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 // We want to custom lower some of our intrinsics.
1613 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1614 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1615 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1616 if (!Subtarget->is64Bit())
1617 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1619 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1620 // handle type legalization for these operations here.
1622 // FIXME: We really should do custom legalization for addition and
1623 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1624 // than generic legalization for 64-bit multiplication-with-overflow, though.
1625 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1626 // Add/Sub/Mul with overflow operations are custom lowered.
1628 setOperationAction(ISD::SADDO, VT, Custom);
1629 setOperationAction(ISD::UADDO, VT, Custom);
1630 setOperationAction(ISD::SSUBO, VT, Custom);
1631 setOperationAction(ISD::USUBO, VT, Custom);
1632 setOperationAction(ISD::SMULO, VT, Custom);
1633 setOperationAction(ISD::UMULO, VT, Custom);
1637 if (!Subtarget->is64Bit()) {
1638 // These libcalls are not available in 32-bit.
1639 setLibcallName(RTLIB::SHL_I128, nullptr);
1640 setLibcallName(RTLIB::SRL_I128, nullptr);
1641 setLibcallName(RTLIB::SRA_I128, nullptr);
1644 // Combine sin / cos into one node or libcall if possible.
1645 if (Subtarget->hasSinCos()) {
1646 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1647 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1648 if (Subtarget->isTargetDarwin()) {
1649 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1650 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1651 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1652 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1656 if (Subtarget->isTargetWin64()) {
1657 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1658 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1659 setOperationAction(ISD::SREM, MVT::i128, Custom);
1660 setOperationAction(ISD::UREM, MVT::i128, Custom);
1661 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1662 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1665 // We have target-specific dag combine patterns for the following nodes:
1666 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1667 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1668 setTargetDAGCombine(ISD::BITCAST);
1669 setTargetDAGCombine(ISD::VSELECT);
1670 setTargetDAGCombine(ISD::SELECT);
1671 setTargetDAGCombine(ISD::SHL);
1672 setTargetDAGCombine(ISD::SRA);
1673 setTargetDAGCombine(ISD::SRL);
1674 setTargetDAGCombine(ISD::OR);
1675 setTargetDAGCombine(ISD::AND);
1676 setTargetDAGCombine(ISD::ADD);
1677 setTargetDAGCombine(ISD::FADD);
1678 setTargetDAGCombine(ISD::FSUB);
1679 setTargetDAGCombine(ISD::FMA);
1680 setTargetDAGCombine(ISD::SUB);
1681 setTargetDAGCombine(ISD::LOAD);
1682 setTargetDAGCombine(ISD::MLOAD);
1683 setTargetDAGCombine(ISD::STORE);
1684 setTargetDAGCombine(ISD::MSTORE);
1685 setTargetDAGCombine(ISD::ZERO_EXTEND);
1686 setTargetDAGCombine(ISD::ANY_EXTEND);
1687 setTargetDAGCombine(ISD::SIGN_EXTEND);
1688 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1689 setTargetDAGCombine(ISD::TRUNCATE);
1690 setTargetDAGCombine(ISD::SINT_TO_FP);
1691 setTargetDAGCombine(ISD::SETCC);
1692 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1693 setTargetDAGCombine(ISD::BUILD_VECTOR);
1694 setTargetDAGCombine(ISD::MUL);
1695 setTargetDAGCombine(ISD::XOR);
1697 computeRegisterProperties(Subtarget->getRegisterInfo());
1699 // On Darwin, -Os means optimize for size without hurting performance,
1700 // do not reduce the limit.
1701 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1702 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1703 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1704 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1705 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1706 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1707 setPrefLoopAlignment(4); // 2^4 bytes.
1709 // Predictable cmov don't hurt on atom because it's in-order.
1710 PredictableSelectIsExpensive = !Subtarget->isAtom();
1711 EnableExtLdPromotion = true;
1712 setPrefFunctionAlignment(4); // 2^4 bytes.
1714 verifyIntrinsicTables();
1717 // This has so far only been implemented for 64-bit MachO.
1718 bool X86TargetLowering::useLoadStackGuardNode() const {
1719 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1722 TargetLoweringBase::LegalizeTypeAction
1723 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1724 if (ExperimentalVectorWideningLegalization &&
1725 VT.getVectorNumElements() != 1 &&
1726 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1727 return TypeWidenVector;
1729 return TargetLoweringBase::getPreferredVectorAction(VT);
1732 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1734 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1736 const unsigned NumElts = VT.getVectorNumElements();
1737 const EVT EltVT = VT.getVectorElementType();
1738 if (VT.is512BitVector()) {
1739 if (Subtarget->hasAVX512())
1740 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1741 EltVT == MVT::f32 || EltVT == MVT::f64)
1743 case 8: return MVT::v8i1;
1744 case 16: return MVT::v16i1;
1746 if (Subtarget->hasBWI())
1747 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1749 case 32: return MVT::v32i1;
1750 case 64: return MVT::v64i1;
1754 if (VT.is256BitVector() || VT.is128BitVector()) {
1755 if (Subtarget->hasVLX())
1756 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1757 EltVT == MVT::f32 || EltVT == MVT::f64)
1759 case 2: return MVT::v2i1;
1760 case 4: return MVT::v4i1;
1761 case 8: return MVT::v8i1;
1763 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1764 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1766 case 8: return MVT::v8i1;
1767 case 16: return MVT::v16i1;
1768 case 32: return MVT::v32i1;
1772 return VT.changeVectorElementTypeToInteger();
1775 /// Helper for getByValTypeAlignment to determine
1776 /// the desired ByVal argument alignment.
1777 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1780 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1781 if (VTy->getBitWidth() == 128)
1783 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1784 unsigned EltAlign = 0;
1785 getMaxByValAlign(ATy->getElementType(), EltAlign);
1786 if (EltAlign > MaxAlign)
1787 MaxAlign = EltAlign;
1788 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1789 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1790 unsigned EltAlign = 0;
1791 getMaxByValAlign(STy->getElementType(i), EltAlign);
1792 if (EltAlign > MaxAlign)
1793 MaxAlign = EltAlign;
1800 /// Return the desired alignment for ByVal aggregate
1801 /// function arguments in the caller parameter area. For X86, aggregates
1802 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1803 /// are at 4-byte boundaries.
1804 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1805 if (Subtarget->is64Bit()) {
1806 // Max of 8 and alignment of type.
1807 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1814 if (Subtarget->hasSSE1())
1815 getMaxByValAlign(Ty, Align);
1819 /// Returns the target specific optimal type for load
1820 /// and store operations as a result of memset, memcpy, and memmove
1821 /// lowering. If DstAlign is zero that means it's safe to destination
1822 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1823 /// means there isn't a need to check it against alignment requirement,
1824 /// probably because the source does not need to be loaded. If 'IsMemset' is
1825 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1826 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1827 /// source is constant so it does not need to be loaded.
1828 /// It returns EVT::Other if the type should be determined using generic
1829 /// target-independent logic.
1831 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1832 unsigned DstAlign, unsigned SrcAlign,
1833 bool IsMemset, bool ZeroMemset,
1835 MachineFunction &MF) const {
1836 const Function *F = MF.getFunction();
1837 if ((!IsMemset || ZeroMemset) &&
1838 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1840 (Subtarget->isUnalignedMemAccessFast() ||
1841 ((DstAlign == 0 || DstAlign >= 16) &&
1842 (SrcAlign == 0 || SrcAlign >= 16)))) {
1844 if (Subtarget->hasInt256())
1846 if (Subtarget->hasFp256())
1849 if (Subtarget->hasSSE2())
1851 if (Subtarget->hasSSE1())
1853 } else if (!MemcpyStrSrc && Size >= 8 &&
1854 !Subtarget->is64Bit() &&
1855 Subtarget->hasSSE2()) {
1856 // Do not use f64 to lower memcpy if source is string constant. It's
1857 // better to use i32 to avoid the loads.
1861 if (Subtarget->is64Bit() && Size >= 8)
1866 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1868 return X86ScalarSSEf32;
1869 else if (VT == MVT::f64)
1870 return X86ScalarSSEf64;
1875 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1880 *Fast = Subtarget->isUnalignedMemAccessFast();
1884 /// Return the entry encoding for a jump table in the
1885 /// current function. The returned value is a member of the
1886 /// MachineJumpTableInfo::JTEntryKind enum.
1887 unsigned X86TargetLowering::getJumpTableEncoding() const {
1888 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1890 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1891 Subtarget->isPICStyleGOT())
1892 return MachineJumpTableInfo::EK_Custom32;
1894 // Otherwise, use the normal jump table encoding heuristics.
1895 return TargetLowering::getJumpTableEncoding();
1899 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1900 const MachineBasicBlock *MBB,
1901 unsigned uid,MCContext &Ctx) const{
1902 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1903 Subtarget->isPICStyleGOT());
1904 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1906 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1907 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1910 /// Returns relocation base for the given PIC jumptable.
1911 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1912 SelectionDAG &DAG) const {
1913 if (!Subtarget->is64Bit())
1914 // This doesn't have SDLoc associated with it, but is not really the
1915 // same as a Register.
1916 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1920 /// This returns the relocation base for the given PIC jumptable,
1921 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1922 const MCExpr *X86TargetLowering::
1923 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1924 MCContext &Ctx) const {
1925 // X86-64 uses RIP relative addressing based on the jump table label.
1926 if (Subtarget->isPICStyleRIPRel())
1927 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1929 // Otherwise, the reference is relative to the PIC base.
1930 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1933 std::pair<const TargetRegisterClass *, uint8_t>
1934 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1936 const TargetRegisterClass *RRC = nullptr;
1938 switch (VT.SimpleTy) {
1940 return TargetLowering::findRepresentativeClass(TRI, VT);
1941 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1942 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1945 RRC = &X86::VR64RegClass;
1947 case MVT::f32: case MVT::f64:
1948 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1949 case MVT::v4f32: case MVT::v2f64:
1950 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1952 RRC = &X86::VR128RegClass;
1955 return std::make_pair(RRC, Cost);
1958 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1959 unsigned &Offset) const {
1960 if (!Subtarget->isTargetLinux())
1963 if (Subtarget->is64Bit()) {
1964 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1966 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1978 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1979 unsigned DestAS) const {
1980 assert(SrcAS != DestAS && "Expected different address spaces!");
1982 return SrcAS < 256 && DestAS < 256;
1985 //===----------------------------------------------------------------------===//
1986 // Return Value Calling Convention Implementation
1987 //===----------------------------------------------------------------------===//
1989 #include "X86GenCallingConv.inc"
1992 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1993 MachineFunction &MF, bool isVarArg,
1994 const SmallVectorImpl<ISD::OutputArg> &Outs,
1995 LLVMContext &Context) const {
1996 SmallVector<CCValAssign, 16> RVLocs;
1997 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1998 return CCInfo.CheckReturn(Outs, RetCC_X86);
2001 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2002 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2007 X86TargetLowering::LowerReturn(SDValue Chain,
2008 CallingConv::ID CallConv, bool isVarArg,
2009 const SmallVectorImpl<ISD::OutputArg> &Outs,
2010 const SmallVectorImpl<SDValue> &OutVals,
2011 SDLoc dl, SelectionDAG &DAG) const {
2012 MachineFunction &MF = DAG.getMachineFunction();
2013 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2015 SmallVector<CCValAssign, 16> RVLocs;
2016 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2017 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2020 SmallVector<SDValue, 6> RetOps;
2021 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2022 // Operand #1 = Bytes To Pop
2023 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
2026 // Copy the result values into the output registers.
2027 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2028 CCValAssign &VA = RVLocs[i];
2029 assert(VA.isRegLoc() && "Can only return in registers!");
2030 SDValue ValToCopy = OutVals[i];
2031 EVT ValVT = ValToCopy.getValueType();
2033 // Promote values to the appropriate types.
2034 if (VA.getLocInfo() == CCValAssign::SExt)
2035 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2036 else if (VA.getLocInfo() == CCValAssign::ZExt)
2037 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2038 else if (VA.getLocInfo() == CCValAssign::AExt)
2039 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2040 else if (VA.getLocInfo() == CCValAssign::BCvt)
2041 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2043 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2044 "Unexpected FP-extend for return value.");
2046 // If this is x86-64, and we disabled SSE, we can't return FP values,
2047 // or SSE or MMX vectors.
2048 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2049 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2050 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2051 report_fatal_error("SSE register return with SSE disabled");
2053 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2054 // llvm-gcc has never done it right and no one has noticed, so this
2055 // should be OK for now.
2056 if (ValVT == MVT::f64 &&
2057 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2058 report_fatal_error("SSE2 register return with SSE2 disabled");
2060 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2061 // the RET instruction and handled by the FP Stackifier.
2062 if (VA.getLocReg() == X86::FP0 ||
2063 VA.getLocReg() == X86::FP1) {
2064 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2065 // change the value to the FP stack register class.
2066 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2067 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2068 RetOps.push_back(ValToCopy);
2069 // Don't emit a copytoreg.
2073 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2074 // which is returned in RAX / RDX.
2075 if (Subtarget->is64Bit()) {
2076 if (ValVT == MVT::x86mmx) {
2077 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2078 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2079 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2081 // If we don't have SSE2 available, convert to v4f32 so the generated
2082 // register is legal.
2083 if (!Subtarget->hasSSE2())
2084 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2089 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2090 Flag = Chain.getValue(1);
2091 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2094 // The x86-64 ABIs require that for returning structs by value we copy
2095 // the sret argument into %rax/%eax (depending on ABI) for the return.
2096 // Win32 requires us to put the sret argument to %eax as well.
2097 // We saved the argument into a virtual register in the entry block,
2098 // so now we copy the value out and into %rax/%eax.
2100 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2101 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2102 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2103 // either case FuncInfo->setSRetReturnReg() will have been called.
2104 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2105 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
2106 "No need for an sret register");
2107 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2110 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2111 X86::RAX : X86::EAX;
2112 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2113 Flag = Chain.getValue(1);
2115 // RAX/EAX now acts like a return value.
2116 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2119 RetOps[0] = Chain; // Update chain.
2121 // Add the flag if we have it.
2123 RetOps.push_back(Flag);
2125 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2128 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2129 if (N->getNumValues() != 1)
2131 if (!N->hasNUsesOfValue(1, 0))
2134 SDValue TCChain = Chain;
2135 SDNode *Copy = *N->use_begin();
2136 if (Copy->getOpcode() == ISD::CopyToReg) {
2137 // If the copy has a glue operand, we conservatively assume it isn't safe to
2138 // perform a tail call.
2139 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2141 TCChain = Copy->getOperand(0);
2142 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2145 bool HasRet = false;
2146 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2148 if (UI->getOpcode() != X86ISD::RET_FLAG)
2150 // If we are returning more than one value, we can definitely
2151 // not make a tail call see PR19530
2152 if (UI->getNumOperands() > 4)
2154 if (UI->getNumOperands() == 4 &&
2155 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2168 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2169 ISD::NodeType ExtendKind) const {
2171 // TODO: Is this also valid on 32-bit?
2172 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2173 ReturnMVT = MVT::i8;
2175 ReturnMVT = MVT::i32;
2177 EVT MinVT = getRegisterType(Context, ReturnMVT);
2178 return VT.bitsLT(MinVT) ? MinVT : VT;
2181 /// Lower the result values of a call into the
2182 /// appropriate copies out of appropriate physical registers.
2185 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2186 CallingConv::ID CallConv, bool isVarArg,
2187 const SmallVectorImpl<ISD::InputArg> &Ins,
2188 SDLoc dl, SelectionDAG &DAG,
2189 SmallVectorImpl<SDValue> &InVals) const {
2191 // Assign locations to each value returned by this call.
2192 SmallVector<CCValAssign, 16> RVLocs;
2193 bool Is64Bit = Subtarget->is64Bit();
2194 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2196 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2198 // Copy all of the result registers out of their specified physreg.
2199 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2200 CCValAssign &VA = RVLocs[i];
2201 EVT CopyVT = VA.getValVT();
2203 // If this is x86-64, and we disabled SSE, we can't return FP values
2204 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2205 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2206 report_fatal_error("SSE register return with SSE disabled");
2209 // If we prefer to use the value in xmm registers, copy it out as f80 and
2210 // use a truncate to move it from fp stack reg to xmm reg.
2211 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2212 isScalarFPTypeInSSEReg(VA.getValVT()))
2215 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2216 CopyVT, InFlag).getValue(1);
2217 SDValue Val = Chain.getValue(0);
2219 if (CopyVT != VA.getValVT())
2220 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2221 // This truncation won't change the value.
2222 DAG.getIntPtrConstant(1));
2224 InFlag = Chain.getValue(2);
2225 InVals.push_back(Val);
2231 //===----------------------------------------------------------------------===//
2232 // C & StdCall & Fast Calling Convention implementation
2233 //===----------------------------------------------------------------------===//
2234 // StdCall calling convention seems to be standard for many Windows' API
2235 // routines and around. It differs from C calling convention just a little:
2236 // callee should clean up the stack, not caller. Symbols should be also
2237 // decorated in some fancy way :) It doesn't support any vector arguments.
2238 // For info on fast calling convention see Fast Calling Convention (tail call)
2239 // implementation LowerX86_32FastCCCallTo.
2241 /// CallIsStructReturn - Determines whether a call uses struct return
2243 enum StructReturnType {
2248 static StructReturnType
2249 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2251 return NotStructReturn;
2253 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2254 if (!Flags.isSRet())
2255 return NotStructReturn;
2256 if (Flags.isInReg())
2257 return RegStructReturn;
2258 return StackStructReturn;
2261 /// Determines whether a function uses struct return semantics.
2262 static StructReturnType
2263 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2265 return NotStructReturn;
2267 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2268 if (!Flags.isSRet())
2269 return NotStructReturn;
2270 if (Flags.isInReg())
2271 return RegStructReturn;
2272 return StackStructReturn;
2275 /// Make a copy of an aggregate at address specified by "Src" to address
2276 /// "Dst" with size and alignment information specified by the specific
2277 /// parameter attribute. The copy will be passed as a byval function parameter.
2279 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2280 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2282 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2284 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2285 /*isVolatile*/false, /*AlwaysInline=*/true,
2286 MachinePointerInfo(), MachinePointerInfo());
2289 /// Return true if the calling convention is one that
2290 /// supports tail call optimization.
2291 static bool IsTailCallConvention(CallingConv::ID CC) {
2292 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2293 CC == CallingConv::HiPE);
2296 /// \brief Return true if the calling convention is a C calling convention.
2297 static bool IsCCallConvention(CallingConv::ID CC) {
2298 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2299 CC == CallingConv::X86_64_SysV);
2302 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2303 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2307 CallingConv::ID CalleeCC = CS.getCallingConv();
2308 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2314 /// Return true if the function is being made into
2315 /// a tailcall target by changing its ABI.
2316 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2317 bool GuaranteedTailCallOpt) {
2318 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2322 X86TargetLowering::LowerMemArgument(SDValue Chain,
2323 CallingConv::ID CallConv,
2324 const SmallVectorImpl<ISD::InputArg> &Ins,
2325 SDLoc dl, SelectionDAG &DAG,
2326 const CCValAssign &VA,
2327 MachineFrameInfo *MFI,
2329 // Create the nodes corresponding to a load from this parameter slot.
2330 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2331 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2332 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2333 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2336 // If value is passed by pointer we have address passed instead of the value
2338 if (VA.getLocInfo() == CCValAssign::Indirect)
2339 ValVT = VA.getLocVT();
2341 ValVT = VA.getValVT();
2343 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2344 // changed with more analysis.
2345 // In case of tail call optimization mark all arguments mutable. Since they
2346 // could be overwritten by lowering of arguments in case of a tail call.
2347 if (Flags.isByVal()) {
2348 unsigned Bytes = Flags.getByValSize();
2349 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2350 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2351 return DAG.getFrameIndex(FI, getPointerTy());
2353 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2354 VA.getLocMemOffset(), isImmutable);
2355 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2356 return DAG.getLoad(ValVT, dl, Chain, FIN,
2357 MachinePointerInfo::getFixedStack(FI),
2358 false, false, false, 0);
2362 // FIXME: Get this from tablegen.
2363 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2364 const X86Subtarget *Subtarget) {
2365 assert(Subtarget->is64Bit());
2367 if (Subtarget->isCallingConvWin64(CallConv)) {
2368 static const MCPhysReg GPR64ArgRegsWin64[] = {
2369 X86::RCX, X86::RDX, X86::R8, X86::R9
2371 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2374 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2375 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2377 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2380 // FIXME: Get this from tablegen.
2381 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2382 CallingConv::ID CallConv,
2383 const X86Subtarget *Subtarget) {
2384 assert(Subtarget->is64Bit());
2385 if (Subtarget->isCallingConvWin64(CallConv)) {
2386 // The XMM registers which might contain var arg parameters are shadowed
2387 // in their paired GPR. So we only need to save the GPR to their home
2389 // TODO: __vectorcall will change this.
2393 const Function *Fn = MF.getFunction();
2394 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2395 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2396 "SSE register cannot be used when SSE is disabled!");
2397 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2398 !Subtarget->hasSSE1())
2399 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2403 static const MCPhysReg XMMArgRegs64Bit[] = {
2404 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2405 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2407 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2411 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2412 CallingConv::ID CallConv,
2414 const SmallVectorImpl<ISD::InputArg> &Ins,
2417 SmallVectorImpl<SDValue> &InVals)
2419 MachineFunction &MF = DAG.getMachineFunction();
2420 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2422 const Function* Fn = MF.getFunction();
2423 if (Fn->hasExternalLinkage() &&
2424 Subtarget->isTargetCygMing() &&
2425 Fn->getName() == "main")
2426 FuncInfo->setForceFramePointer(true);
2428 MachineFrameInfo *MFI = MF.getFrameInfo();
2429 bool Is64Bit = Subtarget->is64Bit();
2430 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2432 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2433 "Var args not supported with calling convention fastcc, ghc or hipe");
2435 // Assign locations to all of the incoming arguments.
2436 SmallVector<CCValAssign, 16> ArgLocs;
2437 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2439 // Allocate shadow area for Win64
2441 CCInfo.AllocateStack(32, 8);
2443 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2445 unsigned LastVal = ~0U;
2447 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2448 CCValAssign &VA = ArgLocs[i];
2449 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2451 assert(VA.getValNo() != LastVal &&
2452 "Don't support value assigned to multiple locs yet");
2454 LastVal = VA.getValNo();
2456 if (VA.isRegLoc()) {
2457 EVT RegVT = VA.getLocVT();
2458 const TargetRegisterClass *RC;
2459 if (RegVT == MVT::i32)
2460 RC = &X86::GR32RegClass;
2461 else if (Is64Bit && RegVT == MVT::i64)
2462 RC = &X86::GR64RegClass;
2463 else if (RegVT == MVT::f32)
2464 RC = &X86::FR32RegClass;
2465 else if (RegVT == MVT::f64)
2466 RC = &X86::FR64RegClass;
2467 else if (RegVT.is512BitVector())
2468 RC = &X86::VR512RegClass;
2469 else if (RegVT.is256BitVector())
2470 RC = &X86::VR256RegClass;
2471 else if (RegVT.is128BitVector())
2472 RC = &X86::VR128RegClass;
2473 else if (RegVT == MVT::x86mmx)
2474 RC = &X86::VR64RegClass;
2475 else if (RegVT == MVT::i1)
2476 RC = &X86::VK1RegClass;
2477 else if (RegVT == MVT::v8i1)
2478 RC = &X86::VK8RegClass;
2479 else if (RegVT == MVT::v16i1)
2480 RC = &X86::VK16RegClass;
2481 else if (RegVT == MVT::v32i1)
2482 RC = &X86::VK32RegClass;
2483 else if (RegVT == MVT::v64i1)
2484 RC = &X86::VK64RegClass;
2486 llvm_unreachable("Unknown argument type!");
2488 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2489 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2491 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2492 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2494 if (VA.getLocInfo() == CCValAssign::SExt)
2495 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2496 DAG.getValueType(VA.getValVT()));
2497 else if (VA.getLocInfo() == CCValAssign::ZExt)
2498 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2499 DAG.getValueType(VA.getValVT()));
2500 else if (VA.getLocInfo() == CCValAssign::BCvt)
2501 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2503 if (VA.isExtInLoc()) {
2504 // Handle MMX values passed in XMM regs.
2505 if (RegVT.isVector())
2506 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2508 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2511 assert(VA.isMemLoc());
2512 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2515 // If value is passed via pointer - do a load.
2516 if (VA.getLocInfo() == CCValAssign::Indirect)
2517 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2518 MachinePointerInfo(), false, false, false, 0);
2520 InVals.push_back(ArgValue);
2523 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2524 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2525 // The x86-64 ABIs require that for returning structs by value we copy
2526 // the sret argument into %rax/%eax (depending on ABI) for the return.
2527 // Win32 requires us to put the sret argument to %eax as well.
2528 // Save the argument into a virtual register so that we can access it
2529 // from the return points.
2530 if (Ins[i].Flags.isSRet()) {
2531 unsigned Reg = FuncInfo->getSRetReturnReg();
2533 MVT PtrTy = getPointerTy();
2534 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2535 FuncInfo->setSRetReturnReg(Reg);
2537 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2538 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2544 unsigned StackSize = CCInfo.getNextStackOffset();
2545 // Align stack specially for tail calls.
2546 if (FuncIsMadeTailCallSafe(CallConv,
2547 MF.getTarget().Options.GuaranteedTailCallOpt))
2548 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2550 // If the function takes variable number of arguments, make a frame index for
2551 // the start of the first vararg value... for expansion of llvm.va_start. We
2552 // can skip this if there are no va_start calls.
2553 if (MFI->hasVAStart() &&
2554 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2555 CallConv != CallingConv::X86_ThisCall))) {
2556 FuncInfo->setVarArgsFrameIndex(
2557 MFI->CreateFixedObject(1, StackSize, true));
2560 // Figure out if XMM registers are in use.
2561 assert(!(MF.getTarget().Options.UseSoftFloat &&
2562 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2563 "SSE register cannot be used when SSE is disabled!");
2565 // 64-bit calling conventions support varargs and register parameters, so we
2566 // have to do extra work to spill them in the prologue.
2567 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2568 // Find the first unallocated argument registers.
2569 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2570 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2571 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2572 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2573 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2574 "SSE register cannot be used when SSE is disabled!");
2576 // Gather all the live in physical registers.
2577 SmallVector<SDValue, 6> LiveGPRs;
2578 SmallVector<SDValue, 8> LiveXMMRegs;
2580 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2581 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2583 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2585 if (!ArgXMMs.empty()) {
2586 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2587 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2588 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2589 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2590 LiveXMMRegs.push_back(
2591 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2596 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2597 // Get to the caller-allocated home save location. Add 8 to account
2598 // for the return address.
2599 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2600 FuncInfo->setRegSaveFrameIndex(
2601 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2602 // Fixup to set vararg frame on shadow area (4 x i64).
2604 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2606 // For X86-64, if there are vararg parameters that are passed via
2607 // registers, then we must store them to their spots on the stack so
2608 // they may be loaded by deferencing the result of va_next.
2609 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2610 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2611 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2612 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2615 // Store the integer parameter registers.
2616 SmallVector<SDValue, 8> MemOps;
2617 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2619 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2620 for (SDValue Val : LiveGPRs) {
2621 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2622 DAG.getIntPtrConstant(Offset));
2624 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2625 MachinePointerInfo::getFixedStack(
2626 FuncInfo->getRegSaveFrameIndex(), Offset),
2628 MemOps.push_back(Store);
2632 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2633 // Now store the XMM (fp + vector) parameter registers.
2634 SmallVector<SDValue, 12> SaveXMMOps;
2635 SaveXMMOps.push_back(Chain);
2636 SaveXMMOps.push_back(ALVal);
2637 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2638 FuncInfo->getRegSaveFrameIndex()));
2639 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2640 FuncInfo->getVarArgsFPOffset()));
2641 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2643 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2644 MVT::Other, SaveXMMOps));
2647 if (!MemOps.empty())
2648 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2651 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2652 // Find the largest legal vector type.
2653 MVT VecVT = MVT::Other;
2654 // FIXME: Only some x86_32 calling conventions support AVX512.
2655 if (Subtarget->hasAVX512() &&
2656 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2657 CallConv == CallingConv::Intel_OCL_BI)))
2658 VecVT = MVT::v16f32;
2659 else if (Subtarget->hasAVX())
2661 else if (Subtarget->hasSSE2())
2664 // We forward some GPRs and some vector types.
2665 SmallVector<MVT, 2> RegParmTypes;
2666 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2667 RegParmTypes.push_back(IntVT);
2668 if (VecVT != MVT::Other)
2669 RegParmTypes.push_back(VecVT);
2671 // Compute the set of forwarded registers. The rest are scratch.
2672 SmallVectorImpl<ForwardedRegister> &Forwards =
2673 FuncInfo->getForwardedMustTailRegParms();
2674 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2676 // Conservatively forward AL on x86_64, since it might be used for varargs.
2677 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2678 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2679 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2682 // Copy all forwards from physical to virtual registers.
2683 for (ForwardedRegister &F : Forwards) {
2684 // FIXME: Can we use a less constrained schedule?
2685 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2686 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2687 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2691 // Some CCs need callee pop.
2692 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2693 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2694 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2696 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2697 // If this is an sret function, the return should pop the hidden pointer.
2698 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2699 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2700 argsAreStructReturn(Ins) == StackStructReturn)
2701 FuncInfo->setBytesToPopOnReturn(4);
2705 // RegSaveFrameIndex is X86-64 only.
2706 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2707 if (CallConv == CallingConv::X86_FastCall ||
2708 CallConv == CallingConv::X86_ThisCall)
2709 // fastcc functions can't have varargs.
2710 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2713 FuncInfo->setArgumentStackSize(StackSize);
2719 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2720 SDValue StackPtr, SDValue Arg,
2721 SDLoc dl, SelectionDAG &DAG,
2722 const CCValAssign &VA,
2723 ISD::ArgFlagsTy Flags) const {
2724 unsigned LocMemOffset = VA.getLocMemOffset();
2725 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2726 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2727 if (Flags.isByVal())
2728 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2730 return DAG.getStore(Chain, dl, Arg, PtrOff,
2731 MachinePointerInfo::getStack(LocMemOffset),
2735 /// Emit a load of return address if tail call
2736 /// optimization is performed and it is required.
2738 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2739 SDValue &OutRetAddr, SDValue Chain,
2740 bool IsTailCall, bool Is64Bit,
2741 int FPDiff, SDLoc dl) const {
2742 // Adjust the Return address stack slot.
2743 EVT VT = getPointerTy();
2744 OutRetAddr = getReturnAddressFrameIndex(DAG);
2746 // Load the "old" Return address.
2747 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2748 false, false, false, 0);
2749 return SDValue(OutRetAddr.getNode(), 1);
2752 /// Emit a store of the return address if tail call
2753 /// optimization is performed and it is required (FPDiff!=0).
2754 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2755 SDValue Chain, SDValue RetAddrFrIdx,
2756 EVT PtrVT, unsigned SlotSize,
2757 int FPDiff, SDLoc dl) {
2758 // Store the return address to the appropriate stack slot.
2759 if (!FPDiff) return Chain;
2760 // Calculate the new stack slot for the return address.
2761 int NewReturnAddrFI =
2762 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2764 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2765 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2766 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2772 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2773 SmallVectorImpl<SDValue> &InVals) const {
2774 SelectionDAG &DAG = CLI.DAG;
2776 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2777 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2778 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2779 SDValue Chain = CLI.Chain;
2780 SDValue Callee = CLI.Callee;
2781 CallingConv::ID CallConv = CLI.CallConv;
2782 bool &isTailCall = CLI.IsTailCall;
2783 bool isVarArg = CLI.IsVarArg;
2785 MachineFunction &MF = DAG.getMachineFunction();
2786 bool Is64Bit = Subtarget->is64Bit();
2787 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2788 StructReturnType SR = callIsStructReturn(Outs);
2789 bool IsSibcall = false;
2790 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2792 if (MF.getTarget().Options.DisableTailCalls)
2795 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2797 // Force this to be a tail call. The verifier rules are enough to ensure
2798 // that we can lower this successfully without moving the return address
2801 } else if (isTailCall) {
2802 // Check if it's really possible to do a tail call.
