resist in the negative half
3679 // of 32bits address space. We may not accept negative offsets, since they may
3680 // be just off and we may accept pretty large positive ones.
3681 if (M == CodeModel::Kernel && Offset >= 0)
3687 /// isCalleePop - Determines whether the callee is required to pop its
3688 /// own arguments. Callee pop is necessary to support tail calls.
3689 bool X86::isCalleePop(CallingConv::ID CallingConv,
3690 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3691 switch (CallingConv) {
3694 case CallingConv::X86_StdCall:
3695 case CallingConv::X86_FastCall:
3696 case CallingConv::X86_ThisCall:
3698 case CallingConv::Fast:
3699 case CallingConv::GHC:
3700 case CallingConv::HiPE:
3707 /// \brief Return true if the condition is an unsigned comparison operation.
3708 static bool isX86CCUnsigned(unsigned X86CC) {
3710 default: llvm_unreachable("Invalid integer condition!");
3711 case X86::COND_E: return true;
3712 case X86::COND_G: return false;
3713 case X86::COND_GE: return false;
3714 case X86::COND_L: return false;
3715 case X86::COND_LE: return false;
3716 case X86::COND_NE: return true;
3717 case X86::COND_B: return true;
3718 case X86::COND_A: return true;
3719 case X86::COND_BE: return true;
3720 case X86::COND_AE: return true;
3722 llvm_unreachable("covered switch fell through?!");
3725 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3726 /// specific condition code, returning the condition code and the LHS/RHS of the
3727 /// comparison to make.
3728 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3729 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3731 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3732 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3733 // X > -1 -> X == 0, jump !sign.
3734 RHS = DAG.getConstant(0, RHS.getValueType());
3735 return X86::COND_NS;
3737 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3738 // X < 0 -> X == 0, jump on sign.
3741 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3743 RHS = DAG.getConstant(0, RHS.getValueType());
3744 return X86::COND_LE;
3748 switch (SetCCOpcode) {
3749 default: llvm_unreachable("Invalid integer condition!");
3750 case ISD::SETEQ: return X86::COND_E;
3751 case ISD::SETGT: return X86::COND_G;
3752 case ISD::SETGE: return X86::COND_GE;
3753 case ISD::SETLT: return X86::COND_L;
3754 case ISD::SETLE: return X86::COND_LE;
3755 case ISD::SETNE: return X86::COND_NE;
3756 case ISD::SETULT: return X86::COND_B;
3757 case ISD::SETUGT: return X86::COND_A;
3758 case ISD::SETULE: return X86::COND_BE;
3759 case ISD::SETUGE: return X86::COND_AE;
3763 // First determine if it is required or is profitable to flip the operands.
3765 // If LHS is a foldable load, but RHS is not, flip the condition.
3766 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3767 !ISD::isNON_EXTLoad(RHS.getNode())) {
3768 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3769 std::swap(LHS, RHS);
3772 switch (SetCCOpcode) {
3778 std::swap(LHS, RHS);
3782 // On a floating point condition, the flags are set as follows:
3784 // 0 | 0 | 0 | X > Y
3785 // 0 | 0 | 1 | X < Y
3786 // 1 | 0 | 0 | X == Y
3787 // 1 | 1 | 1 | unordered
3788 switch (SetCCOpcode) {
3789 default: llvm_unreachable("Condcode should be pre-legalized away");
3791 case ISD::SETEQ: return X86::COND_E;
3792 case ISD::SETOLT: // flipped
3794 case ISD::SETGT: return X86::COND_A;
3795 case ISD::SETOLE: // flipped
3797 case ISD::SETGE: return X86::COND_AE;
3798 case ISD::SETUGT: // flipped
3800 case ISD::SETLT: return X86::COND_B;
3801 case ISD::SETUGE: // flipped
3803 case ISD::SETLE: return X86::COND_BE;
3805 case ISD::SETNE: return X86::COND_NE;
3806 case ISD::SETUO: return X86::COND_P;
3807 case ISD::SETO: return X86::COND_NP;
3809 case ISD::SETUNE: return X86::COND_INVALID;
3813 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3814 /// code. Current x86 isa includes the following FP cmov instructions:
3815 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3816 static bool hasFPCMov(unsigned X86CC) {
3832 /// isFPImmLegal - Returns true if the target can instruction select the
3833 /// specified FP immediate natively. If false, the legalizer will
3834 /// materialize the FP immediate as a load from a constant pool.
3835 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3836 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3837 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3843 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3844 ISD::LoadExtType ExtTy,
3846 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3847 // relocation target a movq or addq instruction: don't let the load shrink.
3848 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3849 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3850 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3851 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3855 /// \brief Returns true if it is beneficial to convert a load of a constant
3856 /// to just the constant itself.
3857 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3859 assert(Ty->isIntegerTy());
3861 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3862 if (BitSize == 0 || BitSize > 64)
3867 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3868 unsigned Index) const {
3869 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3872 return (Index == 0 || Index == ResVT.getVectorNumElements());
3875 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3876 // Speculate cttz only if we can directly use TZCNT.
3877 return Subtarget->hasBMI();
3880 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3881 // Speculate ctlz only if we can directly use LZCNT.
3882 return Subtarget->hasLZCNT();
3885 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3886 /// the specified range (L, H].
3887 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3888 return (Val < 0) || (Val >= Low && Val < Hi);
3891 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3892 /// specified value.
3893 static bool isUndefOrEqual(int Val, int CmpVal) {
3894 return (Val < 0 || Val == CmpVal);
3897 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3898 /// from position Pos and ending in Pos+Size, falls within the specified
3899 /// sequential range (Low, Low+Size]. or is undef.
3900 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3901 unsigned Pos, unsigned Size, int Low) {
3902 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3903 if (!isUndefOrEqual(Mask[i], Low))
3908 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3909 /// the two vector operands have swapped position.
3910 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3911 unsigned NumElems) {
3912 for (unsigned i = 0; i != NumElems; ++i) {
3916 else if (idx < (int)NumElems)
3917 Mask[i] = idx + NumElems;
3919 Mask[i] = idx - NumElems;
3923 /// isVEXTRACTIndex - Return true if the specified
3924 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3925 /// suitable for instruction that extract 128 or 256 bit vectors
3926 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3927 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3928 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3931 // The index should be aligned on a vecWidth-bit boundary.
3933 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3935 MVT VT = N->getSimpleValueType(0);
3936 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3937 bool Result = (Index * ElSize) % vecWidth == 0;
3942 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3943 /// operand specifies a subvector insert that is suitable for input to
3944 /// insertion of 128 or 256-bit subvectors
3945 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3946 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3947 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3949 // The index should be aligned on a vecWidth-bit boundary.
3951 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3953 MVT VT = N->getSimpleValueType(0);
3954 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3955 bool Result = (Index * ElSize) % vecWidth == 0;
3960 bool X86::isVINSERT128Index(SDNode *N) {
3961 return isVINSERTIndex(N, 128);
3964 bool X86::isVINSERT256Index(SDNode *N) {
3965 return isVINSERTIndex(N, 256);
3968 bool X86::isVEXTRACT128Index(SDNode *N) {
3969 return isVEXTRACTIndex(N, 128);
3972 bool X86::isVEXTRACT256Index(SDNode *N) {
3973 return isVEXTRACTIndex(N, 256);
3976 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3977 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3978 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3979 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3982 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3984 MVT VecVT = N->getOperand(0).getSimpleValueType();
3985 MVT ElVT = VecVT.getVectorElementType();
3987 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3988 return Index / NumElemsPerChunk;
3991 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3992 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3993 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3994 llvm_unreachable("Illegal insert subvector for VINSERT");
3997 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3999 MVT VecVT = N->getSimpleValueType(0);
4000 MVT ElVT = VecVT.getVectorElementType();
4002 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4003 return Index / NumElemsPerChunk;
4006 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4007 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4008 /// and VINSERTI128 instructions.
4009 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4010 return getExtractVEXTRACTImmediate(N, 128);
4013 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4014 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4015 /// and VINSERTI64x4 instructions.
4016 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4017 return getExtractVEXTRACTImmediate(N, 256);
4020 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4021 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4022 /// and VINSERTI128 instructions.
4023 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4024 return getInsertVINSERTImmediate(N, 128);
4027 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4028 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4029 /// and VINSERTI64x4 instructions.
4030 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4031 return getInsertVINSERTImmediate(N, 256);
4034 /// isZero - Returns true if Elt is a constant integer zero
4035 static bool isZero(SDValue V) {
4036 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4037 return C && C->isNullValue();
4040 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4042 bool X86::isZeroNode(SDValue Elt) {
4045 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4046 return CFP->getValueAPF().isPosZero();
4050 /// getZeroVector - Returns a vector of specified type with all zero elements.
4052 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4053 SelectionDAG &DAG, SDLoc dl) {
4054 assert(VT.isVector() && "Expected a vector type");
4056 // Always build SSE zero vectors as <4 x i32> bitcasted
4057 // to their dest type. This ensures they get CSE'd.
4059 if (VT.is128BitVector()) { // SSE
4060 if (Subtarget->hasSSE2()) { // SSE2
4061 SDValue Cst = DAG.getConstant(0, MVT::i32);
4062 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4064 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4065 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4067 } else if (VT.is256BitVector()) { // AVX
4068 if (Subtarget->hasInt256()) { // AVX2
4069 SDValue Cst = DAG.getConstant(0, MVT::i32);
4070 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4071 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4073 // 256-bit logic and arithmetic instructions in AVX are all
4074 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4075 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4076 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4077 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4079 } else if (VT.is512BitVector()) { // AVX-512
4080 SDValue Cst = DAG.getConstant(0, MVT::i32);
4081 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4082 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4083 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4084 } else if (VT.getScalarType() == MVT::i1) {
4085 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
4086 SDValue Cst = DAG.getConstant(0, MVT::i1);
4087 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
4088 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4090 llvm_unreachable("Unexpected vector type");
4092 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4095 /// getOnesVector - Returns a vector of specified type with all bits set.
4096 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4097 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4098 /// Then bitcast to their original type, ensuring they get CSE'd.
4099 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4101 assert(VT.isVector() && "Expected a vector type");
4103 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
4105 if (VT.is256BitVector()) {
4106 if (HasInt256) { // AVX2
4107 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4108 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4110 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4111 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4113 } else if (VT.is128BitVector()) {
4114 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4116 llvm_unreachable("Unexpected vector type");
4118 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4121 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4122 /// operation of specified width.
4123 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4125 unsigned NumElems = VT.getVectorNumElements();
4126 SmallVector<int, 8> Mask;
4127 Mask.push_back(NumElems);
4128 for (unsigned i = 1; i != NumElems; ++i)
4130 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4133 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4134 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4136 unsigned NumElems = VT.getVectorNumElements();
4137 SmallVector<int, 8> Mask;
4138 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4140 Mask.push_back(i + NumElems);
4142 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4145 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4146 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4148 unsigned NumElems = VT.getVectorNumElements();
4149 SmallVector<int, 8> Mask;
4150 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4151 Mask.push_back(i + Half);
4152 Mask.push_back(i + NumElems + Half);
4154 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4157 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4158 /// vector of zero or undef vector. This produces a shuffle where the low
4159 /// element of V2 is swizzled into the zero/undef vector, landing at element
4160 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4161 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4163 const X86Subtarget *Subtarget,
4164 SelectionDAG &DAG) {
4165 MVT VT = V2.getSimpleValueType();
4167 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4168 unsigned NumElems = VT.getVectorNumElements();
4169 SmallVector<int, 16> MaskVec;
4170 for (unsigned i = 0; i != NumElems; ++i)
4171 // If this is the insertion idx, put the low elt of V2 here.
4172 MaskVec.push_back(i == Idx ? NumElems : i);
4173 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4176 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4177 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4178 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4179 /// shuffles which use a single input multiple times, and in those cases it will
4180 /// adjust the mask to only have indices within that single input.
4181 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4182 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4183 unsigned NumElems = VT.getVectorNumElements();
4187 bool IsFakeUnary = false;
4188 switch(N->getOpcode()) {
4189 case X86ISD::BLENDI:
4190 ImmN = N->getOperand(N->getNumOperands()-1);
4191 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4194 ImmN = N->getOperand(N->getNumOperands()-1);
4195 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4196 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4198 case X86ISD::UNPCKH:
4199 DecodeUNPCKHMask(VT, Mask);
4200 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4202 case X86ISD::UNPCKL:
4203 DecodeUNPCKLMask(VT, Mask);
4204 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4206 case X86ISD::MOVHLPS:
4207 DecodeMOVHLPSMask(NumElems, Mask);
4208 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4210 case X86ISD::MOVLHPS:
4211 DecodeMOVLHPSMask(NumElems, Mask);
4212 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4214 case X86ISD::PALIGNR:
4215 ImmN = N->getOperand(N->getNumOperands()-1);
4216 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4218 case X86ISD::PSHUFD:
4219 case X86ISD::VPERMILPI:
4220 ImmN = N->getOperand(N->getNumOperands()-1);
4221 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4224 case X86ISD::PSHUFHW:
4225 ImmN = N->getOperand(N->getNumOperands()-1);
4226 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4229 case X86ISD::PSHUFLW:
4230 ImmN = N->getOperand(N->getNumOperands()-1);
4231 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4234 case X86ISD::PSHUFB: {
4236 SDValue MaskNode = N->getOperand(1);
4237 while (MaskNode->getOpcode() == ISD::BITCAST)
4238 MaskNode = MaskNode->getOperand(0);
4240 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4241 // If we have a build-vector, then things are easy.
4242 EVT VT = MaskNode.getValueType();
4243 assert(VT.isVector() &&
4244 "Can't produce a non-vector with a build_vector!");
4245 if (!VT.isInteger())
4248 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4250 SmallVector<uint64_t, 32> RawMask;
4251 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4252 SDValue Op = MaskNode->getOperand(i);
4253 if (Op->getOpcode() == ISD::UNDEF) {
4254 RawMask.push_back((uint64_t)SM_SentinelUndef);
4257 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4260 APInt MaskElement = CN->getAPIntValue();
4262 // We now have to decode the element which could be any integer size and
4263 // extract each byte of it.
4264 for (int j = 0; j < NumBytesPerElement; ++j) {
4265 // Note that this is x86 and so always little endian: the low byte is
4266 // the first byte of the mask.
4267 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4268 MaskElement = MaskElement.lshr(8);
4271 DecodePSHUFBMask(RawMask, Mask);
4275 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4279 SDValue Ptr = MaskLoad->getBasePtr();
4280 if (Ptr->getOpcode() == X86ISD::Wrapper)
4281 Ptr = Ptr->getOperand(0);
4283 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4284 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4287 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4288 DecodePSHUFBMask(C, Mask);
4296 case X86ISD::VPERMI:
4297 ImmN = N->getOperand(N->getNumOperands()-1);
4298 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4303 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4305 case X86ISD::VPERM2X128:
4306 ImmN = N->getOperand(N->getNumOperands()-1);
4307 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4308 if (Mask.empty()) return false;
4310 case X86ISD::MOVSLDUP:
4311 DecodeMOVSLDUPMask(VT, Mask);
4314 case X86ISD::MOVSHDUP:
4315 DecodeMOVSHDUPMask(VT, Mask);
4318 case X86ISD::MOVDDUP:
4319 DecodeMOVDDUPMask(VT, Mask);
4322 case X86ISD::MOVLHPD:
4323 case X86ISD::MOVLPD:
4324 case X86ISD::MOVLPS:
4325 // Not yet implemented
4327 default: llvm_unreachable("unknown target shuffle node");
4330 // If we have a fake unary shuffle, the shuffle mask is spread across two
4331 // inputs that are actually the same node. Re-map the mask to always point
4332 // into the first input.
4335 if (M >= (int)Mask.size())
4341 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4342 /// element of the result of the vector shuffle.
4343 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4346 return SDValue(); // Limit search depth.
4348 SDValue V = SDValue(N, 0);
4349 EVT VT = V.getValueType();
4350 unsigned Opcode = V.getOpcode();
4352 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4353 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4354 int Elt = SV->getMaskElt(Index);
4357 return DAG.getUNDEF(VT.getVectorElementType());
4359 unsigned NumElems = VT.getVectorNumElements();
4360 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4361 : SV->getOperand(1);
4362 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4365 // Recurse into target specific vector shuffles to find scalars.
4366 if (isTargetShuffle(Opcode)) {
4367 MVT ShufVT = V.getSimpleValueType();
4368 unsigned NumElems = ShufVT.getVectorNumElements();
4369 SmallVector<int, 16> ShuffleMask;
4372 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4375 int Elt = ShuffleMask[Index];
4377 return DAG.getUNDEF(ShufVT.getVectorElementType());
4379 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4381 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4385 // Actual nodes that may contain scalar elements
4386 if (Opcode == ISD::BITCAST) {
4387 V = V.getOperand(0);
4388 EVT SrcVT = V.getValueType();
4389 unsigned NumElems = VT.getVectorNumElements();
4391 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4395 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4396 return (Index == 0) ? V.getOperand(0)
4397 : DAG.getUNDEF(VT.getVectorElementType());
4399 if (V.getOpcode() == ISD::BUILD_VECTOR)
4400 return V.getOperand(Index);
4405 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4407 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4408 unsigned NumNonZero, unsigned NumZero,
4410 const X86Subtarget* Subtarget,
4411 const TargetLowering &TLI) {
4418 for (unsigned i = 0; i < 16; ++i) {
4419 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4420 if (ThisIsNonZero && First) {
4422 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4424 V = DAG.getUNDEF(MVT::v8i16);
4429 SDValue ThisElt, LastElt;
4430 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4431 if (LastIsNonZero) {
4432 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4433 MVT::i16, Op.getOperand(i-1));
4435 if (ThisIsNonZero) {
4436 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4437 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4438 ThisElt, DAG.getConstant(8, MVT::i8));
4440 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4444 if (ThisElt.getNode())
4445 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4446 DAG.getIntPtrConstant(i/2));
4450 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4453 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4455 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4456 unsigned NumNonZero, unsigned NumZero,
4458 const X86Subtarget* Subtarget,
4459 const TargetLowering &TLI) {
4466 for (unsigned i = 0; i < 8; ++i) {
4467 bool isNonZero = (NonZeros & (1 << i)) != 0;
4471 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4473 V = DAG.getUNDEF(MVT::v8i16);
4476 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4477 MVT::v8i16, V, Op.getOperand(i),
4478 DAG.getIntPtrConstant(i));
4485 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4486 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4487 const X86Subtarget *Subtarget,
4488 const TargetLowering &TLI) {
4489 // Find all zeroable elements.
4490 std::bitset<4> Zeroable;
4491 for (int i=0; i < 4; ++i) {
4492 SDValue Elt = Op->getOperand(i);
4493 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4495 assert(Zeroable.size() - Zeroable.count() > 1 &&
4496 "We expect at least two non-zero elements!");
4498 // We only know how to deal with build_vector nodes where elements are either
4499 // zeroable or extract_vector_elt with constant index.
4500 SDValue FirstNonZero;
4501 unsigned FirstNonZeroIdx;
4502 for (unsigned i=0; i < 4; ++i) {
4505 SDValue Elt = Op->getOperand(i);
4506 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4507 !isa<ConstantSDNode>(Elt.getOperand(1)))
4509 // Make sure that this node is extracting from a 128-bit vector.
4510 MVT VT = Elt.getOperand(0).getSimpleValueType();
4511 if (!VT.is128BitVector())
4513 if (!FirstNonZero.getNode()) {
4515 FirstNonZeroIdx = i;
4519 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4520 SDValue V1 = FirstNonZero.getOperand(0);
4521 MVT VT = V1.getSimpleValueType();
4523 // See if this build_vector can be lowered as a blend with zero.
4525 unsigned EltMaskIdx, EltIdx;
4527 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4528 if (Zeroable[EltIdx]) {
4529 // The zero vector will be on the right hand side.
4530 Mask[EltIdx] = EltIdx+4;
4534 Elt = Op->getOperand(EltIdx);
4535 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4536 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4537 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4539 Mask[EltIdx] = EltIdx;
4543 // Let the shuffle legalizer deal with blend operations.
4544 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4545 if (V1.getSimpleValueType() != VT)
4546 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4547 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4550 // See if we can lower this build_vector to a INSERTPS.
4551 if (!Subtarget->hasSSE41())
4554 SDValue V2 = Elt.getOperand(0);
4555 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4558 bool CanFold = true;
4559 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4563 SDValue Current = Op->getOperand(i);
4564 SDValue SrcVector = Current->getOperand(0);
4567 CanFold = SrcVector == V1 &&
4568 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4574 assert(V1.getNode() && "Expected at least two non-zero elements!");
4575 if (V1.getSimpleValueType() != MVT::v4f32)
4576 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4577 if (V2.getSimpleValueType() != MVT::v4f32)
4578 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4580 // Ok, we can emit an INSERTPS instruction.
4581 unsigned ZMask = Zeroable.to_ulong();
4583 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4584 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4585 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
4586 DAG.getIntPtrConstant(InsertPSMask));
4587 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
4590 /// Return a vector logical shift node.
4591 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4592 unsigned NumBits, SelectionDAG &DAG,
4593 const TargetLowering &TLI, SDLoc dl) {
4594 assert(VT.is128BitVector() && "Unknown type for VShift");
4595 MVT ShVT = MVT::v2i64;
4596 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4597 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4598 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4599 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4600 SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
4601 return DAG.getNode(ISD::BITCAST, dl, VT,
4602 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4606 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4608 // Check if the scalar load can be widened into a vector load. And if
4609 // the address is "base + cst" see if the cst can be "absorbed" into
4610 // the shuffle mask.
4611 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4612 SDValue Ptr = LD->getBasePtr();
4613 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4615 EVT PVT = LD->getValueType(0);
4616 if (PVT != MVT::i32 && PVT != MVT::f32)
4621 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4622 FI = FINode->getIndex();
4624 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4625 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4626 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4627 Offset = Ptr.getConstantOperandVal(1);
4628 Ptr = Ptr.getOperand(0);
4633 // FIXME: 256-bit vector instructions don't require a strict alignment,
4634 // improve this code to support it better.
4635 unsigned RequiredAlign = VT.getSizeInBits()/8;
4636 SDValue Chain = LD->getChain();
4637 // Make sure the stack object alignment is at least 16 or 32.
4638 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4639 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4640 if (MFI->isFixedObjectIndex(FI)) {
4641 // Can't change the alignment. FIXME: It's possible to compute
4642 // the exact stack offset and reference FI + adjust offset instead.
4643 // If someone *really* cares about this. That's the way to implement it.
4646 MFI->setObjectAlignment(FI, RequiredAlign);
4650 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4651 // Ptr + (Offset & ~15).
4654 if ((Offset % RequiredAlign) & 3)
4656 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4658 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
4659 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4661 int EltNo = (Offset - StartOffset) >> 2;
4662 unsigned NumElems = VT.getVectorNumElements();
4664 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4665 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4666 LD->getPointerInfo().getWithOffset(StartOffset),
4667 false, false, false, 0);
4669 SmallVector<int, 8> Mask(NumElems, EltNo);
4671 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4677 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4678 /// elements can be replaced by a single large load which has the same value as
4679 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4681 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4683 /// FIXME: we'd also like to handle the case where the last elements are zero
4684 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4685 /// There's even a handy isZeroNode for that purpose.
4686 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4687 SDLoc &DL, SelectionDAG &DAG,
4688 bool isAfterLegalize) {
4689 unsigned NumElems = Elts.size();
4691 LoadSDNode *LDBase = nullptr;
4692 unsigned LastLoadedElt = -1U;
4694 // For each element in the initializer, see if we've found a load or an undef.
4695 // If we don't find an initial load element, or later load elements are
4696 // non-consecutive, bail out.
4697 for (unsigned i = 0; i < NumElems; ++i) {
4698 SDValue Elt = Elts[i];
4699 // Look through a bitcast.
4700 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4701 Elt = Elt.getOperand(0);
4702 if (!Elt.getNode() ||
4703 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4706 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4708 LDBase = cast<LoadSDNode>(Elt.getNode());
4712 if (Elt.getOpcode() == ISD::UNDEF)
4715 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4716 EVT LdVT = Elt.getValueType();
4717 // Each loaded element must be the correct fractional portion of the
4718 // requested vector load.
4719 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4721 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4726 // If we have found an entire vector of loads and undefs, then return a large
4727 // load of the entire vector width starting at the base pointer. If we found
4728 // consecutive loads for the low half, generate a vzext_load node.
4729 if (LastLoadedElt == NumElems - 1) {
4730 assert(LDBase && "Did not find base load for merging consecutive loads");
4731 EVT EltVT = LDBase->getValueType(0);
4732 // Ensure that the input vector size for the merged loads matches the
4733 // cumulative size of the input elements.
4734 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4737 if (isAfterLegalize &&
4738 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4741 SDValue NewLd = SDValue();
4743 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4744 LDBase->getPointerInfo(), LDBase->isVolatile(),
4745 LDBase->isNonTemporal(), LDBase->isInvariant(),
4746 LDBase->getAlignment());
4748 if (LDBase->hasAnyUseOfValue(1)) {
4749 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4751 SDValue(NewLd.getNode(), 1));
4752 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4753 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4754 SDValue(NewLd.getNode(), 1));
4760 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4761 //of a v4i32 / v4f32. It's probably worth generalizing.
4762 EVT EltVT = VT.getVectorElementType();
4763 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4764 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4765 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4766 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4768 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4769 LDBase->getPointerInfo(),
4770 LDBase->getAlignment(),
4771 false/*isVolatile*/, true/*ReadMem*/,
4774 // Make sure the newly-created LOAD is in the same position as LDBase in
4775 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4776 // update uses of LDBase's output chain to use the TokenFactor.
4777 if (LDBase->hasAnyUseOfValue(1)) {
4778 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4779 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4780 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4781 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4782 SDValue(ResNode.getNode(), 1));
4785 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4790 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4791 /// to generate a splat value for the following cases:
4792 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4793 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4794 /// a scalar load, or a constant.
4795 /// The VBROADCAST node is returned when a pattern is found,
4796 /// or SDValue() otherwise.
4797 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4798 SelectionDAG &DAG) {
4799 // VBROADCAST requires AVX.
4800 // TODO: Splats could be generated for non-AVX CPUs using SSE
4801 // instructions, but there's less potential gain for only 128-bit vectors.
4802 if (!Subtarget->hasAVX())
4805 MVT VT = Op.getSimpleValueType();
4808 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4809 "Unsupported vector type for broadcast.");
4814 switch (Op.getOpcode()) {
4816 // Unknown pattern found.
4819 case ISD::BUILD_VECTOR: {
4820 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4821 BitVector UndefElements;
4822 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4824 // We need a splat of a single value to use broadcast, and it doesn't
4825 // make any sense if the value is only in one element of the vector.
4826 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4830 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4831 Ld.getOpcode() == ISD::ConstantFP);
4833 // Make sure that all of the users of a non-constant load are from the
4834 // BUILD_VECTOR node.
4835 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4840 case ISD::VECTOR_SHUFFLE: {
4841 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4843 // Shuffles must have a splat mask where the first element is
4845 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4848 SDValue Sc = Op.getOperand(0);
4849 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4850 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4852 if (!Subtarget->hasInt256())
4855 // Use the register form of the broadcast instruction available on AVX2.
4856 if (VT.getSizeInBits() >= 256)
4857 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4858 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
4861 Ld = Sc.getOperand(0);
4862 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4863 Ld.getOpcode() == ISD::ConstantFP);
4865 // The scalar_to_vector node and the suspected
4866 // load node must have exactly one user.
4867 // Constants may have multiple users.
4869 // AVX-512 has register version of the broadcast
4870 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
4871 Ld.getValueType().getSizeInBits() >= 32;
4872 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
4879 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
4880 bool IsGE256 = (VT.getSizeInBits() >= 256);
4882 // When optimizing for size, generate up to 5 extra bytes for a broadcast
4883 // instruction to save 8 or more bytes of constant pool data.
4884 // TODO: If multiple splats are generated to load the same constant,
4885 // it may be detrimental to overall size. There needs to be a way to detect
4886 // that condition to know if this is truly a size win.
4887 const Function *F = DAG.getMachineFunction().getFunction();
4888 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
4890 // Handle broadcasting a single constant scalar from the constant pool
4892 // On Sandybridge (no AVX2), it is still better to load a constant vector
4893 // from the constant pool and not to broadcast it from a scalar.
4894 // But override that restriction when optimizing for size.
4895 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
4896 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
4897 EVT CVT = Ld.getValueType();
4898 assert(!CVT.isVector() && "Must not broadcast a vector type");
4900 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
4901 // For size optimization, also splat v2f64 and v2i64, and for size opt
4902 // with AVX2, also splat i8 and i16.
4903 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
4904 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4905 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
4906 const Constant *C = nullptr;
4907 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
4908 C = CI->getConstantIntValue();
4909 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
4910 C = CF->getConstantFPValue();
4912 assert(C && "Invalid constant type");
4914 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4915 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
4916 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
4917 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
4918 MachinePointerInfo::getConstantPool(),
4919 false, false, false, Alignment);
4921 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4925 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
4927 // Handle AVX2 in-register broadcasts.
4928 if (!IsLoad && Subtarget->hasInt256() &&
4929 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
4930 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4932 // The scalar source must be a normal load.
4936 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4937 (Subtarget->hasVLX() && ScalarSize == 64))
4938 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4940 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
4941 // double since there is no vbroadcastsd xmm
4942 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
4943 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
4944 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4947 // Unsupported broadcast.
4951 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
4952 /// underlying vector and index.
4954 /// Modifies \p ExtractedFromVec to the real vector and returns the real
4956 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
4958 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
4959 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
4962 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
4964 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
4966 // (extract_vector_elt (vector_shuffle<2,u,u,u>
4967 // (extract_subvector (v8f32 %vreg0), Constant<4>),
4970 // In this case the vector is the extract_subvector expression and the index
4971 // is 2, as specified by the shuffle.
4972 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
4973 SDValue ShuffleVec = SVOp->getOperand(0);
4974 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
4975 assert(ShuffleVecVT.getVectorElementType() ==
4976 ExtractedFromVec.getSimpleValueType().getVectorElementType());
4978 int ShuffleIdx = SVOp->getMaskElt(Idx);
4979 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
4980 ExtractedFromVec = ShuffleVec;
4986 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
4987 MVT VT = Op.getSimpleValueType();
4989 // Skip if insert_vec_elt is not supported.
4990 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4991 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
4995 unsigned NumElems = Op.getNumOperands();
4999 SmallVector<unsigned, 4> InsertIndices;
5000 SmallVector<int, 8> Mask(NumElems, -1);
5002 for (unsigned i = 0; i != NumElems; ++i) {
5003 unsigned Opc = Op.getOperand(i).getOpcode();
5005 if (Opc == ISD::UNDEF)
5008 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5009 // Quit if more than 1 elements need inserting.
5010 if (InsertIndices.size() > 1)
5013 InsertIndices.push_back(i);
5017 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5018 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5019 // Quit if non-constant index.
5020 if (!isa<ConstantSDNode>(ExtIdx))
5022 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5024 // Quit if extracted from vector of different type.
5025 if (ExtractedFromVec.getValueType() != VT)
5028 if (!VecIn1.getNode())
5029 VecIn1 = ExtractedFromVec;
5030 else if (VecIn1 != ExtractedFromVec) {
5031 if (!VecIn2.getNode())
5032 VecIn2 = ExtractedFromVec;
5033 else if (VecIn2 != ExtractedFromVec)
5034 // Quit if more than 2 vectors to shuffle
5038 if (ExtractedFromVec == VecIn1)
5040 else if (ExtractedFromVec == VecIn2)
5041 Mask[i] = Idx + NumElems;
5044 if (!VecIn1.getNode())
5047 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5048 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5049 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5050 unsigned Idx = InsertIndices[i];
5051 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5052 DAG.getIntPtrConstant(Idx));
5058 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5060 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5062 MVT VT = Op.getSimpleValueType();
5063 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5064 "Unexpected type in LowerBUILD_VECTORvXi1!");
5067 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5068 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5069 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5070 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5073 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5074 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5075 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5076 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5079 bool AllContants = true;
5080 uint64_t Immediate = 0;
5081 int NonConstIdx = -1;
5082 bool IsSplat = true;
5083 unsigned NumNonConsts = 0;
5084 unsigned NumConsts = 0;
5085 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5086 SDValue In = Op.getOperand(idx);
5087 if (In.getOpcode() == ISD::UNDEF)
5089 if (!isa<ConstantSDNode>(In)) {
5090 AllContants = false;
5095 if (cast<ConstantSDNode>(In)->getZExtValue())
5096 Immediate |= (1ULL << idx);
5098 if (In != Op.getOperand(0))
5103 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5104 DAG.getConstant(Immediate, MVT::i16));
5105 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5106 DAG.getIntPtrConstant(0));
5109 if (NumNonConsts == 1 && NonConstIdx != 0) {
5112 SDValue VecAsImm = DAG.getConstant(Immediate,
5113 MVT::getIntegerVT(VT.getSizeInBits()));
5114 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
5117 DstVec = DAG.getUNDEF(VT);
5118 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5119 Op.getOperand(NonConstIdx),
5120 DAG.getIntPtrConstant(NonConstIdx));
5122 if (!IsSplat && (NonConstIdx != 0))
5123 llvm_unreachable("Unsupported BUILD_VECTOR operation");
5124 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
5127 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5128 DAG.getConstant(-1, SelectVT),
5129 DAG.getConstant(0, SelectVT));
5131 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5132 DAG.getConstant((Immediate | 1), SelectVT),
5133 DAG.getConstant(Immediate, SelectVT));
5134 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
5137 /// \brief Return true if \p N implements a horizontal binop and return the
5138 /// operands for the horizontal binop into V0 and V1.
5140 /// This is a helper function of PerformBUILD_VECTORCombine.
5141 /// This function checks that the build_vector \p N in input implements a
5142 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5143 /// operation to match.
5144 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5145 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5146 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5149 /// This function only analyzes elements of \p N whose indices are
5150 /// in range [BaseIdx, LastIdx).
5151 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5153 unsigned BaseIdx, unsigned LastIdx,
5154 SDValue &V0, SDValue &V1) {
5155 EVT VT = N->getValueType(0);
5157 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5158 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5159 "Invalid Vector in input!");
5161 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5162 bool CanFold = true;
5163 unsigned ExpectedVExtractIdx = BaseIdx;
5164 unsigned NumElts = LastIdx - BaseIdx;
5165 V0 = DAG.getUNDEF(VT);
5166 V1 = DAG.getUNDEF(VT);
5168 // Check if N implements a horizontal binop.
5169 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5170 SDValue Op = N->getOperand(i + BaseIdx);
5173 if (Op->getOpcode() == ISD::UNDEF) {
5174 // Update the expected vector extract index.
5175 if (i * 2 == NumElts)
5176 ExpectedVExtractIdx = BaseIdx;
5177 ExpectedVExtractIdx += 2;
5181 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5186 SDValue Op0 = Op.getOperand(0);
5187 SDValue Op1 = Op.getOperand(1);
5189 // Try to match the following pattern:
5190 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5191 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5192 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5193 Op0.getOperand(0) == Op1.getOperand(0) &&
5194 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5195 isa<ConstantSDNode>(Op1.getOperand(1)));
5199 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5200 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5202 if (i * 2 < NumElts) {
5203 if (V0.getOpcode() == ISD::UNDEF)
5204 V0 = Op0.getOperand(0);
5206 if (V1.getOpcode() == ISD::UNDEF)
5207 V1 = Op0.getOperand(0);
5208 if (i * 2 == NumElts)
5209 ExpectedVExtractIdx = BaseIdx;
5212 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5213 if (I0 == ExpectedVExtractIdx)
5214 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5215 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5216 // Try to match the following dag sequence:
5217 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5218 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5222 ExpectedVExtractIdx += 2;
5228 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5229 /// a concat_vector.
5231 /// This is a helper function of PerformBUILD_VECTORCombine.
5232 /// This function expects two 256-bit vectors called V0 and V1.
5233 /// At first, each vector is split into two separate 128-bit vectors.
5234 /// Then, the resulting 128-bit vectors are used to implement two
5235 /// horizontal binary operations.
5237 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5239 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5240 /// the two new horizontal binop.
5241 /// When Mode is set, the first horizontal binop dag node would take as input
5242 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5243 /// horizontal binop dag node would take as input the lower 128-bit of V1
5244 /// and the upper 128-bit of V1.
5246 /// HADD V0_LO, V0_HI
5247 /// HADD V1_LO, V1_HI
5249 /// Otherwise, the first horizontal binop dag node takes as input the lower
5250 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5251 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5253 /// HADD V0_LO, V1_LO
5254 /// HADD V0_HI, V1_HI
5256 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5257 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5258 /// the upper 128-bits of the result.
5259 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5260 SDLoc DL, SelectionDAG &DAG,
5261 unsigned X86Opcode, bool Mode,
5262 bool isUndefLO, bool isUndefHI) {
5263 EVT VT = V0.getValueType();
5264 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5265 "Invalid nodes in input!");
5267 unsigned NumElts = VT.getVectorNumElements();
5268 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5269 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5270 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5271 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5272 EVT NewVT = V0_LO.getValueType();
5274 SDValue LO = DAG.getUNDEF(NewVT);
5275 SDValue HI = DAG.getUNDEF(NewVT);
5278 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5279 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5280 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5281 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5282 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5284 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5285 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5286 V1_LO->getOpcode() != ISD::UNDEF))
5287 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5289 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5290 V1_HI->getOpcode() != ISD::UNDEF))
5291 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5294 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5297 /// \brief Try to fold a build_vector that performs an 'addsub' into the
5298 /// sequence of 'vadd + vsub + blendi'.
5299 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
5300 const X86Subtarget *Subtarget) {
5302 EVT VT = BV->getValueType(0);
5303 unsigned NumElts = VT.getVectorNumElements();
5304 SDValue InVec0 = DAG.getUNDEF(VT);
5305 SDValue InVec1 = DAG.getUNDEF(VT);
5307 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5308 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5310 // Odd-numbered elements in the input build vector are obtained from
5311 // adding two integer/float elements.
5312 // Even-numbered elements in the input build vector are obtained from
5313 // subtracting two integer/float elements.
5314 unsigned ExpectedOpcode = ISD::FSUB;
5315 unsigned NextExpectedOpcode = ISD::FADD;
5316 bool AddFound = false;
5317 bool SubFound = false;
5319 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5320 SDValue Op = BV->getOperand(i);
5322 // Skip 'undef' values.
5323 unsigned Opcode = Op.getOpcode();
5324 if (Opcode == ISD::UNDEF) {
5325 std::swap(ExpectedOpcode, NextExpectedOpcode);
5329 // Early exit if we found an unexpected opcode.
5330 if (Opcode != ExpectedOpcode)
5333 SDValue Op0 = Op.getOperand(0);
5334 SDValue Op1 = Op.getOperand(1);
5336 // Try to match the following pattern:
5337 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5338 // Early exit if we cannot match that sequence.
5339 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5340 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5341 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5342 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5343 Op0.getOperand(1) != Op1.getOperand(1))
5346 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5350 // We found a valid add/sub node. Update the information accordingly.
5356 // Update InVec0 and InVec1.
5357 if (InVec0.getOpcode() == ISD::UNDEF)
5358 InVec0 = Op0.getOperand(0);
5359 if (InVec1.getOpcode() == ISD::UNDEF)
5360 InVec1 = Op1.getOperand(0);
5362 // Make sure that operands in input to each add/sub node always
5363 // come from a same pair of vectors.
5364 if (InVec0 != Op0.getOperand(0)) {
5365 if (ExpectedOpcode == ISD::FSUB)
5368 // FADD is commutable. Try to commute the operands
5369 // and then test again.
5370 std::swap(Op0, Op1);
5371 if (InVec0 != Op0.getOperand(0))
5375 if (InVec1 != Op1.getOperand(0))
5378 // Update the pair of expected opcodes.
5379 std::swap(ExpectedOpcode, NextExpectedOpcode);
5382 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5383 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5384 InVec1.getOpcode() != ISD::UNDEF)
5385 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5390 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
5391 const X86Subtarget *Subtarget) {
5393 EVT VT = N->getValueType(0);
5394 unsigned NumElts = VT.getVectorNumElements();
5395 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
5396 SDValue InVec0, InVec1;
5398 // Try to match an ADDSUB.
5399 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
5400 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
5401 SDValue Value = matchAddSub(BV, DAG, Subtarget);
5402 if (Value.getNode())
5406 // Try to match horizontal ADD/SUB.
5407 unsigned NumUndefsLO = 0;
5408 unsigned NumUndefsHI = 0;
5409 unsigned Half = NumElts/2;
5411 // Count the number of UNDEF operands in the build_vector in input.
5412 for (unsigned i = 0, e = Half; i != e; ++i)
5413 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5416 for (unsigned i = Half, e = NumElts; i != e; ++i)
5417 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5420 // Early exit if this is either a build_vector of all UNDEFs or all the
5421 // operands but one are UNDEF.
5422 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5425 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5426 // Try to match an SSE3 float HADD/HSUB.
5427 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5428 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5430 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5431 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5432 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5433 // Try to match an SSSE3 integer HADD/HSUB.
5434 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5435 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5437 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5438 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5441 if (!Subtarget->hasAVX())
5444 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5445 // Try to match an AVX horizontal add/sub of packed single/double
5446 // precision floating point values from 256-bit vectors.
5447 SDValue InVec2, InVec3;
5448 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5449 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5450 ((InVec0.getOpcode() == ISD::UNDEF ||
5451 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5452 ((InVec1.getOpcode() == ISD::UNDEF ||
5453 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5454 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5456 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5457 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5458 ((InVec0.getOpcode() == ISD::UNDEF ||
5459 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5460 ((InVec1.getOpcode() == ISD::UNDEF ||
5461 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5462 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5463 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5464 // Try to match an AVX2 horizontal add/sub of signed integers.
5465 SDValue InVec2, InVec3;
5467 bool CanFold = true;
5469 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5470 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5471 ((InVec0.getOpcode() == ISD::UNDEF ||
5472 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5473 ((InVec1.getOpcode() == ISD::UNDEF ||
5474 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5475 X86Opcode = X86ISD::HADD;
5476 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5477 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5478 ((InVec0.getOpcode() == ISD::UNDEF ||
5479 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5480 ((InVec1.getOpcode() == ISD::UNDEF ||
5481 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5482 X86Opcode = X86ISD::HSUB;
5487 // Fold this build_vector into a single horizontal add/sub.
5488 // Do this only if the target has AVX2.
5489 if (Subtarget->hasAVX2())
5490 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5492 // Do not try to expand this build_vector into a pair of horizontal
5493 // add/sub if we can emit a pair of scalar add/sub.
5494 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5497 // Convert this build_vector into a pair of horizontal binop followed by
5499 bool isUndefLO = NumUndefsLO == Half;
5500 bool isUndefHI = NumUndefsHI == Half;
5501 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5502 isUndefLO, isUndefHI);
5506 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5507 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5509 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5510 X86Opcode = X86ISD::HADD;
5511 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5512 X86Opcode = X86ISD::HSUB;
5513 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5514 X86Opcode = X86ISD::FHADD;
5515 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5516 X86Opcode = X86ISD::FHSUB;
5520 // Don't try to expand this build_vector into a pair of horizontal add/sub
5521 // if we can simply emit a pair of scalar add/sub.
5522 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5525 // Convert this build_vector into two horizontal add/sub followed by
5527 bool isUndefLO = NumUndefsLO == Half;
5528 bool isUndefHI = NumUndefsHI == Half;
5529 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5530 isUndefLO, isUndefHI);
5537 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5540 MVT VT = Op.getSimpleValueType();
5541 MVT ExtVT = VT.getVectorElementType();
5542 unsigned NumElems = Op.getNumOperands();
5544 // Generate vectors for predicate vectors.
5545 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5546 return LowerBUILD_VECTORvXi1(Op, DAG);
5548 // Vectors containing all zeros can be matched by pxor and xorps later
5549 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5550 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5551 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5552 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5555 return getZeroVector(VT, Subtarget, DAG, dl);
5558 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5559 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5560 // vpcmpeqd on 256-bit vectors.
5561 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5562 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5565 if (!VT.is512BitVector())
5566 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5569 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
5570 if (Broadcast.getNode())
5573 unsigned EVTBits = ExtVT.getSizeInBits();
5575 unsigned NumZero = 0;
5576 unsigned NumNonZero = 0;
5577 unsigned NonZeros = 0;
5578 bool IsAllConstants = true;
5579 SmallSet<SDValue, 8> Values;
5580 for (unsigned i = 0; i < NumElems; ++i) {
5581 SDValue Elt = Op.getOperand(i);
5582 if (Elt.getOpcode() == ISD::UNDEF)
5585 if (Elt.getOpcode() != ISD::Constant &&
5586 Elt.getOpcode() != ISD::ConstantFP)
5587 IsAllConstants = false;
5588 if (X86::isZeroNode(Elt))
5591 NonZeros |= (1 << i);
5596 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5597 if (NumNonZero == 0)
5598 return DAG.getUNDEF(VT);
5600 // Special case for single non-zero, non-undef, element.
5601 if (NumNonZero == 1) {
5602 unsigned Idx = countTrailingZeros(NonZeros);
5603 SDValue Item = Op.getOperand(Idx);
5605 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5606 // the value are obviously zero, truncate the value to i32 and do the
5607 // insertion that way. Only do this if the value is non-constant or if the
5608 // value is a constant being inserted into element 0. It is cheaper to do
5609 // a constant pool load than it is to do a movd + shuffle.
5610 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5611 (!IsAllConstants || Idx == 0)) {
5612 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5614 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5615 EVT VecVT = MVT::v4i32;
5617 // Truncate the value (which may itself be a constant) to i32, and
5618 // convert it to a vector with movd (S2V+shuffle to zero extend).
5619 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5620 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5622 ISD::BITCAST, dl, VT,
5623 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5627 // If we have a constant or non-constant insertion into the low element of
5628 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5629 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5630 // depending on what the source datatype is.
5633 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5635 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5636 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5637 if (VT.is256BitVector() || VT.is512BitVector()) {
5638 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5639 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5640 Item, DAG.getIntPtrConstant(0));
5642 assert(VT.is128BitVector() && "Expected an SSE value type!");
5643 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5644 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5645 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5648 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5649 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5650 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5651 if (VT.is256BitVector()) {
5652 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5653 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5655 assert(VT.is128BitVector() && "Expected an SSE value type!");
5656 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5658 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5662 // Is it a vector logical left shift?
5663 if (NumElems == 2 && Idx == 1 &&
5664 X86::isZeroNode(Op.getOperand(0)) &&
5665 !X86::isZeroNode(Op.getOperand(1))) {
5666 unsigned NumBits = VT.getSizeInBits();
5667 return getVShift(true, VT,
5668 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5669 VT, Op.getOperand(1)),
5670 NumBits/2, DAG, *this, dl);
5673 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5676 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5677 // is a non-constant being inserted into an element other than the low one,
5678 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5679 // movd/movss) to move this into the low element, then shuffle it into
5681 if (EVTBits == 32) {
5682 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5683 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5687 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5688 if (Values.size() == 1) {
5689 if (EVTBits == 32) {
5690 // Instead of a shuffle like this:
5691 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5692 // Check if it's possible to issue this instead.
5693 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5694 unsigned Idx = countTrailingZeros(NonZeros);
5695 SDValue Item = Op.getOperand(Idx);
5696 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5697 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5702 // A vector full of immediates; various special cases are already
5703 // handled, so this is best done with a single constant-pool load.
5707 // For AVX-length vectors, see if we can use a vector load to get all of the
5708 // elements, otherwise build the individual 128-bit pieces and use
5709 // shuffles to put them in place.
5710 if (VT.is256BitVector() || VT.is512BitVector()) {
5711 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5713 // Check for a build vector of consecutive loads.
5714 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5717 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5719 // Build both the lower and upper subvector.
5720 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5721 makeArrayRef(&V[0], NumElems/2));
5722 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5723 makeArrayRef(&V[NumElems / 2], NumElems/2));
5725 // Recreate the wider vector with the lower and upper part.
5726 if (VT.is256BitVector())
5727 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5728 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5731 // Let legalizer expand 2-wide build_vectors.
5732 if (EVTBits == 64) {
5733 if (NumNonZero == 1) {
5734 // One half is zero or undef.
5735 unsigned Idx = countTrailingZeros(NonZeros);
5736 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5737 Op.getOperand(Idx));
5738 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5743 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5744 if (EVTBits == 8 && NumElems == 16) {
5745 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5747 if (V.getNode()) return V;
5750 if (EVTBits == 16 && NumElems == 8) {
5751 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5753 if (V.getNode()) return V;
5756 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5757 if (EVTBits == 32 && NumElems == 4) {
5758 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
5763 // If element VT is == 32 bits, turn it into a number of shuffles.
5764 SmallVector<SDValue, 8> V(NumElems);
5765 if (NumElems == 4 && NumZero > 0) {
5766 for (unsigned i = 0; i < 4; ++i) {
5767 bool isZero = !(NonZeros & (1 << i));
5769 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5771 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5774 for (unsigned i = 0; i < 2; ++i) {
5775 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5778 V[i] = V[i*2]; // Must be a zero vector.
5781 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5784 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5787 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5792 bool Reverse1 = (NonZeros & 0x3) == 2;
5793 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5797 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5798 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5800 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5803 if (Values.size() > 1 && VT.is128BitVector()) {
5804 // Check for a build vector of consecutive loads.
5805 for (unsigned i = 0; i < NumElems; ++i)
5806 V[i] = Op.getOperand(i);
5808 // Check for elements which are consecutive loads.
5809 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
5813 // Check for a build vector from mostly shuffle plus few inserting.
5814 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5818 // For SSE 4.1, use insertps to put the high elements into the low element.
5819 if (Subtarget->hasSSE41()) {
5821 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5822 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5824 Result = DAG.getUNDEF(VT);
5826 for (unsigned i = 1; i < NumElems; ++i) {
5827 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5828 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5829 Op.getOperand(i), DAG.getIntPtrConstant(i));
5834 // Otherwise, expand into a number of unpckl*, start by extending each of
5835 // our (non-undef) elements to the full vector width with the element in the
5836 // bottom slot of the vector (which generates no code for SSE).
5837 for (unsigned i = 0; i < NumElems; ++i) {
5838 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5839 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5841 V[i] = DAG.getUNDEF(VT);
5844 // Next, we iteratively mix elements, e.g. for v4f32:
5845 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5846 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5847 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5848 unsigned EltStride = NumElems >> 1;
5849 while (EltStride != 0) {
5850 for (unsigned i = 0; i < EltStride; ++i) {
5851 // If V[i+EltStride] is undef and this is the first round of mixing,
5852 // then it is safe to just drop this shuffle: V[i] is already in the
5853 // right place, the one element (since it's the first round) being
5854 // inserted as undef can be dropped. This isn't safe for successive
5855 // rounds because they will permute elements within both vectors.
5856 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5857 EltStride == NumElems/2)
5860 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5869 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5870 // to create 256-bit vectors from two other 128-bit ones.
5871 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5873 MVT ResVT = Op.getSimpleValueType();
5875 assert((ResVT.is256BitVector() ||
5876 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
5878 SDValue V1 = Op.getOperand(0);
5879 SDValue V2 = Op.getOperand(1);
5880 unsigned NumElems = ResVT.getVectorNumElements();
5881 if(ResVT.is256BitVector())
5882 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5884 if (Op.getNumOperands() == 4) {
5885 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
5886 ResVT.getVectorNumElements()/2);
5887 SDValue V3 = Op.getOperand(2);
5888 SDValue V4 = Op.getOperand(3);
5889 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
5890 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
5892 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5895 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5896 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
5897 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
5898 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
5899 Op.getNumOperands() == 4)));
5901 // AVX can use the vinsertf128 instruction to create 256-bit vectors
5902 // from two other 128-bit ones.
5904 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
5905 return LowerAVXCONCAT_VECTORS(Op, DAG);
5909 //===----------------------------------------------------------------------===//
5910 // Vector shuffle lowering
5912 // This is an experimental code path for lowering vector shuffles on x86. It is
5913 // designed to handle arbitrary vector shuffles and blends, gracefully
5914 // degrading performance as necessary. It works hard to recognize idiomatic
5915 // shuffles and lower them to optimal instruction patterns without leaving
5916 // a framework that allows reasonably efficient handling of all vector shuffle
5918 //===----------------------------------------------------------------------===//
5920 /// \brief Tiny helper function to identify a no-op mask.
5922 /// This is a somewhat boring predicate function. It checks whether the mask
5923 /// array input, which is assumed to be a single-input shuffle mask of the kind
5924 /// used by the X86 shuffle instructions (not a fully general
5925 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
5926 /// in-place shuffle are 'no-op's.
5927 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
5928 for (int i = 0, Size = Mask.size(); i < Size; ++i)
5929 if (Mask[i] != -1 && Mask[i] != i)
5934 /// \brief Helper function to classify a mask as a single-input mask.
5936 /// This isn't a generic single-input test because in the vector shuffle
5937 /// lowering we canonicalize single inputs to be the first input operand. This
5938 /// means we can more quickly test for a single input by only checking whether
5939 /// an input from the second operand exists. We also assume that the size of
5940 /// mask corresponds to the size of the input vectors which isn't true in the
5941 /// fully general case.
5942 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
5944 if (M >= (int)Mask.size())
5949 /// \brief Test whether there are elements crossing 128-bit lanes in this
5952 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
5953 /// and we routinely test for these.
5954 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
5955 int LaneSize = 128 / VT.getScalarSizeInBits();
5956 int Size = Mask.size();
5957 for (int i = 0; i < Size; ++i)
5958 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
5963 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
5965 /// This checks a shuffle mask to see if it is performing the same
5966 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
5967 /// that it is also not lane-crossing. It may however involve a blend from the
5968 /// same lane of a second vector.
5970 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
5971 /// non-trivial to compute in the face of undef lanes. The representation is
5972 /// *not* suitable for use with existing 128-bit shuffles as it will contain
5973 /// entries from both V1 and V2 inputs to the wider mask.
5975 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
5976 SmallVectorImpl<int> &RepeatedMask) {
5977 int LaneSize = 128 / VT.getScalarSizeInBits();
5978 RepeatedMask.resize(LaneSize, -1);
5979 int Size = Mask.size();
5980 for (int i = 0; i < Size; ++i) {
5983 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
5984 // This entry crosses lanes, so there is no way to model this shuffle.
5987 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
5988 if (RepeatedMask[i % LaneSize] == -1)
5989 // This is the first non-undef entry in this slot of a 128-bit lane.
5990 RepeatedMask[i % LaneSize] =
5991 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
5992 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
5993 // Found a mismatch with the repeated mask.
5999 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6002 /// This is a fast way to test a shuffle mask against a fixed pattern:
6004 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6006 /// It returns true if the mask is exactly as wide as the argument list, and
6007 /// each element of the mask is either -1 (signifying undef) or the value given
6008 /// in the argument.
6009 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6010 ArrayRef<int> ExpectedMask) {
6011 if (Mask.size() != ExpectedMask.size())
6014 int Size = Mask.size();
6016 // If the values are build vectors, we can look through them to find
6017 // equivalent inputs that make the shuffles equivalent.
6018 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6019 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6021 for (int i = 0; i < Size; ++i)
6022 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6023 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6024 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6025 if (!MaskBV || !ExpectedBV ||
6026 MaskBV->getOperand(Mask[i] % Size) !=
6027 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6034 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6036 /// This helper function produces an 8-bit shuffle immediate corresponding to
6037 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6038 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6041 /// NB: We rely heavily on "undef" masks preserving the input lane.
6042 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
6043 SelectionDAG &DAG) {
6044 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6045 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6046 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6047 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6048 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6051 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6052 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6053 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6054 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6055 return DAG.getConstant(Imm, MVT::i8);
6058 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6060 /// This is used as a fallback approach when first class blend instructions are
6061 /// unavailable. Currently it is only suitable for integer vectors, but could
6062 /// be generalized for floating point vectors if desirable.
6063 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6064 SDValue V2, ArrayRef<int> Mask,
6065 SelectionDAG &DAG) {
6066 assert(VT.isInteger() && "Only supports integer vector types!");
6067 MVT EltVT = VT.getScalarType();
6068 int NumEltBits = EltVT.getSizeInBits();
6069 SDValue Zero = DAG.getConstant(0, EltVT);
6070 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
6071 SmallVector<SDValue, 16> MaskOps;
6072 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6073 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6074 return SDValue(); // Shuffled input!
6075 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6078 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6079 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6080 // We have to cast V2 around.
6081 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6082 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6083 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6084 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6085 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6086 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6089 /// \brief Try to emit a blend instruction for a shuffle.
6091 /// This doesn't do any checks for the availability of instructions for blending
6092 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6093 /// be matched in the backend with the type given. What it does check for is
6094 /// that the shuffle mask is in fact a blend.
6095 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6096 SDValue V2, ArrayRef<int> Mask,
6097 const X86Subtarget *Subtarget,
6098 SelectionDAG &DAG) {
6099 unsigned BlendMask = 0;
6100 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6101 if (Mask[i] >= Size) {
6102 if (Mask[i] != i + Size)
6103 return SDValue(); // Shuffled V2 input!
6104 BlendMask |= 1u << i;
6107 if (Mask[i] >= 0 && Mask[i] != i)
6108 return SDValue(); // Shuffled V1 input!
6110 switch (VT.SimpleTy) {
6115 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6116 DAG.getConstant(BlendMask, MVT::i8));
6120 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6124 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6125 // that instruction.
6126 if (Subtarget->hasAVX2()) {
6127 // Scale the blend by the number of 32-bit dwords per element.
6128 int Scale = VT.getScalarSizeInBits() / 32;
6130 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6131 if (Mask[i] >= Size)
6132 for (int j = 0; j < Scale; ++j)
6133 BlendMask |= 1u << (i * Scale + j);
6135 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6136 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6137 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6138 return DAG.getNode(ISD::BITCAST, DL, VT,
6139 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6140 DAG.getConstant(BlendMask, MVT::i8)));
6144 // For integer shuffles we need to expand the mask and cast the inputs to
6145 // v8i16s prior to blending.
6146 int Scale = 8 / VT.getVectorNumElements();
6148 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6149 if (Mask[i] >= Size)
6150 for (int j = 0; j < Scale; ++j)
6151 BlendMask |= 1u << (i * Scale + j);
6153 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6154 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6155 return DAG.getNode(ISD::BITCAST, DL, VT,
6156 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6157 DAG.getConstant(BlendMask, MVT::i8)));
6161 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6162 SmallVector<int, 8> RepeatedMask;
6163 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6164 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6165 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6167 for (int i = 0; i < 8; ++i)
6168 if (RepeatedMask[i] >= 16)
6169 BlendMask |= 1u << i;
6170 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6171 DAG.getConstant(BlendMask, MVT::i8));
6177 // Scale the blend by the number of bytes per element.
6178 int Scale = VT.getScalarSizeInBits() / 8;
6180 // This form of blend is always done on bytes. Compute the byte vector
6182 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6184 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6185 // mix of LLVM's code generator and the x86 backend. We tell the code
6186 // generator that boolean values in the elements of an x86 vector register
6187 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6188 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6189 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6190 // of the element (the remaining are ignored) and 0 in that high bit would
6191 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6192 // the LLVM model for boolean values in vector elements gets the relevant
6193 // bit set, it is set backwards and over constrained relative to x86's
6195 SmallVector<SDValue, 32> VSELECTMask;
6196 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6197 for (int j = 0; j < Scale; ++j)
6198 VSELECTMask.push_back(
6199 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6200 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
6202 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6203 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6205 ISD::BITCAST, DL, VT,
6206 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6207 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6212 llvm_unreachable("Not a supported integer vector type!");
6216 /// \brief Try to lower as a blend of elements from two inputs followed by
6217 /// a single-input permutation.
6219 /// This matches the pattern where we can blend elements from two inputs and
6220 /// then reduce the shuffle to a single-input permutation.
6221 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6224 SelectionDAG &DAG) {
6225 // We build up the blend mask while checking whether a blend is a viable way
6226 // to reduce the shuffle.
6227 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6228 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6230 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6234 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6236 if (BlendMask[Mask[i] % Size] == -1)
6237 BlendMask[Mask[i] % Size] = Mask[i];
6238 else if (BlendMask[Mask[i] % Size] != Mask[i])
6239 return SDValue(); // Can't blend in the needed input!
6241 PermuteMask[i] = Mask[i] % Size;
6244 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6245 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6248 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6249 /// blends and permutes.
6251 /// This matches the extremely common pattern for handling combined
6252 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6253 /// operations. It will try to pick the best arrangement of shuffles and
6255 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6259 SelectionDAG &DAG) {
6260 // Shuffle the input elements into the desired positions in V1 and V2 and
6261 // blend them together.
6262 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6263 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6264 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6265 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6266 if (Mask[i] >= 0 && Mask[i] < Size) {
6267 V1Mask[i] = Mask[i];
6269 } else if (Mask[i] >= Size) {
6270 V2Mask[i] = Mask[i] - Size;
6271 BlendMask[i] = i + Size;
6274 // Try to lower with the simpler initial blend strategy unless one of the
6275 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6276 // shuffle may be able to fold with a load or other benefit. However, when
6277 // we'll have to do 2x as many shuffles in order to achieve this, blending
6278 // first is a better strategy.
6279 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6280 if (SDValue BlendPerm =
6281 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6284 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6285 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6286 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6289 /// \brief Try to lower a vector shuffle as a byte rotation.
6291 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6292 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6293 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6294 /// try to generically lower a vector shuffle through such an pattern. It
6295 /// does not check for the profitability of lowering either as PALIGNR or
6296 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6297 /// This matches shuffle vectors that look like:
6299 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6301 /// Essentially it concatenates V1 and V2, shifts right by some number of
6302 /// elements, and takes the low elements as the result. Note that while this is
6303 /// specified as a *right shift* because x86 is little-endian, it is a *left
6304 /// rotate* of the vector lanes.
6305 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6308 const X86Subtarget *Subtarget,
6309 SelectionDAG &DAG) {
6310 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6312 int NumElts = Mask.size();
6313 int NumLanes = VT.getSizeInBits() / 128;
6314 int NumLaneElts = NumElts / NumLanes;
6316 // We need to detect various ways of spelling a rotation:
6317 // [11, 12, 13, 14, 15, 0, 1, 2]
6318 // [-1, 12, 13, 14, -1, -1, 1, -1]
6319 // [-1, -1, -1, -1, -1, -1, 1, 2]
6320 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6321 // [-1, 4, 5, 6, -1, -1, 9, -1]
6322 // [-1, 4, 5, 6, -1, -1, -1, -1]
6325 for (int l = 0; l < NumElts; l += NumLaneElts) {
6326 for (int i = 0; i < NumLaneElts; ++i) {
6327 if (Mask[l + i] == -1)
6329 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6331 // Get the mod-Size index and lane correct it.
6332 int LaneIdx = (Mask[l + i] % NumElts) - l;
6333 // Make sure it was in this lane.
6334 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6337 // Determine where a rotated vector would have started.
6338 int StartIdx = i - LaneIdx;
6340 // The identity rotation isn't interesting, stop.
6343 // If we found the tail of a vector the rotation must be the missing
6344 // front. If we found the head of a vector, it must be how much of the
6346 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6349 Rotation = CandidateRotation;
6350 else if (Rotation != CandidateRotation)
6351 // The rotations don't match, so we can't match this mask.