2803 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2804 isVarArg, SR != NotStructReturn,
2805 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2806 Outs, OutVals, Ins, DAG);
2808 // Sibcalls are automatically detected tailcalls which do not require
2810 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2817 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2818 "Var args not supported with calling convention fastcc, ghc or hipe");
2820 // Analyze operands of the call, assigning locations to each operand.
2821 SmallVector<CCValAssign, 16> ArgLocs;
2822 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2824 // Allocate shadow area for Win64
2826 CCInfo.AllocateStack(32, 8);
2828 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2830 // Get a count of how many bytes are to be pushed on the stack.
2831 unsigned NumBytes = CCInfo.getNextStackOffset();
2833 // This is a sibcall. The memory operands are available in caller's
2834 // own caller's stack.
2836 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2837 IsTailCallConvention(CallConv))
2838 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2841 if (isTailCall && !IsSibcall && !IsMustTail) {
2842 // Lower arguments at fp - stackoffset + fpdiff.
2843 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2845 FPDiff = NumBytesCallerPushed - NumBytes;
2847 // Set the delta of movement of the returnaddr stackslot.
2848 // But only set if delta is greater than previous delta.
2849 if (FPDiff < X86Info->getTCReturnAddrDelta())
2850 X86Info->setTCReturnAddrDelta(FPDiff);
2853 unsigned NumBytesToPush = NumBytes;
2854 unsigned NumBytesToPop = NumBytes;
2856 // If we have an inalloca argument, all stack space has already been allocated
2857 // for us and be right at the top of the stack. We don't support multiple
2858 // arguments passed in memory when using inalloca.
2859 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2861 if (!ArgLocs.back().isMemLoc())
2862 report_fatal_error("cannot use inalloca attribute on a register "
2864 if (ArgLocs.back().getLocMemOffset() != 0)
2865 report_fatal_error("any parameter with the inalloca attribute must be "
2866 "the only memory argument");
2870 Chain = DAG.getCALLSEQ_START(
2871 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2873 SDValue RetAddrFrIdx;
2874 // Load return address for tail calls.
2875 if (isTailCall && FPDiff)
2876 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2877 Is64Bit, FPDiff, dl);
2879 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2880 SmallVector<SDValue, 8> MemOpChains;
2883 // Walk the register/memloc assignments, inserting copies/loads. In the case
2884 // of tail call optimization arguments are handle later.
2885 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2886 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2887 // Skip inalloca arguments, they have already been written.
2888 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2889 if (Flags.isInAlloca())
2892 CCValAssign &VA = ArgLocs[i];
2893 EVT RegVT = VA.getLocVT();
2894 SDValue Arg = OutVals[i];
2895 bool isByVal = Flags.isByVal();
2897 // Promote the value if needed.
2898 switch (VA.getLocInfo()) {
2899 default: llvm_unreachable("Unknown loc info!");
2900 case CCValAssign::Full: break;
2901 case CCValAssign::SExt:
2902 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2904 case CCValAssign::ZExt:
2905 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2907 case CCValAssign::AExt:
2908 if (RegVT.is128BitVector()) {
2909 // Special case: passing MMX values in XMM registers.
2910 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2911 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2912 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2914 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2916 case CCValAssign::BCvt:
2917 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2919 case CCValAssign::Indirect: {
2920 // Store the argument.
2921 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2922 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2923 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2924 MachinePointerInfo::getFixedStack(FI),
2931 if (VA.isRegLoc()) {
2932 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2933 if (isVarArg && IsWin64) {
2934 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2935 // shadow reg if callee is a varargs function.
2936 unsigned ShadowReg = 0;
2937 switch (VA.getLocReg()) {
2938 case X86::XMM0: ShadowReg = X86::RCX; break;
2939 case X86::XMM1: ShadowReg = X86::RDX; break;
2940 case X86::XMM2: ShadowReg = X86::R8; break;
2941 case X86::XMM3: ShadowReg = X86::R9; break;
2944 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2946 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2947 assert(VA.isMemLoc());
2948 if (!StackPtr.getNode())
2949 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2951 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2952 dl, DAG, VA, Flags));
2956 if (!MemOpChains.empty())
2957 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2959 if (Subtarget->isPICStyleGOT()) {
2960 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2963 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2964 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2966 // If we are tail calling and generating PIC/GOT style code load the
2967 // address of the callee into ECX. The value in ecx is used as target of
2968 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2969 // for tail calls on PIC/GOT architectures. Normally we would just put the
2970 // address of GOT into ebx and then call target@PLT. But for tail calls
2971 // ebx would be restored (since ebx is callee saved) before jumping to the
2974 // Note: The actual moving to ECX is done further down.
2975 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2976 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2977 !G->getGlobal()->hasProtectedVisibility())
2978 Callee = LowerGlobalAddress(Callee, DAG);
2979 else if (isa<ExternalSymbolSDNode>(Callee))
2980 Callee = LowerExternalSymbol(Callee, DAG);
2984 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2985 // From AMD64 ABI document:
2986 // For calls that may call functions that use varargs or stdargs
2987 // (prototype-less calls or calls to functions containing ellipsis (...) in
2988 // the declaration) %al is used as hidden argument to specify the number
2989 // of SSE registers used. The contents of %al do not need to match exactly
2990 // the number of registers, but must be an ubound on the number of SSE
2991 // registers used and is in the range 0 - 8 inclusive.
2993 // Count the number of XMM registers allocated.
2994 static const MCPhysReg XMMArgRegs[] = {
2995 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2996 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2998 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2999 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3000 && "SSE registers cannot be used when SSE is disabled");
3002 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3003 DAG.getConstant(NumXMMRegs, MVT::i8)));
3006 if (isVarArg && IsMustTail) {
3007 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3008 for (const auto &F : Forwards) {
3009 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3010 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3014 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3015 // don't need this because the eligibility check rejects calls that require
3016 // shuffling arguments passed in memory.
3017 if (!IsSibcall && isTailCall) {
3018 // Force all the incoming stack arguments to be loaded from the stack
3019 // before any new outgoing arguments are stored to the stack, because the
3020 // outgoing stack slots may alias the incoming argument stack slots, and
3021 // the alias isn't otherwise explicit. This is slightly more conservative
3022 // than necessary, because it means that each store effectively depends
3023 // on every argument instead of just those arguments it would clobber.
3024 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3026 SmallVector<SDValue, 8> MemOpChains2;
3029 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3030 CCValAssign &VA = ArgLocs[i];
3033 assert(VA.isMemLoc());
3034 SDValue Arg = OutVals[i];
3035 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3036 // Skip inalloca arguments. They don't require any work.
3037 if (Flags.isInAlloca())
3039 // Create frame index.
3040 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3041 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3042 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3043 FIN = DAG.getFrameIndex(FI, getPointerTy());
3045 if (Flags.isByVal()) {
3046 // Copy relative to framepointer.
3047 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3048 if (!StackPtr.getNode())
3049 StackPtr = DAG.getCopyFromReg(Chain, dl,
3050 RegInfo->getStackRegister(),
3052 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3054 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3058 // Store relative to framepointer.
3059 MemOpChains2.push_back(
3060 DAG.getStore(ArgChain, dl, Arg, FIN,
3061 MachinePointerInfo::getFixedStack(FI),
3066 if (!MemOpChains2.empty())
3067 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3069 // Store the return address to the appropriate stack slot.
3070 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3071 getPointerTy(), RegInfo->getSlotSize(),
3075 // Build a sequence of copy-to-reg nodes chained together with token chain
3076 // and flag operands which copy the outgoing args into registers.
3078 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3079 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3080 RegsToPass[i].second, InFlag);
3081 InFlag = Chain.getValue(1);
3084 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3085 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3086 // In the 64-bit large code model, we have to make all calls
3087 // through a register, since the call instruction's 32-bit
3088 // pc-relative offset may not be large enough to hold the whole
3090 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3091 // If the callee is a GlobalAddress node (quite common, every direct call
3092 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3094 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3096 // We should use extra load for direct calls to dllimported functions in
3098 const GlobalValue *GV = G->getGlobal();
3099 if (!GV->hasDLLImportStorageClass()) {
3100 unsigned char OpFlags = 0;
3101 bool ExtraLoad = false;
3102 unsigned WrapperKind = ISD::DELETED_NODE;
3104 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3105 // external symbols most go through the PLT in PIC mode. If the symbol
3106 // has hidden or protected visibility, or if it is static or local, then
3107 // we don't need to use the PLT - we can directly call it.
3108 if (Subtarget->isTargetELF() &&
3109 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3110 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3111 OpFlags = X86II::MO_PLT;
3112 } else if (Subtarget->isPICStyleStubAny() &&
3113 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3114 (!Subtarget->getTargetTriple().isMacOSX() ||
3115 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3116 // PC-relative references to external symbols should go through $stub,
3117 // unless we're building with the leopard linker or later, which
3118 // automatically synthesizes these stubs.
3119 OpFlags = X86II::MO_DARWIN_STUB;
3120 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3121 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3122 // If the function is marked as non-lazy, generate an indirect call
3123 // which loads from the GOT directly. This avoids runtime overhead
3124 // at the cost of eager binding (and one extra byte of encoding).
3125 OpFlags = X86II::MO_GOTPCREL;
3126 WrapperKind = X86ISD::WrapperRIP;
3130 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3131 G->getOffset(), OpFlags);
3133 // Add a wrapper if needed.
3134 if (WrapperKind != ISD::DELETED_NODE)
3135 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3136 // Add extra indirection if needed.
3138 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3139 MachinePointerInfo::getGOT(),
3140 false, false, false, 0);
3142 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3143 unsigned char OpFlags = 0;
3145 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3146 // external symbols should go through the PLT.
3147 if (Subtarget->isTargetELF() &&
3148 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3149 OpFlags = X86II::MO_PLT;
3150 } else if (Subtarget->isPICStyleStubAny() &&
3151 (!Subtarget->getTargetTriple().isMacOSX() ||
3152 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3153 // PC-relative references to external symbols should go through $stub,
3154 // unless we're building with the leopard linker or later, which
3155 // automatically synthesizes these stubs.
3156 OpFlags = X86II::MO_DARWIN_STUB;
3159 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3161 } else if (Subtarget->isTarget64BitILP32() &&
3162 Callee->getValueType(0) == MVT::i32) {
3163 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3164 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3167 // Returns a chain & a flag for retval copy to use.
3168 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3169 SmallVector<SDValue, 8> Ops;
3171 if (!IsSibcall && isTailCall) {
3172 Chain = DAG.getCALLSEQ_END(Chain,
3173 DAG.getIntPtrConstant(NumBytesToPop, true),
3174 DAG.getIntPtrConstant(0, true), InFlag, dl);
3175 InFlag = Chain.getValue(1);
3178 Ops.push_back(Chain);
3179 Ops.push_back(Callee);
3182 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3184 // Add argument registers to the end of the list so that they are known live
3186 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3187 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3188 RegsToPass[i].second.getValueType()));
3190 // Add a register mask operand representing the call-preserved registers.
3191 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3192 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3193 assert(Mask && "Missing call preserved mask for calling convention");
3194 Ops.push_back(DAG.getRegisterMask(Mask));
3196 if (InFlag.getNode())
3197 Ops.push_back(InFlag);
3201 //// If this is the first return lowered for this function, add the regs
3202 //// to the liveout set for the function.
3203 // This isn't right, although it's probably harmless on x86; liveouts
3204 // should be computed from returns not tail calls. Consider a void
3205 // function making a tail call to a function returning int.
3206 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3209 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3210 InFlag = Chain.getValue(1);
3212 // Create the CALLSEQ_END node.
3213 unsigned NumBytesForCalleeToPop;
3214 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3215 DAG.getTarget().Options.GuaranteedTailCallOpt))
3216 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3217 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3218 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3219 SR == StackStructReturn)
3220 // If this is a call to a struct-return function, the callee
3221 // pops the hidden struct pointer, so we have to push it back.
3222 // This is common for Darwin/X86, Linux & Mingw32 targets.
3223 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3224 NumBytesForCalleeToPop = 4;
3226 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3228 // Returns a flag for retval copy to use.
3230 Chain = DAG.getCALLSEQ_END(Chain,
3231 DAG.getIntPtrConstant(NumBytesToPop, true),
3232 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3235 InFlag = Chain.getValue(1);
3238 // Handle result values, copying them out of physregs into vregs that we
3240 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3241 Ins, dl, DAG, InVals);
3244 //===----------------------------------------------------------------------===//
3245 // Fast Calling Convention (tail call) implementation
3246 //===----------------------------------------------------------------------===//
3248 // Like std call, callee cleans arguments, convention except that ECX is
3249 // reserved for storing the tail called function address. Only 2 registers are
3250 // free for argument passing (inreg). Tail call optimization is performed
3252 // * tailcallopt is enabled
3253 // * caller/callee are fastcc
3254 // On X86_64 architecture with GOT-style position independent code only local
3255 // (within module) calls are supported at the moment.
3256 // To keep the stack aligned according to platform abi the function
3257 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3258 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3259 // If a tail called function callee has more arguments than the caller the
3260 // caller needs to make sure that there is room to move the RETADDR to. This is
3261 // achieved by reserving an area the size of the argument delta right after the
3262 // original RETADDR, but before the saved framepointer or the spilled registers
3263 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3275 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3276 /// for a 16 byte align requirement.
3278 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3279 SelectionDAG& DAG) const {
3280 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3281 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3282 unsigned StackAlignment = TFI.getStackAlignment();
3283 uint64_t AlignMask = StackAlignment - 1;
3284 int64_t Offset = StackSize;
3285 unsigned SlotSize = RegInfo->getSlotSize();
3286 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3287 // Number smaller than 12 so just add the difference.
3288 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3290 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3291 Offset = ((~AlignMask) & Offset) + StackAlignment +
3292 (StackAlignment-SlotSize);
3297 /// MatchingStackOffset - Return true if the given stack call argument is
3298 /// already available in the same position (relatively) of the caller's
3299 /// incoming argument stack.
3301 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3302 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3303 const X86InstrInfo *TII) {
3304 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3306 if (Arg.getOpcode() == ISD::CopyFromReg) {
3307 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3308 if (!TargetRegisterInfo::isVirtualRegister(VR))
3310 MachineInstr *Def = MRI->getVRegDef(VR);
3313 if (!Flags.isByVal()) {
3314 if (!TII->isLoadFromStackSlot(Def, FI))
3317 unsigned Opcode = Def->getOpcode();
3318 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3319 Opcode == X86::LEA64_32r) &&
3320 Def->getOperand(1).isFI()) {
3321 FI = Def->getOperand(1).getIndex();
3322 Bytes = Flags.getByValSize();
3326 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3327 if (Flags.isByVal())
3328 // ByVal argument is passed in as a pointer but it's now being
3329 // dereferenced. e.g.
3330 // define @foo(%struct.X* %A) {
3331 // tail call @bar(%struct.X* byval %A)
3334 SDValue Ptr = Ld->getBasePtr();
3335 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3338 FI = FINode->getIndex();
3339 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3340 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3341 FI = FINode->getIndex();
3342 Bytes = Flags.getByValSize();
3346 assert(FI != INT_MAX);
3347 if (!MFI->isFixedObjectIndex(FI))
3349 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3352 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3353 /// for tail call optimization. Targets which want to do tail call
3354 /// optimization should implement this function.
3356 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3357 CallingConv::ID CalleeCC,
3359 bool isCalleeStructRet,
3360 bool isCallerStructRet,
3362 const SmallVectorImpl<ISD::OutputArg> &Outs,
3363 const SmallVectorImpl<SDValue> &OutVals,
3364 const SmallVectorImpl<ISD::InputArg> &Ins,
3365 SelectionDAG &DAG) const {
3366 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3369 // If -tailcallopt is specified, make fastcc functions tail-callable.
3370 const MachineFunction &MF = DAG.getMachineFunction();
3371 const Function *CallerF = MF.getFunction();
3373 // If the function return type is x86_fp80 and the callee return type is not,
3374 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3375 // perform a tailcall optimization here.
3376 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3379 CallingConv::ID CallerCC = CallerF->getCallingConv();
3380 bool CCMatch = CallerCC == CalleeCC;
3381 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3382 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3384 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3385 if (IsTailCallConvention(CalleeCC) && CCMatch)
3390 // Look for obvious safe cases to perform tail call optimization that do not
3391 // require ABI changes. This is what gcc calls sibcall.
3393 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3394 // emit a special epilogue.
3395 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3396 if (RegInfo->needsStackRealignment(MF))
3399 // Also avoid sibcall optimization if either caller or callee uses struct
3400 // return semantics.
3401 if (isCalleeStructRet || isCallerStructRet)
3404 // An stdcall/thiscall caller is expected to clean up its arguments; the
3405 // callee isn't going to do that.
3406 // FIXME: this is more restrictive than needed. We could produce a tailcall
3407 // when the stack adjustment matches. For example, with a thiscall that takes
3408 // only one argument.
3409 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3410 CallerCC == CallingConv::X86_ThisCall))
3413 // Do not sibcall optimize vararg calls unless all arguments are passed via
3415 if (isVarArg && !Outs.empty()) {
3417 // Optimizing for varargs on Win64 is unlikely to be safe without
3418 // additional testing.
3419 if (IsCalleeWin64 || IsCallerWin64)
3422 SmallVector<CCValAssign, 16> ArgLocs;
3423 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3426 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3427 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3428 if (!ArgLocs[i].isRegLoc())
3432 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3433 // stack. Therefore, if it's not used by the call it is not safe to optimize
3434 // this into a sibcall.
3435 bool Unused = false;
3436 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3443 SmallVector<CCValAssign, 16> RVLocs;
3444 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3446 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3447 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3448 CCValAssign &VA = RVLocs[i];
3449 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3454 // If the calling conventions do not match, then we'd better make sure the
3455 // results are returned in the same way as what the caller expects.
3457 SmallVector<CCValAssign, 16> RVLocs1;
3458 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3460 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3462 SmallVector<CCValAssign, 16> RVLocs2;
3463 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3465 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3467 if (RVLocs1.size() != RVLocs2.size())
3469 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3470 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3472 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3474 if (RVLocs1[i].isRegLoc()) {
3475 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3478 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3484 // If the callee takes no arguments then go on to check the results of the
3486 if (!Outs.empty()) {
3487 // Check if stack adjustment is needed. For now, do not do this if any
3488 // argument is passed on the stack.
3489 SmallVector<CCValAssign, 16> ArgLocs;
3490 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3493 // Allocate shadow area for Win64
3495 CCInfo.AllocateStack(32, 8);
3497 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3498 if (CCInfo.getNextStackOffset()) {
3499 MachineFunction &MF = DAG.getMachineFunction();
3500 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3503 // Check if the arguments are already laid out in the right way as
3504 // the caller's fixed stack objects.
3505 MachineFrameInfo *MFI = MF.getFrameInfo();
3506 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3507 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3508 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3509 CCValAssign &VA = ArgLocs[i];
3510 SDValue Arg = OutVals[i];
3511 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3512 if (VA.getLocInfo() == CCValAssign::Indirect)
3514 if (!VA.isRegLoc()) {
3515 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3522 // If the tailcall address may be in a register, then make sure it's
3523 // possible to register allocate for it. In 32-bit, the call address can
3524 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3525 // callee-saved registers are restored. These happen to be the same
3526 // registers used to pass 'inreg' arguments so watch out for those.
3527 if (!Subtarget->is64Bit() &&
3528 ((!isa<GlobalAddressSDNode>(Callee) &&
3529 !isa<ExternalSymbolSDNode>(Callee)) ||
3530 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3531 unsigned NumInRegs = 0;
3532 // In PIC we need an extra register to formulate the address computation
3534 unsigned MaxInRegs =
3535 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3537 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3538 CCValAssign &VA = ArgLocs[i];
3541 unsigned Reg = VA.getLocReg();
3544 case X86::EAX: case X86::EDX: case X86::ECX:
3545 if (++NumInRegs == MaxInRegs)
3557 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3558 const TargetLibraryInfo *libInfo) const {
3559 return X86::createFastISel(funcInfo, libInfo);
3562 //===----------------------------------------------------------------------===//
3563 // Other Lowering Hooks
3564 //===----------------------------------------------------------------------===//
3566 static bool MayFoldLoad(SDValue Op) {
3567 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3570 static bool MayFoldIntoStore(SDValue Op) {
3571 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3574 static bool isTargetShuffle(unsigned Opcode) {
3576 default: return false;
3577 case X86ISD::BLENDI:
3578 case X86ISD::PSHUFB:
3579 case X86ISD::PSHUFD:
3580 case X86ISD::PSHUFHW:
3581 case X86ISD::PSHUFLW:
3583 case X86ISD::PALIGNR:
3584 case X86ISD::MOVLHPS:
3585 case X86ISD::MOVLHPD:
3586 case X86ISD::MOVHLPS:
3587 case X86ISD::MOVLPS:
3588 case X86ISD::MOVLPD:
3589 case X86ISD::MOVSHDUP:
3590 case X86ISD::MOVSLDUP:
3591 case X86ISD::MOVDDUP:
3594 case X86ISD::UNPCKL:
3595 case X86ISD::UNPCKH:
3596 case X86ISD::VPERMILPI:
3597 case X86ISD::VPERM2X128:
3598 case X86ISD::VPERMI:
3603 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3604 SDValue V1, unsigned TargetMask,
3605 SelectionDAG &DAG) {
3607 default: llvm_unreachable("Unknown x86 shuffle node");
3608 case X86ISD::PSHUFD:
3609 case X86ISD::PSHUFHW:
3610 case X86ISD::PSHUFLW:
3611 case X86ISD::VPERMILPI:
3612 case X86ISD::VPERMI:
3613 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3617 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3618 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3620 default: llvm_unreachable("Unknown x86 shuffle node");
3621 case X86ISD::MOVLHPS:
3622 case X86ISD::MOVLHPD:
3623 case X86ISD::MOVHLPS:
3624 case X86ISD::MOVLPS:
3625 case X86ISD::MOVLPD:
3628 case X86ISD::UNPCKL:
3629 case X86ISD::UNPCKH:
3630 return DAG.getNode(Opc, dl, VT, V1, V2);
3634 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3635 MachineFunction &MF = DAG.getMachineFunction();
3636 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3637 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3638 int ReturnAddrIndex = FuncInfo->getRAIndex();
3640 if (ReturnAddrIndex == 0) {
3641 // Set up a frame object for the return address.
3642 unsigned SlotSize = RegInfo->getSlotSize();
3643 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3646 FuncInfo->setRAIndex(ReturnAddrIndex);
3649 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3652 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3653 bool hasSymbolicDisplacement) {
3654 // Offset should fit into 32 bit immediate field.
3655 if (!isInt<32>(Offset))
3658 // If we don't have a symbolic displacement - we don't have any extra
3660 if (!hasSymbolicDisplacement)
3663 // FIXME: Some tweaks might be needed for medium code model.
3664 if (M != CodeModel::Small && M != CodeModel::Kernel)
3667 // For small code model we assume that latest object is 16MB before end of 31
3668 // bits boundary. We may also accept pretty large negative constants knowing
3669 // that all objects are in the positive half of address space.
3670 if (M == CodeModel::Small && Offset < 16*1024*1024)
3673 // For kernel code model we know that all object resist in the negative half
3674 // of 32bits address space. We may not accept negative offsets, since they may
3675 // be just off and we may accept pretty large positive ones.
3676 if (M == CodeModel::Kernel && Offset >= 0)
3682 /// isCalleePop - Determines whether the callee is required to pop its
3683 /// own arguments. Callee pop is necessary to support tail calls.
3684 bool X86::isCalleePop(CallingConv::ID CallingConv,
3685 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3686 switch (CallingConv) {
3689 case CallingConv::X86_StdCall:
3690 case CallingConv::X86_FastCall:
3691 case CallingConv::X86_ThisCall:
3693 case CallingConv::Fast:
3694 case CallingConv::GHC:
3695 case CallingConv::HiPE:
3702 /// \brief Return true if the condition is an unsigned comparison operation.
3703 static bool isX86CCUnsigned(unsigned X86CC) {
3705 default: llvm_unreachable("Invalid integer condition!");
3706 case X86::COND_E: return true;
3707 case X86::COND_G: return false;
3708 case X86::COND_GE: return false;
3709 case X86::COND_L: return false;
3710 case X86::COND_LE: return false;
3711 case X86::COND_NE: return true;
3712 case X86::COND_B: return true;
3713 case X86::COND_A: return true;
3714 case X86::COND_BE: return true;
3715 case X86::COND_AE: return true;
3717 llvm_unreachable("covered switch fell through?!");
3720 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3721 /// specific condition code, returning the condition code and the LHS/RHS of the
3722 /// comparison to make.
3723 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3724 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3726 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3727 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3728 // X > -1 -> X == 0, jump !sign.
3729 RHS = DAG.getConstant(0, RHS.getValueType());
3730 return X86::COND_NS;
3732 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3733 // X < 0 -> X == 0, jump on sign.
3736 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3738 RHS = DAG.getConstant(0, RHS.getValueType());
3739 return X86::COND_LE;
3743 switch (SetCCOpcode) {
3744 default: llvm_unreachable("Invalid integer condition!");
3745 case ISD::SETEQ: return X86::COND_E;
3746 case ISD::SETGT: return X86::COND_G;
3747 case ISD::SETGE: return X86::COND_GE;
3748 case ISD::SETLT: return X86::COND_L;
3749 case ISD::SETLE: return X86::COND_LE;
3750 case ISD::SETNE: return X86::COND_NE;
3751 case ISD::SETULT: return X86::COND_B;
3752 case ISD::SETUGT: return X86::COND_A;
3753 case ISD::SETULE: return X86::COND_BE;
3754 case ISD::SETUGE: return X86::COND_AE;
3758 // First determine if it is required or is profitable to flip the operands.
3760 // If LHS is a foldable load, but RHS is not, flip the condition.
3761 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3762 !ISD::isNON_EXTLoad(RHS.getNode())) {
3763 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3764 std::swap(LHS, RHS);
3767 switch (SetCCOpcode) {
3773 std::swap(LHS, RHS);
3777 // On a floating point condition, the flags are set as follows:
3779 // 0 | 0 | 0 | X > Y
3780 // 0 | 0 | 1 | X < Y
3781 // 1 | 0 | 0 | X == Y
3782 // 1 | 1 | 1 | unordered
3783 switch (SetCCOpcode) {
3784 default: llvm_unreachable("Condcode should be pre-legalized away");
3786 case ISD::SETEQ: return X86::COND_E;
3787 case ISD::SETOLT: // flipped
3789 case ISD::SETGT: return X86::COND_A;
3790 case ISD::SETOLE: // flipped
3792 case ISD::SETGE: return X86::COND_AE;
3793 case ISD::SETUGT: // flipped
3795 case ISD::SETLT: return X86::COND_B;
3796 case ISD::SETUGE: // flipped
3798 case ISD::SETLE: return X86::COND_BE;
3800 case ISD::SETNE: return X86::COND_NE;
3801 case ISD::SETUO: return X86::COND_P;
3802 case ISD::SETO: return X86::COND_NP;
3804 case ISD::SETUNE: return X86::COND_INVALID;
3808 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3809 /// code. Current x86 isa includes the following FP cmov instructions:
3810 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3811 static bool hasFPCMov(unsigned X86CC) {
3827 /// isFPImmLegal - Returns true if the target can instruction select the
3828 /// specified FP immediate natively. If false, the legalizer will
3829 /// materialize the FP immediate as a load from a constant pool.
3830 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3831 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3832 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3838 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3839 ISD::LoadExtType ExtTy,
3841 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3842 // relocation target a movq or addq instruction: don't let the load shrink.
3843 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3844 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3845 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3846 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3850 /// \brief Returns true if it is beneficial to convert a load of a constant
3851 /// to just the constant itself.
3852 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3854 assert(Ty->isIntegerTy());
3856 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3857 if (BitSize == 0 || BitSize > 64)
3862 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3863 unsigned Index) const {
3864 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3867 return (Index == 0 || Index == ResVT.getVectorNumElements());
3870 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3871 // Speculate cttz only if we can directly use TZCNT.
3872 return Subtarget->hasBMI();
3875 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3876 // Speculate ctlz only if we can directly use LZCNT.
3877 return Subtarget->hasLZCNT();
3880 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3881 /// the specified range (L, H].
3882 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3883 return (Val < 0) || (Val >= Low && Val < Hi);
3886 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3887 /// specified value.
3888 static bool isUndefOrEqual(int Val, int CmpVal) {
3889 return (Val < 0 || Val == CmpVal);
3892 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3893 /// from position Pos and ending in Pos+Size, falls within the specified
3894 /// sequential range (Low, Low+Size]. or is undef.
3895 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3896 unsigned Pos, unsigned Size, int Low) {
3897 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3898 if (!isUndefOrEqual(Mask[i], Low))
3903 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3904 /// the two vector operands have swapped position.
3905 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3906 unsigned NumElems) {
3907 for (unsigned i = 0; i != NumElems; ++i) {
3911 else if (idx < (int)NumElems)
3912 Mask[i] = idx + NumElems;
3914 Mask[i] = idx - NumElems;
3918 /// isVEXTRACTIndex - Return true if the specified
3919 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3920 /// suitable for instruction that extract 128 or 256 bit vectors
3921 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3922 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3923 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3926 // The index should be aligned on a vecWidth-bit boundary.
3928 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3930 MVT VT = N->getSimpleValueType(0);
3931 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3932 bool Result = (Index * ElSize) % vecWidth == 0;
3937 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3938 /// operand specifies a subvector insert that is suitable for input to
3939 /// insertion of 128 or 256-bit subvectors
3940 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3941 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3942 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3944 // The index should be aligned on a vecWidth-bit boundary.
3946 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3948 MVT VT = N->getSimpleValueType(0);
3949 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3950 bool Result = (Index * ElSize) % vecWidth == 0;
3955 bool X86::isVINSERT128Index(SDNode *N) {
3956 return isVINSERTIndex(N, 128);
3959 bool X86::isVINSERT256Index(SDNode *N) {
3960 return isVINSERTIndex(N, 256);
3963 bool X86::isVEXTRACT128Index(SDNode *N) {
3964 return isVEXTRACTIndex(N, 128);
3967 bool X86::isVEXTRACT256Index(SDNode *N) {
3968 return isVEXTRACTIndex(N, 256);
3971 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3972 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3973 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3974 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3977 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3979 MVT VecVT = N->getOperand(0).getSimpleValueType();
3980 MVT ElVT = VecVT.getVectorElementType();
3982 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3983 return Index / NumElemsPerChunk;
3986 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3987 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3988 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3989 llvm_unreachable("Illegal insert subvector for VINSERT");
3992 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3994 MVT VecVT = N->getSimpleValueType(0);
3995 MVT ElVT = VecVT.getVectorElementType();
3997 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3998 return Index / NumElemsPerChunk;
4001 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4002 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4003 /// and VINSERTI128 instructions.
4004 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4005 return getExtractVEXTRACTImmediate(N, 128);
4008 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4009 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4010 /// and VINSERTI64x4 instructions.
4011 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4012 return getExtractVEXTRACTImmediate(N, 256);
4015 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4016 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4017 /// and VINSERTI128 instructions.
4018 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4019 return getInsertVINSERTImmediate(N, 128);
4022 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4023 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4024 /// and VINSERTI64x4 instructions.