6354 // Compute which value this mask is pointing at.
6355 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6357 // Compute which of the two target values this index should be assigned
6358 // to. This reflects whether the high elements are remaining or the low
6359 // elements are remaining.
6360 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6362 // Either set up this value if we've not encountered it before, or check
6363 // that it remains consistent.
6366 else if (TargetV != MaskV)
6367 // This may be a rotation, but it pulls from the inputs in some
6368 // unsupported interleaving.
6373 // Check that we successfully analyzed the mask, and normalize the results.
6374 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6375 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6381 // The actual rotate instruction rotates bytes, so we need to scale the
6382 // rotation based on how many bytes are in the vector lane.
6383 int Scale = 16 / NumLaneElts;
6385 // SSSE3 targets can use the palignr instruction.
6386 if (Subtarget->hasSSSE3()) {
6387 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6388 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6389 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6390 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6392 return DAG.getNode(ISD::BITCAST, DL, VT,
6393 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6394 DAG.getConstant(Rotation * Scale, MVT::i8)));
6397 assert(VT.getSizeInBits() == 128 &&
6398 "Rotate-based lowering only supports 128-bit lowering!");
6399 assert(Mask.size() <= 16 &&
6400 "Can shuffle at most 16 bytes in a 128-bit vector!");
6402 // Default SSE2 implementation
6403 int LoByteShift = 16 - Rotation * Scale;
6404 int HiByteShift = Rotation * Scale;
6406 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6407 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6408 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6410 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6411 DAG.getConstant(LoByteShift, MVT::i8));
6412 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6413 DAG.getConstant(HiByteShift, MVT::i8));
6414 return DAG.getNode(ISD::BITCAST, DL, VT,
6415 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6418 /// \brief Compute whether each element of a shuffle is zeroable.
6420 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6421 /// Either it is an undef element in the shuffle mask, the element of the input
6422 /// referenced is undef, or the element of the input referenced is known to be
6423 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6424 /// as many lanes with this technique as possible to simplify the remaining
6426 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6427 SDValue V1, SDValue V2) {
6428 SmallBitVector Zeroable(Mask.size(), false);
6430 while (V1.getOpcode() == ISD::BITCAST)
6431 V1 = V1->getOperand(0);
6432 while (V2.getOpcode() == ISD::BITCAST)
6433 V2 = V2->getOperand(0);
6435 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6436 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6438 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6440 // Handle the easy cases.
6441 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6446 // If this is an index into a build_vector node (which has the same number
6447 // of elements), dig out the input value and use it.
6448 SDValue V = M < Size ? V1 : V2;
6449 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6452 SDValue Input = V.getOperand(M % Size);
6453 // The UNDEF opcode check really should be dead code here, but not quite
6454 // worth asserting on (it isn't invalid, just unexpected).
6455 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6462 /// \brief Try to emit a bitmask instruction for a shuffle.
6464 /// This handles cases where we can model a blend exactly as a bitmask due to
6465 /// one of the inputs being zeroable.
6466 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6467 SDValue V2, ArrayRef<int> Mask,
6468 SelectionDAG &DAG) {
6469 MVT EltVT = VT.getScalarType();
6470 int NumEltBits = EltVT.getSizeInBits();
6471 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6472 SDValue Zero = DAG.getConstant(0, IntEltVT);
6473 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
6474 if (EltVT.isFloatingPoint()) {
6475 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6476 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6478 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6479 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6481 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6484 if (Mask[i] % Size != i)
6485 return SDValue(); // Not a blend.
6487 V = Mask[i] < Size ? V1 : V2;
6488 else if (V != (Mask[i] < Size ? V1 : V2))
6489 return SDValue(); // Can only let one input through the mask.
6491 VMaskOps[i] = AllOnes;
6494 return SDValue(); // No non-zeroable elements!
6496 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6497 V = DAG.getNode(VT.isFloatingPoint()
6498 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6503 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6505 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6506 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6507 /// matches elements from one of the input vectors shuffled to the left or
6508 /// right with zeroable elements 'shifted in'. It handles both the strictly
6509 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6512 /// PSHL : (little-endian) left bit shift.
6513 /// [ zz, 0, zz, 2 ]
6514 /// [ -1, 4, zz, -1 ]
6515 /// PSRL : (little-endian) right bit shift.
6517 /// [ -1, -1, 7, zz]
6518 /// PSLLDQ : (little-endian) left byte shift
6519 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6520 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6521 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6522 /// PSRLDQ : (little-endian) right byte shift
6523 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6524 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6525 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6526 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6527 SDValue V2, ArrayRef<int> Mask,
6528 SelectionDAG &DAG) {
6529 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6531 int Size = Mask.size();
6532 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6534 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6535 for (int i = 0; i < Size; i += Scale)
6536 for (int j = 0; j < Shift; ++j)
6537 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6543 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6544 for (int i = 0; i != Size; i += Scale) {
6545 unsigned Pos = Left ? i + Shift : i;
6546 unsigned Low = Left ? i : i + Shift;
6547 unsigned Len = Scale - Shift;
6548 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6549 Low + (V == V1 ? 0 : Size)))
6553 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6554 bool ByteShift = ShiftEltBits > 64;
6555 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6556 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6557 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6559 // Normalize the scale for byte shifts to still produce an i64 element
6561 Scale = ByteShift ? Scale / 2 : Scale;
6563 // We need to round trip through the appropriate type for the shift.
6564 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6565 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6566 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6567 "Illegal integer vector type");
6568 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6570 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
6571 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6574 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6575 // keep doubling the size of the integer elements up to that. We can
6576 // then shift the elements of the integer vector by whole multiples of
6577 // their width within the elements of the larger integer vector. Test each
6578 // multiple to see if we can find a match with the moved element indices
6579 // and that the shifted in elements are all zeroable.
6580 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6581 for (int Shift = 1; Shift != Scale; ++Shift)
6582 for (bool Left : {true, false})
6583 if (CheckZeros(Shift, Scale, Left))
6584 for (SDValue V : {V1, V2})
6585 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6592 /// \brief Lower a vector shuffle as a zero or any extension.
6594 /// Given a specific number of elements, element bit width, and extension
6595 /// stride, produce either a zero or any extension based on the available
6596 /// features of the subtarget.
6597 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6598 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6599 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6600 assert(Scale > 1 && "Need a scale to extend.");
6601 int NumElements = VT.getVectorNumElements();
6602 int EltBits = VT.getScalarSizeInBits();
6603 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6604 "Only 8, 16, and 32 bit elements can be extended.");
6605 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6607 // Found a valid zext mask! Try various lowering strategies based on the
6608 // input type and available ISA extensions.
6609 if (Subtarget->hasSSE41()) {
6610 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6611 NumElements / Scale);
6612 return DAG.getNode(ISD::BITCAST, DL, VT,
6613 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6616 // For any extends we can cheat for larger element sizes and use shuffle
6617 // instructions that can fold with a load and/or copy.
6618 if (AnyExt && EltBits == 32) {
6619 int PSHUFDMask[4] = {0, -1, 1, -1};
6621 ISD::BITCAST, DL, VT,
6622 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6623 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6624 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
6626 if (AnyExt && EltBits == 16 && Scale > 2) {
6627 int PSHUFDMask[4] = {0, -1, 0, -1};
6628 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6629 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6630 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
6631 int PSHUFHWMask[4] = {1, -1, -1, -1};
6633 ISD::BITCAST, DL, VT,
6634 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6635 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6636 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
6639 // If this would require more than 2 unpack instructions to expand, use
6640 // pshufb when available. We can only use more than 2 unpack instructions
6641 // when zero extending i8 elements which also makes it easier to use pshufb.
6642 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6643 assert(NumElements == 16 && "Unexpected byte vector width!");
6644 SDValue PSHUFBMask[16];
6645 for (int i = 0; i < 16; ++i)
6647 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
6648 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6649 return DAG.getNode(ISD::BITCAST, DL, VT,
6650 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6651 DAG.getNode(ISD::BUILD_VECTOR, DL,
6652 MVT::v16i8, PSHUFBMask)));
6655 // Otherwise emit a sequence of unpacks.
6657 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6658 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6659 : getZeroVector(InputVT, Subtarget, DAG, DL);
6660 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6661 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6665 } while (Scale > 1);
6666 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6669 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6671 /// This routine will try to do everything in its power to cleverly lower
6672 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6673 /// check for the profitability of this lowering, it tries to aggressively
6674 /// match this pattern. It will use all of the micro-architectural details it
6675 /// can to emit an efficient lowering. It handles both blends with all-zero
6676 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6677 /// masking out later).
6679 /// The reason we have dedicated lowering for zext-style shuffles is that they
6680 /// are both incredibly common and often quite performance sensitive.
6681 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6682 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6683 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6684 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6686 int Bits = VT.getSizeInBits();
6687 int NumElements = VT.getVectorNumElements();
6688 assert(VT.getScalarSizeInBits() <= 32 &&
6689 "Exceeds 32-bit integer zero extension limit");
6690 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6692 // Define a helper function to check a particular ext-scale and lower to it if
6694 auto Lower = [&](int Scale) -> SDValue {
6697 for (int i = 0; i < NumElements; ++i) {
6699 continue; // Valid anywhere but doesn't tell us anything.
6700 if (i % Scale != 0) {
6701 // Each of the extended elements need to be zeroable.
6705 // We no longer are in the anyext case.
6710 // Each of the base elements needs to be consecutive indices into the
6711 // same input vector.
6712 SDValue V = Mask[i] < NumElements ? V1 : V2;
6715 else if (InputV != V)
6716 return SDValue(); // Flip-flopping inputs.
6718 if (Mask[i] % NumElements != i / Scale)
6719 return SDValue(); // Non-consecutive strided elements.
6722 // If we fail to find an input, we have a zero-shuffle which should always
6723 // have already been handled.
6724 // FIXME: Maybe handle this here in case during blending we end up with one?
6728 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6729 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6732 // The widest scale possible for extending is to a 64-bit integer.
6733 assert(Bits % 64 == 0 &&
6734 "The number of bits in a vector must be divisible by 64 on x86!");
6735 int NumExtElements = Bits / 64;
6737 // Each iteration, try extending the elements half as much, but into twice as
6739 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6740 assert(NumElements % NumExtElements == 0 &&
6741 "The input vector size must be divisible by the extended size.");
6742 if (SDValue V = Lower(NumElements / NumExtElements))
6746 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6750 // Returns one of the source operands if the shuffle can be reduced to a
6751 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6752 auto CanZExtLowHalf = [&]() {
6753 for (int i = NumElements / 2; i != NumElements; ++i)
6756 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6758 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
6763 if (SDValue V = CanZExtLowHalf()) {
6764 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
6765 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
6766 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6769 // No viable ext lowering found.
6773 /// \brief Try to get a scalar value for a specific element of a vector.
6775 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
6776 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
6777 SelectionDAG &DAG) {
6778 MVT VT = V.getSimpleValueType();
6779 MVT EltVT = VT.getVectorElementType();
6780 while (V.getOpcode() == ISD::BITCAST)
6781 V = V.getOperand(0);
6782 // If the bitcasts shift the element size, we can't extract an equivalent
6784 MVT NewVT = V.getSimpleValueType();
6785 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
6788 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6789 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
6790 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
6795 /// \brief Helper to test for a load that can be folded with x86 shuffles.
6797 /// This is particularly important because the set of instructions varies
6798 /// significantly based on whether the operand is a load or not.
6799 static bool isShuffleFoldableLoad(SDValue V) {
6800 while (V.getOpcode() == ISD::BITCAST)
6801 V = V.getOperand(0);
6803 return ISD::isNON_EXTLoad(V.getNode());
6806 /// \brief Try to lower insertion of a single element into a zero vector.
6808 /// This is a common pattern that we have especially efficient patterns to lower
6809 /// across all subtarget feature sets.
6810 static SDValue lowerVectorShuffleAsElementInsertion(
6811 MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6812 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6813 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6815 MVT EltVT = VT.getVectorElementType();
6817 int V2Index = std::find_if(Mask.begin(), Mask.end(),
6818 [&Mask](int M) { return M >= (int)Mask.size(); }) -
6820 bool IsV1Zeroable = true;
6821 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6822 if (i != V2Index && !Zeroable[i]) {
6823 IsV1Zeroable = false;
6827 // Check for a single input from a SCALAR_TO_VECTOR node.
6828 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
6829 // all the smarts here sunk into that routine. However, the current
6830 // lowering of BUILD_VECTOR makes that nearly impossible until the old
6831 // vector shuffle lowering is dead.
6832 if (SDValue V2S = getScalarValueForVectorElement(
6833 V2, Mask[V2Index] - Mask.size(), DAG)) {
6834 // We need to zext the scalar if it is smaller than an i32.
6835 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
6836 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
6837 // Using zext to expand a narrow element won't work for non-zero
6842 // Zero-extend directly to i32.
6844 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
6846 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
6847 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
6848 EltVT == MVT::i16) {
6849 // Either not inserting from the low element of the input or the input
6850 // element size is too small to use VZEXT_MOVL to clear the high bits.
6854 if (!IsV1Zeroable) {
6855 // If V1 can't be treated as a zero vector we have fewer options to lower
6856 // this. We can't support integer vectors or non-zero targets cheaply, and
6857 // the V1 elements can't be permuted in any way.
6858 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
6859 if (!VT.isFloatingPoint() || V2Index != 0)
6861 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
6862 V1Mask[V2Index] = -1;
6863 if (!isNoopShuffleMask(V1Mask))
6865 // This is essentially a special case blend operation, but if we have
6866 // general purpose blend operations, they are always faster. Bail and let
6867 // the rest of the lowering handle these as blends.
6868 if (Subtarget->hasSSE41())
6871 // Otherwise, use MOVSD or MOVSS.
6872 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
6873 "Only two types of floating point element types to handle!");
6874 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
6878 // This lowering only works for the low element with floating point vectors.
6879 if (VT.isFloatingPoint() && V2Index != 0)
6882 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
6884 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6887 // If we have 4 or fewer lanes we can cheaply shuffle the element into
6888 // the desired position. Otherwise it is more efficient to do a vector
6889 // shift left. We know that we can do a vector shift left because all
6890 // the inputs are zero.
6891 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
6892 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
6893 V2Shuffle[V2Index] = 0;
6894 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
6896 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
6898 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
6900 V2Index * EltVT.getSizeInBits()/8,
6901 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
6902 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6908 /// \brief Try to lower broadcast of a single element.
6910 /// For convenience, this code also bundles all of the subtarget feature set
6911 /// filtering. While a little annoying to re-dispatch on type here, there isn't
6912 /// a convenient way to factor it out.
6913 static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V,
6915 const X86Subtarget *Subtarget,
6916 SelectionDAG &DAG) {
6917 if (!Subtarget->hasAVX())
6919 if (VT.isInteger() && !Subtarget->hasAVX2())
6922 // Check that the mask is a broadcast.
6923 int BroadcastIdx = -1;
6925 if (M >= 0 && BroadcastIdx == -1)
6927 else if (M >= 0 && M != BroadcastIdx)
6930 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
6931 "a sorted mask where the broadcast "
6934 // Go up the chain of (vector) values to try and find a scalar load that
6935 // we can combine with the broadcast.
6937 switch (V.getOpcode()) {
6938 case ISD::CONCAT_VECTORS: {
6939 int OperandSize = Mask.size() / V.getNumOperands();
6940 V = V.getOperand(BroadcastIdx / OperandSize);
6941 BroadcastIdx %= OperandSize;
6945 case ISD::INSERT_SUBVECTOR: {
6946 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
6947 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
6951 int BeginIdx = (int)ConstantIdx->getZExtValue();
6953 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
6954 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
6955 BroadcastIdx -= BeginIdx;
6966 // Check if this is a broadcast of a scalar. We special case lowering
6967 // for scalars so that we can more effectively fold with loads.
6968 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6969 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
6970 V = V.getOperand(BroadcastIdx);
6972 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
6974 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
6976 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
6977 // We can't broadcast from a vector register w/o AVX2, and we can only
6978 // broadcast from the zero-element of a vector register.
6982 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
6985 // Check for whether we can use INSERTPS to perform the shuffle. We only use
6986 // INSERTPS when the V1 elements are already in the correct locations
6987 // because otherwise we can just always use two SHUFPS instructions which
6988 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
6989 // perform INSERTPS if a single V1 element is out of place and all V2
6990 // elements are zeroable.
6991 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
6993 SelectionDAG &DAG) {
6994 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
6995 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
6996 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
6997 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
6999 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7002 int V1DstIndex = -1;
7003 int V2DstIndex = -1;
7004 bool V1UsedInPlace = false;
7006 for (int i = 0; i < 4; ++i) {
7007 // Synthesize a zero mask from the zeroable elements (includes undefs).
7013 // Flag if we use any V1 inputs in place.
7015 V1UsedInPlace = true;
7019 // We can only insert a single non-zeroable element.
7020 if (V1DstIndex != -1 || V2DstIndex != -1)
7024 // V1 input out of place for insertion.
7027 // V2 input for insertion.
7032 // Don't bother if we have no (non-zeroable) element for insertion.
7033 if (V1DstIndex == -1 && V2DstIndex == -1)
7036 // Determine element insertion src/dst indices. The src index is from the
7037 // start of the inserted vector, not the start of the concatenated vector.
7038 unsigned V2SrcIndex = 0;
7039 if (V1DstIndex != -1) {
7040 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7041 // and don't use the original V2 at all.
7042 V2SrcIndex = Mask[V1DstIndex];
7043 V2DstIndex = V1DstIndex;
7046 V2SrcIndex = Mask[V2DstIndex] - 4;
7049 // If no V1 inputs are used in place, then the result is created only from
7050 // the zero mask and the V2 insertion - so remove V1 dependency.
7052 V1 = DAG.getUNDEF(MVT::v4f32);
7054 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7055 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7057 // Insert the V2 element into the desired position.
7059 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7060 DAG.getConstant(InsertPSMask, MVT::i8));
7063 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7064 /// UNPCK instruction.
7066 /// This specifically targets cases where we end up with alternating between
7067 /// the two inputs, and so can permute them into something that feeds a single
7068 /// UNPCK instruction. Note that this routine only targets integer vectors
7069 /// because for floating point vectors we have a generalized SHUFPS lowering
7070 /// strategy that handles everything that doesn't *exactly* match an unpack,
7071 /// making this clever lowering unnecessary.
7072 static SDValue lowerVectorShuffleAsUnpack(MVT VT, SDLoc DL, SDValue V1,
7073 SDValue V2, ArrayRef<int> Mask,
7074 SelectionDAG &DAG) {
7075 assert(!VT.isFloatingPoint() &&
7076 "This routine only supports integer vectors.");
7077 assert(!isSingleInputShuffleMask(Mask) &&
7078 "This routine should only be used when blending two inputs.");
7079 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7081 int Size = Mask.size();
7083 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7084 return M >= 0 && M % Size < Size / 2;
7086 int NumHiInputs = std::count_if(
7087 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7089 bool UnpackLo = NumLoInputs >= NumHiInputs;
7091 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7092 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7093 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7095 for (int i = 0; i < Size; ++i) {
7099 // Each element of the unpack contains Scale elements from this mask.
7100 int UnpackIdx = i / Scale;
7102 // We only handle the case where V1 feeds the first slots of the unpack.
7103 // We rely on canonicalization to ensure this is the case.
7104 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7107 // Setup the mask for this input. The indexing is tricky as we have to
7108 // handle the unpack stride.
7109 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7110 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7114 // If we will have to shuffle both inputs to use the unpack, check whether
7115 // we can just unpack first and shuffle the result. If so, skip this unpack.
7116 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7117 !isNoopShuffleMask(V2Mask))
7120 // Shuffle the inputs into place.
7121 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7122 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7124 // Cast the inputs to the type we will use to unpack them.
7125 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7126 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7128 // Unpack the inputs and cast the result back to the desired type.
7129 return DAG.getNode(ISD::BITCAST, DL, VT,
7130 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7131 DL, UnpackVT, V1, V2));
7134 // We try each unpack from the largest to the smallest to try and find one
7135 // that fits this mask.
7136 int OrigNumElements = VT.getVectorNumElements();
7137 int OrigScalarSize = VT.getScalarSizeInBits();
7138 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7139 int Scale = ScalarSize / OrigScalarSize;
7140 int NumElements = OrigNumElements / Scale;
7141 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7142 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7146 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7148 if (NumLoInputs == 0 || NumHiInputs == 0) {
7149 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7150 "We have to have *some* inputs!");
7151 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7153 // FIXME: We could consider the total complexity of the permute of each
7154 // possible unpacking. Or at the least we should consider how many
7155 // half-crossings are created.
7156 // FIXME: We could consider commuting the unpacks.
7158 SmallVector<int, 32> PermMask;
7159 PermMask.assign(Size, -1);
7160 for (int i = 0; i < Size; ++i) {
7164 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7167 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7169 return DAG.getVectorShuffle(
7170 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7172 DAG.getUNDEF(VT), PermMask);
7178 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7180 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7181 /// support for floating point shuffles but not integer shuffles. These
7182 /// instructions will incur a domain crossing penalty on some chips though so
7183 /// it is better to avoid lowering through this for integer vectors where
7185 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7186 const X86Subtarget *Subtarget,
7187 SelectionDAG &DAG) {
7189 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7190 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7191 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7192 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7193 ArrayRef<int> Mask = SVOp->getMask();
7194 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7196 if (isSingleInputShuffleMask(Mask)) {
7197 // Use low duplicate instructions for masks that match their pattern.
7198 if (Subtarget->hasSSE3())
7199 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7200 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7202 // Straight shuffle of a single input vector. Simulate this by using the
7203 // single input as both of the "inputs" to this instruction..
7204 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7206 if (Subtarget->hasAVX()) {
7207 // If we have AVX, we can use VPERMILPS which will allow folding a load
7208 // into the shuffle.
7209 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7210 DAG.getConstant(SHUFPDMask, MVT::i8));
7213 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
7214 DAG.getConstant(SHUFPDMask, MVT::i8));
7216 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7217 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7219 // If we have a single input, insert that into V1 if we can do so cheaply.
7220 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7221 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7222 MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
7224 // Try inverting the insertion since for v2 masks it is easy to do and we
7225 // can't reliably sort the mask one way or the other.
7226 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7227 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7228 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7229 MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG))
7233 // Try to use one of the special instruction patterns to handle two common
7234 // blend patterns if a zero-blend above didn't work.
7235 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7236 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7237 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7238 // We can either use a special instruction to load over the low double or
7239 // to move just the low double.
7241 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7243 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7245 if (Subtarget->hasSSE41())
7246 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7250 // Use dedicated unpack instructions for masks that match their pattern.
7251 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7252 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7253 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7254 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7256 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7257 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
7258 DAG.getConstant(SHUFPDMask, MVT::i8));
7261 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7263 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7264 /// the integer unit to minimize domain crossing penalties. However, for blends
7265 /// it falls back to the floating point shuffle operation with appropriate bit
7267 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7268 const X86Subtarget *Subtarget,
7269 SelectionDAG &DAG) {
7271 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7272 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7273 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7274 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7275 ArrayRef<int> Mask = SVOp->getMask();
7276 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7278 if (isSingleInputShuffleMask(Mask)) {
7279 // Check for being able to broadcast a single element.
7280 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1,
7281 Mask, Subtarget, DAG))
7284 // Straight shuffle of a single input vector. For everything from SSE2
7285 // onward this has a single fast instruction with no scary immediates.
7286 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7287 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7288 int WidenedMask[4] = {
7289 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7290 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7292 ISD::BITCAST, DL, MVT::v2i64,
7293 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7294 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7296 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7297 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7298 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7299 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7301 // If we have a blend of two PACKUS operations an the blend aligns with the
7302 // low and half halves, we can just merge the PACKUS operations. This is
7303 // particularly important as it lets us merge shuffles that this routine itself
7305 auto GetPackNode = [](SDValue V) {
7306 while (V.getOpcode() == ISD::BITCAST)
7307 V = V.getOperand(0);
7309 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7311 if (SDValue V1Pack = GetPackNode(V1))
7312 if (SDValue V2Pack = GetPackNode(V2))
7313 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7314 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7315 Mask[0] == 0 ? V1Pack.getOperand(0)
7316 : V1Pack.getOperand(1),
7317 Mask[1] == 2 ? V2Pack.getOperand(0)
7318 : V2Pack.getOperand(1)));
7320 // Try to use shift instructions.
7322 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7325 // When loading a scalar and then shuffling it into a vector we can often do
7326 // the insertion cheaply.
7327 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7328 MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
7330 // Try inverting the insertion since for v2 masks it is easy to do and we
7331 // can't reliably sort the mask one way or the other.
7332 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7333 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7334 MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG))
7337 // We have different paths for blend lowering, but they all must use the
7338 // *exact* same predicate.
7339 bool IsBlendSupported = Subtarget->hasSSE41();
7340 if (IsBlendSupported)
7341 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7345 // Use dedicated unpack instructions for masks that match their pattern.
7346 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7347 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7348 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7349 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7351 // Try to use byte rotation instructions.
7352 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7353 if (Subtarget->hasSSSE3())
7354 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7355 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7358 // If we have direct support for blends, we should lower by decomposing into
7359 // a permute. That will be faster than the domain cross.
7360 if (IsBlendSupported)
7361 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7364 // We implement this with SHUFPD which is pretty lame because it will likely
7365 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7366 // However, all the alternatives are still more cycles and newer chips don't
7367 // have this problem. It would be really nice if x86 had better shuffles here.
7368 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7369 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7370 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7371 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7374 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7376 /// This is used to disable more specialized lowerings when the shufps lowering
7377 /// will happen to be efficient.
7378 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7379 // This routine only handles 128-bit shufps.
7380 assert(Mask.size() == 4 && "Unsupported mask size!");
7382 // To lower with a single SHUFPS we need to have the low half and high half
7383 // each requiring a single input.
7384 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7386 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7392 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7394 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7395 /// It makes no assumptions about whether this is the *best* lowering, it simply
7397 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7398 ArrayRef<int> Mask, SDValue V1,
7399 SDValue V2, SelectionDAG &DAG) {
7400 SDValue LowV = V1, HighV = V2;
7401 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7404 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7406 if (NumV2Elements == 1) {
7408 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7411 // Compute the index adjacent to V2Index and in the same half by toggling
7413 int V2AdjIndex = V2Index ^ 1;
7415 if (Mask[V2AdjIndex] == -1) {
7416 // Handles all the cases where we have a single V2 element and an undef.
7417 // This will only ever happen in the high lanes because we commute the
7418 // vector otherwise.
7420 std::swap(LowV, HighV);
7421 NewMask[V2Index] -= 4;
7423 // Handle the case where the V2 element ends up adjacent to a V1 element.
7424 // To make this work, blend them together as the first step.
7425 int V1Index = V2AdjIndex;
7426 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7427 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7428 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7430 // Now proceed to reconstruct the final blend as we have the necessary
7431 // high or low half formed.
7438 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7439 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7441 } else if (NumV2Elements == 2) {
7442 if (Mask[0] < 4 && Mask[1] < 4) {
7443 // Handle the easy case where we have V1 in the low lanes and V2 in the
7447 } else if (Mask[2] < 4 && Mask[3] < 4) {
7448 // We also handle the reversed case because this utility may get called
7449 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7450 // arrange things in the right direction.
7456 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7457 // trying to place elements directly, just blend them and set up the final
7458 // shuffle to place them.
7460 // The first two blend mask elements are for V1, the second two are for
7462 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7463 Mask[2] < 4 ? Mask[2] : Mask[3],
7464 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7465 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7466 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7467 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7469 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7472 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7473 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7474 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7475 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7478 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7479 getV4X86ShuffleImm8ForMask(NewMask, DAG));
7482 /// \brief Lower 4-lane 32-bit floating point shuffles.
7484 /// Uses instructions exclusively from the floating point unit to minimize
7485 /// domain crossing penalties, as these are sufficient to implement all v4f32
7487 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7488 const X86Subtarget *Subtarget,
7489 SelectionDAG &DAG) {
7491 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7492 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7493 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7494 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7495 ArrayRef<int> Mask = SVOp->getMask();
7496 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7499 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7501 if (NumV2Elements == 0) {
7502 // Check for being able to broadcast a single element.
7503 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1,
7504 Mask, Subtarget, DAG))
7507 // Use even/odd duplicate instructions for masks that match their pattern.
7508 if (Subtarget->hasSSE3()) {
7509 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7510 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7511 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7512 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7515 if (Subtarget->hasAVX()) {
7516 // If we have AVX, we can use VPERMILPS which will allow folding a load
7517 // into the shuffle.
7518 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7519 getV4X86ShuffleImm8ForMask(Mask, DAG));
7522 // Otherwise, use a straight shuffle of a single input vector. We pass the
7523 // input vector to both operands to simulate this with a SHUFPS.
7524 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7525 getV4X86ShuffleImm8ForMask(Mask, DAG));
7528 // There are special ways we can lower some single-element blends. However, we
7529 // have custom ways we can lower more complex single-element blends below that
7530 // we defer to if both this and BLENDPS fail to match, so restrict this to
7531 // when the V2 input is targeting element 0 of the mask -- that is the fast
7533 if (NumV2Elements == 1 && Mask[0] >= 4)
7534 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
7535 Mask, Subtarget, DAG))
7538 if (Subtarget->hasSSE41()) {
7539 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7543 // Use INSERTPS if we can complete the shuffle efficiently.
7544 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7547 if (!isSingleSHUFPSMask(Mask))
7548 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7549 DL, MVT::v4f32, V1, V2, Mask, DAG))
7553 // Use dedicated unpack instructions for masks that match their pattern.
7554 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7555 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7556 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7557 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7558 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7559 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7560 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7561 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7563 // Otherwise fall back to a SHUFPS lowering strategy.
7564 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7567 /// \brief Lower 4-lane i32 vector shuffles.
7569 /// We try to handle these with integer-domain shuffles where we can, but for
7570 /// blends we use the floating point domain blend instructions.
7571 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7572 const X86Subtarget *Subtarget,
7573 SelectionDAG &DAG) {
7575 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7576 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7577 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7578 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7579 ArrayRef<int> Mask = SVOp->getMask();
7580 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7582 // Whenever we can lower this as a zext, that instruction is strictly faster
7583 // than any alternative. It also allows us to fold memory operands into the
7584 // shuffle in many cases.
7585 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7586 Mask, Subtarget, DAG))
7590 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7592 if (NumV2Elements == 0) {
7593 // Check for being able to broadcast a single element.
7594 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1,
7595 Mask, Subtarget, DAG))
7598 // Straight shuffle of a single input vector. For everything from SSE2
7599 // onward this has a single fast instruction with no scary immediates.
7600 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7601 // but we aren't actually going to use the UNPCK instruction because doing
7602 // so prevents folding a load into this instruction or making a copy.
7603 const int UnpackLoMask[] = {0, 0, 1, 1};
7604 const int UnpackHiMask[] = {2, 2, 3, 3};
7605 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7606 Mask = UnpackLoMask;
7607 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7608 Mask = UnpackHiMask;
7610 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7611 getV4X86ShuffleImm8ForMask(Mask, DAG));
7614 // Try to use shift instructions.
7616 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7619 // There are special ways we can lower some single-element blends.
7620 if (NumV2Elements == 1)
7621 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
7622 Mask, Subtarget, DAG))
7625 // We have different paths for blend lowering, but they all must use the
7626 // *exact* same predicate.
7627 bool IsBlendSupported = Subtarget->hasSSE41();
7628 if (IsBlendSupported)
7629 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7633 if (SDValue Masked =
7634 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7637 // Use dedicated unpack instructions for masks that match their pattern.
7638 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7639 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7640 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7641 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7642 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7643 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7644 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7645 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7647 // Try to use byte rotation instructions.
7648 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7649 if (Subtarget->hasSSSE3())
7650 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7651 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7654 // If we have direct support for blends, we should lower by decomposing into
7655 // a permute. That will be faster than the domain cross.
7656 if (IsBlendSupported)
7657 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7660 // Try to lower by permuting the inputs into an unpack instruction.
7661 if (SDValue Unpack =
7662 lowerVectorShuffleAsUnpack(MVT::v4i32, DL, V1, V2, Mask, DAG))
7665 // We implement this with SHUFPS because it can blend from two vectors.
7666 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7667 // up the inputs, bypassing domain shift penalties that we would encur if we
7668 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7670 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7671 DAG.getVectorShuffle(
7673 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7674 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7677 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7678 /// shuffle lowering, and the most complex part.
7680 /// The lowering strategy is to try to form pairs of input lanes which are
7681 /// targeted at the same half of the final vector, and then use a dword shuffle
7682 /// to place them onto the right half, and finally unpack the paired lanes into
7683 /// their final position.
7685 /// The exact breakdown of how to form these dword pairs and align them on the
7686 /// correct sides is really tricky. See the comments within the function for
7687 /// more of the details.
7688 static SDValue lowerV8I16SingleInputVectorShuffle(
7689 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
7690 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7691 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7692 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7693 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7695 SmallVector<int, 4> LoInputs;
7696 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7697 [](int M) { return M >= 0; });
7698 std::sort(LoInputs.begin(), LoInputs.end());
7699 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7700 SmallVector<int, 4> HiInputs;
7701 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7702 [](int M) { return M >= 0; });
7703 std::sort(HiInputs.begin(), HiInputs.end());
7704 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7706 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7707 int NumHToL = LoInputs.size() - NumLToL;
7709 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7710 int NumHToH = HiInputs.size() - NumLToH;
7711 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7712 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7713 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7714 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7716 // Check for being able to broadcast a single element.
7717 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V,
7718 Mask, Subtarget, DAG))
7721 // Try to use shift instructions.
7723 lowerVectorShuffleAsShift(DL, MVT::v8i16, V, V, Mask, DAG))
7726 // Use dedicated unpack instructions for masks that match their pattern.
7727 if (isShuffleEquivalent(V, V, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
7728 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
7729 if (isShuffleEquivalent(V, V, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
7730 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
7732 // Try to use byte rotation instructions.
7733 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7734 DL, MVT::v8i16, V, V, Mask, Subtarget, DAG))
7737 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7738 // such inputs we can swap two of the dwords across the half mark and end up
7739 // with <=2 inputs to each half in each half. Once there, we can fall through
7740 // to the generic code below. For example:
7742 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7743 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7745 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7746 // and an existing 2-into-2 on the other half. In this case we may have to
7747 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7748 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7749 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7750 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7751 // half than the one we target for fixing) will be fixed when we re-enter this
7752 // path. We will also combine away any sequence of PSHUFD instructions that
7753 // result into a single instruction. Here is an example of the tricky case:
7755 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7756 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7758 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7760 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7761 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7763 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7764 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7766 // The result is fine to be handled by the generic logic.
7767 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
7768 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
7769 int AOffset, int BOffset) {
7770 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
7771 "Must call this with A having 3 or 1 inputs from the A half.");
7772 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
7773 "Must call this with B having 1 or 3 inputs from the B half.");
7774 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
7775 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
7777 // Compute the index of dword with only one word among the three inputs in
7778 // a half by taking the sum of the half with three inputs and subtracting
7779 // the sum of the actual three inputs. The difference is the remaining
7782 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
7783 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
7784 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
7785 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
7786 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
7787 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
7788 int TripleNonInputIdx =
7789 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
7790 TripleDWord = TripleNonInputIdx / 2;
7792 // We use xor with one to compute the adjacent DWord to whichever one the
7794 OneInputDWord = (OneInput / 2) ^ 1;
7796 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
7797 // and BToA inputs. If there is also such a problem with the BToB and AToB
7798 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
7799 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
7800 // is essential that we don't *create* a 3<-1 as then we might oscillate.
7801 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
7802 // Compute how many inputs will be flipped by swapping these DWords. We
7804 // to balance this to ensure we don't form a 3-1 shuffle in the other
7806 int NumFlippedAToBInputs =
7807 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
7808 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
7809 int NumFlippedBToBInputs =
7810 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
7811 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
7812 if ((NumFlippedAToBInputs == 1 &&
7813 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
7814 (NumFlippedBToBInputs == 1 &&
7815 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
7816 // We choose whether to fix the A half or B half based on whether that
7817 // half has zero flipped inputs. At zero, we may not be able to fix it
7818 // with that half. We also bias towards fixing the B half because that
7819 // will more commonly be the high half, and we have to bias one way.
7820 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
7821 ArrayRef<int> Inputs) {
7822 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
7823 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
7824 PinnedIdx ^ 1) != Inputs.end();
7825 // Determine whether the free index is in the flipped dword or the
7826 // unflipped dword based on where the pinned index is. We use this bit
7827 // in an xor to conditionally select the adjacent dword.
7828 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
7829 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7830 FixFreeIdx) != Inputs.end();
7831 if (IsFixIdxInput == IsFixFreeIdxInput)
7833 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7834 FixFreeIdx) != Inputs.end();
7835 assert(IsFixIdxInput != IsFixFreeIdxInput &&
7836 "We need to be changing the number of flipped inputs!");
7837 int PSHUFHalfMask[] = {0, 1, 2, 3};
7838 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
7839 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
7841 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
7844 if (M != -1 && M == FixIdx)
7846 else if (M != -1 && M == FixFreeIdx)
7849 if (NumFlippedBToBInputs != 0) {
7851 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7852 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
7854 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
7856 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7857 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
7862 int PSHUFDMask[] = {0, 1, 2, 3};
7863 PSHUFDMask[ADWord] = BDWord;
7864 PSHUFDMask[BDWord] = ADWord;
7865 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7866 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7867 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
7868 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7870 // Adjust the mask to match the new locations of A and B.
7872 if (M != -1 && M/2 == ADWord)
7873 M = 2 * BDWord + M % 2;
7874 else if (M != -1 && M/2 == BDWord)
7875 M = 2 * ADWord + M % 2;
7877 // Recurse back into this routine to re-compute state now that this isn't
7878 // a 3 and 1 problem.
7879 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
7882 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
7883 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
7884 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
7885 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
7887 // At this point there are at most two inputs to the low and high halves from
7888 // each half. That means the inputs can always be grouped into dwords and
7889 // those dwords can then be moved to the correct half with a dword shuffle.
7890 // We use at most one low and one high word shuffle to collect these paired
7891 // inputs into dwords, and finally a dword shuffle to place them.
7892 int PSHUFLMask[4] = {-1, -1, -1, -1};
7893 int PSHUFHMask[4] = {-1, -1, -1, -1};
7894 int PSHUFDMask[4] = {-1, -1, -1, -1};
7896 // First fix the masks for all the inputs that are staying in their
7897 // original halves. This will then dictate the targets of the cross-half
7899 auto fixInPlaceInputs =
7900 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
7901 MutableArrayRef<int> SourceHalfMask,
7902 MutableArrayRef<int> HalfMask, int HalfOffset) {
7903 if (InPlaceInputs.empty())
7905 if (InPlaceInputs.size() == 1) {
7906 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7907 InPlaceInputs[0] - HalfOffset;
7908 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
7911 if (IncomingInputs.empty()) {
7912 // Just fix all of the in place inputs.
7913 for (int Input : InPlaceInputs) {
7914 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
7915 PSHUFDMask[Input / 2] = Input / 2;
7920 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
7921 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7922 InPlaceInputs[0] - HalfOffset;
7923 // Put the second input next to the first so that they are packed into
7924 // a dword. We find the adjacent index by toggling the low bit.
7925 int AdjIndex = InPlaceInputs[0] ^ 1;
7926 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
7927 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
7928 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
7930 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
7931 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
7933 // Now gather the cross-half inputs and place them into a free dword of
7934 // their target half.
7935 // FIXME: This operation could almost certainly be simplified dramatically to
7936 // look more like the 3-1 fixing operation.
7937 auto moveInputsToRightHalf = [&PSHUFDMask](
7938 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
7939 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
7940 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
7942 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
7943 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
7945 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
7947 int LowWord = Word & ~1;
7948 int HighWord = Word | 1;
7949 return isWordClobbered(SourceHalfMask, LowWord) ||
7950 isWordClobbered(SourceHalfMask, HighWord);
7953 if (IncomingInputs.empty())
7956 if (ExistingInputs.empty()) {
7957 // Map any dwords with inputs from them into the right half.
7958 for (int Input : IncomingInputs) {
7959 // If the source half mask maps over the inputs, turn those into
7960 // swaps and use the swapped lane.
7961 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
7962 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
7963 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
7964 Input - SourceOffset;
7965 // We have to swap the uses in our half mask in one sweep.
7966 for (int &M : HalfMask)
7967 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
7969 else if (M == Input)
7970 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7972 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
7973 Input - SourceOffset &&
7974 "Previous placement doesn't match!");
7976 // Note that this correctly re-maps both when we do a swap and when
7977 // we observe the other side of the swap above. We rely on that to
7978 // avoid swapping the members of the input list directly.
7979 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7982 // Map the input's dword into the correct half.
7983 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
7984 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
7986 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
7988 "Previous placement doesn't match!");
7991 // And just directly shift any other-half mask elements to be same-half
7992 // as we will have mirrored the dword containing the element into the
7993 // same position within that half.
7994 for (int &M : HalfMask)
7995 if (M >= SourceOffset && M < SourceOffset + 4) {
7996 M = M - SourceOffset + DestOffset;
7997 assert(M >= 0 && "This should never wrap below zero!");
8002 // Ensure we have the input in a viable dword of its current half. This
8003 // is particularly tricky because the original position may be clobbered
8004 // by inputs being moved and *staying* in that half.
8005 if (IncomingInputs.size() == 1) {
8006 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8007 int InputFixed = std::find(std::begin(SourceHalfMask),
8008 std::end(SourceHalfMask), -1) -
8009 std::begin(SourceHalfMask) + SourceOffset;
8010 SourceHalfMask[InputFixed - SourceOffset] =
8011 IncomingInputs[0] - SourceOffset;
8012 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8014 IncomingInputs[0] = InputFixed;
8016 } else if (IncomingInputs.size() == 2) {
8017 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8018 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8019 // We have two non-adjacent or clobbered inputs we need to extract from
8020 // the source half. To do this, we need to map them into some adjacent
8021 // dword slot in the source mask.
8022 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8023 IncomingInputs[1] - SourceOffset};
8025 // If there is a free slot in the source half mask adjacent to one of
8026 // the inputs, place the other input in it. We use (Index XOR 1) to
8027 // compute an adjacent index.
8028 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8029 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8030 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8031 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8032 InputsFixed[1] = InputsFixed[0] ^ 1;
8033 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8034 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8035 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8036 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8037 InputsFixed[0] = InputsFixed[1] ^ 1;
8038 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8039 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8040 // The two inputs are in the same DWord but it is clobbered and the
8041 // adjacent DWord isn't used at all. Move both inputs to the free
8043 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8044 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8045 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8046 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8048 // The only way we hit this point is if there is no clobbering
8049 // (because there are no off-half inputs to this half) and there is no
8050 // free slot adjacent to one of the inputs. In this case, we have to
8051 // swap an input with a non-input.
8052 for (int i = 0; i < 4; ++i)
8053 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8054 "We can't handle any clobbers here!");
8055 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8056 "Cannot have adjacent inputs here!");
8058 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8059 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8061 // We also have to update the final source mask in this case because
8062 // it may need to undo the above swap.
8063 for (int &M : FinalSourceHalfMask)
8064 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8065 M = InputsFixed[1] + SourceOffset;
8066 else if (M == InputsFixed[1] + SourceOffset)
8067 M = (InputsFixed[0] ^ 1) + SourceOffset;
8069 InputsFixed[1] = InputsFixed[0] ^ 1;
8072 // Point everything at the fixed inputs.
8073 for (int &M : HalfMask)
8074 if (M == IncomingInputs[0])
8075 M = InputsFixed[0] + SourceOffset;
8076 else if (M == IncomingInputs[1])
8077 M = InputsFixed[1] + SourceOffset;
8079 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8080 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8083 llvm_unreachable("Unhandled input size!");
8086 // Now hoist the DWord down to the right half.
8087 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8088 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8089 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8090 for (int &M : HalfMask)
8091 for (int Input : IncomingInputs)
8093 M = FreeDWord * 2 + Input % 2;
8095 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8096 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8097 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8098 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8100 // Now enact all the shuffles we've computed to move the inputs into their
8102 if (!isNoopShuffleMask(PSHUFLMask))
8103 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8104 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
8105 if (!isNoopShuffleMask(PSHUFHMask))
8106 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8107 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
8108 if (!isNoopShuffleMask(PSHUFDMask))
8109 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8110 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8111 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8112 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8114 // At this point, each half should contain all its inputs, and we can then
8115 // just shuffle them into their final position.
8116 assert(std::count_if(LoMask.begin(), LoMask.end(),
8117 [](int M) { return M >= 4; }) == 0 &&
8118 "Failed to lift all the high half inputs to the low mask!");
8119 assert(std::count_if(HiMask.begin(), HiMask.end(),
8120 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8121 "Failed to lift all the low half inputs to the high mask!");
8123 // Do a half shuffle for the low mask.
8124 if (!isNoopShuffleMask(LoMask))
8125 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8126 getV4X86ShuffleImm8ForMask(LoMask, DAG));
8128 // Do a half shuffle with the high mask after shifting its values down.
8129 for (int &M : HiMask)
8132 if (!isNoopShuffleMask(HiMask))
8133 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8134 getV4X86ShuffleImm8ForMask(HiMask, DAG));
8139 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8140 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8141 SDValue V2, ArrayRef<int> Mask,
8142 SelectionDAG &DAG, bool &V1InUse,
8144 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8150 int Size = Mask.size();
8151 int Scale = 16 / Size;
8152 for (int i = 0; i < 16; ++i) {
8153 if (Mask[i / Scale] == -1) {
8154 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8156 const int ZeroMask = 0x80;
8157 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8159 int V2Idx = Mask[i / Scale] < Size
8161 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8162 if (Zeroable[i / Scale])
8163 V1Idx = V2Idx = ZeroMask;
8164 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
8165 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
8166 V1InUse |= (ZeroMask != V1Idx);
8167 V2InUse |= (ZeroMask != V2Idx);
8172 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8173 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8174 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8176 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8177 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8178 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8180 // If we need shuffled inputs from both, blend the two.
8182 if (V1InUse && V2InUse)
8183 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8185 V = V1InUse ? V1 : V2;
8187 // Cast the result back to the correct type.
8188 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8191 /// \brief Generic lowering of 8-lane i16 shuffles.
8193 /// This handles both single-input shuffles and combined shuffle/blends with
8194 /// two inputs. The single input shuffles are immediately delegated to
8195 /// a dedicated lowering routine.
8197 /// The blends are lowered in one of three fundamental ways. If there are few
8198 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8199 /// of the input is significantly cheaper when lowered as an interleaving of
8200 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8201 /// halves of the inputs separately (making them have relatively few inputs)
8202 /// and then concatenate them.
8203 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8204 const X86Subtarget *Subtarget,
8205 SelectionDAG &DAG) {
8207 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8208 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8209 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8210 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8211 ArrayRef<int> OrigMask = SVOp->getMask();
8212 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8213 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8214 MutableArrayRef<int> Mask(MaskStorage);
8216 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8218 // Whenever we can lower this as a zext, that instruction is strictly faster
8219 // than any alternative.
8220 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8221 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8224 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8226 auto isV2 = [](int M) { return M >= 8; };
8228 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8230 if (NumV2Inputs == 0)
8231 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
8233 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8234 "All single-input shuffles should be canonicalized to be V1-input "
8237 // Try to use shift instructions.
8239 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8242 // There are special ways we can lower some single-element blends.
8243 if (NumV2Inputs == 1)
8244 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
8245 Mask, Subtarget, DAG))
8248 // We have different paths for blend lowering, but they all must use the
8249 // *exact* same predicate.
8250 bool IsBlendSupported = Subtarget->hasSSE41();
8251 if (IsBlendSupported)
8252 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8256 if (SDValue Masked =
8257 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8260 // Use dedicated unpack instructions for masks that match their pattern.
8261 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8262 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8263 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8264 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8266 // Try to use byte rotation instructions.
8267 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8268 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8271 if (SDValue BitBlend =
8272 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8275 if (SDValue Unpack =
8276 lowerVectorShuffleAsUnpack(MVT::v8i16, DL, V1, V2, Mask, DAG))
8279 // If we can't directly blend but can use PSHUFB, that will be better as it
8280 // can both shuffle and set up the inefficient blend.
8281 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8282 bool V1InUse, V2InUse;
8283 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8287 // We can always bit-blend if we have to so the fallback strategy is to
8288 // decompose into single-input permutes and blends.
8289 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8293 /// \brief Check whether a compaction lowering can be done by dropping even
8294 /// elements and compute how many times even elements must be dropped.
8296 /// This handles shuffles which take every Nth element where N is a power of
8297 /// two. Example shuffle masks:
8299 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8300 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8301 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8302 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8303 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8304 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8306 /// Any of these lanes can of course be undef.
8308 /// This routine only supports N <= 3.
8309 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8312 /// \returns N above, or the number of times even elements must be dropped if
8313 /// there is such a number. Otherwise returns zero.
8314 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8315 // Figure out whether we're looping over two inputs or just one.
8316 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8318 // The modulus for the shuffle vector entries is based on whether this is
8319 // a single input or not.
8320 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8321 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8322 "We should only be called with masks with a power-of-2 size!");
8324 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8326 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8327 // and 2^3 simultaneously. This is because we may have ambiguity with
8328 // partially undef inputs.
8329 bool ViableForN[3] = {true, true, true};
8331 for (int i = 0, e = Mask.size(); i < e; ++i) {
8332 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8337 bool IsAnyViable = false;
8338 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8339 if (ViableForN[j]) {
8342 // The shuffle mask must be equal to (i * 2^N) % M.
8343 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8346 ViableForN[j] = false;
8348 // Early exit if we exhaust the possible powers of two.
8353 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8357 // Return 0 as there is no viable power of two.
8361 /// \brief Generic lowering of v16i8 shuffles.
8363 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8364 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8365 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8366 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8368 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8369 const X86Subtarget *Subtarget,
8370 SelectionDAG &DAG) {
8372 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8373 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8374 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8375 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8376 ArrayRef<int> Mask = SVOp->getMask();
8377 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8379 // Try to use shift instructions.
8381 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8384 // Try to use byte rotation instructions.
8385 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8386 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8389 // Try to use a zext lowering.
8390 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8391 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8395 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8397 // For single-input shuffles, there are some nicer lowering tricks we can use.
8398 if (NumV2Elements == 0) {
8399 // Check for being able to broadcast a single element.
8400 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1,
8401 Mask, Subtarget, DAG))
8404 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8405 // Notably, this handles splat and partial-splat shuffles more efficiently.
8406 // However, it only makes sense if the pre-duplication shuffle simplifies
8407 // things significantly. Currently, this means we need to be able to
8408 // express the pre-duplication shuffle as an i16 shuffle.
8410 // FIXME: We should check for other patterns which can be widened into an
8411 // i16 shuffle as well.
8412 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8413 for (int i = 0; i < 16; i += 2)
8414 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8419 auto tryToWidenViaDuplication = [&]() -> SDValue {
8420 if (!canWidenViaDuplication(Mask))
8422 SmallVector<int, 4> LoInputs;
8423 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8424 [](int M) { return M >= 0 && M < 8; });
8425 std::sort(LoInputs.begin(), LoInputs.end());
8426 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8428 SmallVector<int, 4> HiInputs;
8429 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8430 [](int M) { return M >= 8; });
8431 std::sort(HiInputs.begin(), HiInputs.end());
8432 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8435 bool TargetLo = LoInputs.size() >= HiInputs.size();
8436 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8437 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8439 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8440 SmallDenseMap<int, int, 8> LaneMap;
8441 for (int I : InPlaceInputs) {
8442 PreDupI16Shuffle[I/2] = I/2;
8445 int j = TargetLo ? 0 : 4, je = j + 4;
8446 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8447 // Check if j is already a shuffle of this input. This happens when
8448 // there are two adjacent bytes after we move the low one.
8449 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8450 // If we haven't yet mapped the input, search for a slot into which
8452 while (j < je && PreDupI16Shuffle[j] != -1)
8456 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8459 // Map this input with the i16 shuffle.
8460 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8463 // Update the lane map based on the mapping we ended up with.
8464 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8467 ISD::BITCAST, DL, MVT::v16i8,
8468 DAG.getVectorShuffle(MVT::v8i16, DL,
8469 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8470 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8472 // Unpack the bytes to form the i16s that will be shuffled into place.
8473 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8474 MVT::v16i8, V1, V1);
8476 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8477 for (int i = 0; i < 16; ++i)
8478 if (Mask[i] != -1) {
8479 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8480 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8481 if (PostDupI16Shuffle[i / 2] == -1)
8482 PostDupI16Shuffle[i / 2] = MappedMask;
8484 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8485 "Conflicting entrties in the original shuffle!");
8488 ISD::BITCAST, DL, MVT::v16i8,
8489 DAG.getVectorShuffle(MVT::v8i16, DL,
8490 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8491 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8493 if (SDValue V = tryToWidenViaDuplication())
8497 // Use dedicated unpack instructions for masks that match their pattern.
8498 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8499 0, 16, 1, 17, 2, 18, 3, 19,
8501 4, 20, 5, 21, 6, 22, 7, 23}))
8502 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8503 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8504 8, 24, 9, 25, 10, 26, 11, 27,
8506 12, 28, 13, 29, 14, 30, 15, 31}))
8507 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8509 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8510 // with PSHUFB. It is important to do this before we attempt to generate any
8511 // blends but after all of the single-input lowerings. If the single input
8512 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8513 // want to preserve that and we can DAG combine any longer sequences into
8514 // a PSHUFB in the end. But once we start blending from multiple inputs,
8515 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8516 // and there are *very* few patterns that would actually be faster than the
8517 // PSHUFB approach because of its ability to zero lanes.
8519 // FIXME: The only exceptions to the above are blends which are exact
8520 // interleavings with direct instructions supporting them. We currently don't
8521 // handle those well here.
8522 if (Subtarget->hasSSSE3()) {
8523 bool V1InUse = false;
8524 bool V2InUse = false;
8526 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8527 DAG, V1InUse, V2InUse);
8529 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8530 // do so. This avoids using them to handle blends-with-zero which is
8531 // important as a single pshufb is significantly faster for that.
8532 if (V1InUse && V2InUse) {
8533 if (Subtarget->hasSSE41())
8534 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8535 Mask, Subtarget, DAG))
8538 // We can use an unpack to do the blending rather than an or in some
8539 // cases. Even though the or may be (very minorly) more efficient, we
8540 // preference this lowering because there are common cases where part of
8541 // the complexity of the shuffles goes away when we do the final blend as
8543 // FIXME: It might be worth trying to detect if the unpack-feeding
8544 // shuffles will both be pshufb, in which case we shouldn't bother with
8546 if (SDValue Unpack =
8547 lowerVectorShuffleAsUnpack(MVT::v16i8, DL, V1, V2, Mask, DAG))
8554 // There are special ways we can lower some single-element blends.
8555 if (NumV2Elements == 1)
8556 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2,
8557 Mask, Subtarget, DAG))
8560 if (SDValue BitBlend =
8561 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8564 // Check whether a compaction lowering can be done. This handles shuffles
8565 // which take every Nth element for some even N. See the helper function for
8568 // We special case these as they can be particularly efficiently handled with
8569 // the PACKUSB instruction on x86 and they show up in common patterns of
8570 // rearranging bytes to truncate wide elements.
8571 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8572 // NumEvenDrops is the power of two stride of the elements. Another way of
8573 // thinking about it is that we need to drop the even elements this many
8574 // times to get the original input.
8575 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8577 // First we need to zero all the dropped bytes.
8578 assert(NumEvenDrops <= 3 &&
8579 "No support for dropping even elements more than 3 times.");
8580 // We use the mask type to pick which bytes are preserved based on how many
8581 // elements are dropped.
8582 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8583 SDValue ByteClearMask =
8584 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8585 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
8586 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8588 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8590 // Now pack things back together.
8591 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8592 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8593 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8594 for (int i = 1; i < NumEvenDrops; ++i) {
8595 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8596 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8602 // Handle multi-input cases by blending single-input shuffles.
8603 if (NumV2Elements > 0)
8604 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8607 // The fallback path for single-input shuffles widens this into two v8i16
8608 // vectors with unpacks, shuffles those, and then pulls them back together
8612 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8613 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8614 for (int i = 0; i < 16; ++i)
8616 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8618 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8620 SDValue VLoHalf, VHiHalf;
8621 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8622 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8624 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8625 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8626 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8627 [](int M) { return M >= 0 && M % 2 == 1; })) {
8628 // Use a mask to drop the high bytes.
8629 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8630 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8631 DAG.getConstant(0x00FF, MVT::v8i16));
8633 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8634 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8636 // Squash the masks to point directly into VLoHalf.
8637 for (int &M : LoBlendMask)
8640 for (int &M : HiBlendMask)
8644 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8645 // VHiHalf so that we can blend them as i16s.
8646 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8647 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8648 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8649 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8652 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8653 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8655 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8658 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8660 /// This routine breaks down the specific type of 128-bit shuffle and
8661 /// dispatches to the lowering routines accordingly.
8662 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8663 MVT VT, const X86Subtarget *Subtarget,
8664 SelectionDAG &DAG) {
8665 switch (VT.SimpleTy) {
8667 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8669 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8671 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8673 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8675 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8677 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8680 llvm_unreachable("Unimplemented!");
8684 /// \brief Helper function to test whether a shuffle mask could be
8685 /// simplified by widening the elements being shuffled.
8687 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8688 /// leaves it in an unspecified state.
8690 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8691 /// shuffle masks. The latter have the special property of a '-2' representing
8692 /// a zero-ed lane of a vector.
8693 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8694 SmallVectorImpl<int> &WidenedMask) {
8695 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8696 // If both elements are undef, its trivial.
8697 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8698 WidenedMask.push_back(SM_SentinelUndef);
8702 // Check for an undef mask and a mask value properly aligned to fit with
8703 // a pair of values. If we find such a case, use the non-undef mask's value.
8704 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8705 WidenedMask.push_back(Mask[i + 1] / 2);
8708 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8709 WidenedMask.push_back(Mask[i] / 2);
8713 // When zeroing, we need to spread the zeroing across both lanes to widen.
8714 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8715 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8716 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8717 WidenedMask.push_back(SM_SentinelZero);
8723 // Finally check if the two mask values are adjacent and aligned with
8725 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8726 WidenedMask.push_back(Mask[i] / 2);
8730 // Otherwise we can't safely widen the elements used in this shuffle.
8733 assert(WidenedMask.size() == Mask.size() / 2 &&
8734 "Incorrect size of mask after widening the elements!");
8739 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8741 /// This routine just extracts two subvectors, shuffles them independently, and
8742 /// then concatenates them back together. This should work effectively with all
8743 /// AVX vector shuffle types.
8744 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8745 SDValue V2, ArrayRef<int> Mask,
8746 SelectionDAG &DAG) {
8747 assert(VT.getSizeInBits() >= 256 &&
8748 "Only for 256-bit or wider vector shuffles!");
8749 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
8750 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
8752 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
8753 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
8755 int NumElements = VT.getVectorNumElements();
8756 int SplitNumElements = NumElements / 2;
8757 MVT ScalarVT = VT.getScalarType();
8758 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
8760 // Rather than splitting build-vectors, just build two narrower build
8761 // vectors. This helps shuffling with splats and zeros.
8762 auto SplitVector = [&](SDValue V) {
8763 while (V.getOpcode() == ISD::BITCAST)
8764 V = V->getOperand(0);
8766 MVT OrigVT = V.getSimpleValueType();
8767 int OrigNumElements = OrigVT.getVectorNumElements();
8768 int OrigSplitNumElements = OrigNumElements / 2;
8769 MVT OrigScalarVT = OrigVT.getScalarType();
8770 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
8774 auto *BV = dyn_cast<BuildVectorSDNode>(V);
8776 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8777 DAG.getIntPtrConstant(0));
8778 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8779 DAG.getIntPtrConstant(OrigSplitNumElements));
8782 SmallVector<SDValue, 16> LoOps, HiOps;
8783 for (int i = 0; i < OrigSplitNumElements; ++i) {
8784 LoOps.push_back(BV->getOperand(i));
8785 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
8787 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
8788 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
8790 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
8791 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
8794 SDValue LoV1, HiV1, LoV2, HiV2;
8795 std::tie(LoV1, HiV1) = SplitVector(V1);
8796 std::tie(LoV2, HiV2) = SplitVector(V2);
8798 // Now create two 4-way blends of these half-width vectors.
8799 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
8800 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
8801 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
8802 for (int i = 0; i < SplitNumElements; ++i) {
8803 int M = HalfMask[i];
8804 if (M >= NumElements) {
8805 if (M >= NumElements + SplitNumElements)
8809 V2BlendMask.push_back(M - NumElements);
8810 V1BlendMask.push_back(-1);
8811 BlendMask.push_back(SplitNumElements + i);
8812 } else if (M >= 0) {
8813 if (M >= SplitNumElements)
8817 V2BlendMask.push_back(-1);
8818 V1BlendMask.push_back(M);
8819 BlendMask.push_back(i);
8821 V2BlendMask.push_back(-1);
8822 V1BlendMask.push_back(-1);
8823 BlendMask.push_back(-1);
8827 // Because the lowering happens after all combining takes place, we need to
8828 // manually combine these blend masks as much as possible so that we create
8829 // a minimal number of high-level vector shuffle nodes.
8831 // First try just blending the halves of V1 or V2.
8832 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
8833 return DAG.getUNDEF(SplitVT);
8834 if (!UseLoV2 && !UseHiV2)
8835 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8836 if (!UseLoV1 && !UseHiV1)
8837 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8839 SDValue V1Blend, V2Blend;
8840 if (UseLoV1 && UseHiV1) {
8842 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8844 // We only use half of V1 so map the usage down into the final blend mask.
8845 V1Blend = UseLoV1 ? LoV1 : HiV1;
8846 for (int i = 0; i < SplitNumElements; ++i)
8847 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
8848 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
8850 if (UseLoV2 && UseHiV2) {
8852 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8854 // We only use half of V2 so map the usage down into the final blend mask.
8855 V2Blend = UseLoV2 ? LoV2 : HiV2;
8856 for (int i = 0; i < SplitNumElements; ++i)
8857 if (BlendMask[i] >= SplitNumElements)
8858 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
8860 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
8862 SDValue Lo = HalfBlend(LoMask);
8863 SDValue Hi = HalfBlend(HiMask);
8864 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
8867 /// \brief Either split a vector in halves or decompose the shuffles and the
8870 /// This is provided as a good fallback for many lowerings of non-single-input
8871 /// shuffles with more than one 128-bit lane. In those cases, we want to select
8872 /// between splitting the shuffle into 128-bit components and stitching those
8873 /// back together vs. extracting the single-input shuffles and blending those
8875 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
8876 SDValue V2, ArrayRef<int> Mask,
8877 SelectionDAG &DAG) {
8878 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
8879 "lower single-input shuffles as it "
8880 "could then recurse on itself.");
8881 int Size = Mask.size();
8883 // If this can be modeled as a broadcast of two elements followed by a blend,
8884 // prefer that lowering. This is especially important because broadcasts can
8885 // often fold with memory operands.