4025 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4026 return getInsertVINSERTImmediate(N, 256);
4029 /// isZero - Returns true if Elt is a constant integer zero
4030 static bool isZero(SDValue V) {
4031 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4032 return C && C->isNullValue();
4035 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4037 bool X86::isZeroNode(SDValue Elt) {
4040 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4041 return CFP->getValueAPF().isPosZero();
4045 /// getZeroVector - Returns a vector of specified type with all zero elements.
4047 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4048 SelectionDAG &DAG, SDLoc dl) {
4049 assert(VT.isVector() && "Expected a vector type");
4051 // Always build SSE zero vectors as <4 x i32> bitcasted
4052 // to their dest type. This ensures they get CSE'd.
4054 if (VT.is128BitVector()) { // SSE
4055 if (Subtarget->hasSSE2()) { // SSE2
4056 SDValue Cst = DAG.getConstant(0, MVT::i32);
4057 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4059 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4060 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4062 } else if (VT.is256BitVector()) { // AVX
4063 if (Subtarget->hasInt256()) { // AVX2
4064 SDValue Cst = DAG.getConstant(0, MVT::i32);
4065 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4066 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4068 // 256-bit logic and arithmetic instructions in AVX are all
4069 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4070 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4071 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4072 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4074 } else if (VT.is512BitVector()) { // AVX-512
4075 SDValue Cst = DAG.getConstant(0, MVT::i32);
4076 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4077 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4078 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4079 } else if (VT.getScalarType() == MVT::i1) {
4080 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
4081 SDValue Cst = DAG.getConstant(0, MVT::i1);
4082 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
4083 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4085 llvm_unreachable("Unexpected vector type");
4087 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4090 /// getOnesVector - Returns a vector of specified type with all bits set.
4091 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4092 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4093 /// Then bitcast to their original type, ensuring they get CSE'd.
4094 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4096 assert(VT.isVector() && "Expected a vector type");
4098 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
4100 if (VT.is256BitVector()) {
4101 if (HasInt256) { // AVX2
4102 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4103 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4105 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4106 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4108 } else if (VT.is128BitVector()) {
4109 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4111 llvm_unreachable("Unexpected vector type");
4113 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4116 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4117 /// operation of specified width.
4118 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4120 unsigned NumElems = VT.getVectorNumElements();
4121 SmallVector<int, 8> Mask;
4122 Mask.push_back(NumElems);
4123 for (unsigned i = 1; i != NumElems; ++i)
4125 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4128 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4129 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4131 unsigned NumElems = VT.getVectorNumElements();
4132 SmallVector<int, 8> Mask;
4133 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4135 Mask.push_back(i + NumElems);
4137 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4140 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4141 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4143 unsigned NumElems = VT.getVectorNumElements();
4144 SmallVector<int, 8> Mask;
4145 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4146 Mask.push_back(i + Half);
4147 Mask.push_back(i + NumElems + Half);
4149 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4152 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4153 /// vector of zero or undef vector. This produces a shuffle where the low
4154 /// element of V2 is swizzled into the zero/undef vector, landing at element
4155 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4156 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4158 const X86Subtarget *Subtarget,
4159 SelectionDAG &DAG) {
4160 MVT VT = V2.getSimpleValueType();
4162 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4163 unsigned NumElems = VT.getVectorNumElements();
4164 SmallVector<int, 16> MaskVec;
4165 for (unsigned i = 0; i != NumElems; ++i)
4166 // If this is the insertion idx, put the low elt of V2 here.
4167 MaskVec.push_back(i == Idx ? NumElems : i);
4168 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4171 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4172 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4173 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4174 /// shuffles which use a single input multiple times, and in those cases it will
4175 /// adjust the mask to only have indices within that single input.
4176 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4177 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4178 unsigned NumElems = VT.getVectorNumElements();
4182 bool IsFakeUnary = false;
4183 switch(N->getOpcode()) {
4184 case X86ISD::BLENDI:
4185 ImmN = N->getOperand(N->getNumOperands()-1);
4186 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4189 ImmN = N->getOperand(N->getNumOperands()-1);
4190 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4191 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4193 case X86ISD::UNPCKH:
4194 DecodeUNPCKHMask(VT, Mask);
4195 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4197 case X86ISD::UNPCKL:
4198 DecodeUNPCKLMask(VT, Mask);
4199 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4201 case X86ISD::MOVHLPS:
4202 DecodeMOVHLPSMask(NumElems, Mask);
4203 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4205 case X86ISD::MOVLHPS:
4206 DecodeMOVLHPSMask(NumElems, Mask);
4207 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4209 case X86ISD::PALIGNR:
4210 ImmN = N->getOperand(N->getNumOperands()-1);
4211 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4213 case X86ISD::PSHUFD:
4214 case X86ISD::VPERMILPI:
4215 ImmN = N->getOperand(N->getNumOperands()-1);
4216 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4219 case X86ISD::PSHUFHW:
4220 ImmN = N->getOperand(N->getNumOperands()-1);
4221 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4224 case X86ISD::PSHUFLW:
4225 ImmN = N->getOperand(N->getNumOperands()-1);
4226 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4229 case X86ISD::PSHUFB: {
4231 SDValue MaskNode = N->getOperand(1);
4232 while (MaskNode->getOpcode() == ISD::BITCAST)
4233 MaskNode = MaskNode->getOperand(0);
4235 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4236 // If we have a build-vector, then things are easy.
4237 EVT VT = MaskNode.getValueType();
4238 assert(VT.isVector() &&
4239 "Can't produce a non-vector with a build_vector!");
4240 if (!VT.isInteger())
4243 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4245 SmallVector<uint64_t, 32> RawMask;
4246 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4247 SDValue Op = MaskNode->getOperand(i);
4248 if (Op->getOpcode() == ISD::UNDEF) {
4249 RawMask.push_back((uint64_t)SM_SentinelUndef);
4252 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4255 APInt MaskElement = CN->getAPIntValue();
4257 // We now have to decode the element which could be any integer size and
4258 // extract each byte of it.
4259 for (int j = 0; j < NumBytesPerElement; ++j) {
4260 // Note that this is x86 and so always little endian: the low byte is
4261 // the first byte of the mask.
4262 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4263 MaskElement = MaskElement.lshr(8);
4266 DecodePSHUFBMask(RawMask, Mask);
4270 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4274 SDValue Ptr = MaskLoad->getBasePtr();
4275 if (Ptr->getOpcode() == X86ISD::Wrapper)
4276 Ptr = Ptr->getOperand(0);
4278 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4279 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4282 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4283 DecodePSHUFBMask(C, Mask);
4291 case X86ISD::VPERMI:
4292 ImmN = N->getOperand(N->getNumOperands()-1);
4293 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4298 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4300 case X86ISD::VPERM2X128:
4301 ImmN = N->getOperand(N->getNumOperands()-1);
4302 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4303 if (Mask.empty()) return false;
4305 case X86ISD::MOVSLDUP:
4306 DecodeMOVSLDUPMask(VT, Mask);
4309 case X86ISD::MOVSHDUP:
4310 DecodeMOVSHDUPMask(VT, Mask);
4313 case X86ISD::MOVDDUP:
4314 DecodeMOVDDUPMask(VT, Mask);
4317 case X86ISD::MOVLHPD:
4318 case X86ISD::MOVLPD:
4319 case X86ISD::MOVLPS:
4320 // Not yet implemented
4322 default: llvm_unreachable("unknown target shuffle node");
4325 // If we have a fake unary shuffle, the shuffle mask is spread across two
4326 // inputs that are actually the same node. Re-map the mask to always point
4327 // into the first input.
4330 if (M >= (int)Mask.size())
4336 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4337 /// element of the result of the vector shuffle.
4338 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4341 return SDValue(); // Limit search depth.
4343 SDValue V = SDValue(N, 0);
4344 EVT VT = V.getValueType();
4345 unsigned Opcode = V.getOpcode();
4347 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4348 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4349 int Elt = SV->getMaskElt(Index);
4352 return DAG.getUNDEF(VT.getVectorElementType());
4354 unsigned NumElems = VT.getVectorNumElements();
4355 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4356 : SV->getOperand(1);
4357 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4360 // Recurse into target specific vector shuffles to find scalars.
4361 if (isTargetShuffle(Opcode)) {
4362 MVT ShufVT = V.getSimpleValueType();
4363 unsigned NumElems = ShufVT.getVectorNumElements();
4364 SmallVector<int, 16> ShuffleMask;
4367 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4370 int Elt = ShuffleMask[Index];
4372 return DAG.getUNDEF(ShufVT.getVectorElementType());
4374 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4376 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4380 // Actual nodes that may contain scalar elements
4381 if (Opcode == ISD::BITCAST) {
4382 V = V.getOperand(0);
4383 EVT SrcVT = V.getValueType();
4384 unsigned NumElems = VT.getVectorNumElements();
4386 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4390 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4391 return (Index == 0) ? V.getOperand(0)
4392 : DAG.getUNDEF(VT.getVectorElementType());
4394 if (V.getOpcode() == ISD::BUILD_VECTOR)
4395 return V.getOperand(Index);
4400 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4402 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4403 unsigned NumNonZero, unsigned NumZero,
4405 const X86Subtarget* Subtarget,
4406 const TargetLowering &TLI) {
4413 for (unsigned i = 0; i < 16; ++i) {
4414 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4415 if (ThisIsNonZero && First) {
4417 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4419 V = DAG.getUNDEF(MVT::v8i16);
4424 SDValue ThisElt, LastElt;
4425 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4426 if (LastIsNonZero) {
4427 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4428 MVT::i16, Op.getOperand(i-1));
4430 if (ThisIsNonZero) {
4431 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4432 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4433 ThisElt, DAG.getConstant(8, MVT::i8));
4435 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4439 if (ThisElt.getNode())
4440 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4441 DAG.getIntPtrConstant(i/2));
4445 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4448 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4450 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4451 unsigned NumNonZero, unsigned NumZero,
4453 const X86Subtarget* Subtarget,
4454 const TargetLowering &TLI) {
4461 for (unsigned i = 0; i < 8; ++i) {
4462 bool isNonZero = (NonZeros & (1 << i)) != 0;
4466 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4468 V = DAG.getUNDEF(MVT::v8i16);
4471 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4472 MVT::v8i16, V, Op.getOperand(i),
4473 DAG.getIntPtrConstant(i));
4480 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4481 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4482 const X86Subtarget *Subtarget,
4483 const TargetLowering &TLI) {
4484 // Find all zeroable elements.
4485 std::bitset<4> Zeroable;
4486 for (int i=0; i < 4; ++i) {
4487 SDValue Elt = Op->getOperand(i);
4488 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4490 assert(Zeroable.size() - Zeroable.count() > 1 &&
4491 "We expect at least two non-zero elements!");
4493 // We only know how to deal with build_vector nodes where elements are either
4494 // zeroable or extract_vector_elt with constant index.
4495 SDValue FirstNonZero;
4496 unsigned FirstNonZeroIdx;
4497 for (unsigned i=0; i < 4; ++i) {
4500 SDValue Elt = Op->getOperand(i);
4501 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4502 !isa<ConstantSDNode>(Elt.getOperand(1)))
4504 // Make sure that this node is extracting from a 128-bit vector.
4505 MVT VT = Elt.getOperand(0).getSimpleValueType();
4506 if (!VT.is128BitVector())
4508 if (!FirstNonZero.getNode()) {
4510 FirstNonZeroIdx = i;
4514 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4515 SDValue V1 = FirstNonZero.getOperand(0);
4516 MVT VT = V1.getSimpleValueType();
4518 // See if this build_vector can be lowered as a blend with zero.
4520 unsigned EltMaskIdx, EltIdx;
4522 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4523 if (Zeroable[EltIdx]) {
4524 // The zero vector will be on the right hand side.
4525 Mask[EltIdx] = EltIdx+4;
4529 Elt = Op->getOperand(EltIdx);
4530 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4531 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4532 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4534 Mask[EltIdx] = EltIdx;
4538 // Let the shuffle legalizer deal with blend operations.
4539 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4540 if (V1.getSimpleValueType() != VT)
4541 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4542 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4545 // See if we can lower this build_vector to a INSERTPS.
4546 if (!Subtarget->hasSSE41())
4549 SDValue V2 = Elt.getOperand(0);
4550 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4553 bool CanFold = true;
4554 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4558 SDValue Current = Op->getOperand(i);
4559 SDValue SrcVector = Current->getOperand(0);
4562 CanFold = SrcVector == V1 &&
4563 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4569 assert(V1.getNode() && "Expected at least two non-zero elements!");
4570 if (V1.getSimpleValueType() != MVT::v4f32)
4571 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4572 if (V2.getSimpleValueType() != MVT::v4f32)
4573 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4575 // Ok, we can emit an INSERTPS instruction.
4576 unsigned ZMask = Zeroable.to_ulong();
4578 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4579 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4580 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
4581 DAG.getIntPtrConstant(InsertPSMask));
4582 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
4585 /// Return a vector logical shift node.
4586 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4587 unsigned NumBits, SelectionDAG &DAG,
4588 const TargetLowering &TLI, SDLoc dl) {
4589 assert(VT.is128BitVector() && "Unknown type for VShift");
4590 MVT ShVT = MVT::v2i64;
4591 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4592 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4593 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4594 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4595 SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
4596 return DAG.getNode(ISD::BITCAST, dl, VT,
4597 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4601 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4603 // Check if the scalar load can be widened into a vector load. And if
4604 // the address is "base + cst" see if the cst can be "absorbed" into
4605 // the shuffle mask.
4606 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4607 SDValue Ptr = LD->getBasePtr();
4608 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4610 EVT PVT = LD->getValueType(0);
4611 if (PVT != MVT::i32 && PVT != MVT::f32)
4616 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4617 FI = FINode->getIndex();
4619 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4620 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4621 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4622 Offset = Ptr.getConstantOperandVal(1);
4623 Ptr = Ptr.getOperand(0);
4628 // FIXME: 256-bit vector instructions don't require a strict alignment,
4629 // improve this code to support it better.
4630 unsigned RequiredAlign = VT.getSizeInBits()/8;
4631 SDValue Chain = LD->getChain();
4632 // Make sure the stack object alignment is at least 16 or 32.
4633 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4634 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4635 if (MFI->isFixedObjectIndex(FI)) {
4636 // Can't change the alignment. FIXME: It's possible to compute
4637 // the exact stack offset and reference FI + adjust offset instead.
4638 // If someone *really* cares about this. That's the way to implement it.
4641 MFI->setObjectAlignment(FI, RequiredAlign);
4645 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4646 // Ptr + (Offset & ~15).
4649 if ((Offset % RequiredAlign) & 3)
4651 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4653 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
4654 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4656 int EltNo = (Offset - StartOffset) >> 2;
4657 unsigned NumElems = VT.getVectorNumElements();
4659 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4660 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4661 LD->getPointerInfo().getWithOffset(StartOffset),
4662 false, false, false, 0);
4664 SmallVector<int, 8> Mask(NumElems, EltNo);
4666 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4672 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4673 /// elements can be replaced by a single large load which has the same value as
4674 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4676 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4678 /// FIXME: we'd also like to handle the case where the last elements are zero
4679 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4680 /// There's even a handy isZeroNode for that purpose.
4681 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4682 SDLoc &DL, SelectionDAG &DAG,
4683 bool isAfterLegalize) {
4684 unsigned NumElems = Elts.size();
4686 LoadSDNode *LDBase = nullptr;
4687 unsigned LastLoadedElt = -1U;
4689 // For each element in the initializer, see if we've found a load or an undef.
4690 // If we don't find an initial load element, or later load elements are
4691 // non-consecutive, bail out.
4692 for (unsigned i = 0; i < NumElems; ++i) {
4693 SDValue Elt = Elts[i];
4694 // Look through a bitcast.
4695 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4696 Elt = Elt.getOperand(0);
4697 if (!Elt.getNode() ||
4698 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4701 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4703 LDBase = cast<LoadSDNode>(Elt.getNode());
4707 if (Elt.getOpcode() == ISD::UNDEF)
4710 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4711 EVT LdVT = Elt.getValueType();
4712 // Each loaded element must be the correct fractional portion of the
4713 // requested vector load.
4714 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4716 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4721 // If we have found an entire vector of loads and undefs, then return a large
4722 // load of the entire vector width starting at the base pointer. If we found
4723 // consecutive loads for the low half, generate a vzext_load node.
4724 if (LastLoadedElt == NumElems - 1) {
4725 assert(LDBase && "Did not find base load for merging consecutive loads");
4726 EVT EltVT = LDBase->getValueType(0);
4727 // Ensure that the input vector size for the merged loads matches the
4728 // cumulative size of the input elements.
4729 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4732 if (isAfterLegalize &&
4733 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4736 SDValue NewLd = SDValue();
4738 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4739 LDBase->getPointerInfo(), LDBase->isVolatile(),
4740 LDBase->isNonTemporal(), LDBase->isInvariant(),
4741 LDBase->getAlignment());
4743 if (LDBase->hasAnyUseOfValue(1)) {
4744 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4746 SDValue(NewLd.getNode(), 1));
4747 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4748 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4749 SDValue(NewLd.getNode(), 1));
4755 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4756 //of a v4i32 / v4f32. It's probably worth generalizing.
4757 EVT EltVT = VT.getVectorElementType();
4758 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4759 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4760 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4761 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4763 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4764 LDBase->getPointerInfo(),
4765 LDBase->getAlignment(),
4766 false/*isVolatile*/, true/*ReadMem*/,
4769 // Make sure the newly-created LOAD is in the same position as LDBase in
4770 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4771 // update uses of LDBase's output chain to use the TokenFactor.
4772 if (LDBase->hasAnyUseOfValue(1)) {
4773 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4774 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4775 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4776 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4777 SDValue(ResNode.getNode(), 1));
4780 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4785 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4786 /// to generate a splat value for the following cases:
4787 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4788 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4789 /// a scalar load, or a constant.
4790 /// The VBROADCAST node is returned when a pattern is found,
4791 /// or SDValue() otherwise.
4792 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4793 SelectionDAG &DAG) {
4794 // VBROADCAST requires AVX.
4795 // TODO: Splats could be generated for non-AVX CPUs using SSE
4796 // instructions, but there's less potential gain for only 128-bit vectors.
4797 if (!Subtarget->hasAVX())
4800 MVT VT = Op.getSimpleValueType();
4803 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4804 "Unsupported vector type for broadcast.");
4809 switch (Op.getOpcode()) {
4811 // Unknown pattern found.
4814 case ISD::BUILD_VECTOR: {
4815 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4816 BitVector UndefElements;
4817 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4819 // We need a splat of a single value to use broadcast, and it doesn't
4820 // make any sense if the value is only in one element of the vector.
4821 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4825 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4826 Ld.getOpcode() == ISD::ConstantFP);
4828 // Make sure that all of the users of a non-constant load are from the
4829 // BUILD_VECTOR node.
4830 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4835 case ISD::VECTOR_SHUFFLE: {
4836 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4838 // Shuffles must have a splat mask where the first element is
4840 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4843 SDValue Sc = Op.getOperand(0);
4844 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4845 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4847 if (!Subtarget->hasInt256())
4850 // Use the register form of the broadcast instruction available on AVX2.
4851 if (VT.getSizeInBits() >= 256)
4852 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4853 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
4856 Ld = Sc.getOperand(0);
4857 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4858 Ld.getOpcode() == ISD::ConstantFP);
4860 // The scalar_to_vector node and the suspected
4861 // load node must have exactly one user.
4862 // Constants may have multiple users.
4864 // AVX-512 has register version of the broadcast
4865 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
4866 Ld.getValueType().getSizeInBits() >= 32;
4867 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
4874 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
4875 bool IsGE256 = (VT.getSizeInBits() >= 256);
4877 // When optimizing for size, generate up to 5 extra bytes for a broadcast
4878 // instruction to save 8 or more bytes of constant pool data.
4879 // TODO: If multiple splats are generated to load the same constant,
4880 // it may be detrimental to overall size. There needs to be a way to detect
4881 // that condition to know if this is truly a size win.
4882 const Function *F = DAG.getMachineFunction().getFunction();
4883 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
4885 // Handle broadcasting a single constant scalar from the constant pool
4887 // On Sandybridge (no AVX2), it is still better to load a constant vector
4888 // from the constant pool and not to broadcast it from a scalar.
4889 // But override that restriction when optimizing for size.
4890 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
4891 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
4892 EVT CVT = Ld.getValueType();
4893 assert(!CVT.isVector() && "Must not broadcast a vector type");
4895 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
4896 // For size optimization, also splat v2f64 and v2i64, and for size opt
4897 // with AVX2, also splat i8 and i16.
4898 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
4899 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4900 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
4901 const Constant *C = nullptr;
4902 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
4903 C = CI->getConstantIntValue();
4904 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
4905 C = CF->getConstantFPValue();
4907 assert(C && "Invalid constant type");
4909 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4910 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
4911 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
4912 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
4913 MachinePointerInfo::getConstantPool(),
4914 false, false, false, Alignment);
4916 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4920 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
4922 // Handle AVX2 in-register broadcasts.
4923 if (!IsLoad && Subtarget->hasInt256() &&
4924 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
4925 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4927 // The scalar source must be a normal load.
4931 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4932 (Subtarget->hasVLX() && ScalarSize == 64))
4933 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4935 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
4936 // double since there is no vbroadcastsd xmm
4937 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
4938 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
4939 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4942 // Unsupported broadcast.
4946 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
4947 /// underlying vector and index.
4949 /// Modifies \p ExtractedFromVec to the real vector and returns the real
4951 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
4953 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
4954 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
4957 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
4959 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
4961 // (extract_vector_elt (vector_shuffle<2,u,u,u>
4962 // (extract_subvector (v8f32 %vreg0), Constant<4>),
4965 // In this case the vector is the extract_subvector expression and the index
4966 // is 2, as specified by the shuffle.
4967 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
4968 SDValue ShuffleVec = SVOp->getOperand(0);
4969 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
4970 assert(ShuffleVecVT.getVectorElementType() ==
4971 ExtractedFromVec.getSimpleValueType().getVectorElementType());
4973 int ShuffleIdx = SVOp->getMaskElt(Idx);
4974 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
4975 ExtractedFromVec = ShuffleVec;
4981 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
4982 MVT VT = Op.getSimpleValueType();
4984 // Skip if insert_vec_elt is not supported.
4985 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4986 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
4990 unsigned NumElems = Op.getNumOperands();
4994 SmallVector<unsigned, 4> InsertIndices;
4995 SmallVector<int, 8> Mask(NumElems, -1);
4997 for (unsigned i = 0; i != NumElems; ++i) {
4998 unsigned Opc = Op.getOperand(i).getOpcode();
5000 if (Opc == ISD::UNDEF)
5003 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5004 // Quit if more than 1 elements need inserting.
5005 if (InsertIndices.size() > 1)
5008 InsertIndices.push_back(i);
5012 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5013 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5014 // Quit if non-constant index.
5015 if (!isa<ConstantSDNode>(ExtIdx))
5017 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5019 // Quit if extracted from vector of different type.
5020 if (ExtractedFromVec.getValueType() != VT)
5023 if (!VecIn1.getNode())
5024 VecIn1 = ExtractedFromVec;
5025 else if (VecIn1 != ExtractedFromVec) {
5026 if (!VecIn2.getNode())
5027 VecIn2 = ExtractedFromVec;
5028 else if (VecIn2 != ExtractedFromVec)
5029 // Quit if more than 2 vectors to shuffle
5033 if (ExtractedFromVec == VecIn1)
5035 else if (ExtractedFromVec == VecIn2)
5036 Mask[i] = Idx + NumElems;
5039 if (!VecIn1.getNode())
5042 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5043 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5044 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5045 unsigned Idx = InsertIndices[i];
5046 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5047 DAG.getIntPtrConstant(Idx));
5053 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5055 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5057 MVT VT = Op.getSimpleValueType();
5058 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5059 "Unexpected type in LowerBUILD_VECTORvXi1!");
5062 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5063 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5064 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5065 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5068 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5069 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5070 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5071 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5074 bool AllContants = true;
5075 uint64_t Immediate = 0;
5076 int NonConstIdx = -1;
5077 bool IsSplat = true;
5078 unsigned NumNonConsts = 0;
5079 unsigned NumConsts = 0;
5080 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5081 SDValue In = Op.getOperand(idx);
5082 if (In.getOpcode() == ISD::UNDEF)
5084 if (!isa<ConstantSDNode>(In)) {
5085 AllContants = false;
5090 if (cast<ConstantSDNode>(In)->getZExtValue())
5091 Immediate |= (1ULL << idx);
5093 if (In != Op.getOperand(0))
5098 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5099 DAG.getConstant(Immediate, MVT::i16));
5100 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5101 DAG.getIntPtrConstant(0));
5104 if (NumNonConsts == 1 && NonConstIdx != 0) {
5107 SDValue VecAsImm = DAG.getConstant(Immediate,
5108 MVT::getIntegerVT(VT.getSizeInBits()));
5109 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
5112 DstVec = DAG.getUNDEF(VT);
5113 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5114 Op.getOperand(NonConstIdx),
5115 DAG.getIntPtrConstant(NonConstIdx));
5117 if (!IsSplat && (NonConstIdx != 0))
5118 llvm_unreachable("Unsupported BUILD_VECTOR operation");
5119 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
5122 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5123 DAG.getConstant(-1, SelectVT),
5124 DAG.getConstant(0, SelectVT));
5126 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5127 DAG.getConstant((Immediate | 1), SelectVT),
5128 DAG.getConstant(Immediate, SelectVT));
5129 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
5132 /// \brief Return true if \p N implements a horizontal binop and return the
5133 /// operands for the horizontal binop into V0 and V1.
5135 /// This is a helper function of PerformBUILD_VECTORCombine.
5136 /// This function checks that the build_vector \p N in input implements a
5137 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5138 /// operation to match.
5139 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5140 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5141 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5144 /// This function only analyzes elements of \p N whose indices are
5145 /// in range [BaseIdx, LastIdx).
5146 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5148 unsigned BaseIdx, unsigned LastIdx,
5149 SDValue &V0, SDValue &V1) {
5150 EVT VT = N->getValueType(0);
5152 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5153 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5154 "Invalid Vector in input!");
5156 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5157 bool CanFold = true;
5158 unsigned ExpectedVExtractIdx = BaseIdx;
5159 unsigned NumElts = LastIdx - BaseIdx;
5160 V0 = DAG.getUNDEF(VT);
5161 V1 = DAG.getUNDEF(VT);
5163 // Check if N implements a horizontal binop.
5164 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5165 SDValue Op = N->getOperand(i + BaseIdx);
5168 if (Op->getOpcode() == ISD::UNDEF) {
5169 // Update the expected vector extract index.
5170 if (i * 2 == NumElts)
5171 ExpectedVExtractIdx = BaseIdx;
5172 ExpectedVExtractIdx += 2;
5176 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5181 SDValue Op0 = Op.getOperand(0);
5182 SDValue Op1 = Op.getOperand(1);
5184 // Try to match the following pattern:
5185 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5186 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5187 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5188 Op0.getOperand(0) == Op1.getOperand(0) &&
5189 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5190 isa<ConstantSDNode>(Op1.getOperand(1)));
5194 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5195 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5197 if (i * 2 < NumElts) {
5198 if (V0.getOpcode() == ISD::UNDEF)
5199 V0 = Op0.getOperand(0);
5201 if (V1.getOpcode() == ISD::UNDEF)
5202 V1 = Op0.getOperand(0);
5203 if (i * 2 == NumElts)
5204 ExpectedVExtractIdx = BaseIdx;
5207 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5208 if (I0 == ExpectedVExtractIdx)
5209 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5210 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5211 // Try to match the following dag sequence:
5212 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5213 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5217 ExpectedVExtractIdx += 2;
5223 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5224 /// a concat_vector.
5226 /// This is a helper function of PerformBUILD_VECTORCombine.
5227 /// This function expects two 256-bit vectors called V0 and V1.
5228 /// At first, each vector is split into two separate 128-bit vectors.
5229 /// Then, the resulting 128-bit vectors are used to implement two
5230 /// horizontal binary operations.
5232 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5234 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5235 /// the two new horizontal binop.
5236 /// When Mode is set, the first horizontal binop dag node would take as input
5237 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5238 /// horizontal binop dag node would take as input the lower 128-bit of V1
5239 /// and the upper 128-bit of V1.
5241 /// HADD V0_LO, V0_HI
5242 /// HADD V1_LO, V1_HI
5244 /// Otherwise, the first horizontal binop dag node takes as input the lower
5245 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5246 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5248 /// HADD V0_LO, V1_LO
5249 /// HADD V0_HI, V1_HI
5251 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5252 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5253 /// the upper 128-bits of the result.
5254 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5255 SDLoc DL, SelectionDAG &DAG,
5256 unsigned X86Opcode, bool Mode,
5257 bool isUndefLO, bool isUndefHI) {
5258 EVT VT = V0.getValueType();
5259 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5260 "Invalid nodes in input!");
5262 unsigned NumElts = VT.getVectorNumElements();
5263 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5264 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5265 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5266 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5267 EVT NewVT = V0_LO.getValueType();
5269 SDValue LO = DAG.getUNDEF(NewVT);
5270 SDValue HI = DAG.getUNDEF(NewVT);
5273 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5274 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5275 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5276 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5277 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5279 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5280 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5281 V1_LO->getOpcode() != ISD::UNDEF))
5282 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5284 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5285 V1_HI->getOpcode() != ISD::UNDEF))
5286 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5289 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5292 /// \brief Try to fold a build_vector that performs an 'addsub' into the
5293 /// sequence of 'vadd + vsub + blendi'.
5294 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
5295 const X86Subtarget *Subtarget) {
5297 EVT VT = BV->getValueType(0);
5298 unsigned NumElts = VT.getVectorNumElements();
5299 SDValue InVec0 = DAG.getUNDEF(VT);
5300 SDValue InVec1 = DAG.getUNDEF(VT);
5302 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5303 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5305 // Odd-numbered elements in the input build vector are obtained from
5306 // adding two integer/float elements.
5307 // Even-numbered elements in the input build vector are obtained from
5308 // subtracting two integer/float elements.
5309 unsigned ExpectedOpcode = ISD::FSUB;
5310 unsigned NextExpectedOpcode = ISD::FADD;
5311 bool AddFound = false;
5312 bool SubFound = false;
5314 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5315 SDValue Op = BV->getOperand(i);
5317 // Skip 'undef' values.
5318 unsigned Opcode = Op.getOpcode();
5319 if (Opcode == ISD::UNDEF) {
5320 std::swap(ExpectedOpcode, NextExpectedOpcode);
5324 // Early exit if we found an unexpected opcode.
5325 if (Opcode != ExpectedOpcode)
5328 SDValue Op0 = Op.getOperand(0);
5329 SDValue Op1 = Op.getOperand(1);
5331 // Try to match the following pattern:
5332 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5333 // Early exit if we cannot match that sequence.
5334 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5335 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5336 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5337 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5338 Op0.getOperand(1) != Op1.getOperand(1))
5341 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5345 // We found a valid add/sub node. Update the information accordingly.
5351 // Update InVec0 and InVec1.
5352 if (InVec0.getOpcode() == ISD::UNDEF)
5353 InVec0 = Op0.getOperand(0);
5354 if (InVec1.getOpcode() == ISD::UNDEF)
5355 InVec1 = Op1.getOperand(0);
5357 // Make sure that operands in input to each add/sub node always
5358 // come from a same pair of vectors.
5359 if (InVec0 != Op0.getOperand(0)) {
5360 if (ExpectedOpcode == ISD::FSUB)
5363 // FADD is commutable. Try to commute the operands
5364 // and then test again.
5365 std::swap(Op0, Op1);
5366 if (InVec0 != Op0.getOperand(0))
5370 if (InVec1 != Op1.getOperand(0))
5373 // Update the pair of expected opcodes.
5374 std::swap(ExpectedOpcode, NextExpectedOpcode);
5377 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5378 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5379 InVec1.getOpcode() != ISD::UNDEF)
5380 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5385 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
5386 const X86Subtarget *Subtarget) {
5388 EVT VT = N->getValueType(0);
5389 unsigned NumElts = VT.getVectorNumElements();
5390 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
5391 SDValue InVec0, InVec1;
5393 // Try to match an ADDSUB.