8886 auto DoBothBroadcast = [&] {
8887 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
8890 if (V2BroadcastIdx == -1)
8891 V2BroadcastIdx = M - Size;
8892 else if (M - Size != V2BroadcastIdx)
8894 } else if (M >= 0) {
8895 if (V1BroadcastIdx == -1)
8897 else if (M != V1BroadcastIdx)
8902 if (DoBothBroadcast())
8903 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
8906 // If the inputs all stem from a single 128-bit lane of each input, then we
8907 // split them rather than blending because the split will decompose to
8908 // unusually few instructions.
8909 int LaneCount = VT.getSizeInBits() / 128;
8910 int LaneSize = Size / LaneCount;
8911 SmallBitVector LaneInputs[2];
8912 LaneInputs[0].resize(LaneCount, false);
8913 LaneInputs[1].resize(LaneCount, false);
8914 for (int i = 0; i < Size; ++i)
8916 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
8917 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
8918 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8920 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
8921 // that the decomposed single-input shuffles don't end up here.
8922 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8925 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
8926 /// a permutation and blend of those lanes.
8928 /// This essentially blends the out-of-lane inputs to each lane into the lane
8929 /// from a permuted copy of the vector. This lowering strategy results in four
8930 /// instructions in the worst case for a single-input cross lane shuffle which
8931 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
8932 /// of. Special cases for each particular shuffle pattern should be handled
8933 /// prior to trying this lowering.
8934 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
8935 SDValue V1, SDValue V2,
8937 SelectionDAG &DAG) {
8938 // FIXME: This should probably be generalized for 512-bit vectors as well.
8939 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
8940 int LaneSize = Mask.size() / 2;
8942 // If there are only inputs from one 128-bit lane, splitting will in fact be
8943 // less expensive. The flags track wether the given lane contains an element
8944 // that crosses to another lane.
8945 bool LaneCrossing[2] = {false, false};
8946 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8947 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
8948 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
8949 if (!LaneCrossing[0] || !LaneCrossing[1])
8950 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8952 if (isSingleInputShuffleMask(Mask)) {
8953 SmallVector<int, 32> FlippedBlendMask;
8954 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8955 FlippedBlendMask.push_back(
8956 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
8958 : Mask[i] % LaneSize +
8959 (i / LaneSize) * LaneSize + Size));
8961 // Flip the vector, and blend the results which should now be in-lane. The
8962 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
8963 // 5 for the high source. The value 3 selects the high half of source 2 and
8964 // the value 2 selects the low half of source 2. We only use source 2 to
8965 // allow folding it into a memory operand.
8966 unsigned PERMMask = 3 | 2 << 4;
8967 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
8968 V1, DAG.getConstant(PERMMask, MVT::i8));
8969 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
8972 // This now reduces to two single-input shuffles of V1 and V2 which at worst
8973 // will be handled by the above logic and a blend of the results, much like
8974 // other patterns in AVX.
8975 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8978 /// \brief Handle lowering 2-lane 128-bit shuffles.
8979 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8980 SDValue V2, ArrayRef<int> Mask,
8981 const X86Subtarget *Subtarget,
8982 SelectionDAG &DAG) {
8983 // Blends are faster and handle all the non-lane-crossing cases.
8984 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
8988 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
8989 VT.getVectorNumElements() / 2);
8990 // Check for patterns which can be matched with a single insert of a 128-bit
8992 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1}) ||
8993 isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
8994 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
8995 DAG.getIntPtrConstant(0));
8996 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
8997 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
8998 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9000 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 6, 7})) {
9001 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9002 DAG.getIntPtrConstant(0));
9003 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
9004 DAG.getIntPtrConstant(2));
9005 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9008 // Otherwise form a 128-bit permutation.
9009 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
9010 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
9011 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9012 DAG.getConstant(PermMask, MVT::i8));
9015 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9016 /// shuffling each lane.
9018 /// This will only succeed when the result of fixing the 128-bit lanes results
9019 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9020 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9021 /// the lane crosses early and then use simpler shuffles within each lane.
9023 /// FIXME: It might be worthwhile at some point to support this without
9024 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9025 /// in x86 only floating point has interesting non-repeating shuffles, and even
9026 /// those are still *marginally* more expensive.
9027 static SDValue lowerVectorShuffleByMerging128BitLanes(
9028 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9029 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9030 assert(!isSingleInputShuffleMask(Mask) &&
9031 "This is only useful with multiple inputs.");
9033 int Size = Mask.size();
9034 int LaneSize = 128 / VT.getScalarSizeInBits();
9035 int NumLanes = Size / LaneSize;
9036 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9038 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9039 // check whether the in-128-bit lane shuffles share a repeating pattern.
9040 SmallVector<int, 4> Lanes;
9041 Lanes.resize(NumLanes, -1);
9042 SmallVector<int, 4> InLaneMask;
9043 InLaneMask.resize(LaneSize, -1);
9044 for (int i = 0; i < Size; ++i) {
9048 int j = i / LaneSize;
9051 // First entry we've seen for this lane.
9052 Lanes[j] = Mask[i] / LaneSize;
9053 } else if (Lanes[j] != Mask[i] / LaneSize) {
9054 // This doesn't match the lane selected previously!
9058 // Check that within each lane we have a consistent shuffle mask.
9059 int k = i % LaneSize;
9060 if (InLaneMask[k] < 0) {
9061 InLaneMask[k] = Mask[i] % LaneSize;
9062 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9063 // This doesn't fit a repeating in-lane mask.
9068 // First shuffle the lanes into place.
9069 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9070 VT.getSizeInBits() / 64);
9071 SmallVector<int, 8> LaneMask;
9072 LaneMask.resize(NumLanes * 2, -1);
9073 for (int i = 0; i < NumLanes; ++i)
9074 if (Lanes[i] >= 0) {
9075 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9076 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9079 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9080 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9081 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9083 // Cast it back to the type we actually want.
9084 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9086 // Now do a simple shuffle that isn't lane crossing.
9087 SmallVector<int, 8> NewMask;
9088 NewMask.resize(Size, -1);
9089 for (int i = 0; i < Size; ++i)
9091 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9092 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9093 "Must not introduce lane crosses at this point!");
9095 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9098 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9101 /// This returns true if the elements from a particular input are already in the
9102 /// slot required by the given mask and require no permutation.
9103 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9104 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9105 int Size = Mask.size();
9106 for (int i = 0; i < Size; ++i)
9107 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9113 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9115 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9116 /// isn't available.
9117 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9118 const X86Subtarget *Subtarget,
9119 SelectionDAG &DAG) {
9121 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9122 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9123 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9124 ArrayRef<int> Mask = SVOp->getMask();
9125 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9127 SmallVector<int, 4> WidenedMask;
9128 if (canWidenShuffleElements(Mask, WidenedMask))
9129 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9132 if (isSingleInputShuffleMask(Mask)) {
9133 // Check for being able to broadcast a single element.
9134 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1,
9135 Mask, Subtarget, DAG))
9138 // Use low duplicate instructions for masks that match their pattern.
9139 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9140 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9142 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9143 // Non-half-crossing single input shuffles can be lowerid with an
9144 // interleaved permutation.
9145 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9146 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9147 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9148 DAG.getConstant(VPERMILPMask, MVT::i8));
9151 // With AVX2 we have direct support for this permutation.
9152 if (Subtarget->hasAVX2())
9153 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9154 getV4X86ShuffleImm8ForMask(Mask, DAG));
9156 // Otherwise, fall back.
9157 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9161 // X86 has dedicated unpack instructions that can handle specific blend
9162 // operations: UNPCKH and UNPCKL.
9163 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9164 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9165 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9166 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9167 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9168 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9169 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9170 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9172 // If we have a single input to the zero element, insert that into V1 if we
9173 // can do so cheaply.
9175 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
9176 if (NumV2Elements == 1 && Mask[0] >= 4)
9177 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9178 MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
9181 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9185 // Check if the blend happens to exactly fit that of SHUFPD.
9186 if ((Mask[0] == -1 || Mask[0] < 2) &&
9187 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9188 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9189 (Mask[3] == -1 || Mask[3] >= 6)) {
9190 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9191 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9192 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9193 DAG.getConstant(SHUFPDMask, MVT::i8));
9195 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9196 (Mask[1] == -1 || Mask[1] < 2) &&
9197 (Mask[2] == -1 || Mask[2] >= 6) &&
9198 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9199 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9200 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9201 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9202 DAG.getConstant(SHUFPDMask, MVT::i8));
9205 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9206 // shuffle. However, if we have AVX2 and either inputs are already in place,
9207 // we will be able to shuffle even across lanes the other input in a single
9208 // instruction so skip this pattern.
9209 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9210 isShuffleMaskInputInPlace(1, Mask))))
9211 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9212 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9215 // If we have AVX2 then we always want to lower with a blend because an v4 we
9216 // can fully permute the elements.
9217 if (Subtarget->hasAVX2())
9218 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9221 // Otherwise fall back on generic lowering.
9222 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9225 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9227 /// This routine is only called when we have AVX2 and thus a reasonable
9228 /// instruction set for v4i64 shuffling..
9229 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9230 const X86Subtarget *Subtarget,
9231 SelectionDAG &DAG) {
9233 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9234 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9235 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9236 ArrayRef<int> Mask = SVOp->getMask();
9237 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9238 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9240 SmallVector<int, 4> WidenedMask;
9241 if (canWidenShuffleElements(Mask, WidenedMask))
9242 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9245 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9249 // Check for being able to broadcast a single element.
9250 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1,
9251 Mask, Subtarget, DAG))
9254 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9255 // use lower latency instructions that will operate on both 128-bit lanes.
9256 SmallVector<int, 2> RepeatedMask;
9257 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9258 if (isSingleInputShuffleMask(Mask)) {
9259 int PSHUFDMask[] = {-1, -1, -1, -1};
9260 for (int i = 0; i < 2; ++i)
9261 if (RepeatedMask[i] >= 0) {
9262 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9263 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9266 ISD::BITCAST, DL, MVT::v4i64,
9267 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9268 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9269 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9273 // AVX2 provides a direct instruction for permuting a single input across
9275 if (isSingleInputShuffleMask(Mask))
9276 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9277 getV4X86ShuffleImm8ForMask(Mask, DAG));
9279 // Try to use shift instructions.
9281 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9284 // Use dedicated unpack instructions for masks that match their pattern.
9285 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9286 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9287 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9288 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9289 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9290 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9291 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9292 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9294 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9295 // shuffle. However, if we have AVX2 and either inputs are already in place,
9296 // we will be able to shuffle even across lanes the other input in a single
9297 // instruction so skip this pattern.
9298 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9299 isShuffleMaskInputInPlace(1, Mask))))
9300 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9301 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9304 // Otherwise fall back on generic blend lowering.
9305 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9309 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9311 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9312 /// isn't available.
9313 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9314 const X86Subtarget *Subtarget,
9315 SelectionDAG &DAG) {
9317 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9318 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9319 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9320 ArrayRef<int> Mask = SVOp->getMask();
9321 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9323 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9327 // Check for being able to broadcast a single element.
9328 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1,
9329 Mask, Subtarget, DAG))
9332 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9333 // options to efficiently lower the shuffle.
9334 SmallVector<int, 4> RepeatedMask;
9335 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9336 assert(RepeatedMask.size() == 4 &&
9337 "Repeated masks must be half the mask width!");
9339 // Use even/odd duplicate instructions for masks that match their pattern.
9340 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9341 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9342 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9343 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9345 if (isSingleInputShuffleMask(Mask))
9346 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9347 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9349 // Use dedicated unpack instructions for masks that match their pattern.
9350 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9351 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9352 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9353 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9354 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9355 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9356 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9357 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9359 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9360 // have already handled any direct blends. We also need to squash the
9361 // repeated mask into a simulated v4f32 mask.
9362 for (int i = 0; i < 4; ++i)
9363 if (RepeatedMask[i] >= 8)
9364 RepeatedMask[i] -= 4;
9365 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9368 // If we have a single input shuffle with different shuffle patterns in the
9369 // two 128-bit lanes use the variable mask to VPERMILPS.
9370 if (isSingleInputShuffleMask(Mask)) {
9371 SDValue VPermMask[8];
9372 for (int i = 0; i < 8; ++i)
9373 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9374 : DAG.getConstant(Mask[i], MVT::i32);
9375 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9377 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9378 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9380 if (Subtarget->hasAVX2())
9381 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9382 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9383 DAG.getNode(ISD::BUILD_VECTOR, DL,
9384 MVT::v8i32, VPermMask)),
9387 // Otherwise, fall back.
9388 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9392 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9394 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9395 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9398 // If we have AVX2 then we always want to lower with a blend because at v8 we
9399 // can fully permute the elements.
9400 if (Subtarget->hasAVX2())
9401 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9404 // Otherwise fall back on generic lowering.
9405 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9408 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9410 /// This routine is only called when we have AVX2 and thus a reasonable
9411 /// instruction set for v8i32 shuffling..
9412 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9413 const X86Subtarget *Subtarget,
9414 SelectionDAG &DAG) {
9416 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9417 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9418 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9419 ArrayRef<int> Mask = SVOp->getMask();
9420 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9421 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9423 // Whenever we can lower this as a zext, that instruction is strictly faster
9424 // than any alternative. It also allows us to fold memory operands into the
9425 // shuffle in many cases.
9426 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9427 Mask, Subtarget, DAG))
9430 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9434 // Check for being able to broadcast a single element.
9435 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1,
9436 Mask, Subtarget, DAG))
9439 // If the shuffle mask is repeated in each 128-bit lane we can use more
9440 // efficient instructions that mirror the shuffles across the two 128-bit
9442 SmallVector<int, 4> RepeatedMask;
9443 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9444 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9445 if (isSingleInputShuffleMask(Mask))
9446 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9447 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9449 // Use dedicated unpack instructions for masks that match their pattern.
9450 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9451 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9452 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9453 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9454 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9455 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9456 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9457 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9460 // Try to use shift instructions.
9462 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9465 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9466 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9469 // If the shuffle patterns aren't repeated but it is a single input, directly
9470 // generate a cross-lane VPERMD instruction.
9471 if (isSingleInputShuffleMask(Mask)) {
9472 SDValue VPermMask[8];
9473 for (int i = 0; i < 8; ++i)
9474 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9475 : DAG.getConstant(Mask[i], MVT::i32);
9477 X86ISD::VPERMV, DL, MVT::v8i32,
9478 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9481 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9483 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9484 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9487 // Otherwise fall back on generic blend lowering.
9488 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9492 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9494 /// This routine is only called when we have AVX2 and thus a reasonable
9495 /// instruction set for v16i16 shuffling..
9496 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9497 const X86Subtarget *Subtarget,
9498 SelectionDAG &DAG) {
9500 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9501 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9502 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9503 ArrayRef<int> Mask = SVOp->getMask();
9504 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9505 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9507 // Whenever we can lower this as a zext, that instruction is strictly faster
9508 // than any alternative. It also allows us to fold memory operands into the
9509 // shuffle in many cases.
9510 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9511 Mask, Subtarget, DAG))
9514 // Check for being able to broadcast a single element.
9515 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1,
9516 Mask, Subtarget, DAG))
9519 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9523 // Use dedicated unpack instructions for masks that match their pattern.
9524 if (isShuffleEquivalent(V1, V2, Mask,
9525 {// First 128-bit lane:
9526 0, 16, 1, 17, 2, 18, 3, 19,
9527 // Second 128-bit lane:
9528 8, 24, 9, 25, 10, 26, 11, 27}))
9529 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9530 if (isShuffleEquivalent(V1, V2, Mask,
9531 {// First 128-bit lane:
9532 4, 20, 5, 21, 6, 22, 7, 23,
9533 // Second 128-bit lane:
9534 12, 28, 13, 29, 14, 30, 15, 31}))
9535 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9537 // Try to use shift instructions.
9539 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9542 // Try to use byte rotation instructions.
9543 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9544 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9547 if (isSingleInputShuffleMask(Mask)) {
9548 // There are no generalized cross-lane shuffle operations available on i16
9550 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9551 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9554 SDValue PSHUFBMask[32];
9555 for (int i = 0; i < 16; ++i) {
9556 if (Mask[i] == -1) {
9557 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9561 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9562 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9563 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
9564 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
9567 ISD::BITCAST, DL, MVT::v16i16,
9569 X86ISD::PSHUFB, DL, MVT::v32i8,
9570 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9571 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9574 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9576 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9577 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9580 // Otherwise fall back on generic lowering.
9581 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9584 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9586 /// This routine is only called when we have AVX2 and thus a reasonable
9587 /// instruction set for v32i8 shuffling..
9588 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9589 const X86Subtarget *Subtarget,
9590 SelectionDAG &DAG) {
9592 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9593 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9594 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9595 ArrayRef<int> Mask = SVOp->getMask();
9596 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9597 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9599 // Whenever we can lower this as a zext, that instruction is strictly faster
9600 // than any alternative. It also allows us to fold memory operands into the
9601 // shuffle in many cases.
9602 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9603 Mask, Subtarget, DAG))
9606 // Check for being able to broadcast a single element.
9607 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1,
9608 Mask, Subtarget, DAG))
9611 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9615 // Use dedicated unpack instructions for masks that match their pattern.
9616 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9618 if (isShuffleEquivalent(
9620 {// First 128-bit lane:
9621 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9622 // Second 128-bit lane:
9623 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9624 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9625 if (isShuffleEquivalent(
9627 {// First 128-bit lane:
9628 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9629 // Second 128-bit lane:
9630 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9631 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9633 // Try to use shift instructions.
9635 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9638 // Try to use byte rotation instructions.
9639 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9640 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9643 if (isSingleInputShuffleMask(Mask)) {
9644 // There are no generalized cross-lane shuffle operations available on i8
9646 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9647 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9650 SDValue PSHUFBMask[32];
9651 for (int i = 0; i < 32; ++i)
9654 ? DAG.getUNDEF(MVT::i8)
9655 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
9658 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9659 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9662 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9664 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9665 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9668 // Otherwise fall back on generic lowering.
9669 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9672 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9674 /// This routine either breaks down the specific type of a 256-bit x86 vector
9675 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9676 /// together based on the available instructions.
9677 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9678 MVT VT, const X86Subtarget *Subtarget,
9679 SelectionDAG &DAG) {
9681 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9682 ArrayRef<int> Mask = SVOp->getMask();
9684 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9685 // check for those subtargets here and avoid much of the subtarget querying in
9686 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9687 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9688 // floating point types there eventually, just immediately cast everything to
9689 // a float and operate entirely in that domain.
9690 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9691 int ElementBits = VT.getScalarSizeInBits();
9692 if (ElementBits < 32)
9693 // No floating point type available, decompose into 128-bit vectors.
9694 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9696 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9697 VT.getVectorNumElements());
9698 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9699 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9700 return DAG.getNode(ISD::BITCAST, DL, VT,
9701 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9704 switch (VT.SimpleTy) {
9706 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9708 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9710 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9712 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9714 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9716 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9719 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9723 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
9724 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9725 const X86Subtarget *Subtarget,
9726 SelectionDAG &DAG) {
9728 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9729 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9730 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9731 ArrayRef<int> Mask = SVOp->getMask();
9732 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9734 // X86 has dedicated unpack instructions that can handle specific blend
9735 // operations: UNPCKH and UNPCKL.
9736 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9737 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
9738 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9739 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
9741 // FIXME: Implement direct support for this type!
9742 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
9745 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
9746 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9747 const X86Subtarget *Subtarget,
9748 SelectionDAG &DAG) {
9750 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9751 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9752 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9753 ArrayRef<int> Mask = SVOp->getMask();
9754 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9756 // Use dedicated unpack instructions for masks that match their pattern.
9757 if (isShuffleEquivalent(V1, V2, Mask,
9758 {// First 128-bit lane.
9759 0, 16, 1, 17, 4, 20, 5, 21,
9760 // Second 128-bit lane.
9761 8, 24, 9, 25, 12, 28, 13, 29}))
9762 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
9763 if (isShuffleEquivalent(V1, V2, Mask,
9764 {// First 128-bit lane.
9765 2, 18, 3, 19, 6, 22, 7, 23,
9766 // Second 128-bit lane.
9767 10, 26, 11, 27, 14, 30, 15, 31}))
9768 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
9770 // FIXME: Implement direct support for this type!
9771 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
9774 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
9775 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9776 const X86Subtarget *Subtarget,
9777 SelectionDAG &DAG) {
9779 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9780 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9781 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9782 ArrayRef<int> Mask = SVOp->getMask();
9783 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9785 // X86 has dedicated unpack instructions that can handle specific blend
9786 // operations: UNPCKH and UNPCKL.
9787 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9788 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
9789 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9790 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
9792 // FIXME: Implement direct support for this type!
9793 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
9796 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
9797 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9798 const X86Subtarget *Subtarget,
9799 SelectionDAG &DAG) {
9801 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9802 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9803 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9804 ArrayRef<int> Mask = SVOp->getMask();
9805 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9807 // Use dedicated unpack instructions for masks that match their pattern.
9808 if (isShuffleEquivalent(V1, V2, Mask,
9809 {// First 128-bit lane.
9810 0, 16, 1, 17, 4, 20, 5, 21,
9811 // Second 128-bit lane.
9812 8, 24, 9, 25, 12, 28, 13, 29}))
9813 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
9814 if (isShuffleEquivalent(V1, V2, Mask,
9815 {// First 128-bit lane.
9816 2, 18, 3, 19, 6, 22, 7, 23,
9817 // Second 128-bit lane.
9818 10, 26, 11, 27, 14, 30, 15, 31}))
9819 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
9821 // FIXME: Implement direct support for this type!
9822 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
9825 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
9826 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9827 const X86Subtarget *Subtarget,
9828 SelectionDAG &DAG) {
9830 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9831 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9832 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9833 ArrayRef<int> Mask = SVOp->getMask();
9834 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9835 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
9837 // FIXME: Implement direct support for this type!
9838 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
9841 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
9842 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9843 const X86Subtarget *Subtarget,
9844 SelectionDAG &DAG) {
9846 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9847 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9848 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9849 ArrayRef<int> Mask = SVOp->getMask();
9850 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
9851 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
9853 // FIXME: Implement direct support for this type!
9854 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
9857 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
9859 /// This routine either breaks down the specific type of a 512-bit x86 vector
9860 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
9861 /// together based on the available instructions.
9862 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9863 MVT VT, const X86Subtarget *Subtarget,
9864 SelectionDAG &DAG) {
9866 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9867 ArrayRef<int> Mask = SVOp->getMask();
9868 assert(Subtarget->hasAVX512() &&
9869 "Cannot lower 512-bit vectors w/ basic ISA!");
9871 // Check for being able to broadcast a single element.
9872 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
9873 Mask, Subtarget, DAG))
9876 // Dispatch to each element type for lowering. If we don't have supprot for
9877 // specific element type shuffles at 512 bits, immediately split them and
9878 // lower them. Each lowering routine of a given type is allowed to assume that
9879 // the requisite ISA extensions for that element type are available.
9880 switch (VT.SimpleTy) {
9882 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9884 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9886 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9888 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9890 if (Subtarget->hasBWI())
9891 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9894 if (Subtarget->hasBWI())
9895 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9899 llvm_unreachable("Not a valid 512-bit x86 vector type!");
9902 // Otherwise fall back on splitting.
9903 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9906 /// \brief Top-level lowering for x86 vector shuffles.
9908 /// This handles decomposition, canonicalization, and lowering of all x86
9909 /// vector shuffles. Most of the specific lowering strategies are encapsulated
9910 /// above in helper routines. The canonicalization attempts to widen shuffles
9911 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
9912 /// s.t. only one of the two inputs needs to be tested, etc.
9913 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
9914 SelectionDAG &DAG) {
9915 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9916 ArrayRef<int> Mask = SVOp->getMask();
9917 SDValue V1 = Op.getOperand(0);
9918 SDValue V2 = Op.getOperand(1);
9919 MVT VT = Op.getSimpleValueType();
9920 int NumElements = VT.getVectorNumElements();
9923 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
9925 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
9926 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
9927 if (V1IsUndef && V2IsUndef)
9928 return DAG.getUNDEF(VT);
9930 // When we create a shuffle node we put the UNDEF node to second operand,
9931 // but in some cases the first operand may be transformed to UNDEF.
9932 // In this case we should just commute the node.
9934 return DAG.getCommutedVectorShuffle(*SVOp);
9936 // Check for non-undef masks pointing at an undef vector and make the masks
9937 // undef as well. This makes it easier to match the shuffle based solely on
9941 if (M >= NumElements) {
9942 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
9943 for (int &M : NewMask)
9944 if (M >= NumElements)
9946 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
9949 // We actually see shuffles that are entirely re-arrangements of a set of
9950 // zero inputs. This mostly happens while decomposing complex shuffles into
9951 // simple ones. Directly lower these as a buildvector of zeros.
9952 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9954 return getZeroVector(VT, Subtarget, DAG, dl);
9956 // Try to collapse shuffles into using a vector type with fewer elements but
9957 // wider element types. We cap this to not form integers or floating point
9958 // elements wider than 64 bits, but it might be interesting to form i128
9959 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
9960 SmallVector<int, 16> WidenedMask;
9961 if (VT.getScalarSizeInBits() < 64 &&
9962 canWidenShuffleElements(Mask, WidenedMask)) {
9963 MVT NewEltVT = VT.isFloatingPoint()
9964 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
9965 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
9966 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
9967 // Make sure that the new vector type is legal. For example, v2f64 isn't
9969 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
9970 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
9971 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
9972 return DAG.getNode(ISD::BITCAST, dl, VT,
9973 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
9977 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
9978 for (int M : SVOp->getMask())
9981 else if (M < NumElements)
9986 // Commute the shuffle as needed such that more elements come from V1 than
9987 // V2. This allows us to match the shuffle pattern strictly on how many
9988 // elements come from V1 without handling the symmetric cases.
9989 if (NumV2Elements > NumV1Elements)
9990 return DAG.getCommutedVectorShuffle(*SVOp);
9992 // When the number of V1 and V2 elements are the same, try to minimize the
9993 // number of uses of V2 in the low half of the vector. When that is tied,
9994 // ensure that the sum of indices for V1 is equal to or lower than the sum
9995 // indices for V2. When those are equal, try to ensure that the number of odd
9996 // indices for V1 is lower than the number of odd indices for V2.
9997 if (NumV1Elements == NumV2Elements) {
9998 int LowV1Elements = 0, LowV2Elements = 0;
9999 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10000 if (M >= NumElements)
10004 if (LowV2Elements > LowV1Elements) {
10005 return DAG.getCommutedVectorShuffle(*SVOp);
10006 } else if (LowV2Elements == LowV1Elements) {
10007 int SumV1Indices = 0, SumV2Indices = 0;
10008 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10009 if (SVOp->getMask()[i] >= NumElements)
10011 else if (SVOp->getMask()[i] >= 0)
10013 if (SumV2Indices < SumV1Indices) {
10014 return DAG.getCommutedVectorShuffle(*SVOp);
10015 } else if (SumV2Indices == SumV1Indices) {
10016 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10017 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10018 if (SVOp->getMask()[i] >= NumElements)
10019 NumV2OddIndices += i % 2;
10020 else if (SVOp->getMask()[i] >= 0)
10021 NumV1OddIndices += i % 2;
10022 if (NumV2OddIndices < NumV1OddIndices)
10023 return DAG.getCommutedVectorShuffle(*SVOp);
10028 // For each vector width, delegate to a specialized lowering routine.
10029 if (VT.getSizeInBits() == 128)
10030 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10032 if (VT.getSizeInBits() == 256)
10033 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10035 // Force AVX-512 vectors to be scalarized for now.
10036 // FIXME: Implement AVX-512 support!
10037 if (VT.getSizeInBits() == 512)
10038 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10040 llvm_unreachable("Unimplemented!");
10043 // This function assumes its argument is a BUILD_VECTOR of constants or
10044 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10046 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10047 unsigned &MaskValue) {
10049 unsigned NumElems = BuildVector->getNumOperands();
10050 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10051 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10052 unsigned NumElemsInLane = NumElems / NumLanes;
10054 // Blend for v16i16 should be symetric for the both lanes.
10055 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10056 SDValue EltCond = BuildVector->getOperand(i);
10057 SDValue SndLaneEltCond =
10058 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10060 int Lane1Cond = -1, Lane2Cond = -1;
10061 if (isa<ConstantSDNode>(EltCond))
10062 Lane1Cond = !isZero(EltCond);
10063 if (isa<ConstantSDNode>(SndLaneEltCond))
10064 Lane2Cond = !isZero(SndLaneEltCond);
10066 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10067 // Lane1Cond != 0, means we want the first argument.
10068 // Lane1Cond == 0, means we want the second argument.
10069 // The encoding of this argument is 0 for the first argument, 1
10070 // for the second. Therefore, invert the condition.
10071 MaskValue |= !Lane1Cond << i;
10072 else if (Lane1Cond < 0)
10073 MaskValue |= !Lane2Cond << i;
10080 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10081 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10082 const X86Subtarget *Subtarget,
10083 SelectionDAG &DAG) {
10084 SDValue Cond = Op.getOperand(0);
10085 SDValue LHS = Op.getOperand(1);
10086 SDValue RHS = Op.getOperand(2);
10088 MVT VT = Op.getSimpleValueType();
10090 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10092 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10094 // Only non-legal VSELECTs reach this lowering, convert those into generic
10095 // shuffles and re-use the shuffle lowering path for blends.
10096 SmallVector<int, 32> Mask;
10097 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10098 SDValue CondElt = CondBV->getOperand(i);
10100 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10102 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10105 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10106 // A vselect where all conditions and data are constants can be optimized into
10107 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10108 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10109 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10110 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10113 // Try to lower this to a blend-style vector shuffle. This can handle all
10114 // constant condition cases.
10115 SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG);
10116 if (BlendOp.getNode())
10119 // Variable blends are only legal from SSE4.1 onward.
10120 if (!Subtarget->hasSSE41())
10123 // Some types for vselect were previously set to Expand, not Legal or
10124 // Custom. Return an empty SDValue so we fall-through to Expand, after
10125 // the Custom lowering phase.
10126 MVT VT = Op.getSimpleValueType();
10127 switch (VT.SimpleTy) {
10132 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10137 // We couldn't create a "Blend with immediate" node.
10138 // This node should still be legal, but we'll have to emit a blendv*
10143 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10144 MVT VT = Op.getSimpleValueType();
10147 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10150 if (VT.getSizeInBits() == 8) {
10151 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10152 Op.getOperand(0), Op.getOperand(1));
10153 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10154 DAG.getValueType(VT));
10155 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10158 if (VT.getSizeInBits() == 16) {
10159 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10160 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10162 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10163 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10164 DAG.getNode(ISD::BITCAST, dl,
10167 Op.getOperand(1)));
10168 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10169 Op.getOperand(0), Op.getOperand(1));
10170 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10171 DAG.getValueType(VT));
10172 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10175 if (VT == MVT::f32) {
10176 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10177 // the result back to FR32 register. It's only worth matching if the
10178 // result has a single use which is a store or a bitcast to i32. And in
10179 // the case of a store, it's not worth it if the index is a constant 0,
10180 // because a MOVSSmr can be used instead, which is smaller and faster.
10181 if (!Op.hasOneUse())
10183 SDNode *User = *Op.getNode()->use_begin();
10184 if ((User->getOpcode() != ISD::STORE ||
10185 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10186 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10187 (User->getOpcode() != ISD::BITCAST ||
10188 User->getValueType(0) != MVT::i32))
10190 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10191 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10194 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10197 if (VT == MVT::i32 || VT == MVT::i64) {
10198 // ExtractPS/pextrq works with constant index.
10199 if (isa<ConstantSDNode>(Op.getOperand(1)))
10205 /// Extract one bit from mask vector, like v16i1 or v8i1.
10206 /// AVX-512 feature.
10208 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10209 SDValue Vec = Op.getOperand(0);
10211 MVT VecVT = Vec.getSimpleValueType();
10212 SDValue Idx = Op.getOperand(1);
10213 MVT EltVT = Op.getSimpleValueType();
10215 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10216 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10217 "Unexpected vector type in ExtractBitFromMaskVector");
10219 // variable index can't be handled in mask registers,
10220 // extend vector to VR512
10221 if (!isa<ConstantSDNode>(Idx)) {
10222 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10223 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10224 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10225 ExtVT.getVectorElementType(), Ext, Idx);
10226 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10229 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10230 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10231 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10232 rc = getRegClassFor(MVT::v16i1);
10233 unsigned MaxSift = rc->getSize()*8 - 1;
10234 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10235 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10236 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10237 DAG.getConstant(MaxSift, MVT::i8));
10238 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10239 DAG.getIntPtrConstant(0));
10243 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10244 SelectionDAG &DAG) const {
10246 SDValue Vec = Op.getOperand(0);
10247 MVT VecVT = Vec.getSimpleValueType();
10248 SDValue Idx = Op.getOperand(1);
10250 if (Op.getSimpleValueType() == MVT::i1)
10251 return ExtractBitFromMaskVector(Op, DAG);
10253 if (!isa<ConstantSDNode>(Idx)) {
10254 if (VecVT.is512BitVector() ||
10255 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10256 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10259 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10260 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10261 MaskEltVT.getSizeInBits());
10263 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10264 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10265 getZeroVector(MaskVT, Subtarget, DAG, dl),
10266 Idx, DAG.getConstant(0, getPointerTy()));
10267 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10268 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10269 Perm, DAG.getConstant(0, getPointerTy()));
10274 // If this is a 256-bit vector result, first extract the 128-bit vector and
10275 // then extract the element from the 128-bit vector.
10276 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10278 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10279 // Get the 128-bit vector.
10280 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10281 MVT EltVT = VecVT.getVectorElementType();
10283 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10285 //if (IdxVal >= NumElems/2)
10286 // IdxVal -= NumElems/2;
10287 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10288 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10289 DAG.getConstant(IdxVal, MVT::i32));
10292 assert(VecVT.is128BitVector() && "Unexpected vector length");
10294 if (Subtarget->hasSSE41()) {
10295 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10300 MVT VT = Op.getSimpleValueType();
10301 // TODO: handle v16i8.
10302 if (VT.getSizeInBits() == 16) {
10303 SDValue Vec = Op.getOperand(0);
10304 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10306 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10307 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10308 DAG.getNode(ISD::BITCAST, dl,
10310 Op.getOperand(1)));
10311 // Transform it so it match pextrw which produces a 32-bit result.
10312 MVT EltVT = MVT::i32;
10313 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10314 Op.getOperand(0), Op.getOperand(1));
10315 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10316 DAG.getValueType(VT));
10317 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10320 if (VT.getSizeInBits() == 32) {
10321 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10325 // SHUFPS the element to the lowest double word, then movss.
10326 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10327 MVT VVT = Op.getOperand(0).getSimpleValueType();
10328 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10329 DAG.getUNDEF(VVT), Mask);
10330 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10331 DAG.getIntPtrConstant(0));
10334 if (VT.getSizeInBits() == 64) {
10335 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10336 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10337 // to match extract_elt for f64.
10338 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10342 // UNPCKHPD the element to the lowest double word, then movsd.
10343 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10344 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10345 int Mask[2] = { 1, -1 };
10346 MVT VVT = Op.getOperand(0).getSimpleValueType();
10347 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10348 DAG.getUNDEF(VVT), Mask);
10349 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10350 DAG.getIntPtrConstant(0));
10356 /// Insert one bit to mask vector, like v16i1 or v8i1.
10357 /// AVX-512 feature.
10359 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10361 SDValue Vec = Op.getOperand(0);
10362 SDValue Elt = Op.getOperand(1);
10363 SDValue Idx = Op.getOperand(2);
10364 MVT VecVT = Vec.getSimpleValueType();
10366 if (!isa<ConstantSDNode>(Idx)) {
10367 // Non constant index. Extend source and destination,
10368 // insert element and then truncate the result.
10369 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10370 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10371 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10372 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10373 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10374 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10377 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10378 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10379 if (Vec.getOpcode() == ISD::UNDEF)
10380 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10381 DAG.getConstant(IdxVal, MVT::i8));
10382 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10383 unsigned MaxSift = rc->getSize()*8 - 1;
10384 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10385 DAG.getConstant(MaxSift, MVT::i8));
10386 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10387 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10388 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10391 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10392 SelectionDAG &DAG) const {
10393 MVT VT = Op.getSimpleValueType();
10394 MVT EltVT = VT.getVectorElementType();
10396 if (EltVT == MVT::i1)
10397 return InsertBitToMaskVector(Op, DAG);
10400 SDValue N0 = Op.getOperand(0);
10401 SDValue N1 = Op.getOperand(1);
10402 SDValue N2 = Op.getOperand(2);
10403 if (!isa<ConstantSDNode>(N2))
10405 auto *N2C = cast<ConstantSDNode>(N2);
10406 unsigned IdxVal = N2C->getZExtValue();
10408 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10409 // into that, and then insert the subvector back into the result.
10410 if (VT.is256BitVector() || VT.is512BitVector()) {
10411 // Get the desired 128-bit vector half.
10412 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10414 // Insert the element into the desired half.
10415 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10416 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10418 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10419 DAG.getConstant(IdxIn128, MVT::i32));
10421 // Insert the changed part back to the 256-bit vector
10422 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10424 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10426 if (Subtarget->hasSSE41()) {
10427 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10429 if (VT == MVT::v8i16) {
10430 Opc = X86ISD::PINSRW;
10432 assert(VT == MVT::v16i8);
10433 Opc = X86ISD::PINSRB;
10436 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10438 if (N1.getValueType() != MVT::i32)
10439 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10440 if (N2.getValueType() != MVT::i32)
10441 N2 = DAG.getIntPtrConstant(IdxVal);
10442 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10445 if (EltVT == MVT::f32) {
10446 // Bits [7:6] of the constant are the source select. This will always be
10447 // zero here. The DAG Combiner may combine an extract_elt index into
10449 // bits. For example (insert (extract, 3), 2) could be matched by
10451 // the '3' into bits [7:6] of X86ISD::INSERTPS.
10452 // Bits [5:4] of the constant are the destination select. This is the
10453 // value of the incoming immediate.
10454 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10455 // combine either bitwise AND or insert of float 0.0 to set these bits.
10456 N2 = DAG.getIntPtrConstant(IdxVal << 4);
10457 // Create this as a scalar to vector..
10458 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10459 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10462 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10463 // PINSR* works with constant index.
10468 if (EltVT == MVT::i8)
10471 if (EltVT.getSizeInBits() == 16) {
10472 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10473 // as its second argument.
10474 if (N1.getValueType() != MVT::i32)
10475 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10476 if (N2.getValueType() != MVT::i32)
10477 N2 = DAG.getIntPtrConstant(IdxVal);
10478 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10483 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10485 MVT OpVT = Op.getSimpleValueType();
10487 // If this is a 256-bit vector result, first insert into a 128-bit
10488 // vector and then insert into the 256-bit vector.
10489 if (!OpVT.is128BitVector()) {
10490 // Insert into a 128-bit vector.
10491 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10492 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10493 OpVT.getVectorNumElements() / SizeFactor);
10495 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10497 // Insert the 128-bit vector.
10498 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10501 if (OpVT == MVT::v1i64 &&
10502 Op.getOperand(0).getValueType() == MVT::i64)
10503 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10505 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10506 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10507 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10508 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10511 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10512 // a simple subregister reference or explicit instructions to grab
10513 // upper bits of a vector.
10514 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10515 SelectionDAG &DAG) {
10517 SDValue In = Op.getOperand(0);
10518 SDValue Idx = Op.getOperand(1);
10519 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10520 MVT ResVT = Op.getSimpleValueType();
10521 MVT InVT = In.getSimpleValueType();
10523 if (Subtarget->hasFp256()) {
10524 if (ResVT.is128BitVector() &&
10525 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10526 isa<ConstantSDNode>(Idx)) {
10527 return Extract128BitVector(In, IdxVal, DAG, dl);
10529 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10530 isa<ConstantSDNode>(Idx)) {
10531 return Extract256BitVector(In, IdxVal, DAG, dl);
10537 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10538 // simple superregister reference or explicit instructions to insert
10539 // the upper bits of a vector.
10540 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10541 SelectionDAG &DAG) {
10542 if (!Subtarget->hasAVX())
10546 SDValue Vec = Op.getOperand(0);
10547 SDValue SubVec = Op.getOperand(1);
10548 SDValue Idx = Op.getOperand(2);
10550 if (!isa<ConstantSDNode>(Idx))
10553 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10554 MVT OpVT = Op.getSimpleValueType();
10555 MVT SubVecVT = SubVec.getSimpleValueType();
10557 // Fold two 16-byte subvector loads into one 32-byte load:
10558 // (insert_subvector (insert_subvector undef, (load addr), 0),
10559 // (load addr + 16), Elts/2)
10561 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10562 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10563 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10564 !Subtarget->isUnalignedMem32Slow()) {
10565 SDValue SubVec2 = Vec.getOperand(1);
10566 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10567 if (Idx2->getZExtValue() == 0) {
10568 SDValue Ops[] = { SubVec2, SubVec };
10569 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10576 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10577 SubVecVT.is128BitVector())
10578 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10580 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10581 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10586 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10587 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10588 // one of the above mentioned nodes. It has to be wrapped because otherwise
10589 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10590 // be used to form addressing mode. These wrapped nodes will be selected
10593 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10594 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10596 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10597 // global base reg.
10598 unsigned char OpFlag = 0;
10599 unsigned WrapperKind = X86ISD::Wrapper;
10600 CodeModel::Model M = DAG.getTarget().getCodeModel();
10602 if (Subtarget->isPICStyleRIPRel() &&
10603 (M == CodeModel::Small || M == CodeModel::Kernel))
10604 WrapperKind = X86ISD::WrapperRIP;
10605 else if (Subtarget->isPICStyleGOT())
10606 OpFlag = X86II::MO_GOTOFF;
10607 else if (Subtarget->isPICStyleStubPIC())
10608 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10610 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10611 CP->getAlignment(),
10612 CP->getOffset(), OpFlag);
10614 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10615 // With PIC, the address is actually $g + Offset.
10617 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10618 DAG.getNode(X86ISD::GlobalBaseReg,
10619 SDLoc(), getPointerTy()),
10626 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10627 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10629 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10630 // global base reg.
10631 unsigned char OpFlag = 0;
10632 unsigned WrapperKind = X86ISD::Wrapper;
10633 CodeModel::Model M = DAG.getTarget().getCodeModel();
10635 if (Subtarget->isPICStyleRIPRel() &&
10636 (M == CodeModel::Small || M == CodeModel::Kernel))
10637 WrapperKind = X86ISD::WrapperRIP;
10638 else if (Subtarget->isPICStyleGOT())
10639 OpFlag = X86II::MO_GOTOFF;
10640 else if (Subtarget->isPICStyleStubPIC())
10641 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10643 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10646 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10648 // With PIC, the address is actually $g + Offset.
10650 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10651 DAG.getNode(X86ISD::GlobalBaseReg,
10652 SDLoc(), getPointerTy()),
10659 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10660 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10662 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10663 // global base reg.
10664 unsigned char OpFlag = 0;
10665 unsigned WrapperKind = X86ISD::Wrapper;
10666 CodeModel::Model M = DAG.getTarget().getCodeModel();
10668 if (Subtarget->isPICStyleRIPRel() &&
10669 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10670 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10671 OpFlag = X86II::MO_GOTPCREL;
10672 WrapperKind = X86ISD::WrapperRIP;
10673 } else if (Subtarget->isPICStyleGOT()) {
10674 OpFlag = X86II::MO_GOT;
10675 } else if (Subtarget->isPICStyleStubPIC()) {
10676 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10677 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10678 OpFlag = X86II::MO_DARWIN_NONLAZY;
10681 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
10684 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10686 // With PIC, the address is actually $g + Offset.
10687 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
10688 !Subtarget->is64Bit()) {
10689 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10690 DAG.getNode(X86ISD::GlobalBaseReg,
10691 SDLoc(), getPointerTy()),
10695 // For symbols that require a load from a stub to get the address, emit the
10697 if (isGlobalStubReference(OpFlag))
10698 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
10699 MachinePointerInfo::getGOT(), false, false, false, 0);
10705 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
10706 // Create the TargetBlockAddressAddress node.
10707 unsigned char OpFlags =
10708 Subtarget->ClassifyBlockAddressReference();
10709 CodeModel::Model M = DAG.getTarget().getCodeModel();
10710 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
10711 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
10713 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
10716 if (Subtarget->isPICStyleRIPRel() &&
10717 (M == CodeModel::Small || M == CodeModel::Kernel))
10718 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10720 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10722 // With PIC, the address is actually $g + Offset.
10723 if (isGlobalRelativeToPICBase(OpFlags)) {
10724 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10725 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10733 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
10734 int64_t Offset, SelectionDAG &DAG) const {
10735 // Create the TargetGlobalAddress node, folding in the constant
10736 // offset if it is legal.
10737 unsigned char OpFlags =
10738 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
10739 CodeModel::Model M = DAG.getTarget().getCodeModel();
10741 if (OpFlags == X86II::MO_NO_FLAG &&
10742 X86::isOffsetSuitableForCodeModel(Offset, M)) {
10743 // A direct static reference to a global.
10744 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
10747 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
10750 if (Subtarget->isPICStyleRIPRel() &&
10751 (M == CodeModel::Small || M == CodeModel::Kernel))
10752 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10754 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10756 // With PIC, the address is actually $g + Offset.
10757 if (isGlobalRelativeToPICBase(OpFlags)) {
10758 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10759 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10763 // For globals that require a load from a stub to get the address, emit the
10765 if (isGlobalStubReference(OpFlags))
10766 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
10767 MachinePointerInfo::getGOT(), false, false, false, 0);
10769 // If there was a non-zero offset that we didn't fold, create an explicit
10770 // addition for it.
10772 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
10773 DAG.getConstant(Offset, getPointerTy()));
10779 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
10780 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
10781 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
10782 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
10786 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
10787 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
10788 unsigned char OperandFlags, bool LocalDynamic = false) {
10789 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10790 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10792 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10793 GA->getValueType(0),
10797 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
10801 SDValue Ops[] = { Chain, TGA, *InFlag };
10802 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10804 SDValue Ops[] = { Chain, TGA };
10805 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10808 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
10809 MFI->setAdjustsStack(true);
10810 MFI->setHasCalls(true);
10812 SDValue Flag = Chain.getValue(1);
10813 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
10816 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
10818 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10821 SDLoc dl(GA); // ? function entry point might be better
10822 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10823 DAG.getNode(X86ISD::GlobalBaseReg,
10824 SDLoc(), PtrVT), InFlag);
10825 InFlag = Chain.getValue(1);
10827 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
10830 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
10832 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10834 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
10835 X86::RAX, X86II::MO_TLSGD);
10838 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
10844 // Get the start address of the TLS block for this module.
10845 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
10846 .getInfo<X86MachineFunctionInfo>();
10847 MFI->incNumLocalDynamicTLSAccesses();
10851 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
10852 X86II::MO_TLSLD, /*LocalDynamic=*/true);
10855 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10856 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
10857 InFlag = Chain.getValue(1);
10858 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
10859 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
10862 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
10866 unsigned char OperandFlags = X86II::MO_DTPOFF;
10867 unsigned WrapperKind = X86ISD::Wrapper;
10868 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10869 GA->getValueType(0),
10870 GA->getOffset(), OperandFlags);
10871 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10873 // Add x@dtpoff with the base.
10874 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
10877 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
10878 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10879 const EVT PtrVT, TLSModel::Model model,
10880 bool is64Bit, bool isPIC) {
10883 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
10884 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
10885 is64Bit ? 257 : 256));
10887 SDValue ThreadPointer =
10888 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
10889 MachinePointerInfo(Ptr), false, false, false, 0);
10891 unsigned char OperandFlags = 0;
10892 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
10894 unsigned WrapperKind = X86ISD::Wrapper;
10895 if (model == TLSModel::LocalExec) {
10896 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
10897 } else if (model == TLSModel::InitialExec) {
10899 OperandFlags = X86II::MO_GOTTPOFF;
10900 WrapperKind = X86ISD::WrapperRIP;
10902 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
10905 llvm_unreachable("Unexpected model");
10908 // emit "addl x@ntpoff,%eax" (local exec)
10909 // or "addl x@indntpoff,%eax" (initial exec)
10910 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
10912 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
10913 GA->getOffset(), OperandFlags);
10914 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10916 if (model == TLSModel::InitialExec) {
10917 if (isPIC && !is64Bit) {
10918 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
10919 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
10923 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
10924 MachinePointerInfo::getGOT(), false, false, false, 0);
10927 // The address of the thread local variable is the add of the thread
10928 // pointer with the offset of the variable.
10929 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
10933 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
10935 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
10936 const GlobalValue *GV = GA->getGlobal();
10938 if (Subtarget->isTargetELF()) {
10939 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
10942 case TLSModel::GeneralDynamic:
10943 if (Subtarget->is64Bit())
10944 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
10945 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
10946 case TLSModel::LocalDynamic:
10947 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
10948 Subtarget->is64Bit());
10949 case TLSModel::InitialExec:
10950 case TLSModel::LocalExec:
10951 return LowerToTLSExecModel(
10952 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
10953 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
10955 llvm_unreachable("Unknown TLS model.");
10958 if (Subtarget->isTargetDarwin()) {
10959 // Darwin only has one model of TLS. Lower to that.
10960 unsigned char OpFlag = 0;
10961 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
10962 X86ISD::WrapperRIP : X86ISD::Wrapper;
10964 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10965 // global base reg.
10966 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
10967 !Subtarget->is64Bit();
10969 OpFlag = X86II::MO_TLVP_PIC_BASE;
10971 OpFlag = X86II::MO_TLVP;
10973 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
10974 GA->getValueType(0),
10975 GA->getOffset(), OpFlag);
10976 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10978 // With PIC32, the address is actually $g + Offset.
10980 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10981 DAG.getNode(X86ISD::GlobalBaseReg,
10982 SDLoc(), getPointerTy()),
10985 // Lowering the machine isd will make sure everything is in the right
10987 SDValue Chain = DAG.getEntryNode();
10988 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10989 SDValue Args[] = { Chain, Offset };
10990 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
10992 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
10993 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10994 MFI->setAdjustsStack(true);
10996 // And our return value (tls address) is in the standard call return value
10998 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10999 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11000 Chain.getValue(1));
11003 if (Subtarget->isTargetKnownWindowsMSVC() ||
11004 Subtarget->isTargetWindowsGNU()) {
11005 // Just use the implicit TLS architecture
11006 // Need to generate someting similar to:
11007 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11009 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11010 // mov rcx, qword [rdx+rcx*8]
11011 // mov eax, .tls$:tlsvar
11012 // [rax+rcx] contains the address
11013 // Windows 64bit: gs:0x58
11014 // Windows 32bit: fs:__tls_array
11017 SDValue Chain = DAG.getEntryNode();
11019 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11020 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11021 // use its literal value of 0x2C.
11022 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11023 ? Type::getInt8PtrTy(*DAG.getContext(),
11025 : Type::getInt32PtrTy(*DAG.getContext(),
11029 Subtarget->is64Bit()
11030 ? DAG.getIntPtrConstant(0x58)
11031 : (Subtarget->isTargetWindowsGNU()
11032 ? DAG.getIntPtrConstant(0x2C)
11033 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11035 SDValue ThreadPointer =
11036 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11037 MachinePointerInfo(Ptr), false, false, false, 0);
11039 // Load the _tls_index variable
11040 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11041 if (Subtarget->is64Bit())
11042 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
11043 IDX, MachinePointerInfo(), MVT::i32,
11044 false, false, false, 0);
11046 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11047 false, false, false, 0);
11049 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
11051 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11053 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11054 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11055 false, false, false, 0);
11057 // Get the offset of start of .tls section
11058 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11059 GA->getValueType(0),
11060 GA->getOffset(), X86II::MO_SECREL);
11061 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11063 // The address of the thread local variable is the add of the thread
11064 // pointer with the offset of the variable.
11065 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11068 llvm_unreachable("TLS not implemented for this target.");
11071 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11072 /// and take a 2 x i32 value to shift plus a shift amount.
11073 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11074 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11075 MVT VT = Op.getSimpleValueType();
11076 unsigned VTBits = VT.getSizeInBits();
11078 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11079 SDValue ShOpLo = Op.getOperand(0);
11080 SDValue ShOpHi = Op.getOperand(1);
11081 SDValue ShAmt = Op.getOperand(2);
11082 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11083 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11085 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11086 DAG.getConstant(VTBits - 1, MVT::i8));
11087 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11088 DAG.getConstant(VTBits - 1, MVT::i8))
11089 : DAG.getConstant(0, VT);
11091 SDValue Tmp2, Tmp3;
11092 if (Op.getOpcode() == ISD::SHL_PARTS) {
11093 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11094 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11096 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11097 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11100 // If the shift amount is larger or equal than the width of a part we can't
11101 // rely on the results of shld/shrd. Insert a test and select the appropriate
11102 // values for large shift amounts.
11103 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11104 DAG.getConstant(VTBits, MVT::i8));
11105 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11106 AndNode, DAG.getConstant(0, MVT::i8));
11109 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
11110 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11111 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11113 if (Op.getOpcode() == ISD::SHL_PARTS) {
11114 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11115 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11117 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11118 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11121 SDValue Ops[2] = { Lo, Hi };
11122 return DAG.getMergeValues(Ops, dl);
11125 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11126 SelectionDAG &DAG) const {
11127 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11130 if (SrcVT.isVector()) {
11131 if (SrcVT.getVectorElementType() == MVT::i1) {
11132 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11133 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11134 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11135 Op.getOperand(0)));
11140 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11141 "Unknown SINT_TO_FP to lower!");
11143 // These are really Legal; return the operand so the caller accepts it as
11145 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11147 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11148 Subtarget->is64Bit()) {
11152 unsigned Size = SrcVT.getSizeInBits()/8;
11153 MachineFunction &MF = DAG.getMachineFunction();
11154 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11155 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11156 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11158 MachinePointerInfo::getFixedStack(SSFI),
11160 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11163 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11165 SelectionDAG &DAG) const {
11169 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11171 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11173 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11175 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11177 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11178 MachineMemOperand *MMO;
11180 int SSFI = FI->getIndex();
11182 DAG.getMachineFunction()
11183 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11184 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11186 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11187 StackSlot = StackSlot.getOperand(1);
11189 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11190 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11192 Tys, Ops, SrcVT, MMO);
11195 Chain = Result.getValue(1);
11196 SDValue InFlag = Result.getValue(2);
11198 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11199 // shouldn't be necessary except that RFP cannot be live across
11200 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11201 MachineFunction &MF = DAG.getMachineFunction();
11202 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11203 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11204 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11205 Tys = DAG.getVTList(MVT::Other);
11207 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11209 MachineMemOperand *MMO =
11210 DAG.getMachineFunction()
11211 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11212 MachineMemOperand::MOStore, SSFISize, SSFISize);
11214 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11215 Ops, Op.getValueType(), MMO);
11216 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11217 MachinePointerInfo::getFixedStack(SSFI),
11218 false, false, false, 0);
11224 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11225 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11226 SelectionDAG &DAG) const {
11227 // This algorithm is not obvious. Here it is what we're trying to output:
11230 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11231 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11233 haddpd %xmm0, %xmm0
11235 pshufd $0x4e, %xmm0, %xmm1
11241 LLVMContext *Context = DAG.getContext();
11243 // Build some magic constants.
11244 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11245 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11246 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11248 SmallVector<Constant*,2> CV1;
11250 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11251 APInt(64, 0x4330000000000000ULL))));
11253 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11254 APInt(64, 0x4530000000000000ULL))));
11255 Constant *C1 = ConstantVector::get(CV1);
11256 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11258 // Load the 64-bit value into an XMM register.
11259 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11261 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11262 MachinePointerInfo::getConstantPool(),
11263 false, false, false, 16);
11264 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11265 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11268 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11269 MachinePointerInfo::getConstantPool(),
11270 false, false, false, 16);
11271 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11272 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11275 if (Subtarget->hasSSE3()) {
11276 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11277 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11279 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11280 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11282 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11283 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11287 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11288 DAG.getIntPtrConstant(0));
11291 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11292 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11293 SelectionDAG &DAG) const {
11295 // FP constant to bias correct the final result.
11296 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
11299 // Load the 32-bit value into an XMM register.
11300 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11303 // Zero out the upper parts of the register.
11304 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11306 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11307 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11308 DAG.getIntPtrConstant(0));
11310 // Or the load with the bias.
11311 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11312 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11313 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11314 MVT::v2f64, Load)),
11315 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11316 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11317 MVT::v2f64, Bias)));
11318 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11319 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11320 DAG.getIntPtrConstant(0));
11322 // Subtract the bias.
11323 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11325 // Handle final rounding.
11326 EVT DestVT = Op.getValueType();
11328 if (DestVT.bitsLT(MVT::f64))
11329 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11330 DAG.getIntPtrConstant(0));
11331 if (DestVT.bitsGT(MVT::f64))
11332 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11334 // Handle final rounding.
11338 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11339 const X86Subtarget &Subtarget) {
11340 // The algorithm is the following:
11341 // #ifdef __SSE4_1__
11342 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11343 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11344 // (uint4) 0x53000000, 0xaa);
11346 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11347 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11349 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11350 // return (float4) lo + fhi;
11353 SDValue V = Op->getOperand(0);
11354 EVT VecIntVT = V.getValueType();
11355 bool Is128 = VecIntVT == MVT::v4i32;
11356 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11357 // If we convert to something else than the supported type, e.g., to v4f64,
11359 if (VecFloatVT != Op->getValueType(0))
11362 unsigned NumElts = VecIntVT.getVectorNumElements();
11363 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11364 "Unsupported custom type");
11365 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11367 // In the #idef/#else code, we have in common:
11368 // - The vector of constants:
11374 // Create the splat vector for 0x4b000000.
11375 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
11376 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11377 CstLow, CstLow, CstLow, CstLow};
11378 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11379 makeArrayRef(&CstLowArray[0], NumElts));
11380 // Create the splat vector for 0x53000000.
11381 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
11382 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11383 CstHigh, CstHigh, CstHigh, CstHigh};
11384 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11385 makeArrayRef(&CstHighArray[0], NumElts));
11387 // Create the right shift.
11388 SDValue CstShift = DAG.getConstant(16, MVT::i32);
11389 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11390 CstShift, CstShift, CstShift, CstShift};
11391 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11392 makeArrayRef(&CstShiftArray[0], NumElts));
11393 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11396 if (Subtarget.hasSSE41()) {
11397 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11398 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11399 SDValue VecCstLowBitcast =
11400 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11401 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11402 // Low will be bitcasted right away, so do not bother bitcasting back to its
11404 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11405 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
11406 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11407 // (uint4) 0x53000000, 0xaa);
11408 SDValue VecCstHighBitcast =
11409 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11410 SDValue VecShiftBitcast =
11411 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11412 // High will be bitcasted right away, so do not bother bitcasting back to
11413 // its original type.
11414 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11415 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
11417 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
11418 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11419 CstMask, CstMask, CstMask);
11420 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11421 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11422 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11424 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11425 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11428 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11429 SDValue CstFAdd = DAG.getConstantFP(
11430 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
11431 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11432 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11433 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11434 makeArrayRef(&CstFAddArray[0], NumElts));
11436 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11437 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11439 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11440 // return (float4) lo + fhi;
11441 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11442 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11445 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11446 SelectionDAG &DAG) const {
11447 SDValue N0 = Op.getOperand(0);
11448 MVT SVT = N0.getSimpleValueType();
11451 switch (SVT.SimpleTy) {
11453 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11458 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11459 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11460 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11464 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11466 llvm_unreachable(nullptr);
11469 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11470 SelectionDAG &DAG) const {
11471 SDValue N0 = Op.getOperand(0);
11474 if (Op.getValueType().isVector())
11475 return lowerUINT_TO_FP_vec(Op, DAG);
11477 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11478 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11479 // the optimization here.
11480 if (DAG.SignBitIsZero(N0))
11481 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11483 MVT SrcVT = N0.getSimpleValueType();
11484 MVT DstVT = Op.getSimpleValueType();
11485 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11486 return LowerUINT_TO_FP_i64(Op, DAG);
11487 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11488 return LowerUINT_TO_FP_i32(Op, DAG);
11489 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11492 // Make a 64-bit buffer, and use it to build an FILD.
11493 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11494 if (SrcVT == MVT::i32) {
11495 SDValue WordOff = DAG.getConstant(4, getPointerTy());
11496 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11497 getPointerTy(), StackSlot, WordOff);
11498 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11499 StackSlot, MachinePointerInfo(),
11501 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
11502 OffsetSlot, MachinePointerInfo(),
11504 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11508 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11509 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11510 StackSlot, MachinePointerInfo(),
11512 // For i64 source, we need to add the appropriate power of 2 if the input
11513 // was negative. This is the same as the optimization in
11514 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11515 // we must be careful to do the computation in x87 extended precision, not
11516 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11517 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11518 MachineMemOperand *MMO =
11519 DAG.getMachineFunction()
11520 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11521 MachineMemOperand::MOLoad, 8, 8);
11523 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11524 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11525 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11528 APInt FF(32, 0x5F800000ULL);
11530 // Check whether the sign bit is set.
11531 SDValue SignSet = DAG.getSetCC(dl,
11532 getSetCCResultType(*DAG.getContext(), MVT::i64),
11533 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
11536 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11537 SDValue FudgePtr = DAG.getConstantPool(
11538 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11541 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11542 SDValue Zero = DAG.getIntPtrConstant(0);
11543 SDValue Four = DAG.getIntPtrConstant(4);
11544 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11546 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11548 // Load the value out, extending it from f32 to f80.
11549 // FIXME: Avoid the extend by constructing the right constant pool?
11550 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11551 FudgePtr, MachinePointerInfo::getConstantPool(),
11552 MVT::f32, false, false, false, 4);
11553 // Extend everything to 80 bits to force it to be done on x87.
11554 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11555 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
11558 std::pair<SDValue,SDValue>
11559 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11560 bool IsSigned, bool IsReplace) const {
11563 EVT DstTy = Op.getValueType();
11565 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11566 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11570 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11571 DstTy.getSimpleVT() >= MVT::i16 &&
11572 "Unknown FP_TO_INT to lower!");
11574 // These are really Legal.
11575 if (DstTy == MVT::i32 &&
11576 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11577 return std::make_pair(SDValue(), SDValue());
11578 if (Subtarget->is64Bit() &&
11579 DstTy == MVT::i64 &&
11580 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11581 return std::make_pair(SDValue(), SDValue());
11583 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11584 // stack slot, or into the FTOL runtime function.
11585 MachineFunction &MF = DAG.getMachineFunction();
11586 unsigned MemSize = DstTy.getSizeInBits()/8;
11587 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11588 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11591 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11592 Opc = X86ISD::WIN_FTOL;
11594 switch (DstTy.getSimpleVT().SimpleTy) {
11595 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11596 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11597 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11598 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11601 SDValue Chain = DAG.getEntryNode();
11602 SDValue Value = Op.getOperand(0);
11603 EVT TheVT = Op.getOperand(0).getValueType();
11604 // FIXME This causes a redundant load/store if the SSE-class value is already
11605 // in memory, such as if it is on the callstack.
11606 if (isScalarFPTypeInSSEReg(TheVT)) {
11607 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11608 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11609 MachinePointerInfo::getFixedStack(SSFI),
11611 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11613 Chain, StackSlot, DAG.getValueType(TheVT)
11616 MachineMemOperand *MMO =
11617 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11618 MachineMemOperand::MOLoad, MemSize, MemSize);
11619 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11620 Chain = Value.getValue(1);
11621 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11622 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11625 MachineMemOperand *MMO =
11626 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11627 MachineMemOperand::MOStore, MemSize, MemSize);
11629 if (Opc != X86ISD::WIN_FTOL) {
11630 // Build the FP_TO_INT*_IN_MEM
11631 SDValue Ops[] = { Chain, Value, StackSlot };
11632 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11634 return std::make_pair(FIST, StackSlot);
11636 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11637 DAG.getVTList(MVT::Other, MVT::Glue),
11639 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11640 MVT::i32, ftol.getValue(1));
11641 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11642 MVT::i32, eax.getValue(2));
11643 SDValue Ops[] = { eax, edx };
11644 SDValue pair = IsReplace
11645 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11646 : DAG.getMergeValues(Ops, DL);
11647 return std::make_pair(pair, SDValue());
11651 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11652 const X86Subtarget *Subtarget) {
11653 MVT VT = Op->getSimpleValueType(0);
11654 SDValue In = Op->getOperand(0);
11655 MVT InVT = In.getSimpleValueType();
11658 // Optimize vectors in AVX mode:
11661 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11662 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11663 // Concat upper and lower parts.
11666 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11667 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11668 // Concat upper and lower parts.
11671 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11672 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11673 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
11676 if (Subtarget->hasInt256())
11677 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
11679 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
11680 SDValue Undef = DAG.getUNDEF(InVT);
11681 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
11682 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11683 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11685 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
11686 VT.getVectorNumElements()/2);
11688 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
11689 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
11691 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
11694 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
11695 SelectionDAG &DAG) {
11696 MVT VT = Op->getSimpleValueType(0);
11697 SDValue In = Op->getOperand(0);
11698 MVT InVT = In.getSimpleValueType();
11700 unsigned int NumElts = VT.getVectorNumElements();
11701 if (NumElts != 8 && NumElts != 16)
11704 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
11705 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
11707 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
11708 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11709 // Now we have only mask extension
11710 assert(InVT.getVectorElementType() == MVT::i1);
11711 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
11712 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11713 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
11714 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11715 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11716 MachinePointerInfo::getConstantPool(),
11717 false, false, false, Alignment);
11719 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
11720 if (VT.is512BitVector())
11722 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
11725 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11726 SelectionDAG &DAG) {
11727 if (Subtarget->hasFp256()) {
11728 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11736 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11737 SelectionDAG &DAG) {
11739 MVT VT = Op.getSimpleValueType();
11740 SDValue In = Op.getOperand(0);
11741 MVT SVT = In.getSimpleValueType();
11743 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
11744 return LowerZERO_EXTEND_AVX512(Op, DAG);
11746 if (Subtarget->hasFp256()) {
11747 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11752 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
11753 VT.getVectorNumElements() != SVT.getVectorNumElements());
11757 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
11759 MVT VT = Op.getSimpleValueType();
11760 SDValue In = Op.getOperand(0);
11761 MVT InVT = In.getSimpleValueType();
11763 if (VT == MVT::i1) {
11764 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
11765 "Invalid scalar TRUNCATE operation");
11766 if (InVT.getSizeInBits() >= 32)
11768 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
11769 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
11771 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
11772 "Invalid TRUNCATE operation");
11774 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
11775 if (VT.getVectorElementType().getSizeInBits() >=8)
11776 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
11778 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
11779 unsigned NumElts = InVT.getVectorNumElements();
11780 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
11781 if (InVT.getSizeInBits() < 512) {
11782 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
11783 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
11787 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
11788 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11789 SDValue CP = DAG.getConstantPool(C, getPointerTy());
11790 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11791 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11792 MachinePointerInfo::getConstantPool(),
11793 false, false, false, Alignment);
11794 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
11795 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
11796 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
11799 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
11800 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
11801 if (Subtarget->hasInt256()) {
11802 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
11803 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
11804 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
11806 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
11807 DAG.getIntPtrConstant(0));
11810 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11811 DAG.getIntPtrConstant(0));
11812 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11813 DAG.getIntPtrConstant(2));
11814 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11815 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11816 static const int ShufMask[] = {0, 2, 4, 6};
11817 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
11820 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
11821 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
11822 if (Subtarget->hasInt256()) {
11823 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
11825 SmallVector<SDValue,32> pshufbMask;
11826 for (unsigned i = 0; i < 2; ++i) {
11827 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
11828 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
11829 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
11830 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
11831 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
11832 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
11833 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
11834 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
11835 for (unsigned j = 0; j < 8; ++j)
11836 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
11838 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
11839 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
11840 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
11842 static const int ShufMask[] = {0, 2, -1, -1};
11843 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
11845 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11846 DAG.getIntPtrConstant(0));
11847 return DAG.getNode(ISD::BITCAST, DL, VT, In);
11850 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11851 DAG.getIntPtrConstant(0));
11853 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11854 DAG.getIntPtrConstant(4));
11856 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
11857 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
11859 // The PSHUFB mask:
11860 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
11861 -1, -1, -1, -1, -1, -1, -1, -1};
11863 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
11864 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
11865 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
11867 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11868 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11870 // The MOVLHPS Mask:
11871 static const int ShufMask2[] = {0, 1, 4, 5};
11872 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
11873 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
11876 // Handle truncation of V256 to V128 using shuffles.
11877 if (!VT.is128BitVector() || !InVT.is256BitVector())
11880 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
11882 unsigned NumElems = VT.getVectorNumElements();
11883 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
11885 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
11886 // Prepare truncation shuffle mask
11887 for (unsigned i = 0; i != NumElems; ++i)
11888 MaskVec[i] = i * 2;
11889 SDValue V = DAG.getVectorShuffle(NVT, DL,
11890 DAG.getNode(ISD::BITCAST, DL, NVT, In),
11891 DAG.getUNDEF(NVT), &MaskVec[0]);
11892 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
11893 DAG.getIntPtrConstant(0));
11896 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
11897 SelectionDAG &DAG) const {
11898 assert(!Op.getSimpleValueType().isVector());
11900 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11901 /*IsSigned=*/ true, /*IsReplace=*/ false);
11902 SDValue FIST = Vals.first, StackSlot = Vals.second;
11903 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
11904 if (!FIST.getNode()) return Op;
11906 if (StackSlot.getNode())
11907 // Load the result.
11908 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11909 FIST, StackSlot, MachinePointerInfo(),
11910 false, false, false, 0);
11912 // The node is the result.
11916 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
11917 SelectionDAG &DAG) const {
11918 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11919 /*IsSigned=*/ false, /*IsReplace=*/ false);
11920 SDValue FIST = Vals.first, StackSlot = Vals.second;
11921 assert(FIST.getNode() && "Unexpected failure");
11923 if (StackSlot.getNode())
11924 // Load the result.
11925 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11926 FIST, StackSlot, MachinePointerInfo(),
11927 false, false, false, 0);
11929 // The node is the result.
11933 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
11935 MVT VT = Op.getSimpleValueType();
11936 SDValue In = Op.getOperand(0);
11937 MVT SVT = In.getSimpleValueType();
11939 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
11941 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
11942 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
11943 In, DAG.getUNDEF(SVT)));
11946 /// The only differences between FABS and FNEG are the mask and the logic op.
11947 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
11948 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
11949 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
11950 "Wrong opcode for lowering FABS or FNEG.");
11952 bool IsFABS = (Op.getOpcode() == ISD::FABS);
11954 // If this is a FABS and it has an FNEG user, bail out to fold the combination
11955 // into an FNABS. We'll lower the FABS after that if it is still in use.
11957 for (SDNode *User : Op->uses())
11958 if (User->getOpcode() == ISD::FNEG)
11961 SDValue Op0 = Op.getOperand(0);
11962 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
11965 MVT VT = Op.getSimpleValueType();
11966 // Assume scalar op for initialization; update for vector if needed.
11967 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
11968 // generate a 16-byte vector constant and logic op even for the scalar case.
11969 // Using a 16-byte mask allows folding the load of the mask with
11970 // the logic op, so it can save (~4 bytes) on code size.
11972 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
11973 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
11974 // decide if we should generate a 16-byte constant mask when we only need 4 or
11975 // 8 bytes for the scalar case.
11976 if (VT.isVector()) {
11977 EltVT = VT.getVectorElementType();
11978 NumElts = VT.getVectorNumElements();
11981 unsigned EltBits = EltVT.getSizeInBits();
11982 LLVMContext *Context = DAG.getContext();
11983 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
11985 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
11986 Constant *C = ConstantInt::get(*Context, MaskElt);
11987 C = ConstantVector::getSplat(NumElts, C);
11988 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11989 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
11990 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11991 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11992 MachinePointerInfo::getConstantPool(),
11993 false, false, false, Alignment);
11995 if (VT.isVector()) {
11996 // For a vector, cast operands to a vector type, perform the logic op,
11997 // and cast the result back to the original value type.
11998 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
11999 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12000 SDValue Operand = IsFNABS ?
12001 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12002 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12003 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12004 return DAG.getNode(ISD::BITCAST, dl, VT,
12005 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12008 // If not vector, then scalar.
12009 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12010 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12011 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12014 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12015 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12016 LLVMContext *Context = DAG.getContext();
12017 SDValue Op0 = Op.getOperand(0);
12018 SDValue Op1 = Op.getOperand(1);
12020 MVT VT = Op.getSimpleValueType();
12021 MVT SrcVT = Op1.getSimpleValueType();
12023 // If second operand is smaller, extend it first.
12024 if (SrcVT.bitsLT(VT)) {
12025 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12028 // And if it is bigger, shrink it first.
12029 if (SrcVT.bitsGT(VT)) {
12030 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
12034 // At this point the operands and the result should have the same
12035 // type, and that won't be f80 since that is not custom lowered.
12037 const fltSemantics &Sem =
12038 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12039 const unsigned SizeInBits = VT.getSizeInBits();
12041 SmallVector<Constant *, 4> CV(
12042 VT == MVT::f64 ? 2 : 4,
12043 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12045 // First, clear all bits but the sign bit from the second operand (sign).
12046 CV[0] = ConstantFP::get(*Context,
12047 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12048 Constant *C = ConstantVector::get(CV);
12049 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12050 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12051 MachinePointerInfo::getConstantPool(),
12052 false, false, false, 16);
12053 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12055 // Next, clear the sign bit from the first operand (magnitude).
12056 // If it's a constant, we can clear it here.
12057 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12058 APFloat APF = Op0CN->getValueAPF();
12059 // If the magnitude is a positive zero, the sign bit alone is enough.
12060 if (APF.isPosZero())
12063 CV[0] = ConstantFP::get(*Context, APF);
12065 CV[0] = ConstantFP::get(
12067 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12069 C = ConstantVector::get(CV);
12070 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12071 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12072 MachinePointerInfo::getConstantPool(),
12073 false, false, false, 16);
12074 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12075 if (!isa<ConstantFPSDNode>(Op0))
12076 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12078 // OR the magnitude value with the sign bit.
12079 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12082 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12083 SDValue N0 = Op.getOperand(0);
12085 MVT VT = Op.getSimpleValueType();
12087 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12088 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12089 DAG.getConstant(1, VT));
12090 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
12093 // Check whether an OR'd tree is PTEST-able.
12094 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12095 SelectionDAG &DAG) {
12096 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12098 if (!Subtarget->hasSSE41())
12101 if (!Op->hasOneUse())
12104 SDNode *N = Op.getNode();
12107 SmallVector<SDValue, 8> Opnds;
12108 DenseMap<SDValue, unsigned> VecInMap;
12109 SmallVector<SDValue, 8> VecIns;
12110 EVT VT = MVT::Other;
12112 // Recognize a special case where a vector is casted into wide integer to
12114 Opnds.push_back(N->getOperand(0));
12115 Opnds.push_back(N->getOperand(1));
12117 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12118 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12119 // BFS traverse all OR'd operands.
12120 if (I->getOpcode() == ISD::OR) {
12121 Opnds.push_back(I->getOperand(0));
12122 Opnds.push_back(I->getOperand(1));
12123 // Re-evaluate the number of nodes to be traversed.
12124 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12128 // Quit if a non-EXTRACT_VECTOR_ELT
12129 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12132 // Quit if without a constant index.
12133 SDValue Idx = I->getOperand(1);
12134 if (!isa<ConstantSDNode>(Idx))
12137 SDValue ExtractedFromVec = I->getOperand(0);
12138 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12139 if (M == VecInMap.end()) {
12140 VT = ExtractedFromVec.getValueType();
12141 // Quit if not 128/256-bit vector.
12142 if (!VT.is128BitVector() && !VT.is256BitVector())
12144 // Quit if not the same type.
12145 if (VecInMap.begin() != VecInMap.end() &&
12146 VT != VecInMap.begin()->first.getValueType())
12148 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12149 VecIns.push_back(ExtractedFromVec);
12151 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12154 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12155 "Not extracted from 128-/256-bit vector.");
12157 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12159 for (DenseMap<SDValue, unsigned>::const_iterator
12160 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12161 // Quit if not all elements are used.
12162 if (I->second != FullMask)
12166 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12168 // Cast all vectors into TestVT for PTEST.
12169 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12170 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12172 // If more than one full vectors are evaluated, OR them first before PTEST.
12173 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12174 // Each iteration will OR 2 nodes and append the result until there is only
12175 // 1 node left, i.e. the final OR'd value of all vectors.
12176 SDValue LHS = VecIns[Slot];
12177 SDValue RHS = VecIns[Slot + 1];
12178 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12181 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12182 VecIns.back(), VecIns.back());
12185 /// \brief return true if \c Op has a use that doesn't just read flags.
12186 static bool hasNonFlagsUse(SDValue Op) {
12187 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12189 SDNode *User = *UI;
12190 unsigned UOpNo = UI.getOperandNo();
12191 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12192 // Look pass truncate.
12193 UOpNo = User->use_begin().getOperandNo();
12194 User = *User->use_begin();
12197 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12198 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12204 /// Emit nodes that will be selected as "test Op0,Op0", or something
12206 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12207 SelectionDAG &DAG) const {
12208 if (Op.getValueType() == MVT::i1) {
12209 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12210 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12211 DAG.getConstant(0, MVT::i8));
12213 // CF and OF aren't always set the way we want. Determine which
12214 // of these we need.
12215 bool NeedCF = false;
12216 bool NeedOF = false;
12219 case X86::COND_A: case X86::COND_AE:
12220 case X86::COND_B: case X86::COND_BE:
12223 case X86::COND_G: case X86::COND_GE:
12224 case X86::COND_L: case X86::COND_LE:
12225 case X86::COND_O: case X86::COND_NO: {
12226 // Check if we really need to set the
12227 // Overflow flag. If NoSignedWrap is present
12228 // that is not actually needed.
12229 switch (Op->getOpcode()) {
12234 const BinaryWithFlagsSDNode *BinNode =
12235 cast<BinaryWithFlagsSDNode>(Op.getNode());
12236 if (BinNode->hasNoSignedWrap())
12246 // See if we can use the EFLAGS value from the operand instead of
12247 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12248 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12249 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12250 // Emit a CMP with 0, which is the TEST pattern.
12251 //if (Op.getValueType() == MVT::i1)
12252 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12253 // DAG.getConstant(0, MVT::i1));
12254 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12255 DAG.getConstant(0, Op.getValueType()));
12257 unsigned Opcode = 0;
12258 unsigned NumOperands = 0;
12260 // Truncate operations may prevent the merge of the SETCC instruction
12261 // and the arithmetic instruction before it. Attempt to truncate the operands
12262 // of the arithmetic instruction and use a reduced bit-width instruction.
12263 bool NeedTruncation = false;
12264 SDValue ArithOp = Op;
12265 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12266 SDValue Arith = Op->getOperand(0);
12267 // Both the trunc and the arithmetic op need to have one user each.
12268 if (Arith->hasOneUse())
12269 switch (Arith.getOpcode()) {
12276 NeedTruncation = true;
12282 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12283 // which may be the result of a CAST. We use the variable 'Op', which is the
12284 // non-casted variable when we check for possible users.
12285 switch (ArithOp.getOpcode()) {
12287 // Due to an isel shortcoming, be conservative if this add is likely to be
12288 // selected as part of a load-modify-store instruction. When the root node
12289 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12290 // uses of other nodes in the match, such as the ADD in this case. This
12291 // leads to the ADD being left around and reselected, with the result being
12292 // two adds in the output. Alas, even if none our users are stores, that
12293 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12294 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12295 // climbing the DAG back to the root, and it doesn't seem to be worth the
12297 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12298 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12299 if (UI->getOpcode() != ISD::CopyToReg &&
12300 UI->getOpcode() != ISD::SETCC &&
12301 UI->getOpcode() != ISD::STORE)
12304 if (ConstantSDNode *C =
12305 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12306 // An add of one will be selected as an INC.
12307 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12308 Opcode = X86ISD::INC;
12313 // An add of negative one (subtract of one) will be selected as a DEC.
12314 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12315 Opcode = X86ISD::DEC;
12321 // Otherwise use a regular EFLAGS-setting add.
12322 Opcode = X86ISD::ADD;
12327 // If we have a constant logical shift that's only used in a comparison
12328 // against zero turn it into an equivalent AND. This allows turning it into
12329 // a TEST instruction later.
12330 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12331 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12332 EVT VT = Op.getValueType();
12333 unsigned BitWidth = VT.getSizeInBits();
12334 unsigned ShAmt = Op->getConstantOperandVal(1);
12335 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12337 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12338 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12339 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12340 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12342 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12343 DAG.getConstant(Mask, VT));
12344 DAG.ReplaceAllUsesWith(Op, New);
12350 // If the primary and result isn't used, don't bother using X86ISD::AND,
12351 // because a TEST instruction will be better.
12352 if (!hasNonFlagsUse(Op))
12358 // Due to the ISEL shortcoming noted above, be conservative if this op is
12359 // likely to be selected as part of a load-modify-store instruction.
12360 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12361 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12362 if (UI->getOpcode() == ISD::STORE)
12365 // Otherwise use a regular EFLAGS-setting instruction.
12366 switch (ArithOp.getOpcode()) {
12367 default: llvm_unreachable("unexpected operator!");
12368 case ISD::SUB: Opcode = X86ISD::SUB; break;
12369 case ISD::XOR: Opcode = X86ISD::XOR; break;
12370 case ISD::AND: Opcode = X86ISD::AND; break;
12372 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12373 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12374 if (EFLAGS.getNode())
12377 Opcode = X86ISD::OR;
12391 return SDValue(Op.getNode(), 1);
12397 // If we found that truncation is beneficial, perform the truncation and
12399 if (NeedTruncation) {
12400 EVT VT = Op.getValueType();
12401 SDValue WideVal = Op->getOperand(0);
12402 EVT WideVT = WideVal.getValueType();
12403 unsigned ConvertedOp = 0;
12404 // Use a target machine opcode to prevent further DAGCombine
12405 // optimizations that may separate the arithmetic operations
12406 // from the setcc node.
12407 switch (WideVal.getOpcode()) {
12409 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12410 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12411 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12412 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12413 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12417 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12418 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12419 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12420 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12421 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12427 // Emit a CMP with 0, which is the TEST pattern.
12428 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12429 DAG.getConstant(0, Op.getValueType()));
12431 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12432 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12434 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12435 DAG.ReplaceAllUsesWith(Op, New);
12436 return SDValue(New.getNode(), 1);
12439 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12441 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12442 SDLoc dl, SelectionDAG &DAG) const {
12443 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12444 if (C->getAPIntValue() == 0)
12445 return EmitTest(Op0, X86CC, dl, DAG);
12447 if (Op0.getValueType() == MVT::i1)
12448 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12451 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12452 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12453 // Do the comparison at i32 if it's smaller, besides the Atom case.
12454 // This avoids subregister aliasing issues. Keep the smaller reference
12455 // if we're optimizing for size, however, as that'll allow better folding
12456 // of memory operations.
12457 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12458 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12459 Attribute::MinSize) &&
12460 !Subtarget->isAtom()) {
12461 unsigned ExtendOp =
12462 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12463 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12464 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12466 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12467 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12468 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12470 return SDValue(Sub.getNode(), 1);
12472 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12475 /// Convert a comparison if required by the subtarget.
12476 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12477 SelectionDAG &DAG) const {
12478 // If the subtarget does not support the FUCOMI instruction, floating-point
12479 // comparisons have to be converted.
12480 if (Subtarget->hasCMov() ||
12481 Cmp.getOpcode() != X86ISD::CMP ||
12482 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12483 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12486 // The instruction selector will select an FUCOM instruction instead of
12487 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12488 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12489 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12491 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12492 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12493 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12494 DAG.getConstant(8, MVT::i8));
12495 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12496 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12499 /// The minimum architected relative accuracy is 2^-12. We need one
12500 /// Newton-Raphson step to have a good float result (24 bits of precision).
12501 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12502 DAGCombinerInfo &DCI,
12503 unsigned &RefinementSteps,
12504 bool &UseOneConstNR) const {
12505 // FIXME: We should use instruction latency models to calculate the cost of
12506 // each potential sequence, but this is very hard to do reliably because
12507 // at least Intel's Core* chips have variable timing based on the number of
12508 // significant digits in the divisor and/or sqrt operand.
12509 if (!Subtarget->useSqrtEst())
12512 EVT VT = Op.getValueType();
12514 // SSE1 has rsqrtss and rsqrtps.
12515 // TODO: Add support for AVX512 (v16f32).
12516 // It is likely not profitable to do this for f64 because a double-precision
12517 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12518 // instructions: convert to single, rsqrtss, convert back to double, refine
12519 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12520 // along with FMA, this could be a throughput win.
12521 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12522 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12523 RefinementSteps = 1;
12524 UseOneConstNR = false;
12525 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12530 /// The minimum architected relative accuracy is 2^-12. We need one
12531 /// Newton-Raphson step to have a good float result (24 bits of precision).
12532 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12533 DAGCombinerInfo &DCI,
12534 unsigned &RefinementSteps) const {
12535 // FIXME: We should use instruction latency models to calculate the cost of
12536 // each potential sequence, but this is very hard to do reliably because
12537 // at least Intel's Core* chips have variable timing based on the number of
12538 // significant digits in the divisor.
12539 if (!Subtarget->useReciprocalEst())
12542 EVT VT = Op.getValueType();
12544 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12545 // TODO: Add support for AVX512 (v16f32).
12546 // It is likely not profitable to do this for f64 because a double-precision
12547 // reciprocal estimate with refinement on x86 prior to FMA requires
12548 // 15 instructions: convert to single, rcpss, convert back to double, refine
12549 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12550 // along with FMA, this could be a throughput win.
12551 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12552 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12553 RefinementSteps = ReciprocalEstimateRefinementSteps;
12554 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12559 static bool isAllOnes(SDValue V) {
12560 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12561 return C && C->isAllOnesValue();
12564 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12565 /// if it's possible.
12566 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12567 SDLoc dl, SelectionDAG &DAG) const {
12568 SDValue Op0 = And.getOperand(0);
12569 SDValue Op1 = And.getOperand(1);
12570 if (Op0.getOpcode() == ISD::TRUNCATE)
12571 Op0 = Op0.getOperand(0);
12572 if (Op1.getOpcode() == ISD::TRUNCATE)
12573 Op1 = Op1.getOperand(0);
12576 if (Op1.getOpcode() == ISD::SHL)
12577 std::swap(Op0, Op1);
12578 if (Op0.getOpcode() == ISD::SHL) {
12579 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12580 if (And00C->getZExtValue() == 1) {
12581 // If we looked past a truncate, check that it's only truncating away
12583 unsigned BitWidth = Op0.getValueSizeInBits();
12584 unsigned AndBitWidth = And.getValueSizeInBits();
12585 if (BitWidth > AndBitWidth) {
12587 DAG.computeKnownBits(Op0, Zeros, Ones);
12588 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12592 RHS = Op0.getOperand(1);
12594 } else if (Op1.getOpcode() == ISD::Constant) {
12595 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
12596 uint64_t AndRHSVal = AndRHS->getZExtValue();
12597 SDValue AndLHS = Op0;
12599 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
12600 LHS = AndLHS.getOperand(0);
12601 RHS = AndLHS.getOperand(1);
12604 // Use BT if the immediate can't be encoded in a TEST instruction.
12605 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
12607 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
12611 if (LHS.getNode()) {
12612 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12613 // instruction. Since the shift amount is in-range-or-undefined, we know
12614 // that doing a bittest on the i32 value is ok. We extend to i32 because
12615 // the encoding for the i16 version is larger than the i32 version.
12616 // Also promote i16 to i32 for performance / code size reason.
12617 if (LHS.getValueType() == MVT::i8 ||
12618 LHS.getValueType() == MVT::i16)
12619 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12621 // If the operand types disagree, extend the shift amount to match. Since
12622 // BT ignores high bits (like shifts) we can use anyextend.
12623 if (LHS.getValueType() != RHS.getValueType())
12624 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12626 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12627 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12628 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12629 DAG.getConstant(Cond, MVT::i8), BT);
12635 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12637 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12642 // SSE Condition code mapping:
12651 switch (SetCCOpcode) {
12652 default: llvm_unreachable("Unexpected SETCC condition");
12654 case ISD::SETEQ: SSECC = 0; break;
12656 case ISD::SETGT: Swap = true; // Fallthrough
12658 case ISD::SETOLT: SSECC = 1; break;
12660 case ISD::SETGE: Swap = true; // Fallthrough
12662 case ISD::SETOLE: SSECC = 2; break;
12663 case ISD::SETUO: SSECC = 3; break;
12665 case ISD::SETNE: SSECC = 4; break;
12666 case ISD::SETULE: Swap = true; // Fallthrough
12667 case ISD::SETUGE: SSECC = 5; break;
12668 case ISD::SETULT: Swap = true; // Fallthrough
12669 case ISD::SETUGT: SSECC = 6; break;
12670 case ISD::SETO: SSECC = 7; break;
12672 case ISD::SETONE: SSECC = 8; break;
12675 std::swap(Op0, Op1);
12680 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
12681 // ones, and then concatenate the result back.
12682 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
12683 MVT VT = Op.getSimpleValueType();
12685 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
12686 "Unsupported value type for operation");
12688 unsigned NumElems = VT.getVectorNumElements();
12690 SDValue CC = Op.getOperand(2);
12692 // Extract the LHS vectors
12693 SDValue LHS = Op.getOperand(0);
12694 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12695 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12697 // Extract the RHS vectors
12698 SDValue RHS = Op.getOperand(1);
12699 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
12700 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
12702 // Issue the operation on the smaller types and concatenate the result back
12703 MVT EltVT = VT.getVectorElementType();
12704 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12705 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
12706 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
12707 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
12710 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
12711 const X86Subtarget *Subtarget) {
12712 SDValue Op0 = Op.getOperand(0);
12713 SDValue Op1 = Op.getOperand(1);
12714 SDValue CC = Op.getOperand(2);
12715 MVT VT = Op.getSimpleValueType();
12718 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
12719 Op.getValueType().getScalarType() == MVT::i1 &&
12720 "Cannot set masked compare for this operation");
12722 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12724 bool Unsigned = false;
12727 switch (SetCCOpcode) {
12728 default: llvm_unreachable("Unexpected SETCC condition");
12729 case ISD::SETNE: SSECC = 4; break;
12730 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
12731 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
12732 case ISD::SETLT: Swap = true; //fall-through
12733 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
12734 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
12735 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
12736 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
12737 case ISD::SETULE: Unsigned = true; //fall-through
12738 case ISD::SETLE: SSECC = 2; break;
12742 std::swap(Op0, Op1);
12744 return DAG.getNode(Opc, dl, VT, Op0, Op1);
12745 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
12746 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12747 DAG.getConstant(SSECC, MVT::i8));
12750 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
12751 /// operand \p Op1. If non-trivial (for example because it's not constant)
12752 /// return an empty value.
12753 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
12755 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
12759 MVT VT = Op1.getSimpleValueType();
12760 MVT EVT = VT.getVectorElementType();
12761 unsigned n = VT.getVectorNumElements();
12762 SmallVector<SDValue, 8> ULTOp1;
12764 for (unsigned i = 0; i < n; ++i) {
12765 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
12766 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
12769 // Avoid underflow.
12770 APInt Val = Elt->getAPIntValue();
12774 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
12777 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
12780 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
12781 SelectionDAG &DAG) {
12782 SDValue Op0 = Op.getOperand(0);
12783 SDValue Op1 = Op.getOperand(1);
12784 SDValue CC = Op.getOperand(2);
12785 MVT VT = Op.getSimpleValueType();
12786 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12787 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
12792 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
12793 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
12796 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
12797 unsigned Opc = X86ISD::CMPP;
12798 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
12799 assert(VT.getVectorNumElements() <= 16);
12800 Opc = X86ISD::CMPM;
12802 // In the two special cases we can't handle, emit two comparisons.
12805 unsigned CombineOpc;
12806 if (SetCCOpcode == ISD::SETUEQ) {
12807 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
12809 assert(SetCCOpcode == ISD::SETONE);
12810 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
12813 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12814 DAG.getConstant(CC0, MVT::i8));
12815 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12816 DAG.getConstant(CC1, MVT::i8));
12817 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
12819 // Handle all other FP comparisons here.
12820 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12821 DAG.getConstant(SSECC, MVT::i8));
12824 // Break 256-bit integer vector compare into smaller ones.
12825 if (VT.is256BitVector() && !Subtarget->hasInt256())
12826 return Lower256IntVSETCC(Op, DAG);
12828 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
12829 EVT OpVT = Op1.getValueType();
12830 if (Subtarget->hasAVX512()) {
12831 if (Op1.getValueType().is512BitVector() ||
12832 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
12833 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
12834 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
12836 // In AVX-512 architecture setcc returns mask with i1 elements,
12837 // But there is no compare instruction for i8 and i16 elements in KNL.
12838 // We are not talking about 512-bit operands in this case, these
12839 // types are illegal.
12841 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
12842 OpVT.getVectorElementType().getSizeInBits() >= 8))
12843 return DAG.getNode(ISD::TRUNCATE, dl, VT,
12844 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
12847 // We are handling one of the integer comparisons here. Since SSE only has
12848 // GT and EQ comparisons for integer, swapping operands and multiple
12849 // operations may be required for some comparisons.
12851 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
12852 bool Subus = false;
12854 switch (SetCCOpcode) {
12855 default: llvm_unreachable("Unexpected SETCC condition");
12856 case ISD::SETNE: Invert = true;
12857 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
12858 case ISD::SETLT: Swap = true;
12859 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
12860 case ISD::SETGE: Swap = true;
12861 case ISD::SETLE: Opc = X86ISD::PCMPGT;
12862 Invert = true; break;
12863 case ISD::SETULT: Swap = true;
12864 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
12865 FlipSigns = true; break;
12866 case ISD::SETUGE: Swap = true;
12867 case ISD::SETULE: Opc = X86ISD::PCMPGT;
12868 FlipSigns = true; Invert = true; break;
12871 // Special case: Use min/max operations for SETULE/SETUGE
12872 MVT VET = VT.getVectorElementType();
12874 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
12875 || (Subtarget->hasSSE2() && (VET == MVT::i8));
12878 switch (SetCCOpcode) {
12880 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
12881 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
12884 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
12887 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
12888 if (!MinMax && hasSubus) {
12889 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
12891 // t = psubus Op0, Op1
12892 // pcmpeq t, <0..0>
12893 switch (SetCCOpcode) {
12895 case ISD::SETULT: {
12896 // If the comparison is against a constant we can turn this into a
12897 // setule. With psubus, setule does not require a swap. This is
12898 // beneficial because the constant in the register is no longer
12899 // destructed as the destination so it can be hoisted out of a loop.
12900 // Only do this pre-AVX since vpcmp* is no longer destructive.
12901 if (Subtarget->hasAVX())
12903 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
12904 if (ULEOp1.getNode()) {
12906 Subus = true; Invert = false; Swap = false;
12910 // Psubus is better than flip-sign because it requires no inversion.
12911 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
12912 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
12916 Opc = X86ISD::SUBUS;
12922 std::swap(Op0, Op1);
12924 // Check that the operation in question is available (most are plain SSE2,
12925 // but PCMPGTQ and PCMPEQQ have different requirements).
12926 if (VT == MVT::v2i64) {
12927 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
12928 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
12930 // First cast everything to the right type.
12931 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12932 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12934 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12935 // bits of the inputs before performing those operations. The lower
12936 // compare is always unsigned.
12939 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
12941 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
12942 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
12943 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
12944 Sign, Zero, Sign, Zero);
12946 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
12947 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
12949 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
12950 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
12951 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
12953 // Create masks for only the low parts/high parts of the 64 bit integers.
12954 static const int MaskHi[] = { 1, 1, 3, 3 };
12955 static const int MaskLo[] = { 0, 0, 2, 2 };
12956 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
12957 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
12958 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
12960 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
12961 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
12964 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12966 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12969 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
12970 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
12971 // pcmpeqd + pshufd + pand.
12972 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
12974 // First cast everything to the right type.
12975 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12976 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12979 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
12981 // Make sure the lower and upper halves are both all-ones.
12982 static const int Mask[] = { 1, 0, 3, 2 };
12983 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
12984 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
12987 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12989 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12993 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12994 // bits of the inputs before performing those operations.
12996 EVT EltVT = VT.getVectorElementType();
12997 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
12998 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
12999 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13002 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13004 // If the logical-not of the result is required, perform that now.
13006 Result = DAG.getNOT(dl, Result, VT);
13009 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13012 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13013 getZeroVector(VT, Subtarget, DAG, dl));
13018 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13020 MVT VT = Op.getSimpleValueType();
13022 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13024 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13025 && "SetCC type must be 8-bit or 1-bit integer");
13026 SDValue Op0 = Op.getOperand(0);
13027 SDValue Op1 = Op.getOperand(1);
13029 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13031 // Optimize to BT if possible.
13032 // Lower (X & (1 << N)) == 0 to BT(X, N).
13033 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13034 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13035 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13036 Op1.getOpcode() == ISD::Constant &&
13037 cast<ConstantSDNode>(Op1)->isNullValue() &&
13038 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13039 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13040 if (NewSetCC.getNode()) {
13042 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13047 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13049 if (Op1.getOpcode() == ISD::Constant &&
13050 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13051 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13052 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13054 // If the input is a setcc, then reuse the input setcc or use a new one with
13055 // the inverted condition.
13056 if (Op0.getOpcode() == X86ISD::SETCC) {
13057 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13058 bool Invert = (CC == ISD::SETNE) ^
13059 cast<ConstantSDNode>(Op1)->isNullValue();
13063 CCode = X86::GetOppositeBranchCondition(CCode);
13064 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13065 DAG.getConstant(CCode, MVT::i8),
13066 Op0.getOperand(1));
13068 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13072 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13073 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13074 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13076 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13077 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
13080 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13081 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
13082 if (X86CC == X86::COND_INVALID)
13085 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13086 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13087 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13088 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
13090 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13094 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13095 static bool isX86LogicalCmp(SDValue Op) {
13096 unsigned Opc = Op.getNode()->getOpcode();
13097 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13098 Opc == X86ISD::SAHF)
13100 if (Op.getResNo() == 1 &&
13101 (Opc == X86ISD::ADD ||
13102 Opc == X86ISD::SUB ||
13103 Opc == X86ISD::ADC ||
13104 Opc == X86ISD::SBB ||
13105 Opc == X86ISD::SMUL ||
13106 Opc == X86ISD::UMUL ||
13107 Opc == X86ISD::INC ||
13108 Opc == X86ISD::DEC ||
13109 Opc == X86ISD::OR ||
13110 Opc == X86ISD::XOR ||
13111 Opc == X86ISD::AND))
13114 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13120 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13121 if (V.getOpcode() != ISD::TRUNCATE)
13124 SDValue VOp0 = V.getOperand(0);
13125 unsigned InBits = VOp0.getValueSizeInBits();
13126 unsigned Bits = V.getValueSizeInBits();
13127 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13130 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13131 bool addTest = true;
13132 SDValue Cond = Op.getOperand(0);
13133 SDValue Op1 = Op.getOperand(1);
13134 SDValue Op2 = Op.getOperand(2);
13136 EVT VT = Op1.getValueType();
13139 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13140 // are available. Otherwise fp cmovs get lowered into a less efficient branch
13141 // sequence later on.
13142 if (Cond.getOpcode() == ISD::SETCC &&
13143 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13144 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13145 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13146 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13147 int SSECC = translateX86FSETCC(
13148 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13151 if (Subtarget->hasAVX512()) {
13152 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13153 DAG.getConstant(SSECC, MVT::i8));
13154 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13156 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13157 DAG.getConstant(SSECC, MVT::i8));
13158 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13159 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13160 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13164 if (Cond.getOpcode() == ISD::SETCC) {
13165 SDValue NewCond = LowerSETCC(Cond, DAG);
13166 if (NewCond.getNode())
13170 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13171 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13172 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13173 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13174 if (Cond.getOpcode() == X86ISD::SETCC &&
13175 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13176 isZero(Cond.getOperand(1).getOperand(1))) {
13177 SDValue Cmp = Cond.getOperand(1);
13179 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13181 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13182 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13183 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13185 SDValue CmpOp0 = Cmp.getOperand(0);
13186 // Apply further optimizations for special cases
13187 // (select (x != 0), -1, 0) -> neg & sbb
13188 // (select (x == 0), 0, -1) -> neg & sbb
13189 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13190 if (YC->isNullValue() &&
13191 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13192 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13193 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13194 DAG.getConstant(0, CmpOp0.getValueType()),
13196 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13197 DAG.getConstant(X86::COND_B, MVT::i8),
13198 SDValue(Neg.getNode(), 1));
13202 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13203 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
13204 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13206 SDValue Res = // Res = 0 or -1.
13207 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13208 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
13210 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13211 Res = DAG.getNOT(DL, Res, Res.getValueType());
13213 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13214 if (!N2C || !N2C->isNullValue())
13215 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13220 // Look past (and (setcc_carry (cmp ...)), 1).
13221 if (Cond.getOpcode() == ISD::AND &&
13222 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13223 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13224 if (C && C->getAPIntValue() == 1)
13225 Cond = Cond.getOperand(0);
13228 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13229 // setting operand in place of the X86ISD::SETCC.
13230 unsigned CondOpcode = Cond.getOpcode();
13231 if (CondOpcode == X86ISD::SETCC ||
13232 CondOpcode == X86ISD::SETCC_CARRY) {
13233 CC = Cond.getOperand(0);
13235 SDValue Cmp = Cond.getOperand(1);
13236 unsigned Opc = Cmp.getOpcode();
13237 MVT VT = Op.getSimpleValueType();
13239 bool IllegalFPCMov = false;
13240 if (VT.isFloatingPoint() && !VT.isVector() &&
13241 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13242 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13244 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13245 Opc == X86ISD::BT) { // FIXME
13249 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13250 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13251 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13252 Cond.getOperand(0).getValueType() != MVT::i8)) {
13253 SDValue LHS = Cond.getOperand(0);
13254 SDValue RHS = Cond.getOperand(1);
13255 unsigned X86Opcode;
13258 switch (CondOpcode) {
13259 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13260 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13261 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13262 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13263 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13264 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13265 default: llvm_unreachable("unexpected overflowing operator");
13267 if (CondOpcode == ISD::UMULO)
13268 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13271 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13273 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13275 if (CondOpcode == ISD::UMULO)
13276 Cond = X86Op.getValue(2);
13278 Cond = X86Op.getValue(1);
13280 CC = DAG.getConstant(X86Cond, MVT::i8);
13285 // Look pass the truncate if the high bits are known zero.
13286 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13287 Cond = Cond.getOperand(0);
13289 // We know the result of AND is compared against zero. Try to match
13291 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13292 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13293 if (NewSetCC.getNode()) {
13294 CC = NewSetCC.getOperand(0);
13295 Cond = NewSetCC.getOperand(1);
13302 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13303 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13306 // a < b ? -1 : 0 -> RES = ~setcc_carry
13307 // a < b ? 0 : -1 -> RES = setcc_carry
13308 // a >= b ? -1 : 0 -> RES = setcc_carry
13309 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13310 if (Cond.getOpcode() == X86ISD::SUB) {
13311 Cond = ConvertCmpIfNecessary(Cond, DAG);
13312 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13314 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13315 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13316 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13317 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
13318 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13319 return DAG.getNOT(DL, Res, Res.getValueType());
13324 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13325 // widen the cmov and push the truncate through. This avoids introducing a new
13326 // branch during isel and doesn't add any extensions.
13327 if (Op.getValueType() == MVT::i8 &&
13328 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13329 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13330 if (T1.getValueType() == T2.getValueType() &&
13331 // Blacklist CopyFromReg to avoid partial register stalls.
13332 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13333 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13334 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13335 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13339 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13340 // condition is true.
13341 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13342 SDValue Ops[] = { Op2, Op1, CC, Cond };
13343 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13346 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13347 SelectionDAG &DAG) {
13348 MVT VT = Op->getSimpleValueType(0);
13349 SDValue In = Op->getOperand(0);
13350 MVT InVT = In.getSimpleValueType();
13351 MVT VTElt = VT.getVectorElementType();
13352 MVT InVTElt = InVT.getVectorElementType();
13356 if ((InVTElt == MVT::i1) &&
13357 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13358 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13360 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13361 VTElt.getSizeInBits() <= 16)) ||
13363 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13364 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13366 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13367 VTElt.getSizeInBits() >= 32))))
13368 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13370 unsigned int NumElts = VT.getVectorNumElements();
13372 if (NumElts != 8 && NumElts != 16)
13375 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13376 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13377 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13378 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13381 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13382 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13384 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
13385 Constant *C = ConstantInt::get(*DAG.getContext(),
13386 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
13388 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
13389 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13390 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
13391 MachinePointerInfo::getConstantPool(),
13392 false, false, false, Alignment);
13393 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
13394 if (VT.is512BitVector())
13396 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
13399 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13400 SelectionDAG &DAG) {
13401 MVT VT = Op->getSimpleValueType(0);
13402 SDValue In = Op->getOperand(0);
13403 MVT InVT = In.getSimpleValueType();
13406 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13407 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13409 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13410 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13411 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13414 if (Subtarget->hasInt256())
13415 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13417 // Optimize vectors in AVX mode
13418 // Sign extend v8i16 to v8i32 and
13421 // Divide input vector into two parts
13422 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13423 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13424 // concat the vectors to original VT
13426 unsigned NumElems = InVT.getVectorNumElements();
13427 SDValue Undef = DAG.getUNDEF(InVT);
13429 SmallVector<int,8> ShufMask1(NumElems, -1);
13430 for (unsigned i = 0; i != NumElems/2; ++i)
13433 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13435 SmallVector<int,8> ShufMask2(NumElems, -1);
13436 for (unsigned i = 0; i != NumElems/2; ++i)
13437 ShufMask2[i] = i + NumElems/2;
13439 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13441 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13442 VT.getVectorNumElements()/2);
13444 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13445 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13447 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13450 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13451 // may emit an illegal shuffle but the expansion is still better than scalar
13452 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13453 // we'll emit a shuffle and a arithmetic shift.
13454 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13455 // TODO: It is possible to support ZExt by zeroing the undef values during
13456 // the shuffle phase or after the shuffle.
13457 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13458 SelectionDAG &DAG) {
13459 MVT RegVT = Op.getSimpleValueType();
13460 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13461 assert(RegVT.isInteger() &&
13462 "We only custom lower integer vector sext loads.");
13464 // Nothing useful we can do without SSE2 shuffles.
13465 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13467 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13469 EVT MemVT = Ld->getMemoryVT();
13470 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13471 unsigned RegSz = RegVT.getSizeInBits();
13473 ISD::LoadExtType Ext = Ld->getExtensionType();
13475 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13476 && "Only anyext and sext are currently implemented.");
13477 assert(MemVT != RegVT && "Cannot extend to the same type");
13478 assert(MemVT.isVector() && "Must load a vector from memory");
13480 unsigned NumElems = RegVT.getVectorNumElements();
13481 unsigned MemSz = MemVT.getSizeInBits();
13482 assert(RegSz > MemSz && "Register size must be greater than the mem size");
13484 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
13485 // The only way in which we have a legal 256-bit vector result but not the
13486 // integer 256-bit operations needed to directly lower a sextload is if we
13487 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
13488 // a 128-bit vector and a normal sign_extend to 256-bits that should get
13489 // correctly legalized. We do this late to allow the canonical form of
13490 // sextload to persist throughout the rest of the DAG combiner -- it wants
13491 // to fold together any extensions it can, and so will fuse a sign_extend
13492 // of an sextload into a sextload targeting a wider value.
13494 if (MemSz == 128) {
13495 // Just switch this to a normal load.
13496 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
13497 "it must be a legal 128-bit vector "
13499 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
13500 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
13501 Ld->isInvariant(), Ld->getAlignment());
13503 assert(MemSz < 128 &&
13504 "Can't extend a type wider than 128 bits to a 256 bit vector!");
13505 // Do an sext load to a 128-bit vector type. We want to use the same
13506 // number of elements, but elements half as wide. This will end up being
13507 // recursively lowered by this routine, but will succeed as we definitely
13508 // have all the necessary features if we're using AVX1.
13510 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
13511 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
13513 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
13514 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
13515 Ld->isNonTemporal(), Ld->isInvariant(),
13516 Ld->getAlignment());
13519 // Replace chain users with the new chain.
13520 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
13521 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
13523 // Finally, do a normal sign-extend to the desired register.
13524 return DAG.getSExtOrTrunc(Load, dl, RegVT);
13527 // All sizes must be a power of two.
13528 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
13529 "Non-power-of-two elements are not custom lowered!");
13531 // Attempt to load the original value using scalar loads.
13532 // Find the largest scalar type that divides the total loaded size.
13533 MVT SclrLoadTy = MVT::i8;
13534 for (MVT Tp : MVT::integer_valuetypes()) {
13535 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
13540 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
13541 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
13543 SclrLoadTy = MVT::f64;
13545 // Calculate the number of scalar loads that we need to perform
13546 // in order to load our vector from memory.
13547 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
13549 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
13550 "Can only lower sext loads with a single scalar load!");
13552 unsigned loadRegZize = RegSz;
13553 if (Ext == ISD::SEXTLOAD && RegSz == 256)
13556 // Represent our vector as a sequence of elements which are the
13557 // largest scalar that we can load.
13558 EVT LoadUnitVecVT = EVT::getVectorVT(
13559 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
13561 // Represent the data using the same element type that is stored in
13562 // memory. In practice, we ''widen'' MemVT.
13564 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
13565 loadRegZize / MemVT.getScalarType().getSizeInBits());
13567 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
13568 "Invalid vector type");
13570 // We can't shuffle using an illegal type.
13571 assert(TLI.isTypeLegal(WideVecVT) &&
13572 "We only lower types that form legal widened vector types");
13574 SmallVector<SDValue, 8> Chains;
13575 SDValue Ptr = Ld->getBasePtr();
13576 SDValue Increment =
13577 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
13578 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
13580 for (unsigned i = 0; i < NumLoads; ++i) {
13581 // Perform a single load.
13582 SDValue ScalarLoad =
13583 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
13584 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
13585 Ld->getAlignment());
13586 Chains.push_back(ScalarLoad.getValue(1));
13587 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
13588 // another round of DAGCombining.
13590 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
13592 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
13593 ScalarLoad, DAG.getIntPtrConstant(i));
13595 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
13598 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
13600 // Bitcast the loaded value to a vector of the original element type, in
13601 // the size of the target vector type.
13602 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
13603 unsigned SizeRatio = RegSz / MemSz;
13605 if (Ext == ISD::SEXTLOAD) {
13606 // If we have SSE4.1, we can directly emit a VSEXT node.
13607 if (Subtarget->hasSSE41()) {
13608 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
13609 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13613 // Otherwise we'll shuffle the small elements in the high bits of the
13614 // larger type and perform an arithmetic shift. If the shift is not legal
13615 // it's better to scalarize.
13616 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
13617 "We can't implement a sext load without an arithmetic right shift!");
13619 // Redistribute the loaded elements into the different locations.
13620 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13621 for (unsigned i = 0; i != NumElems; ++i)
13622 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
13624 SDValue Shuff = DAG.getVectorShuffle(
13625 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13627 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13629 // Build the arithmetic shift.
13630 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
13631 MemVT.getVectorElementType().getSizeInBits();
13633 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
13635 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13639 // Redistribute the loaded elements into the different locations.
13640 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13641 for (unsigned i = 0; i != NumElems; ++i)
13642 ShuffleVec[i * SizeRatio] = i;
13644 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
13645 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13647 // Bitcast to the requested type.
13648 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13649 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13653 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
13654 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
13655 // from the AND / OR.
13656 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
13657 Opc = Op.getOpcode();
13658 if (Opc != ISD::OR && Opc != ISD::AND)
13660 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13661 Op.getOperand(0).hasOneUse() &&
13662 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
13663 Op.getOperand(1).hasOneUse());
13666 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
13667 // 1 and that the SETCC node has a single use.
13668 static bool isXor1OfSetCC(SDValue Op) {
13669 if (Op.getOpcode() != ISD::XOR)
13671 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
13672 if (N1C && N1C->getAPIntValue() == 1) {
13673 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13674 Op.getOperand(0).hasOneUse();
13679 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
13680 bool addTest = true;
13681 SDValue Chain = Op.getOperand(0);
13682 SDValue Cond = Op.getOperand(1);
13683 SDValue Dest = Op.getOperand(2);
13686 bool Inverted = false;
13688 if (Cond.getOpcode() == ISD::SETCC) {
13689 // Check for setcc([su]{add,sub,mul}o == 0).
13690 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
13691 isa<ConstantSDNode>(Cond.getOperand(1)) &&
13692 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
13693 Cond.getOperand(0).getResNo() == 1 &&
13694 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
13695 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
13696 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
13697 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
13698 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
13699 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
13701 Cond = Cond.getOperand(0);
13703 SDValue NewCond = LowerSETCC(Cond, DAG);
13704 if (NewCond.getNode())
13709 // FIXME: LowerXALUO doesn't handle these!!
13710 else if (Cond.getOpcode() == X86ISD::ADD ||
13711 Cond.getOpcode() == X86ISD::SUB ||
13712 Cond.getOpcode() == X86ISD::SMUL ||
13713 Cond.getOpcode() == X86ISD::UMUL)
13714 Cond = LowerXALUO(Cond, DAG);
13717 // Look pass (and (setcc_carry (cmp ...)), 1).
13718 if (Cond.getOpcode() == ISD::AND &&
13719 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13720 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13721 if (C && C->getAPIntValue() == 1)
13722 Cond = Cond.getOperand(0);
13725 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13726 // setting operand in place of the X86ISD::SETCC.
13727 unsigned CondOpcode = Cond.getOpcode();
13728 if (CondOpcode == X86ISD::SETCC ||
13729 CondOpcode == X86ISD::SETCC_CARRY) {
13730 CC = Cond.getOperand(0);
13732 SDValue Cmp = Cond.getOperand(1);
13733 unsigned Opc = Cmp.getOpcode();
13734 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
13735 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
13739 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
13743 // These can only come from an arithmetic instruction with overflow,
13744 // e.g. SADDO, UADDO.
13745 Cond = Cond.getNode()->getOperand(1);
13751 CondOpcode = Cond.getOpcode();
13752 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13753 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13754 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13755 Cond.getOperand(0).getValueType() != MVT::i8)) {
13756 SDValue LHS = Cond.getOperand(0);
13757 SDValue RHS = Cond.getOperand(1);
13758 unsigned X86Opcode;
13761 // Keep this in sync with LowerXALUO, otherwise we might create redundant
13762 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
13764 switch (CondOpcode) {
13765 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13767 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13769 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
13772 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13773 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13775 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13777 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
13780 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13781 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13782 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13783 default: llvm_unreachable("unexpected overflowing operator");
13786 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
13787 if (CondOpcode == ISD::UMULO)
13788 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13791 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13793 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
13795 if (CondOpcode == ISD::UMULO)
13796 Cond = X86Op.getValue(2);
13798 Cond = X86Op.getValue(1);
13800 CC = DAG.getConstant(X86Cond, MVT::i8);
13804 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
13805 SDValue Cmp = Cond.getOperand(0).getOperand(1);
13806 if (CondOpc == ISD::OR) {
13807 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
13808 // two branches instead of an explicit OR instruction with a
13810 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13811 isX86LogicalCmp(Cmp)) {
13812 CC = Cond.getOperand(0).getOperand(0);
13813 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13814 Chain, Dest, CC, Cmp);
13815 CC = Cond.getOperand(1).getOperand(0);
13819 } else { // ISD::AND
13820 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
13821 // two branches instead of an explicit AND instruction with a
13822 // separate test. However, we only do this if this block doesn't
13823 // have a fall-through edge, because this requires an explicit
13824 // jmp when the condition is false.
13825 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13826 isX86LogicalCmp(Cmp) &&
13827 Op.getNode()->hasOneUse()) {
13828 X86::CondCode CCode =
13829 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13830 CCode = X86::GetOppositeBranchCondition(CCode);
13831 CC = DAG.getConstant(CCode, MVT::i8);
13832 SDNode *User = *Op.getNode()->use_begin();
13833 // Look for an unconditional branch following this conditional branch.
13834 // We need this because we need to reverse the successors in order
13835 // to implement FCMP_OEQ.
13836 if (User->getOpcode() == ISD::BR) {
13837 SDValue FalseBB = User->getOperand(1);
13839 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13840 assert(NewBR == User);
13844 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13845 Chain, Dest, CC, Cmp);
13846 X86::CondCode CCode =
13847 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
13848 CCode = X86::GetOppositeBranchCondition(CCode);
13849 CC = DAG.getConstant(CCode, MVT::i8);
13855 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
13856 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
13857 // It should be transformed during dag combiner except when the condition
13858 // is set by a arithmetics with overflow node.
13859 X86::CondCode CCode =
13860 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13861 CCode = X86::GetOppositeBranchCondition(CCode);
13862 CC = DAG.getConstant(CCode, MVT::i8);
13863 Cond = Cond.getOperand(0).getOperand(1);
13865 } else if (Cond.getOpcode() == ISD::SETCC &&
13866 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
13867 // For FCMP_OEQ, we can emit
13868 // two branches instead of an explicit AND instruction with a
13869 // separate test. However, we only do this if this block doesn't
13870 // have a fall-through edge, because this requires an explicit
13871 // jmp when the condition is false.
13872 if (Op.getNode()->hasOneUse()) {
13873 SDNode *User = *Op.getNode()->use_begin();
13874 // Look for an unconditional branch following this conditional branch.
13875 // We need this because we need to reverse the successors in order
13876 // to implement FCMP_OEQ.
13877 if (User->getOpcode() == ISD::BR) {
13878 SDValue FalseBB = User->getOperand(1);
13880 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13881 assert(NewBR == User);
13885 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13886 Cond.getOperand(0), Cond.getOperand(1));
13887 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13888 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13889 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13890 Chain, Dest, CC, Cmp);
13891 CC = DAG.getConstant(X86::COND_P, MVT::i8);
13896 } else if (Cond.getOpcode() == ISD::SETCC &&
13897 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
13898 // For FCMP_UNE, we can emit
13899 // two branches instead of an explicit AND instruction with a
13900 // separate test. However, we only do this if this block doesn't
13901 // have a fall-through edge, because this requires an explicit
13902 // jmp when the condition is false.
13903 if (Op.getNode()->hasOneUse()) {
13904 SDNode *User = *Op.getNode()->use_begin();
13905 // Look for an unconditional branch following this conditional branch.
13906 // We need this because we need to reverse the successors in order
13907 // to implement FCMP_UNE.
13908 if (User->getOpcode() == ISD::BR) {
13909 SDValue FalseBB = User->getOperand(1);
13911 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13912 assert(NewBR == User);
13915 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13916 Cond.getOperand(0), Cond.getOperand(1));
13917 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13918 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13919 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13920 Chain, Dest, CC, Cmp);
13921 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
13931 // Look pass the truncate if the high bits are known zero.
13932 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13933 Cond = Cond.getOperand(0);
13935 // We know the result of AND is compared against zero. Try to match
13937 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13938 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
13939 if (NewSetCC.getNode()) {
13940 CC = NewSetCC.getOperand(0);
13941 Cond = NewSetCC.getOperand(1);
13948 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
13949 CC = DAG.getConstant(X86Cond, MVT::i8);
13950 Cond = EmitTest(Cond, X86Cond, dl, DAG);
13952 Cond = ConvertCmpIfNecessary(Cond, DAG);
13953 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13954 Chain, Dest, CC, Cond);
13957 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
13958 // Calls to _alloca are needed to probe the stack when allocating more than 4k
13959 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
13960 // that the guard pages used by the OS virtual memory manager are allocated in
13961 // correct sequence.
13963 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
13964 SelectionDAG &DAG) const {
13965 MachineFunction &MF = DAG.getMachineFunction();
13966 bool SplitStack = MF.shouldSplitStack();
13967 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
13972 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13973 SDNode* Node = Op.getNode();
13975 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
13976 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
13977 " not tell us which reg is the stack pointer!");
13978 EVT VT = Node->getValueType(0);
13979 SDValue Tmp1 = SDValue(Node, 0);
13980 SDValue Tmp2 = SDValue(Node, 1);
13981 SDValue Tmp3 = Node->getOperand(2);
13982 SDValue Chain = Tmp1.getOperand(0);
13984 // Chain the dynamic stack allocation so that it doesn't modify the stack
13985 // pointer when other instructions are using the stack.
13986 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
13989 SDValue Size = Tmp2.getOperand(1);
13990 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
13991 Chain = SP.getValue(1);
13992 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
13993 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
13994 unsigned StackAlign = TFI.getStackAlignment();
13995 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
13996 if (Align > StackAlign)
13997 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
13998 DAG.getConstant(-(uint64_t)Align, VT));
13999 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14001 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
14002 DAG.getIntPtrConstant(0, true), SDValue(),
14005 SDValue Ops[2] = { Tmp1, Tmp2 };
14006 return DAG.getMergeValues(Ops, dl);
14010 SDValue Chain = Op.getOperand(0);
14011 SDValue Size = Op.getOperand(1);
14012 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14013 EVT VT = Op.getNode()->getValueType(0);
14015 bool Is64Bit = Subtarget->is64Bit();
14016 EVT SPTy = getPointerTy();
14019 MachineRegisterInfo &MRI = MF.getRegInfo();
14022 // The 64 bit implementation of segmented stacks needs to clobber both r10
14023 // r11. This makes it impossible to use it along with nested parameters.
14024 const Function *F = MF.getFunction();
14026 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14028 if (I->hasNestAttr())
14029 report_fatal_error("Cannot use segmented stacks with functions that "
14030 "have nested arguments.");
14033 const TargetRegisterClass *AddrRegClass =
14034 getRegClassFor(getPointerTy());
14035 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14036 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14037 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14038 DAG.getRegister(Vreg, SPTy));
14039 SDValue Ops1[2] = { Value, Chain };
14040 return DAG.getMergeValues(Ops1, dl);
14043 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14045 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14046 Flag = Chain.getValue(1);
14047 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14049 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14051 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14052 unsigned SPReg = RegInfo->getStackRegister();
14053 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14054 Chain = SP.getValue(1);
14057 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14058 DAG.getConstant(-(uint64_t)Align, VT));
14059 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14062 SDValue Ops1[2] = { SP, Chain };
14063 return DAG.getMergeValues(Ops1, dl);
14067 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14068 MachineFunction &MF = DAG.getMachineFunction();
14069 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14071 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14074 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14075 // vastart just stores the address of the VarArgsFrameIndex slot into the
14076 // memory location argument.
14077 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14079 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14080 MachinePointerInfo(SV), false, false, 0);
14084 // gp_offset (0 - 6 * 8)
14085 // fp_offset (48 - 48 + 8 * 16)
14086 // overflow_arg_area (point to parameters coming in memory).
14088 SmallVector<SDValue, 8> MemOps;
14089 SDValue FIN = Op.getOperand(1);
14091 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14092 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14094 FIN, MachinePointerInfo(SV), false, false, 0);
14095 MemOps.push_back(Store);
14098 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14099 FIN, DAG.getIntPtrConstant(4));
14100 Store = DAG.getStore(Op.getOperand(0), DL,
14101 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
14103 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14104 MemOps.push_back(Store);
14106 // Store ptr to overflow_arg_area
14107 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14108 FIN, DAG.getIntPtrConstant(4));
14109 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14111 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14112 MachinePointerInfo(SV, 8),
14114 MemOps.push_back(Store);
14116 // Store ptr to reg_save_area.
14117 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14118 FIN, DAG.getIntPtrConstant(8));
14119 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14121 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14122 MachinePointerInfo(SV, 16), false, false, 0);
14123 MemOps.push_back(Store);
14124 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14127 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14128 assert(Subtarget->is64Bit() &&
14129 "LowerVAARG only handles 64-bit va_arg!");
14130 assert((Subtarget->isTargetLinux() ||
14131 Subtarget->isTargetDarwin()) &&
14132 "Unhandled target in LowerVAARG");
14133 assert(Op.getNode()->getNumOperands() == 4);
14134 SDValue Chain = Op.getOperand(0);
14135 SDValue SrcPtr = Op.getOperand(1);
14136 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14137 unsigned Align = Op.getConstantOperandVal(3);
14140 EVT ArgVT = Op.getNode()->getValueType(0);
14141 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14142 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14145 // Decide which area this value should be read from.
14146 // TODO: Implement the AMD64 ABI in its entirety. This simple
14147 // selection mechanism works only for the basic types.
14148 if (ArgVT == MVT::f80) {
14149 llvm_unreachable("va_arg for f80 not yet implemented");
14150 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14151 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14152 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14153 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14155 llvm_unreachable("Unhandled argument type in LowerVAARG");
14158 if (ArgMode == 2) {
14159 // Sanity Check: Make sure using fp_offset makes sense.
14160 assert(!DAG.getTarget().Options.UseSoftFloat &&
14161 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14162 Attribute::NoImplicitFloat)) &&
14163 Subtarget->hasSSE1());
14166 // Insert VAARG_64 node into the DAG
14167 // VAARG_64 returns two values: Variable Argument Address, Chain
14168 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, MVT::i32),
14169 DAG.getConstant(ArgMode, MVT::i8),
14170 DAG.getConstant(Align, MVT::i32)};
14171 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14172 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14173 VTs, InstOps, MVT::i64,
14174 MachinePointerInfo(SV),
14176 /*Volatile=*/false,
14178 /*WriteMem=*/true);
14179 Chain = VAARG.getValue(1);
14181 // Load the next argument and return it
14182 return DAG.getLoad(ArgVT, dl,
14185 MachinePointerInfo(),
14186 false, false, false, 0);
14189 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14190 SelectionDAG &DAG) {
14191 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14192 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14193 SDValue Chain = Op.getOperand(0);
14194 SDValue DstPtr = Op.getOperand(1);
14195 SDValue SrcPtr = Op.getOperand(2);
14196 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14197 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14200 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14201 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
14203 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14206 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14207 // amount is a constant. Takes immediate version of shift as input.
14208 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14209 SDValue SrcOp, uint64_t ShiftAmt,
14210 SelectionDAG &DAG) {
14211 MVT ElementType = VT.getVectorElementType();
14213 // Fold this packed shift into its first operand if ShiftAmt is 0.