5394 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
5395 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
5396 SDValue Value = matchAddSub(BV, DAG, Subtarget);
5397 if (Value.getNode())
5401 // Try to match horizontal ADD/SUB.
5402 unsigned NumUndefsLO = 0;
5403 unsigned NumUndefsHI = 0;
5404 unsigned Half = NumElts/2;
5406 // Count the number of UNDEF operands in the build_vector in input.
5407 for (unsigned i = 0, e = Half; i != e; ++i)
5408 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5411 for (unsigned i = Half, e = NumElts; i != e; ++i)
5412 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5415 // Early exit if this is either a build_vector of all UNDEFs or all the
5416 // operands but one are UNDEF.
5417 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5420 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5421 // Try to match an SSE3 float HADD/HSUB.
5422 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5423 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5425 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5426 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5427 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5428 // Try to match an SSSE3 integer HADD/HSUB.
5429 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5430 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5432 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5433 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5436 if (!Subtarget->hasAVX())
5439 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5440 // Try to match an AVX horizontal add/sub of packed single/double
5441 // precision floating point values from 256-bit vectors.
5442 SDValue InVec2, InVec3;
5443 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5444 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5445 ((InVec0.getOpcode() == ISD::UNDEF ||
5446 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5447 ((InVec1.getOpcode() == ISD::UNDEF ||
5448 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5449 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5451 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5452 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5453 ((InVec0.getOpcode() == ISD::UNDEF ||
5454 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5455 ((InVec1.getOpcode() == ISD::UNDEF ||
5456 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5457 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5458 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5459 // Try to match an AVX2 horizontal add/sub of signed integers.
5460 SDValue InVec2, InVec3;
5462 bool CanFold = true;
5464 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5465 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5466 ((InVec0.getOpcode() == ISD::UNDEF ||
5467 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5468 ((InVec1.getOpcode() == ISD::UNDEF ||
5469 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5470 X86Opcode = X86ISD::HADD;
5471 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5472 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5473 ((InVec0.getOpcode() == ISD::UNDEF ||
5474 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5475 ((InVec1.getOpcode() == ISD::UNDEF ||
5476 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5477 X86Opcode = X86ISD::HSUB;
5482 // Fold this build_vector into a single horizontal add/sub.
5483 // Do this only if the target has AVX2.
5484 if (Subtarget->hasAVX2())
5485 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5487 // Do not try to expand this build_vector into a pair of horizontal
5488 // add/sub if we can emit a pair of scalar add/sub.
5489 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5492 // Convert this build_vector into a pair of horizontal binop followed by
5494 bool isUndefLO = NumUndefsLO == Half;
5495 bool isUndefHI = NumUndefsHI == Half;
5496 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5497 isUndefLO, isUndefHI);
5501 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5502 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5504 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5505 X86Opcode = X86ISD::HADD;
5506 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5507 X86Opcode = X86ISD::HSUB;
5508 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5509 X86Opcode = X86ISD::FHADD;
5510 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5511 X86Opcode = X86ISD::FHSUB;
5515 // Don't try to expand this build_vector into a pair of horizontal add/sub
5516 // if we can simply emit a pair of scalar add/sub.
5517 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5520 // Convert this build_vector into two horizontal add/sub followed by
5522 bool isUndefLO = NumUndefsLO == Half;
5523 bool isUndefHI = NumUndefsHI == Half;
5524 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5525 isUndefLO, isUndefHI);
5532 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5535 MVT VT = Op.getSimpleValueType();
5536 MVT ExtVT = VT.getVectorElementType();
5537 unsigned NumElems = Op.getNumOperands();
5539 // Generate vectors for predicate vectors.
5540 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5541 return LowerBUILD_VECTORvXi1(Op, DAG);
5543 // Vectors containing all zeros can be matched by pxor and xorps later
5544 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5545 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5546 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5547 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5550 return getZeroVector(VT, Subtarget, DAG, dl);
5553 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5554 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5555 // vpcmpeqd on 256-bit vectors.
5556 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5557 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5560 if (!VT.is512BitVector())
5561 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5564 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
5565 if (Broadcast.getNode())
5568 unsigned EVTBits = ExtVT.getSizeInBits();
5570 unsigned NumZero = 0;
5571 unsigned NumNonZero = 0;
5572 unsigned NonZeros = 0;
5573 bool IsAllConstants = true;
5574 SmallSet<SDValue, 8> Values;
5575 for (unsigned i = 0; i < NumElems; ++i) {
5576 SDValue Elt = Op.getOperand(i);
5577 if (Elt.getOpcode() == ISD::UNDEF)
5580 if (Elt.getOpcode() != ISD::Constant &&
5581 Elt.getOpcode() != ISD::ConstantFP)
5582 IsAllConstants = false;
5583 if (X86::isZeroNode(Elt))
5586 NonZeros |= (1 << i);
5591 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5592 if (NumNonZero == 0)
5593 return DAG.getUNDEF(VT);
5595 // Special case for single non-zero, non-undef, element.
5596 if (NumNonZero == 1) {
5597 unsigned Idx = countTrailingZeros(NonZeros);
5598 SDValue Item = Op.getOperand(Idx);
5600 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5601 // the value are obviously zero, truncate the value to i32 and do the
5602 // insertion that way. Only do this if the value is non-constant or if the
5603 // value is a constant being inserted into element 0. It is cheaper to do
5604 // a constant pool load than it is to do a movd + shuffle.
5605 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5606 (!IsAllConstants || Idx == 0)) {
5607 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5609 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5610 EVT VecVT = MVT::v4i32;
5612 // Truncate the value (which may itself be a constant) to i32, and
5613 // convert it to a vector with movd (S2V+shuffle to zero extend).
5614 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5615 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5617 ISD::BITCAST, dl, VT,
5618 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5622 // If we have a constant or non-constant insertion into the low element of
5623 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5624 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5625 // depending on what the source datatype is.
5628 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5630 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5631 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5632 if (VT.is256BitVector() || VT.is512BitVector()) {
5633 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5634 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5635 Item, DAG.getIntPtrConstant(0));
5637 assert(VT.is128BitVector() && "Expected an SSE value type!");
5638 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5639 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5640 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5643 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5644 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5645 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5646 if (VT.is256BitVector()) {
5647 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5648 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5650 assert(VT.is128BitVector() && "Expected an SSE value type!");
5651 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5653 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5657 // Is it a vector logical left shift?
5658 if (NumElems == 2 && Idx == 1 &&
5659 X86::isZeroNode(Op.getOperand(0)) &&
5660 !X86::isZeroNode(Op.getOperand(1))) {
5661 unsigned NumBits = VT.getSizeInBits();
5662 return getVShift(true, VT,
5663 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5664 VT, Op.getOperand(1)),
5665 NumBits/2, DAG, *this, dl);
5668 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5671 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5672 // is a non-constant being inserted into an element other than the low one,
5673 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5674 // movd/movss) to move this into the low element, then shuffle it into
5676 if (EVTBits == 32) {
5677 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5678 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5682 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5683 if (Values.size() == 1) {
5684 if (EVTBits == 32) {
5685 // Instead of a shuffle like this:
5686 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5687 // Check if it's possible to issue this instead.
5688 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5689 unsigned Idx = countTrailingZeros(NonZeros);
5690 SDValue Item = Op.getOperand(Idx);
5691 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5692 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5697 // A vector full of immediates; various special cases are already
5698 // handled, so this is best done with a single constant-pool load.
5702 // For AVX-length vectors, see if we can use a vector load to get all of the
5703 // elements, otherwise build the individual 128-bit pieces and use
5704 // shuffles to put them in place.
5705 if (VT.is256BitVector() || VT.is512BitVector()) {
5706 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5708 // Check for a build vector of consecutive loads.
5709 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5712 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5714 // Build both the lower and upper subvector.
5715 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5716 makeArrayRef(&V[0], NumElems/2));
5717 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5718 makeArrayRef(&V[NumElems / 2], NumElems/2));
5720 // Recreate the wider vector with the lower and upper part.
5721 if (VT.is256BitVector())
5722 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5723 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5726 // Let legalizer expand 2-wide build_vectors.
5727 if (EVTBits == 64) {
5728 if (NumNonZero == 1) {
5729 // One half is zero or undef.
5730 unsigned Idx = countTrailingZeros(NonZeros);
5731 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5732 Op.getOperand(Idx));
5733 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5738 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5739 if (EVTBits == 8 && NumElems == 16) {
5740 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5742 if (V.getNode()) return V;
5745 if (EVTBits == 16 && NumElems == 8) {
5746 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5748 if (V.getNode()) return V;
5751 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5752 if (EVTBits == 32 && NumElems == 4) {
5753 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
5758 // If element VT is == 32 bits, turn it into a number of shuffles.
5759 SmallVector<SDValue, 8> V(NumElems);
5760 if (NumElems == 4 && NumZero > 0) {
5761 for (unsigned i = 0; i < 4; ++i) {
5762 bool isZero = !(NonZeros & (1 << i));
5764 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5766 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5769 for (unsigned i = 0; i < 2; ++i) {
5770 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5773 V[i] = V[i*2]; // Must be a zero vector.
5776 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5779 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5782 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5787 bool Reverse1 = (NonZeros & 0x3) == 2;
5788 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5792 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5793 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5795 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5798 if (Values.size() > 1 && VT.is128BitVector()) {
5799 // Check for a build vector of consecutive loads.
5800 for (unsigned i = 0; i < NumElems; ++i)
5801 V[i] = Op.getOperand(i);
5803 // Check for elements which are consecutive loads.
5804 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
5808 // Check for a build vector from mostly shuffle plus few inserting.
5809 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5813 // For SSE 4.1, use insertps to put the high elements into the low element.
5814 if (Subtarget->hasSSE41()) {
5816 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5817 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5819 Result = DAG.getUNDEF(VT);
5821 for (unsigned i = 1; i < NumElems; ++i) {
5822 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5823 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5824 Op.getOperand(i), DAG.getIntPtrConstant(i));
5829 // Otherwise, expand into a number of unpckl*, start by extending each of
5830 // our (non-undef) elements to the full vector width with the element in the
5831 // bottom slot of the vector (which generates no code for SSE).
5832 for (unsigned i = 0; i < NumElems; ++i) {
5833 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5834 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5836 V[i] = DAG.getUNDEF(VT);
5839 // Next, we iteratively mix elements, e.g. for v4f32:
5840 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5841 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5842 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5843 unsigned EltStride = NumElems >> 1;
5844 while (EltStride != 0) {
5845 for (unsigned i = 0; i < EltStride; ++i) {
5846 // If V[i+EltStride] is undef and this is the first round of mixing,
5847 // then it is safe to just drop this shuffle: V[i] is already in the
5848 // right place, the one element (since it's the first round) being
5849 // inserted as undef can be dropped. This isn't safe for successive
5850 // rounds because they will permute elements within both vectors.
5851 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5852 EltStride == NumElems/2)
5855 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5864 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5865 // to create 256-bit vectors from two other 128-bit ones.
5866 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5868 MVT ResVT = Op.getSimpleValueType();
5870 assert((ResVT.is256BitVector() ||
5871 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
5873 SDValue V1 = Op.getOperand(0);
5874 SDValue V2 = Op.getOperand(1);
5875 unsigned NumElems = ResVT.getVectorNumElements();
5876 if(ResVT.is256BitVector())
5877 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5879 if (Op.getNumOperands() == 4) {
5880 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
5881 ResVT.getVectorNumElements()/2);
5882 SDValue V3 = Op.getOperand(2);
5883 SDValue V4 = Op.getOperand(3);
5884 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
5885 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
5887 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5890 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5891 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
5892 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
5893 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
5894 Op.getNumOperands() == 4)));
5896 // AVX can use the vinsertf128 instruction to create 256-bit vectors
5897 // from two other 128-bit ones.
5899 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
5900 return LowerAVXCONCAT_VECTORS(Op, DAG);
5904 //===----------------------------------------------------------------------===//
5905 // Vector shuffle lowering
5907 // This is an experimental code path for lowering vector shuffles on x86. It is
5908 // designed to handle arbitrary vector shuffles and blends, gracefully
5909 // degrading performance as necessary. It works hard to recognize idiomatic
5910 // shuffles and lower them to optimal instruction patterns without leaving
5911 // a framework that allows reasonably efficient handling of all vector shuffle
5913 //===----------------------------------------------------------------------===//
5915 /// \brief Tiny helper function to identify a no-op mask.
5917 /// This is a somewhat boring predicate function. It checks whether the mask
5918 /// array input, which is assumed to be a single-input shuffle mask of the kind
5919 /// used by the X86 shuffle instructions (not a fully general
5920 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
5921 /// in-place shuffle are 'no-op's.
5922 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
5923 for (int i = 0, Size = Mask.size(); i < Size; ++i)
5924 if (Mask[i] != -1 && Mask[i] != i)
5929 /// \brief Helper function to classify a mask as a single-input mask.
5931 /// This isn't a generic single-input test because in the vector shuffle
5932 /// lowering we canonicalize single inputs to be the first input operand. This
5933 /// means we can more quickly test for a single input by only checking whether
5934 /// an input from the second operand exists. We also assume that the size of
5935 /// mask corresponds to the size of the input vectors which isn't true in the
5936 /// fully general case.
5937 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
5939 if (M >= (int)Mask.size())
5944 /// \brief Test whether there are elements crossing 128-bit lanes in this
5947 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
5948 /// and we routinely test for these.
5949 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
5950 int LaneSize = 128 / VT.getScalarSizeInBits();
5951 int Size = Mask.size();
5952 for (int i = 0; i < Size; ++i)
5953 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
5958 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
5960 /// This checks a shuffle mask to see if it is performing the same
5961 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
5962 /// that it is also not lane-crossing. It may however involve a blend from the
5963 /// same lane of a second vector.
5965 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
5966 /// non-trivial to compute in the face of undef lanes. The representation is
5967 /// *not* suitable for use with existing 128-bit shuffles as it will contain
5968 /// entries from both V1 and V2 inputs to the wider mask.
5970 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
5971 SmallVectorImpl<int> &RepeatedMask) {
5972 int LaneSize = 128 / VT.getScalarSizeInBits();
5973 RepeatedMask.resize(LaneSize, -1);
5974 int Size = Mask.size();
5975 for (int i = 0; i < Size; ++i) {
5978 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
5979 // This entry crosses lanes, so there is no way to model this shuffle.
5982 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
5983 if (RepeatedMask[i % LaneSize] == -1)
5984 // This is the first non-undef entry in this slot of a 128-bit lane.
5985 RepeatedMask[i % LaneSize] =
5986 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
5987 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
5988 // Found a mismatch with the repeated mask.
5994 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
5997 /// This is a fast way to test a shuffle mask against a fixed pattern:
5999 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6001 /// It returns true if the mask is exactly as wide as the argument list, and
6002 /// each element of the mask is either -1 (signifying undef) or the value given
6003 /// in the argument.
6004 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6005 ArrayRef<int> ExpectedMask) {
6006 if (Mask.size() != ExpectedMask.size())
6009 int Size = Mask.size();
6011 // If the values are build vectors, we can look through them to find
6012 // equivalent inputs that make the shuffles equivalent.
6013 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6014 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6016 for (int i = 0; i < Size; ++i)
6017 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6018 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6019 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6020 if (!MaskBV || !ExpectedBV ||
6021 MaskBV->getOperand(Mask[i] % Size) !=
6022 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6029 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6031 /// This helper function produces an 8-bit shuffle immediate corresponding to
6032 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6033 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6036 /// NB: We rely heavily on "undef" masks preserving the input lane.
6037 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
6038 SelectionDAG &DAG) {
6039 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6040 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6041 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6042 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6043 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6046 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6047 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6048 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6049 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6050 return DAG.getConstant(Imm, MVT::i8);
6053 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6055 /// This is used as a fallback approach when first class blend instructions are
6056 /// unavailable. Currently it is only suitable for integer vectors, but could
6057 /// be generalized for floating point vectors if desirable.
6058 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6059 SDValue V2, ArrayRef<int> Mask,
6060 SelectionDAG &DAG) {
6061 assert(VT.isInteger() && "Only supports integer vector types!");
6062 MVT EltVT = VT.getScalarType();
6063 int NumEltBits = EltVT.getSizeInBits();
6064 SDValue Zero = DAG.getConstant(0, EltVT);
6065 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
6066 SmallVector<SDValue, 16> MaskOps;
6067 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6068 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6069 return SDValue(); // Shuffled input!
6070 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6073 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6074 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6075 // We have to cast V2 around.
6076 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6077 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6078 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6079 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6080 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6081 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6084 /// \brief Try to emit a blend instruction for a shuffle.
6086 /// This doesn't do any checks for the availability of instructions for blending
6087 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6088 /// be matched in the backend with the type given. What it does check for is
6089 /// that the shuffle mask is in fact a blend.
6090 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6091 SDValue V2, ArrayRef<int> Mask,
6092 const X86Subtarget *Subtarget,
6093 SelectionDAG &DAG) {
6094 unsigned BlendMask = 0;
6095 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6096 if (Mask[i] >= Size) {
6097 if (Mask[i] != i + Size)
6098 return SDValue(); // Shuffled V2 input!
6099 BlendMask |= 1u << i;
6102 if (Mask[i] >= 0 && Mask[i] != i)
6103 return SDValue(); // Shuffled V1 input!
6105 switch (VT.SimpleTy) {
6110 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6111 DAG.getConstant(BlendMask, MVT::i8));
6115 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6119 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6120 // that instruction.
6121 if (Subtarget->hasAVX2()) {
6122 // Scale the blend by the number of 32-bit dwords per element.
6123 int Scale = VT.getScalarSizeInBits() / 32;
6125 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6126 if (Mask[i] >= Size)
6127 for (int j = 0; j < Scale; ++j)
6128 BlendMask |= 1u << (i * Scale + j);
6130 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6131 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6132 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6133 return DAG.getNode(ISD::BITCAST, DL, VT,
6134 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6135 DAG.getConstant(BlendMask, MVT::i8)));
6139 // For integer shuffles we need to expand the mask and cast the inputs to
6140 // v8i16s prior to blending.
6141 int Scale = 8 / VT.getVectorNumElements();
6143 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6144 if (Mask[i] >= Size)
6145 for (int j = 0; j < Scale; ++j)
6146 BlendMask |= 1u << (i * Scale + j);
6148 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6149 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6150 return DAG.getNode(ISD::BITCAST, DL, VT,
6151 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6152 DAG.getConstant(BlendMask, MVT::i8)));
6156 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6157 SmallVector<int, 8> RepeatedMask;
6158 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6159 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6160 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6162 for (int i = 0; i < 8; ++i)
6163 if (RepeatedMask[i] >= 16)
6164 BlendMask |= 1u << i;
6165 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6166 DAG.getConstant(BlendMask, MVT::i8));
6172 // Scale the blend by the number of bytes per element.
6173 int Scale = VT.getScalarSizeInBits() / 8;
6175 // This form of blend is always done on bytes. Compute the byte vector
6177 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6179 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6180 // mix of LLVM's code generator and the x86 backend. We tell the code
6181 // generator that boolean values in the elements of an x86 vector register
6182 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6183 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6184 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6185 // of the element (the remaining are ignored) and 0 in that high bit would
6186 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6187 // the LLVM model for boolean values in vector elements gets the relevant
6188 // bit set, it is set backwards and over constrained relative to x86's
6190 SmallVector<SDValue, 32> VSELECTMask;
6191 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6192 for (int j = 0; j < Scale; ++j)
6193 VSELECTMask.push_back(
6194 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6195 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
6197 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6198 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6200 ISD::BITCAST, DL, VT,
6201 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6202 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6207 llvm_unreachable("Not a supported integer vector type!");
6211 /// \brief Try to lower as a blend of elements from two inputs followed by
6212 /// a single-input permutation.
6214 /// This matches the pattern where we can blend elements from two inputs and
6215 /// then reduce the shuffle to a single-input permutation.
6216 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6219 SelectionDAG &DAG) {
6220 // We build up the blend mask while checking whether a blend is a viable way
6221 // to reduce the shuffle.
6222 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6223 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6225 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6229 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6231 if (BlendMask[Mask[i] % Size] == -1)
6232 BlendMask[Mask[i] % Size] = Mask[i];
6233 else if (BlendMask[Mask[i] % Size] != Mask[i])
6234 return SDValue(); // Can't blend in the needed input!
6236 PermuteMask[i] = Mask[i] % Size;
6239 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6240 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6243 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6244 /// blends and permutes.
6246 /// This matches the extremely common pattern for handling combined
6247 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6248 /// operations. It will try to pick the best arrangement of shuffles and
6250 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6254 SelectionDAG &DAG) {
6255 // Shuffle the input elements into the desired positions in V1 and V2 and
6256 // blend them together.
6257 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6258 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6259 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6260 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6261 if (Mask[i] >= 0 && Mask[i] < Size) {
6262 V1Mask[i] = Mask[i];
6264 } else if (Mask[i] >= Size) {
6265 V2Mask[i] = Mask[i] - Size;
6266 BlendMask[i] = i + Size;
6269 // Try to lower with the simpler initial blend strategy unless one of the
6270 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6271 // shuffle may be able to fold with a load or other benefit. However, when
6272 // we'll have to do 2x as many shuffles in order to achieve this, blending
6273 // first is a better strategy.
6274 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6275 if (SDValue BlendPerm =
6276 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6279 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6280 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6281 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6284 /// \brief Try to lower a vector shuffle as a byte rotation.
6286 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6287 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6288 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6289 /// try to generically lower a vector shuffle through such an pattern. It
6290 /// does not check for the profitability of lowering either as PALIGNR or
6291 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6292 /// This matches shuffle vectors that look like:
6294 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6296 /// Essentially it concatenates V1 and V2, shifts right by some number of
6297 /// elements, and takes the low elements as the result. Note that while this is
6298 /// specified as a *right shift* because x86 is little-endian, it is a *left
6299 /// rotate* of the vector lanes.
6300 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6303 const X86Subtarget *Subtarget,
6304 SelectionDAG &DAG) {
6305 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6307 int NumElts = Mask.size();
6308 int NumLanes = VT.getSizeInBits() / 128;
6309 int NumLaneElts = NumElts / NumLanes;
6311 // We need to detect various ways of spelling a rotation:
6312 // [11, 12, 13, 14, 15, 0, 1, 2]
6313 // [-1, 12, 13, 14, -1, -1, 1, -1]
6314 // [-1, -1, -1, -1, -1, -1, 1, 2]
6315 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6316 // [-1, 4, 5, 6, -1, -1, 9, -1]
6317 // [-1, 4, 5, 6, -1, -1, -1, -1]
6320 for (int l = 0; l < NumElts; l += NumLaneElts) {
6321 for (int i = 0; i < NumLaneElts; ++i) {
6322 if (Mask[l + i] == -1)
6324 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6326 // Get the mod-Size index and lane correct it.
6327 int LaneIdx = (Mask[l + i] % NumElts) - l;
6328 // Make sure it was in this lane.
6329 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6332 // Determine where a rotated vector would have started.
6333 int StartIdx = i - LaneIdx;
6335 // The identity rotation isn't interesting, stop.
6338 // If we found the tail of a vector the rotation must be the missing
6339 // front. If we found the head of a vector, it must be how much of the
6341 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6344 Rotation = CandidateRotation;
6345 else if (Rotation != CandidateRotation)
6346 // The rotations don't match, so we can't match this mask.
6349 // Compute which value this mask is pointing at.
6350 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6352 // Compute which of the two target values this index should be assigned
6353 // to. This reflects whether the high elements are remaining or the low
6354 // elements are remaining.
6355 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6357 // Either set up this value if we've not encountered it before, or check
6358 // that it remains consistent.
6361 else if (TargetV != MaskV)
6362 // This may be a rotation, but it pulls from the inputs in some
6363 // unsupported interleaving.
6368 // Check that we successfully analyzed the mask, and normalize the results.
6369 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6370 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6376 // The actual rotate instruction rotates bytes, so we need to scale the
6377 // rotation based on how many bytes are in the vector lane.
6378 int Scale = 16 / NumLaneElts;
6380 // SSSE3 targets can use the palignr instruction.
6381 if (Subtarget->hasSSSE3()) {
6382 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6383 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6384 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6385 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6387 return DAG.getNode(ISD::BITCAST, DL, VT,
6388 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6389 DAG.getConstant(Rotation * Scale, MVT::i8)));
6392 assert(VT.getSizeInBits() == 128 &&
6393 "Rotate-based lowering only supports 128-bit lowering!");
6394 assert(Mask.size() <= 16 &&
6395 "Can shuffle at most 16 bytes in a 128-bit vector!");
6397 // Default SSE2 implementation
6398 int LoByteShift = 16 - Rotation * Scale;
6399 int HiByteShift = Rotation * Scale;
6401 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6402 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6403 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6405 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6406 DAG.getConstant(LoByteShift, MVT::i8));
6407 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6408 DAG.getConstant(HiByteShift, MVT::i8));
6409 return DAG.getNode(ISD::BITCAST, DL, VT,
6410 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6413 /// \brief Compute whether each element of a shuffle is zeroable.
6415 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6416 /// Either it is an undef element in the shuffle mask, the element of the input
6417 /// referenced is undef, or the element of the input referenced is known to be
6418 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6419 /// as many lanes with this technique as possible to simplify the remaining
6421 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6422 SDValue V1, SDValue V2) {
6423 SmallBitVector Zeroable(Mask.size(), false);
6425 while (V1.getOpcode() == ISD::BITCAST)
6426 V1 = V1->getOperand(0);
6427 while (V2.getOpcode() == ISD::BITCAST)
6428 V2 = V2->getOperand(0);
6430 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6431 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6433 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6435 // Handle the easy cases.
6436 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6441 // If this is an index into a build_vector node (which has the same number
6442 // of elements), dig out the input value and use it.
6443 SDValue V = M < Size ? V1 : V2;
6444 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6447 SDValue Input = V.getOperand(M % Size);
6448 // The UNDEF opcode check really should be dead code here, but not quite
6449 // worth asserting on (it isn't invalid, just unexpected).
6450 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6457 /// \brief Try to emit a bitmask instruction for a shuffle.
6459 /// This handles cases where we can model a blend exactly as a bitmask due to
6460 /// one of the inputs being zeroable.
6461 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6462 SDValue V2, ArrayRef<int> Mask,
6463 SelectionDAG &DAG) {
6464 MVT EltVT = VT.getScalarType();
6465 int NumEltBits = EltVT.getSizeInBits();
6466 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6467 SDValue Zero = DAG.getConstant(0, IntEltVT);
6468 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
6469 if (EltVT.isFloatingPoint()) {
6470 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6471 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6473 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6474 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6476 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6479 if (Mask[i] % Size != i)
6480 return SDValue(); // Not a blend.
6482 V = Mask[i] < Size ? V1 : V2;
6483 else if (V != (Mask[i] < Size ? V1 : V2))
6484 return SDValue(); // Can only let one input through the mask.
6486 VMaskOps[i] = AllOnes;
6489 return SDValue(); // No non-zeroable elements!
6491 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6492 V = DAG.getNode(VT.isFloatingPoint()
6493 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6498 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6500 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6501 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6502 /// matches elements from one of the input vectors shuffled to the left or
6503 /// right with zeroable elements 'shifted in'. It handles both the strictly
6504 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6507 /// PSHL : (little-endian) left bit shift.
6508 /// [ zz, 0, zz, 2 ]
6509 /// [ -1, 4, zz, -1 ]
6510 /// PSRL : (little-endian) right bit shift.
6512 /// [ -1, -1, 7, zz]
6513 /// PSLLDQ : (little-endian) left byte shift
6514 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6515 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6516 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6517 /// PSRLDQ : (little-endian) right byte shift
6518 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6519 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6520 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6521 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6522 SDValue V2, ArrayRef<int> Mask,
6523 SelectionDAG &DAG) {
6524 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6526 int Size = Mask.size();
6527 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6529 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6530 for (int i = 0; i < Size; i += Scale)
6531 for (int j = 0; j < Shift; ++j)
6532 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6538 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6539 for (int i = 0; i != Size; i += Scale) {
6540 unsigned Pos = Left ? i + Shift : i;
6541 unsigned Low = Left ? i : i + Shift;
6542 unsigned Len = Scale - Shift;
6543 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6544 Low + (V == V1 ? 0 : Size)))
6548 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6549 bool ByteShift = ShiftEltBits > 64;
6550 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6551 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6552 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6554 // Normalize the scale for byte shifts to still produce an i64 element
6556 Scale = ByteShift ? Scale / 2 : Scale;
6558 // We need to round trip through the appropriate type for the shift.
6559 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6560 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6561 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6562 "Illegal integer vector type");
6563 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6565 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
6566 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6569 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6570 // keep doubling the size of the integer elements up to that. We can
6571 // then shift the elements of the integer vector by whole multiples of
6572 // their width within the elements of the larger integer vector. Test each
6573 // multiple to see if we can find a match with the moved element indices
6574 // and that the shifted in elements are all zeroable.
6575 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6576 for (int Shift = 1; Shift != Scale; ++Shift)
6577 for (bool Left : {true, false})
6578 if (CheckZeros(Shift, Scale, Left))
6579 for (SDValue V : {V1, V2})
6580 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6587 /// \brief Lower a vector shuffle as a zero or any extension.
6589 /// Given a specific number of elements, element bit width, and extension
6590 /// stride, produce either a zero or any extension based on the available
6591 /// features of the subtarget.
6592 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6593 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6594 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6595 assert(Scale > 1 && "Need a scale to extend.");
6596 int NumElements = VT.getVectorNumElements();
6597 int EltBits = VT.getScalarSizeInBits();
6598 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6599 "Only 8, 16, and 32 bit elements can be extended.");
6600 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6602 // Found a valid zext mask! Try various lowering strategies based on the
6603 // input type and available ISA extensions.
6604 if (Subtarget->hasSSE41()) {
6605 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6606 NumElements / Scale);
6607 return DAG.getNode(ISD::BITCAST, DL, VT,
6608 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6611 // For any extends we can cheat for larger element sizes and use shuffle
6612 // instructions that can fold with a load and/or copy.
6613 if (AnyExt && EltBits == 32) {
6614 int PSHUFDMask[4] = {0, -1, 1, -1};
6616 ISD::BITCAST, DL, VT,
6617 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6618 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6619 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
6621 if (AnyExt && EltBits == 16 && Scale > 2) {
6622 int PSHUFDMask[4] = {0, -1, 0, -1};
6623 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6624 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6625 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
6626 int PSHUFHWMask[4] = {1, -1, -1, -1};
6628 ISD::BITCAST, DL, VT,
6629 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6630 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6631 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
6634 // If this would require more than 2 unpack instructions to expand, use
6635 // pshufb when available. We can only use more than 2 unpack instructions
6636 // when zero extending i8 elements which also makes it easier to use pshufb.
6637 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6638 assert(NumElements == 16 && "Unexpected byte vector width!");
6639 SDValue PSHUFBMask[16];
6640 for (int i = 0; i < 16; ++i)
6642 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
6643 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6644 return DAG.getNode(ISD::BITCAST, DL, VT,
6645 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6646 DAG.getNode(ISD::BUILD_VECTOR, DL,
6647 MVT::v16i8, PSHUFBMask)));
6650 // Otherwise emit a sequence of unpacks.
6652 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6653 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6654 : getZeroVector(InputVT, Subtarget, DAG, DL);
6655 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6656 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6660 } while (Scale > 1);
6661 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6664 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6666 /// This routine will try to do everything in its power to cleverly lower
6667 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6668 /// check for the profitability of this lowering, it tries to aggressively
6669 /// match this pattern. It will use all of the micro-architectural details it
6670 /// can to emit an efficient lowering. It handles both blends with all-zero
6671 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6672 /// masking out later).
6674 /// The reason we have dedicated lowering for zext-style shuffles is that they
6675 /// are both incredibly common and often quite performance sensitive.
6676 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6677 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6678 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6679 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6681 int Bits = VT.getSizeInBits();
6682 int NumElements = VT.getVectorNumElements();
6683 assert(VT.getScalarSizeInBits() <= 32 &&
6684 "Exceeds 32-bit integer zero extension limit");
6685 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6687 // Define a helper function to check a particular ext-scale and lower to it if
6689 auto Lower = [&](int Scale) -> SDValue {
6692 for (int i = 0; i < NumElements; ++i) {
6694 continue; // Valid anywhere but doesn't tell us anything.
6695 if (i % Scale != 0) {
6696 // Each of the extended elements need to be zeroable.
6700 // We no longer are in the anyext case.
6705 // Each of the base elements needs to be consecutive indices into the
6706 // same input vector.
6707 SDValue V = Mask[i] < NumElements ? V1 : V2;
6710 else if (InputV != V)
6711 return SDValue(); // Flip-flopping inputs.
6713 if (Mask[i] % NumElements != i / Scale)
6714 return SDValue(); // Non-consecutive strided elements.
6717 // If we fail to find an input, we have a zero-shuffle which should always
6718 // have already been handled.
6719 // FIXME: Maybe handle this here in case during blending we end up with one?
6723 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6724 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6727 // The widest scale possible for extending is to a 64-bit integer.
6728 assert(Bits % 64 == 0 &&
6729 "The number of bits in a vector must be divisible by 64 on x86!");
6730 int NumExtElements = Bits / 64;
6732 // Each iteration, try extending the elements half as much, but into twice as
6734 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6735 assert(NumElements % NumExtElements == 0 &&
6736 "The input vector size must be divisible by the extended size.");
6737 if (SDValue V = Lower(NumElements / NumExtElements))
6741 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6745 // Returns one of the source operands if the shuffle can be reduced to a
6746 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6747 auto CanZExtLowHalf = [&]() {
6748 for (int i = NumElements / 2; i != NumElements; ++i)
6751 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6753 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
6758 if (SDValue V = CanZExtLowHalf()) {
6759 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
6760 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
6761 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6764 // No viable ext lowering found.
6768 /// \brief Try to get a scalar value for a specific element of a vector.
6770 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
6771 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
6772 SelectionDAG &DAG) {
6773 MVT VT = V.getSimpleValueType();
6774 MVT EltVT = VT.getVectorElementType();
6775 while (V.getOpcode() == ISD::BITCAST)
6776 V = V.getOperand(0);
6777 // If the bitcasts shift the element size, we can't extract an equivalent
6779 MVT NewVT = V.getSimpleValueType();
6780 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
6783 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6784 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
6785 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
6790 /// \brief Helper to test for a load that can be folded with x86 shuffles.
6792 /// This is particularly important because the set of instructions varies
6793 /// significantly based on whether the operand is a load or not.
6794 static bool isShuffleFoldableLoad(SDValue V) {
6795 while (V.getOpcode() == ISD::BITCAST)
6796 V = V.getOperand(0);
6798 return ISD::isNON_EXTLoad(V.getNode());
6801 /// \brief Try to lower insertion of a single element into a zero vector.
6803 /// This is a common pattern that we have especially efficient patterns to lower
6804 /// across all subtarget feature sets.
6805 static SDValue lowerVectorShuffleAsElementInsertion(
6806 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6807 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6808 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6810 MVT EltVT = VT.getVectorElementType();
6812 int V2Index = std::find_if(Mask.begin(), Mask.end(),
6813 [&Mask](int M) { return M >= (int)Mask.size(); }) -
6815 bool IsV1Zeroable = true;
6816 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6817 if (i != V2Index && !Zeroable[i]) {
6818 IsV1Zeroable = false;
6822 // Check for a single input from a SCALAR_TO_VECTOR node.
6823 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
6824 // all the smarts here sunk into that routine. However, the current
6825 // lowering of BUILD_VECTOR makes that nearly impossible until the old
6826 // vector shuffle lowering is dead.
6827 if (SDValue V2S = getScalarValueForVectorElement(
6828 V2, Mask[V2Index] - Mask.size(), DAG)) {
6829 // We need to zext the scalar if it is smaller than an i32.
6830 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
6831 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
6832 // Using zext to expand a narrow element won't work for non-zero
6837 // Zero-extend directly to i32.
6839 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
6841 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
6842 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
6843 EltVT == MVT::i16) {
6844 // Either not inserting from the low element of the input or the input
6845 // element size is too small to use VZEXT_MOVL to clear the high bits.
6849 if (!IsV1Zeroable) {
6850 // If V1 can't be treated as a zero vector we have fewer options to lower
6851 // this. We can't support integer vectors or non-zero targets cheaply, and
6852 // the V1 elements can't be permuted in any way.
6853 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
6854 if (!VT.isFloatingPoint() || V2Index != 0)
6856 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
6857 V1Mask[V2Index] = -1;
6858 if (!isNoopShuffleMask(V1Mask))
6860 // This is essentially a special case blend operation, but if we have
6861 // general purpose blend operations, they are always faster. Bail and let
6862 // the rest of the lowering handle these as blends.
6863 if (Subtarget->hasSSE41())
6866 // Otherwise, use MOVSD or MOVSS.
6867 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
6868 "Only two types of floating point element types to handle!");
6869 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
6873 // This lowering only works for the low element with floating point vectors.
6874 if (VT.isFloatingPoint() && V2Index != 0)
6877 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
6879 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6882 // If we have 4 or fewer lanes we can cheaply shuffle the element into
6883 // the desired position. Otherwise it is more efficient to do a vector
6884 // shift left. We know that we can do a vector shift left because all
6885 // the inputs are zero.
6886 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
6887 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
6888 V2Shuffle[V2Index] = 0;
6889 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
6891 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
6893 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
6895 V2Index * EltVT.getSizeInBits()/8,
6896 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
6897 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6903 /// \brief Try to lower broadcast of a single element.
6905 /// For convenience, this code also bundles all of the subtarget feature set
6906 /// filtering. While a little annoying to re-dispatch on type here, there isn't
6907 /// a convenient way to factor it out.
6908 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
6910 const X86Subtarget *Subtarget,
6911 SelectionDAG &DAG) {
6912 if (!Subtarget->hasAVX())
6914 if (VT.isInteger() && !Subtarget->hasAVX2())
6917 // Check that the mask is a broadcast.
6918 int BroadcastIdx = -1;
6920 if (M >= 0 && BroadcastIdx == -1)
6922 else if (M >= 0 && M != BroadcastIdx)
6925 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
6926 "a sorted mask where the broadcast "
6929 // Go up the chain of (vector) values to try and find a scalar load that
6930 // we can combine with the broadcast.
6932 switch (V.getOpcode()) {
6933 case ISD::CONCAT_VECTORS: {
6934 int OperandSize = Mask.size() / V.getNumOperands();
6935 V = V.getOperand(BroadcastIdx / OperandSize);
6936 BroadcastIdx %= OperandSize;
6940 case ISD::INSERT_SUBVECTOR: {
6941 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
6942 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
6946 int BeginIdx = (int)ConstantIdx->getZExtValue();
6948 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
6949 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
6950 BroadcastIdx -= BeginIdx;
6961 // Check if this is a broadcast of a scalar. We special case lowering
6962 // for scalars so that we can more effectively fold with loads.
6963 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6964 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
6965 V = V.getOperand(BroadcastIdx);
6967 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
6969 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
6971 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
6972 // We can't broadcast from a vector register w/o AVX2, and we can only
6973 // broadcast from the zero-element of a vector register.
6977 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
6980 // Check for whether we can use INSERTPS to perform the shuffle. We only use
6981 // INSERTPS when the V1 elements are already in the correct locations
6982 // because otherwise we can just always use two SHUFPS instructions which
6983 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
6984 // perform INSERTPS if a single V1 element is out of place and all V2
6985 // elements are zeroable.
6986 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
6988 SelectionDAG &DAG) {
6989 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
6990 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
6991 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
6992 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
6994 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6997 int V1DstIndex = -1;
6998 int V2DstIndex = -1;
6999 bool V1UsedInPlace = false;
7001 for (int i = 0; i < 4; ++i) {
7002 // Synthesize a zero mask from the zeroable elements (includes undefs).
7008 // Flag if we use any V1 inputs in place.
7010 V1UsedInPlace = true;
7014 // We can only insert a single non-zeroable element.
7015 if (V1DstIndex != -1 || V2DstIndex != -1)
7019 // V1 input out of place for insertion.
7022 // V2 input for insertion.
7027 // Don't bother if we have no (non-zeroable) element for insertion.
7028 if (V1DstIndex == -1 && V2DstIndex == -1)
7031 // Determine element insertion src/dst indices. The src index is from the
7032 // start of the inserted vector, not the start of the concatenated vector.
7033 unsigned V2SrcIndex = 0;
7034 if (V1DstIndex != -1) {
7035 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7036 // and don't use the original V2 at all.
7037 V2SrcIndex = Mask[V1DstIndex];
7038 V2DstIndex = V1DstIndex;
7041 V2SrcIndex = Mask[V2DstIndex] - 4;
7044 // If no V1 inputs are used in place, then the result is created only from
7045 // the zero mask and the V2 insertion - so remove V1 dependency.
7047 V1 = DAG.getUNDEF(MVT::v4f32);
7049 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7050 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7052 // Insert the V2 element into the desired position.
7054 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7055 DAG.getConstant(InsertPSMask, MVT::i8));
7058 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7059 /// UNPCK instruction.
7061 /// This specifically targets cases where we end up with alternating between
7062 /// the two inputs, and so can permute them into something that feeds a single
7063 /// UNPCK instruction. Note that this routine only targets integer vectors
7064 /// because for floating point vectors we have a generalized SHUFPS lowering
7065 /// strategy that handles everything that doesn't *exactly* match an unpack,
7066 /// making this clever lowering unnecessary.
7067 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7068 SDValue V2, ArrayRef<int> Mask,
7069 SelectionDAG &DAG) {
7070 assert(!VT.isFloatingPoint() &&
7071 "This routine only supports integer vectors.");
7072 assert(!isSingleInputShuffleMask(Mask) &&
7073 "This routine should only be used when blending two inputs.");
7074 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7076 int Size = Mask.size();
7078 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7079 return M >= 0 && M % Size < Size / 2;
7081 int NumHiInputs = std::count_if(
7082 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7084 bool UnpackLo = NumLoInputs >= NumHiInputs;
7086 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7087 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7088 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7090 for (int i = 0; i < Size; ++i) {
7094 // Each element of the unpack contains Scale elements from this mask.
7095 int UnpackIdx = i / Scale;
7097 // We only handle the case where V1 feeds the first slots of the unpack.
7098 // We rely on canonicalization to ensure this is the case.
7099 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7102 // Setup the mask for this input. The indexing is tricky as we have to
7103 // handle the unpack stride.
7104 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7105 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7109 // If we will have to shuffle both inputs to use the unpack, check whether
7110 // we can just unpack first and shuffle the result. If so, skip this unpack.
7111 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7112 !isNoopShuffleMask(V2Mask))
7115 // Shuffle the inputs into place.
7116 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7117 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7119 // Cast the inputs to the type we will use to unpack them.
7120 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7121 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7123 // Unpack the inputs and cast the result back to the desired type.
7124 return DAG.getNode(ISD::BITCAST, DL, VT,
7125 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7126 DL, UnpackVT, V1, V2));
7129 // We try each unpack from the largest to the smallest to try and find one
7130 // that fits this mask.
7131 int OrigNumElements = VT.getVectorNumElements();
7132 int OrigScalarSize = VT.getScalarSizeInBits();
7133 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7134 int Scale = ScalarSize / OrigScalarSize;
7135 int NumElements = OrigNumElements / Scale;
7136 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7137 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7141 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7143 if (NumLoInputs == 0 || NumHiInputs == 0) {
7144 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7145 "We have to have *some* inputs!");
7146 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7148 // FIXME: We could consider the total complexity of the permute of each
7149 // possible unpacking. Or at the least we should consider how many
7150 // half-crossings are created.
7151 // FIXME: We could consider commuting the unpacks.
7153 SmallVector<int, 32> PermMask;
7154 PermMask.assign(Size, -1);
7155 for (int i = 0; i < Size; ++i) {
7159 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7162 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7164 return DAG.getVectorShuffle(
7165 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7167 DAG.getUNDEF(VT), PermMask);
7173 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7175 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7176 /// support for floating point shuffles but not integer shuffles. These
7177 /// instructions will incur a domain crossing penalty on some chips though so
7178 /// it is better to avoid lowering through this for integer vectors where
7180 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7181 const X86Subtarget *Subtarget,
7182 SelectionDAG &DAG) {
7184 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7185 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7186 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7187 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7188 ArrayRef<int> Mask = SVOp->getMask();
7189 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7191 if (isSingleInputShuffleMask(Mask)) {
7192 // Use low duplicate instructions for masks that match their pattern.
7193 if (Subtarget->hasSSE3())
7194 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7195 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7197 // Straight shuffle of a single input vector. Simulate this by using the
7198 // single input as both of the "inputs" to this instruction..
7199 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7201 if (Subtarget->hasAVX()) {
7202 // If we have AVX, we can use VPERMILPS which will allow folding a load
7203 // into the shuffle.
7204 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7205 DAG.getConstant(SHUFPDMask, MVT::i8));
7208 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
7209 DAG.getConstant(SHUFPDMask, MVT::i8));
7211 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7212 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7214 // If we have a single input, insert that into V1 if we can do so cheaply.
7215 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7216 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7217 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7219 // Try inverting the insertion since for v2 masks it is easy to do and we
7220 // can't reliably sort the mask one way or the other.
7221 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7222 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7223 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7224 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7228 // Try to use one of the special instruction patterns to handle two common
7229 // blend patterns if a zero-blend above didn't work.
7230 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7231 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7232 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7233 // We can either use a special instruction to load over the low double or
7234 // to move just the low double.
7236 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7238 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7240 if (Subtarget->hasSSE41())
7241 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7245 // Use dedicated unpack instructions for masks that match their pattern.
7246 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7247 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7248 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7249 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7251 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7252 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
7253 DAG.getConstant(SHUFPDMask, MVT::i8));
7256 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7258 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7259 /// the integer unit to minimize domain crossing penalties. However, for blends
7260 /// it falls back to the floating point shuffle operation with appropriate bit
7262 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7263 const X86Subtarget *Subtarget,
7264 SelectionDAG &DAG) {
7266 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7267 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7268 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7269 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7270 ArrayRef<int> Mask = SVOp->getMask();
7271 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7273 if (isSingleInputShuffleMask(Mask)) {
7274 // Check for being able to broadcast a single element.
7275 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7276 Mask, Subtarget, DAG))
7279 // Straight shuffle of a single input vector. For everything from SSE2
7280 // onward this has a single fast instruction with no scary immediates.
7281 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7282 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7283 int WidenedMask[4] = {
7284 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7285 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7287 ISD::BITCAST, DL, MVT::v2i64,
7288 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7289 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7291 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7292 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7293 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7294 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7296 // If we have a blend of two PACKUS operations an the blend aligns with the
7297 // low and half halves, we can just merge the PACKUS operations. This is
7298 // particularly important as it lets us merge shuffles that this routine itself
7300 auto GetPackNode = [](SDValue V) {
7301 while (V.getOpcode() == ISD::BITCAST)
7302 V = V.getOperand(0);
7304 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7306 if (SDValue V1Pack = GetPackNode(V1))
7307 if (SDValue V2Pack = GetPackNode(V2))
7308 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7309 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7310 Mask[0] == 0 ? V1Pack.getOperand(0)
7311 : V1Pack.getOperand(1),
7312 Mask[1] == 2 ? V2Pack.getOperand(0)
7313 : V2Pack.getOperand(1)));
7315 // Try to use shift instructions.
7317 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7320 // When loading a scalar and then shuffling it into a vector we can often do
7321 // the insertion cheaply.
7322 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7323 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7325 // Try inverting the insertion since for v2 masks it is easy to do and we
7326 // can't reliably sort the mask one way or the other.
7327 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7328 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7329 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7332 // We have different paths for blend lowering, but they all must use the
7333 // *exact* same predicate.
7334 bool IsBlendSupported = Subtarget->hasSSE41();
7335 if (IsBlendSupported)
7336 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7340 // Use dedicated unpack instructions for masks that match their pattern.
7341 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7342 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7343 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7344 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7346 // Try to use byte rotation instructions.
7347 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7348 if (Subtarget->hasSSSE3())
7349 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7350 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7353 // If we have direct support for blends, we should lower by decomposing into
7354 // a permute. That will be faster than the domain cross.
7355 if (IsBlendSupported)
7356 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7359 // We implement this with SHUFPD which is pretty lame because it will likely
7360 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7361 // However, all the alternatives are still more cycles and newer chips don't
7362 // have this problem. It would be really nice if x86 had better shuffles here.
7363 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7364 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7365 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7366 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7369 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7371 /// This is used to disable more specialized lowerings when the shufps lowering
7372 /// will happen to be efficient.
7373 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7374 // This routine only handles 128-bit shufps.
7375 assert(Mask.size() == 4 && "Unsupported mask size!");
7377 // To lower with a single SHUFPS we need to have the low half and high half
7378 // each requiring a single input.
7379 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7381 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7387 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7389 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7390 /// It makes no assumptions about whether this is the *best* lowering, it simply
7392 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7393 ArrayRef<int> Mask, SDValue V1,
7394 SDValue V2, SelectionDAG &DAG) {
7395 SDValue LowV = V1, HighV = V2;
7396 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7399 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7401 if (NumV2Elements == 1) {
7403 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7406 // Compute the index adjacent to V2Index and in the same half by toggling
7408 int V2AdjIndex = V2Index ^ 1;
7410 if (Mask[V2AdjIndex] == -1) {
7411 // Handles all the cases where we have a single V2 element and an undef.
7412 // This will only ever happen in the high lanes because we commute the
7413 // vector otherwise.
7415 std::swap(LowV, HighV);
7416 NewMask[V2Index] -= 4;
7418 // Handle the case where the V2 element ends up adjacent to a V1 element.
7419 // To make this work, blend them together as the first step.
7420 int V1Index = V2AdjIndex;
7421 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7422 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7423 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7425 // Now proceed to reconstruct the final blend as we have the necessary
7426 // high or low half formed.
7433 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7434 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7436 } else if (NumV2Elements == 2) {
7437 if (Mask[0] < 4 && Mask[1] < 4) {
7438 // Handle the easy case where we have V1 in the low lanes and V2 in the
7442 } else if (Mask[2] < 4 && Mask[3] < 4) {
7443 // We also handle the reversed case because this utility may get called
7444 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7445 // arrange things in the right direction.
7451 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7452 // trying to place elements directly, just blend them and set up the final
7453 // shuffle to place them.
7455 // The first two blend mask elements are for V1, the second two are for
7457 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7458 Mask[2] < 4 ? Mask[2] : Mask[3],
7459 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7460 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7461 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7462 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7464 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7467 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7468 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7469 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7470 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7473 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7474 getV4X86ShuffleImm8ForMask(NewMask, DAG));
7477 /// \brief Lower 4-lane 32-bit floating point shuffles.
7479 /// Uses instructions exclusively from the floating point unit to minimize
7480 /// domain crossing penalties, as these are sufficient to implement all v4f32
7482 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7483 const X86Subtarget *Subtarget,
7484 SelectionDAG &DAG) {
7486 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7487 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7488 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7489 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7490 ArrayRef<int> Mask = SVOp->getMask();
7491 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7494 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7496 if (NumV2Elements == 0) {
7497 // Check for being able to broadcast a single element.
7498 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7499 Mask, Subtarget, DAG))
7502 // Use even/odd duplicate instructions for masks that match their pattern.
7503 if (Subtarget->hasSSE3()) {
7504 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7505 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7506 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7507 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7510 if (Subtarget->hasAVX()) {
7511 // If we have AVX, we can use VPERMILPS which will allow folding a load
7512 // into the shuffle.
7513 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7514 getV4X86ShuffleImm8ForMask(Mask, DAG));
7517 // Otherwise, use a straight shuffle of a single input vector. We pass the
7518 // input vector to both operands to simulate this with a SHUFPS.
7519 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7520 getV4X86ShuffleImm8ForMask(Mask, DAG));
7523 // There are special ways we can lower some single-element blends. However, we
7524 // have custom ways we can lower more complex single-element blends below that
7525 // we defer to if both this and BLENDPS fail to match, so restrict this to
7526 // when the V2 input is targeting element 0 of the mask -- that is the fast
7528 if (NumV2Elements == 1 && Mask[0] >= 4)
7529 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7530 Mask, Subtarget, DAG))
7533 if (Subtarget->hasSSE41()) {
7534 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7538 // Use INSERTPS if we can complete the shuffle efficiently.
7539 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7542 if (!isSingleSHUFPSMask(Mask))
7543 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7544 DL, MVT::v4f32, V1, V2, Mask, DAG))
7548 // Use dedicated unpack instructions for masks that match their pattern.
7549 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7550 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7551 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7552 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7553 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7554 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7555 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7556 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7558 // Otherwise fall back to a SHUFPS lowering strategy.
7559 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7562 /// \brief Lower 4-lane i32 vector shuffles.
7564 /// We try to handle these with integer-domain shuffles where we can, but for
7565 /// blends we use the floating point domain blend instructions.
7566 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7567 const X86Subtarget *Subtarget,
7568 SelectionDAG &DAG) {
7570 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7571 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7572 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7573 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7574 ArrayRef<int> Mask = SVOp->getMask();
7575 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7577 // Whenever we can lower this as a zext, that instruction is strictly faster
7578 // than any alternative. It also allows us to fold memory operands into the
7579 // shuffle in many cases.
7580 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7581 Mask, Subtarget, DAG))
7585 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7587 if (NumV2Elements == 0) {
7588 // Check for being able to broadcast a single element.
7589 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7590 Mask, Subtarget, DAG))
7593 // Straight shuffle of a single input vector. For everything from SSE2
7594 // onward this has a single fast instruction with no scary immediates.
7595 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7596 // but we aren't actually going to use the UNPCK instruction because doing
7597 // so prevents folding a load into this instruction or making a copy.
7598 const int UnpackLoMask[] = {0, 0, 1, 1};
7599 const int UnpackHiMask[] = {2, 2, 3, 3};
7600 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7601 Mask = UnpackLoMask;
7602 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7603 Mask = UnpackHiMask;
7605 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7606 getV4X86ShuffleImm8ForMask(Mask, DAG));
7609 // Try to use shift instructions.
7611 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7614 // There are special ways we can lower some single-element blends.
7615 if (NumV2Elements == 1)
7616 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7617 Mask, Subtarget, DAG))
7620 // We have different paths for blend lowering, but they all must use the
7621 // *exact* same predicate.
7622 bool IsBlendSupported = Subtarget->hasSSE41();
7623 if (IsBlendSupported)
7624 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7628 if (SDValue Masked =
7629 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7632 // Use dedicated unpack instructions for masks that match their pattern.
7633 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7634 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7635 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7636 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7637 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7638 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7639 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7640 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7642 // Try to use byte rotation instructions.
7643 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7644 if (Subtarget->hasSSSE3())
7645 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7646 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7649 // If we have direct support for blends, we should lower by decomposing into
7650 // a permute. That will be faster than the domain cross.
7651 if (IsBlendSupported)
7652 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7655 // Try to lower by permuting the inputs into an unpack instruction.
7656 if (SDValue Unpack =
7657 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7660 // We implement this with SHUFPS because it can blend from two vectors.
7661 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7662 // up the inputs, bypassing domain shift penalties that we would encur if we
7663 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7665 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7666 DAG.getVectorShuffle(
7668 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7669 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7672 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7673 /// shuffle lowering, and the most complex part.
7675 /// The lowering strategy is to try to form pairs of input lanes which are
7676 /// targeted at the same half of the final vector, and then use a dword shuffle
7677 /// to place them onto the right half, and finally unpack the paired lanes into
7678 /// their final position.
7680 /// The exact breakdown of how to form these dword pairs and align them on the
7681 /// correct sides is really tricky. See the comments within the function for
7682 /// more of the details.
7683 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7684 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
7685 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7686 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7687 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7688 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7690 SmallVector<int, 4> LoInputs;
7691 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7692 [](int M) { return M >= 0; });
7693 std::sort(LoInputs.begin(), LoInputs.end());
7694 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7695 SmallVector<int, 4> HiInputs;
7696 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7697 [](int M) { return M >= 0; });
7698 std::sort(HiInputs.begin(), HiInputs.end());
7699 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7701 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7702 int NumHToL = LoInputs.size() - NumLToL;
7704 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7705 int NumHToH = HiInputs.size() - NumLToH;
7706 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7707 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7708 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7709 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7711 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7712 // such inputs we can swap two of the dwords across the half mark and end up
7713 // with <=2 inputs to each half in each half. Once there, we can fall through
7714 // to the generic code below. For example:
7716 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7717 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7719 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7720 // and an existing 2-into-2 on the other half. In this case we may have to
7721 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7722 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7723 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7724 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7725 // half than the one we target for fixing) will be fixed when we re-enter this
7726 // path. We will also combine away any sequence of PSHUFD instructions that
7727 // result into a single instruction. Here is an example of the tricky case:
7729 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7730 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7732 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7734 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7735 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7737 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7738 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7740 // The result is fine to be handled by the generic logic.
7741 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
7742 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
7743 int AOffset, int BOffset) {
7744 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
7745 "Must call this with A having 3 or 1 inputs from the A half.");
7746 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
7747 "Must call this with B having 1 or 3 inputs from the B half.");
7748 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
7749 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
7751 // Compute the index of dword with only one word among the three inputs in
7752 // a half by taking the sum of the half with three inputs and subtracting
7753 // the sum of the actual three inputs. The difference is the remaining
7756 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
7757 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
7758 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
7759 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
7760 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
7761 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
7762 int TripleNonInputIdx =
7763 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
7764 TripleDWord = TripleNonInputIdx / 2;
7766 // We use xor with one to compute the adjacent DWord to whichever one the
7768 OneInputDWord = (OneInput / 2) ^ 1;
7770 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
7771 // and BToA inputs. If there is also such a problem with the BToB and AToB
7772 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
7773 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
7774 // is essential that we don't *create* a 3<-1 as then we might oscillate.
7775 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
7776 // Compute how many inputs will be flipped by swapping these DWords. We
7778 // to balance this to ensure we don't form a 3-1 shuffle in the other
7780 int NumFlippedAToBInputs =
7781 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
7782 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
7783 int NumFlippedBToBInputs =
7784 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
7785 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
7786 if ((NumFlippedAToBInputs == 1 &&
7787 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
7788 (NumFlippedBToBInputs == 1 &&
7789 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
7790 // We choose whether to fix the A half or B half based on whether that
7791 // half has zero flipped inputs. At zero, we may not be able to fix it
7792 // with that half. We also bias towards fixing the B half because that
7793 // will more commonly be the high half, and we have to bias one way.
7794 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
7795 ArrayRef<int> Inputs) {
7796 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
7797 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
7798 PinnedIdx ^ 1) != Inputs.end();
7799 // Determine whether the free index is in the flipped dword or the
7800 // unflipped dword based on where the pinned index is. We use this bit
7801 // in an xor to conditionally select the adjacent dword.
7802 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
7803 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7804 FixFreeIdx) != Inputs.end();
7805 if (IsFixIdxInput == IsFixFreeIdxInput)
7807 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7808 FixFreeIdx) != Inputs.end();
7809 assert(IsFixIdxInput != IsFixFreeIdxInput &&
7810 "We need to be changing the number of flipped inputs!");
7811 int PSHUFHalfMask[] = {0, 1, 2, 3};
7812 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
7813 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
7815 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
7818 if (M != -1 && M == FixIdx)
7820 else if (M != -1 && M == FixFreeIdx)
7823 if (NumFlippedBToBInputs != 0) {
7825 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7826 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
7828 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
7830 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7831 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
7836 int PSHUFDMask[] = {0, 1, 2, 3};
7837 PSHUFDMask[ADWord] = BDWord;
7838 PSHUFDMask[BDWord] = ADWord;
7839 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7840 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7841 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
7842 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7844 // Adjust the mask to match the new locations of A and B.
7846 if (M != -1 && M/2 == ADWord)
7847 M = 2 * BDWord + M % 2;
7848 else if (M != -1 && M/2 == BDWord)
7849 M = 2 * ADWord + M % 2;
7851 // Recurse back into this routine to re-compute state now that this isn't
7852 // a 3 and 1 problem.
7853 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
7856 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
7857 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
7858 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
7859 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
7861 // At this point there are at most two inputs to the low and high halves from
7862 // each half. That means the inputs can always be grouped into dwords and
7863 // those dwords can then be moved to the correct half with a dword shuffle.
7864 // We use at most one low and one high word shuffle to collect these paired
7865 // inputs into dwords, and finally a dword shuffle to place them.
7866 int PSHUFLMask[4] = {-1, -1, -1, -1};
7867 int PSHUFHMask[4] = {-1, -1, -1, -1};
7868 int PSHUFDMask[4] = {-1, -1, -1, -1};
7870 // First fix the masks for all the inputs that are staying in their
7871 // original halves. This will then dictate the targets of the cross-half
7873 auto fixInPlaceInputs =
7874 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
7875 MutableArrayRef<int> SourceHalfMask,
7876 MutableArrayRef<int> HalfMask, int HalfOffset) {
7877 if (InPlaceInputs.empty())
7879 if (InPlaceInputs.size() == 1) {
7880 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7881 InPlaceInputs[0] - HalfOffset;
7882 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
7885 if (IncomingInputs.empty()) {
7886 // Just fix all of the in place inputs.
7887 for (int Input : InPlaceInputs) {
7888 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
7889 PSHUFDMask[Input / 2] = Input / 2;
7894 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
7895 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7896 InPlaceInputs[0] - HalfOffset;
7897 // Put the second input next to the first so that they are packed into
7898 // a dword. We find the adjacent index by toggling the low bit.
7899 int AdjIndex = InPlaceInputs[0] ^ 1;
7900 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
7901 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
7902 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
7904 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
7905 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
7907 // Now gather the cross-half inputs and place them into a free dword of
7908 // their target half.
7909 // FIXME: This operation could almost certainly be simplified dramatically to
7910 // look more like the 3-1 fixing operation.
7911 auto moveInputsToRightHalf = [&PSHUFDMask](
7912 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
7913 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
7914 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
7916 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
7917 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
7919 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
7921 int LowWord = Word & ~1;
7922 int HighWord = Word | 1;
7923 return isWordClobbered(SourceHalfMask, LowWord) ||
7924 isWordClobbered(SourceHalfMask, HighWord);
7927 if (IncomingInputs.empty())
7930 if (ExistingInputs.empty()) {
7931 // Map any dwords with inputs from them into the right half.
7932 for (int Input : IncomingInputs) {
7933 // If the source half mask maps over the inputs, turn those into
7934 // swaps and use the swapped lane.
7935 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
7936 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
7937 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
7938 Input - SourceOffset;
7939 // We have to swap the uses in our half mask in one sweep.
7940 for (int &M : HalfMask)
7941 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
7943 else if (M == Input)
7944 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7946 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
7947 Input - SourceOffset &&
7948 "Previous placement doesn't match!");
7950 // Note that this correctly re-maps both when we do a swap and when
7951 // we observe the other side of the swap above. We rely on that to
7952 // avoid swapping the members of the input list directly.
7953 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7956 // Map the input's dword into the correct half.
7957 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
7958 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
7960 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
7962 "Previous placement doesn't match!");
7965 // And just directly shift any other-half mask elements to be same-half
7966 // as we will have mirrored the dword containing the element into the
7967 // same position within that half.