14217 // Check for ShiftAmt >= element width
14218 if (ShiftAmt >= ElementType.getSizeInBits()) {
14219 if (Opc == X86ISD::VSRAI)
14220 ShiftAmt = ElementType.getSizeInBits() - 1;
14222 return DAG.getConstant(0, VT);
14225 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14226 && "Unknown target vector shift-by-constant node");
14228 // Fold this packed vector shift into a build vector if SrcOp is a
14229 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14230 if (VT == SrcOp.getSimpleValueType() &&
14231 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14232 SmallVector<SDValue, 8> Elts;
14233 unsigned NumElts = SrcOp->getNumOperands();
14234 ConstantSDNode *ND;
14237 default: llvm_unreachable(nullptr);
14238 case X86ISD::VSHLI:
14239 for (unsigned i=0; i!=NumElts; ++i) {
14240 SDValue CurrentOp = SrcOp->getOperand(i);
14241 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14242 Elts.push_back(CurrentOp);
14245 ND = cast<ConstantSDNode>(CurrentOp);
14246 const APInt &C = ND->getAPIntValue();
14247 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
14250 case X86ISD::VSRLI:
14251 for (unsigned i=0; i!=NumElts; ++i) {
14252 SDValue CurrentOp = SrcOp->getOperand(i);
14253 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14254 Elts.push_back(CurrentOp);
14257 ND = cast<ConstantSDNode>(CurrentOp);
14258 const APInt &C = ND->getAPIntValue();
14259 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
14262 case X86ISD::VSRAI:
14263 for (unsigned i=0; i!=NumElts; ++i) {
14264 SDValue CurrentOp = SrcOp->getOperand(i);
14265 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14266 Elts.push_back(CurrentOp);
14269 ND = cast<ConstantSDNode>(CurrentOp);
14270 const APInt &C = ND->getAPIntValue();
14271 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
14276 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14279 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
14282 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14283 // may or may not be a constant. Takes immediate version of shift as input.
14284 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14285 SDValue SrcOp, SDValue ShAmt,
14286 SelectionDAG &DAG) {
14287 MVT SVT = ShAmt.getSimpleValueType();
14288 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14290 // Catch shift-by-constant.
14291 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14292 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14293 CShAmt->getZExtValue(), DAG);
14295 // Change opcode to non-immediate version
14297 default: llvm_unreachable("Unknown target vector shift node");
14298 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14299 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14300 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14303 const X86Subtarget &Subtarget =
14304 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14305 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14306 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14307 // Let the shuffle legalizer expand this shift amount node.
14308 SDValue Op0 = ShAmt.getOperand(0);
14309 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14310 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14312 // Need to build a vector containing shift amount.
14313 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14314 SmallVector<SDValue, 4> ShOps;
14315 ShOps.push_back(ShAmt);
14316 if (SVT == MVT::i32) {
14317 ShOps.push_back(DAG.getConstant(0, SVT));
14318 ShOps.push_back(DAG.getUNDEF(SVT));
14320 ShOps.push_back(DAG.getUNDEF(SVT));
14322 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14323 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14326 // The return type has to be a 128-bit type with the same element
14327 // type as the input type.
14328 MVT EltVT = VT.getVectorElementType();
14329 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14331 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14332 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14335 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14336 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14337 /// necessary casting for \p Mask when lowering masking intrinsics.
14338 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14339 SDValue PreservedSrc,
14340 const X86Subtarget *Subtarget,
14341 SelectionDAG &DAG) {
14342 EVT VT = Op.getValueType();
14343 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14344 MVT::i1, VT.getVectorNumElements());
14345 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14346 Mask.getValueType().getSizeInBits());
14349 assert(MaskVT.isSimple() && "invalid mask type");
14351 if (isAllOnes(Mask))
14354 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14355 // are extracted by EXTRACT_SUBVECTOR.
14356 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14357 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14358 DAG.getIntPtrConstant(0));
14360 switch (Op.getOpcode()) {
14362 case X86ISD::PCMPEQM:
14363 case X86ISD::PCMPGTM:
14365 case X86ISD::CMPMU:
14366 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14368 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14369 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14370 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14373 /// \brief Creates an SDNode for a predicated scalar operation.
14374 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14375 /// The mask is comming as MVT::i8 and it should be truncated
14376 /// to MVT::i1 while lowering masking intrinsics.
14377 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14378 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14379 /// a scalar instruction.
14380 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14381 SDValue PreservedSrc,
14382 const X86Subtarget *Subtarget,
14383 SelectionDAG &DAG) {
14384 if (isAllOnes(Mask))
14387 EVT VT = Op.getValueType();
14389 // The mask should be of type MVT::i1
14390 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14392 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14393 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14394 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14397 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14398 SelectionDAG &DAG) {
14400 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14401 EVT VT = Op.getValueType();
14402 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14404 switch(IntrData->Type) {
14405 case INTR_TYPE_1OP:
14406 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14407 case INTR_TYPE_2OP:
14408 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14410 case INTR_TYPE_3OP:
14411 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14412 Op.getOperand(2), Op.getOperand(3));
14413 case INTR_TYPE_1OP_MASK_RM: {
14414 SDValue Src = Op.getOperand(1);
14415 SDValue Src0 = Op.getOperand(2);
14416 SDValue Mask = Op.getOperand(3);
14417 SDValue RoundingMode = Op.getOperand(4);
14418 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14420 Mask, Src0, Subtarget, DAG);
14422 case INTR_TYPE_SCALAR_MASK_RM: {
14423 SDValue Src1 = Op.getOperand(1);
14424 SDValue Src2 = Op.getOperand(2);
14425 SDValue Src0 = Op.getOperand(3);
14426 SDValue Mask = Op.getOperand(4);
14427 SDValue RoundingMode = Op.getOperand(5);
14428 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14430 Mask, Src0, Subtarget, DAG);
14432 case INTR_TYPE_2OP_MASK: {
14433 SDValue Src1 = Op.getOperand(1);
14434 SDValue Src2 = Op.getOperand(2);
14435 SDValue PassThru = Op.getOperand(3);
14436 SDValue Mask = Op.getOperand(4);
14437 // We specify 2 possible opcodes for intrinsics with rounding modes.
14438 // First, we check if the intrinsic may have non-default rounding mode,
14439 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14440 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14441 if (IntrWithRoundingModeOpcode != 0) {
14442 SDValue Rnd = Op.getOperand(5);
14443 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14444 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14445 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14446 dl, Op.getValueType(),
14448 Mask, PassThru, Subtarget, DAG);
14451 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14453 Mask, PassThru, Subtarget, DAG);
14455 case FMA_OP_MASK: {
14456 SDValue Src1 = Op.getOperand(1);
14457 SDValue Src2 = Op.getOperand(2);
14458 SDValue Src3 = Op.getOperand(3);
14459 SDValue Mask = Op.getOperand(4);
14460 // We specify 2 possible opcodes for intrinsics with rounding modes.
14461 // First, we check if the intrinsic may have non-default rounding mode,
14462 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14463 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14464 if (IntrWithRoundingModeOpcode != 0) {
14465 SDValue Rnd = Op.getOperand(5);
14466 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14467 X86::STATIC_ROUNDING::CUR_DIRECTION)
14468 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14469 dl, Op.getValueType(),
14470 Src1, Src2, Src3, Rnd),
14471 Mask, Src1, Subtarget, DAG);
14473 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
14474 dl, Op.getValueType(),
14476 Mask, Src1, Subtarget, DAG);
14479 case CMP_MASK_CC: {
14480 // Comparison intrinsics with masks.
14481 // Example of transformation:
14482 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
14483 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
14485 // (v8i1 (insert_subvector undef,
14486 // (v2i1 (and (PCMPEQM %a, %b),
14487 // (extract_subvector
14488 // (v8i1 (bitcast %mask)), 0))), 0))))
14489 EVT VT = Op.getOperand(1).getValueType();
14490 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14491 VT.getVectorNumElements());
14492 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
14493 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14494 Mask.getValueType().getSizeInBits());
14496 if (IntrData->Type == CMP_MASK_CC) {
14497 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14498 Op.getOperand(2), Op.getOperand(3));
14500 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
14501 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14504 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
14505 DAG.getTargetConstant(0, MaskVT),
14507 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
14508 DAG.getUNDEF(BitcastVT), CmpMask,
14509 DAG.getIntPtrConstant(0));
14510 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
14512 case COMI: { // Comparison intrinsics
14513 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
14514 SDValue LHS = Op.getOperand(1);
14515 SDValue RHS = Op.getOperand(2);
14516 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
14517 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
14518 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
14519 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14520 DAG.getConstant(X86CC, MVT::i8), Cond);
14521 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14524 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
14525 Op.getOperand(1), Op.getOperand(2), DAG);
14527 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
14528 Op.getSimpleValueType(),
14530 Op.getOperand(2), DAG),
14531 Op.getOperand(4), Op.getOperand(3), Subtarget,
14533 case COMPRESS_EXPAND_IN_REG: {
14534 SDValue Mask = Op.getOperand(3);
14535 SDValue DataToCompress = Op.getOperand(1);
14536 SDValue PassThru = Op.getOperand(2);
14537 if (isAllOnes(Mask)) // return data as is
14538 return Op.getOperand(1);
14539 EVT VT = Op.getValueType();
14540 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14541 VT.getVectorNumElements());
14542 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14543 Mask.getValueType().getSizeInBits());
14545 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14546 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14547 DAG.getIntPtrConstant(0));
14549 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
14553 SDValue Mask = Op.getOperand(3);
14554 EVT VT = Op.getValueType();
14555 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14556 VT.getVectorNumElements());
14557 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14558 Mask.getValueType().getSizeInBits());
14560 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14561 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14562 DAG.getIntPtrConstant(0));
14563 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
14572 default: return SDValue(); // Don't custom lower most intrinsics.
14574 case Intrinsic::x86_avx512_mask_valign_q_512:
14575 case Intrinsic::x86_avx512_mask_valign_d_512:
14576 // Vector source operands are swapped.
14577 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
14578 Op.getValueType(), Op.getOperand(2),
14581 Op.getOperand(5), Op.getOperand(4),
14584 // ptest and testp intrinsics. The intrinsic these come from are designed to
14585 // return an integer value, not just an instruction so lower it to the ptest
14586 // or testp pattern and a setcc for the result.
14587 case Intrinsic::x86_sse41_ptestz:
14588 case Intrinsic::x86_sse41_ptestc:
14589 case Intrinsic::x86_sse41_ptestnzc:
14590 case Intrinsic::x86_avx_ptestz_256:
14591 case Intrinsic::x86_avx_ptestc_256:
14592 case Intrinsic::x86_avx_ptestnzc_256:
14593 case Intrinsic::x86_avx_vtestz_ps:
14594 case Intrinsic::x86_avx_vtestc_ps:
14595 case Intrinsic::x86_avx_vtestnzc_ps:
14596 case Intrinsic::x86_avx_vtestz_pd:
14597 case Intrinsic::x86_avx_vtestc_pd:
14598 case Intrinsic::x86_avx_vtestnzc_pd:
14599 case Intrinsic::x86_avx_vtestz_ps_256:
14600 case Intrinsic::x86_avx_vtestc_ps_256:
14601 case Intrinsic::x86_avx_vtestnzc_ps_256:
14602 case Intrinsic::x86_avx_vtestz_pd_256:
14603 case Intrinsic::x86_avx_vtestc_pd_256:
14604 case Intrinsic::x86_avx_vtestnzc_pd_256: {
14605 bool IsTestPacked = false;
14608 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
14609 case Intrinsic::x86_avx_vtestz_ps:
14610 case Intrinsic::x86_avx_vtestz_pd:
14611 case Intrinsic::x86_avx_vtestz_ps_256:
14612 case Intrinsic::x86_avx_vtestz_pd_256:
14613 IsTestPacked = true; // Fallthrough
14614 case Intrinsic::x86_sse41_ptestz:
14615 case Intrinsic::x86_avx_ptestz_256:
14617 X86CC = X86::COND_E;
14619 case Intrinsic::x86_avx_vtestc_ps:
14620 case Intrinsic::x86_avx_vtestc_pd:
14621 case Intrinsic::x86_avx_vtestc_ps_256:
14622 case Intrinsic::x86_avx_vtestc_pd_256:
14623 IsTestPacked = true; // Fallthrough
14624 case Intrinsic::x86_sse41_ptestc:
14625 case Intrinsic::x86_avx_ptestc_256:
14627 X86CC = X86::COND_B;
14629 case Intrinsic::x86_avx_vtestnzc_ps:
14630 case Intrinsic::x86_avx_vtestnzc_pd:
14631 case Intrinsic::x86_avx_vtestnzc_ps_256:
14632 case Intrinsic::x86_avx_vtestnzc_pd_256:
14633 IsTestPacked = true; // Fallthrough
14634 case Intrinsic::x86_sse41_ptestnzc:
14635 case Intrinsic::x86_avx_ptestnzc_256:
14637 X86CC = X86::COND_A;
14641 SDValue LHS = Op.getOperand(1);
14642 SDValue RHS = Op.getOperand(2);
14643 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
14644 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
14645 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14646 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
14647 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14649 case Intrinsic::x86_avx512_kortestz_w:
14650 case Intrinsic::x86_avx512_kortestc_w: {
14651 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
14652 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
14653 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
14654 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14655 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
14656 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
14657 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14660 case Intrinsic::x86_sse42_pcmpistria128:
14661 case Intrinsic::x86_sse42_pcmpestria128:
14662 case Intrinsic::x86_sse42_pcmpistric128:
14663 case Intrinsic::x86_sse42_pcmpestric128:
14664 case Intrinsic::x86_sse42_pcmpistrio128:
14665 case Intrinsic::x86_sse42_pcmpestrio128:
14666 case Intrinsic::x86_sse42_pcmpistris128:
14667 case Intrinsic::x86_sse42_pcmpestris128:
14668 case Intrinsic::x86_sse42_pcmpistriz128:
14669 case Intrinsic::x86_sse42_pcmpestriz128: {
14673 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14674 case Intrinsic::x86_sse42_pcmpistria128:
14675 Opcode = X86ISD::PCMPISTRI;
14676 X86CC = X86::COND_A;
14678 case Intrinsic::x86_sse42_pcmpestria128:
14679 Opcode = X86ISD::PCMPESTRI;
14680 X86CC = X86::COND_A;
14682 case Intrinsic::x86_sse42_pcmpistric128:
14683 Opcode = X86ISD::PCMPISTRI;
14684 X86CC = X86::COND_B;
14686 case Intrinsic::x86_sse42_pcmpestric128:
14687 Opcode = X86ISD::PCMPESTRI;
14688 X86CC = X86::COND_B;
14690 case Intrinsic::x86_sse42_pcmpistrio128:
14691 Opcode = X86ISD::PCMPISTRI;
14692 X86CC = X86::COND_O;
14694 case Intrinsic::x86_sse42_pcmpestrio128:
14695 Opcode = X86ISD::PCMPESTRI;
14696 X86CC = X86::COND_O;
14698 case Intrinsic::x86_sse42_pcmpistris128:
14699 Opcode = X86ISD::PCMPISTRI;
14700 X86CC = X86::COND_S;
14702 case Intrinsic::x86_sse42_pcmpestris128:
14703 Opcode = X86ISD::PCMPESTRI;
14704 X86CC = X86::COND_S;
14706 case Intrinsic::x86_sse42_pcmpistriz128:
14707 Opcode = X86ISD::PCMPISTRI;
14708 X86CC = X86::COND_E;
14710 case Intrinsic::x86_sse42_pcmpestriz128:
14711 Opcode = X86ISD::PCMPESTRI;
14712 X86CC = X86::COND_E;
14715 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14716 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14717 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
14718 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14719 DAG.getConstant(X86CC, MVT::i8),
14720 SDValue(PCMP.getNode(), 1));
14721 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14724 case Intrinsic::x86_sse42_pcmpistri128:
14725 case Intrinsic::x86_sse42_pcmpestri128: {
14727 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
14728 Opcode = X86ISD::PCMPISTRI;
14730 Opcode = X86ISD::PCMPESTRI;
14732 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14733 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14734 return DAG.getNode(Opcode, dl, VTs, NewOps);
14739 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14740 SDValue Src, SDValue Mask, SDValue Base,
14741 SDValue Index, SDValue ScaleOp, SDValue Chain,
14742 const X86Subtarget * Subtarget) {
14744 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14745 assert(C && "Invalid scale type");
14746 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14747 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14748 Index.getSimpleValueType().getVectorNumElements());
14750 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14752 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14754 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14755 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
14756 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14757 SDValue Segment = DAG.getRegister(0, MVT::i32);
14758 if (Src.getOpcode() == ISD::UNDEF)
14759 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
14760 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14761 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14762 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
14763 return DAG.getMergeValues(RetOps, dl);
14766 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14767 SDValue Src, SDValue Mask, SDValue Base,
14768 SDValue Index, SDValue ScaleOp, SDValue Chain) {
14770 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14771 assert(C && "Invalid scale type");
14772 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14773 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14774 SDValue Segment = DAG.getRegister(0, MVT::i32);
14775 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14776 Index.getSimpleValueType().getVectorNumElements());
14778 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14780 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14782 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14783 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
14784 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
14785 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14786 return SDValue(Res, 1);
14789 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14790 SDValue Mask, SDValue Base, SDValue Index,
14791 SDValue ScaleOp, SDValue Chain) {
14793 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14794 assert(C && "Invalid scale type");
14795 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14796 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14797 SDValue Segment = DAG.getRegister(0, MVT::i32);
14799 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
14801 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14803 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14805 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14806 //SDVTList VTs = DAG.getVTList(MVT::Other);
14807 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14808 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
14809 return SDValue(Res, 0);
14812 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
14813 // read performance monitor counters (x86_rdpmc).
14814 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
14815 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14816 SmallVectorImpl<SDValue> &Results) {
14817 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14818 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14821 // The ECX register is used to select the index of the performance counter
14823 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
14825 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
14827 // Reads the content of a 64-bit performance counter and returns it in the
14828 // registers EDX:EAX.
14829 if (Subtarget->is64Bit()) {
14830 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14831 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14834 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14835 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14838 Chain = HI.getValue(1);
14840 if (Subtarget->is64Bit()) {
14841 // The EAX register is loaded with the low-order 32 bits. The EDX register
14842 // is loaded with the supported high-order bits of the counter.
14843 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14844 DAG.getConstant(32, MVT::i8));
14845 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14846 Results.push_back(Chain);
14850 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14851 SDValue Ops[] = { LO, HI };
14852 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14853 Results.push_back(Pair);
14854 Results.push_back(Chain);
14857 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
14858 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
14859 // also used to custom lower READCYCLECOUNTER nodes.
14860 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
14861 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14862 SmallVectorImpl<SDValue> &Results) {
14863 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14864 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
14867 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
14868 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
14869 // and the EAX register is loaded with the low-order 32 bits.
14870 if (Subtarget->is64Bit()) {
14871 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14872 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14875 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14876 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14879 SDValue Chain = HI.getValue(1);
14881 if (Opcode == X86ISD::RDTSCP_DAG) {
14882 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14884 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
14885 // the ECX register. Add 'ecx' explicitly to the chain.
14886 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
14888 // Explicitly store the content of ECX at the location passed in input
14889 // to the 'rdtscp' intrinsic.
14890 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
14891 MachinePointerInfo(), false, false, 0);
14894 if (Subtarget->is64Bit()) {
14895 // The EDX register is loaded with the high-order 32 bits of the MSR, and
14896 // the EAX register is loaded with the low-order 32 bits.
14897 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14898 DAG.getConstant(32, MVT::i8));
14899 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14900 Results.push_back(Chain);
14904 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14905 SDValue Ops[] = { LO, HI };
14906 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14907 Results.push_back(Pair);
14908 Results.push_back(Chain);
14911 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
14912 SelectionDAG &DAG) {
14913 SmallVector<SDValue, 2> Results;
14915 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
14917 return DAG.getMergeValues(Results, DL);
14921 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14922 SelectionDAG &DAG) {
14923 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
14925 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
14930 switch(IntrData->Type) {
14932 llvm_unreachable("Unknown Intrinsic Type");
14936 // Emit the node with the right value type.
14937 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
14938 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
14940 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
14941 // Otherwise return the value from Rand, which is always 0, casted to i32.
14942 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
14943 DAG.getConstant(1, Op->getValueType(1)),
14944 DAG.getConstant(X86::COND_B, MVT::i32),
14945 SDValue(Result.getNode(), 1) };
14946 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
14947 DAG.getVTList(Op->getValueType(1), MVT::Glue),
14950 // Return { result, isValid, chain }.
14951 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
14952 SDValue(Result.getNode(), 2));
14955 //gather(v1, mask, index, base, scale);
14956 SDValue Chain = Op.getOperand(0);
14957 SDValue Src = Op.getOperand(2);
14958 SDValue Base = Op.getOperand(3);
14959 SDValue Index = Op.getOperand(4);
14960 SDValue Mask = Op.getOperand(5);
14961 SDValue Scale = Op.getOperand(6);
14962 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
14966 //scatter(base, mask, index, v1, scale);
14967 SDValue Chain = Op.getOperand(0);
14968 SDValue Base = Op.getOperand(2);
14969 SDValue Mask = Op.getOperand(3);
14970 SDValue Index = Op.getOperand(4);
14971 SDValue Src = Op.getOperand(5);
14972 SDValue Scale = Op.getOperand(6);
14973 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
14976 SDValue Hint = Op.getOperand(6);
14978 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
14979 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
14980 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
14981 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
14982 SDValue Chain = Op.getOperand(0);
14983 SDValue Mask = Op.getOperand(2);
14984 SDValue Index = Op.getOperand(3);
14985 SDValue Base = Op.getOperand(4);
14986 SDValue Scale = Op.getOperand(5);
14987 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
14989 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
14991 SmallVector<SDValue, 2> Results;
14992 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
14993 return DAG.getMergeValues(Results, dl);
14995 // Read Performance Monitoring Counters.
14997 SmallVector<SDValue, 2> Results;
14998 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
14999 return DAG.getMergeValues(Results, dl);
15001 // XTEST intrinsics.
15003 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15004 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15005 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15006 DAG.getConstant(X86::COND_NE, MVT::i8),
15008 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15009 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15010 Ret, SDValue(InTrans.getNode(), 1));
15014 SmallVector<SDValue, 2> Results;
15015 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15016 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15017 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15018 DAG.getConstant(-1, MVT::i8));
15019 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15020 Op.getOperand(4), GenCF.getValue(1));
15021 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15022 Op.getOperand(5), MachinePointerInfo(),
15024 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15025 DAG.getConstant(X86::COND_B, MVT::i8),
15027 Results.push_back(SetCC);
15028 Results.push_back(Store);
15029 return DAG.getMergeValues(Results, dl);
15031 case COMPRESS_TO_MEM: {
15033 SDValue Mask = Op.getOperand(4);
15034 SDValue DataToCompress = Op.getOperand(3);
15035 SDValue Addr = Op.getOperand(2);
15036 SDValue Chain = Op.getOperand(0);
15038 if (isAllOnes(Mask)) // return just a store
15039 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15040 MachinePointerInfo(), false, false, 0);
15042 EVT VT = DataToCompress.getValueType();
15043 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15044 VT.getVectorNumElements());
15045 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15046 Mask.getValueType().getSizeInBits());
15047 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15048 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15049 DAG.getIntPtrConstant(0));
15051 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15052 DataToCompress, DAG.getUNDEF(VT));
15053 return DAG.getStore(Chain, dl, Compressed, Addr,
15054 MachinePointerInfo(), false, false, 0);
15056 case EXPAND_FROM_MEM: {
15058 SDValue Mask = Op.getOperand(4);
15059 SDValue PathThru = Op.getOperand(3);
15060 SDValue Addr = Op.getOperand(2);
15061 SDValue Chain = Op.getOperand(0);
15062 EVT VT = Op.getValueType();
15064 if (isAllOnes(Mask)) // return just a load
15065 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15067 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15068 VT.getVectorNumElements());
15069 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15070 Mask.getValueType().getSizeInBits());
15071 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15072 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15073 DAG.getIntPtrConstant(0));
15075 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15076 false, false, false, 0);
15078 SDValue Results[] = {
15079 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15081 return DAG.getMergeValues(Results, dl);
15086 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15087 SelectionDAG &DAG) const {
15088 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15089 MFI->setReturnAddressIsTaken(true);
15091 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15094 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15096 EVT PtrVT = getPointerTy();
15099 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15100 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15101 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
15102 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15103 DAG.getNode(ISD::ADD, dl, PtrVT,
15104 FrameAddr, Offset),
15105 MachinePointerInfo(), false, false, false, 0);
15108 // Just load the return address.
15109 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15110 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15111 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15114 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15115 MachineFunction &MF = DAG.getMachineFunction();
15116 MachineFrameInfo *MFI = MF.getFrameInfo();
15117 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15118 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15119 EVT VT = Op.getValueType();
15121 MFI->setFrameAddressIsTaken(true);
15123 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15124 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15125 // is not possible to crawl up the stack without looking at the unwind codes
15127 int FrameAddrIndex = FuncInfo->getFAIndex();
15128 if (!FrameAddrIndex) {
15129 // Set up a frame object for the return address.
15130 unsigned SlotSize = RegInfo->getSlotSize();
15131 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15132 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
15133 FuncInfo->setFAIndex(FrameAddrIndex);
15135 return DAG.getFrameIndex(FrameAddrIndex, VT);
15138 unsigned FrameReg =
15139 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15140 SDLoc dl(Op); // FIXME probably not meaningful
15141 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15142 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15143 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15144 "Invalid Frame Register!");
15145 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15147 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15148 MachinePointerInfo(),
15149 false, false, false, 0);
15153 // FIXME? Maybe this could be a TableGen attribute on some registers and
15154 // this table could be generated automatically from RegInfo.
15155 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15157 unsigned Reg = StringSwitch<unsigned>(RegName)
15158 .Case("esp", X86::ESP)
15159 .Case("rsp", X86::RSP)
15163 report_fatal_error("Invalid register name global variable");
15166 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15167 SelectionDAG &DAG) const {
15168 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15169 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
15172 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15173 SDValue Chain = Op.getOperand(0);
15174 SDValue Offset = Op.getOperand(1);
15175 SDValue Handler = Op.getOperand(2);
15178 EVT PtrVT = getPointerTy();
15179 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15180 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15181 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15182 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15183 "Invalid Frame Register!");
15184 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15185 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15187 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15188 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
15189 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15190 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15192 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15194 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15195 DAG.getRegister(StoreAddrReg, PtrVT));
15198 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15199 SelectionDAG &DAG) const {
15201 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15202 DAG.getVTList(MVT::i32, MVT::Other),
15203 Op.getOperand(0), Op.getOperand(1));
15206 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15207 SelectionDAG &DAG) const {
15209 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15210 Op.getOperand(0), Op.getOperand(1));
15213 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15214 return Op.getOperand(0);
15217 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15218 SelectionDAG &DAG) const {
15219 SDValue Root = Op.getOperand(0);
15220 SDValue Trmp = Op.getOperand(1); // trampoline
15221 SDValue FPtr = Op.getOperand(2); // nested function
15222 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15225 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15226 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15228 if (Subtarget->is64Bit()) {
15229 SDValue OutChains[6];
15231 // Large code-model.
15232 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15233 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15235 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15236 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15238 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15240 // Load the pointer to the nested function into R11.
15241 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15242 SDValue Addr = Trmp;
15243 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15244 Addr, MachinePointerInfo(TrmpAddr),
15247 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15248 DAG.getConstant(2, MVT::i64));
15249 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15250 MachinePointerInfo(TrmpAddr, 2),
15253 // Load the 'nest' parameter value into R10.
15254 // R10 is specified in X86CallingConv.td
15255 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15256 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15257 DAG.getConstant(10, MVT::i64));
15258 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15259 Addr, MachinePointerInfo(TrmpAddr, 10),
15262 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15263 DAG.getConstant(12, MVT::i64));
15264 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15265 MachinePointerInfo(TrmpAddr, 12),
15268 // Jump to the nested function.
15269 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15270 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15271 DAG.getConstant(20, MVT::i64));
15272 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15273 Addr, MachinePointerInfo(TrmpAddr, 20),
15276 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15277 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15278 DAG.getConstant(22, MVT::i64));
15279 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
15280 MachinePointerInfo(TrmpAddr, 22),
15283 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15285 const Function *Func =
15286 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15287 CallingConv::ID CC = Func->getCallingConv();
15292 llvm_unreachable("Unsupported calling convention");
15293 case CallingConv::C:
15294 case CallingConv::X86_StdCall: {
15295 // Pass 'nest' parameter in ECX.
15296 // Must be kept in sync with X86CallingConv.td
15297 NestReg = X86::ECX;
15299 // Check that ECX wasn't needed by an 'inreg' parameter.
15300 FunctionType *FTy = Func->getFunctionType();
15301 const AttributeSet &Attrs = Func->getAttributes();
15303 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15304 unsigned InRegCount = 0;
15307 for (FunctionType::param_iterator I = FTy->param_begin(),
15308 E = FTy->param_end(); I != E; ++I, ++Idx)
15309 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15310 // FIXME: should only count parameters that are lowered to integers.
15311 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15313 if (InRegCount > 2) {
15314 report_fatal_error("Nest register in use - reduce number of inreg"
15320 case CallingConv::X86_FastCall:
15321 case CallingConv::X86_ThisCall:
15322 case CallingConv::Fast:
15323 // Pass 'nest' parameter in EAX.
15324 // Must be kept in sync with X86CallingConv.td
15325 NestReg = X86::EAX;
15329 SDValue OutChains[4];
15330 SDValue Addr, Disp;
15332 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15333 DAG.getConstant(10, MVT::i32));
15334 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15336 // This is storing the opcode for MOV32ri.
15337 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15338 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15339 OutChains[0] = DAG.getStore(Root, dl,
15340 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
15341 Trmp, MachinePointerInfo(TrmpAddr),
15344 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15345 DAG.getConstant(1, MVT::i32));
15346 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15347 MachinePointerInfo(TrmpAddr, 1),
15350 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15351 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15352 DAG.getConstant(5, MVT::i32));
15353 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
15354 MachinePointerInfo(TrmpAddr, 5),
15357 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15358 DAG.getConstant(6, MVT::i32));
15359 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15360 MachinePointerInfo(TrmpAddr, 6),
15363 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15367 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15368 SelectionDAG &DAG) const {
15370 The rounding mode is in bits 11:10 of FPSR, and has the following
15372 00 Round to nearest
15377 FLT_ROUNDS, on the other hand, expects the following:
15384 To perform the conversion, we do:
15385 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15388 MachineFunction &MF = DAG.getMachineFunction();
15389 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15390 unsigned StackAlignment = TFI.getStackAlignment();
15391 MVT VT = Op.getSimpleValueType();
15394 // Save FP Control Word to stack slot
15395 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15396 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15398 MachineMemOperand *MMO =
15399 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15400 MachineMemOperand::MOStore, 2, 2);
15402 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15403 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15404 DAG.getVTList(MVT::Other),
15405 Ops, MVT::i16, MMO);
15407 // Load FP Control Word from stack slot
15408 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15409 MachinePointerInfo(), false, false, false, 0);
15411 // Transform as necessary
15413 DAG.getNode(ISD::SRL, DL, MVT::i16,
15414 DAG.getNode(ISD::AND, DL, MVT::i16,
15415 CWD, DAG.getConstant(0x800, MVT::i16)),
15416 DAG.getConstant(11, MVT::i8));
15418 DAG.getNode(ISD::SRL, DL, MVT::i16,
15419 DAG.getNode(ISD::AND, DL, MVT::i16,
15420 CWD, DAG.getConstant(0x400, MVT::i16)),
15421 DAG.getConstant(9, MVT::i8));
15424 DAG.getNode(ISD::AND, DL, MVT::i16,
15425 DAG.getNode(ISD::ADD, DL, MVT::i16,
15426 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
15427 DAG.getConstant(1, MVT::i16)),
15428 DAG.getConstant(3, MVT::i16));
15430 return DAG.getNode((VT.getSizeInBits() < 16 ?
15431 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
15434 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
15435 MVT VT = Op.getSimpleValueType();
15437 unsigned NumBits = VT.getSizeInBits();
15440 Op = Op.getOperand(0);
15441 if (VT == MVT::i8) {
15442 // Zero extend to i32 since there is not an i8 bsr.
15444 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15447 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
15448 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15449 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15451 // If src is zero (i.e. bsr sets ZF), returns NumBits.
15454 DAG.getConstant(NumBits+NumBits-1, OpVT),
15455 DAG.getConstant(X86::COND_E, MVT::i8),
15458 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
15460 // Finally xor with NumBits-1.
15461 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15464 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15468 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
15469 MVT VT = Op.getSimpleValueType();
15471 unsigned NumBits = VT.getSizeInBits();
15474 Op = Op.getOperand(0);
15475 if (VT == MVT::i8) {
15476 // Zero extend to i32 since there is not an i8 bsr.
15478 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15481 // Issue a bsr (scan bits in reverse).
15482 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15483 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15485 // And xor with NumBits-1.
15486 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15489 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15493 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
15494 MVT VT = Op.getSimpleValueType();
15495 unsigned NumBits = VT.getSizeInBits();
15497 Op = Op.getOperand(0);
15499 // Issue a bsf (scan bits forward) which also sets EFLAGS.
15500 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
15501 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
15503 // If src is zero (i.e. bsf sets ZF), returns NumBits.
15506 DAG.getConstant(NumBits, VT),
15507 DAG.getConstant(X86::COND_E, MVT::i8),
15510 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
15513 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
15514 // ones, and then concatenate the result back.
15515 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
15516 MVT VT = Op.getSimpleValueType();
15518 assert(VT.is256BitVector() && VT.isInteger() &&
15519 "Unsupported value type for operation");
15521 unsigned NumElems = VT.getVectorNumElements();
15524 // Extract the LHS vectors
15525 SDValue LHS = Op.getOperand(0);
15526 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15527 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15529 // Extract the RHS vectors
15530 SDValue RHS = Op.getOperand(1);
15531 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15532 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15534 MVT EltVT = VT.getVectorElementType();
15535 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15537 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15538 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15539 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15542 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15543 assert(Op.getSimpleValueType().is256BitVector() &&
15544 Op.getSimpleValueType().isInteger() &&
15545 "Only handle AVX 256-bit vector integer operation");
15546 return Lower256IntArith(Op, DAG);
15549 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15550 assert(Op.getSimpleValueType().is256BitVector() &&
15551 Op.getSimpleValueType().isInteger() &&
15552 "Only handle AVX 256-bit vector integer operation");
15553 return Lower256IntArith(Op, DAG);
15556 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
15557 SelectionDAG &DAG) {
15559 MVT VT = Op.getSimpleValueType();
15561 // Decompose 256-bit ops into smaller 128-bit ops.
15562 if (VT.is256BitVector() && !Subtarget->hasInt256())
15563 return Lower256IntArith(Op, DAG);
15565 SDValue A = Op.getOperand(0);
15566 SDValue B = Op.getOperand(1);
15568 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
15569 if (VT == MVT::v4i32) {
15570 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
15571 "Should not custom lower when pmuldq is available!");
15573 // Extract the odd parts.
15574 static const int UnpackMask[] = { 1, -1, 3, -1 };
15575 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
15576 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
15578 // Multiply the even parts.
15579 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
15580 // Now multiply odd parts.
15581 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
15583 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
15584 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
15586 // Merge the two vectors back together with a shuffle. This expands into 2
15588 static const int ShufMask[] = { 0, 4, 2, 6 };
15589 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
15592 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
15593 "Only know how to lower V2I64/V4I64/V8I64 multiply");
15595 // Ahi = psrlqi(a, 32);
15596 // Bhi = psrlqi(b, 32);
15598 // AloBlo = pmuludq(a, b);
15599 // AloBhi = pmuludq(a, Bhi);
15600 // AhiBlo = pmuludq(Ahi, b);
15602 // AloBhi = psllqi(AloBhi, 32);
15603 // AhiBlo = psllqi(AhiBlo, 32);
15604 // return AloBlo + AloBhi + AhiBlo;
15606 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
15607 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
15609 // Bit cast to 32-bit vectors for MULUDQ
15610 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
15611 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
15612 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
15613 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
15614 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
15615 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
15617 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
15618 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
15619 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
15621 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
15622 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
15624 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
15625 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
15628 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
15629 assert(Subtarget->isTargetWin64() && "Unexpected target");
15630 EVT VT = Op.getValueType();
15631 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
15632 "Unexpected return type for lowering");
15636 switch (Op->getOpcode()) {
15637 default: llvm_unreachable("Unexpected request for libcall!");
15638 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
15639 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
15640 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
15641 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
15642 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
15643 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
15647 SDValue InChain = DAG.getEntryNode();
15649 TargetLowering::ArgListTy Args;
15650 TargetLowering::ArgListEntry Entry;
15651 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
15652 EVT ArgVT = Op->getOperand(i).getValueType();
15653 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
15654 "Unexpected argument type for lowering");
15655 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
15656 Entry.Node = StackPtr;
15657 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
15659 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15660 Entry.Ty = PointerType::get(ArgTy,0);
15661 Entry.isSExt = false;
15662 Entry.isZExt = false;
15663 Args.push_back(Entry);
15666 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
15669 TargetLowering::CallLoweringInfo CLI(DAG);
15670 CLI.setDebugLoc(dl).setChain(InChain)
15671 .setCallee(getLibcallCallingConv(LC),
15672 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
15673 Callee, std::move(Args), 0)
15674 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
15676 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
15677 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
15680 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
15681 SelectionDAG &DAG) {
15682 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
15683 EVT VT = Op0.getValueType();
15686 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
15687 (VT == MVT::v8i32 && Subtarget->hasInt256()));
15689 // PMULxD operations multiply each even value (starting at 0) of LHS with
15690 // the related value of RHS and produce a widen result.
15691 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15692 // => <2 x i64> <ae|cg>
15694 // In other word, to have all the results, we need to perform two PMULxD:
15695 // 1. one with the even values.
15696 // 2. one with the odd values.
15697 // To achieve #2, with need to place the odd values at an even position.
15699 // Place the odd value at an even position (basically, shift all values 1
15700 // step to the left):
15701 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
15702 // <a|b|c|d> => <b|undef|d|undef>
15703 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
15704 // <e|f|g|h> => <f|undef|h|undef>
15705 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
15707 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
15709 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
15710 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
15712 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
15713 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15714 // => <2 x i64> <ae|cg>
15715 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
15716 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
15717 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
15718 // => <2 x i64> <bf|dh>
15719 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
15720 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
15722 // Shuffle it back into the right order.
15723 SDValue Highs, Lows;
15724 if (VT == MVT::v8i32) {
15725 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
15726 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15727 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
15728 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15730 const int HighMask[] = {1, 5, 3, 7};
15731 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15732 const int LowMask[] = {0, 4, 2, 6};
15733 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15736 // If we have a signed multiply but no PMULDQ fix up the high parts of a
15737 // unsigned multiply.
15738 if (IsSigned && !Subtarget->hasSSE41()) {
15740 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
15741 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
15742 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
15743 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
15744 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
15746 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
15747 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
15750 // The first result of MUL_LOHI is actually the low value, followed by the
15752 SDValue Ops[] = {Lows, Highs};
15753 return DAG.getMergeValues(Ops, dl);
15756 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
15757 const X86Subtarget *Subtarget) {
15758 MVT VT = Op.getSimpleValueType();
15760 SDValue R = Op.getOperand(0);
15761 SDValue Amt = Op.getOperand(1);
15763 // Optimize shl/srl/sra with constant shift amount.
15764 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
15765 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
15766 uint64_t ShiftAmt = ShiftConst->getZExtValue();
15768 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
15769 (Subtarget->hasInt256() &&
15770 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15771 (Subtarget->hasAVX512() &&
15772 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15773 if (Op.getOpcode() == ISD::SHL)
15774 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15776 if (Op.getOpcode() == ISD::SRL)
15777 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15779 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
15780 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15784 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
15785 unsigned NumElts = VT.getVectorNumElements();
15786 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
15788 if (Op.getOpcode() == ISD::SHL) {
15789 // Make a large shift.
15790 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
15792 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
15793 // Zero out the rightmost bits.
15794 SmallVector<SDValue, 32> V(
15795 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), MVT::i8));
15796 return DAG.getNode(ISD::AND, dl, VT, SHL,
15797 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15799 if (Op.getOpcode() == ISD::SRL) {
15800 // Make a large shift.
15801 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
15803 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
15804 // Zero out the leftmost bits.
15805 SmallVector<SDValue, 32> V(
15806 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, MVT::i8));
15807 return DAG.getNode(ISD::AND, dl, VT, SRL,
15808 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15810 if (Op.getOpcode() == ISD::SRA) {
15811 if (ShiftAmt == 7) {
15812 // R s>> 7 === R s< 0
15813 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15814 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
15817 // R s>> a === ((R u>> a) ^ m) - m
15818 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
15819 SmallVector<SDValue, 32> V(NumElts,
15820 DAG.getConstant(128 >> ShiftAmt, MVT::i8));
15821 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
15822 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
15823 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
15826 llvm_unreachable("Unknown shift opcode.");
15831 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15832 if (!Subtarget->is64Bit() &&
15833 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
15834 Amt.getOpcode() == ISD::BITCAST &&
15835 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15836 Amt = Amt.getOperand(0);
15837 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15838 VT.getVectorNumElements();
15839 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
15840 uint64_t ShiftAmt = 0;
15841 for (unsigned i = 0; i != Ratio; ++i) {
15842 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
15846 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
15848 // Check remaining shift amounts.
15849 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15850 uint64_t ShAmt = 0;
15851 for (unsigned j = 0; j != Ratio; ++j) {
15852 ConstantSDNode *C =
15853 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
15857 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
15859 if (ShAmt != ShiftAmt)
15862 switch (Op.getOpcode()) {
15864 llvm_unreachable("Unknown shift opcode!");
15866 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15869 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15872 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15880 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
15881 const X86Subtarget* Subtarget) {
15882 MVT VT = Op.getSimpleValueType();
15884 SDValue R = Op.getOperand(0);
15885 SDValue Amt = Op.getOperand(1);
15887 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
15888 VT == MVT::v4i32 || VT == MVT::v8i16 ||
15889 (Subtarget->hasInt256() &&
15890 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
15891 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15892 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15894 EVT EltVT = VT.getVectorElementType();
15896 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
15897 // Check if this build_vector node is doing a splat.
15898 // If so, then set BaseShAmt equal to the splat value.
15899 BaseShAmt = BV->getSplatValue();
15900 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
15901 BaseShAmt = SDValue();
15903 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
15904 Amt = Amt.getOperand(0);
15906 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
15907 if (SVN && SVN->isSplat()) {
15908 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
15909 SDValue InVec = Amt.getOperand(0);
15910 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
15911 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
15912 "Unexpected shuffle index found!");
15913 BaseShAmt = InVec.getOperand(SplatIdx);
15914 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
15915 if (ConstantSDNode *C =
15916 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
15917 if (C->getZExtValue() == SplatIdx)
15918 BaseShAmt = InVec.getOperand(1);
15923 // Avoid introducing an extract element from a shuffle.
15924 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
15925 DAG.getIntPtrConstant(SplatIdx));
15929 if (BaseShAmt.getNode()) {
15930 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
15931 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
15932 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
15933 else if (EltVT.bitsLT(MVT::i32))
15934 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
15936 switch (Op.getOpcode()) {
15938 llvm_unreachable("Unknown shift opcode!");
15940 switch (VT.SimpleTy) {
15941 default: return SDValue();
15950 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
15953 switch (VT.SimpleTy) {
15954 default: return SDValue();
15961 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
15964 switch (VT.SimpleTy) {
15965 default: return SDValue();
15974 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
15980 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15981 if (!Subtarget->is64Bit() &&
15982 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
15983 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
15984 Amt.getOpcode() == ISD::BITCAST &&
15985 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15986 Amt = Amt.getOperand(0);
15987 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15988 VT.getVectorNumElements();
15989 std::vector<SDValue> Vals(Ratio);
15990 for (unsigned i = 0; i != Ratio; ++i)
15991 Vals[i] = Amt.getOperand(i);
15992 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15993 for (unsigned j = 0; j != Ratio; ++j)
15994 if (Vals[j] != Amt.getOperand(i + j))
15997 switch (Op.getOpcode()) {
15999 llvm_unreachable("Unknown shift opcode!");
16001 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
16003 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
16005 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
16012 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16013 SelectionDAG &DAG) {
16014 MVT VT = Op.getSimpleValueType();
16016 SDValue R = Op.getOperand(0);
16017 SDValue Amt = Op.getOperand(1);
16020 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16021 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16023 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
16027 V = LowerScalarVariableShift(Op, DAG, Subtarget);
16031 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
16033 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
16034 if (Subtarget->hasInt256()) {
16035 if (Op.getOpcode() == ISD::SRL &&
16036 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16037 VT == MVT::v4i64 || VT == MVT::v8i32))
16039 if (Op.getOpcode() == ISD::SHL &&
16040 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16041 VT == MVT::v4i64 || VT == MVT::v8i32))
16043 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
16047 // If possible, lower this packed shift into a vector multiply instead of
16048 // expanding it into a sequence of scalar shifts.
16049 // Do this only if the vector shift count is a constant build_vector.
16050 if (Op.getOpcode() == ISD::SHL &&
16051 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16052 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16053 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16054 SmallVector<SDValue, 8> Elts;
16055 EVT SVT = VT.getScalarType();
16056 unsigned SVTBits = SVT.getSizeInBits();
16057 const APInt &One = APInt(SVTBits, 1);
16058 unsigned NumElems = VT.getVectorNumElements();
16060 for (unsigned i=0; i !=NumElems; ++i) {
16061 SDValue Op = Amt->getOperand(i);
16062 if (Op->getOpcode() == ISD::UNDEF) {
16063 Elts.push_back(Op);
16067 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16068 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16069 uint64_t ShAmt = C.getZExtValue();
16070 if (ShAmt >= SVTBits) {
16071 Elts.push_back(DAG.getUNDEF(SVT));
16074 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
16076 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16077 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16080 // Lower SHL with variable shift amount.
16081 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16082 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
16084 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
16085 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16086 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16087 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16090 // If possible, lower this shift as a sequence of two shifts by
16091 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16093 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16095 // Could be rewritten as:
16096 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16098 // The advantage is that the two shifts from the example would be
16099 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16100 // the vector shift into four scalar shifts plus four pairs of vector
16102 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16103 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16104 unsigned TargetOpcode = X86ISD::MOVSS;
16105 bool CanBeSimplified;
16106 // The splat value for the first packed shift (the 'X' from the example).
16107 SDValue Amt1 = Amt->getOperand(0);
16108 // The splat value for the second packed shift (the 'Y' from the example).
16109 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16110 Amt->getOperand(2);
16112 // See if it is possible to replace this node with a sequence of
16113 // two shifts followed by a MOVSS/MOVSD
16114 if (VT == MVT::v4i32) {
16115 // Check if it is legal to use a MOVSS.
16116 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16117 Amt2 == Amt->getOperand(3);
16118 if (!CanBeSimplified) {
16119 // Otherwise, check if we can still simplify this node using a MOVSD.
16120 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16121 Amt->getOperand(2) == Amt->getOperand(3);
16122 TargetOpcode = X86ISD::MOVSD;
16123 Amt2 = Amt->getOperand(2);
16126 // Do similar checks for the case where the machine value type
16128 CanBeSimplified = Amt1 == Amt->getOperand(1);
16129 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16130 CanBeSimplified = Amt2 == Amt->getOperand(i);
16132 if (!CanBeSimplified) {
16133 TargetOpcode = X86ISD::MOVSD;
16134 CanBeSimplified = true;
16135 Amt2 = Amt->getOperand(4);
16136 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16137 CanBeSimplified = Amt1 == Amt->getOperand(i);
16138 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16139 CanBeSimplified = Amt2 == Amt->getOperand(j);
16143 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16144 isa<ConstantSDNode>(Amt2)) {
16145 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16146 EVT CastVT = MVT::v4i32;
16148 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
16149 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16151 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
16152 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16153 if (TargetOpcode == X86ISD::MOVSD)
16154 CastVT = MVT::v2i64;
16155 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16156 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16157 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16159 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16163 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16164 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
16167 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
16168 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
16170 // Turn 'a' into a mask suitable for VSELECT
16171 SDValue VSelM = DAG.getConstant(0x80, VT);
16172 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16173 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16175 SDValue CM1 = DAG.getConstant(0x0f, VT);
16176 SDValue CM2 = DAG.getConstant(0x3f, VT);
16178 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
16179 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
16180 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
16181 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16182 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16185 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16186 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16187 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16189 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
16190 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
16191 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
16192 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16193 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16196 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16197 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16198 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16200 // return VSELECT(r, r+r, a);
16201 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16202 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16206 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16207 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16208 // solution better.
16209 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16210 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16212 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16213 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16214 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16215 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16216 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16219 // Decompose 256-bit shifts into smaller 128-bit shifts.
16220 if (VT.is256BitVector()) {
16221 unsigned NumElems = VT.getVectorNumElements();
16222 MVT EltVT = VT.getVectorElementType();
16223 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16225 // Extract the two vectors
16226 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16227 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16229 // Recreate the shift amount vectors
16230 SDValue Amt1, Amt2;
16231 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16232 // Constant shift amount
16233 SmallVector<SDValue, 4> Amt1Csts;
16234 SmallVector<SDValue, 4> Amt2Csts;
16235 for (unsigned i = 0; i != NumElems/2; ++i)
16236 Amt1Csts.push_back(Amt->getOperand(i));
16237 for (unsigned i = NumElems/2; i != NumElems; ++i)
16238 Amt2Csts.push_back(Amt->getOperand(i));
16240 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16241 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16243 // Variable shift amount
16244 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16245 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16248 // Issue new vector shifts for the smaller types
16249 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16250 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16252 // Concatenate the result back
16253 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16259 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16260 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16261 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16262 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16263 // has only one use.
16264 SDNode *N = Op.getNode();
16265 SDValue LHS = N->getOperand(0);
16266 SDValue RHS = N->getOperand(1);
16267 unsigned BaseOp = 0;
16270 switch (Op.getOpcode()) {
16271 default: llvm_unreachable("Unknown ovf instruction!");
16273 // A subtract of one will be selected as a INC. Note that INC doesn't
16274 // set CF, so we can't do this for UADDO.
16275 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16277 BaseOp = X86ISD::INC;
16278 Cond = X86::COND_O;
16281 BaseOp = X86ISD::ADD;
16282 Cond = X86::COND_O;
16285 BaseOp = X86ISD::ADD;
16286 Cond = X86::COND_B;
16289 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16290 // set CF, so we can't do this for USUBO.
16291 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16293 BaseOp = X86ISD::DEC;
16294 Cond = X86::COND_O;
16297 BaseOp = X86ISD::SUB;
16298 Cond = X86::COND_O;
16301 BaseOp = X86ISD::SUB;
16302 Cond = X86::COND_B;
16305 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16306 Cond = X86::COND_O;
16308 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16309 if (N->getValueType(0) == MVT::i8) {
16310 BaseOp = X86ISD::UMUL8;
16311 Cond = X86::COND_O;
16314 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16316 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16319 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16320 DAG.getConstant(X86::COND_O, MVT::i32),
16321 SDValue(Sum.getNode(), 2));
16323 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16327 // Also sets EFLAGS.
16328 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16329 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16332 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16333 DAG.getConstant(Cond, MVT::i32),
16334 SDValue(Sum.getNode(), 1));
16336 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16339 // Sign extension of the low part of vector elements. This may be used either
16340 // when sign extend instructions are not available or if the vector element
16341 // sizes already match the sign-extended size. If the vector elements are in
16342 // their pre-extended size and sign extend instructions are available, that will
16343 // be handled by LowerSIGN_EXTEND.
16344 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
16345 SelectionDAG &DAG) const {
16347 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
16348 MVT VT = Op.getSimpleValueType();
16350 if (!Subtarget->hasSSE2() || !VT.isVector())
16353 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
16354 ExtraVT.getScalarType().getSizeInBits();
16356 switch (VT.SimpleTy) {
16357 default: return SDValue();
16360 if (!Subtarget->hasFp256())
16362 if (!Subtarget->hasInt256()) {
16363 // needs to be split
16364 unsigned NumElems = VT.getVectorNumElements();
16366 // Extract the LHS vectors
16367 SDValue LHS = Op.getOperand(0);
16368 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16369 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16371 MVT EltVT = VT.getVectorElementType();
16372 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16374 EVT ExtraEltVT = ExtraVT.getVectorElementType();
16375 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
16376 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
16378 SDValue Extra = DAG.getValueType(ExtraVT);
16380 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
16381 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
16383 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
16388 SDValue Op0 = Op.getOperand(0);
16390 // This is a sign extension of some low part of vector elements without
16391 // changing the size of the vector elements themselves:
16392 // Shift-Left + Shift-Right-Algebraic.
16393 SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
16395 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
16401 /// Returns true if the operand type is exactly twice the native width, and
16402 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16403 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16404 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16405 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16406 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16409 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16410 else if (OpWidth == 128)
16411 return Subtarget->hasCmpxchg16b();
16416 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16417 return needsCmpXchgNb(SI->getValueOperand()->getType());
16420 // Note: this turns large loads into lock cmpxchg8b/16b.
16421 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16422 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16423 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16424 return needsCmpXchgNb(PTy->getElementType());
16427 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16428 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16429 const Type *MemType = AI->getType();
16431 // If the operand is too big, we must see if cmpxchg8/16b is available
16432 // and default to library calls otherwise.
16433 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16434 return needsCmpXchgNb(MemType);
16436 AtomicRMWInst::BinOp Op = AI->getOperation();
16439 llvm_unreachable("Unknown atomic operation");
16440 case AtomicRMWInst::Xchg:
16441 case AtomicRMWInst::Add:
16442 case AtomicRMWInst::Sub:
16443 // It's better to use xadd, xsub or xchg for these in all cases.
16445 case AtomicRMWInst::Or:
16446 case AtomicRMWInst::And:
16447 case AtomicRMWInst::Xor:
16448 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16449 // prefix to a normal instruction for these operations.
16450 return !AI->use_empty();
16451 case AtomicRMWInst::Nand:
16452 case AtomicRMWInst::Max:
16453 case AtomicRMWInst::Min:
16454 case AtomicRMWInst::UMax:
16455 case AtomicRMWInst::UMin:
16456 // These always require a non-trivial set of data operations on x86. We must
16457 // use a cmpxchg loop.
16462 static bool hasMFENCE(const X86Subtarget& Subtarget) {
16463 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
16464 // no-sse2). There isn't any reason to disable it if the target processor
16466 return Subtarget.hasSSE2() || Subtarget.is64Bit();
16470 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
16471 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16472 const Type *MemType = AI->getType();
16473 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
16474 // there is no benefit in turning such RMWs into loads, and it is actually
16475 // harmful as it introduces a mfence.
16476 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16479 auto Builder = IRBuilder<>(AI);
16480 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16481 auto SynchScope = AI->getSynchScope();
16482 // We must restrict the ordering to avoid generating loads with Release or
16483 // ReleaseAcquire orderings.
16484 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
16485 auto Ptr = AI->getPointerOperand();
16487 // Before the load we need a fence. Here is an example lifted from
16488 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
16491 // x.store(1, relaxed);
16492 // r1 = y.fetch_add(0, release);
16494 // y.fetch_add(42, acquire);
16495 // r2 = x.load(relaxed);
16496 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
16497 // lowered to just a load without a fence. A mfence flushes the store buffer,
16498 // making the optimization clearly correct.
16499 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
16500 // otherwise, we might be able to be more agressive on relaxed idempotent
16501 // rmw. In practice, they do not look useful, so we don't try to be
16502 // especially clever.
16503 if (SynchScope == SingleThread) {
16504 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
16505 // the IR level, so we must wrap it in an intrinsic.
16507 } else if (hasMFENCE(*Subtarget)) {
16508 Function *MFence = llvm::Intrinsic::getDeclaration(M,
16509 Intrinsic::x86_sse2_mfence);
16510 Builder.CreateCall(MFence);
16512 // FIXME: it might make sense to use a locked operation here but on a
16513 // different cache-line to prevent cache-line bouncing. In practice it
16514 // is probably a small win, and x86 processors without mfence are rare
16515 // enough that we do not bother.
16519 // Finally we can emit the atomic load.
16520 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
16521 AI->getType()->getPrimitiveSizeInBits());
16522 Loaded->setAtomic(Order, SynchScope);
16523 AI->replaceAllUsesWith(Loaded);
16524 AI->eraseFromParent();
16528 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
16529 SelectionDAG &DAG) {
16531 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
16532 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
16533 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
16534 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
16536 // The only fence that needs an instruction is a sequentially-consistent
16537 // cross-thread fence.
16538 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
16539 if (hasMFENCE(*Subtarget))
16540 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
16542 SDValue Chain = Op.getOperand(0);
16543 SDValue Zero = DAG.getConstant(0, MVT::i32);
16545 DAG.getRegister(X86::ESP, MVT::i32), // Base
16546 DAG.getTargetConstant(1, MVT::i8), // Scale
16547 DAG.getRegister(0, MVT::i32), // Index
16548 DAG.getTargetConstant(0, MVT::i32), // Disp
16549 DAG.getRegister(0, MVT::i32), // Segment.
16553 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
16554 return SDValue(Res, 0);
16557 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
16558 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
16561 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
16562 SelectionDAG &DAG) {
16563 MVT T = Op.getSimpleValueType();
16567 switch(T.SimpleTy) {
16568 default: llvm_unreachable("Invalid value type!");
16569 case MVT::i8: Reg = X86::AL; size = 1; break;
16570 case MVT::i16: Reg = X86::AX; size = 2; break;
16571 case MVT::i32: Reg = X86::EAX; size = 4; break;
16573 assert(Subtarget->is64Bit() && "Node not type legal!");
16574 Reg = X86::RAX; size = 8;
16577 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
16578 Op.getOperand(2), SDValue());
16579 SDValue Ops[] = { cpIn.getValue(0),
16582 DAG.getTargetConstant(size, MVT::i8),
16583 cpIn.getValue(1) };
16584 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16585 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
16586 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
16590 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
16591 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
16592 MVT::i32, cpOut.getValue(2));
16593 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
16594 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
16596 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
16597 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
16598 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
16602 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
16603 SelectionDAG &DAG) {
16604 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
16605 MVT DstVT = Op.getSimpleValueType();
16607 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
16608 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16609 if (DstVT != MVT::f64)
16610 // This conversion needs to be expanded.
16613 SDValue InVec = Op->getOperand(0);
16615 unsigned NumElts = SrcVT.getVectorNumElements();
16616 EVT SVT = SrcVT.getVectorElementType();
16618 // Widen the vector in input in the case of MVT::v2i32.
16619 // Example: from MVT::v2i32 to MVT::v4i32.
16620 SmallVector<SDValue, 16> Elts;
16621 for (unsigned i = 0, e = NumElts; i != e; ++i)
16622 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
16623 DAG.getIntPtrConstant(i)));
16625 // Explicitly mark the extra elements as Undef.
16626 Elts.append(NumElts, DAG.getUNDEF(SVT));
16628 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
16629 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
16630 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
16631 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
16632 DAG.getIntPtrConstant(0));
16635 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
16636 Subtarget->hasMMX() && "Unexpected custom BITCAST");
16637 assert((DstVT == MVT::i64 ||
16638 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
16639 "Unexpected custom BITCAST");
16640 // i64 <=> MMX conversions are Legal.
16641 if (SrcVT==MVT::i64 && DstVT.isVector())
16643 if (DstVT==MVT::i64 && SrcVT.isVector())
16645 // MMX <=> MMX conversions are Legal.
16646 if (SrcVT.isVector() && DstVT.isVector())
16648 // All other conversions need to be expanded.
16652 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
16653 SelectionDAG &DAG) {
16654 SDNode *Node = Op.getNode();
16657 Op = Op.getOperand(0);
16658 EVT VT = Op.getValueType();
16659 assert((VT.is128BitVector() || VT.is256BitVector()) &&
16660 "CTPOP lowering only implemented for 128/256-bit wide vector types");
16662 unsigned NumElts = VT.getVectorNumElements();
16663 EVT EltVT = VT.getVectorElementType();
16664 unsigned Len = EltVT.getSizeInBits();
16666 // This is the vectorized version of the "best" algorithm from
16667 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
16668 // with a minor tweak to use a series of adds + shifts instead of vector
16669 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
16671 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
16672 // v8i32 => Always profitable
16674 // FIXME: There a couple of possible improvements:
16676 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
16677 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
16679 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
16680 "CTPOP not implemented for this vector element type.");
16682 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
16683 // extra legalization.
16684 bool NeedsBitcast = EltVT == MVT::i32;
16685 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
16687 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
16688 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
16689 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
16691 // v = v - ((v >> 1) & 0x55555555...)
16692 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
16693 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
16694 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
16696 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16698 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
16699 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
16701 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
16703 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
16704 if (VT != And.getValueType())
16705 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16706 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
16708 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
16709 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
16710 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
16711 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
16712 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
16714 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
16715 if (NeedsBitcast) {
16716 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16717 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
16718 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
16721 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
16722 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
16723 if (VT != AndRHS.getValueType()) {
16724 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
16725 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
16727 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
16729 // v = (v + (v >> 4)) & 0x0F0F0F0F...
16730 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
16731 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
16732 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
16733 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16735 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
16736 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
16737 if (NeedsBitcast) {
16738 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16739 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
16741 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
16742 if (VT != And.getValueType())
16743 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16745 // The algorithm mentioned above uses:
16746 // v = (v * 0x01010101...) >> (Len - 8)
16748 // Change it to use vector adds + vector shifts which yield faster results on
16749 // Haswell than using vector integer multiplication.
16751 // For i32 elements:
16752 // v = v + (v >> 8)
16753 // v = v + (v >> 16)
16755 // For i64 elements:
16756 // v = v + (v >> 8)
16757 // v = v + (v >> 16)
16758 // v = v + (v >> 32)
16761 SmallVector<SDValue, 8> Csts;
16762 for (unsigned i = 8; i <= Len/2; i *= 2) {
16763 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
16764 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
16765 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
16766 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16770 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
16771 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
16772 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
16773 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
16774 if (NeedsBitcast) {
16775 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16776 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
16778 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
16779 if (VT != And.getValueType())
16780 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16785 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
16786 SDNode *Node = Op.getNode();
16788 EVT T = Node->getValueType(0);
16789 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
16790 DAG.getConstant(0, T), Node->getOperand(2));
16791 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
16792 cast<AtomicSDNode>(Node)->getMemoryVT(),
16793 Node->getOperand(0),
16794 Node->getOperand(1), negOp,
16795 cast<AtomicSDNode>(Node)->getMemOperand(),
16796 cast<AtomicSDNode>(Node)->getOrdering(),
16797 cast<AtomicSDNode>(Node)->getSynchScope());
16800 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
16801 SDNode *Node = Op.getNode();
16803 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
16805 // Convert seq_cst store -> xchg
16806 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
16807 // FIXME: On 32-bit, store -> fist or movq would be more efficient
16808 // (The only way to get a 16-byte store is cmpxchg16b)
16809 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
16810 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
16811 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
16812 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
16813 cast<AtomicSDNode>(Node)->getMemoryVT(),
16814 Node->getOperand(0),
16815 Node->getOperand(1), Node->getOperand(2),
16816 cast<AtomicSDNode>(Node)->getMemOperand(),
16817 cast<AtomicSDNode>(Node)->getOrdering(),
16818 cast<AtomicSDNode>(Node)->getSynchScope());
16819 return Swap.getValue(1);
16821 // Other atomic stores have a simple pattern.