7968 for (int &M : HalfMask)
7969 if (M >= SourceOffset && M < SourceOffset + 4) {
7970 M = M - SourceOffset + DestOffset;
7971 assert(M >= 0 && "This should never wrap below zero!");
7976 // Ensure we have the input in a viable dword of its current half. This
7977 // is particularly tricky because the original position may be clobbered
7978 // by inputs being moved and *staying* in that half.
7979 if (IncomingInputs.size() == 1) {
7980 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
7981 int InputFixed = std::find(std::begin(SourceHalfMask),
7982 std::end(SourceHalfMask), -1) -
7983 std::begin(SourceHalfMask) + SourceOffset;
7984 SourceHalfMask[InputFixed - SourceOffset] =
7985 IncomingInputs[0] - SourceOffset;
7986 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
7988 IncomingInputs[0] = InputFixed;
7990 } else if (IncomingInputs.size() == 2) {
7991 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
7992 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
7993 // We have two non-adjacent or clobbered inputs we need to extract from
7994 // the source half. To do this, we need to map them into some adjacent
7995 // dword slot in the source mask.
7996 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
7997 IncomingInputs[1] - SourceOffset};
7999 // If there is a free slot in the source half mask adjacent to one of
8000 // the inputs, place the other input in it. We use (Index XOR 1) to
8001 // compute an adjacent index.
8002 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8003 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8004 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8005 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8006 InputsFixed[1] = InputsFixed[0] ^ 1;
8007 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8008 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8009 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8010 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8011 InputsFixed[0] = InputsFixed[1] ^ 1;
8012 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8013 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8014 // The two inputs are in the same DWord but it is clobbered and the
8015 // adjacent DWord isn't used at all. Move both inputs to the free
8017 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8018 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8019 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8020 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8022 // The only way we hit this point is if there is no clobbering
8023 // (because there are no off-half inputs to this half) and there is no
8024 // free slot adjacent to one of the inputs. In this case, we have to
8025 // swap an input with a non-input.
8026 for (int i = 0; i < 4; ++i)
8027 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8028 "We can't handle any clobbers here!");
8029 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8030 "Cannot have adjacent inputs here!");
8032 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8033 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8035 // We also have to update the final source mask in this case because
8036 // it may need to undo the above swap.
8037 for (int &M : FinalSourceHalfMask)
8038 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8039 M = InputsFixed[1] + SourceOffset;
8040 else if (M == InputsFixed[1] + SourceOffset)
8041 M = (InputsFixed[0] ^ 1) + SourceOffset;
8043 InputsFixed[1] = InputsFixed[0] ^ 1;
8046 // Point everything at the fixed inputs.
8047 for (int &M : HalfMask)
8048 if (M == IncomingInputs[0])
8049 M = InputsFixed[0] + SourceOffset;
8050 else if (M == IncomingInputs[1])
8051 M = InputsFixed[1] + SourceOffset;
8053 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8054 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8057 llvm_unreachable("Unhandled input size!");
8060 // Now hoist the DWord down to the right half.
8061 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8062 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8063 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8064 for (int &M : HalfMask)
8065 for (int Input : IncomingInputs)
8067 M = FreeDWord * 2 + Input % 2;
8069 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8070 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8071 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8072 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8074 // Now enact all the shuffles we've computed to move the inputs into their
8076 if (!isNoopShuffleMask(PSHUFLMask))
8077 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8078 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
8079 if (!isNoopShuffleMask(PSHUFHMask))
8080 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8081 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
8082 if (!isNoopShuffleMask(PSHUFDMask))
8083 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8084 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8085 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8086 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8088 // At this point, each half should contain all its inputs, and we can then
8089 // just shuffle them into their final position.
8090 assert(std::count_if(LoMask.begin(), LoMask.end(),
8091 [](int M) { return M >= 4; }) == 0 &&
8092 "Failed to lift all the high half inputs to the low mask!");
8093 assert(std::count_if(HiMask.begin(), HiMask.end(),
8094 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8095 "Failed to lift all the low half inputs to the high mask!");
8097 // Do a half shuffle for the low mask.
8098 if (!isNoopShuffleMask(LoMask))
8099 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8100 getV4X86ShuffleImm8ForMask(LoMask, DAG));
8102 // Do a half shuffle with the high mask after shifting its values down.
8103 for (int &M : HiMask)
8106 if (!isNoopShuffleMask(HiMask))
8107 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8108 getV4X86ShuffleImm8ForMask(HiMask, DAG));
8113 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8114 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8115 SDValue V2, ArrayRef<int> Mask,
8116 SelectionDAG &DAG, bool &V1InUse,
8118 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8124 int Size = Mask.size();
8125 int Scale = 16 / Size;
8126 for (int i = 0; i < 16; ++i) {
8127 if (Mask[i / Scale] == -1) {
8128 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8130 const int ZeroMask = 0x80;
8131 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8133 int V2Idx = Mask[i / Scale] < Size
8135 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8136 if (Zeroable[i / Scale])
8137 V1Idx = V2Idx = ZeroMask;
8138 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
8139 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
8140 V1InUse |= (ZeroMask != V1Idx);
8141 V2InUse |= (ZeroMask != V2Idx);
8146 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8147 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8148 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8150 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8151 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8152 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8154 // If we need shuffled inputs from both, blend the two.
8156 if (V1InUse && V2InUse)
8157 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8159 V = V1InUse ? V1 : V2;
8161 // Cast the result back to the correct type.
8162 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8165 /// \brief Generic lowering of 8-lane i16 shuffles.
8167 /// This handles both single-input shuffles and combined shuffle/blends with
8168 /// two inputs. The single input shuffles are immediately delegated to
8169 /// a dedicated lowering routine.
8171 /// The blends are lowered in one of three fundamental ways. If there are few
8172 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8173 /// of the input is significantly cheaper when lowered as an interleaving of
8174 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8175 /// halves of the inputs separately (making them have relatively few inputs)
8176 /// and then concatenate them.
8177 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8178 const X86Subtarget *Subtarget,
8179 SelectionDAG &DAG) {
8181 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8182 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8183 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8184 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8185 ArrayRef<int> OrigMask = SVOp->getMask();
8186 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8187 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8188 MutableArrayRef<int> Mask(MaskStorage);
8190 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8192 // Whenever we can lower this as a zext, that instruction is strictly faster
8193 // than any alternative.
8194 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8195 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8198 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8200 auto isV2 = [](int M) { return M >= 8; };
8202 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8204 if (NumV2Inputs == 0) {
8205 // Check for being able to broadcast a single element.
8206 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8207 Mask, Subtarget, DAG))
8210 // Try to use shift instructions.
8212 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8215 // Use dedicated unpack instructions for masks that match their pattern.
8216 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8217 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8218 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8219 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8221 // Try to use byte rotation instructions.
8222 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8223 Mask, Subtarget, DAG))
8226 return lowerV8I16GeneralSingleInputVectorShuffle(DL, V1, Mask, Subtarget,
8230 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8231 "All single-input shuffles should be canonicalized to be V1-input "
8234 // Try to use shift instructions.
8236 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8239 // There are special ways we can lower some single-element blends.
8240 if (NumV2Inputs == 1)
8241 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8242 Mask, Subtarget, DAG))
8245 // We have different paths for blend lowering, but they all must use the
8246 // *exact* same predicate.
8247 bool IsBlendSupported = Subtarget->hasSSE41();
8248 if (IsBlendSupported)
8249 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8253 if (SDValue Masked =
8254 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8257 // Use dedicated unpack instructions for masks that match their pattern.
8258 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8259 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8260 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8261 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8263 // Try to use byte rotation instructions.
8264 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8265 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8268 if (SDValue BitBlend =
8269 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8272 if (SDValue Unpack =
8273 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8276 // If we can't directly blend but can use PSHUFB, that will be better as it
8277 // can both shuffle and set up the inefficient blend.
8278 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8279 bool V1InUse, V2InUse;
8280 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8284 // We can always bit-blend if we have to so the fallback strategy is to
8285 // decompose into single-input permutes and blends.
8286 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8290 /// \brief Check whether a compaction lowering can be done by dropping even
8291 /// elements and compute how many times even elements must be dropped.
8293 /// This handles shuffles which take every Nth element where N is a power of
8294 /// two. Example shuffle masks:
8296 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8297 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8298 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8299 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8300 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8301 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8303 /// Any of these lanes can of course be undef.
8305 /// This routine only supports N <= 3.
8306 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8309 /// \returns N above, or the number of times even elements must be dropped if
8310 /// there is such a number. Otherwise returns zero.
8311 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8312 // Figure out whether we're looping over two inputs or just one.
8313 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8315 // The modulus for the shuffle vector entries is based on whether this is
8316 // a single input or not.
8317 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8318 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8319 "We should only be called with masks with a power-of-2 size!");
8321 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8323 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8324 // and 2^3 simultaneously. This is because we may have ambiguity with
8325 // partially undef inputs.
8326 bool ViableForN[3] = {true, true, true};
8328 for (int i = 0, e = Mask.size(); i < e; ++i) {
8329 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8334 bool IsAnyViable = false;
8335 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8336 if (ViableForN[j]) {
8339 // The shuffle mask must be equal to (i * 2^N) % M.
8340 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8343 ViableForN[j] = false;
8345 // Early exit if we exhaust the possible powers of two.
8350 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8354 // Return 0 as there is no viable power of two.
8358 /// \brief Generic lowering of v16i8 shuffles.
8360 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8361 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8362 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8363 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8365 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8366 const X86Subtarget *Subtarget,
8367 SelectionDAG &DAG) {
8369 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8370 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8371 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8372 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8373 ArrayRef<int> Mask = SVOp->getMask();
8374 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8376 // Try to use shift instructions.
8378 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8381 // Try to use byte rotation instructions.
8382 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8383 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8386 // Try to use a zext lowering.
8387 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8388 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8392 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8394 // For single-input shuffles, there are some nicer lowering tricks we can use.
8395 if (NumV2Elements == 0) {
8396 // Check for being able to broadcast a single element.
8397 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8398 Mask, Subtarget, DAG))
8401 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8402 // Notably, this handles splat and partial-splat shuffles more efficiently.
8403 // However, it only makes sense if the pre-duplication shuffle simplifies
8404 // things significantly. Currently, this means we need to be able to
8405 // express the pre-duplication shuffle as an i16 shuffle.
8407 // FIXME: We should check for other patterns which can be widened into an
8408 // i16 shuffle as well.
8409 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8410 for (int i = 0; i < 16; i += 2)
8411 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8416 auto tryToWidenViaDuplication = [&]() -> SDValue {
8417 if (!canWidenViaDuplication(Mask))
8419 SmallVector<int, 4> LoInputs;
8420 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8421 [](int M) { return M >= 0 && M < 8; });
8422 std::sort(LoInputs.begin(), LoInputs.end());
8423 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8425 SmallVector<int, 4> HiInputs;
8426 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8427 [](int M) { return M >= 8; });
8428 std::sort(HiInputs.begin(), HiInputs.end());
8429 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8432 bool TargetLo = LoInputs.size() >= HiInputs.size();
8433 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8434 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8436 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8437 SmallDenseMap<int, int, 8> LaneMap;
8438 for (int I : InPlaceInputs) {
8439 PreDupI16Shuffle[I/2] = I/2;
8442 int j = TargetLo ? 0 : 4, je = j + 4;
8443 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8444 // Check if j is already a shuffle of this input. This happens when
8445 // there are two adjacent bytes after we move the low one.
8446 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8447 // If we haven't yet mapped the input, search for a slot into which
8449 while (j < je && PreDupI16Shuffle[j] != -1)
8453 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8456 // Map this input with the i16 shuffle.
8457 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8460 // Update the lane map based on the mapping we ended up with.
8461 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8464 ISD::BITCAST, DL, MVT::v16i8,
8465 DAG.getVectorShuffle(MVT::v8i16, DL,
8466 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8467 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8469 // Unpack the bytes to form the i16s that will be shuffled into place.
8470 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8471 MVT::v16i8, V1, V1);
8473 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8474 for (int i = 0; i < 16; ++i)
8475 if (Mask[i] != -1) {
8476 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8477 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8478 if (PostDupI16Shuffle[i / 2] == -1)
8479 PostDupI16Shuffle[i / 2] = MappedMask;
8481 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8482 "Conflicting entrties in the original shuffle!");
8485 ISD::BITCAST, DL, MVT::v16i8,
8486 DAG.getVectorShuffle(MVT::v8i16, DL,
8487 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8488 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8490 if (SDValue V = tryToWidenViaDuplication())
8494 // Use dedicated unpack instructions for masks that match their pattern.
8495 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8496 0, 16, 1, 17, 2, 18, 3, 19,
8498 4, 20, 5, 21, 6, 22, 7, 23}))
8499 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8500 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8501 8, 24, 9, 25, 10, 26, 11, 27,
8503 12, 28, 13, 29, 14, 30, 15, 31}))
8504 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8506 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8507 // with PSHUFB. It is important to do this before we attempt to generate any
8508 // blends but after all of the single-input lowerings. If the single input
8509 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8510 // want to preserve that and we can DAG combine any longer sequences into
8511 // a PSHUFB in the end. But once we start blending from multiple inputs,
8512 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8513 // and there are *very* few patterns that would actually be faster than the
8514 // PSHUFB approach because of its ability to zero lanes.
8516 // FIXME: The only exceptions to the above are blends which are exact
8517 // interleavings with direct instructions supporting them. We currently don't
8518 // handle those well here.
8519 if (Subtarget->hasSSSE3()) {
8520 bool V1InUse = false;
8521 bool V2InUse = false;
8523 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8524 DAG, V1InUse, V2InUse);
8526 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8527 // do so. This avoids using them to handle blends-with-zero which is
8528 // important as a single pshufb is significantly faster for that.
8529 if (V1InUse && V2InUse) {
8530 if (Subtarget->hasSSE41())
8531 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8532 Mask, Subtarget, DAG))
8535 // We can use an unpack to do the blending rather than an or in some
8536 // cases. Even though the or may be (very minorly) more efficient, we
8537 // preference this lowering because there are common cases where part of
8538 // the complexity of the shuffles goes away when we do the final blend as
8540 // FIXME: It might be worth trying to detect if the unpack-feeding
8541 // shuffles will both be pshufb, in which case we shouldn't bother with
8543 if (SDValue Unpack =
8544 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8551 // There are special ways we can lower some single-element blends.
8552 if (NumV2Elements == 1)
8553 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8554 Mask, Subtarget, DAG))
8557 if (SDValue BitBlend =
8558 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8561 // Check whether a compaction lowering can be done. This handles shuffles
8562 // which take every Nth element for some even N. See the helper function for
8565 // We special case these as they can be particularly efficiently handled with
8566 // the PACKUSB instruction on x86 and they show up in common patterns of
8567 // rearranging bytes to truncate wide elements.
8568 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8569 // NumEvenDrops is the power of two stride of the elements. Another way of
8570 // thinking about it is that we need to drop the even elements this many
8571 // times to get the original input.
8572 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8574 // First we need to zero all the dropped bytes.
8575 assert(NumEvenDrops <= 3 &&
8576 "No support for dropping even elements more than 3 times.");
8577 // We use the mask type to pick which bytes are preserved based on how many
8578 // elements are dropped.
8579 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8580 SDValue ByteClearMask =
8581 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8582 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
8583 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8585 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8587 // Now pack things back together.
8588 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8589 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8590 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8591 for (int i = 1; i < NumEvenDrops; ++i) {
8592 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8593 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8599 // Handle multi-input cases by blending single-input shuffles.
8600 if (NumV2Elements > 0)
8601 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8604 // The fallback path for single-input shuffles widens this into two v8i16
8605 // vectors with unpacks, shuffles those, and then pulls them back together
8609 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8610 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8611 for (int i = 0; i < 16; ++i)
8613 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8615 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8617 SDValue VLoHalf, VHiHalf;
8618 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8619 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8621 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8622 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8623 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8624 [](int M) { return M >= 0 && M % 2 == 1; })) {
8625 // Use a mask to drop the high bytes.
8626 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8627 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8628 DAG.getConstant(0x00FF, MVT::v8i16));
8630 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8631 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8633 // Squash the masks to point directly into VLoHalf.
8634 for (int &M : LoBlendMask)
8637 for (int &M : HiBlendMask)
8641 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8642 // VHiHalf so that we can blend them as i16s.
8643 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8644 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8645 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8646 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8649 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8650 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8652 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8655 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8657 /// This routine breaks down the specific type of 128-bit shuffle and
8658 /// dispatches to the lowering routines accordingly.
8659 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8660 MVT VT, const X86Subtarget *Subtarget,
8661 SelectionDAG &DAG) {
8662 switch (VT.SimpleTy) {
8664 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8666 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8668 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8670 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8672 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8674 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8677 llvm_unreachable("Unimplemented!");
8681 /// \brief Helper function to test whether a shuffle mask could be
8682 /// simplified by widening the elements being shuffled.
8684 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8685 /// leaves it in an unspecified state.
8687 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8688 /// shuffle masks. The latter have the special property of a '-2' representing
8689 /// a zero-ed lane of a vector.
8690 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8691 SmallVectorImpl<int> &WidenedMask) {
8692 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8693 // If both elements are undef, its trivial.
8694 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8695 WidenedMask.push_back(SM_SentinelUndef);
8699 // Check for an undef mask and a mask value properly aligned to fit with
8700 // a pair of values. If we find such a case, use the non-undef mask's value.
8701 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8702 WidenedMask.push_back(Mask[i + 1] / 2);
8705 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8706 WidenedMask.push_back(Mask[i] / 2);
8710 // When zeroing, we need to spread the zeroing across both lanes to widen.
8711 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8712 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8713 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8714 WidenedMask.push_back(SM_SentinelZero);
8720 // Finally check if the two mask values are adjacent and aligned with
8722 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8723 WidenedMask.push_back(Mask[i] / 2);
8727 // Otherwise we can't safely widen the elements used in this shuffle.
8730 assert(WidenedMask.size() == Mask.size() / 2 &&
8731 "Incorrect size of mask after widening the elements!");
8736 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8738 /// This routine just extracts two subvectors, shuffles them independently, and
8739 /// then concatenates them back together. This should work effectively with all
8740 /// AVX vector shuffle types.
8741 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8742 SDValue V2, ArrayRef<int> Mask,
8743 SelectionDAG &DAG) {
8744 assert(VT.getSizeInBits() >= 256 &&
8745 "Only for 256-bit or wider vector shuffles!");
8746 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
8747 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
8749 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
8750 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
8752 int NumElements = VT.getVectorNumElements();
8753 int SplitNumElements = NumElements / 2;
8754 MVT ScalarVT = VT.getScalarType();
8755 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
8757 // Rather than splitting build-vectors, just build two narrower build
8758 // vectors. This helps shuffling with splats and zeros.
8759 auto SplitVector = [&](SDValue V) {
8760 while (V.getOpcode() == ISD::BITCAST)
8761 V = V->getOperand(0);
8763 MVT OrigVT = V.getSimpleValueType();
8764 int OrigNumElements = OrigVT.getVectorNumElements();
8765 int OrigSplitNumElements = OrigNumElements / 2;
8766 MVT OrigScalarVT = OrigVT.getScalarType();
8767 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
8771 auto *BV = dyn_cast<BuildVectorSDNode>(V);
8773 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8774 DAG.getIntPtrConstant(0));
8775 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8776 DAG.getIntPtrConstant(OrigSplitNumElements));
8779 SmallVector<SDValue, 16> LoOps, HiOps;
8780 for (int i = 0; i < OrigSplitNumElements; ++i) {
8781 LoOps.push_back(BV->getOperand(i));
8782 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
8784 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
8785 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
8787 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
8788 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
8791 SDValue LoV1, HiV1, LoV2, HiV2;
8792 std::tie(LoV1, HiV1) = SplitVector(V1);
8793 std::tie(LoV2, HiV2) = SplitVector(V2);
8795 // Now create two 4-way blends of these half-width vectors.
8796 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
8797 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
8798 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
8799 for (int i = 0; i < SplitNumElements; ++i) {
8800 int M = HalfMask[i];
8801 if (M >= NumElements) {
8802 if (M >= NumElements + SplitNumElements)
8806 V2BlendMask.push_back(M - NumElements);
8807 V1BlendMask.push_back(-1);
8808 BlendMask.push_back(SplitNumElements + i);
8809 } else if (M >= 0) {
8810 if (M >= SplitNumElements)
8814 V2BlendMask.push_back(-1);
8815 V1BlendMask.push_back(M);
8816 BlendMask.push_back(i);
8818 V2BlendMask.push_back(-1);
8819 V1BlendMask.push_back(-1);
8820 BlendMask.push_back(-1);
8824 // Because the lowering happens after all combining takes place, we need to
8825 // manually combine these blend masks as much as possible so that we create
8826 // a minimal number of high-level vector shuffle nodes.
8828 // First try just blending the halves of V1 or V2.
8829 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
8830 return DAG.getUNDEF(SplitVT);
8831 if (!UseLoV2 && !UseHiV2)
8832 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8833 if (!UseLoV1 && !UseHiV1)
8834 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8836 SDValue V1Blend, V2Blend;
8837 if (UseLoV1 && UseHiV1) {
8839 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8841 // We only use half of V1 so map the usage down into the final blend mask.
8842 V1Blend = UseLoV1 ? LoV1 : HiV1;
8843 for (int i = 0; i < SplitNumElements; ++i)
8844 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
8845 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
8847 if (UseLoV2 && UseHiV2) {
8849 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8851 // We only use half of V2 so map the usage down into the final blend mask.
8852 V2Blend = UseLoV2 ? LoV2 : HiV2;
8853 for (int i = 0; i < SplitNumElements; ++i)
8854 if (BlendMask[i] >= SplitNumElements)
8855 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
8857 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
8859 SDValue Lo = HalfBlend(LoMask);
8860 SDValue Hi = HalfBlend(HiMask);
8861 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
8864 /// \brief Either split a vector in halves or decompose the shuffles and the
8867 /// This is provided as a good fallback for many lowerings of non-single-input
8868 /// shuffles with more than one 128-bit lane. In those cases, we want to select
8869 /// between splitting the shuffle into 128-bit components and stitching those
8870 /// back together vs. extracting the single-input shuffles and blending those
8872 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
8873 SDValue V2, ArrayRef<int> Mask,
8874 SelectionDAG &DAG) {
8875 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
8876 "lower single-input shuffles as it "
8877 "could then recurse on itself.");
8878 int Size = Mask.size();
8880 // If this can be modeled as a broadcast of two elements followed by a blend,
8881 // prefer that lowering. This is especially important because broadcasts can
8882 // often fold with memory operands.
8883 auto DoBothBroadcast = [&] {
8884 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
8887 if (V2BroadcastIdx == -1)
8888 V2BroadcastIdx = M - Size;
8889 else if (M - Size != V2BroadcastIdx)
8891 } else if (M >= 0) {
8892 if (V1BroadcastIdx == -1)
8894 else if (M != V1BroadcastIdx)
8899 if (DoBothBroadcast())
8900 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
8903 // If the inputs all stem from a single 128-bit lane of each input, then we
8904 // split them rather than blending because the split will decompose to
8905 // unusually few instructions.
8906 int LaneCount = VT.getSizeInBits() / 128;
8907 int LaneSize = Size / LaneCount;
8908 SmallBitVector LaneInputs[2];
8909 LaneInputs[0].resize(LaneCount, false);
8910 LaneInputs[1].resize(LaneCount, false);
8911 for (int i = 0; i < Size; ++i)
8913 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
8914 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
8915 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8917 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
8918 // that the decomposed single-input shuffles don't end up here.
8919 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8922 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
8923 /// a permutation and blend of those lanes.
8925 /// This essentially blends the out-of-lane inputs to each lane into the lane
8926 /// from a permuted copy of the vector. This lowering strategy results in four
8927 /// instructions in the worst case for a single-input cross lane shuffle which
8928 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
8929 /// of. Special cases for each particular shuffle pattern should be handled
8930 /// prior to trying this lowering.
8931 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
8932 SDValue V1, SDValue V2,
8934 SelectionDAG &DAG) {
8935 // FIXME: This should probably be generalized for 512-bit vectors as well.
8936 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
8937 int LaneSize = Mask.size() / 2;
8939 // If there are only inputs from one 128-bit lane, splitting will in fact be
8940 // less expensive. The flags track wether the given lane contains an element
8941 // that crosses to another lane.
8942 bool LaneCrossing[2] = {false, false};
8943 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8944 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
8945 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
8946 if (!LaneCrossing[0] || !LaneCrossing[1])
8947 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8949 if (isSingleInputShuffleMask(Mask)) {
8950 SmallVector<int, 32> FlippedBlendMask;
8951 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8952 FlippedBlendMask.push_back(
8953 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
8955 : Mask[i] % LaneSize +
8956 (i / LaneSize) * LaneSize + Size));
8958 // Flip the vector, and blend the results which should now be in-lane. The
8959 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
8960 // 5 for the high source. The value 3 selects the high half of source 2 and
8961 // the value 2 selects the low half of source 2. We only use source 2 to
8962 // allow folding it into a memory operand.
8963 unsigned PERMMask = 3 | 2 << 4;
8964 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
8965 V1, DAG.getConstant(PERMMask, MVT::i8));
8966 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
8969 // This now reduces to two single-input shuffles of V1 and V2 which at worst
8970 // will be handled by the above logic and a blend of the results, much like
8971 // other patterns in AVX.
8972 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8975 /// \brief Handle lowering 2-lane 128-bit shuffles.
8976 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8977 SDValue V2, ArrayRef<int> Mask,
8978 const X86Subtarget *Subtarget,
8979 SelectionDAG &DAG) {
8980 // Blends are faster and handle all the non-lane-crossing cases.
8981 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
8985 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
8986 VT.getVectorNumElements() / 2);
8987 // Check for patterns which can be matched with a single insert of a 128-bit
8989 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1}) ||
8990 isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
8991 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
8992 DAG.getIntPtrConstant(0));
8993 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
8994 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
8995 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
8997 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 6, 7})) {
8998 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
8999 DAG.getIntPtrConstant(0));
9000 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
9001 DAG.getIntPtrConstant(2));
9002 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9005 // Otherwise form a 128-bit permutation.
9006 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
9007 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
9008 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9009 DAG.getConstant(PermMask, MVT::i8));
9012 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9013 /// shuffling each lane.
9015 /// This will only succeed when the result of fixing the 128-bit lanes results
9016 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9017 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9018 /// the lane crosses early and then use simpler shuffles within each lane.
9020 /// FIXME: It might be worthwhile at some point to support this without
9021 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9022 /// in x86 only floating point has interesting non-repeating shuffles, and even
9023 /// those are still *marginally* more expensive.
9024 static SDValue lowerVectorShuffleByMerging128BitLanes(
9025 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9026 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9027 assert(!isSingleInputShuffleMask(Mask) &&
9028 "This is only useful with multiple inputs.");
9030 int Size = Mask.size();
9031 int LaneSize = 128 / VT.getScalarSizeInBits();
9032 int NumLanes = Size / LaneSize;
9033 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9035 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9036 // check whether the in-128-bit lane shuffles share a repeating pattern.
9037 SmallVector<int, 4> Lanes;
9038 Lanes.resize(NumLanes, -1);
9039 SmallVector<int, 4> InLaneMask;
9040 InLaneMask.resize(LaneSize, -1);
9041 for (int i = 0; i < Size; ++i) {
9045 int j = i / LaneSize;
9048 // First entry we've seen for this lane.
9049 Lanes[j] = Mask[i] / LaneSize;
9050 } else if (Lanes[j] != Mask[i] / LaneSize) {
9051 // This doesn't match the lane selected previously!
9055 // Check that within each lane we have a consistent shuffle mask.
9056 int k = i % LaneSize;
9057 if (InLaneMask[k] < 0) {
9058 InLaneMask[k] = Mask[i] % LaneSize;
9059 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9060 // This doesn't fit a repeating in-lane mask.
9065 // First shuffle the lanes into place.
9066 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9067 VT.getSizeInBits() / 64);
9068 SmallVector<int, 8> LaneMask;
9069 LaneMask.resize(NumLanes * 2, -1);
9070 for (int i = 0; i < NumLanes; ++i)
9071 if (Lanes[i] >= 0) {
9072 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9073 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9076 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9077 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9078 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9080 // Cast it back to the type we actually want.
9081 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9083 // Now do a simple shuffle that isn't lane crossing.
9084 SmallVector<int, 8> NewMask;
9085 NewMask.resize(Size, -1);
9086 for (int i = 0; i < Size; ++i)
9088 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9089 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9090 "Must not introduce lane crosses at this point!");
9092 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9095 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9098 /// This returns true if the elements from a particular input are already in the
9099 /// slot required by the given mask and require no permutation.
9100 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9101 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9102 int Size = Mask.size();
9103 for (int i = 0; i < Size; ++i)
9104 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9110 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9112 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9113 /// isn't available.
9114 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9115 const X86Subtarget *Subtarget,
9116 SelectionDAG &DAG) {
9118 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9119 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9120 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9121 ArrayRef<int> Mask = SVOp->getMask();
9122 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9124 SmallVector<int, 4> WidenedMask;
9125 if (canWidenShuffleElements(Mask, WidenedMask))
9126 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9129 if (isSingleInputShuffleMask(Mask)) {
9130 // Check for being able to broadcast a single element.
9131 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9132 Mask, Subtarget, DAG))
9135 // Use low duplicate instructions for masks that match their pattern.
9136 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9137 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9139 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9140 // Non-half-crossing single input shuffles can be lowerid with an
9141 // interleaved permutation.
9142 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9143 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9144 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9145 DAG.getConstant(VPERMILPMask, MVT::i8));
9148 // With AVX2 we have direct support for this permutation.
9149 if (Subtarget->hasAVX2())
9150 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9151 getV4X86ShuffleImm8ForMask(Mask, DAG));
9153 // Otherwise, fall back.
9154 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9158 // X86 has dedicated unpack instructions that can handle specific blend
9159 // operations: UNPCKH and UNPCKL.
9160 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9161 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9162 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9163 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9164 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9165 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9166 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9167 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9169 // If we have a single input to the zero element, insert that into V1 if we
9170 // can do so cheaply.
9172 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
9173 if (NumV2Elements == 1 && Mask[0] >= 4)
9174 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9175 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9178 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9182 // Check if the blend happens to exactly fit that of SHUFPD.
9183 if ((Mask[0] == -1 || Mask[0] < 2) &&
9184 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9185 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9186 (Mask[3] == -1 || Mask[3] >= 6)) {
9187 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9188 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9189 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9190 DAG.getConstant(SHUFPDMask, MVT::i8));
9192 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9193 (Mask[1] == -1 || Mask[1] < 2) &&
9194 (Mask[2] == -1 || Mask[2] >= 6) &&
9195 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9196 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9197 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9198 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9199 DAG.getConstant(SHUFPDMask, MVT::i8));
9202 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9203 // shuffle. However, if we have AVX2 and either inputs are already in place,
9204 // we will be able to shuffle even across lanes the other input in a single
9205 // instruction so skip this pattern.
9206 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9207 isShuffleMaskInputInPlace(1, Mask))))
9208 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9209 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9212 // If we have AVX2 then we always want to lower with a blend because an v4 we
9213 // can fully permute the elements.
9214 if (Subtarget->hasAVX2())
9215 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9218 // Otherwise fall back on generic lowering.
9219 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9222 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9224 /// This routine is only called when we have AVX2 and thus a reasonable
9225 /// instruction set for v4i64 shuffling..
9226 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9227 const X86Subtarget *Subtarget,
9228 SelectionDAG &DAG) {
9230 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9231 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9232 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9233 ArrayRef<int> Mask = SVOp->getMask();
9234 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9235 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9237 SmallVector<int, 4> WidenedMask;
9238 if (canWidenShuffleElements(Mask, WidenedMask))
9239 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9242 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9246 // Check for being able to broadcast a single element.
9247 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9248 Mask, Subtarget, DAG))
9251 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9252 // use lower latency instructions that will operate on both 128-bit lanes.