16825 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
16826 EVT VT = Op.getNode()->getSimpleValueType(0);
16828 // Let legalize expand this if it isn't a legal type yet.
16829 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
16832 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16835 bool ExtraOp = false;
16836 switch (Op.getOpcode()) {
16837 default: llvm_unreachable("Invalid code");
16838 case ISD::ADDC: Opc = X86ISD::ADD; break;
16839 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
16840 case ISD::SUBC: Opc = X86ISD::SUB; break;
16841 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
16845 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16847 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16848 Op.getOperand(1), Op.getOperand(2));
16851 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
16852 SelectionDAG &DAG) {
16853 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
16855 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
16856 // which returns the values as { float, float } (in XMM0) or
16857 // { double, double } (which is returned in XMM0, XMM1).
16859 SDValue Arg = Op.getOperand(0);
16860 EVT ArgVT = Arg.getValueType();
16861 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16863 TargetLowering::ArgListTy Args;
16864 TargetLowering::ArgListEntry Entry;
16868 Entry.isSExt = false;
16869 Entry.isZExt = false;
16870 Args.push_back(Entry);
16872 bool isF64 = ArgVT == MVT::f64;
16873 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
16874 // the small struct {f32, f32} is returned in (eax, edx). For f64,
16875 // the results are returned via SRet in memory.
16876 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
16877 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16878 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
16880 Type *RetTy = isF64
16881 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
16882 : (Type*)VectorType::get(ArgTy, 4);
16884 TargetLowering::CallLoweringInfo CLI(DAG);
16885 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
16886 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
16888 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
16891 // Returned in xmm0 and xmm1.
16892 return CallResult.first;
16894 // Returned in bits 0:31 and 32:64 xmm0.
16895 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16896 CallResult.first, DAG.getIntPtrConstant(0));
16897 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16898 CallResult.first, DAG.getIntPtrConstant(1));
16899 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
16900 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
16903 /// LowerOperation - Provide custom lowering hooks for some operations.
16905 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
16906 switch (Op.getOpcode()) {
16907 default: llvm_unreachable("Should not custom lower this!");
16908 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
16909 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
16910 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
16911 return LowerCMP_SWAP(Op, Subtarget, DAG);
16912 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
16913 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
16914 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
16915 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
16916 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
16917 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
16918 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
16919 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
16920 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
16921 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
16922 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
16923 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
16924 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
16925 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
16926 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
16927 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
16928 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
16929 case ISD::SHL_PARTS:
16930 case ISD::SRA_PARTS:
16931 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
16932 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
16933 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
16934 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
16935 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
16936 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
16937 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
16938 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
16939 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
16940 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
16941 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
16943 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
16944 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
16945 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
16946 case ISD::SETCC: return LowerSETCC(Op, DAG);
16947 case ISD::SELECT: return LowerSELECT(Op, DAG);
16948 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
16949 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
16950 case ISD::VASTART: return LowerVASTART(Op, DAG);
16951 case ISD::VAARG: return LowerVAARG(Op, DAG);
16952 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
16953 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
16954 case ISD::INTRINSIC_VOID:
16955 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
16956 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
16957 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
16958 case ISD::FRAME_TO_ARGS_OFFSET:
16959 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
16960 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
16961 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
16962 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
16963 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
16964 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
16965 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
16966 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
16967 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
16968 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
16969 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
16970 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
16971 case ISD::UMUL_LOHI:
16972 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
16975 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
16981 case ISD::UMULO: return LowerXALUO(Op, DAG);
16982 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
16983 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
16987 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
16988 case ISD::ADD: return LowerADD(Op, DAG);
16989 case ISD::SUB: return LowerSUB(Op, DAG);
16990 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
16994 /// ReplaceNodeResults - Replace a node with an illegal result type
16995 /// with a new node built out of custom code.
16996 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
16997 SmallVectorImpl<SDValue>&Results,
16998 SelectionDAG &DAG) const {
17000 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17001 switch (N->getOpcode()) {
17003 llvm_unreachable("Do not know how to custom type legalize this operation!");
17004 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
17005 case X86ISD::FMINC:
17007 case X86ISD::FMAXC:
17008 case X86ISD::FMAX: {
17009 EVT VT = N->getValueType(0);
17010 if (VT != MVT::v2f32)
17011 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
17012 SDValue UNDEF = DAG.getUNDEF(VT);
17013 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17014 N->getOperand(0), UNDEF);
17015 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
17016 N->getOperand(1), UNDEF);
17017 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
17020 case ISD::SIGN_EXTEND_INREG:
17025 // We don't want to expand or promote these.
17032 case ISD::UDIVREM: {
17033 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
17034 Results.push_back(V);
17037 case ISD::FP_TO_SINT:
17038 case ISD::FP_TO_UINT: {
17039 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
17041 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
17044 std::pair<SDValue,SDValue> Vals =
17045 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17046 SDValue FIST = Vals.first, StackSlot = Vals.second;
17047 if (FIST.getNode()) {
17048 EVT VT = N->getValueType(0);
17049 // Return a load from the stack slot.
17050 if (StackSlot.getNode())
17051 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17052 MachinePointerInfo(),
17053 false, false, false, 0));
17055 Results.push_back(FIST);
17059 case ISD::UINT_TO_FP: {
17060 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17061 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17062 N->getValueType(0) != MVT::v2f32)
17064 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17066 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
17068 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17069 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17070 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17071 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17072 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17073 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17076 case ISD::FP_ROUND: {
17077 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17079 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17080 Results.push_back(V);
17083 case ISD::INTRINSIC_W_CHAIN: {
17084 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17086 default : llvm_unreachable("Do not know how to custom type "
17087 "legalize this intrinsic operation!");
17088 case Intrinsic::x86_rdtsc:
17089 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17091 case Intrinsic::x86_rdtscp:
17092 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17094 case Intrinsic::x86_rdpmc:
17095 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17098 case ISD::READCYCLECOUNTER: {
17099 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17102 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17103 EVT T = N->getValueType(0);
17104 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17105 bool Regs64bit = T == MVT::i128;
17106 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17107 SDValue cpInL, cpInH;
17108 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17109 DAG.getConstant(0, HalfT));
17110 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17111 DAG.getConstant(1, HalfT));
17112 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17113 Regs64bit ? X86::RAX : X86::EAX,
17115 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17116 Regs64bit ? X86::RDX : X86::EDX,
17117 cpInH, cpInL.getValue(1));
17118 SDValue swapInL, swapInH;
17119 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17120 DAG.getConstant(0, HalfT));
17121 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17122 DAG.getConstant(1, HalfT));
17123 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17124 Regs64bit ? X86::RBX : X86::EBX,
17125 swapInL, cpInH.getValue(1));
17126 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17127 Regs64bit ? X86::RCX : X86::ECX,
17128 swapInH, swapInL.getValue(1));
17129 SDValue Ops[] = { swapInH.getValue(0),
17131 swapInH.getValue(1) };
17132 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17133 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17134 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17135 X86ISD::LCMPXCHG8_DAG;
17136 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17137 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17138 Regs64bit ? X86::RAX : X86::EAX,
17139 HalfT, Result.getValue(1));
17140 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17141 Regs64bit ? X86::RDX : X86::EDX,
17142 HalfT, cpOutL.getValue(2));
17143 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17145 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17146 MVT::i32, cpOutH.getValue(2));
17148 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17149 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17150 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17152 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17153 Results.push_back(Success);
17154 Results.push_back(EFLAGS.getValue(1));
17157 case ISD::ATOMIC_SWAP:
17158 case ISD::ATOMIC_LOAD_ADD:
17159 case ISD::ATOMIC_LOAD_SUB:
17160 case ISD::ATOMIC_LOAD_AND:
17161 case ISD::ATOMIC_LOAD_OR:
17162 case ISD::ATOMIC_LOAD_XOR:
17163 case ISD::ATOMIC_LOAD_NAND:
17164 case ISD::ATOMIC_LOAD_MIN:
17165 case ISD::ATOMIC_LOAD_MAX:
17166 case ISD::ATOMIC_LOAD_UMIN:
17167 case ISD::ATOMIC_LOAD_UMAX:
17168 case ISD::ATOMIC_LOAD: {
17169 // Delegate to generic TypeLegalization. Situations we can really handle
17170 // should have already been dealt with by AtomicExpandPass.cpp.
17173 case ISD::BITCAST: {
17174 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17175 EVT DstVT = N->getValueType(0);
17176 EVT SrcVT = N->getOperand(0)->getValueType(0);
17178 if (SrcVT != MVT::f64 ||
17179 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17182 unsigned NumElts = DstVT.getVectorNumElements();
17183 EVT SVT = DstVT.getVectorElementType();
17184 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17185 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17186 MVT::v2f64, N->getOperand(0));
17187 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17189 if (ExperimentalVectorWideningLegalization) {
17190 // If we are legalizing vectors by widening, we already have the desired
17191 // legal vector type, just return it.
17192 Results.push_back(ToVecInt);
17196 SmallVector<SDValue, 8> Elts;
17197 for (unsigned i = 0, e = NumElts; i != e; ++i)
17198 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17199 ToVecInt, DAG.getIntPtrConstant(i)));
17201 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17206 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17208 default: return nullptr;
17209 case X86ISD::BSF: return "X86ISD::BSF";
17210 case X86ISD::BSR: return "X86ISD::BSR";
17211 case X86ISD::SHLD: return "X86ISD::SHLD";
17212 case X86ISD::SHRD: return "X86ISD::SHRD";
17213 case X86ISD::FAND: return "X86ISD::FAND";
17214 case X86ISD::FANDN: return "X86ISD::FANDN";
17215 case X86ISD::FOR: return "X86ISD::FOR";
17216 case X86ISD::FXOR: return "X86ISD::FXOR";
17217 case X86ISD::FSRL: return "X86ISD::FSRL";
17218 case X86ISD::FILD: return "X86ISD::FILD";
17219 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17220 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17221 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17222 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17223 case X86ISD::FLD: return "X86ISD::FLD";
17224 case X86ISD::FST: return "X86ISD::FST";
17225 case X86ISD::CALL: return "X86ISD::CALL";
17226 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17227 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17228 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17229 case X86ISD::BT: return "X86ISD::BT";
17230 case X86ISD::CMP: return "X86ISD::CMP";
17231 case X86ISD::COMI: return "X86ISD::COMI";
17232 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17233 case X86ISD::CMPM: return "X86ISD::CMPM";
17234 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17235 case X86ISD::SETCC: return "X86ISD::SETCC";
17236 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17237 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17238 case X86ISD::CMOV: return "X86ISD::CMOV";
17239 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17240 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17241 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17242 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17243 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17244 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17245 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17246 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17247 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17248 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17249 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17250 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17251 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17252 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17253 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17254 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17255 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17256 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17257 case X86ISD::HADD: return "X86ISD::HADD";
17258 case X86ISD::HSUB: return "X86ISD::HSUB";
17259 case X86ISD::FHADD: return "X86ISD::FHADD";
17260 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17261 case X86ISD::UMAX: return "X86ISD::UMAX";
17262 case X86ISD::UMIN: return "X86ISD::UMIN";
17263 case X86ISD::SMAX: return "X86ISD::SMAX";
17264 case X86ISD::SMIN: return "X86ISD::SMIN";
17265 case X86ISD::FMAX: return "X86ISD::FMAX";
17266 case X86ISD::FMIN: return "X86ISD::FMIN";
17267 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17268 case X86ISD::FMINC: return "X86ISD::FMINC";
17269 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17270 case X86ISD::FRCP: return "X86ISD::FRCP";
17271 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17272 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17273 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17274 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17275 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17276 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17277 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17278 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17279 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17280 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17281 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17282 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17283 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17284 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17285 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17286 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17287 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17288 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17289 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17290 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17291 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17292 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17293 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17294 case X86ISD::VSHL: return "X86ISD::VSHL";
17295 case X86ISD::VSRL: return "X86ISD::VSRL";
17296 case X86ISD::VSRA: return "X86ISD::VSRA";
17297 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17298 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17299 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17300 case X86ISD::CMPP: return "X86ISD::CMPP";
17301 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17302 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17303 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17304 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17305 case X86ISD::ADD: return "X86ISD::ADD";
17306 case X86ISD::SUB: return "X86ISD::SUB";
17307 case X86ISD::ADC: return "X86ISD::ADC";
17308 case X86ISD::SBB: return "X86ISD::SBB";
17309 case X86ISD::SMUL: return "X86ISD::SMUL";
17310 case X86ISD::UMUL: return "X86ISD::UMUL";
17311 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17312 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17313 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17314 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17315 case X86ISD::INC: return "X86ISD::INC";
17316 case X86ISD::DEC: return "X86ISD::DEC";
17317 case X86ISD::OR: return "X86ISD::OR";
17318 case X86ISD::XOR: return "X86ISD::XOR";
17319 case X86ISD::AND: return "X86ISD::AND";
17320 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17321 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17322 case X86ISD::PTEST: return "X86ISD::PTEST";
17323 case X86ISD::TESTP: return "X86ISD::TESTP";
17324 case X86ISD::TESTM: return "X86ISD::TESTM";
17325 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17326 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17327 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17328 case X86ISD::PACKUS: return "X86ISD::PACKUS";
17329 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
17330 case X86ISD::VALIGN: return "X86ISD::VALIGN";
17331 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
17332 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
17333 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
17334 case X86ISD::SHUFP: return "X86ISD::SHUFP";
17335 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
17336 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
17337 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
17338 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
17339 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
17340 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
17341 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
17342 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
17343 case X86ISD::MOVSD: return "X86ISD::MOVSD";
17344 case X86ISD::MOVSS: return "X86ISD::MOVSS";
17345 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
17346 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
17347 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
17348 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
17349 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
17350 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
17351 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
17352 case X86ISD::VPERMV: return "X86ISD::VPERMV";
17353 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
17354 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
17355 case X86ISD::VPERMI: return "X86ISD::VPERMI";
17356 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
17357 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
17358 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
17359 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
17360 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
17361 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
17362 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
17363 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
17364 case X86ISD::SAHF: return "X86ISD::SAHF";
17365 case X86ISD::RDRAND: return "X86ISD::RDRAND";
17366 case X86ISD::RDSEED: return "X86ISD::RDSEED";
17367 case X86ISD::FMADD: return "X86ISD::FMADD";
17368 case X86ISD::FMSUB: return "X86ISD::FMSUB";
17369 case X86ISD::FNMADD: return "X86ISD::FNMADD";
17370 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
17371 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
17372 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
17373 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
17374 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
17375 case X86ISD::XTEST: return "X86ISD::XTEST";
17376 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
17377 case X86ISD::EXPAND: return "X86ISD::EXPAND";
17378 case X86ISD::SELECT: return "X86ISD::SELECT";
17379 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
17380 case X86ISD::RCP28: return "X86ISD::RCP28";
17381 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
17382 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
17383 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
17384 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
17385 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
17389 // isLegalAddressingMode - Return true if the addressing mode represented
17390 // by AM is legal for this target, for a load/store of the specified type.
17391 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
17393 // X86 supports extremely general addressing modes.
17394 CodeModel::Model M = getTargetMachine().getCodeModel();
17395 Reloc::Model R = getTargetMachine().getRelocationModel();
17397 // X86 allows a sign-extended 32-bit immediate field as a displacement.
17398 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
17403 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
17405 // If a reference to this global requires an extra load, we can't fold it.
17406 if (isGlobalStubReference(GVFlags))
17409 // If BaseGV requires a register for the PIC base, we cannot also have a
17410 // BaseReg specified.
17411 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
17414 // If lower 4G is not available, then we must use rip-relative addressing.
17415 if ((M != CodeModel::Small || R != Reloc::Static) &&
17416 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
17420 switch (AM.Scale) {
17426 // These scales always work.
17431 // These scales are formed with basereg+scalereg. Only accept if there is
17436 default: // Other stuff never works.
17443 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
17444 unsigned Bits = Ty->getScalarSizeInBits();
17446 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
17447 // particularly cheaper than those without.
17451 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
17452 // variable shifts just as cheap as scalar ones.
17453 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
17456 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
17457 // fully general vector.
17461 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
17462 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17464 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
17465 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
17466 return NumBits1 > NumBits2;
17469 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
17470 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17473 if (!isTypeLegal(EVT::getEVT(Ty1)))
17476 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
17478 // Assuming the caller doesn't have a zeroext or signext return parameter,
17479 // truncation all the way down to i1 is valid.
17483 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
17484 return isInt<32>(Imm);
17487 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
17488 // Can also use sub to handle negated immediates.
17489 return isInt<32>(Imm);
17492 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
17493 if (!VT1.isInteger() || !VT2.isInteger())
17495 unsigned NumBits1 = VT1.getSizeInBits();
17496 unsigned NumBits2 = VT2.getSizeInBits();
17497 return NumBits1 > NumBits2;
17500 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
17501 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17502 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
17505 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
17506 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17507 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
17510 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
17511 EVT VT1 = Val.getValueType();
17512 if (isZExtFree(VT1, VT2))
17515 if (Val.getOpcode() != ISD::LOAD)
17518 if (!VT1.isSimple() || !VT1.isInteger() ||
17519 !VT2.isSimple() || !VT2.isInteger())
17522 switch (VT1.getSimpleVT().SimpleTy) {
17527 // X86 has 8, 16, and 32-bit zero-extending loads.
17534 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
17537 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
17538 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
17541 VT = VT.getScalarType();
17543 if (!VT.isSimple())
17546 switch (VT.getSimpleVT().SimpleTy) {
17557 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
17558 // i16 instructions are longer (0x66 prefix) and potentially slower.
17559 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
17562 /// isShuffleMaskLegal - Targets can use this to indicate that they only
17563 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
17564 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
17565 /// are assumed to be legal.
17567 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
17569 if (!VT.isSimple())
17572 // Very little shuffling can be done for 64-bit vectors right now.
17573 if (VT.getSizeInBits() == 64)
17576 // We only care that the types being shuffled are legal. The lowering can
17577 // handle any possible shuffle mask that results.
17578 return isTypeLegal(VT.getSimpleVT());
17582 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
17584 // Just delegate to the generic legality, clear masks aren't special.
17585 return isShuffleMaskLegal(Mask, VT);
17588 //===----------------------------------------------------------------------===//
17589 // X86 Scheduler Hooks
17590 //===----------------------------------------------------------------------===//
17592 /// Utility function to emit xbegin specifying the start of an RTM region.
17593 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
17594 const TargetInstrInfo *TII) {
17595 DebugLoc DL = MI->getDebugLoc();
17597 const BasicBlock *BB = MBB->getBasicBlock();
17598 MachineFunction::iterator I = MBB;
17601 // For the v = xbegin(), we generate
17612 MachineBasicBlock *thisMBB = MBB;
17613 MachineFunction *MF = MBB->getParent();
17614 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
17615 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
17616 MF->insert(I, mainMBB);
17617 MF->insert(I, sinkMBB);
17619 // Transfer the remainder of BB and its successor edges to sinkMBB.
17620 sinkMBB->splice(sinkMBB->begin(), MBB,
17621 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17622 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
17626 // # fallthrough to mainMBB
17627 // # abortion to sinkMBB
17628 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
17629 thisMBB->addSuccessor(mainMBB);
17630 thisMBB->addSuccessor(sinkMBB);
17634 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
17635 mainMBB->addSuccessor(sinkMBB);
17638 // EAX is live into the sinkMBB
17639 sinkMBB->addLiveIn(X86::EAX);
17640 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
17641 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17644 MI->eraseFromParent();
17648 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
17649 // or XMM0_V32I8 in AVX all of this code can be replaced with that
17650 // in the .td file.
17651 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
17652 const TargetInstrInfo *TII) {
17654 switch (MI->getOpcode()) {
17655 default: llvm_unreachable("illegal opcode!");
17656 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
17657 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
17658 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
17659 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
17660 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
17661 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
17662 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
17663 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
17666 DebugLoc dl = MI->getDebugLoc();
17667 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17669 unsigned NumArgs = MI->getNumOperands();
17670 for (unsigned i = 1; i < NumArgs; ++i) {
17671 MachineOperand &Op = MI->getOperand(i);
17672 if (!(Op.isReg() && Op.isImplicit()))
17673 MIB.addOperand(Op);
17675 if (MI->hasOneMemOperand())
17676 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17678 BuildMI(*BB, MI, dl,
17679 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17680 .addReg(X86::XMM0);
17682 MI->eraseFromParent();
17686 // FIXME: Custom handling because TableGen doesn't support multiple implicit
17687 // defs in an instruction pattern
17688 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
17689 const TargetInstrInfo *TII) {
17691 switch (MI->getOpcode()) {
17692 default: llvm_unreachable("illegal opcode!");
17693 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
17694 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
17695 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
17696 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
17697 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
17698 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
17699 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
17700 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
17703 DebugLoc dl = MI->getDebugLoc();
17704 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17706 unsigned NumArgs = MI->getNumOperands(); // remove the results
17707 for (unsigned i = 1; i < NumArgs; ++i) {
17708 MachineOperand &Op = MI->getOperand(i);
17709 if (!(Op.isReg() && Op.isImplicit()))
17710 MIB.addOperand(Op);
17712 if (MI->hasOneMemOperand())
17713 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17715 BuildMI(*BB, MI, dl,
17716 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17719 MI->eraseFromParent();
17723 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
17724 const X86Subtarget *Subtarget) {
17725 DebugLoc dl = MI->getDebugLoc();
17726 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17727 // Address into RAX/EAX, other two args into ECX, EDX.
17728 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
17729 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
17730 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
17731 for (int i = 0; i < X86::AddrNumOperands; ++i)
17732 MIB.addOperand(MI->getOperand(i));
17734 unsigned ValOps = X86::AddrNumOperands;
17735 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
17736 .addReg(MI->getOperand(ValOps).getReg());
17737 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
17738 .addReg(MI->getOperand(ValOps+1).getReg());
17740 // The instruction doesn't actually take any operands though.
17741 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
17743 MI->eraseFromParent(); // The pseudo is gone now.
17747 MachineBasicBlock *
17748 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
17749 MachineBasicBlock *MBB) const {
17750 // Emit va_arg instruction on X86-64.
17752 // Operands to this pseudo-instruction:
17753 // 0 ) Output : destination address (reg)
17754 // 1-5) Input : va_list address (addr, i64mem)
17755 // 6 ) ArgSize : Size (in bytes) of vararg type
17756 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
17757 // 8 ) Align : Alignment of type
17758 // 9 ) EFLAGS (implicit-def)
17760 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
17761 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
17763 unsigned DestReg = MI->getOperand(0).getReg();
17764 MachineOperand &Base = MI->getOperand(1);
17765 MachineOperand &Scale = MI->getOperand(2);
17766 MachineOperand &Index = MI->getOperand(3);
17767 MachineOperand &Disp = MI->getOperand(4);
17768 MachineOperand &Segment = MI->getOperand(5);
17769 unsigned ArgSize = MI->getOperand(6).getImm();
17770 unsigned ArgMode = MI->getOperand(7).getImm();
17771 unsigned Align = MI->getOperand(8).getImm();
17773 // Memory Reference
17774 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
17775 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17776 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17778 // Machine Information
17779 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17780 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
17781 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
17782 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
17783 DebugLoc DL = MI->getDebugLoc();
17785 // struct va_list {
17788 // i64 overflow_area (address)
17789 // i64 reg_save_area (address)
17791 // sizeof(va_list) = 24
17792 // alignment(va_list) = 8
17794 unsigned TotalNumIntRegs = 6;
17795 unsigned TotalNumXMMRegs = 8;
17796 bool UseGPOffset = (ArgMode == 1);
17797 bool UseFPOffset = (ArgMode == 2);
17798 unsigned MaxOffset = TotalNumIntRegs * 8 +
17799 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
17801 /* Align ArgSize to a multiple of 8 */
17802 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
17803 bool NeedsAlign = (Align > 8);
17805 MachineBasicBlock *thisMBB = MBB;
17806 MachineBasicBlock *overflowMBB;
17807 MachineBasicBlock *offsetMBB;
17808 MachineBasicBlock *endMBB;
17810 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
17811 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
17812 unsigned OffsetReg = 0;
17814 if (!UseGPOffset && !UseFPOffset) {
17815 // If we only pull from the overflow region, we don't create a branch.
17816 // We don't need to alter control flow.
17817 OffsetDestReg = 0; // unused
17818 OverflowDestReg = DestReg;
17820 offsetMBB = nullptr;
17821 overflowMBB = thisMBB;
17824 // First emit code to check if gp_offset (or fp_offset) is below the bound.
17825 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
17826 // If not, pull from overflow_area. (branch to overflowMBB)
17831 // offsetMBB overflowMBB
17836 // Registers for the PHI in endMBB
17837 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
17838 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
17840 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17841 MachineFunction *MF = MBB->getParent();
17842 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17843 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17844 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17846 MachineFunction::iterator MBBIter = MBB;
17849 // Insert the new basic blocks
17850 MF->insert(MBBIter, offsetMBB);
17851 MF->insert(MBBIter, overflowMBB);
17852 MF->insert(MBBIter, endMBB);
17854 // Transfer the remainder of MBB and its successor edges to endMBB.
17855 endMBB->splice(endMBB->begin(), thisMBB,
17856 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
17857 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
17859 // Make offsetMBB and overflowMBB successors of thisMBB
17860 thisMBB->addSuccessor(offsetMBB);
17861 thisMBB->addSuccessor(overflowMBB);
17863 // endMBB is a successor of both offsetMBB and overflowMBB
17864 offsetMBB->addSuccessor(endMBB);
17865 overflowMBB->addSuccessor(endMBB);
17867 // Load the offset value into a register
17868 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17869 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
17873 .addDisp(Disp, UseFPOffset ? 4 : 0)
17874 .addOperand(Segment)
17875 .setMemRefs(MMOBegin, MMOEnd);
17877 // Check if there is enough room left to pull this argument.
17878 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
17880 .addImm(MaxOffset + 8 - ArgSizeA8);
17882 // Branch to "overflowMBB" if offset >= max
17883 // Fall through to "offsetMBB" otherwise
17884 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
17885 .addMBB(overflowMBB);
17888 // In offsetMBB, emit code to use the reg_save_area.
17890 assert(OffsetReg != 0);
17892 // Read the reg_save_area address.
17893 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
17894 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
17899 .addOperand(Segment)
17900 .setMemRefs(MMOBegin, MMOEnd);
17902 // Zero-extend the offset
17903 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
17904 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
17907 .addImm(X86::sub_32bit);
17909 // Add the offset to the reg_save_area to get the final address.
17910 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
17911 .addReg(OffsetReg64)
17912 .addReg(RegSaveReg);
17914 // Compute the offset for the next argument
17915 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17916 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
17918 .addImm(UseFPOffset ? 16 : 8);
17920 // Store it back into the va_list.
17921 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
17925 .addDisp(Disp, UseFPOffset ? 4 : 0)
17926 .addOperand(Segment)
17927 .addReg(NextOffsetReg)
17928 .setMemRefs(MMOBegin, MMOEnd);
17931 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
17936 // Emit code to use overflow area
17939 // Load the overflow_area address into a register.
17940 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
17941 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
17946 .addOperand(Segment)
17947 .setMemRefs(MMOBegin, MMOEnd);
17949 // If we need to align it, do so. Otherwise, just copy the address
17950 // to OverflowDestReg.
17952 // Align the overflow address
17953 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
17954 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
17956 // aligned_addr = (addr + (align-1)) & ~(align-1)
17957 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
17958 .addReg(OverflowAddrReg)
17961 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
17963 .addImm(~(uint64_t)(Align-1));
17965 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
17966 .addReg(OverflowAddrReg);
17969 // Compute the next overflow address after this argument.
17970 // (the overflow address should be kept 8-byte aligned)
17971 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
17972 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
17973 .addReg(OverflowDestReg)
17974 .addImm(ArgSizeA8);
17976 // Store the new overflow address.
17977 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
17982 .addOperand(Segment)
17983 .addReg(NextAddrReg)
17984 .setMemRefs(MMOBegin, MMOEnd);
17986 // If we branched, emit the PHI to the front of endMBB.
17988 BuildMI(*endMBB, endMBB->begin(), DL,
17989 TII->get(X86::PHI), DestReg)
17990 .addReg(OffsetDestReg).addMBB(offsetMBB)
17991 .addReg(OverflowDestReg).addMBB(overflowMBB);
17994 // Erase the pseudo instruction
17995 MI->eraseFromParent();
18000 MachineBasicBlock *
18001 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
18003 MachineBasicBlock *MBB) const {
18004 // Emit code to save XMM registers to the stack. The ABI says that the
18005 // number of registers to save is given in %al, so it's theoretically
18006 // possible to do an indirect jump trick to avoid saving all of them,
18007 // however this code takes a simpler approach and just executes all
18008 // of the stores if %al is non-zero. It's less code, and it's probably
18009 // easier on the hardware branch predictor, and stores aren't all that
18010 // expensive anyway.
18012 // Create the new basic blocks. One block contains all the XMM stores,
18013 // and one block is the final destination regardless of whether any
18014 // stores were performed.
18015 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
18016 MachineFunction *F = MBB->getParent();
18017 MachineFunction::iterator MBBIter = MBB;
18019 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
18020 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
18021 F->insert(MBBIter, XMMSaveMBB);
18022 F->insert(MBBIter, EndMBB);
18024 // Transfer the remainder of MBB and its successor edges to EndMBB.
18025 EndMBB->splice(EndMBB->begin(), MBB,
18026 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18027 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
18029 // The original block will now fall through to the XMM save block.
18030 MBB->addSuccessor(XMMSaveMBB);
18031 // The XMMSaveMBB will fall through to the end block.
18032 XMMSaveMBB->addSuccessor(EndMBB);
18034 // Now add the instructions.
18035 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18036 DebugLoc DL = MI->getDebugLoc();
18038 unsigned CountReg = MI->getOperand(0).getReg();
18039 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
18040 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
18042 if (!Subtarget->isTargetWin64()) {
18043 // If %al is 0, branch around the XMM save block.
18044 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
18045 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18046 MBB->addSuccessor(EndMBB);
18049 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18050 // that was just emitted, but clearly shouldn't be "saved".
18051 assert((MI->getNumOperands() <= 3 ||
18052 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18053 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18054 && "Expected last argument to be EFLAGS");
18055 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18056 // In the XMM save block, save all the XMM argument registers.
18057 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18058 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18059 MachineMemOperand *MMO =
18060 F->getMachineMemOperand(
18061 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18062 MachineMemOperand::MOStore,
18063 /*Size=*/16, /*Align=*/16);
18064 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18065 .addFrameIndex(RegSaveFrameIndex)
18066 .addImm(/*Scale=*/1)
18067 .addReg(/*IndexReg=*/0)
18068 .addImm(/*Disp=*/Offset)
18069 .addReg(/*Segment=*/0)
18070 .addReg(MI->getOperand(i).getReg())
18071 .addMemOperand(MMO);
18074 MI->eraseFromParent(); // The pseudo instruction is gone now.
18079 // The EFLAGS operand of SelectItr might be missing a kill marker
18080 // because there were multiple uses of EFLAGS, and ISel didn't know
18081 // which to mark. Figure out whether SelectItr should have had a
18082 // kill marker, and set it if it should. Returns the correct kill
18084 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18085 MachineBasicBlock* BB,
18086 const TargetRegisterInfo* TRI) {
18087 // Scan forward through BB for a use/def of EFLAGS.
18088 MachineBasicBlock::iterator miI(std::next(SelectItr));
18089 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18090 const MachineInstr& mi = *miI;
18091 if (mi.readsRegister(X86::EFLAGS))
18093 if (mi.definesRegister(X86::EFLAGS))
18094 break; // Should have kill-flag - update below.
18097 // If we hit the end of the block, check whether EFLAGS is live into a
18099 if (miI == BB->end()) {
18100 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18101 sEnd = BB->succ_end();
18102 sItr != sEnd; ++sItr) {
18103 MachineBasicBlock* succ = *sItr;
18104 if (succ->isLiveIn(X86::EFLAGS))
18109 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18110 // out. SelectMI should have a kill flag on EFLAGS.
18111 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18115 MachineBasicBlock *
18116 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18117 MachineBasicBlock *BB) const {
18118 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18119 DebugLoc DL = MI->getDebugLoc();
18121 // To "insert" a SELECT_CC instruction, we actually have to insert the
18122 // diamond control-flow pattern. The incoming instruction knows the
18123 // destination vreg to set, the condition code register to branch on, the
18124 // true/false values to select between, and a branch opcode to use.
18125 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18126 MachineFunction::iterator It = BB;
18132 // cmpTY ccX, r1, r2
18134 // fallthrough --> copy0MBB
18135 MachineBasicBlock *thisMBB = BB;
18136 MachineFunction *F = BB->getParent();
18137 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18138 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18139 F->insert(It, copy0MBB);
18140 F->insert(It, sinkMBB);
18142 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18143 // live into the sink and copy blocks.
18144 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18145 if (!MI->killsRegister(X86::EFLAGS) &&
18146 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
18147 copy0MBB->addLiveIn(X86::EFLAGS);
18148 sinkMBB->addLiveIn(X86::EFLAGS);
18151 // Transfer the remainder of BB and its successor edges to sinkMBB.
18152 sinkMBB->splice(sinkMBB->begin(), BB,
18153 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18154 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18156 // Add the true and fallthrough blocks as its successors.
18157 BB->addSuccessor(copy0MBB);
18158 BB->addSuccessor(sinkMBB);
18160 // Create the conditional branch instruction.
18162 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18163 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18166 // %FalseValue = ...
18167 // # fallthrough to sinkMBB
18168 copy0MBB->addSuccessor(sinkMBB);
18171 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18173 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18174 TII->get(X86::PHI), MI->getOperand(0).getReg())
18175 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18176 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18178 MI->eraseFromParent(); // The pseudo instruction is gone now.
18182 MachineBasicBlock *
18183 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18184 MachineBasicBlock *BB) const {
18185 MachineFunction *MF = BB->getParent();
18186 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18187 DebugLoc DL = MI->getDebugLoc();
18188 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18190 assert(MF->shouldSplitStack());
18192 const bool Is64Bit = Subtarget->is64Bit();
18193 const bool IsLP64 = Subtarget->isTarget64BitLP64();
18195 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
18196 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
18199 // ... [Till the alloca]
18200 // If stacklet is not large enough, jump to mallocMBB
18203 // Allocate by subtracting from RSP
18204 // Jump to continueMBB
18207 // Allocate by call to runtime
18211 // [rest of original BB]
18214 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18215 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18216 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18218 MachineRegisterInfo &MRI = MF->getRegInfo();
18219 const TargetRegisterClass *AddrRegClass =
18220 getRegClassFor(getPointerTy());
18222 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18223 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18224 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
18225 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
18226 sizeVReg = MI->getOperand(1).getReg(),
18227 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
18229 MachineFunction::iterator MBBIter = BB;
18232 MF->insert(MBBIter, bumpMBB);
18233 MF->insert(MBBIter, mallocMBB);
18234 MF->insert(MBBIter, continueMBB);
18236 continueMBB->splice(continueMBB->begin(), BB,
18237 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18238 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
18240 // Add code to the main basic block to check if the stack limit has been hit,
18241 // and if so, jump to mallocMBB otherwise to bumpMBB.
18242 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
18243 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
18244 .addReg(tmpSPVReg).addReg(sizeVReg);
18245 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
18246 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
18247 .addReg(SPLimitVReg);
18248 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
18250 // bumpMBB simply decreases the stack pointer, since we know the current
18251 // stacklet has enough space.
18252 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
18253 .addReg(SPLimitVReg);
18254 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
18255 .addReg(SPLimitVReg);
18256 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18258 // Calls into a routine in libgcc to allocate more space from the heap.
18259 const uint32_t *RegMask =
18260 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18262 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
18264 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18265 .addExternalSymbol("__morestack_allocate_stack_space")
18266 .addRegMask(RegMask)
18267 .addReg(X86::RDI, RegState::Implicit)
18268 .addReg(X86::RAX, RegState::ImplicitDefine);
18269 } else if (Is64Bit) {
18270 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
18272 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18273 .addExternalSymbol("__morestack_allocate_stack_space")
18274 .addRegMask(RegMask)
18275 .addReg(X86::EDI, RegState::Implicit)
18276 .addReg(X86::EAX, RegState::ImplicitDefine);
18278 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
18280 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
18281 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
18282 .addExternalSymbol("__morestack_allocate_stack_space")
18283 .addRegMask(RegMask)
18284 .addReg(X86::EAX, RegState::ImplicitDefine);
18288 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
18291 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
18292 .addReg(IsLP64 ? X86::RAX : X86::EAX);
18293 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18295 // Set up the CFG correctly.
18296 BB->addSuccessor(bumpMBB);
18297 BB->addSuccessor(mallocMBB);
18298 mallocMBB->addSuccessor(continueMBB);
18299 bumpMBB->addSuccessor(continueMBB);
18301 // Take care of the PHI nodes.
18302 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
18303 MI->getOperand(0).getReg())
18304 .addReg(mallocPtrVReg).addMBB(mallocMBB)
18305 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
18307 // Delete the original pseudo instruction.
18308 MI->eraseFromParent();
18311 return continueMBB;
18314 MachineBasicBlock *
18315 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
18316 MachineBasicBlock *BB) const {
18317 DebugLoc DL = MI->getDebugLoc();
18319 assert(!Subtarget->isTargetMachO());
18321 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
18323 MI->eraseFromParent(); // The pseudo instruction is gone now.
18327 MachineBasicBlock *
18328 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
18329 MachineBasicBlock *BB) const {
18330 // This is pretty easy. We're taking the value that we received from
18331 // our load from the relocation, sticking it in either RDI (x86-64)
18332 // or EAX and doing an indirect call. The return value will then
18333 // be in the normal return register.
18334 MachineFunction *F = BB->getParent();
18335 const X86InstrInfo *TII = Subtarget->getInstrInfo();
18336 DebugLoc DL = MI->getDebugLoc();
18338 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
18339 assert(MI->getOperand(3).isGlobal() && "This should be a global");
18341 // Get a register mask for the lowered call.
18342 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
18343 // proper register mask.
18344 const uint32_t *RegMask =
18345 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18346 if (Subtarget->is64Bit()) {
18347 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18348 TII->get(X86::MOV64rm), X86::RDI)
18350 .addImm(0).addReg(0)
18351 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18352 MI->getOperand(3).getTargetFlags())
18354 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
18355 addDirectMem(MIB, X86::RDI);
18356 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
18357 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
18358 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18359 TII->get(X86::MOV32rm), X86::EAX)
18361 .addImm(0).addReg(0)
18362 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18363 MI->getOperand(3).getTargetFlags())
18365 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18366 addDirectMem(MIB, X86::EAX);
18367 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18369 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18370 TII->get(X86::MOV32rm), X86::EAX)
18371 .addReg(TII->getGlobalBaseReg(F))
18372 .addImm(0).addReg(0)
18373 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18374 MI->getOperand(3).getTargetFlags())
18376 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18377 addDirectMem(MIB, X86::EAX);
18378 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18381 MI->eraseFromParent(); // The pseudo instruction is gone now.
18385 MachineBasicBlock *
18386 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
18387 MachineBasicBlock *MBB) const {
18388 DebugLoc DL = MI->getDebugLoc();
18389 MachineFunction *MF = MBB->getParent();
18390 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18391 MachineRegisterInfo &MRI = MF->getRegInfo();
18393 const BasicBlock *BB = MBB->getBasicBlock();
18394 MachineFunction::iterator I = MBB;
18397 // Memory Reference
18398 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18399 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18402 unsigned MemOpndSlot = 0;
18404 unsigned CurOp = 0;
18406 DstReg = MI->getOperand(CurOp++).getReg();
18407 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
18408 assert(RC->hasType(MVT::i32) && "Invalid destination!");
18409 unsigned mainDstReg = MRI.createVirtualRegister(RC);
18410 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
18412 MemOpndSlot = CurOp;
18414 MVT PVT = getPointerTy();
18415 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18416 "Invalid Pointer Size!");
18418 // For v = setjmp(buf), we generate
18421 // buf[LabelOffset] = restoreMBB
18422 // SjLjSetup restoreMBB
18428 // v = phi(main, restore)
18431 // if base pointer being used, load it from frame
18434 MachineBasicBlock *thisMBB = MBB;
18435 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18436 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18437 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
18438 MF->insert(I, mainMBB);
18439 MF->insert(I, sinkMBB);
18440 MF->push_back(restoreMBB);
18442 MachineInstrBuilder MIB;
18444 // Transfer the remainder of BB and its successor edges to sinkMBB.
18445 sinkMBB->splice(sinkMBB->begin(), MBB,
18446 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18447 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18450 unsigned PtrStoreOpc = 0;
18451 unsigned LabelReg = 0;
18452 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18453 Reloc::Model RM = MF->getTarget().getRelocationModel();
18454 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
18455 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
18457 // Prepare IP either in reg or imm.
18458 if (!UseImmLabel) {
18459 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
18460 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
18461 LabelReg = MRI.createVirtualRegister(PtrRC);
18462 if (Subtarget->is64Bit()) {
18463 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
18467 .addMBB(restoreMBB)
18470 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
18471 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
18472 .addReg(XII->getGlobalBaseReg(MF))
18475 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
18479 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
18481 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
18482 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18483 if (i == X86::AddrDisp)
18484 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
18486 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
18489 MIB.addReg(LabelReg);
18491 MIB.addMBB(restoreMBB);
18492 MIB.setMemRefs(MMOBegin, MMOEnd);
18494 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
18495 .addMBB(restoreMBB);
18497 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18498 MIB.addRegMask(RegInfo->getNoPreservedMask());
18499 thisMBB->addSuccessor(mainMBB);
18500 thisMBB->addSuccessor(restoreMBB);
18504 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
18505 mainMBB->addSuccessor(sinkMBB);
18508 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18509 TII->get(X86::PHI), DstReg)
18510 .addReg(mainDstReg).addMBB(mainMBB)
18511 .addReg(restoreDstReg).addMBB(restoreMBB);
18514 if (RegInfo->hasBasePointer(*MF)) {
18515 const bool Uses64BitFramePtr =
18516 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
18517 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
18518 X86FI->setRestoreBasePointer(MF);
18519 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
18520 unsigned BasePtr = RegInfo->getBaseRegister();
18521 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
18522 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
18523 FramePtr, true, X86FI->getRestoreBasePointerOffset())
18524 .setMIFlag(MachineInstr::FrameSetup);
18526 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
18527 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
18528 restoreMBB->addSuccessor(sinkMBB);
18530 MI->eraseFromParent();
18534 MachineBasicBlock *
18535 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
18536 MachineBasicBlock *MBB) const {
18537 DebugLoc DL = MI->getDebugLoc();
18538 MachineFunction *MF = MBB->getParent();
18539 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18540 MachineRegisterInfo &MRI = MF->getRegInfo();
18542 // Memory Reference
18543 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18544 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18546 MVT PVT = getPointerTy();
18547 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18548 "Invalid Pointer Size!");
18550 const TargetRegisterClass *RC =
18551 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
18552 unsigned Tmp = MRI.createVirtualRegister(RC);
18553 // Since FP is only updated here but NOT referenced, it's treated as GPR.
18554 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18555 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
18556 unsigned SP = RegInfo->getStackRegister();
18558 MachineInstrBuilder MIB;
18560 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18561 const int64_t SPOffset = 2 * PVT.getStoreSize();
18563 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
18564 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
18567 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
18568 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
18569 MIB.addOperand(MI->getOperand(i));
18570 MIB.setMemRefs(MMOBegin, MMOEnd);
18572 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
18573 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18574 if (i == X86::AddrDisp)
18575 MIB.addDisp(MI->getOperand(i), LabelOffset);
18577 MIB.addOperand(MI->getOperand(i));
18579 MIB.setMemRefs(MMOBegin, MMOEnd);
18581 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
18582 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18583 if (i == X86::AddrDisp)
18584 MIB.addDisp(MI->getOperand(i), SPOffset);
18586 MIB.addOperand(MI->getOperand(i));
18588 MIB.setMemRefs(MMOBegin, MMOEnd);
18590 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
18592 MI->eraseFromParent();
18596 // Replace 213-type (isel default) FMA3 instructions with 231-type for
18597 // accumulator loops. Writing back to the accumulator allows the coalescer
18598 // to remove extra copies in the loop.
18599 MachineBasicBlock *
18600 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
18601 MachineBasicBlock *MBB) const {
18602 MachineOperand &AddendOp = MI->getOperand(3);
18604 // Bail out early if the addend isn't a register - we can't switch these.
18605 if (!AddendOp.isReg())
18608 MachineFunction &MF = *MBB->getParent();
18609 MachineRegisterInfo &MRI = MF.getRegInfo();
18611 // Check whether the addend is defined by a PHI:
18612 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
18613 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
18614 if (!AddendDef.isPHI())
18617 // Look for the following pattern:
18619 // %addend = phi [%entry, 0], [%loop, %result]
18621 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
18625 // %addend = phi [%entry, 0], [%loop, %result]
18627 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
18629 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
18630 assert(AddendDef.getOperand(i).isReg());
18631 MachineOperand PHISrcOp = AddendDef.getOperand(i);
18632 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
18633 if (&PHISrcInst == MI) {
18634 // Found a matching instruction.
18635 unsigned NewFMAOpc = 0;
18636 switch (MI->getOpcode()) {
18637 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
18638 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
18639 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
18640 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
18641 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
18642 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
18643 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
18644 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
18645 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
18646 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
18647 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
18648 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
18649 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
18650 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
18651 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
18652 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
18653 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
18654 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
18655 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
18656 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
18658 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
18659 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
18660 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
18661 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
18662 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
18663 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
18664 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
18665 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
18666 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
18667 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
18668 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
18669 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
18670 default: llvm_unreachable("Unrecognized FMA variant.");
18673 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
18674 MachineInstrBuilder MIB =
18675 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
18676 .addOperand(MI->getOperand(0))
18677 .addOperand(MI->getOperand(3))
18678 .addOperand(MI->getOperand(2))
18679 .addOperand(MI->getOperand(1));
18680 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
18681 MI->eraseFromParent();
18688 MachineBasicBlock *
18689 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
18690 MachineBasicBlock *BB) const {
18691 switch (MI->getOpcode()) {
18692 default: llvm_unreachable("Unexpected instr type to insert");
18693 case X86::TAILJMPd64:
18694 case X86::TAILJMPr64:
18695 case X86::TAILJMPm64:
18696 case X86::TAILJMPd64_REX:
18697 case X86::TAILJMPr64_REX:
18698 case X86::TAILJMPm64_REX:
18699 llvm_unreachable("TAILJMP64 would not be touched here.");
18700 case X86::TCRETURNdi64:
18701 case X86::TCRETURNri64:
18702 case X86::TCRETURNmi64:
18704 case X86::WIN_ALLOCA:
18705 return EmitLoweredWinAlloca(MI, BB);
18706 case X86::SEG_ALLOCA_32:
18707 case X86::SEG_ALLOCA_64:
18708 return EmitLoweredSegAlloca(MI, BB);
18709 case X86::TLSCall_32:
18710 case X86::TLSCall_64:
18711 return EmitLoweredTLSCall(MI, BB);
18712 case X86::CMOV_GR8:
18713 case X86::CMOV_FR32:
18714 case X86::CMOV_FR64:
18715 case X86::CMOV_V4F32:
18716 case X86::CMOV_V2F64:
18717 case X86::CMOV_V2I64:
18718 case X86::CMOV_V8F32:
18719 case X86::CMOV_V4F64:
18720 case X86::CMOV_V4I64:
18721 case X86::CMOV_V16F32:
18722 case X86::CMOV_V8F64:
18723 case X86::CMOV_V8I64:
18724 case X86::CMOV_GR16:
18725 case X86::CMOV_GR32:
18726 case X86::CMOV_RFP32:
18727 case X86::CMOV_RFP64:
18728 case X86::CMOV_RFP80:
18729 return EmitLoweredSelect(MI, BB);
18731 case X86::FP32_TO_INT16_IN_MEM:
18732 case X86::FP32_TO_INT32_IN_MEM:
18733 case X86::FP32_TO_INT64_IN_MEM:
18734 case X86::FP64_TO_INT16_IN_MEM:
18735 case X86::FP64_TO_INT32_IN_MEM:
18736 case X86::FP64_TO_INT64_IN_MEM:
18737 case X86::FP80_TO_INT16_IN_MEM:
18738 case X86::FP80_TO_INT32_IN_MEM:
18739 case X86::FP80_TO_INT64_IN_MEM: {
18740 MachineFunction *F = BB->getParent();
18741 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18742 DebugLoc DL = MI->getDebugLoc();
18744 // Change the floating point control register to use "round towards zero"
18745 // mode when truncating to an integer value.
18746 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
18747 addFrameReference(BuildMI(*BB, MI, DL,
18748 TII->get(X86::FNSTCW16m)), CWFrameIdx);
18750 // Load the old value of the high byte of the control word...
18752 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
18753 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
18756 // Set the high part to be round to zero...
18757 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
18760 // Reload the modified control word now...
18761 addFrameReference(BuildMI(*BB, MI, DL,
18762 TII->get(X86::FLDCW16m)), CWFrameIdx);
18764 // Restore the memory image of control word to original value
18765 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
18768 // Get the X86 opcode to use.
18770 switch (MI->getOpcode()) {
18771 default: llvm_unreachable("illegal opcode!");
18772 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
18773 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
18774 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
18775 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
18776 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
18777 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
18778 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
18779 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
18780 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
18784 MachineOperand &Op = MI->getOperand(0);
18786 AM.BaseType = X86AddressMode::RegBase;
18787 AM.Base.Reg = Op.getReg();
18789 AM.BaseType = X86AddressMode::FrameIndexBase;
18790 AM.Base.FrameIndex = Op.getIndex();
18792 Op = MI->getOperand(1);
18794 AM.Scale = Op.getImm();
18795 Op = MI->getOperand(2);
18797 AM.IndexReg = Op.getImm();
18798 Op = MI->getOperand(3);
18799 if (Op.isGlobal()) {
18800 AM.GV = Op.getGlobal();
18802 AM.Disp = Op.getImm();
18804 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
18805 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
18807 // Reload the original control word now.
18808 addFrameReference(BuildMI(*BB, MI, DL,
18809 TII->get(X86::FLDCW16m)), CWFrameIdx);
18811 MI->eraseFromParent(); // The pseudo instruction is gone now.
18814 // String/text processing lowering.
18815 case X86::PCMPISTRM128REG:
18816 case X86::VPCMPISTRM128REG:
18817 case X86::PCMPISTRM128MEM:
18818 case X86::VPCMPISTRM128MEM:
18819 case X86::PCMPESTRM128REG:
18820 case X86::VPCMPESTRM128REG:
18821 case X86::PCMPESTRM128MEM:
18822 case X86::VPCMPESTRM128MEM:
18823 assert(Subtarget->hasSSE42() &&
18824 "Target must have SSE4.2 or AVX features enabled");
18825 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
18827 // String/text processing lowering.
18828 case X86::PCMPISTRIREG:
18829 case X86::VPCMPISTRIREG:
18830 case X86::PCMPISTRIMEM:
18831 case X86::VPCMPISTRIMEM:
18832 case X86::PCMPESTRIREG:
18833 case X86::VPCMPESTRIREG:
18834 case X86::PCMPESTRIMEM:
18835 case X86::VPCMPESTRIMEM:
18836 assert(Subtarget->hasSSE42() &&
18837 "Target must have SSE4.2 or AVX features enabled");
18838 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
18840 // Thread synchronization.
18842 return EmitMonitor(MI, BB, Subtarget);
18846 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
18848 case X86::VASTART_SAVE_XMM_REGS:
18849 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
18851 case X86::VAARG_64:
18852 return EmitVAARG64WithCustomInserter(MI, BB);
18854 case X86::EH_SjLj_SetJmp32:
18855 case X86::EH_SjLj_SetJmp64:
18856 return emitEHSjLjSetJmp(MI, BB);
18858 case X86::EH_SjLj_LongJmp32:
18859 case X86::EH_SjLj_LongJmp64:
18860 return emitEHSjLjLongJmp(MI, BB);
18862 case TargetOpcode::STATEPOINT:
18863 // As an implementation detail, STATEPOINT shares the STACKMAP format at
18864 // this point in the process. We diverge later.
18865 return emitPatchPoint(MI, BB);
18867 case TargetOpcode::STACKMAP:
18868 case TargetOpcode::PATCHPOINT:
18869 return emitPatchPoint(MI, BB);
18871 case X86::VFMADDPDr213r:
18872 case X86::VFMADDPSr213r:
18873 case X86::VFMADDSDr213r:
18874 case X86::VFMADDSSr213r:
18875 case X86::VFMSUBPDr213r:
18876 case X86::VFMSUBPSr213r:
18877 case X86::VFMSUBSDr213r:
18878 case X86::VFMSUBSSr213r:
18879 case X86::VFNMADDPDr213r:
18880 case X86::VFNMADDPSr213r:
18881 case X86::VFNMADDSDr213r:
18882 case X86::VFNMADDSSr213r:
18883 case X86::VFNMSUBPDr213r:
18884 case X86::VFNMSUBPSr213r:
18885 case X86::VFNMSUBSDr213r:
18886 case X86::VFNMSUBSSr213r:
18887 case X86::VFMADDSUBPDr213r:
18888 case X86::VFMADDSUBPSr213r:
18889 case X86::VFMSUBADDPDr213r:
18890 case X86::VFMSUBADDPSr213r:
18891 case X86::VFMADDPDr213rY:
18892 case X86::VFMADDPSr213rY:
18893 case X86::VFMSUBPDr213rY:
18894 case X86::VFMSUBPSr213rY:
18895 case X86::VFNMADDPDr213rY:
18896 case X86::VFNMADDPSr213rY:
18897 case X86::VFNMSUBPDr213rY:
18898 case X86::VFNMSUBPSr213rY:
18899 case X86::VFMADDSUBPDr213rY:
18900 case X86::VFMADDSUBPSr213rY:
18901 case X86::VFMSUBADDPDr213rY:
18902 case X86::VFMSUBADDPSr213rY:
18903 return emitFMA3Instr(MI, BB);
18907 //===----------------------------------------------------------------------===//
18908 // X86 Optimization Hooks
18909 //===----------------------------------------------------------------------===//
18911 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
18914 const SelectionDAG &DAG,
18915 unsigned Depth) const {
18916 unsigned BitWidth = KnownZero.getBitWidth();
18917 unsigned Opc = Op.getOpcode();
18918 assert((Opc >= ISD::BUILTIN_OP_END ||
18919 Opc == ISD::INTRINSIC_WO_CHAIN ||
18920 Opc == ISD::INTRINSIC_W_CHAIN ||
18921 Opc == ISD::INTRINSIC_VOID) &&
18922 "Should use MaskedValueIsZero if you don't know whether Op"
18923 " is a target node!");
18925 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
18939 // These nodes' second result is a boolean.
18940 if (Op.getResNo() == 0)
18943 case X86ISD::SETCC:
18944 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
18946 case ISD::INTRINSIC_WO_CHAIN: {
18947 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18948 unsigned NumLoBits = 0;
18951 case Intrinsic::x86_sse_movmsk_ps:
18952 case Intrinsic::x86_avx_movmsk_ps_256:
18953 case Intrinsic::x86_sse2_movmsk_pd:
18954 case Intrinsic::x86_avx_movmsk_pd_256:
18955 case Intrinsic::x86_mmx_pmovmskb:
18956 case Intrinsic::x86_sse2_pmovmskb_128:
18957 case Intrinsic::x86_avx2_pmovmskb: {
18958 // High bits of movmskp{s|d}, pmovmskb are known zero.
18960 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
18961 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
18962 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
18963 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
18964 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
18965 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
18966 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
18967 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
18969 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
18978 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
18980 const SelectionDAG &,
18981 unsigned Depth) const {
18982 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
18983 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
18984 return Op.getValueType().getScalarType().getSizeInBits();
18990 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
18991 /// node is a GlobalAddress + offset.
18992 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
18993 const GlobalValue* &GA,
18994 int64_t &Offset) const {
18995 if (N->getOpcode() == X86ISD::Wrapper) {
18996 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
18997 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
18998 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
19002 return TargetLowering::isGAPlusOffset(N, GA, Offset);
19005 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
19006 /// same as extracting the high 128-bit part of 256-bit vector and then
19007 /// inserting the result into the low part of a new 256-bit vector
19008 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
19009 EVT VT = SVOp->getValueType(0);
19010 unsigned NumElems = VT.getVectorNumElements();
19012 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19013 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
19014 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19015 SVOp->getMaskElt(j) >= 0)
19021 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
19022 /// same as extracting the low 128-bit part of 256-bit vector and then
19023 /// inserting the result into the high part of a new 256-bit vector
19024 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
19025 EVT VT = SVOp->getValueType(0);
19026 unsigned NumElems = VT.getVectorNumElements();
19028 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19029 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
19030 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
19031 SVOp->getMaskElt(j) >= 0)
19037 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
19038 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
19039 TargetLowering::DAGCombinerInfo &DCI,
19040 const X86Subtarget* Subtarget) {
19042 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19043 SDValue V1 = SVOp->getOperand(0);
19044 SDValue V2 = SVOp->getOperand(1);
19045 EVT VT = SVOp->getValueType(0);
19046 unsigned NumElems = VT.getVectorNumElements();
19048 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19049 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19053 // V UNDEF BUILD_VECTOR UNDEF
19055 // CONCAT_VECTOR CONCAT_VECTOR
19058 // RESULT: V + zero extended
19060 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19061 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19062 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19065 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19068 // To match the shuffle mask, the first half of the mask should
19069 // be exactly the first vector, and all the rest a splat with the
19070 // first element of the second one.
19071 for (unsigned i = 0; i != NumElems/2; ++i)
19072 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19073 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19076 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19077 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19078 if (Ld->hasNUsesOfValue(1, 0)) {
19079 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19080 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19082 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19084 Ld->getPointerInfo(),
19085 Ld->getAlignment(),
19086 false/*isVolatile*/, true/*ReadMem*/,
19087 false/*WriteMem*/);
19089 // Make sure the newly-created LOAD is in the same position as Ld in
19090 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19091 // and update uses of Ld's output chain to use the TokenFactor.
19092 if (Ld->hasAnyUseOfValue(1)) {
19093 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19094 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19095 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19096 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19097 SDValue(ResNode.getNode(), 1));
19100 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19104 // Emit a zeroed vector and insert the desired subvector on its
19106 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19107 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19108 return DCI.CombineTo(N, InsV);
19111 //===--------------------------------------------------------------------===//
19112 // Combine some shuffles into subvector extracts and inserts:
19115 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19116 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19117 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19118 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19119 return DCI.CombineTo(N, InsV);
19122 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19123 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19124 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19125 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19126 return DCI.CombineTo(N, InsV);
19132 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19135 /// This is the leaf of the recursive combinine below. When we have found some
19136 /// chain of single-use x86 shuffle instructions and accumulated the combined
19137 /// shuffle mask represented by them, this will try to pattern match that mask
19138 /// into either a single instruction if there is a special purpose instruction
19139 /// for this operation, or into a PSHUFB instruction which is a fully general
19140 /// instruction but should only be used to replace chains over a certain depth.
19141 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19142 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19143 TargetLowering::DAGCombinerInfo &DCI,
19144 const X86Subtarget *Subtarget) {
19145 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19147 // Find the operand that enters the chain. Note that multiple uses are OK
19148 // here, we're not going to remove the operand we find.
19149 SDValue Input = Op.getOperand(0);
19150 while (Input.getOpcode() == ISD::BITCAST)
19151 Input = Input.getOperand(0);
19153 MVT VT = Input.getSimpleValueType();
19154 MVT RootVT = Root.getSimpleValueType();
19157 // Just remove no-op shuffle masks.
19158 if (Mask.size() == 1) {
19159 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19164 // Use the float domain if the operand type is a floating point type.
19165 bool FloatDomain = VT.isFloatingPoint();
19167 // For floating point shuffles, we don't have free copies in the shuffle
19168 // instructions or the ability to load as part of the instruction, so
19169 // canonicalize their shuffles to UNPCK or MOV variants.
19171 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19172 // vectors because it can have a load folded into it that UNPCK cannot. This
19173 // doesn't preclude something switching to the shorter encoding post-RA.
19175 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
19176 bool Lo = Mask.equals(0, 0);
19179 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19180 // is no slower than UNPCKLPD but has the option to fold the input operand
19181 // into even an unaligned memory load.
19182 if (Lo && Subtarget->hasSSE3()) {
19183 Shuffle = X86ISD::MOVDDUP;
19184 ShuffleVT = MVT::v2f64;
19186 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
19187 // than the UNPCK variants.
19188 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
19189 ShuffleVT = MVT::v4f32;
19191 if (Depth == 1 && Root->getOpcode() == Shuffle)
19192 return false; // Nothing to do!
19193 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19194 DCI.AddToWorklist(Op.getNode());
19195 if (Shuffle == X86ISD::MOVDDUP)
19196 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19198 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19199 DCI.AddToWorklist(Op.getNode());
19200 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19204 if (Subtarget->hasSSE3() &&
19205 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
19206 bool Lo = Mask.equals(0, 0, 2, 2);
19207 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
19208 MVT ShuffleVT = MVT::v4f32;
19209 if (Depth == 1 && Root->getOpcode() == Shuffle)
19210 return false; // Nothing to do!
19211 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19212 DCI.AddToWorklist(Op.getNode());
19213 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19214 DCI.AddToWorklist(Op.getNode());
19215 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19219 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
19220 bool Lo = Mask.equals(0, 0, 1, 1);
19221 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19222 MVT ShuffleVT = MVT::v4f32;
19223 if (Depth == 1 && Root->getOpcode() == Shuffle)
19224 return false; // Nothing to do!
19225 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19226 DCI.AddToWorklist(Op.getNode());
19227 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19228 DCI.AddToWorklist(Op.getNode());
19229 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19235 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
19236 // variants as none of these have single-instruction variants that are
19237 // superior to the UNPCK formulation.
19238 if (!FloatDomain &&
19239 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
19240 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
19241 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
19242 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
19244 bool Lo = Mask[0] == 0;
19245 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19246 if (Depth == 1 && Root->getOpcode() == Shuffle)
19247 return false; // Nothing to do!
19249 switch (Mask.size()) {
19251 ShuffleVT = MVT::v8i16;
19254 ShuffleVT = MVT::v16i8;
19257 llvm_unreachable("Impossible mask size!");
19259 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19260 DCI.AddToWorklist(Op.getNode());
19261 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19262 DCI.AddToWorklist(Op.getNode());
19263 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19268 // Don't try to re-form single instruction chains under any circumstances now
19269 // that we've done encoding canonicalization for them.
19273 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
19274 // can replace them with a single PSHUFB instruction profitably. Intel's
19275 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
19276 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
19277 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
19278 SmallVector<SDValue, 16> PSHUFBMask;
19279 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
19280 int Ratio = 16 / Mask.size();
19281 for (unsigned i = 0; i < 16; ++i) {
19282 if (Mask[i / Ratio] == SM_SentinelUndef) {
19283 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
19286 int M = Mask[i / Ratio] != SM_SentinelZero
19287 ? Ratio * Mask[i / Ratio] + i % Ratio
19289 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
19291 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
19292 DCI.AddToWorklist(Op.getNode());
19293 SDValue PSHUFBMaskOp =
19294 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
19295 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
19296 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
19297 DCI.AddToWorklist(Op.getNode());
19298 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19303 // Failed to find any combines.
19307 /// \brief Fully generic combining of x86 shuffle instructions.
19309 /// This should be the last combine run over the x86 shuffle instructions. Once
19310 /// they have been fully optimized, this will recursively consider all chains
19311 /// of single-use shuffle instructions, build a generic model of the cumulative
19312 /// shuffle operation, and check for simpler instructions which implement this
19313 /// operation. We use this primarily for two purposes:
19315 /// 1) Collapse generic shuffles to specialized single instructions when
19316 /// equivalent. In most cases, this is just an encoding size win, but
19317 /// sometimes we will collapse multiple generic shuffles into a single
19318 /// special-purpose shuffle.
19319 /// 2) Look for sequences of shuffle instructions with 3 or more total
19320 /// instructions, and replace them with the slightly more expensive SSSE3
19321 /// PSHUFB instruction if available. We do this as the last combining step
19322 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
19323 /// a suitable short sequence of other instructions. The PHUFB will either
19324 /// use a register or have to read from memory and so is slightly (but only
19325 /// slightly) more expensive than the other shuffle instructions.
19327 /// Because this is inherently a quadratic operation (for each shuffle in
19328 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
19329 /// This should never be an issue in practice as the shuffle lowering doesn't
19330 /// produce sequences of more than 8 instructions.
19332 /// FIXME: We will currently miss some cases where the redundant shuffling
19333 /// would simplify under the threshold for PSHUFB formation because of
19334 /// combine-ordering. To fix this, we should do the redundant instruction
19335 /// combining in this recursive walk.
19336 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
19337 ArrayRef<int> RootMask,
19338 int Depth, bool HasPSHUFB,
19340 TargetLowering::DAGCombinerInfo &DCI,
19341 const X86Subtarget *Subtarget) {
19342 // Bound the depth of our recursive combine because this is ultimately
19343 // quadratic in nature.
19347 // Directly rip through bitcasts to find the underlying operand.
19348 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
19349 Op = Op.getOperand(0);
19351 MVT VT = Op.getSimpleValueType();
19352 if (!VT.isVector())
19353 return false; // Bail if we hit a non-vector.
19354 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
19355 // version should be added.
19356 if (VT.getSizeInBits() != 128)
19359 assert(Root.getSimpleValueType().isVector() &&
19360 "Shuffles operate on vector types!");
19361 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
19362 "Can only combine shuffles of the same vector register size.");
19364 if (!isTargetShuffle(Op.getOpcode()))
19366 SmallVector<int, 16> OpMask;
19368 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
19369 // We only can combine unary shuffles which we can decode the mask for.
19370 if (!HaveMask || !IsUnary)
19373 assert(VT.getVectorNumElements() == OpMask.size() &&
19374 "Different mask size from vector size!");
19375 assert(((RootMask.size() > OpMask.size() &&
19376 RootMask.size() % OpMask.size() == 0) ||
19377 (OpMask.size() > RootMask.size() &&
19378 OpMask.size() % RootMask.size() == 0) ||
19379 OpMask.size() == RootMask.size()) &&
19380 "The smaller number of elements must divide the larger.");
19381 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
19382 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
19383 assert(((RootRatio == 1 && OpRatio == 1) ||
19384 (RootRatio == 1) != (OpRatio == 1)) &&
19385 "Must not have a ratio for both incoming and op masks!");
19387 SmallVector<int, 16> Mask;
19388 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
19390 // Merge this shuffle operation's mask into our accumulated mask. Note that
19391 // this shuffle's mask will be the first applied to the input, followed by the
19392 // root mask to get us all the way to the root value arrangement. The reason
19393 // for this order is that we are recursing up the operation chain.
19394 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
19395 int RootIdx = i / RootRatio;
19396 if (RootMask[RootIdx] < 0) {
19397 // This is a zero or undef lane, we're done.
19398 Mask.push_back(RootMask[RootIdx]);
19402 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
19403 int OpIdx = RootMaskedIdx / OpRatio;
19404 if (OpMask[OpIdx] < 0) {
19405 // The incoming lanes are zero or undef, it doesn't matter which ones we
19407 Mask.push_back(OpMask[OpIdx]);
19411 // Ok, we have non-zero lanes, map them through.
19412 Mask.push_back(OpMask[OpIdx] * OpRatio +
19413 RootMaskedIdx % OpRatio);
19416 // See if we can recurse into the operand to combine more things.
19417 switch (Op.getOpcode()) {
19418 case X86ISD::PSHUFB:
19420 case X86ISD::PSHUFD:
19421 case X86ISD::PSHUFHW:
19422 case X86ISD::PSHUFLW:
19423 if (Op.getOperand(0).hasOneUse() &&
19424 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19425 HasPSHUFB, DAG, DCI, Subtarget))
19429 case X86ISD::UNPCKL:
19430 case X86ISD::UNPCKH:
19431 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
19432 // We can't check for single use, we have to check that this shuffle is the only user.
19433 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
19434 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19435 HasPSHUFB, DAG, DCI, Subtarget))
19440 // Minor canonicalization of the accumulated shuffle mask to make it easier
19441 // to match below. All this does is detect masks with squential pairs of
19442 // elements, and shrink them to the half-width mask. It does this in a loop
19443 // so it will reduce the size of the mask to the minimal width mask which
19444 // performs an equivalent shuffle.
19445 SmallVector<int, 16> WidenedMask;
19446 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
19447 Mask = std::move(WidenedMask);
19448 WidenedMask.clear();
19451 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
19455 /// \brief Get the PSHUF-style mask from PSHUF node.
19457 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
19458 /// PSHUF-style masks that can be reused with such instructions.
19459 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
19460 SmallVector<int, 4> Mask;
19462 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
19466 switch (N.getOpcode()) {
19467 case X86ISD::PSHUFD:
19469 case X86ISD::PSHUFLW:
19472 case X86ISD::PSHUFHW:
19473 Mask.erase(Mask.begin(), Mask.begin() + 4);
19474 for (int &M : Mask)
19478 llvm_unreachable("No valid shuffle instruction found!");
19482 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
19484 /// We walk up the chain and look for a combinable shuffle, skipping over
19485 /// shuffles that we could hoist this shuffle's transformation past without
19486 /// altering anything.
19488 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
19490 TargetLowering::DAGCombinerInfo &DCI) {
19491 assert(N.getOpcode() == X86ISD::PSHUFD &&
19492 "Called with something other than an x86 128-bit half shuffle!");
19495 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
19496 // of the shuffles in the chain so that we can form a fresh chain to replace
19498 SmallVector<SDValue, 8> Chain;
19499 SDValue V = N.getOperand(0);
19500 for (; V.hasOneUse(); V = V.getOperand(0)) {
19501 switch (V.getOpcode()) {
19503 return SDValue(); // Nothing combined!
19506 // Skip bitcasts as we always know the type for the target specific
19510 case X86ISD::PSHUFD:
19511 // Found another dword shuffle.
19514 case X86ISD::PSHUFLW:
19515 // Check that the low words (being shuffled) are the identity in the
19516 // dword shuffle, and the high words are self-contained.
19517 if (Mask[0] != 0 || Mask[1] != 1 ||
19518 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
19521 Chain.push_back(V);
19524 case X86ISD::PSHUFHW:
19525 // Check that the high words (being shuffled) are the identity in the
19526 // dword shuffle, and the low words are self-contained.
19527 if (Mask[2] != 2 || Mask[3] != 3 ||
19528 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
19531 Chain.push_back(V);
19534 case X86ISD::UNPCKL:
19535 case X86ISD::UNPCKH:
19536 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
19537 // shuffle into a preceding word shuffle.
19538 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
19541 // Search for a half-shuffle which we can combine with.
19542 unsigned CombineOp =
19543 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
19544 if (V.getOperand(0) != V.getOperand(1) ||
19545 !V->isOnlyUserOf(V.getOperand(0).getNode()))
19547 Chain.push_back(V);
19548 V = V.getOperand(0);
19550 switch (V.getOpcode()) {
19552 return SDValue(); // Nothing to combine.
19554 case X86ISD::PSHUFLW:
19555 case X86ISD::PSHUFHW:
19556 if (V.getOpcode() == CombineOp)
19559 Chain.push_back(V);
19563 V = V.getOperand(0);
19567 } while (V.hasOneUse());
19570 // Break out of the loop if we break out of the switch.
19574 if (!V.hasOneUse())
19575 // We fell out of the loop without finding a viable combining instruction.
19578 // Merge this node's mask and our incoming mask.
19579 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19580 for (int &M : Mask)
19582 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
19583 getV4X86ShuffleImm8ForMask(Mask, DAG));
19585 // Rebuild the chain around this new shuffle.
19586 while (!Chain.empty()) {
19587 SDValue W = Chain.pop_back_val();
19589 if (V.getValueType() != W.getOperand(0).getValueType())
19590 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
19592 switch (W.getOpcode()) {
19594 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
19596 case X86ISD::UNPCKL:
19597 case X86ISD::UNPCKH:
19598 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
19601 case X86ISD::PSHUFD:
19602 case X86ISD::PSHUFLW:
19603 case X86ISD::PSHUFHW:
19604 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
19608 if (V.getValueType() != N.getValueType())
19609 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
19611 // Return the new chain to replace N.
19615 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
19617 /// We walk up the chain, skipping shuffles of the other half and looking
19618 /// through shuffles which switch halves trying to find a shuffle of the same
19619 /// pair of dwords.
19620 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
19622 TargetLowering::DAGCombinerInfo &DCI) {
19624 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
19625 "Called with something other than an x86 128-bit half shuffle!");
19627 unsigned CombineOpcode = N.getOpcode();
19629 // Walk up a single-use chain looking for a combinable shuffle.
19630 SDValue V = N.getOperand(0);
19631 for (; V.hasOneUse(); V = V.getOperand(0)) {
19632 switch (V.getOpcode()) {
19634 return false; // Nothing combined!
19637 // Skip bitcasts as we always know the type for the target specific
19641 case X86ISD::PSHUFLW:
19642 case X86ISD::PSHUFHW:
19643 if (V.getOpcode() == CombineOpcode)
19646 // Other-half shuffles are no-ops.
19649 // Break out of the loop if we break out of the switch.
19653 if (!V.hasOneUse())
19654 // We fell out of the loop without finding a viable combining instruction.
19657 // Combine away the bottom node as its shuffle will be accumulated into
19658 // a preceding shuffle.
19659 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19661 // Record the old value.
19664 // Merge this node's mask and our incoming mask (adjusted to account for all
19665 // the pshufd instructions encountered).
19666 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19667 for (int &M : Mask)
19669 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
19670 getV4X86ShuffleImm8ForMask(Mask, DAG));
19672 // Check that the shuffles didn't cancel each other out. If not, we need to
19673 // combine to the new one.
19675 // Replace the combinable shuffle with the combined one, updating all users
19676 // so that we re-evaluate the chain here.
19677 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
19682 /// \brief Try to combine x86 target specific shuffles.
19683 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
19684 TargetLowering::DAGCombinerInfo &DCI,
19685 const X86Subtarget *Subtarget) {
19687 MVT VT = N.getSimpleValueType();
19688 SmallVector<int, 4> Mask;
19690 switch (N.getOpcode()) {
19691 case X86ISD::PSHUFD:
19692 case X86ISD::PSHUFLW:
19693 case X86ISD::PSHUFHW:
19694 Mask = getPSHUFShuffleMask(N);
19695 assert(Mask.size() == 4);
19701 // Nuke no-op shuffles that show up after combining.
19702 if (isNoopShuffleMask(Mask))
19703 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19705 // Look for simplifications involving one or two shuffle instructions.
19706 SDValue V = N.getOperand(0);
19707 switch (N.getOpcode()) {
19710 case X86ISD::PSHUFLW:
19711 case X86ISD::PSHUFHW:
19712 assert(VT == MVT::v8i16);
19715 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
19716 return SDValue(); // We combined away this shuffle, so we're done.
19718 // See if this reduces to a PSHUFD which is no more expensive and can
19719 // combine with more operations. Note that it has to at least flip the
19720 // dwords as otherwise it would have been removed as a no-op.
19721 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
19722 int DMask[] = {0, 1, 2, 3};
19723 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
19724 DMask[DOffset + 0] = DOffset + 1;
19725 DMask[DOffset + 1] = DOffset + 0;
19726 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
19727 DCI.AddToWorklist(V.getNode());
19728 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
19729 getV4X86ShuffleImm8ForMask(DMask, DAG));
19730 DCI.AddToWorklist(V.getNode());
19731 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
19734 // Look for shuffle patterns which can be implemented as a single unpack.
19735 // FIXME: This doesn't handle the location of the PSHUFD generically, and
19736 // only works when we have a PSHUFD followed by two half-shuffles.
19737 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
19738 (V.getOpcode() == X86ISD::PSHUFLW ||
19739 V.getOpcode() == X86ISD::PSHUFHW) &&
19740 V.getOpcode() != N.getOpcode() &&
19742 SDValue D = V.getOperand(0);
19743 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
19744 D = D.getOperand(0);
19745 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
19746 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19747 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
19748 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19749 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19751 for (int i = 0; i < 4; ++i) {
19752 WordMask[i + NOffset] = Mask[i] + NOffset;
19753 WordMask[i + VOffset] = VMask[i] + VOffset;
19755 // Map the word mask through the DWord mask.
19757 for (int i = 0; i < 8; ++i)
19758 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
19759 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
19760 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
19761 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
19762 std::begin(UnpackLoMask)) ||
19763 std::equal(std::begin(MappedMask), std::end(MappedMask),
19764 std::begin(UnpackHiMask))) {
19765 // We can replace all three shuffles with an unpack.
19766 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
19767 DCI.AddToWorklist(V.getNode());
19768 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
19770 DL, MVT::v8i16, V, V);
19777 case X86ISD::PSHUFD:
19778 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
19787 /// \brief Try to combine a shuffle into a target-specific add-sub node.
19789 /// We combine this directly on the abstract vector shuffle nodes so it is
19790 /// easier to generically match. We also insert dummy vector shuffle nodes for
19791 /// the operands which explicitly discard the lanes which are unused by this
19792 /// operation to try to flow through the rest of the combiner the fact that
19793 /// they're unused.
19794 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
19796 EVT VT = N->getValueType(0);
19798 // We only handle target-independent shuffles.
19799 // FIXME: It would be easy and harmless to use the target shuffle mask
19800 // extraction tool to support more.
19801 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
19804 auto *SVN = cast<ShuffleVectorSDNode>(N);
19805 ArrayRef<int> Mask = SVN->getMask();
19806 SDValue V1 = N->getOperand(0);
19807 SDValue V2 = N->getOperand(1);
19809 // We require the first shuffle operand to be the SUB node, and the second to
19810 // be the ADD node.
19811 // FIXME: We should support the commuted patterns.
19812 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
19815 // If there are other uses of these operations we can't fold them.
19816 if (!V1->hasOneUse() || !V2->hasOneUse())
19819 // Ensure that both operations have the same operands. Note that we can
19820 // commute the FADD operands.
19821 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
19822 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
19823 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
19826 // We're looking for blends between FADD and FSUB nodes. We insist on these
19827 // nodes being lined up in a specific expected pattern.
19828 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
19829 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
19830 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
19833 // Only specific types are legal at this point, assert so we notice if and
19834 // when these change.
19835 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
19836 VT == MVT::v4f64) &&
19837 "Unknown vector type encountered!");
19839 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
19842 /// PerformShuffleCombine - Performs several different shuffle combines.
19843 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
19844 TargetLowering::DAGCombinerInfo &DCI,
19845 const X86Subtarget *Subtarget) {
19847 SDValue N0 = N->getOperand(0);
19848 SDValue N1 = N->getOperand(1);
19849 EVT VT = N->getValueType(0);
19851 // Don't create instructions with illegal types after legalize types has run.
19852 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19853 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
19856 // If we have legalized the vector types, look for blends of FADD and FSUB
19857 // nodes that we can fuse into an ADDSUB node.
19858 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
19859 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
19862 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
19863 if (Subtarget->hasFp256() && VT.is256BitVector() &&
19864 N->getOpcode() == ISD::VECTOR_SHUFFLE)
19865 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
19867 // During Type Legalization, when promoting illegal vector types,
19868 // the backend might introduce new shuffle dag nodes and bitcasts.
19870 // This code performs the following transformation:
19871 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
19872 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
19874 // We do this only if both the bitcast and the BINOP dag nodes have
19875 // one use. Also, perform this transformation only if the new binary
19876 // operation is legal. This is to avoid introducing dag nodes that
19877 // potentially need to be further expanded (or custom lowered) into a
19878 // less optimal sequence of dag nodes.
19879 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
19880 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
19881 N0.getOpcode() == ISD::BITCAST) {
19882 SDValue BC0 = N0.getOperand(0);
19883 EVT SVT = BC0.getValueType();
19884 unsigned Opcode = BC0.getOpcode();
19885 unsigned NumElts = VT.getVectorNumElements();
19887 if (BC0.hasOneUse() && SVT.isVector() &&
19888 SVT.getVectorNumElements() * 2 == NumElts &&
19889 TLI.isOperationLegal(Opcode, VT)) {
19890 bool CanFold = false;
19902 unsigned SVTNumElts = SVT.getVectorNumElements();
19903 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19904 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
19905 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
19906 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
19907 CanFold = SVOp->getMaskElt(i) < 0;
19910 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
19911 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
19912 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
19913 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
19918 // Only handle 128 wide vector from here on.
19919 if (!VT.is128BitVector())
19922 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
19923 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
19924 // consecutive, non-overlapping, and in the right order.
19925 SmallVector<SDValue, 16> Elts;
19926 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
19927 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
19929 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
19933 if (isTargetShuffle(N->getOpcode())) {
19935 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
19936 if (Shuffle.getNode())
19939 // Try recursively combining arbitrary sequences of x86 shuffle
19940 // instructions into higher-order shuffles. We do this after combining
19941 // specific PSHUF instruction sequences into their minimal form so that we
19942 // can evaluate how many specialized shuffle instructions are involved in
19943 // a particular chain.
19944 SmallVector<int, 1> NonceMask; // Just a placeholder.
19945 NonceMask.push_back(0);
19946 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
19947 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
19949 return SDValue(); // This routine will use CombineTo to replace N.
19955 /// PerformTruncateCombine - Converts truncate operation to
19956 /// a sequence of vector shuffle operations.
19957 /// It is possible when we truncate 256-bit vector to 128-bit vector
19958 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
19959 TargetLowering::DAGCombinerInfo &DCI,
19960 const X86Subtarget *Subtarget) {
19964 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
19965 /// specific shuffle of a load can be folded into a single element load.
19966 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
19967 /// shuffles have been custom lowered so we need to handle those here.
19968 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
19969 TargetLowering::DAGCombinerInfo &DCI) {
19970 if (DCI.isBeforeLegalizeOps())
19973 SDValue InVec = N->getOperand(0);
19974 SDValue EltNo = N->getOperand(1);
19976 if (!isa<ConstantSDNode>(EltNo))
19979 EVT OriginalVT = InVec.getValueType();
19981 if (InVec.getOpcode() == ISD::BITCAST) {
19982 // Don't duplicate a load with other uses.
19983 if (!InVec.hasOneUse())
19985 EVT BCVT = InVec.getOperand(0).getValueType();
19986 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
19988 InVec = InVec.getOperand(0);
19991 EVT CurrentVT = InVec.getValueType();
19993 if (!isTargetShuffle(InVec.getOpcode()))
19996 // Don't duplicate a load with other uses.
19997 if (!InVec.hasOneUse())
20000 SmallVector<int, 16> ShuffleMask;
20002 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
20003 ShuffleMask, UnaryShuffle))
20006 // Select the input vector, guarding against out of range extract vector.
20007 unsigned NumElems = CurrentVT.getVectorNumElements();
20008 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
20009 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
20010 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
20011 : InVec.getOperand(1);
20013 // If inputs to shuffle are the same for both ops, then allow 2 uses
20014 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
20015 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
20017 if (LdNode.getOpcode() == ISD::BITCAST) {
20018 // Don't duplicate a load with other uses.
20019 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
20022 AllowedUses = 1; // only allow 1 load use if we have a bitcast
20023 LdNode = LdNode.getOperand(0);
20026 if (!ISD::isNormalLoad(LdNode.getNode()))
20029 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
20031 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
20034 EVT EltVT = N->getValueType(0);
20035 // If there's a bitcast before the shuffle, check if the load type and
20036 // alignment is valid.
20037 unsigned Align = LN0->getAlignment();
20038 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20039 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
20040 EltVT.getTypeForEVT(*DAG.getContext()));
20042 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
20045 // All checks match so transform back to vector_shuffle so that DAG combiner
20046 // can finish the job
20049 // Create shuffle node taking into account the case that its a unary shuffle
20050 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20051 : InVec.getOperand(1);
20052 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20053 InVec.getOperand(0), Shuffle,
20055 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20056 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20060 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20061 /// special and don't usually play with other vector types, it's better to
20062 /// handle them early to be sure we emit efficient code by avoiding
20063 /// store-load conversions.
20064 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20065 if (N->getValueType(0) != MVT::x86mmx ||
20066 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20067 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20070 SDValue V = N->getOperand(0);
20071 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20072 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20073 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20074 N->getValueType(0), V.getOperand(0));
20079 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20080 /// generation and convert it from being a bunch of shuffles and extracts
20081 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20082 /// storing the value and loading scalars back, while for x64 we should
20083 /// use 64-bit extracts and shifts.
20084 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20085 TargetLowering::DAGCombinerInfo &DCI) {
20086 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20087 if (NewOp.getNode())
20090 SDValue InputVector = N->getOperand(0);
20092 // Detect mmx to i32 conversion through a v2i32 elt extract.
20093 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20094 N->getValueType(0) == MVT::i32 &&
20095 InputVector.getValueType() == MVT::v2i32) {
20097 // The bitcast source is a direct mmx result.
20098 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20099 if (MMXSrc.getValueType() == MVT::x86mmx)
20100 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20101 N->getValueType(0),
20102 InputVector.getNode()->getOperand(0));
20104 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20105 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20106 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20107 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20108 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20109 MMXSrcOp.getValueType() == MVT::v1i64 &&
20110 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20111 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20112 N->getValueType(0),
20113 MMXSrcOp.getOperand(0));
20116 // Only operate on vectors of 4 elements, where the alternative shuffling
20117 // gets to be more expensive.
20118 if (InputVector.getValueType() != MVT::v4i32)
20121 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20122 // single use which is a sign-extend or zero-extend, and all elements are
20124 SmallVector<SDNode *, 4> Uses;
20125 unsigned ExtractedElements = 0;
20126 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20127 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20128 if (UI.getUse().getResNo() != InputVector.getResNo())
20131 SDNode *Extract = *UI;
20132 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20135 if (Extract->getValueType(0) != MVT::i32)
20137 if (!Extract->hasOneUse())
20139 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20140 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20142 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20145 // Record which element was extracted.
20146 ExtractedElements |=
20147 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20149 Uses.push_back(Extract);
20152 // If not all the elements were used, this may not be worthwhile.
20153 if (ExtractedElements != 15)
20156 // Ok, we've now decided to do the transformation.
20157 // If 64-bit shifts are legal, use the extract-shift sequence,
20158 // otherwise bounce the vector off the cache.
20159 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20161 SDLoc dl(InputVector);
20163 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20164 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20165 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20166 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20167 DAG.getConstant(0, VecIdxTy));
20168 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20169 DAG.getConstant(1, VecIdxTy));
20171 SDValue ShAmt = DAG.getConstant(32,
20172 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20173 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20174 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20175 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20176 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20177 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20178 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
20180 // Store the value to a temporary stack slot.
20181 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
20182 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
20183 MachinePointerInfo(), false, false, 0);
20185 EVT ElementType = InputVector.getValueType().getVectorElementType();
20186 unsigned EltSize = ElementType.getSizeInBits() / 8;
20188 // Replace each use (extract) with a load of the appropriate element.
20189 for (unsigned i = 0; i < 4; ++i) {
20190 uint64_t Offset = EltSize * i;
20191 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
20193 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
20194 StackPtr, OffsetVal);
20196 // Load the scalar.
20197 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
20198 ScalarAddr, MachinePointerInfo(),
20199 false, false, false, 0);
20204 // Replace the extracts
20205 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
20206 UE = Uses.end(); UI != UE; ++UI) {
20207 SDNode *Extract = *UI;
20209 SDValue Idx = Extract->getOperand(1);
20210 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
20211 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
20214 // The replacement was made in place; don't return anything.
20218 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
20219 static std::pair<unsigned, bool>
20220 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
20221 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
20222 if (!VT.isVector())
20223 return std::make_pair(0, false);
20225 bool NeedSplit = false;
20226 switch (VT.getSimpleVT().SimpleTy) {
20227 default: return std::make_pair(0, false);
20230 if (!Subtarget->hasVLX())
20231 return std::make_pair(0, false);
20235 if (!Subtarget->hasBWI())
20236 return std::make_pair(0, false);
20240 if (!Subtarget->hasAVX512())
20241 return std::make_pair(0, false);
20246 if (!Subtarget->hasAVX2())
20248 if (!Subtarget->hasAVX())
20249 return std::make_pair(0, false);
20254 if (!Subtarget->hasSSE2())
20255 return std::make_pair(0, false);
20258 // SSE2 has only a small subset of the operations.
20259 bool hasUnsigned = Subtarget->hasSSE41() ||
20260 (Subtarget->hasSSE2() && VT == MVT::v16i8);
20261 bool hasSigned = Subtarget->hasSSE41() ||
20262 (Subtarget->hasSSE2() && VT == MVT::v8i16);
20264 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20267 // Check for x CC y ? x : y.
20268 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20269 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20274 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20277 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20280 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20283 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20285 // Check for x CC y ? y : x -- a min/max with reversed arms.
20286 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20287 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20292 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20295 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20298 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20301 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20305 return std::make_pair(Opc, NeedSplit);
20309 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
20310 const X86Subtarget *Subtarget) {
20312 SDValue Cond = N->getOperand(0);
20313 SDValue LHS = N->getOperand(1);
20314 SDValue RHS = N->getOperand(2);
20316 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
20317 SDValue CondSrc = Cond->getOperand(0);
20318 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
20319 Cond = CondSrc->getOperand(0);
20322 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
20325 // A vselect where all conditions and data are constants can be optimized into
20326 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
20327 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
20328 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
20331 unsigned MaskValue = 0;
20332 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
20335 MVT VT = N->getSimpleValueType(0);
20336 unsigned NumElems = VT.getVectorNumElements();
20337 SmallVector<int, 8> ShuffleMask(NumElems, -1);
20338 for (unsigned i = 0; i < NumElems; ++i) {
20339 // Be sure we emit undef where we can.
20340 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
20341 ShuffleMask[i] = -1;
20343 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
20346 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20347 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
20349 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
20352 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
20354 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
20355 TargetLowering::DAGCombinerInfo &DCI,
20356 const X86Subtarget *Subtarget) {
20358 SDValue Cond = N->getOperand(0);
20359 // Get the LHS/RHS of the select.
20360 SDValue LHS = N->getOperand(1);
20361 SDValue RHS = N->getOperand(2);
20362 EVT VT = LHS.getValueType();
20363 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20365 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
20366 // instructions match the semantics of the common C idiom x<y?x:y but not
20367 // x<=y?x:y, because of how they handle negative zero (which can be
20368 // ignored in unsafe-math mode).
20369 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
20370 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
20371 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
20372 (Subtarget->hasSSE2() ||
20373 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
20374 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20376 unsigned Opcode = 0;
20377 // Check for x CC y ? x : y.
20378 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20379 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20383 // Converting this to a min would handle NaNs incorrectly, and swapping
20384 // the operands would cause it to handle comparisons between positive
20385 // and negative zero incorrectly.
20386 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20387 if (!DAG.getTarget().Options.UnsafeFPMath &&
20388 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20390 std::swap(LHS, RHS);
20392 Opcode = X86ISD::FMIN;
20395 // Converting this to a min would handle comparisons between positive
20396 // and negative zero incorrectly.
20397 if (!DAG.getTarget().Options.UnsafeFPMath &&
20398 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20400 Opcode = X86ISD::FMIN;
20403 // Converting this to a min would handle both negative zeros and NaNs
20404 // incorrectly, but we can swap the operands to fix both.
20405 std::swap(LHS, RHS);
20409 Opcode = X86ISD::FMIN;
20413 // Converting this to a max would handle comparisons between positive
20414 // and negative zero incorrectly.
20415 if (!DAG.getTarget().Options.UnsafeFPMath &&
20416 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20418 Opcode = X86ISD::FMAX;
20421 // Converting this to a max would handle NaNs incorrectly, and swapping
20422 // the operands would cause it to handle comparisons between positive
20423 // and negative zero incorrectly.
20424 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20425 if (!DAG.getTarget().Options.UnsafeFPMath &&
20426 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20428 std::swap(LHS, RHS);
20430 Opcode = X86ISD::FMAX;
20433 // Converting this to a max would handle both negative zeros and NaNs
20434 // incorrectly, but we can swap the operands to fix both.
20435 std::swap(LHS, RHS);
20439 Opcode = X86ISD::FMAX;
20442 // Check for x CC y ? y : x -- a min/max with reversed arms.
20443 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20444 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20448 // Converting this to a min would handle comparisons between positive
20449 // and negative zero incorrectly, and swapping the operands would
20450 // cause it to handle NaNs incorrectly.
20451 if (!DAG.getTarget().Options.UnsafeFPMath &&
20452 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
20453 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20455 std::swap(LHS, RHS);
20457 Opcode = X86ISD::FMIN;
20460 // Converting this to a min would handle NaNs incorrectly.
20461 if (!DAG.getTarget().Options.UnsafeFPMath &&
20462 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
20464 Opcode = X86ISD::FMIN;
20467 // Converting this to a min would handle both negative zeros and NaNs
20468 // incorrectly, but we can swap the operands to fix both.
20469 std::swap(LHS, RHS);
20473 Opcode = X86ISD::FMIN;
20477 // Converting this to a max would handle NaNs incorrectly.
20478 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20480 Opcode = X86ISD::FMAX;
20483 // Converting this to a max would handle comparisons between positive
20484 // and negative zero incorrectly, and swapping the operands would
20485 // cause it to handle NaNs incorrectly.
20486 if (!DAG.getTarget().Options.UnsafeFPMath &&
20487 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
20488 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20490 std::swap(LHS, RHS);
20492 Opcode = X86ISD::FMAX;
20495 // Converting this to a max would handle both negative zeros and NaNs
20496 // incorrectly, but we can swap the operands to fix both.
20497 std::swap(LHS, RHS);
20501 Opcode = X86ISD::FMAX;
20507 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
20510 EVT CondVT = Cond.getValueType();
20511 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
20512 CondVT.getVectorElementType() == MVT::i1) {
20513 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
20514 // lowering on KNL. In this case we convert it to
20515 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
20516 // The same situation for all 128 and 256-bit vectors of i8 and i16.
20517 // Since SKX these selects have a proper lowering.
20518 EVT OpVT = LHS.getValueType();
20519 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
20520 (OpVT.getVectorElementType() == MVT::i8 ||
20521 OpVT.getVectorElementType() == MVT::i16) &&
20522 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
20523 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
20524 DCI.AddToWorklist(Cond.getNode());
20525 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
20528 // If this is a select between two integer constants, try to do some
20530 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
20531 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
20532 // Don't do this for crazy integer types.
20533 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
20534 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
20535 // so that TrueC (the true value) is larger than FalseC.
20536 bool NeedsCondInvert = false;
20538 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
20539 // Efficiently invertible.
20540 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
20541 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
20542 isa<ConstantSDNode>(Cond.getOperand(1))))) {
20543 NeedsCondInvert = true;
20544 std::swap(TrueC, FalseC);
20547 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
20548 if (FalseC->getAPIntValue() == 0 &&
20549 TrueC->getAPIntValue().isPowerOf2()) {
20550 if (NeedsCondInvert) // Invert the condition if needed.
20551 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20552 DAG.getConstant(1, Cond.getValueType()));
20554 // Zero extend the condition if needed.
20555 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
20557 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
20558 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
20559 DAG.getConstant(ShAmt, MVT::i8));
20562 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
20563 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
20564 if (NeedsCondInvert) // Invert the condition if needed.
20565 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20566 DAG.getConstant(1, Cond.getValueType()));
20568 // Zero extend the condition if needed.
20569 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
20570 FalseC->getValueType(0), Cond);
20571 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20572 SDValue(FalseC, 0));
20575 // Optimize cases that will turn into an LEA instruction. This requires
20576 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
20577 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
20578 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
20579 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
20581 bool isFastMultiplier = false;
20583 switch ((unsigned char)Diff) {
20585 case 1: // result = add base, cond
20586 case 2: // result = lea base( , cond*2)
20587 case 3: // result = lea base(cond, cond*2)
20588 case 4: // result = lea base( , cond*4)
20589 case 5: // result = lea base(cond, cond*4)
20590 case 8: // result = lea base( , cond*8)
20591 case 9: // result = lea base(cond, cond*8)
20592 isFastMultiplier = true;
20597 if (isFastMultiplier) {
20598 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
20599 if (NeedsCondInvert) // Invert the condition if needed.
20600 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20601 DAG.getConstant(1, Cond.getValueType()));
20603 // Zero extend the condition if needed.
20604 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
20606 // Scale the condition by the difference.
20608 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
20609 DAG.getConstant(Diff, Cond.getValueType()));
20611 // Add the base if non-zero.
20612 if (FalseC->getAPIntValue() != 0)
20613 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20614 SDValue(FalseC, 0));
20621 // Canonicalize max and min:
20622 // (x > y) ? x : y -> (x >= y) ? x : y
20623 // (x < y) ? x : y -> (x <= y) ? x : y
20624 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
20625 // the need for an extra compare
20626 // against zero. e.g.
20627 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
20629 // testl %edi, %edi
20631 // cmovgl %edi, %eax
20635 // cmovsl %eax, %edi
20636 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
20637 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20638 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20639 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20644 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
20645 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
20646 Cond.getOperand(0), Cond.getOperand(1), NewCC);
20647 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
20652 // Early exit check
20653 if (!TLI.isTypeLegal(VT))
20656 // Match VSELECTs into subs with unsigned saturation.
20657 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
20658 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
20659 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
20660 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
20661 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20663 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
20664 // left side invert the predicate to simplify logic below.
20666 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
20668 CC = ISD::getSetCCInverse(CC, true);
20669 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
20673 if (Other.getNode() && Other->getNumOperands() == 2 &&
20674 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
20675 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
20676 SDValue CondRHS = Cond->getOperand(1);
20678 // Look for a general sub with unsigned saturation first.
20679 // x >= y ? x-y : 0 --> subus x, y
20680 // x > y ? x-y : 0 --> subus x, y
20681 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
20682 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
20683 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
20685 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
20686 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
20687 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
20688 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
20689 // If the RHS is a constant we have to reverse the const
20690 // canonicalization.
20691 // x > C-1 ? x+-C : 0 --> subus x, C
20692 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
20693 CondRHSConst->getAPIntValue() ==
20694 (-OpRHSConst->getAPIntValue() - 1))
20695 return DAG.getNode(
20696 X86ISD::SUBUS, DL, VT, OpLHS,
20697 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
20699 // Another special case: If C was a sign bit, the sub has been
20700 // canonicalized into a xor.
20701 // FIXME: Would it be better to use computeKnownBits to determine
20702 // whether it's safe to decanonicalize the xor?
20703 // x s< 0 ? x^C : 0 --> subus x, C
20704 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
20705 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
20706 OpRHSConst->getAPIntValue().isSignBit())
20707 // Note that we have to rebuild the RHS constant here to ensure we
20708 // don't rely on particular values of undef lanes.
20709 return DAG.getNode(
20710 X86ISD::SUBUS, DL, VT, OpLHS,
20711 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
20716 // Try to match a min/max vector operation.
20717 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
20718 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
20719 unsigned Opc = ret.first;
20720 bool NeedSplit = ret.second;
20722 if (Opc && NeedSplit) {
20723 unsigned NumElems = VT.getVectorNumElements();
20724 // Extract the LHS vectors
20725 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
20726 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
20728 // Extract the RHS vectors
20729 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
20730 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
20732 // Create min/max for each subvector
20733 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
20734 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
20736 // Merge the result
20737 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
20739 return DAG.getNode(Opc, DL, VT, LHS, RHS);
20742 // Simplify vector selection if condition value type matches vselect
20744 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
20745 assert(Cond.getValueType().isVector() &&
20746 "vector select expects a vector selector!");
20748 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
20749 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
20751 // Try invert the condition if true value is not all 1s and false value
20753 if (!TValIsAllOnes && !FValIsAllZeros &&
20754 // Check if the selector will be produced by CMPP*/PCMP*
20755 Cond.getOpcode() == ISD::SETCC &&
20756 // Check if SETCC has already been promoted
20757 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
20758 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
20759 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
20761 if (TValIsAllZeros || FValIsAllOnes) {
20762 SDValue CC = Cond.getOperand(2);
20763 ISD::CondCode NewCC =
20764 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
20765 Cond.getOperand(0).getValueType().isInteger());
20766 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
20767 std::swap(LHS, RHS);
20768 TValIsAllOnes = FValIsAllOnes;
20769 FValIsAllZeros = TValIsAllZeros;
20773 if (TValIsAllOnes || FValIsAllZeros) {
20776 if (TValIsAllOnes && FValIsAllZeros)
20778 else if (TValIsAllOnes)
20779 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
20780 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
20781 else if (FValIsAllZeros)
20782 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
20783 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
20785 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
20789 // If we know that this node is legal then we know that it is going to be
20790 // matched by one of the SSE/AVX BLEND instructions. These instructions only
20791 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
20792 // to simplify previous instructions.
20793 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
20794 !DCI.isBeforeLegalize() &&
20795 // We explicitly check against SSE4.1, v8i16 and v16i16 because, although
20796 // vselect nodes may be marked as Custom, they might only be legal when
20797 // Cond is a build_vector of constants. This will be taken care in
20798 // a later condition.
20799 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) &&
20800 Subtarget->hasSSE41() && VT != MVT::v16i16 && VT != MVT::v8i16) &&
20801 // Don't optimize vector of constants. Those are handled by
20802 // the generic code and all the bits must be properly set for
20803 // the generic optimizer.
20804 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
20805 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
20807 // Don't optimize vector selects that map to mask-registers.
20811 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
20812 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
20814 APInt KnownZero, KnownOne;
20815 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
20816 DCI.isBeforeLegalizeOps());
20817 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
20818 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
20820 // If we changed the computation somewhere in the DAG, this change
20821 // will affect all users of Cond.
20822 // Make sure it is fine and update all the nodes so that we do not
20823 // use the generic VSELECT anymore. Otherwise, we may perform
20824 // wrong optimizations as we messed up with the actual expectation
20825 // for the vector boolean values.
20826 if (Cond != TLO.Old) {
20827 // Check all uses of that condition operand to check whether it will be
20828 // consumed by non-BLEND instructions, which may depend on all bits are
20830 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20832 if (I->getOpcode() != ISD::VSELECT)
20833 // TODO: Add other opcodes eventually lowered into BLEND.
20836 // Update all the users of the condition, before committing the change,
20837 // so that the VSELECT optimizations that expect the correct vector
20838 // boolean value will not be triggered.
20839 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20841 DAG.ReplaceAllUsesOfValueWith(
20843 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
20844 Cond, I->getOperand(1), I->getOperand(2)));
20845 DCI.CommitTargetLoweringOpt(TLO);
20848 // At this point, only Cond is changed. Change the condition
20849 // just for N to keep the opportunity to optimize all other
20850 // users their own way.
20851 DAG.ReplaceAllUsesOfValueWith(
20853 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
20854 TLO.New, N->getOperand(1), N->getOperand(2)));
20859 // We should generate an X86ISD::BLENDI from a vselect if its argument
20860 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
20861 // constants. This specific pattern gets generated when we split a
20862 // selector for a 512 bit vector in a machine without AVX512 (but with
20863 // 256-bit vectors), during legalization:
20865 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
20867 // Iff we find this pattern and the build_vectors are built from
20868 // constants, we translate the vselect into a shuffle_vector that we
20869 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
20870 if ((N->getOpcode() == ISD::VSELECT ||
20871 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
20872 !DCI.isBeforeLegalize()) {
20873 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
20874 if (Shuffle.getNode())
20881 // Check whether a boolean test is testing a boolean value generated by
20882 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
20885 // Simplify the following patterns:
20886 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
20887 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
20888 // to (Op EFLAGS Cond)
20890 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
20891 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
20892 // to (Op EFLAGS !Cond)
20894 // where Op could be BRCOND or CMOV.
20896 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
20897 // Quit if not CMP and SUB with its value result used.
20898 if (Cmp.getOpcode() != X86ISD::CMP &&
20899 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
20902 // Quit if not used as a boolean value.
20903 if (CC != X86::COND_E && CC != X86::COND_NE)
20906 // Check CMP operands. One of them should be 0 or 1 and the other should be
20907 // an SetCC or extended from it.
20908 SDValue Op1 = Cmp.getOperand(0);
20909 SDValue Op2 = Cmp.getOperand(1);
20912 const ConstantSDNode* C = nullptr;
20913 bool needOppositeCond = (CC == X86::COND_E);
20914 bool checkAgainstTrue = false; // Is it a comparison against 1?
20916 if ((C = dyn_cast<ConstantSDNode>(Op1)))
20918 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
20920 else // Quit if all operands are not constants.
20923 if (C->getZExtValue() == 1) {
20924 needOppositeCond = !needOppositeCond;
20925 checkAgainstTrue = true;
20926 } else if (C->getZExtValue() != 0)
20927 // Quit if the constant is neither 0 or 1.
20930 bool truncatedToBoolWithAnd = false;
20931 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
20932 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
20933 SetCC.getOpcode() == ISD::TRUNCATE ||
20934 SetCC.getOpcode() == ISD::AND) {
20935 if (SetCC.getOpcode() == ISD::AND) {
20937 ConstantSDNode *CS;
20938 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
20939 CS->getZExtValue() == 1)
20941 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
20942 CS->getZExtValue() == 1)
20946 SetCC = SetCC.getOperand(OpIdx);
20947 truncatedToBoolWithAnd = true;
20949 SetCC = SetCC.getOperand(0);
20952 switch (SetCC.getOpcode()) {
20953 case X86ISD::SETCC_CARRY:
20954 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
20955 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
20956 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
20957 // truncated to i1 using 'and'.
20958 if (checkAgainstTrue && !truncatedToBoolWithAnd)
20960 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
20961 "Invalid use of SETCC_CARRY!");
20963 case X86ISD::SETCC:
20964 // Set the condition code or opposite one if necessary.
20965 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
20966 if (needOppositeCond)
20967 CC = X86::GetOppositeBranchCondition(CC);
20968 return SetCC.getOperand(1);
20969 case X86ISD::CMOV: {
20970 // Check whether false/true value has canonical one, i.e. 0 or 1.
20971 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
20972 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
20973 // Quit if true value is not a constant.
20976 // Quit if false value is not a constant.
20978 SDValue Op = SetCC.getOperand(0);
20979 // Skip 'zext' or 'trunc' node.
20980 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
20981 Op.getOpcode() == ISD::TRUNCATE)
20982 Op = Op.getOperand(0);
20983 // A special case for rdrand/rdseed, where 0 is set if false cond is
20985 if ((Op.getOpcode() != X86ISD::RDRAND &&
20986 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
20989 // Quit if false value is not the constant 0 or 1.
20990 bool FValIsFalse = true;
20991 if (FVal && FVal->getZExtValue() != 0) {
20992 if (FVal->getZExtValue() != 1)
20994 // If FVal is 1, opposite cond is needed.
20995 needOppositeCond = !needOppositeCond;
20996 FValIsFalse = false;
20998 // Quit if TVal is not the constant opposite of FVal.
20999 if (FValIsFalse && TVal->getZExtValue() != 1)
21001 if (!FValIsFalse && TVal->getZExtValue() != 0)
21003 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
21004 if (needOppositeCond)
21005 CC = X86::GetOppositeBranchCondition(CC);
21006 return SetCC.getOperand(3);
21013 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
21014 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
21015 TargetLowering::DAGCombinerInfo &DCI,
21016 const X86Subtarget *Subtarget) {
21019 // If the flag operand isn't dead, don't touch this CMOV.
21020 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
21023 SDValue FalseOp = N->getOperand(0);
21024 SDValue TrueOp = N->getOperand(1);
21025 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
21026 SDValue Cond = N->getOperand(3);
21028 if (CC == X86::COND_E || CC == X86::COND_NE) {
21029 switch (Cond.getOpcode()) {
21033 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
21034 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
21035 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21041 Flags = checkBoolTestSetCCCombine(Cond, CC);
21042 if (Flags.getNode() &&
21043 // Extra check as FCMOV only supports a subset of X86 cond.
21044 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21045 SDValue Ops[] = { FalseOp, TrueOp,
21046 DAG.getConstant(CC, MVT::i8), Flags };
21047 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21050 // If this is a select between two integer constants, try to do some
21051 // optimizations. Note that the operands are ordered the opposite of SELECT
21053 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21054 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21055 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21056 // larger than FalseC (the false value).
21057 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21058 CC = X86::GetOppositeBranchCondition(CC);
21059 std::swap(TrueC, FalseC);
21060 std::swap(TrueOp, FalseOp);
21063 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21064 // This is efficient for any integer data type (including i8/i16) and
21066 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21067 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21068 DAG.getConstant(CC, MVT::i8), Cond);
21070 // Zero extend the condition if needed.
21071 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21073 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21074 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21075 DAG.getConstant(ShAmt, MVT::i8));
21076 if (N->getNumValues() == 2) // Dead flag value?
21077 return DCI.CombineTo(N, Cond, SDValue());
21081 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21082 // for any integer data type, including i8/i16.
21083 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21084 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21085 DAG.getConstant(CC, MVT::i8), Cond);
21087 // Zero extend the condition if needed.
21088 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21089 FalseC->getValueType(0), Cond);
21090 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21091 SDValue(FalseC, 0));
21093 if (N->getNumValues() == 2) // Dead flag value?
21094 return DCI.CombineTo(N, Cond, SDValue());
21098 // Optimize cases that will turn into an LEA instruction. This requires
21099 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21100 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21101 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21102 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21104 bool isFastMultiplier = false;
21106 switch ((unsigned char)Diff) {
21108 case 1: // result = add base, cond
21109 case 2: // result = lea base( , cond*2)
21110 case 3: // result = lea base(cond, cond*2)
21111 case 4: // result = lea base( , cond*4)
21112 case 5: // result = lea base(cond, cond*4)
21113 case 8: // result = lea base( , cond*8)
21114 case 9: // result = lea base(cond, cond*8)
21115 isFastMultiplier = true;
21120 if (isFastMultiplier) {
21121 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21122 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21123 DAG.getConstant(CC, MVT::i8), Cond);
21124 // Zero extend the condition if needed.
21125 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21127 // Scale the condition by the difference.
21129 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21130 DAG.getConstant(Diff, Cond.getValueType()));
21132 // Add the base if non-zero.
21133 if (FalseC->getAPIntValue() != 0)
21134 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21135 SDValue(FalseC, 0));
21136 if (N->getNumValues() == 2) // Dead flag value?
21137 return DCI.CombineTo(N, Cond, SDValue());
21144 // Handle these cases:
21145 // (select (x != c), e, c) -> select (x != c), e, x),
21146 // (select (x == c), c, e) -> select (x == c), x, e)
21147 // where the c is an integer constant, and the "select" is the combination
21148 // of CMOV and CMP.
21150 // The rationale for this change is that the conditional-move from a constant
21151 // needs two instructions, however, conditional-move from a register needs
21152 // only one instruction.
21154 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
21155 // some instruction-combining opportunities. This opt needs to be
21156 // postponed as late as possible.
21158 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
21159 // the DCI.xxxx conditions are provided to postpone the optimization as
21160 // late as possible.
21162 ConstantSDNode *CmpAgainst = nullptr;
21163 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
21164 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
21165 !isa<ConstantSDNode>(Cond.getOperand(0))) {
21167 if (CC == X86::COND_NE &&
21168 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
21169 CC = X86::GetOppositeBranchCondition(CC);
21170 std::swap(TrueOp, FalseOp);
21173 if (CC == X86::COND_E &&
21174 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
21175 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
21176 DAG.getConstant(CC, MVT::i8), Cond };
21177 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
21185 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
21186 const X86Subtarget *Subtarget) {
21187 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
21189 default: return SDValue();
21190 // SSE/AVX/AVX2 blend intrinsics.
21191 case Intrinsic::x86_avx2_pblendvb:
21192 case Intrinsic::x86_avx2_pblendw:
21193 case Intrinsic::x86_avx2_pblendd_128:
21194 case Intrinsic::x86_avx2_pblendd_256:
21195 // Don't try to simplify this intrinsic if we don't have AVX2.
21196 if (!Subtarget->hasAVX2())
21199 case Intrinsic::x86_avx_blend_pd_256:
21200 case Intrinsic::x86_avx_blend_ps_256:
21201 case Intrinsic::x86_avx_blendv_pd_256:
21202 case Intrinsic::x86_avx_blendv_ps_256:
21203 // Don't try to simplify this intrinsic if we don't have AVX.
21204 if (!Subtarget->hasAVX())
21207 case Intrinsic::x86_sse41_pblendw:
21208 case Intrinsic::x86_sse41_blendpd:
21209 case Intrinsic::x86_sse41_blendps:
21210 case Intrinsic::x86_sse41_blendvps:
21211 case Intrinsic::x86_sse41_blendvpd:
21212 case Intrinsic::x86_sse41_pblendvb: {
21213 SDValue Op0 = N->getOperand(1);
21214 SDValue Op1 = N->getOperand(2);
21215 SDValue Mask = N->getOperand(3);
21217 // Don't try to simplify this intrinsic if we don't have SSE4.1.
21218 if (!Subtarget->hasSSE41())
21221 // fold (blend A, A, Mask) -> A
21224 // fold (blend A, B, allZeros) -> A
21225 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
21227 // fold (blend A, B, allOnes) -> B
21228 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
21231 // Simplify the case where the mask is a constant i32 value.
21232 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
21233 if (C->isNullValue())
21235 if (C->isAllOnesValue())
21242 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
21243 case Intrinsic::x86_sse2_psrai_w:
21244 case Intrinsic::x86_sse2_psrai_d:
21245 case Intrinsic::x86_avx2_psrai_w:
21246 case Intrinsic::x86_avx2_psrai_d:
21247 case Intrinsic::x86_sse2_psra_w:
21248 case Intrinsic::x86_sse2_psra_d:
21249 case Intrinsic::x86_avx2_psra_w:
21250 case Intrinsic::x86_avx2_psra_d: {
21251 SDValue Op0 = N->getOperand(1);
21252 SDValue Op1 = N->getOperand(2);
21253 EVT VT = Op0.getValueType();
21254 assert(VT.isVector() && "Expected a vector type!");
21256 if (isa<BuildVectorSDNode>(Op1))
21257 Op1 = Op1.getOperand(0);
21259 if (!isa<ConstantSDNode>(Op1))
21262 EVT SVT = VT.getVectorElementType();
21263 unsigned SVTBits = SVT.getSizeInBits();
21265 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
21266 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
21267 uint64_t ShAmt = C.getZExtValue();
21269 // Don't try to convert this shift into a ISD::SRA if the shift
21270 // count is bigger than or equal to the element size.
21271 if (ShAmt >= SVTBits)
21274 // Trivial case: if the shift count is zero, then fold this
21275 // into the first operand.
21279 // Replace this packed shift intrinsic with a target independent
21281 SDValue Splat = DAG.getConstant(C, VT);
21282 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
21287 /// PerformMulCombine - Optimize a single multiply with constant into two
21288 /// in order to implement it with two cheaper instructions, e.g.
21289 /// LEA + SHL, LEA + LEA.
21290 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
21291 TargetLowering::DAGCombinerInfo &DCI) {
21292 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
21295 EVT VT = N->getValueType(0);
21296 if (VT != MVT::i64 && VT != MVT::i32)
21299 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
21302 uint64_t MulAmt = C->getZExtValue();
21303 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
21306 uint64_t MulAmt1 = 0;
21307 uint64_t MulAmt2 = 0;
21308 if ((MulAmt % 9) == 0) {
21310 MulAmt2 = MulAmt / 9;
21311 } else if ((MulAmt % 5) == 0) {
21313 MulAmt2 = MulAmt / 5;
21314 } else if ((MulAmt % 3) == 0) {
21316 MulAmt2 = MulAmt / 3;
21319 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
21322 if (isPowerOf2_64(MulAmt2) &&
21323 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
21324 // If second multiplifer is pow2, issue it first. We want the multiply by
21325 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
21327 std::swap(MulAmt1, MulAmt2);
21330 if (isPowerOf2_64(MulAmt1))
21331 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
21332 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
21334 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
21335 DAG.getConstant(MulAmt1, VT));
21337 if (isPowerOf2_64(MulAmt2))
21338 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
21339 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
21341 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
21342 DAG.getConstant(MulAmt2, VT));
21344 // Do not add new nodes to DAG combiner worklist.
21345 DCI.CombineTo(N, NewMul, false);
21350 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
21351 SDValue N0 = N->getOperand(0);
21352 SDValue N1 = N->getOperand(1);
21353 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
21354 EVT VT = N0.getValueType();
21356 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
21357 // since the result of setcc_c is all zero's or all ones.
21358 if (VT.isInteger() && !VT.isVector() &&
21359 N1C && N0.getOpcode() == ISD::AND &&
21360 N0.getOperand(1).getOpcode() == ISD::Constant) {
21361 SDValue N00 = N0.getOperand(0);
21362 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
21363 ((N00.getOpcode() == ISD::ANY_EXTEND ||
21364 N00.getOpcode() == ISD::ZERO_EXTEND) &&
21365 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
21366 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
21367 APInt ShAmt = N1C->getAPIntValue();
21368 Mask = Mask.shl(ShAmt);
21370 return DAG.getNode(ISD::AND, SDLoc(N), VT,
21371 N00, DAG.getConstant(Mask, VT));
21375 // Hardware support for vector shifts is sparse which makes us scalarize the
21376 // vector operations in many cases. Also, on sandybridge ADD is faster than
21378 // (shl V, 1) -> add V,V
21379 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
21380 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
21381 assert(N0.getValueType().isVector() && "Invalid vector shift type");
21382 // We shift all of the values by one. In many cases we do not have
21383 // hardware support for this operation. This is better expressed as an ADD
21385 if (N1SplatC->getZExtValue() == 1)
21386 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
21392 /// \brief Returns a vector of 0s if the node in input is a vector logical
21393 /// shift by a constant amount which is known to be bigger than or equal
21394 /// to the vector element size in bits.
21395 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
21396 const X86Subtarget *Subtarget) {
21397 EVT VT = N->getValueType(0);
21399 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
21400 (!Subtarget->hasInt256() ||
21401 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
21404 SDValue Amt = N->getOperand(1);
21406 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
21407 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
21408 APInt ShiftAmt = AmtSplat->getAPIntValue();
21409 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
21411 // SSE2/AVX2 logical shifts always return a vector of 0s
21412 // if the shift amount is bigger than or equal to
21413 // the element size. The constant shift amount will be
21414 // encoded as a 8-bit immediate.
21415 if (ShiftAmt.trunc(8).uge(MaxAmount))
21416 return getZeroVector(VT, Subtarget, DAG, DL);
21422 /// PerformShiftCombine - Combine shifts.
21423 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
21424 TargetLowering::DAGCombinerInfo &DCI,
21425 const X86Subtarget *Subtarget) {
21426 if (N->getOpcode() == ISD::SHL) {
21427 SDValue V = PerformSHLCombine(N, DAG);
21428 if (V.getNode()) return V;
21431 if (N->getOpcode() != ISD::SRA) {
21432 // Try to fold this logical shift into a zero vector.
21433 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
21434 if (V.getNode()) return V;
21440 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
21441 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
21442 // and friends. Likewise for OR -> CMPNEQSS.
21443 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
21444 TargetLowering::DAGCombinerInfo &DCI,
21445 const X86Subtarget *Subtarget) {
21448 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
21449 // we're requiring SSE2 for both.
21450 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
21451 SDValue N0 = N->getOperand(0);
21452 SDValue N1 = N->getOperand(1);
21453 SDValue CMP0 = N0->getOperand(1);
21454 SDValue CMP1 = N1->getOperand(1);
21457 // The SETCCs should both refer to the same CMP.
21458 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
21461 SDValue CMP00 = CMP0->getOperand(0);
21462 SDValue CMP01 = CMP0->getOperand(1);
21463 EVT VT = CMP00.getValueType();
21465 if (VT == MVT::f32 || VT == MVT::f64) {
21466 bool ExpectingFlags = false;
21467 // Check for any users that want flags:
21468 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
21469 !ExpectingFlags && UI != UE; ++UI)
21470 switch (UI->getOpcode()) {
21475 ExpectingFlags = true;
21477 case ISD::CopyToReg:
21478 case ISD::SIGN_EXTEND:
21479 case ISD::ZERO_EXTEND:
21480 case ISD::ANY_EXTEND:
21484 if (!ExpectingFlags) {
21485 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
21486 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
21488 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
21489 X86::CondCode tmp = cc0;
21494 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
21495 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
21496 // FIXME: need symbolic constants for these magic numbers.
21497 // See X86ATTInstPrinter.cpp:printSSECC().
21498 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
21499 if (Subtarget->hasAVX512()) {
21500 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
21501 CMP01, DAG.getConstant(x86cc, MVT::i8));
21502 if (N->getValueType(0) != MVT::i1)
21503 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
21507 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
21508 CMP00.getValueType(), CMP00, CMP01,
21509 DAG.getConstant(x86cc, MVT::i8));
21511 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
21512 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
21514 if (is64BitFP && !Subtarget->is64Bit()) {
21515 // On a 32-bit target, we cannot bitcast the 64-bit float to a
21516 // 64-bit integer, since that's not a legal type. Since
21517 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
21518 // bits, but can do this little dance to extract the lowest 32 bits
21519 // and work with those going forward.
21520 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
21522 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
21524 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
21525 Vector32, DAG.getIntPtrConstant(0));
21529 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
21530 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
21531 DAG.getConstant(1, IntVT));
21532 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
21533 return OneBitOfTruth;
21541 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
21542 /// so it can be folded inside ANDNP.
21543 static bool CanFoldXORWithAllOnes(const SDNode *N) {
21544 EVT VT = N->getValueType(0);
21546 // Match direct AllOnes for 128 and 256-bit vectors
21547 if (ISD::isBuildVectorAllOnes(N))
21550 // Look through a bit convert.
21551 if (N->getOpcode() == ISD::BITCAST)
21552 N = N->getOperand(0).getNode();
21554 // Sometimes the operand may come from a insert_subvector building a 256-bit
21556 if (VT.is256BitVector() &&
21557 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
21558 SDValue V1 = N->getOperand(0);
21559 SDValue V2 = N->getOperand(1);
21561 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
21562 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
21563 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
21564 ISD::isBuildVectorAllOnes(V2.getNode()))
21571 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
21572 // register. In most cases we actually compare or select YMM-sized registers
21573 // and mixing the two types creates horrible code. This method optimizes
21574 // some of the transition sequences.
21575 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
21576 TargetLowering::DAGCombinerInfo &DCI,
21577 const X86Subtarget *Subtarget) {
21578 EVT VT = N->getValueType(0);
21579 if (!VT.is256BitVector())
21582 assert((N->getOpcode() == ISD::ANY_EXTEND ||
21583 N->getOpcode() == ISD::ZERO_EXTEND ||
21584 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
21586 SDValue Narrow = N->getOperand(0);
21587 EVT NarrowVT = Narrow->getValueType(0);
21588 if (!NarrowVT.is128BitVector())
21591 if (Narrow->getOpcode() != ISD::XOR &&
21592 Narrow->getOpcode() != ISD::AND &&
21593 Narrow->getOpcode() != ISD::OR)
21596 SDValue N0 = Narrow->getOperand(0);
21597 SDValue N1 = Narrow->getOperand(1);
21600 // The Left side has to be a trunc.
21601 if (N0.getOpcode() != ISD::TRUNCATE)
21604 // The type of the truncated inputs.
21605 EVT WideVT = N0->getOperand(0)->getValueType(0);
21609 // The right side has to be a 'trunc' or a constant vector.
21610 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
21611 ConstantSDNode *RHSConstSplat = nullptr;
21612 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
21613 RHSConstSplat = RHSBV->getConstantSplatNode();
21614 if (!RHSTrunc && !RHSConstSplat)
21617 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21619 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
21622 // Set N0 and N1 to hold the inputs to the new wide operation.
21623 N0 = N0->getOperand(0);
21624 if (RHSConstSplat) {
21625 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
21626 SDValue(RHSConstSplat, 0));
21627 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
21628 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
21629 } else if (RHSTrunc) {
21630 N1 = N1->getOperand(0);
21633 // Generate the wide operation.
21634 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
21635 unsigned Opcode = N->getOpcode();
21637 case ISD::ANY_EXTEND:
21639 case ISD::ZERO_EXTEND: {
21640 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
21641 APInt Mask = APInt::getAllOnesValue(InBits);
21642 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
21643 return DAG.getNode(ISD::AND, DL, VT,
21644 Op, DAG.getConstant(Mask, VT));
21646 case ISD::SIGN_EXTEND:
21647 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
21648 Op, DAG.getValueType(NarrowVT));
21650 llvm_unreachable("Unexpected opcode");
21654 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
21655 TargetLowering::DAGCombinerInfo &DCI,
21656 const X86Subtarget *Subtarget) {
21657 SDValue N0 = N->getOperand(0);
21658 SDValue N1 = N->getOperand(1);
21661 // A vector zext_in_reg may be represented as a shuffle,
21662 // feeding into a bitcast (this represents anyext) feeding into
21663 // an and with a mask.
21664 // We'd like to try to combine that into a shuffle with zero
21665 // plus a bitcast, removing the and.
21666 if (N0.getOpcode() != ISD::BITCAST ||
21667 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
21670 // The other side of the AND should be a splat of 2^C, where C
21671 // is the number of bits in the source type.
21672 if (N1.getOpcode() == ISD::BITCAST)
21673 N1 = N1.getOperand(0);
21674 if (N1.getOpcode() != ISD::BUILD_VECTOR)
21676 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
21678 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
21679 EVT SrcType = Shuffle->getValueType(0);
21681 // We expect a single-source shuffle
21682 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
21685 unsigned SrcSize = SrcType.getScalarSizeInBits();
21687 APInt SplatValue, SplatUndef;
21688 unsigned SplatBitSize;
21690 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
21691 SplatBitSize, HasAnyUndefs))
21694 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
21695 // Make sure the splat matches the mask we expect
21696 if (SplatBitSize > ResSize ||
21697 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
21700 // Make sure the input and output size make sense
21701 if (SrcSize >= ResSize || ResSize % SrcSize)
21704 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
21705 // The number of u's between each two values depends on the ratio between
21706 // the source and dest type.
21707 unsigned ZextRatio = ResSize / SrcSize;
21708 bool IsZext = true;
21709 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
21710 if (i % ZextRatio) {
21711 if (Shuffle->getMaskElt(i) > 0) {
21717 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
21718 // Expected element number
21728 // Ok, perform the transformation - replace the shuffle with
21729 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
21730 // (instead of undef) where the k elements come from the zero vector.