9253 SmallVector<int, 2> RepeatedMask;
9254 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9255 if (isSingleInputShuffleMask(Mask)) {
9256 int PSHUFDMask[] = {-1, -1, -1, -1};
9257 for (int i = 0; i < 2; ++i)
9258 if (RepeatedMask[i] >= 0) {
9259 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9260 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9263 ISD::BITCAST, DL, MVT::v4i64,
9264 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9265 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9266 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9270 // AVX2 provides a direct instruction for permuting a single input across
9272 if (isSingleInputShuffleMask(Mask))
9273 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9274 getV4X86ShuffleImm8ForMask(Mask, DAG));
9276 // Try to use shift instructions.
9278 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9281 // Use dedicated unpack instructions for masks that match their pattern.
9282 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9283 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9284 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9285 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9286 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9287 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9288 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9289 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9291 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9292 // shuffle. However, if we have AVX2 and either inputs are already in place,
9293 // we will be able to shuffle even across lanes the other input in a single
9294 // instruction so skip this pattern.
9295 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9296 isShuffleMaskInputInPlace(1, Mask))))
9297 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9298 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9301 // Otherwise fall back on generic blend lowering.
9302 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9306 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9308 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9309 /// isn't available.
9310 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9311 const X86Subtarget *Subtarget,
9312 SelectionDAG &DAG) {
9314 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9315 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9316 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9317 ArrayRef<int> Mask = SVOp->getMask();
9318 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9320 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9324 // Check for being able to broadcast a single element.
9325 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9326 Mask, Subtarget, DAG))
9329 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9330 // options to efficiently lower the shuffle.
9331 SmallVector<int, 4> RepeatedMask;
9332 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9333 assert(RepeatedMask.size() == 4 &&
9334 "Repeated masks must be half the mask width!");
9336 // Use even/odd duplicate instructions for masks that match their pattern.
9337 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9338 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9339 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9340 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9342 if (isSingleInputShuffleMask(Mask))
9343 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9344 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9346 // Use dedicated unpack instructions for masks that match their pattern.
9347 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9348 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9349 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9350 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9351 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9352 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9353 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9354 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9356 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9357 // have already handled any direct blends. We also need to squash the
9358 // repeated mask into a simulated v4f32 mask.
9359 for (int i = 0; i < 4; ++i)
9360 if (RepeatedMask[i] >= 8)
9361 RepeatedMask[i] -= 4;
9362 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9365 // If we have a single input shuffle with different shuffle patterns in the
9366 // two 128-bit lanes use the variable mask to VPERMILPS.
9367 if (isSingleInputShuffleMask(Mask)) {
9368 SDValue VPermMask[8];
9369 for (int i = 0; i < 8; ++i)
9370 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9371 : DAG.getConstant(Mask[i], MVT::i32);
9372 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9374 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9375 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9377 if (Subtarget->hasAVX2())
9378 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9379 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9380 DAG.getNode(ISD::BUILD_VECTOR, DL,
9381 MVT::v8i32, VPermMask)),
9384 // Otherwise, fall back.
9385 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9389 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9391 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9392 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9395 // If we have AVX2 then we always want to lower with a blend because at v8 we
9396 // can fully permute the elements.
9397 if (Subtarget->hasAVX2())
9398 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9401 // Otherwise fall back on generic lowering.
9402 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9405 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9407 /// This routine is only called when we have AVX2 and thus a reasonable
9408 /// instruction set for v8i32 shuffling..
9409 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9410 const X86Subtarget *Subtarget,
9411 SelectionDAG &DAG) {
9413 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9414 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9415 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9416 ArrayRef<int> Mask = SVOp->getMask();
9417 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9418 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9420 // Whenever we can lower this as a zext, that instruction is strictly faster
9421 // than any alternative. It also allows us to fold memory operands into the
9422 // shuffle in many cases.
9423 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9424 Mask, Subtarget, DAG))
9427 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9431 // Check for being able to broadcast a single element.
9432 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9433 Mask, Subtarget, DAG))
9436 // If the shuffle mask is repeated in each 128-bit lane we can use more
9437 // efficient instructions that mirror the shuffles across the two 128-bit
9439 SmallVector<int, 4> RepeatedMask;
9440 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9441 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9442 if (isSingleInputShuffleMask(Mask))
9443 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9444 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9446 // Use dedicated unpack instructions for masks that match their pattern.
9447 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9448 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9449 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9450 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9451 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9452 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9453 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9454 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9457 // Try to use shift instructions.
9459 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9462 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9463 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9466 // If the shuffle patterns aren't repeated but it is a single input, directly
9467 // generate a cross-lane VPERMD instruction.
9468 if (isSingleInputShuffleMask(Mask)) {
9469 SDValue VPermMask[8];
9470 for (int i = 0; i < 8; ++i)
9471 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9472 : DAG.getConstant(Mask[i], MVT::i32);
9474 X86ISD::VPERMV, DL, MVT::v8i32,
9475 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9478 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9480 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9481 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9484 // Otherwise fall back on generic blend lowering.
9485 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9489 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9491 /// This routine is only called when we have AVX2 and thus a reasonable
9492 /// instruction set for v16i16 shuffling..
9493 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9494 const X86Subtarget *Subtarget,
9495 SelectionDAG &DAG) {
9497 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9498 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9499 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9500 ArrayRef<int> Mask = SVOp->getMask();
9501 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9502 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9504 // Whenever we can lower this as a zext, that instruction is strictly faster
9505 // than any alternative. It also allows us to fold memory operands into the
9506 // shuffle in many cases.
9507 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9508 Mask, Subtarget, DAG))
9511 // Check for being able to broadcast a single element.
9512 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9513 Mask, Subtarget, DAG))
9516 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9520 // Use dedicated unpack instructions for masks that match their pattern.
9521 if (isShuffleEquivalent(V1, V2, Mask,
9522 {// First 128-bit lane:
9523 0, 16, 1, 17, 2, 18, 3, 19,
9524 // Second 128-bit lane:
9525 8, 24, 9, 25, 10, 26, 11, 27}))
9526 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9527 if (isShuffleEquivalent(V1, V2, Mask,
9528 {// First 128-bit lane:
9529 4, 20, 5, 21, 6, 22, 7, 23,
9530 // Second 128-bit lane:
9531 12, 28, 13, 29, 14, 30, 15, 31}))
9532 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9534 // Try to use shift instructions.
9536 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9539 // Try to use byte rotation instructions.
9540 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9541 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9544 if (isSingleInputShuffleMask(Mask)) {
9545 // There are no generalized cross-lane shuffle operations available on i16
9547 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9548 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9551 SDValue PSHUFBMask[32];
9552 for (int i = 0; i < 16; ++i) {
9553 if (Mask[i] == -1) {
9554 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9558 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9559 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9560 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
9561 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
9564 ISD::BITCAST, DL, MVT::v16i16,
9566 X86ISD::PSHUFB, DL, MVT::v32i8,
9567 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9568 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9571 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9573 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9574 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9577 // Otherwise fall back on generic lowering.
9578 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9581 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9583 /// This routine is only called when we have AVX2 and thus a reasonable
9584 /// instruction set for v32i8 shuffling..
9585 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9586 const X86Subtarget *Subtarget,
9587 SelectionDAG &DAG) {
9589 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9590 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9591 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9592 ArrayRef<int> Mask = SVOp->getMask();
9593 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9594 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9596 // Whenever we can lower this as a zext, that instruction is strictly faster
9597 // than any alternative. It also allows us to fold memory operands into the
9598 // shuffle in many cases.
9599 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9600 Mask, Subtarget, DAG))
9603 // Check for being able to broadcast a single element.
9604 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9605 Mask, Subtarget, DAG))
9608 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9612 // Use dedicated unpack instructions for masks that match their pattern.
9613 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9615 if (isShuffleEquivalent(
9617 {// First 128-bit lane:
9618 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9619 // Second 128-bit lane:
9620 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9621 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9622 if (isShuffleEquivalent(
9624 {// First 128-bit lane:
9625 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9626 // Second 128-bit lane:
9627 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9628 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9630 // Try to use shift instructions.
9632 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9635 // Try to use byte rotation instructions.
9636 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9637 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9640 if (isSingleInputShuffleMask(Mask)) {
9641 // There are no generalized cross-lane shuffle operations available on i8
9643 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9644 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9647 SDValue PSHUFBMask[32];
9648 for (int i = 0; i < 32; ++i)
9651 ? DAG.getUNDEF(MVT::i8)
9652 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
9655 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9656 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9659 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9661 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9662 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9665 // Otherwise fall back on generic lowering.
9666 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9669 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9671 /// This routine either breaks down the specific type of a 256-bit x86 vector
9672 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9673 /// together based on the available instructions.
9674 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9675 MVT VT, const X86Subtarget *Subtarget,
9676 SelectionDAG &DAG) {
9678 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9679 ArrayRef<int> Mask = SVOp->getMask();
9681 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9682 // check for those subtargets here and avoid much of the subtarget querying in
9683 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9684 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9685 // floating point types there eventually, just immediately cast everything to
9686 // a float and operate entirely in that domain.
9687 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9688 int ElementBits = VT.getScalarSizeInBits();
9689 if (ElementBits < 32)
9690 // No floating point type available, decompose into 128-bit vectors.
9691 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9693 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9694 VT.getVectorNumElements());
9695 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9696 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9697 return DAG.getNode(ISD::BITCAST, DL, VT,
9698 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9701 switch (VT.SimpleTy) {
9703 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9705 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9707 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9709 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9711 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9713 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9716 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9720 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
9721 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9722 const X86Subtarget *Subtarget,
9723 SelectionDAG &DAG) {
9725 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9726 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9727 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9728 ArrayRef<int> Mask = SVOp->getMask();
9729 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9731 // X86 has dedicated unpack instructions that can handle specific blend
9732 // operations: UNPCKH and UNPCKL.
9733 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9734 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
9735 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9736 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
9738 // FIXME: Implement direct support for this type!
9739 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
9742 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
9743 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9744 const X86Subtarget *Subtarget,
9745 SelectionDAG &DAG) {
9747 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9748 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9749 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9750 ArrayRef<int> Mask = SVOp->getMask();
9751 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9753 // Use dedicated unpack instructions for masks that match their pattern.
9754 if (isShuffleEquivalent(V1, V2, Mask,
9755 {// First 128-bit lane.
9756 0, 16, 1, 17, 4, 20, 5, 21,
9757 // Second 128-bit lane.
9758 8, 24, 9, 25, 12, 28, 13, 29}))
9759 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
9760 if (isShuffleEquivalent(V1, V2, Mask,
9761 {// First 128-bit lane.
9762 2, 18, 3, 19, 6, 22, 7, 23,
9763 // Second 128-bit lane.
9764 10, 26, 11, 27, 14, 30, 15, 31}))
9765 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
9767 // FIXME: Implement direct support for this type!
9768 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
9771 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
9772 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9773 const X86Subtarget *Subtarget,
9774 SelectionDAG &DAG) {
9776 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9777 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9778 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9779 ArrayRef<int> Mask = SVOp->getMask();
9780 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9782 // X86 has dedicated unpack instructions that can handle specific blend
9783 // operations: UNPCKH and UNPCKL.
9784 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9785 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
9786 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9787 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
9789 // FIXME: Implement direct support for this type!
9790 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
9793 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
9794 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9795 const X86Subtarget *Subtarget,
9796 SelectionDAG &DAG) {
9798 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9799 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9800 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9801 ArrayRef<int> Mask = SVOp->getMask();
9802 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9804 // Use dedicated unpack instructions for masks that match their pattern.
9805 if (isShuffleEquivalent(V1, V2, Mask,
9806 {// First 128-bit lane.
9807 0, 16, 1, 17, 4, 20, 5, 21,
9808 // Second 128-bit lane.
9809 8, 24, 9, 25, 12, 28, 13, 29}))
9810 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
9811 if (isShuffleEquivalent(V1, V2, Mask,
9812 {// First 128-bit lane.
9813 2, 18, 3, 19, 6, 22, 7, 23,
9814 // Second 128-bit lane.
9815 10, 26, 11, 27, 14, 30, 15, 31}))
9816 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
9818 // FIXME: Implement direct support for this type!
9819 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
9822 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
9823 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9824 const X86Subtarget *Subtarget,
9825 SelectionDAG &DAG) {
9827 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9828 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9829 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9830 ArrayRef<int> Mask = SVOp->getMask();
9831 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9832 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
9834 // FIXME: Implement direct support for this type!
9835 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
9838 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
9839 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9840 const X86Subtarget *Subtarget,
9841 SelectionDAG &DAG) {
9843 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9844 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9845 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9846 ArrayRef<int> Mask = SVOp->getMask();
9847 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
9848 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
9850 // FIXME: Implement direct support for this type!
9851 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
9854 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
9856 /// This routine either breaks down the specific type of a 512-bit x86 vector
9857 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
9858 /// together based on the available instructions.
9859 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9860 MVT VT, const X86Subtarget *Subtarget,
9861 SelectionDAG &DAG) {
9863 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9864 ArrayRef<int> Mask = SVOp->getMask();
9865 assert(Subtarget->hasAVX512() &&
9866 "Cannot lower 512-bit vectors w/ basic ISA!");
9868 // Check for being able to broadcast a single element.
9869 if (SDValue Broadcast =
9870 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
9873 // Dispatch to each element type for lowering. If we don't have supprot for
9874 // specific element type shuffles at 512 bits, immediately split them and
9875 // lower them. Each lowering routine of a given type is allowed to assume that
9876 // the requisite ISA extensions for that element type are available.
9877 switch (VT.SimpleTy) {
9879 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9881 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9883 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9885 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9887 if (Subtarget->hasBWI())
9888 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9891 if (Subtarget->hasBWI())
9892 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9896 llvm_unreachable("Not a valid 512-bit x86 vector type!");
9899 // Otherwise fall back on splitting.
9900 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9903 /// \brief Top-level lowering for x86 vector shuffles.
9905 /// This handles decomposition, canonicalization, and lowering of all x86
9906 /// vector shuffles. Most of the specific lowering strategies are encapsulated
9907 /// above in helper routines. The canonicalization attempts to widen shuffles
9908 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
9909 /// s.t. only one of the two inputs needs to be tested, etc.
9910 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
9911 SelectionDAG &DAG) {
9912 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9913 ArrayRef<int> Mask = SVOp->getMask();
9914 SDValue V1 = Op.getOperand(0);
9915 SDValue V2 = Op.getOperand(1);
9916 MVT VT = Op.getSimpleValueType();
9917 int NumElements = VT.getVectorNumElements();
9920 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
9922 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
9923 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
9924 if (V1IsUndef && V2IsUndef)
9925 return DAG.getUNDEF(VT);
9927 // When we create a shuffle node we put the UNDEF node to second operand,
9928 // but in some cases the first operand may be transformed to UNDEF.
9929 // In this case we should just commute the node.
9931 return DAG.getCommutedVectorShuffle(*SVOp);
9933 // Check for non-undef masks pointing at an undef vector and make the masks
9934 // undef as well. This makes it easier to match the shuffle based solely on
9938 if (M >= NumElements) {
9939 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
9940 for (int &M : NewMask)
9941 if (M >= NumElements)
9943 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
9946 // We actually see shuffles that are entirely re-arrangements of a set of
9947 // zero inputs. This mostly happens while decomposing complex shuffles into
9948 // simple ones. Directly lower these as a buildvector of zeros.
9949 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9951 return getZeroVector(VT, Subtarget, DAG, dl);
9953 // Try to collapse shuffles into using a vector type with fewer elements but
9954 // wider element types. We cap this to not form integers or floating point
9955 // elements wider than 64 bits, but it might be interesting to form i128
9956 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
9957 SmallVector<int, 16> WidenedMask;
9958 if (VT.getScalarSizeInBits() < 64 &&
9959 canWidenShuffleElements(Mask, WidenedMask)) {
9960 MVT NewEltVT = VT.isFloatingPoint()
9961 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
9962 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
9963 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
9964 // Make sure that the new vector type is legal. For example, v2f64 isn't
9966 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
9967 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
9968 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
9969 return DAG.getNode(ISD::BITCAST, dl, VT,
9970 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
9974 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
9975 for (int M : SVOp->getMask())
9978 else if (M < NumElements)
9983 // Commute the shuffle as needed such that more elements come from V1 than
9984 // V2. This allows us to match the shuffle pattern strictly on how many
9985 // elements come from V1 without handling the symmetric cases.
9986 if (NumV2Elements > NumV1Elements)
9987 return DAG.getCommutedVectorShuffle(*SVOp);
9989 // When the number of V1 and V2 elements are the same, try to minimize the
9990 // number of uses of V2 in the low half of the vector. When that is tied,
9991 // ensure that the sum of indices for V1 is equal to or lower than the sum
9992 // indices for V2. When those are equal, try to ensure that the number of odd
9993 // indices for V1 is lower than the number of odd indices for V2.
9994 if (NumV1Elements == NumV2Elements) {
9995 int LowV1Elements = 0, LowV2Elements = 0;
9996 for (int M : SVOp->getMask().slice(0, NumElements / 2))
9997 if (M >= NumElements)
10001 if (LowV2Elements > LowV1Elements) {
10002 return DAG.getCommutedVectorShuffle(*SVOp);
10003 } else if (LowV2Elements == LowV1Elements) {
10004 int SumV1Indices = 0, SumV2Indices = 0;
10005 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10006 if (SVOp->getMask()[i] >= NumElements)
10008 else if (SVOp->getMask()[i] >= 0)
10010 if (SumV2Indices < SumV1Indices) {
10011 return DAG.getCommutedVectorShuffle(*SVOp);
10012 } else if (SumV2Indices == SumV1Indices) {
10013 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10014 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10015 if (SVOp->getMask()[i] >= NumElements)
10016 NumV2OddIndices += i % 2;
10017 else if (SVOp->getMask()[i] >= 0)
10018 NumV1OddIndices += i % 2;
10019 if (NumV2OddIndices < NumV1OddIndices)
10020 return DAG.getCommutedVectorShuffle(*SVOp);
10025 // For each vector width, delegate to a specialized lowering routine.
10026 if (VT.getSizeInBits() == 128)
10027 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10029 if (VT.getSizeInBits() == 256)
10030 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10032 // Force AVX-512 vectors to be scalarized for now.
10033 // FIXME: Implement AVX-512 support!
10034 if (VT.getSizeInBits() == 512)
10035 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10037 llvm_unreachable("Unimplemented!");
10040 // This function assumes its argument is a BUILD_VECTOR of constants or
10041 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10043 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10044 unsigned &MaskValue) {
10046 unsigned NumElems = BuildVector->getNumOperands();
10047 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10048 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10049 unsigned NumElemsInLane = NumElems / NumLanes;
10051 // Blend for v16i16 should be symetric for the both lanes.
10052 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10053 SDValue EltCond = BuildVector->getOperand(i);
10054 SDValue SndLaneEltCond =
10055 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10057 int Lane1Cond = -1, Lane2Cond = -1;
10058 if (isa<ConstantSDNode>(EltCond))
10059 Lane1Cond = !isZero(EltCond);
10060 if (isa<ConstantSDNode>(SndLaneEltCond))
10061 Lane2Cond = !isZero(SndLaneEltCond);
10063 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10064 // Lane1Cond != 0, means we want the first argument.
10065 // Lane1Cond == 0, means we want the second argument.
10066 // The encoding of this argument is 0 for the first argument, 1
10067 // for the second. Therefore, invert the condition.
10068 MaskValue |= !Lane1Cond << i;
10069 else if (Lane1Cond < 0)
10070 MaskValue |= !Lane2Cond << i;
10077 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10078 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10079 const X86Subtarget *Subtarget,
10080 SelectionDAG &DAG) {
10081 SDValue Cond = Op.getOperand(0);
10082 SDValue LHS = Op.getOperand(1);
10083 SDValue RHS = Op.getOperand(2);
10085 MVT VT = Op.getSimpleValueType();
10087 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10089 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10091 // Only non-legal VSELECTs reach this lowering, convert those into generic
10092 // shuffles and re-use the shuffle lowering path for blends.
10093 SmallVector<int, 32> Mask;
10094 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10095 SDValue CondElt = CondBV->getOperand(i);
10097 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10099 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10102 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10103 // A vselect where all conditions and data are constants can be optimized into
10104 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10105 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10106 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10107 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10110 // Try to lower this to a blend-style vector shuffle. This can handle all
10111 // constant condition cases.
10112 SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG);
10113 if (BlendOp.getNode())
10116 // Variable blends are only legal from SSE4.1 onward.
10117 if (!Subtarget->hasSSE41())
10120 // Some types for vselect were previously set to Expand, not Legal or
10121 // Custom. Return an empty SDValue so we fall-through to Expand, after
10122 // the Custom lowering phase.
10123 MVT VT = Op.getSimpleValueType();
10124 switch (VT.SimpleTy) {
10129 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10134 // We couldn't create a "Blend with immediate" node.
10135 // This node should still be legal, but we'll have to emit a blendv*
10140 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10141 MVT VT = Op.getSimpleValueType();
10144 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10147 if (VT.getSizeInBits() == 8) {
10148 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10149 Op.getOperand(0), Op.getOperand(1));
10150 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10151 DAG.getValueType(VT));
10152 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10155 if (VT.getSizeInBits() == 16) {
10156 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10157 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10159 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10160 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10161 DAG.getNode(ISD::BITCAST, dl,
10164 Op.getOperand(1)));
10165 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10166 Op.getOperand(0), Op.getOperand(1));
10167 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10168 DAG.getValueType(VT));
10169 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10172 if (VT == MVT::f32) {
10173 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10174 // the result back to FR32 register. It's only worth matching if the
10175 // result has a single use which is a store or a bitcast to i32. And in
10176 // the case of a store, it's not worth it if the index is a constant 0,
10177 // because a MOVSSmr can be used instead, which is smaller and faster.
10178 if (!Op.hasOneUse())
10180 SDNode *User = *Op.getNode()->use_begin();
10181 if ((User->getOpcode() != ISD::STORE ||
10182 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10183 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10184 (User->getOpcode() != ISD::BITCAST ||
10185 User->getValueType(0) != MVT::i32))
10187 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10188 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10191 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10194 if (VT == MVT::i32 || VT == MVT::i64) {
10195 // ExtractPS/pextrq works with constant index.
10196 if (isa<ConstantSDNode>(Op.getOperand(1)))
10202 /// Extract one bit from mask vector, like v16i1 or v8i1.
10203 /// AVX-512 feature.
10205 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10206 SDValue Vec = Op.getOperand(0);
10208 MVT VecVT = Vec.getSimpleValueType();
10209 SDValue Idx = Op.getOperand(1);
10210 MVT EltVT = Op.getSimpleValueType();
10212 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10213 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10214 "Unexpected vector type in ExtractBitFromMaskVector");
10216 // variable index can't be handled in mask registers,
10217 // extend vector to VR512
10218 if (!isa<ConstantSDNode>(Idx)) {
10219 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10220 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10221 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10222 ExtVT.getVectorElementType(), Ext, Idx);
10223 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10226 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10227 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10228 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10229 rc = getRegClassFor(MVT::v16i1);
10230 unsigned MaxSift = rc->getSize()*8 - 1;
10231 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10232 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10233 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10234 DAG.getConstant(MaxSift, MVT::i8));
10235 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10236 DAG.getIntPtrConstant(0));
10240 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10241 SelectionDAG &DAG) const {
10243 SDValue Vec = Op.getOperand(0);
10244 MVT VecVT = Vec.getSimpleValueType();
10245 SDValue Idx = Op.getOperand(1);
10247 if (Op.getSimpleValueType() == MVT::i1)
10248 return ExtractBitFromMaskVector(Op, DAG);
10250 if (!isa<ConstantSDNode>(Idx)) {
10251 if (VecVT.is512BitVector() ||
10252 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10253 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10256 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10257 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10258 MaskEltVT.getSizeInBits());
10260 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10261 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10262 getZeroVector(MaskVT, Subtarget, DAG, dl),
10263 Idx, DAG.getConstant(0, getPointerTy()));
10264 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10265 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10266 Perm, DAG.getConstant(0, getPointerTy()));
10271 // If this is a 256-bit vector result, first extract the 128-bit vector and
10272 // then extract the element from the 128-bit vector.
10273 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10275 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10276 // Get the 128-bit vector.
10277 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10278 MVT EltVT = VecVT.getVectorElementType();
10280 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10282 //if (IdxVal >= NumElems/2)
10283 // IdxVal -= NumElems/2;
10284 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10285 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10286 DAG.getConstant(IdxVal, MVT::i32));
10289 assert(VecVT.is128BitVector() && "Unexpected vector length");
10291 if (Subtarget->hasSSE41()) {
10292 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10297 MVT VT = Op.getSimpleValueType();
10298 // TODO: handle v16i8.
10299 if (VT.getSizeInBits() == 16) {
10300 SDValue Vec = Op.getOperand(0);
10301 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10303 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10304 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10305 DAG.getNode(ISD::BITCAST, dl,
10307 Op.getOperand(1)));
10308 // Transform it so it match pextrw which produces a 32-bit result.
10309 MVT EltVT = MVT::i32;
10310 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10311 Op.getOperand(0), Op.getOperand(1));
10312 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10313 DAG.getValueType(VT));
10314 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10317 if (VT.getSizeInBits() == 32) {
10318 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10322 // SHUFPS the element to the lowest double word, then movss.
10323 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10324 MVT VVT = Op.getOperand(0).getSimpleValueType();
10325 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10326 DAG.getUNDEF(VVT), Mask);
10327 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10328 DAG.getIntPtrConstant(0));
10331 if (VT.getSizeInBits() == 64) {
10332 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10333 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10334 // to match extract_elt for f64.
10335 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10339 // UNPCKHPD the element to the lowest double word, then movsd.
10340 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10341 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10342 int Mask[2] = { 1, -1 };
10343 MVT VVT = Op.getOperand(0).getSimpleValueType();
10344 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10345 DAG.getUNDEF(VVT), Mask);
10346 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10347 DAG.getIntPtrConstant(0));
10353 /// Insert one bit to mask vector, like v16i1 or v8i1.
10354 /// AVX-512 feature.
10356 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10358 SDValue Vec = Op.getOperand(0);
10359 SDValue Elt = Op.getOperand(1);
10360 SDValue Idx = Op.getOperand(2);
10361 MVT VecVT = Vec.getSimpleValueType();
10363 if (!isa<ConstantSDNode>(Idx)) {
10364 // Non constant index. Extend source and destination,
10365 // insert element and then truncate the result.
10366 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10367 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10368 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10369 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10370 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10371 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10374 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10375 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10376 if (Vec.getOpcode() == ISD::UNDEF)
10377 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10378 DAG.getConstant(IdxVal, MVT::i8));
10379 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10380 unsigned MaxSift = rc->getSize()*8 - 1;
10381 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10382 DAG.getConstant(MaxSift, MVT::i8));
10383 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10384 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10385 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10388 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10389 SelectionDAG &DAG) const {
10390 MVT VT = Op.getSimpleValueType();
10391 MVT EltVT = VT.getVectorElementType();
10393 if (EltVT == MVT::i1)
10394 return InsertBitToMaskVector(Op, DAG);
10397 SDValue N0 = Op.getOperand(0);
10398 SDValue N1 = Op.getOperand(1);
10399 SDValue N2 = Op.getOperand(2);
10400 if (!isa<ConstantSDNode>(N2))
10402 auto *N2C = cast<ConstantSDNode>(N2);
10403 unsigned IdxVal = N2C->getZExtValue();
10405 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10406 // into that, and then insert the subvector back into the result.
10407 if (VT.is256BitVector() || VT.is512BitVector()) {
10408 // Get the desired 128-bit vector half.
10409 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10411 // Insert the element into the desired half.
10412 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10413 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10415 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10416 DAG.getConstant(IdxIn128, MVT::i32));
10418 // Insert the changed part back to the 256-bit vector
10419 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10421 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10423 if (Subtarget->hasSSE41()) {
10424 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10426 if (VT == MVT::v8i16) {
10427 Opc = X86ISD::PINSRW;
10429 assert(VT == MVT::v16i8);
10430 Opc = X86ISD::PINSRB;
10433 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10435 if (N1.getValueType() != MVT::i32)
10436 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10437 if (N2.getValueType() != MVT::i32)
10438 N2 = DAG.getIntPtrConstant(IdxVal);
10439 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10442 if (EltVT == MVT::f32) {
10443 // Bits [7:6] of the constant are the source select. This will always be
10444 // zero here. The DAG Combiner may combine an extract_elt index into
10446 // bits. For example (insert (extract, 3), 2) could be matched by
10448 // the '3' into bits [7:6] of X86ISD::INSERTPS.
10449 // Bits [5:4] of the constant are the destination select. This is the
10450 // value of the incoming immediate.
10451 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10452 // combine either bitwise AND or insert of float 0.0 to set these bits.
10453 N2 = DAG.getIntPtrConstant(IdxVal << 4);
10454 // Create this as a scalar to vector..
10455 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10456 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10459 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10460 // PINSR* works with constant index.
10465 if (EltVT == MVT::i8)
10468 if (EltVT.getSizeInBits() == 16) {
10469 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10470 // as its second argument.
10471 if (N1.getValueType() != MVT::i32)
10472 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10473 if (N2.getValueType() != MVT::i32)
10474 N2 = DAG.getIntPtrConstant(IdxVal);
10475 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10480 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10482 MVT OpVT = Op.getSimpleValueType();
10484 // If this is a 256-bit vector result, first insert into a 128-bit
10485 // vector and then insert into the 256-bit vector.
10486 if (!OpVT.is128BitVector()) {
10487 // Insert into a 128-bit vector.
10488 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10489 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10490 OpVT.getVectorNumElements() / SizeFactor);
10492 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10494 // Insert the 128-bit vector.
10495 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10498 if (OpVT == MVT::v1i64 &&
10499 Op.getOperand(0).getValueType() == MVT::i64)
10500 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10502 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10503 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10504 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10505 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10508 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10509 // a simple subregister reference or explicit instructions to grab
10510 // upper bits of a vector.
10511 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10512 SelectionDAG &DAG) {
10514 SDValue In = Op.getOperand(0);
10515 SDValue Idx = Op.getOperand(1);
10516 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10517 MVT ResVT = Op.getSimpleValueType();
10518 MVT InVT = In.getSimpleValueType();
10520 if (Subtarget->hasFp256()) {
10521 if (ResVT.is128BitVector() &&
10522 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10523 isa<ConstantSDNode>(Idx)) {
10524 return Extract128BitVector(In, IdxVal, DAG, dl);
10526 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10527 isa<ConstantSDNode>(Idx)) {
10528 return Extract256BitVector(In, IdxVal, DAG, dl);
10534 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10535 // simple superregister reference or explicit instructions to insert
10536 // the upper bits of a vector.
10537 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10538 SelectionDAG &DAG) {
10539 if (!Subtarget->hasAVX())
10543 SDValue Vec = Op.getOperand(0);
10544 SDValue SubVec = Op.getOperand(1);
10545 SDValue Idx = Op.getOperand(2);
10547 if (!isa<ConstantSDNode>(Idx))
10550 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10551 MVT OpVT = Op.getSimpleValueType();
10552 MVT SubVecVT = SubVec.getSimpleValueType();
10554 // Fold two 16-byte subvector loads into one 32-byte load:
10555 // (insert_subvector (insert_subvector undef, (load addr), 0),
10556 // (load addr + 16), Elts/2)
10558 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10559 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10560 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10561 !Subtarget->isUnalignedMem32Slow()) {
10562 SDValue SubVec2 = Vec.getOperand(1);
10563 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10564 if (Idx2->getZExtValue() == 0) {
10565 SDValue Ops[] = { SubVec2, SubVec };
10566 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10573 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10574 SubVecVT.is128BitVector())
10575 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10577 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10578 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10583 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10584 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10585 // one of the above mentioned nodes. It has to be wrapped because otherwise
10586 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10587 // be used to form addressing mode. These wrapped nodes will be selected
10590 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10591 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10593 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10594 // global base reg.