21731 SmallVector<int, 8> Mask;
21732 unsigned NumElems = SrcType.getVectorNumElements();
21733 for (unsigned i = 0; i < NumElems; ++i)
21735 Mask.push_back(NumElems);
21737 Mask.push_back(i / ZextRatio);
21739 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
21740 Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
21741 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
21744 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
21745 TargetLowering::DAGCombinerInfo &DCI,
21746 const X86Subtarget *Subtarget) {
21747 if (DCI.isBeforeLegalizeOps())
21750 SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget);
21751 if (Zext.getNode())
21754 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21758 EVT VT = N->getValueType(0);
21759 SDValue N0 = N->getOperand(0);
21760 SDValue N1 = N->getOperand(1);
21763 // Create BEXTR instructions
21764 // BEXTR is ((X >> imm) & (2**size-1))
21765 if (VT == MVT::i32 || VT == MVT::i64) {
21766 // Check for BEXTR.
21767 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
21768 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
21769 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
21770 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
21771 if (MaskNode && ShiftNode) {
21772 uint64_t Mask = MaskNode->getZExtValue();
21773 uint64_t Shift = ShiftNode->getZExtValue();
21774 if (isMask_64(Mask)) {
21775 uint64_t MaskSize = countPopulation(Mask);
21776 if (Shift + MaskSize <= VT.getSizeInBits())
21777 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
21778 DAG.getConstant(Shift | (MaskSize << 8), VT));
21786 // Want to form ANDNP nodes:
21787 // 1) In the hopes of then easily combining them with OR and AND nodes
21788 // to form PBLEND/PSIGN.
21789 // 2) To match ANDN packed intrinsics
21790 if (VT != MVT::v2i64 && VT != MVT::v4i64)
21793 // Check LHS for vnot
21794 if (N0.getOpcode() == ISD::XOR &&
21795 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
21796 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
21797 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
21799 // Check RHS for vnot
21800 if (N1.getOpcode() == ISD::XOR &&
21801 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
21802 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
21803 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
21808 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
21809 TargetLowering::DAGCombinerInfo &DCI,
21810 const X86Subtarget *Subtarget) {
21811 if (DCI.isBeforeLegalizeOps())
21814 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21818 SDValue N0 = N->getOperand(0);
21819 SDValue N1 = N->getOperand(1);
21820 EVT VT = N->getValueType(0);
21822 // look for psign/blend
21823 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
21824 if (!Subtarget->hasSSSE3() ||
21825 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
21828 // Canonicalize pandn to RHS
21829 if (N0.getOpcode() == X86ISD::ANDNP)
21831 // or (and (m, y), (pandn m, x))
21832 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
21833 SDValue Mask = N1.getOperand(0);
21834 SDValue X = N1.getOperand(1);
21836 if (N0.getOperand(0) == Mask)
21837 Y = N0.getOperand(1);
21838 if (N0.getOperand(1) == Mask)
21839 Y = N0.getOperand(0);
21841 // Check to see if the mask appeared in both the AND and ANDNP and
21845 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
21846 // Look through mask bitcast.
21847 if (Mask.getOpcode() == ISD::BITCAST)
21848 Mask = Mask.getOperand(0);
21849 if (X.getOpcode() == ISD::BITCAST)
21850 X = X.getOperand(0);
21851 if (Y.getOpcode() == ISD::BITCAST)
21852 Y = Y.getOperand(0);
21854 EVT MaskVT = Mask.getValueType();
21856 // Validate that the Mask operand is a vector sra node.
21857 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
21858 // there is no psrai.b
21859 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
21860 unsigned SraAmt = ~0;
21861 if (Mask.getOpcode() == ISD::SRA) {
21862 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
21863 if (auto *AmtConst = AmtBV->getConstantSplatNode())
21864 SraAmt = AmtConst->getZExtValue();
21865 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
21866 SDValue SraC = Mask.getOperand(1);
21867 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
21869 if ((SraAmt + 1) != EltBits)
21874 // Now we know we at least have a plendvb with the mask val. See if
21875 // we can form a psignb/w/d.
21876 // psign = x.type == y.type == mask.type && y = sub(0, x);
21877 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
21878 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
21879 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
21880 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
21881 "Unsupported VT for PSIGN");
21882 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
21883 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21885 // PBLENDVB only available on SSE 4.1
21886 if (!Subtarget->hasSSE41())
21889 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
21891 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
21892 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
21893 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
21894 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
21895 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21899 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
21902 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
21903 MachineFunction &MF = DAG.getMachineFunction();
21905 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
21907 // SHLD/SHRD instructions have lower register pressure, but on some
21908 // platforms they have higher latency than the equivalent
21909 // series of shifts/or that would otherwise be generated.
21910 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
21911 // have higher latencies and we are not optimizing for size.
21912 if (!OptForSize && Subtarget->isSHLDSlow())
21915 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
21917 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
21919 if (!N0.hasOneUse() || !N1.hasOneUse())
21922 SDValue ShAmt0 = N0.getOperand(1);
21923 if (ShAmt0.getValueType() != MVT::i8)
21925 SDValue ShAmt1 = N1.getOperand(1);
21926 if (ShAmt1.getValueType() != MVT::i8)
21928 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
21929 ShAmt0 = ShAmt0.getOperand(0);
21930 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
21931 ShAmt1 = ShAmt1.getOperand(0);
21934 unsigned Opc = X86ISD::SHLD;
21935 SDValue Op0 = N0.getOperand(0);
21936 SDValue Op1 = N1.getOperand(0);
21937 if (ShAmt0.getOpcode() == ISD::SUB) {
21938 Opc = X86ISD::SHRD;
21939 std::swap(Op0, Op1);
21940 std::swap(ShAmt0, ShAmt1);
21943 unsigned Bits = VT.getSizeInBits();
21944 if (ShAmt1.getOpcode() == ISD::SUB) {
21945 SDValue Sum = ShAmt1.getOperand(0);
21946 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
21947 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
21948 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
21949 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
21950 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
21951 return DAG.getNode(Opc, DL, VT,
21953 DAG.getNode(ISD::TRUNCATE, DL,
21956 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
21957 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
21959 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
21960 return DAG.getNode(Opc, DL, VT,
21961 N0.getOperand(0), N1.getOperand(0),
21962 DAG.getNode(ISD::TRUNCATE, DL,
21969 // Generate NEG and CMOV for integer abs.
21970 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
21971 EVT VT = N->getValueType(0);
21973 // Since X86 does not have CMOV for 8-bit integer, we don't convert
21974 // 8-bit integer abs to NEG and CMOV.
21975 if (VT.isInteger() && VT.getSizeInBits() == 8)
21978 SDValue N0 = N->getOperand(0);
21979 SDValue N1 = N->getOperand(1);
21982 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
21983 // and change it to SUB and CMOV.
21984 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
21985 N0.getOpcode() == ISD::ADD &&
21986 N0.getOperand(1) == N1 &&
21987 N1.getOpcode() == ISD::SRA &&
21988 N1.getOperand(0) == N0.getOperand(0))
21989 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
21990 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
21991 // Generate SUB & CMOV.
21992 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
21993 DAG.getConstant(0, VT), N0.getOperand(0));
21995 SDValue Ops[] = { N0.getOperand(0), Neg,
21996 DAG.getConstant(X86::COND_GE, MVT::i8),
21997 SDValue(Neg.getNode(), 1) };
21998 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
22003 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
22004 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
22005 TargetLowering::DAGCombinerInfo &DCI,
22006 const X86Subtarget *Subtarget) {
22007 if (DCI.isBeforeLegalizeOps())
22010 if (Subtarget->hasCMov()) {
22011 SDValue RV = performIntegerAbsCombine(N, DAG);
22019 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
22020 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
22021 TargetLowering::DAGCombinerInfo &DCI,
22022 const X86Subtarget *Subtarget) {
22023 LoadSDNode *Ld = cast<LoadSDNode>(N);
22024 EVT RegVT = Ld->getValueType(0);
22025 EVT MemVT = Ld->getMemoryVT();
22027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22029 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
22030 // into two 16-byte operations.
22031 ISD::LoadExtType Ext = Ld->getExtensionType();
22032 unsigned Alignment = Ld->getAlignment();
22033 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
22034 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22035 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
22036 unsigned NumElems = RegVT.getVectorNumElements();
22040 SDValue Ptr = Ld->getBasePtr();
22041 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
22043 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
22045 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22046 Ld->getPointerInfo(), Ld->isVolatile(),
22047 Ld->isNonTemporal(), Ld->isInvariant(),
22049 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22050 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22051 Ld->getPointerInfo(), Ld->isVolatile(),
22052 Ld->isNonTemporal(), Ld->isInvariant(),
22053 std::min(16U, Alignment));
22054 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22056 Load2.getValue(1));
22058 SDValue NewVec = DAG.getUNDEF(RegVT);
22059 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22060 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22061 return DCI.CombineTo(N, NewVec, TF, true);
22067 /// PerformMLOADCombine - Resolve extending loads
22068 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22069 TargetLowering::DAGCombinerInfo &DCI,
22070 const X86Subtarget *Subtarget) {
22071 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22072 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22075 EVT VT = Mld->getValueType(0);
22076 unsigned NumElems = VT.getVectorNumElements();
22077 EVT LdVT = Mld->getMemoryVT();
22080 assert(LdVT != VT && "Cannot extend to the same type");
22081 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
22082 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
22083 // From, To sizes and ElemCount must be pow of two
22084 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22085 "Unexpected size for extending masked load");
22087 unsigned SizeRatio = ToSz / FromSz;
22088 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
22090 // Create a type on which we perform the shuffle
22091 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22092 LdVT.getScalarType(), NumElems*SizeRatio);
22093 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22095 // Convert Src0 value
22096 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
22097 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
22098 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22099 for (unsigned i = 0; i != NumElems; ++i)
22100 ShuffleVec[i] = i * SizeRatio;
22102 // Can't shuffle using an illegal type.
22103 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22104 && "WideVecVT should be legal");
22105 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
22106 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
22108 // Prepare the new mask
22110 SDValue Mask = Mld->getMask();
22111 if (Mask.getValueType() == VT) {
22112 // Mask and original value have the same type
22113 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22114 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22115 for (unsigned i = 0; i != NumElems; ++i)
22116 ShuffleVec[i] = i * SizeRatio;
22117 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22118 ShuffleVec[i] = NumElems*SizeRatio;
22119 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22120 DAG.getConstant(0, WideVecVT),
22124 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22125 unsigned WidenNumElts = NumElems*SizeRatio;
22126 unsigned MaskNumElts = VT.getVectorNumElements();
22127 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22130 unsigned NumConcat = WidenNumElts / MaskNumElts;
22131 SmallVector<SDValue, 16> Ops(NumConcat);
22132 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22134 for (unsigned i = 1; i != NumConcat; ++i)
22137 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22140 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
22141 Mld->getBasePtr(), NewMask, WideSrc0,
22142 Mld->getMemoryVT(), Mld->getMemOperand(),
22144 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
22145 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
22148 /// PerformMSTORECombine - Resolve truncating stores
22149 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
22150 const X86Subtarget *Subtarget) {
22151 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
22152 if (!Mst->isTruncatingStore())
22155 EVT VT = Mst->getValue().getValueType();
22156 unsigned NumElems = VT.getVectorNumElements();
22157 EVT StVT = Mst->getMemoryVT();
22160 assert(StVT != VT && "Cannot truncate to the same type");
22161 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22162 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22164 // From, To sizes and ElemCount must be pow of two
22165 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22166 "Unexpected size for truncating masked store");
22167 // We are going to use the original vector elt for storing.
22168 // Accumulated smaller vector elements must be a multiple of the store size.
22169 assert (((NumElems * FromSz) % ToSz) == 0 &&
22170 "Unexpected ratio for truncating masked store");
22172 unsigned SizeRatio = FromSz / ToSz;
22173 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22175 // Create a type on which we perform the shuffle
22176 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22177 StVT.getScalarType(), NumElems*SizeRatio);
22179 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22181 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
22182 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22183 for (unsigned i = 0; i != NumElems; ++i)
22184 ShuffleVec[i] = i * SizeRatio;
22186 // Can't shuffle using an illegal type.
22187 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22188 && "WideVecVT should be legal");
22190 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22191 DAG.getUNDEF(WideVecVT),
22195 SDValue Mask = Mst->getMask();
22196 if (Mask.getValueType() == VT) {
22197 // Mask and original value have the same type
22198 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22199 for (unsigned i = 0; i != NumElems; ++i)
22200 ShuffleVec[i] = i * SizeRatio;
22201 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22202 ShuffleVec[i] = NumElems*SizeRatio;
22203 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22204 DAG.getConstant(0, WideVecVT),
22208 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22209 unsigned WidenNumElts = NumElems*SizeRatio;
22210 unsigned MaskNumElts = VT.getVectorNumElements();
22211 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22214 unsigned NumConcat = WidenNumElts / MaskNumElts;
22215 SmallVector<SDValue, 16> Ops(NumConcat);
22216 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22218 for (unsigned i = 1; i != NumConcat; ++i)
22221 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22224 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
22225 NewMask, StVT, Mst->getMemOperand(), false);
22227 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
22228 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
22229 const X86Subtarget *Subtarget) {
22230 StoreSDNode *St = cast<StoreSDNode>(N);
22231 EVT VT = St->getValue().getValueType();
22232 EVT StVT = St->getMemoryVT();
22234 SDValue StoredVal = St->getOperand(1);
22235 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22237 // If we are saving a concatenation of two XMM registers and 32-byte stores
22238 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
22239 unsigned Alignment = St->getAlignment();
22240 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
22241 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22242 StVT == VT && !IsAligned) {
22243 unsigned NumElems = VT.getVectorNumElements();
22247 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
22248 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
22250 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
22251 SDValue Ptr0 = St->getBasePtr();
22252 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
22254 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
22255 St->getPointerInfo(), St->isVolatile(),
22256 St->isNonTemporal(), Alignment);
22257 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
22258 St->getPointerInfo(), St->isVolatile(),
22259 St->isNonTemporal(),
22260 std::min(16U, Alignment));
22261 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
22264 // Optimize trunc store (of multiple scalars) to shuffle and store.
22265 // First, pack all of the elements in one place. Next, store to memory
22266 // in fewer chunks.
22267 if (St->isTruncatingStore() && VT.isVector()) {
22268 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22269 unsigned NumElems = VT.getVectorNumElements();
22270 assert(StVT != VT && "Cannot truncate to the same type");
22271 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22272 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22274 // From, To sizes and ElemCount must be pow of two
22275 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
22276 // We are going to use the original vector elt for storing.
22277 // Accumulated smaller vector elements must be a multiple of the store size.
22278 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
22280 unsigned SizeRatio = FromSz / ToSz;
22282 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22284 // Create a type on which we perform the shuffle
22285 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22286 StVT.getScalarType(), NumElems*SizeRatio);
22288 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22290 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
22291 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
22292 for (unsigned i = 0; i != NumElems; ++i)
22293 ShuffleVec[i] = i * SizeRatio;
22295 // Can't shuffle using an illegal type.
22296 if (!TLI.isTypeLegal(WideVecVT))
22299 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22300 DAG.getUNDEF(WideVecVT),
22302 // At this point all of the data is stored at the bottom of the
22303 // register. We now need to save it to mem.
22305 // Find the largest store unit
22306 MVT StoreType = MVT::i8;
22307 for (MVT Tp : MVT::integer_valuetypes()) {
22308 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
22312 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
22313 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
22314 (64 <= NumElems * ToSz))
22315 StoreType = MVT::f64;
22317 // Bitcast the original vector into a vector of store-size units
22318 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
22319 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
22320 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
22321 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
22322 SmallVector<SDValue, 8> Chains;
22323 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
22324 TLI.getPointerTy());
22325 SDValue Ptr = St->getBasePtr();
22327 // Perform one or more big stores into memory.
22328 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
22329 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
22330 StoreType, ShuffWide,
22331 DAG.getIntPtrConstant(i));
22332 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
22333 St->getPointerInfo(), St->isVolatile(),
22334 St->isNonTemporal(), St->getAlignment());
22335 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22336 Chains.push_back(Ch);
22339 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
22342 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
22343 // the FP state in cases where an emms may be missing.
22344 // A preferable solution to the general problem is to figure out the right
22345 // places to insert EMMS. This qualifies as a quick hack.
22347 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
22348 if (VT.getSizeInBits() != 64)
22351 const Function *F = DAG.getMachineFunction().getFunction();
22352 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
22353 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
22354 && Subtarget->hasSSE2();
22355 if ((VT.isVector() ||
22356 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
22357 isa<LoadSDNode>(St->getValue()) &&
22358 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
22359 St->getChain().hasOneUse() && !St->isVolatile()) {
22360 SDNode* LdVal = St->getValue().getNode();
22361 LoadSDNode *Ld = nullptr;
22362 int TokenFactorIndex = -1;
22363 SmallVector<SDValue, 8> Ops;
22364 SDNode* ChainVal = St->getChain().getNode();
22365 // Must be a store of a load. We currently handle two cases: the load
22366 // is a direct child, and it's under an intervening TokenFactor. It is
22367 // possible to dig deeper under nested TokenFactors.
22368 if (ChainVal == LdVal)
22369 Ld = cast<LoadSDNode>(St->getChain());
22370 else if (St->getValue().hasOneUse() &&
22371 ChainVal->getOpcode() == ISD::TokenFactor) {
22372 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
22373 if (ChainVal->getOperand(i).getNode() == LdVal) {
22374 TokenFactorIndex = i;
22375 Ld = cast<LoadSDNode>(St->getValue());
22377 Ops.push_back(ChainVal->getOperand(i));
22381 if (!Ld || !ISD::isNormalLoad(Ld))
22384 // If this is not the MMX case, i.e. we are just turning i64 load/store
22385 // into f64 load/store, avoid the transformation if there are multiple
22386 // uses of the loaded value.
22387 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
22392 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
22393 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
22395 if (Subtarget->is64Bit() || F64IsLegal) {
22396 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
22397 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
22398 Ld->getPointerInfo(), Ld->isVolatile(),
22399 Ld->isNonTemporal(), Ld->isInvariant(),
22400 Ld->getAlignment());
22401 SDValue NewChain = NewLd.getValue(1);
22402 if (TokenFactorIndex != -1) {
22403 Ops.push_back(NewChain);
22404 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22406 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
22407 St->getPointerInfo(),
22408 St->isVolatile(), St->isNonTemporal(),
22409 St->getAlignment());
22412 // Otherwise, lower to two pairs of 32-bit loads / stores.
22413 SDValue LoAddr = Ld->getBasePtr();
22414 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
22415 DAG.getConstant(4, MVT::i32));
22417 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
22418 Ld->getPointerInfo(),
22419 Ld->isVolatile(), Ld->isNonTemporal(),
22420 Ld->isInvariant(), Ld->getAlignment());
22421 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
22422 Ld->getPointerInfo().getWithOffset(4),
22423 Ld->isVolatile(), Ld->isNonTemporal(),
22425 MinAlign(Ld->getAlignment(), 4));
22427 SDValue NewChain = LoLd.getValue(1);
22428 if (TokenFactorIndex != -1) {
22429 Ops.push_back(LoLd);
22430 Ops.push_back(HiLd);
22431 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22434 LoAddr = St->getBasePtr();
22435 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
22436 DAG.getConstant(4, MVT::i32));
22438 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
22439 St->getPointerInfo(),
22440 St->isVolatile(), St->isNonTemporal(),
22441 St->getAlignment());
22442 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
22443 St->getPointerInfo().getWithOffset(4),
22445 St->isNonTemporal(),
22446 MinAlign(St->getAlignment(), 4));
22447 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
22452 /// Return 'true' if this vector operation is "horizontal"
22453 /// and return the operands for the horizontal operation in LHS and RHS. A
22454 /// horizontal operation performs the binary operation on successive elements
22455 /// of its first operand, then on successive elements of its second operand,
22456 /// returning the resulting values in a vector. For example, if
22457 /// A = < float a0, float a1, float a2, float a3 >
22459 /// B = < float b0, float b1, float b2, float b3 >
22460 /// then the result of doing a horizontal operation on A and B is
22461 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
22462 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
22463 /// A horizontal-op B, for some already available A and B, and if so then LHS is
22464 /// set to A, RHS to B, and the routine returns 'true'.
22465 /// Note that the binary operation should have the property that if one of the
22466 /// operands is UNDEF then the result is UNDEF.
22467 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
22468 // Look for the following pattern: if
22469 // A = < float a0, float a1, float a2, float a3 >
22470 // B = < float b0, float b1, float b2, float b3 >
22472 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
22473 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
22474 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
22475 // which is A horizontal-op B.
22477 // At least one of the operands should be a vector shuffle.
22478 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
22479 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
22482 MVT VT = LHS.getSimpleValueType();
22484 assert((VT.is128BitVector() || VT.is256BitVector()) &&
22485 "Unsupported vector type for horizontal add/sub");
22487 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
22488 // operate independently on 128-bit lanes.
22489 unsigned NumElts = VT.getVectorNumElements();
22490 unsigned NumLanes = VT.getSizeInBits()/128;
22491 unsigned NumLaneElts = NumElts / NumLanes;
22492 assert((NumLaneElts % 2 == 0) &&
22493 "Vector type should have an even number of elements in each lane");
22494 unsigned HalfLaneElts = NumLaneElts/2;
22496 // View LHS in the form
22497 // LHS = VECTOR_SHUFFLE A, B, LMask
22498 // If LHS is not a shuffle then pretend it is the shuffle
22499 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
22500 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
22503 SmallVector<int, 16> LMask(NumElts);
22504 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22505 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
22506 A = LHS.getOperand(0);
22507 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
22508 B = LHS.getOperand(1);
22509 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
22510 std::copy(Mask.begin(), Mask.end(), LMask.begin());
22512 if (LHS.getOpcode() != ISD::UNDEF)
22514 for (unsigned i = 0; i != NumElts; ++i)
22518 // Likewise, view RHS in the form
22519 // RHS = VECTOR_SHUFFLE C, D, RMask
22521 SmallVector<int, 16> RMask(NumElts);
22522 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22523 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
22524 C = RHS.getOperand(0);
22525 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
22526 D = RHS.getOperand(1);
22527 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
22528 std::copy(Mask.begin(), Mask.end(), RMask.begin());
22530 if (RHS.getOpcode() != ISD::UNDEF)
22532 for (unsigned i = 0; i != NumElts; ++i)
22536 // Check that the shuffles are both shuffling the same vectors.
22537 if (!(A == C && B == D) && !(A == D && B == C))
22540 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
22541 if (!A.getNode() && !B.getNode())
22544 // If A and B occur in reverse order in RHS, then "swap" them (which means
22545 // rewriting the mask).
22547 CommuteVectorShuffleMask(RMask, NumElts);
22549 // At this point LHS and RHS are equivalent to
22550 // LHS = VECTOR_SHUFFLE A, B, LMask
22551 // RHS = VECTOR_SHUFFLE A, B, RMask
22552 // Check that the masks correspond to performing a horizontal operation.
22553 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
22554 for (unsigned i = 0; i != NumLaneElts; ++i) {
22555 int LIdx = LMask[i+l], RIdx = RMask[i+l];
22557 // Ignore any UNDEF components.
22558 if (LIdx < 0 || RIdx < 0 ||
22559 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
22560 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
22563 // Check that successive elements are being operated on. If not, this is
22564 // not a horizontal operation.
22565 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
22566 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
22567 if (!(LIdx == Index && RIdx == Index + 1) &&
22568 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
22573 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
22574 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
22578 /// Do target-specific dag combines on floating point adds.
22579 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
22580 const X86Subtarget *Subtarget) {
22581 EVT VT = N->getValueType(0);
22582 SDValue LHS = N->getOperand(0);
22583 SDValue RHS = N->getOperand(1);
22585 // Try to synthesize horizontal adds from adds of shuffles.
22586 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22587 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22588 isHorizontalBinOp(LHS, RHS, true))
22589 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
22593 /// Do target-specific dag combines on floating point subs.
22594 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
22595 const X86Subtarget *Subtarget) {
22596 EVT VT = N->getValueType(0);
22597 SDValue LHS = N->getOperand(0);
22598 SDValue RHS = N->getOperand(1);
22600 // Try to synthesize horizontal subs from subs of shuffles.
22601 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22602 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22603 isHorizontalBinOp(LHS, RHS, false))
22604 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
22608 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
22609 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
22610 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
22612 // F[X]OR(0.0, x) -> x
22613 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22614 if (C->getValueAPF().isPosZero())
22615 return N->getOperand(1);
22617 // F[X]OR(x, 0.0) -> x
22618 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22619 if (C->getValueAPF().isPosZero())
22620 return N->getOperand(0);
22624 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
22625 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
22626 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
22628 // Only perform optimizations if UnsafeMath is used.
22629 if (!DAG.getTarget().Options.UnsafeFPMath)
22632 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
22633 // into FMINC and FMAXC, which are Commutative operations.
22634 unsigned NewOp = 0;
22635 switch (N->getOpcode()) {
22636 default: llvm_unreachable("unknown opcode");
22637 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
22638 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
22641 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
22642 N->getOperand(0), N->getOperand(1));
22645 /// Do target-specific dag combines on X86ISD::FAND nodes.
22646 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
22647 // FAND(0.0, x) -> 0.0
22648 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22649 if (C->getValueAPF().isPosZero())
22650 return N->getOperand(0);
22652 // FAND(x, 0.0) -> 0.0
22653 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22654 if (C->getValueAPF().isPosZero())
22655 return N->getOperand(1);
22660 /// Do target-specific dag combines on X86ISD::FANDN nodes
22661 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
22662 // FANDN(0.0, x) -> x
22663 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22664 if (C->getValueAPF().isPosZero())
22665 return N->getOperand(1);
22667 // FANDN(x, 0.0) -> 0.0
22668 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22669 if (C->getValueAPF().isPosZero())
22670 return N->getOperand(1);
22675 static SDValue PerformBTCombine(SDNode *N,
22677 TargetLowering::DAGCombinerInfo &DCI) {
22678 // BT ignores high bits in the bit index operand.
22679 SDValue Op1 = N->getOperand(1);
22680 if (Op1.hasOneUse()) {
22681 unsigned BitWidth = Op1.getValueSizeInBits();
22682 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
22683 APInt KnownZero, KnownOne;
22684 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
22685 !DCI.isBeforeLegalizeOps());
22686 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22687 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
22688 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
22689 DCI.CommitTargetLoweringOpt(TLO);
22694 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
22695 SDValue Op = N->getOperand(0);
22696 if (Op.getOpcode() == ISD::BITCAST)
22697 Op = Op.getOperand(0);
22698 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
22699 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
22700 VT.getVectorElementType().getSizeInBits() ==
22701 OpVT.getVectorElementType().getSizeInBits()) {
22702 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
22707 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
22708 const X86Subtarget *Subtarget) {
22709 EVT VT = N->getValueType(0);
22710 if (!VT.isVector())
22713 SDValue N0 = N->getOperand(0);
22714 SDValue N1 = N->getOperand(1);
22715 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
22718 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
22719 // both SSE and AVX2 since there is no sign-extended shift right
22720 // operation on a vector with 64-bit elements.
22721 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
22722 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
22723 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
22724 N0.getOpcode() == ISD::SIGN_EXTEND)) {
22725 SDValue N00 = N0.getOperand(0);
22727 // EXTLOAD has a better solution on AVX2,
22728 // it may be replaced with X86ISD::VSEXT node.
22729 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
22730 if (!ISD::isNormalLoad(N00.getNode()))
22733 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
22734 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
22736 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
22742 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
22743 TargetLowering::DAGCombinerInfo &DCI,
22744 const X86Subtarget *Subtarget) {
22745 SDValue N0 = N->getOperand(0);
22746 EVT VT = N->getValueType(0);
22748 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
22749 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
22750 // This exposes the sext to the sdivrem lowering, so that it directly extends
22751 // from AH (which we otherwise need to do contortions to access).
22752 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
22753 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
22755 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22756 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
22757 N0.getOperand(0), N0.getOperand(1));
22758 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22759 return R.getValue(1);
22762 if (!DCI.isBeforeLegalizeOps())
22765 if (!Subtarget->hasFp256())
22768 if (VT.isVector() && VT.getSizeInBits() == 256) {
22769 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22777 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
22778 const X86Subtarget* Subtarget) {
22780 EVT VT = N->getValueType(0);
22782 // Let legalize expand this if it isn't a legal type yet.
22783 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
22786 EVT ScalarVT = VT.getScalarType();
22787 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
22788 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
22791 SDValue A = N->getOperand(0);
22792 SDValue B = N->getOperand(1);
22793 SDValue C = N->getOperand(2);
22795 bool NegA = (A.getOpcode() == ISD::FNEG);
22796 bool NegB = (B.getOpcode() == ISD::FNEG);
22797 bool NegC = (C.getOpcode() == ISD::FNEG);
22799 // Negative multiplication when NegA xor NegB
22800 bool NegMul = (NegA != NegB);
22802 A = A.getOperand(0);
22804 B = B.getOperand(0);
22806 C = C.getOperand(0);
22810 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
22812 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
22814 return DAG.getNode(Opcode, dl, VT, A, B, C);
22817 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
22818 TargetLowering::DAGCombinerInfo &DCI,
22819 const X86Subtarget *Subtarget) {
22820 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
22821 // (and (i32 x86isd::setcc_carry), 1)
22822 // This eliminates the zext. This transformation is necessary because
22823 // ISD::SETCC is always legalized to i8.
22825 SDValue N0 = N->getOperand(0);
22826 EVT VT = N->getValueType(0);
22828 if (N0.getOpcode() == ISD::AND &&
22830 N0.getOperand(0).hasOneUse()) {
22831 SDValue N00 = N0.getOperand(0);
22832 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22833 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22834 if (!C || C->getZExtValue() != 1)
22836 return DAG.getNode(ISD::AND, dl, VT,
22837 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22838 N00.getOperand(0), N00.getOperand(1)),
22839 DAG.getConstant(1, VT));
22843 if (N0.getOpcode() == ISD::TRUNCATE &&
22845 N0.getOperand(0).hasOneUse()) {
22846 SDValue N00 = N0.getOperand(0);
22847 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22848 return DAG.getNode(ISD::AND, dl, VT,
22849 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22850 N00.getOperand(0), N00.getOperand(1)),
22851 DAG.getConstant(1, VT));
22854 if (VT.is256BitVector()) {
22855 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22860 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
22861 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
22862 // This exposes the zext to the udivrem lowering, so that it directly extends
22863 // from AH (which we otherwise need to do contortions to access).
22864 if (N0.getOpcode() == ISD::UDIVREM &&
22865 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
22866 (VT == MVT::i32 || VT == MVT::i64)) {
22867 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22868 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
22869 N0.getOperand(0), N0.getOperand(1));
22870 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22871 return R.getValue(1);
22877 // Optimize x == -y --> x+y == 0
22878 // x != -y --> x+y != 0
22879 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
22880 const X86Subtarget* Subtarget) {
22881 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
22882 SDValue LHS = N->getOperand(0);
22883 SDValue RHS = N->getOperand(1);
22884 EVT VT = N->getValueType(0);
22887 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
22888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
22889 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
22890 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22891 LHS.getValueType(), RHS, LHS.getOperand(1));
22892 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22893 addV, DAG.getConstant(0, addV.getValueType()), CC);
22895 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
22896 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
22897 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
22898 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22899 RHS.getValueType(), LHS, RHS.getOperand(1));
22900 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22901 addV, DAG.getConstant(0, addV.getValueType()), CC);
22904 if (VT.getScalarType() == MVT::i1) {
22905 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
22906 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22907 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
22908 if (!IsSEXT0 && !IsVZero0)
22910 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
22911 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22912 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
22914 if (!IsSEXT1 && !IsVZero1)
22917 if (IsSEXT0 && IsVZero1) {
22918 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
22919 if (CC == ISD::SETEQ)
22920 return DAG.getNOT(DL, LHS.getOperand(0), VT);
22921 return LHS.getOperand(0);
22923 if (IsSEXT1 && IsVZero0) {
22924 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
22925 if (CC == ISD::SETEQ)
22926 return DAG.getNOT(DL, RHS.getOperand(0), VT);
22927 return RHS.getOperand(0);
22934 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
22935 SelectionDAG &DAG) {
22937 MVT VT = Load->getSimpleValueType(0);
22938 MVT EVT = VT.getVectorElementType();
22939 SDValue Addr = Load->getOperand(1);
22940 SDValue NewAddr = DAG.getNode(
22941 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
22942 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
22945 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
22946 DAG.getMachineFunction().getMachineMemOperand(
22947 Load->getMemOperand(), 0, EVT.getStoreSize()));
22951 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
22952 const X86Subtarget *Subtarget) {
22954 MVT VT = N->getOperand(1)->getSimpleValueType(0);
22955 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
22956 "X86insertps is only defined for v4x32");
22958 SDValue Ld = N->getOperand(1);
22959 if (MayFoldLoad(Ld)) {
22960 // Extract the countS bits from the immediate so we can get the proper
22961 // address when narrowing the vector load to a specific element.
22962 // When the second source op is a memory address, interps doesn't use
22963 // countS and just gets an f32 from that address.
22964 unsigned DestIndex =
22965 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
22966 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
22970 // Create this as a scalar to vector to match the instruction pattern.
22971 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
22972 // countS bits are ignored when loading from memory on insertps, which
22973 // means we don't need to explicitly set them to 0.
22974 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
22975 LoadScalarToVector, N->getOperand(2));
22978 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
22979 SDValue V0 = N->getOperand(0);
22980 SDValue V1 = N->getOperand(1);
22982 EVT VT = N->getValueType(0);
22984 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
22985 // operands and changing the mask to 1. This saves us a bunch of
22986 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
22987 // x86InstrInfo knows how to commute this back after instruction selection
22988 // if it would help register allocation.
22990 // TODO: If optimizing for size or a processor that doesn't suffer from
22991 // partial register update stalls, this should be transformed into a MOVSD
22992 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
22994 if (VT == MVT::v2f64)
22995 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
22996 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
22997 SDValue NewMask = DAG.getConstant(1, MVT::i8);
22998 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
23004 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
23005 // as "sbb reg,reg", since it can be extended without zext and produces
23006 // an all-ones bit which is more useful than 0/1 in some cases.
23007 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
23010 return DAG.getNode(ISD::AND, DL, VT,
23011 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23012 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
23013 DAG.getConstant(1, VT));
23014 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
23015 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
23016 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
23017 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
23020 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
23021 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
23022 TargetLowering::DAGCombinerInfo &DCI,
23023 const X86Subtarget *Subtarget) {
23025 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
23026 SDValue EFLAGS = N->getOperand(1);
23028 if (CC == X86::COND_A) {
23029 // Try to convert COND_A into COND_B in an attempt to facilitate
23030 // materializing "setb reg".
23032 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
23033 // cannot take an immediate as its first operand.
23035 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
23036 EFLAGS.getValueType().isInteger() &&
23037 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
23038 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23039 EFLAGS.getNode()->getVTList(),
23040 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23041 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23042 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
23046 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
23047 // a zext and produces an all-ones bit which is more useful than 0/1 in some
23049 if (CC == X86::COND_B)
23050 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23054 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23055 if (Flags.getNode()) {
23056 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23057 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23063 // Optimize branch condition evaluation.
23065 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23066 TargetLowering::DAGCombinerInfo &DCI,
23067 const X86Subtarget *Subtarget) {
23069 SDValue Chain = N->getOperand(0);
23070 SDValue Dest = N->getOperand(1);
23071 SDValue EFLAGS = N->getOperand(3);
23072 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23076 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23077 if (Flags.getNode()) {
23078 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23079 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23086 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23087 SelectionDAG &DAG) {
23088 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23089 // optimize away operation when it's from a constant.
23091 // The general transformation is:
23092 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23093 // AND(VECTOR_CMP(x,y), constant2)
23094 // constant2 = UNARYOP(constant)
23096 // Early exit if this isn't a vector operation, the operand of the
23097 // unary operation isn't a bitwise AND, or if the sizes of the operations
23098 // aren't the same.
23099 EVT VT = N->getValueType(0);
23100 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23101 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23102 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23105 // Now check that the other operand of the AND is a constant. We could
23106 // make the transformation for non-constant splats as well, but it's unclear
23107 // that would be a benefit as it would not eliminate any operations, just
23108 // perform one more step in scalar code before moving to the vector unit.
23109 if (BuildVectorSDNode *BV =
23110 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23111 // Bail out if the vector isn't a constant.
23112 if (!BV->isConstant())
23115 // Everything checks out. Build up the new and improved node.
23117 EVT IntVT = BV->getValueType(0);
23118 // Create a new constant of the appropriate type for the transformed
23120 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23121 // The AND node needs bitcasts to/from an integer vector type around it.
23122 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
23123 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
23124 N->getOperand(0)->getOperand(0), MaskConst);
23125 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
23132 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
23133 const X86Subtarget *Subtarget) {
23134 // First try to optimize away the conversion entirely when it's
23135 // conditionally from a constant. Vectors only.
23136 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
23137 if (Res != SDValue())
23140 // Now move on to more general possibilities.
23141 SDValue Op0 = N->getOperand(0);
23142 EVT InVT = Op0->getValueType(0);
23144 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
23145 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
23147 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
23148 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
23149 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
23152 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
23153 // a 32-bit target where SSE doesn't support i64->FP operations.
23154 if (Op0.getOpcode() == ISD::LOAD) {
23155 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
23156 EVT VT = Ld->getValueType(0);
23157 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
23158 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
23159 !Subtarget->is64Bit() && VT == MVT::i64) {
23160 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
23161 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
23162 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
23169 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
23170 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
23171 X86TargetLowering::DAGCombinerInfo &DCI) {
23172 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
23173 // the result is either zero or one (depending on the input carry bit).
23174 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
23175 if (X86::isZeroNode(N->getOperand(0)) &&
23176 X86::isZeroNode(N->getOperand(1)) &&
23177 // We don't have a good way to replace an EFLAGS use, so only do this when
23179 SDValue(N, 1).use_empty()) {
23181 EVT VT = N->getValueType(0);
23182 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
23183 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
23184 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
23185 DAG.getConstant(X86::COND_B,MVT::i8),
23187 DAG.getConstant(1, VT));
23188 return DCI.CombineTo(N, Res1, CarryOut);
23194 // fold (add Y, (sete X, 0)) -> adc 0, Y
23195 // (add Y, (setne X, 0)) -> sbb -1, Y
23196 // (sub (sete X, 0), Y) -> sbb 0, Y
23197 // (sub (setne X, 0), Y) -> adc -1, Y
23198 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
23201 // Look through ZExts.
23202 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
23203 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
23206 SDValue SetCC = Ext.getOperand(0);
23207 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
23210 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
23211 if (CC != X86::COND_E && CC != X86::COND_NE)
23214 SDValue Cmp = SetCC.getOperand(1);
23215 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
23216 !X86::isZeroNode(Cmp.getOperand(1)) ||
23217 !Cmp.getOperand(0).getValueType().isInteger())
23220 SDValue CmpOp0 = Cmp.getOperand(0);
23221 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
23222 DAG.getConstant(1, CmpOp0.getValueType()));
23224 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
23225 if (CC == X86::COND_NE)
23226 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
23227 DL, OtherVal.getValueType(), OtherVal,
23228 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
23229 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
23230 DL, OtherVal.getValueType(), OtherVal,
23231 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
23234 /// PerformADDCombine - Do target-specific dag combines on integer adds.
23235 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
23236 const X86Subtarget *Subtarget) {
23237 EVT VT = N->getValueType(0);
23238 SDValue Op0 = N->getOperand(0);
23239 SDValue Op1 = N->getOperand(1);
23241 // Try to synthesize horizontal adds from adds of shuffles.
23242 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23243 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23244 isHorizontalBinOp(Op0, Op1, true))
23245 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
23247 return OptimizeConditionalInDecrement(N, DAG);
23250 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
23251 const X86Subtarget *Subtarget) {
23252 SDValue Op0 = N->getOperand(0);
23253 SDValue Op1 = N->getOperand(1);
23255 // X86 can't encode an immediate LHS of a sub. See if we can push the
23256 // negation into a preceding instruction.
23257 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
23258 // If the RHS of the sub is a XOR with one use and a constant, invert the
23259 // immediate. Then add one to the LHS of the sub so we can turn
23260 // X-Y -> X+~Y+1, saving one register.
23261 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
23262 isa<ConstantSDNode>(Op1.getOperand(1))) {
23263 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
23264 EVT VT = Op0.getValueType();
23265 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
23267 DAG.getConstant(~XorC, VT));
23268 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
23269 DAG.getConstant(C->getAPIntValue()+1, VT));
23273 // Try to synthesize horizontal adds from adds of shuffles.
23274 EVT VT = N->getValueType(0);
23275 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23276 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23277 isHorizontalBinOp(Op0, Op1, true))
23278 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
23280 return OptimizeConditionalInDecrement(N, DAG);
23283 /// performVZEXTCombine - Performs build vector combines
23284 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
23285 TargetLowering::DAGCombinerInfo &DCI,
23286 const X86Subtarget *Subtarget) {
23288 MVT VT = N->getSimpleValueType(0);
23289 SDValue Op = N->getOperand(0);
23290 MVT OpVT = Op.getSimpleValueType();
23291 MVT OpEltVT = OpVT.getVectorElementType();
23292 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
23294 // (vzext (bitcast (vzext (x)) -> (vzext x)
23296 while (V.getOpcode() == ISD::BITCAST)
23297 V = V.getOperand(0);
23299 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
23300 MVT InnerVT = V.getSimpleValueType();
23301 MVT InnerEltVT = InnerVT.getVectorElementType();
23303 // If the element sizes match exactly, we can just do one larger vzext. This
23304 // is always an exact type match as vzext operates on integer types.
23305 if (OpEltVT == InnerEltVT) {
23306 assert(OpVT == InnerVT && "Types must match for vzext!");
23307 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
23310 // The only other way we can combine them is if only a single element of the
23311 // inner vzext is used in the input to the outer vzext.
23312 if (InnerEltVT.getSizeInBits() < InputBits)
23315 // In this case, the inner vzext is completely dead because we're going to
23316 // only look at bits inside of the low element. Just do the outer vzext on
23317 // a bitcast of the input to the inner.
23318 return DAG.getNode(X86ISD::VZEXT, DL, VT,
23319 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
23322 // Check if we can bypass extracting and re-inserting an element of an input
23323 // vector. Essentialy:
23324 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
23325 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
23326 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
23327 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
23328 SDValue ExtractedV = V.getOperand(0);
23329 SDValue OrigV = ExtractedV.getOperand(0);
23330 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
23331 if (ExtractIdx->getZExtValue() == 0) {
23332 MVT OrigVT = OrigV.getSimpleValueType();
23333 // Extract a subvector if necessary...
23334 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
23335 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
23336 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
23337 OrigVT.getVectorNumElements() / Ratio);
23338 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
23339 DAG.getIntPtrConstant(0));
23341 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
23342 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
23349 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
23350 DAGCombinerInfo &DCI) const {
23351 SelectionDAG &DAG = DCI.DAG;
23352 switch (N->getOpcode()) {
23354 case ISD::EXTRACT_VECTOR_ELT:
23355 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
23358 case X86ISD::SHRUNKBLEND:
23359 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
23360 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
23361 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
23362 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
23363 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
23364 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
23365 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
23368 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
23369 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
23370 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
23371 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
23372 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
23373 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
23374 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
23375 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
23376 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
23377 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
23378 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
23380 case X86ISD::FOR: return PerformFORCombine(N, DAG);
23382 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
23383 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
23384 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
23385 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
23386 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
23387 case ISD::ANY_EXTEND:
23388 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
23389 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
23390 case ISD::SIGN_EXTEND_INREG:
23391 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
23392 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
23393 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
23394 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
23395 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
23396 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
23397 case X86ISD::SHUFP: // Handle all target specific shuffles
23398 case X86ISD::PALIGNR:
23399 case X86ISD::UNPCKH:
23400 case X86ISD::UNPCKL:
23401 case X86ISD::MOVHLPS:
23402 case X86ISD::MOVLHPS:
23403 case X86ISD::PSHUFB:
23404 case X86ISD::PSHUFD:
23405 case X86ISD::PSHUFHW:
23406 case X86ISD::PSHUFLW:
23407 case X86ISD::MOVSS:
23408 case X86ISD::MOVSD:
23409 case X86ISD::VPERMILPI:
23410 case X86ISD::VPERM2X128:
23411 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
23412 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
23413 case ISD::INTRINSIC_WO_CHAIN:
23414 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
23415 case X86ISD::INSERTPS: {
23416 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
23417 return PerformINSERTPSCombine(N, DAG, Subtarget);
23420 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
23421 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
23427 /// isTypeDesirableForOp - Return true if the target has native support for
23428 /// the specified value type and it is 'desirable' to use the type for the
23429 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
23430 /// instruction encodings are longer and some i16 instructions are slow.
23431 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
23432 if (!isTypeLegal(VT))
23434 if (VT != MVT::i16)
23441 case ISD::SIGN_EXTEND:
23442 case ISD::ZERO_EXTEND:
23443 case ISD::ANY_EXTEND:
23456 /// IsDesirableToPromoteOp - This method query the target whether it is
23457 /// beneficial for dag combiner to promote the specified node. If true, it
23458 /// should return the desired promotion type by reference.
23459 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
23460 EVT VT = Op.getValueType();
23461 if (VT != MVT::i16)
23464 bool Promote = false;
23465 bool Commute = false;
23466 switch (Op.getOpcode()) {
23469 LoadSDNode *LD = cast<LoadSDNode>(Op);
23470 // If the non-extending load has a single use and it's not live out, then it
23471 // might be folded.
23472 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
23473 Op.hasOneUse()*/) {
23474 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
23475 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
23476 // The only case where we'd want to promote LOAD (rather then it being
23477 // promoted as an operand is when it's only use is liveout.
23478 if (UI->getOpcode() != ISD::CopyToReg)
23485 case ISD::SIGN_EXTEND:
23486 case ISD::ZERO_EXTEND:
23487 case ISD::ANY_EXTEND:
23492 SDValue N0 = Op.getOperand(0);
23493 // Look out for (store (shl (load), x)).
23494 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
23507 SDValue N0 = Op.getOperand(0);
23508 SDValue N1 = Op.getOperand(1);
23509 if (!Commute && MayFoldLoad(N1))
23511 // Avoid disabling potential load folding opportunities.
23512 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
23514 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
23524 //===----------------------------------------------------------------------===//
23525 // X86 Inline Assembly Support
23526 //===----------------------------------------------------------------------===//
23529 // Helper to match a string separated by whitespace.
23530 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
23531 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
23533 for (unsigned i = 0, e = args.size(); i != e; ++i) {
23534 StringRef piece(*args[i]);
23535 if (!s.startswith(piece)) // Check if the piece matches.
23538 s = s.substr(piece.size());
23539 StringRef::size_type pos = s.find_first_not_of(" \t");
23540 if (pos == 0) // We matched a prefix.
23548 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
23551 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
23553 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
23554 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
23555 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
23556 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
23558 if (AsmPieces.size() == 3)
23560 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
23567 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
23568 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
23570 std::string AsmStr = IA->getAsmString();
23572 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
23573 if (!Ty || Ty->getBitWidth() % 16 != 0)
23576 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
23577 SmallVector<StringRef, 4> AsmPieces;
23578 SplitString(AsmStr, AsmPieces, ";\n");
23580 switch (AsmPieces.size()) {
23581 default: return false;
23583 // FIXME: this should verify that we are targeting a 486 or better. If not,
23584 // we will turn this bswap into something that will be lowered to logical
23585 // ops instead of emitting the bswap asm. For now, we don't support 486 or
23586 // lower so don't worry about this.
23588 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
23589 matchAsm(AsmPieces[0], "bswapl", "$0") ||
23590 matchAsm(AsmPieces[0], "bswapq", "$0") ||
23591 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
23592 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
23593 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
23594 // No need to check constraints, nothing other than the equivalent of
23595 // "=r,0" would be valid here.
23596 return IntrinsicLowering::LowerToByteSwap(CI);
23599 // rorw $$8, ${0:w} --> llvm.bswap.i16
23600 if (CI->getType()->isIntegerTy(16) &&
23601 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23602 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
23603 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
23605 const std::string &ConstraintsStr = IA->getConstraintString();
23606 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23607 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23608 if (clobbersFlagRegisters(AsmPieces))
23609 return IntrinsicLowering::LowerToByteSwap(CI);
23613 if (CI->getType()->isIntegerTy(32) &&
23614 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23615 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
23616 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
23617 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
23619 const std::string &ConstraintsStr = IA->getConstraintString();
23620 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23621 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23622 if (clobbersFlagRegisters(AsmPieces))
23623 return IntrinsicLowering::LowerToByteSwap(CI);
23626 if (CI->getType()->isIntegerTy(64)) {
23627 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
23628 if (Constraints.size() >= 2 &&
23629 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
23630 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
23631 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
23632 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
23633 matchAsm(AsmPieces[1], "bswap", "%edx") &&
23634 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
23635 return IntrinsicLowering::LowerToByteSwap(CI);
23643 /// getConstraintType - Given a constraint letter, return the type of
23644 /// constraint it is for this target.
23645 X86TargetLowering::ConstraintType
23646 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
23647 if (Constraint.size() == 1) {
23648 switch (Constraint[0]) {
23659 return C_RegisterClass;
23683 return TargetLowering::getConstraintType(Constraint);
23686 /// Examine constraint type and operand type and determine a weight value.
23687 /// This object must already have been set up with the operand type
23688 /// and the current alternative constraint selected.
23689 TargetLowering::ConstraintWeight
23690 X86TargetLowering::getSingleConstraintMatchWeight(
23691 AsmOperandInfo &info, const char *constraint) const {
23692 ConstraintWeight weight = CW_Invalid;
23693 Value *CallOperandVal = info.CallOperandVal;
23694 // If we don't have a value, we can't do a match,
23695 // but allow it at the lowest weight.
23696 if (!CallOperandVal)
23698 Type *type = CallOperandVal->getType();
23699 // Look at the constraint type.
23700 switch (*constraint) {
23702 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
23713 if (CallOperandVal->getType()->isIntegerTy())
23714 weight = CW_SpecificReg;
23719 if (type->isFloatingPointTy())
23720 weight = CW_SpecificReg;
23723 if (type->isX86_MMXTy() && Subtarget->hasMMX())
23724 weight = CW_SpecificReg;
23728 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
23729 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
23730 weight = CW_Register;
23733 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
23734 if (C->getZExtValue() <= 31)
23735 weight = CW_Constant;
23739 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23740 if (C->getZExtValue() <= 63)
23741 weight = CW_Constant;
23745 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23746 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
23747 weight = CW_Constant;
23751 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23752 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
23753 weight = CW_Constant;
23757 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23758 if (C->getZExtValue() <= 3)
23759 weight = CW_Constant;
23763 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23764 if (C->getZExtValue() <= 0xff)
23765 weight = CW_Constant;
23770 if (dyn_cast<ConstantFP>(CallOperandVal)) {
23771 weight = CW_Constant;
23775 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23776 if ((C->getSExtValue() >= -0x80000000LL) &&
23777 (C->getSExtValue() <= 0x7fffffffLL))
23778 weight = CW_Constant;
23782 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23783 if (C->getZExtValue() <= 0xffffffff)
23784 weight = CW_Constant;
23791 /// LowerXConstraint - try to replace an X constraint, which matches anything,
23792 /// with another that has more specific requirements based on the type of the
23793 /// corresponding operand.
23794 const char *X86TargetLowering::
23795 LowerXConstraint(EVT ConstraintVT) const {
23796 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
23797 // 'f' like normal targets.
23798 if (ConstraintVT.isFloatingPoint()) {
23799 if (Subtarget->hasSSE2())
23801 if (Subtarget->hasSSE1())
23805 return TargetLowering::LowerXConstraint(ConstraintVT);
23808 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
23809 /// vector. If it is invalid, don't add anything to Ops.
23810 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
23811 std::string &Constraint,
23812 std::vector<SDValue>&Ops,
23813 SelectionDAG &DAG) const {
23816 // Only support length 1 constraints for now.
23817 if (Constraint.length() > 1) return;
23819 char ConstraintLetter = Constraint[0];
23820 switch (ConstraintLetter) {
23823 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23824 if (C->getZExtValue() <= 31) {
23825 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23831 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23832 if (C->getZExtValue() <= 63) {
23833 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23839 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23840 if (isInt<8>(C->getSExtValue())) {
23841 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23847 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23848 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
23849 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
23850 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
23856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23857 if (C->getZExtValue() <= 3) {
23858 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23864 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23865 if (C->getZExtValue() <= 255) {
23866 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23872 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23873 if (C->getZExtValue() <= 127) {
23874 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23880 // 32-bit signed value
23881 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23882 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23883 C->getSExtValue())) {
23884 // Widen to 64 bits here to get it sign extended.
23885 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
23888 // FIXME gcc accepts some relocatable values here too, but only in certain
23889 // memory models; it's complicated.
23894 // 32-bit unsigned value
23895 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23896 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23897 C->getZExtValue())) {
23898 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23902 // FIXME gcc accepts some relocatable values here too, but only in certain
23903 // memory models; it's complicated.
23907 // Literal immediates are always ok.
23908 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
23909 // Widen to 64 bits here to get it sign extended.
23910 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
23914 // In any sort of PIC mode addresses need to be computed at runtime by
23915 // adding in a register or some sort of table lookup. These can't
23916 // be used as immediates.
23917 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
23920 // If we are in non-pic codegen mode, we allow the address of a global (with
23921 // an optional displacement) to be used with 'i'.
23922 GlobalAddressSDNode *GA = nullptr;
23923 int64_t Offset = 0;
23925 // Match either (GA), (GA+C), (GA+C1+C2), etc.
23927 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
23928 Offset += GA->getOffset();
23930 } else if (Op.getOpcode() == ISD::ADD) {
23931 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23932 Offset += C->getZExtValue();
23933 Op = Op.getOperand(0);
23936 } else if (Op.getOpcode() == ISD::SUB) {
23937 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23938 Offset += -C->getZExtValue();
23939 Op = Op.getOperand(0);
23944 // Otherwise, this isn't something we can handle, reject it.
23948 const GlobalValue *GV = GA->getGlobal();
23949 // If we require an extra load to get this address, as in PIC mode, we
23950 // can't accept it.
23951 if (isGlobalStubReference(
23952 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
23955 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
23956 GA->getValueType(0), Offset);
23961 if (Result.getNode()) {
23962 Ops.push_back(Result);
23965 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
23968 std::pair<unsigned, const TargetRegisterClass*>
23969 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
23971 // First, see if this is a constraint that directly corresponds to an LLVM
23973 if (Constraint.size() == 1) {
23974 // GCC Constraint Letters
23975 switch (Constraint[0]) {
23977 // TODO: Slight differences here in allocation order and leaving
23978 // RIP in the class. Do they matter any more here than they do
23979 // in the normal allocation?
23980 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
23981 if (Subtarget->is64Bit()) {
23982 if (VT == MVT::i32 || VT == MVT::f32)
23983 return std::make_pair(0U, &X86::GR32RegClass);
23984 if (VT == MVT::i16)
23985 return std::make_pair(0U, &X86::GR16RegClass);
23986 if (VT == MVT::i8 || VT == MVT::i1)
23987 return std::make_pair(0U, &X86::GR8RegClass);
23988 if (VT == MVT::i64 || VT == MVT::f64)
23989 return std::make_pair(0U, &X86::GR64RegClass);
23992 // 32-bit fallthrough
23993 case 'Q': // Q_REGS
23994 if (VT == MVT::i32 || VT == MVT::f32)
23995 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
23996 if (VT == MVT::i16)
23997 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
23998 if (VT == MVT::i8 || VT == MVT::i1)
23999 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
24000 if (VT == MVT::i64)
24001 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
24003 case 'r': // GENERAL_REGS
24004 case 'l': // INDEX_REGS
24005 if (VT == MVT::i8 || VT == MVT::i1)
24006 return std::make_pair(0U, &X86::GR8RegClass);
24007 if (VT == MVT::i16)
24008 return std::make_pair(0U, &X86::GR16RegClass);
24009 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
24010 return std::make_pair(0U, &X86::GR32RegClass);
24011 return std::make_pair(0U, &X86::GR64RegClass);
24012 case 'R': // LEGACY_REGS
24013 if (VT == MVT::i8 || VT == MVT::i1)
24014 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
24015 if (VT == MVT::i16)
24016 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
24017 if (VT == MVT::i32 || !Subtarget->is64Bit())
24018 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
24019 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
24020 case 'f': // FP Stack registers.
24021 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
24022 // value to the correct fpstack register class.
24023 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
24024 return std::make_pair(0U, &X86::RFP32RegClass);
24025 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
24026 return std::make_pair(0U, &X86::RFP64RegClass);
24027 return std::make_pair(0U, &X86::RFP80RegClass);
24028 case 'y': // MMX_REGS if MMX allowed.
24029 if (!Subtarget->hasMMX()) break;
24030 return std::make_pair(0U, &X86::VR64RegClass);
24031 case 'Y': // SSE_REGS if SSE2 allowed
24032 if (!Subtarget->hasSSE2()) break;
24034 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
24035 if (!Subtarget->hasSSE1()) break;
24037 switch (VT.SimpleTy) {
24039 // Scalar SSE types.
24042 return std::make_pair(0U, &X86::FR32RegClass);
24045 return std::make_pair(0U, &X86::FR64RegClass);
24053 return std::make_pair(0U, &X86::VR128RegClass);
24061 return std::make_pair(0U, &X86::VR256RegClass);
24066 return std::make_pair(0U, &X86::VR512RegClass);
24072 // Use the default implementation in TargetLowering to convert the register
24073 // constraint into a member of a register class.
24074 std::pair<unsigned, const TargetRegisterClass*> Res;
24075 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
24077 // Not found as a standard register?
24079 // Map st(0) -> st(7) -> ST0
24080 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24081 tolower(Constraint[1]) == 's' &&
24082 tolower(Constraint[2]) == 't' &&
24083 Constraint[3] == '(' &&
24084 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24085 Constraint[5] == ')' &&
24086 Constraint[6] == '}') {
24088 Res.first = X86::FP0+Constraint[4]-'0';
24089 Res.second = &X86::RFP80RegClass;
24093 // GCC allows "st(0)" to be called just plain "st".
24094 if (StringRef("{st}").equals_lower(Constraint)) {
24095 Res.first = X86::FP0;
24096 Res.second = &X86::RFP80RegClass;
24101 if (StringRef("{flags}").equals_lower(Constraint)) {
24102 Res.first = X86::EFLAGS;
24103 Res.second = &X86::CCRRegClass;
24107 // 'A' means EAX + EDX.
24108 if (Constraint == "A") {
24109 Res.first = X86::EAX;
24110 Res.second = &X86::GR32_ADRegClass;
24116 // Otherwise, check to see if this is a register class of the wrong value
24117 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24118 // turn into {ax},{dx}.
24119 if (Res.second->hasType(VT))
24120 return Res; // Correct type already, nothing to do.
24122 // All of the single-register GCC register classes map their values onto
24123 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24124 // really want an 8-bit or 32-bit register, map to the appropriate register
24125 // class and return the appropriate register.
24126 if (Res.second == &X86::GR16RegClass) {
24127 if (VT == MVT::i8 || VT == MVT::i1) {
24128 unsigned DestReg = 0;
24129 switch (Res.first) {
24131 case X86::AX: DestReg = X86::AL; break;
24132 case X86::DX: DestReg = X86::DL; break;
24133 case X86::CX: DestReg = X86::CL; break;
24134 case X86::BX: DestReg = X86::BL; break;
24137 Res.first = DestReg;
24138 Res.second = &X86::GR8RegClass;
24140 } else if (VT == MVT::i32 || VT == MVT::f32) {
24141 unsigned DestReg = 0;
24142 switch (Res.first) {
24144 case X86::AX: DestReg = X86::EAX; break;
24145 case X86::DX: DestReg = X86::EDX; break;
24146 case X86::CX: DestReg = X86::ECX; break;
24147 case X86::BX: DestReg = X86::EBX; break;
24148 case X86::SI: DestReg = X86::ESI; break;
24149 case X86::DI: DestReg = X86::EDI; break;
24150 case X86::BP: DestReg = X86::EBP; break;
24151 case X86::SP: DestReg = X86::ESP; break;
24154 Res.first = DestReg;
24155 Res.second = &X86::GR32RegClass;
24157 } else if (VT == MVT::i64 || VT == MVT::f64) {
24158 unsigned DestReg = 0;
24159 switch (Res.first) {
24161 case X86::AX: DestReg = X86::RAX; break;
24162 case X86::DX: DestReg = X86::RDX; break;
24163 case X86::CX: DestReg = X86::RCX; break;
24164 case X86::BX: DestReg = X86::RBX; break;
24165 case X86::SI: DestReg = X86::RSI; break;
24166 case X86::DI: DestReg = X86::RDI; break;
24167 case X86::BP: DestReg = X86::RBP; break;
24168 case X86::SP: DestReg = X86::RSP; break;
24171 Res.first = DestReg;
24172 Res.second = &X86::GR64RegClass;
24175 } else if (Res.second == &X86::FR32RegClass ||
24176 Res.second == &X86::FR64RegClass ||
24177 Res.second == &X86::VR128RegClass ||
24178 Res.second == &X86::VR256RegClass ||
24179 Res.second == &X86::FR32XRegClass ||
24180 Res.second == &X86::FR64XRegClass ||
24181 Res.second == &X86::VR128XRegClass ||
24182 Res.second == &X86::VR256XRegClass ||
24183 Res.second == &X86::VR512RegClass) {
24184 // Handle references to XMM physical registers that got mapped into the
24185 // wrong class. This can happen with constraints like {xmm0} where the
24186 // target independent register mapper will just pick the first match it can
24187 // find, ignoring the required type.
24189 if (VT == MVT::f32 || VT == MVT::i32)
24190 Res.second = &X86::FR32RegClass;
24191 else if (VT == MVT::f64 || VT == MVT::i64)
24192 Res.second = &X86::FR64RegClass;
24193 else if (X86::VR128RegClass.hasType(VT))
24194 Res.second = &X86::VR128RegClass;
24195 else if (X86::VR256RegClass.hasType(VT))
24196 Res.second = &X86::VR256RegClass;
24197 else if (X86::VR512RegClass.hasType(VT))
24198 Res.second = &X86::VR512RegClass;
24204 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
24206 // Scaling factors are not free at all.
24207 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
24208 // will take 2 allocations in the out of order engine instead of 1
24209 // for plain addressing mode, i.e. inst (reg1).
24211 // vaddps (%rsi,%drx), %ymm0, %ymm1
24212 // Requires two allocations (one for the load, one for the computation)
24214 // vaddps (%rsi), %ymm0, %ymm1
24215 // Requires just 1 allocation, i.e., freeing allocations for other operations
24216 // and having less micro operations to execute.
24218 // For some X86 architectures, this is even worse because for instance for
24219 // stores, the complex addressing mode forces the instruction to use the
24220 // "load" ports instead of the dedicated "store" port.
24221 // E.g., on Haswell:
24222 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
24223 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
24224 if (isLegalAddressingMode(AM, Ty))
24225 // Scale represents reg2 * scale, thus account for 1
24226 // as soon as we use a second register.
24227 return AM.Scale != 0;
24231 bool X86TargetLowering::isTargetFTOL() const {
24232 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();