10595 unsigned char OpFlag = 0;
10596 unsigned WrapperKind = X86ISD::Wrapper;
10597 CodeModel::Model M = DAG.getTarget().getCodeModel();
10599 if (Subtarget->isPICStyleRIPRel() &&
10600 (M == CodeModel::Small || M == CodeModel::Kernel))
10601 WrapperKind = X86ISD::WrapperRIP;
10602 else if (Subtarget->isPICStyleGOT())
10603 OpFlag = X86II::MO_GOTOFF;
10604 else if (Subtarget->isPICStyleStubPIC())
10605 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10607 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10608 CP->getAlignment(),
10609 CP->getOffset(), OpFlag);
10611 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10612 // With PIC, the address is actually $g + Offset.
10614 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10615 DAG.getNode(X86ISD::GlobalBaseReg,
10616 SDLoc(), getPointerTy()),
10623 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10624 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10626 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10627 // global base reg.
10628 unsigned char OpFlag = 0;
10629 unsigned WrapperKind = X86ISD::Wrapper;
10630 CodeModel::Model M = DAG.getTarget().getCodeModel();
10632 if (Subtarget->isPICStyleRIPRel() &&
10633 (M == CodeModel::Small || M == CodeModel::Kernel))
10634 WrapperKind = X86ISD::WrapperRIP;
10635 else if (Subtarget->isPICStyleGOT())
10636 OpFlag = X86II::MO_GOTOFF;
10637 else if (Subtarget->isPICStyleStubPIC())
10638 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10640 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10643 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10645 // With PIC, the address is actually $g + Offset.
10647 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10648 DAG.getNode(X86ISD::GlobalBaseReg,
10649 SDLoc(), getPointerTy()),
10656 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10657 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10659 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10660 // global base reg.
10661 unsigned char OpFlag = 0;
10662 unsigned WrapperKind = X86ISD::Wrapper;
10663 CodeModel::Model M = DAG.getTarget().getCodeModel();
10665 if (Subtarget->isPICStyleRIPRel() &&
10666 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10667 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10668 OpFlag = X86II::MO_GOTPCREL;
10669 WrapperKind = X86ISD::WrapperRIP;
10670 } else if (Subtarget->isPICStyleGOT()) {
10671 OpFlag = X86II::MO_GOT;
10672 } else if (Subtarget->isPICStyleStubPIC()) {
10673 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10674 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10675 OpFlag = X86II::MO_DARWIN_NONLAZY;
10678 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
10681 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10683 // With PIC, the address is actually $g + Offset.
10684 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
10685 !Subtarget->is64Bit()) {
10686 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10687 DAG.getNode(X86ISD::GlobalBaseReg,
10688 SDLoc(), getPointerTy()),
10692 // For symbols that require a load from a stub to get the address, emit the
10694 if (isGlobalStubReference(OpFlag))
10695 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
10696 MachinePointerInfo::getGOT(), false, false, false, 0);
10702 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
10703 // Create the TargetBlockAddressAddress node.
10704 unsigned char OpFlags =
10705 Subtarget->ClassifyBlockAddressReference();
10706 CodeModel::Model M = DAG.getTarget().getCodeModel();
10707 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
10708 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
10710 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
10713 if (Subtarget->isPICStyleRIPRel() &&
10714 (M == CodeModel::Small || M == CodeModel::Kernel))
10715 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10717 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10719 // With PIC, the address is actually $g + Offset.
10720 if (isGlobalRelativeToPICBase(OpFlags)) {
10721 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10722 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10730 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
10731 int64_t Offset, SelectionDAG &DAG) const {
10732 // Create the TargetGlobalAddress node, folding in the constant
10733 // offset if it is legal.
10734 unsigned char OpFlags =
10735 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
10736 CodeModel::Model M = DAG.getTarget().getCodeModel();
10738 if (OpFlags == X86II::MO_NO_FLAG &&
10739 X86::isOffsetSuitableForCodeModel(Offset, M)) {
10740 // A direct static reference to a global.
10741 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
10744 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
10747 if (Subtarget->isPICStyleRIPRel() &&
10748 (M == CodeModel::Small || M == CodeModel::Kernel))
10749 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10751 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10753 // With PIC, the address is actually $g + Offset.
10754 if (isGlobalRelativeToPICBase(OpFlags)) {
10755 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10756 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10760 // For globals that require a load from a stub to get the address, emit the
10762 if (isGlobalStubReference(OpFlags))
10763 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
10764 MachinePointerInfo::getGOT(), false, false, false, 0);
10766 // If there was a non-zero offset that we didn't fold, create an explicit
10767 // addition for it.
10769 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
10770 DAG.getConstant(Offset, getPointerTy()));
10776 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
10777 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
10778 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
10779 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
10783 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
10784 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
10785 unsigned char OperandFlags, bool LocalDynamic = false) {
10786 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10787 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10789 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10790 GA->getValueType(0),
10794 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
10798 SDValue Ops[] = { Chain, TGA, *InFlag };
10799 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10801 SDValue Ops[] = { Chain, TGA };
10802 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10805 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
10806 MFI->setAdjustsStack(true);
10807 MFI->setHasCalls(true);
10809 SDValue Flag = Chain.getValue(1);
10810 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
10813 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
10815 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10818 SDLoc dl(GA); // ? function entry point might be better
10819 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10820 DAG.getNode(X86ISD::GlobalBaseReg,
10821 SDLoc(), PtrVT), InFlag);
10822 InFlag = Chain.getValue(1);
10824 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
10827 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
10829 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10831 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
10832 X86::RAX, X86II::MO_TLSGD);
10835 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
10841 // Get the start address of the TLS block for this module.
10842 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
10843 .getInfo<X86MachineFunctionInfo>();
10844 MFI->incNumLocalDynamicTLSAccesses();
10848 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
10849 X86II::MO_TLSLD, /*LocalDynamic=*/true);
10852 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10853 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
10854 InFlag = Chain.getValue(1);
10855 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
10856 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
10859 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
10863 unsigned char OperandFlags = X86II::MO_DTPOFF;
10864 unsigned WrapperKind = X86ISD::Wrapper;
10865 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10866 GA->getValueType(0),
10867 GA->getOffset(), OperandFlags);
10868 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10870 // Add x@dtpoff with the base.
10871 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
10874 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
10875 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10876 const EVT PtrVT, TLSModel::Model model,
10877 bool is64Bit, bool isPIC) {
10880 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
10881 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
10882 is64Bit ? 257 : 256));
10884 SDValue ThreadPointer =
10885 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
10886 MachinePointerInfo(Ptr), false, false, false, 0);
10888 unsigned char OperandFlags = 0;
10889 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
10891 unsigned WrapperKind = X86ISD::Wrapper;
10892 if (model == TLSModel::LocalExec) {
10893 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
10894 } else if (model == TLSModel::InitialExec) {
10896 OperandFlags = X86II::MO_GOTTPOFF;
10897 WrapperKind = X86ISD::WrapperRIP;
10899 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
10902 llvm_unreachable("Unexpected model");
10905 // emit "addl x@ntpoff,%eax" (local exec)
10906 // or "addl x@indntpoff,%eax" (initial exec)
10907 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
10909 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
10910 GA->getOffset(), OperandFlags);
10911 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10913 if (model == TLSModel::InitialExec) {
10914 if (isPIC && !is64Bit) {
10915 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
10916 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
10920 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
10921 MachinePointerInfo::getGOT(), false, false, false, 0);
10924 // The address of the thread local variable is the add of the thread
10925 // pointer with the offset of the variable.
10926 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
10930 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
10932 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
10933 const GlobalValue *GV = GA->getGlobal();
10935 if (Subtarget->isTargetELF()) {
10936 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
10939 case TLSModel::GeneralDynamic:
10940 if (Subtarget->is64Bit())
10941 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
10942 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
10943 case TLSModel::LocalDynamic:
10944 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
10945 Subtarget->is64Bit());
10946 case TLSModel::InitialExec:
10947 case TLSModel::LocalExec:
10948 return LowerToTLSExecModel(
10949 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
10950 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
10952 llvm_unreachable("Unknown TLS model.");
10955 if (Subtarget->isTargetDarwin()) {
10956 // Darwin only has one model of TLS. Lower to that.
10957 unsigned char OpFlag = 0;
10958 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
10959 X86ISD::WrapperRIP : X86ISD::Wrapper;
10961 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10962 // global base reg.
10963 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
10964 !Subtarget->is64Bit();
10966 OpFlag = X86II::MO_TLVP_PIC_BASE;
10968 OpFlag = X86II::MO_TLVP;
10970 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
10971 GA->getValueType(0),
10972 GA->getOffset(), OpFlag);
10973 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10975 // With PIC32, the address is actually $g + Offset.
10977 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10978 DAG.getNode(X86ISD::GlobalBaseReg,
10979 SDLoc(), getPointerTy()),
10982 // Lowering the machine isd will make sure everything is in the right
10984 SDValue Chain = DAG.getEntryNode();
10985 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10986 SDValue Args[] = { Chain, Offset };
10987 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
10989 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
10990 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10991 MFI->setAdjustsStack(true);
10993 // And our return value (tls address) is in the standard call return value
10995 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10996 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
10997 Chain.getValue(1));
11000 if (Subtarget->isTargetKnownWindowsMSVC() ||
11001 Subtarget->isTargetWindowsGNU()) {
11002 // Just use the implicit TLS architecture
11003 // Need to generate someting similar to:
11004 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11006 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11007 // mov rcx, qword [rdx+rcx*8]
11008 // mov eax, .tls$:tlsvar
11009 // [rax+rcx] contains the address
11010 // Windows 64bit: gs:0x58
11011 // Windows 32bit: fs:__tls_array
11014 SDValue Chain = DAG.getEntryNode();
11016 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11017 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11018 // use its literal value of 0x2C.
11019 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11020 ? Type::getInt8PtrTy(*DAG.getContext(),
11022 : Type::getInt32PtrTy(*DAG.getContext(),
11026 Subtarget->is64Bit()
11027 ? DAG.getIntPtrConstant(0x58)
11028 : (Subtarget->isTargetWindowsGNU()
11029 ? DAG.getIntPtrConstant(0x2C)
11030 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11032 SDValue ThreadPointer =
11033 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11034 MachinePointerInfo(Ptr), false, false, false, 0);
11036 // Load the _tls_index variable
11037 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11038 if (Subtarget->is64Bit())
11039 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
11040 IDX, MachinePointerInfo(), MVT::i32,
11041 false, false, false, 0);
11043 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11044 false, false, false, 0);
11046 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
11048 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11050 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11051 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11052 false, false, false, 0);
11054 // Get the offset of start of .tls section
11055 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11056 GA->getValueType(0),
11057 GA->getOffset(), X86II::MO_SECREL);
11058 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11060 // The address of the thread local variable is the add of the thread
11061 // pointer with the offset of the variable.
11062 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11065 llvm_unreachable("TLS not implemented for this target.");
11068 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11069 /// and take a 2 x i32 value to shift plus a shift amount.
11070 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11071 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11072 MVT VT = Op.getSimpleValueType();
11073 unsigned VTBits = VT.getSizeInBits();
11075 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11076 SDValue ShOpLo = Op.getOperand(0);
11077 SDValue ShOpHi = Op.getOperand(1);
11078 SDValue ShAmt = Op.getOperand(2);
11079 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11080 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11082 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11083 DAG.getConstant(VTBits - 1, MVT::i8));
11084 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11085 DAG.getConstant(VTBits - 1, MVT::i8))
11086 : DAG.getConstant(0, VT);
11088 SDValue Tmp2, Tmp3;
11089 if (Op.getOpcode() == ISD::SHL_PARTS) {
11090 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11091 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11093 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11094 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11097 // If the shift amount is larger or equal than the width of a part we can't
11098 // rely on the results of shld/shrd. Insert a test and select the appropriate
11099 // values for large shift amounts.
11100 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11101 DAG.getConstant(VTBits, MVT::i8));
11102 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11103 AndNode, DAG.getConstant(0, MVT::i8));
11106 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
11107 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11108 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11110 if (Op.getOpcode() == ISD::SHL_PARTS) {
11111 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11112 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11114 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11115 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11118 SDValue Ops[2] = { Lo, Hi };
11119 return DAG.getMergeValues(Ops, dl);
11122 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11123 SelectionDAG &DAG) const {
11124 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11127 if (SrcVT.isVector()) {
11128 if (SrcVT.getVectorElementType() == MVT::i1) {
11129 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11130 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11131 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11132 Op.getOperand(0)));
11137 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11138 "Unknown SINT_TO_FP to lower!");
11140 // These are really Legal; return the operand so the caller accepts it as
11142 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11144 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11145 Subtarget->is64Bit()) {
11149 unsigned Size = SrcVT.getSizeInBits()/8;
11150 MachineFunction &MF = DAG.getMachineFunction();
11151 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11152 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11153 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11155 MachinePointerInfo::getFixedStack(SSFI),
11157 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11160 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11162 SelectionDAG &DAG) const {
11166 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11168 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11170 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11172 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11174 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11175 MachineMemOperand *MMO;
11177 int SSFI = FI->getIndex();
11179 DAG.getMachineFunction()
11180 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11181 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11183 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11184 StackSlot = StackSlot.getOperand(1);
11186 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11187 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11189 Tys, Ops, SrcVT, MMO);
11192 Chain = Result.getValue(1);
11193 SDValue InFlag = Result.getValue(2);
11195 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11196 // shouldn't be necessary except that RFP cannot be live across
11197 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11198 MachineFunction &MF = DAG.getMachineFunction();
11199 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11200 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11201 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11202 Tys = DAG.getVTList(MVT::Other);
11204 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11206 MachineMemOperand *MMO =
11207 DAG.getMachineFunction()
11208 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11209 MachineMemOperand::MOStore, SSFISize, SSFISize);
11211 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11212 Ops, Op.getValueType(), MMO);
11213 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11214 MachinePointerInfo::getFixedStack(SSFI),
11215 false, false, false, 0);
11221 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11222 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11223 SelectionDAG &DAG) const {
11224 // This algorithm is not obvious. Here it is what we're trying to output:
11227 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11228 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11230 haddpd %xmm0, %xmm0
11232 pshufd $0x4e, %xmm0, %xmm1
11238 LLVMContext *Context = DAG.getContext();
11240 // Build some magic constants.
11241 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11242 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11243 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11245 SmallVector<Constant*,2> CV1;
11247 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11248 APInt(64, 0x4330000000000000ULL))));
11250 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11251 APInt(64, 0x4530000000000000ULL))));
11252 Constant *C1 = ConstantVector::get(CV1);
11253 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11255 // Load the 64-bit value into an XMM register.
11256 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11258 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11259 MachinePointerInfo::getConstantPool(),
11260 false, false, false, 16);
11261 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11262 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11265 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11266 MachinePointerInfo::getConstantPool(),
11267 false, false, false, 16);
11268 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11269 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11272 if (Subtarget->hasSSE3()) {
11273 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11274 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11276 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11277 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11279 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11280 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11284 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11285 DAG.getIntPtrConstant(0));
11288 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11289 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11290 SelectionDAG &DAG) const {
11292 // FP constant to bias correct the final result.
11293 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
11296 // Load the 32-bit value into an XMM register.
11297 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11300 // Zero out the upper parts of the register.
11301 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11303 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11304 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11305 DAG.getIntPtrConstant(0));
11307 // Or the load with the bias.
11308 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11309 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11310 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11311 MVT::v2f64, Load)),
11312 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11313 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11314 MVT::v2f64, Bias)));
11315 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11316 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11317 DAG.getIntPtrConstant(0));
11319 // Subtract the bias.
11320 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11322 // Handle final rounding.
11323 EVT DestVT = Op.getValueType();
11325 if (DestVT.bitsLT(MVT::f64))
11326 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11327 DAG.getIntPtrConstant(0));
11328 if (DestVT.bitsGT(MVT::f64))
11329 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11331 // Handle final rounding.
11335 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11336 const X86Subtarget &Subtarget) {
11337 // The algorithm is the following:
11338 // #ifdef __SSE4_1__
11339 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11340 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11341 // (uint4) 0x53000000, 0xaa);
11343 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11344 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11346 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11347 // return (float4) lo + fhi;
11350 SDValue V = Op->getOperand(0);
11351 EVT VecIntVT = V.getValueType();
11352 bool Is128 = VecIntVT == MVT::v4i32;
11353 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11354 // If we convert to something else than the supported type, e.g., to v4f64,
11356 if (VecFloatVT != Op->getValueType(0))
11359 unsigned NumElts = VecIntVT.getVectorNumElements();
11360 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11361 "Unsupported custom type");
11362 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11364 // In the #idef/#else code, we have in common:
11365 // - The vector of constants:
11371 // Create the splat vector for 0x4b000000.
11372 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
11373 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11374 CstLow, CstLow, CstLow, CstLow};
11375 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11376 makeArrayRef(&CstLowArray[0], NumElts));
11377 // Create the splat vector for 0x53000000.
11378 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
11379 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11380 CstHigh, CstHigh, CstHigh, CstHigh};
11381 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11382 makeArrayRef(&CstHighArray[0], NumElts));
11384 // Create the right shift.
11385 SDValue CstShift = DAG.getConstant(16, MVT::i32);
11386 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11387 CstShift, CstShift, CstShift, CstShift};
11388 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11389 makeArrayRef(&CstShiftArray[0], NumElts));
11390 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11393 if (Subtarget.hasSSE41()) {
11394 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11395 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11396 SDValue VecCstLowBitcast =
11397 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11398 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11399 // Low will be bitcasted right away, so do not bother bitcasting back to its
11401 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11402 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
11403 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11404 // (uint4) 0x53000000, 0xaa);
11405 SDValue VecCstHighBitcast =
11406 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11407 SDValue VecShiftBitcast =
11408 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11409 // High will be bitcasted right away, so do not bother bitcasting back to
11410 // its original type.
11411 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11412 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
11414 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
11415 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11416 CstMask, CstMask, CstMask);
11417 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11418 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11419 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11421 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11422 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11425 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11426 SDValue CstFAdd = DAG.getConstantFP(
11427 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
11428 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11429 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11430 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11431 makeArrayRef(&CstFAddArray[0], NumElts));
11433 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11434 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11436 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11437 // return (float4) lo + fhi;
11438 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11439 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11442 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11443 SelectionDAG &DAG) const {
11444 SDValue N0 = Op.getOperand(0);
11445 MVT SVT = N0.getSimpleValueType();
11448 switch (SVT.SimpleTy) {
11450 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11455 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11456 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11457 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11461 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11463 llvm_unreachable(nullptr);
11466 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11467 SelectionDAG &DAG) const {
11468 SDValue N0 = Op.getOperand(0);
11471 if (Op.getValueType().isVector())
11472 return lowerUINT_TO_FP_vec(Op, DAG);
11474 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11475 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11476 // the optimization here.
11477 if (DAG.SignBitIsZero(N0))
11478 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11480 MVT SrcVT = N0.getSimpleValueType();
11481 MVT DstVT = Op.getSimpleValueType();
11482 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11483 return LowerUINT_TO_FP_i64(Op, DAG);
11484 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11485 return LowerUINT_TO_FP_i32(Op, DAG);
11486 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11489 // Make a 64-bit buffer, and use it to build an FILD.
11490 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11491 if (SrcVT == MVT::i32) {
11492 SDValue WordOff = DAG.getConstant(4, getPointerTy());
11493 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11494 getPointerTy(), StackSlot, WordOff);
11495 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11496 StackSlot, MachinePointerInfo(),
11498 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
11499 OffsetSlot, MachinePointerInfo(),
11501 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11505 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11506 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11507 StackSlot, MachinePointerInfo(),
11509 // For i64 source, we need to add the appropriate power of 2 if the input
11510 // was negative. This is the same as the optimization in
11511 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11512 // we must be careful to do the computation in x87 extended precision, not
11513 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11514 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11515 MachineMemOperand *MMO =
11516 DAG.getMachineFunction()
11517 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11518 MachineMemOperand::MOLoad, 8, 8);
11520 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11521 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11522 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11525 APInt FF(32, 0x5F800000ULL);
11527 // Check whether the sign bit is set.
11528 SDValue SignSet = DAG.getSetCC(dl,
11529 getSetCCResultType(*DAG.getContext(), MVT::i64),
11530 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
11533 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11534 SDValue FudgePtr = DAG.getConstantPool(
11535 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11538 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11539 SDValue Zero = DAG.getIntPtrConstant(0);
11540 SDValue Four = DAG.getIntPtrConstant(4);
11541 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11543 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11545 // Load the value out, extending it from f32 to f80.
11546 // FIXME: Avoid the extend by constructing the right constant pool?
11547 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11548 FudgePtr, MachinePointerInfo::getConstantPool(),
11549 MVT::f32, false, false, false, 4);
11550 // Extend everything to 80 bits to force it to be done on x87.
11551 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11552 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
11555 std::pair<SDValue,SDValue>
11556 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11557 bool IsSigned, bool IsReplace) const {
11560 EVT DstTy = Op.getValueType();
11562 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11563 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11567 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11568 DstTy.getSimpleVT() >= MVT::i16 &&
11569 "Unknown FP_TO_INT to lower!");
11571 // These are really Legal.
11572 if (DstTy == MVT::i32 &&
11573 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11574 return std::make_pair(SDValue(), SDValue());
11575 if (Subtarget->is64Bit() &&
11576 DstTy == MVT::i64 &&
11577 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11578 return std::make_pair(SDValue(), SDValue());
11580 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11581 // stack slot, or into the FTOL runtime function.
11582 MachineFunction &MF = DAG.getMachineFunction();
11583 unsigned MemSize = DstTy.getSizeInBits()/8;
11584 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11585 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11588 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11589 Opc = X86ISD::WIN_FTOL;
11591 switch (DstTy.getSimpleVT().SimpleTy) {
11592 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11593 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11594 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11595 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11598 SDValue Chain = DAG.getEntryNode();
11599 SDValue Value = Op.getOperand(0);
11600 EVT TheVT = Op.getOperand(0).getValueType();
11601 // FIXME This causes a redundant load/store if the SSE-class value is already
11602 // in memory, such as if it is on the callstack.
11603 if (isScalarFPTypeInSSEReg(TheVT)) {
11604 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11605 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11606 MachinePointerInfo::getFixedStack(SSFI),
11608 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11610 Chain, StackSlot, DAG.getValueType(TheVT)
11613 MachineMemOperand *MMO =
11614 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11615 MachineMemOperand::MOLoad, MemSize, MemSize);
11616 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11617 Chain = Value.getValue(1);
11618 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11619 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11622 MachineMemOperand *MMO =
11623 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11624 MachineMemOperand::MOStore, MemSize, MemSize);
11626 if (Opc != X86ISD::WIN_FTOL) {
11627 // Build the FP_TO_INT*_IN_MEM
11628 SDValue Ops[] = { Chain, Value, StackSlot };
11629 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11631 return std::make_pair(FIST, StackSlot);
11633 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11634 DAG.getVTList(MVT::Other, MVT::Glue),
11636 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11637 MVT::i32, ftol.getValue(1));
11638 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11639 MVT::i32, eax.getValue(2));
11640 SDValue Ops[] = { eax, edx };
11641 SDValue pair = IsReplace
11642 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11643 : DAG.getMergeValues(Ops, DL);
11644 return std::make_pair(pair, SDValue());
11648 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11649 const X86Subtarget *Subtarget) {
11650 MVT VT = Op->getSimpleValueType(0);
11651 SDValue In = Op->getOperand(0);
11652 MVT InVT = In.getSimpleValueType();
11655 // Optimize vectors in AVX mode:
11658 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11659 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11660 // Concat upper and lower parts.
11663 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11664 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11665 // Concat upper and lower parts.
11668 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11669 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11670 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
11673 if (Subtarget->hasInt256())
11674 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
11676 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
11677 SDValue Undef = DAG.getUNDEF(InVT);
11678 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
11679 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11680 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11682 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
11683 VT.getVectorNumElements()/2);
11685 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
11686 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
11688 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
11691 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
11692 SelectionDAG &DAG) {
11693 MVT VT = Op->getSimpleValueType(0);
11694 SDValue In = Op->getOperand(0);
11695 MVT InVT = In.getSimpleValueType();
11697 unsigned int NumElts = VT.getVectorNumElements();
11698 if (NumElts != 8 && NumElts != 16)
11701 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
11702 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
11704 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
11705 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11706 // Now we have only mask extension
11707 assert(InVT.getVectorElementType() == MVT::i1);
11708 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
11709 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11710 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
11711 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11712 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11713 MachinePointerInfo::getConstantPool(),
11714 false, false, false, Alignment);
11716 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
11717 if (VT.is512BitVector())
11719 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
11722 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11723 SelectionDAG &DAG) {
11724 if (Subtarget->hasFp256()) {
11725 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11733 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11734 SelectionDAG &DAG) {
11736 MVT VT = Op.getSimpleValueType();
11737 SDValue In = Op.getOperand(0);
11738 MVT SVT = In.getSimpleValueType();
11740 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
11741 return LowerZERO_EXTEND_AVX512(Op, DAG);
11743 if (Subtarget->hasFp256()) {
11744 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11749 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
11750 VT.getVectorNumElements() != SVT.getVectorNumElements());
11754 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
11756 MVT VT = Op.getSimpleValueType();
11757 SDValue In = Op.getOperand(0);
11758 MVT InVT = In.getSimpleValueType();
11760 if (VT == MVT::i1) {
11761 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
11762 "Invalid scalar TRUNCATE operation");
11763 if (InVT.getSizeInBits() >= 32)
11765 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
11766 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
11768 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
11769 "Invalid TRUNCATE operation");
11771 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
11772 if (VT.getVectorElementType().getSizeInBits() >=8)
11773 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
11775 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
11776 unsigned NumElts = InVT.getVectorNumElements();
11777 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
11778 if (InVT.getSizeInBits() < 512) {
11779 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
11780 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
11784 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
11785 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11786 SDValue CP = DAG.getConstantPool(C, getPointerTy());
11787 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11788 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11789 MachinePointerInfo::getConstantPool(),
11790 false, false, false, Alignment);
11791 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
11792 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
11793 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
11796 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
11797 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
11798 if (Subtarget->hasInt256()) {
11799 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
11800 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
11801 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
11803 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
11804 DAG.getIntPtrConstant(0));
11807 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11808 DAG.getIntPtrConstant(0));
11809 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11810 DAG.getIntPtrConstant(2));
11811 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11812 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11813 static const int ShufMask[] = {0, 2, 4, 6};
11814 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
11817 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
11818 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
11819 if (Subtarget->hasInt256()) {
11820 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
11822 SmallVector<SDValue,32> pshufbMask;
11823 for (unsigned i = 0; i < 2; ++i) {
11824 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
11825 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
11826 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
11827 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
11828 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
11829 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
11830 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
11831 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
11832 for (unsigned j = 0; j < 8; ++j)
11833 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
11835 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
11836 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
11837 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
11839 static const int ShufMask[] = {0, 2, -1, -1};
11840 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
11842 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11843 DAG.getIntPtrConstant(0));
11844 return DAG.getNode(ISD::BITCAST, DL, VT, In);
11847 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11848 DAG.getIntPtrConstant(0));
11850 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11851 DAG.getIntPtrConstant(4));
11853 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
11854 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
11856 // The PSHUFB mask:
11857 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
11858 -1, -1, -1, -1, -1, -1, -1, -1};
11860 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
11861 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
11862 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
11864 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11865 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11867 // The MOVLHPS Mask:
11868 static const int ShufMask2[] = {0, 1, 4, 5};
11869 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
11870 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
11873 // Handle truncation of V256 to V128 using shuffles.
11874 if (!VT.is128BitVector() || !InVT.is256BitVector())
11877 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
11879 unsigned NumElems = VT.getVectorNumElements();
11880 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
11882 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
11883 // Prepare truncation shuffle mask
11884 for (unsigned i = 0; i != NumElems; ++i)
11885 MaskVec[i] = i * 2;
11886 SDValue V = DAG.getVectorShuffle(NVT, DL,
11887 DAG.getNode(ISD::BITCAST, DL, NVT, In),
11888 DAG.getUNDEF(NVT), &MaskVec[0]);
11889 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
11890 DAG.getIntPtrConstant(0));
11893 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
11894 SelectionDAG &DAG) const {
11895 assert(!Op.getSimpleValueType().isVector());
11897 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11898 /*IsSigned=*/ true, /*IsReplace=*/ false);
11899 SDValue FIST = Vals.first, StackSlot = Vals.second;
11900 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
11901 if (!FIST.getNode()) return Op;
11903 if (StackSlot.getNode())
11904 // Load the result.
11905 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11906 FIST, StackSlot, MachinePointerInfo(),
11907 false, false, false, 0);
11909 // The node is the result.
11913 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
11914 SelectionDAG &DAG) const {
11915 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11916 /*IsSigned=*/ false, /*IsReplace=*/ false);
11917 SDValue FIST = Vals.first, StackSlot = Vals.second;
11918 assert(FIST.getNode() && "Unexpected failure");
11920 if (StackSlot.getNode())
11921 // Load the result.
11922 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11923 FIST, StackSlot, MachinePointerInfo(),
11924 false, false, false, 0);
11926 // The node is the result.
11930 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
11932 MVT VT = Op.getSimpleValueType();
11933 SDValue In = Op.getOperand(0);
11934 MVT SVT = In.getSimpleValueType();
11936 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
11938 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
11939 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
11940 In, DAG.getUNDEF(SVT)));
11943 /// The only differences between FABS and FNEG are the mask and the logic op.
11944 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
11945 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
11946 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
11947 "Wrong opcode for lowering FABS or FNEG.");
11949 bool IsFABS = (Op.getOpcode() == ISD::FABS);
11951 // If this is a FABS and it has an FNEG user, bail out to fold the combination
11952 // into an FNABS. We'll lower the FABS after that if it is still in use.
11954 for (SDNode *User : Op->uses())
11955 if (User->getOpcode() == ISD::FNEG)
11958 SDValue Op0 = Op.getOperand(0);
11959 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
11962 MVT VT = Op.getSimpleValueType();
11963 // Assume scalar op for initialization; update for vector if needed.
11964 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
11965 // generate a 16-byte vector constant and logic op even for the scalar case.
11966 // Using a 16-byte mask allows folding the load of the mask with
11967 // the logic op, so it can save (~4 bytes) on code size.
11969 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
11970 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
11971 // decide if we should generate a 16-byte constant mask when we only need 4 or
11972 // 8 bytes for the scalar case.
11973 if (VT.isVector()) {
11974 EltVT = VT.getVectorElementType();
11975 NumElts = VT.getVectorNumElements();
11978 unsigned EltBits = EltVT.getSizeInBits();
11979 LLVMContext *Context = DAG.getContext();
11980 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
11982 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
11983 Constant *C = ConstantInt::get(*Context, MaskElt);
11984 C = ConstantVector::getSplat(NumElts, C);
11985 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11986 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
11987 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11988 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11989 MachinePointerInfo::getConstantPool(),
11990 false, false, false, Alignment);
11992 if (VT.isVector()) {
11993 // For a vector, cast operands to a vector type, perform the logic op,
11994 // and cast the result back to the original value type.
11995 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
11996 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
11997 SDValue Operand = IsFNABS ?
11998 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
11999 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12000 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12001 return DAG.getNode(ISD::BITCAST, dl, VT,
12002 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12005 // If not vector, then scalar.
12006 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12007 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12008 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12011 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12012 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12013 LLVMContext *Context = DAG.getContext();
12014 SDValue Op0 = Op.getOperand(0);
12015 SDValue Op1 = Op.getOperand(1);
12017 MVT VT = Op.getSimpleValueType();
12018 MVT SrcVT = Op1.getSimpleValueType();
12020 // If second operand is smaller, extend it first.
12021 if (SrcVT.bitsLT(VT)) {
12022 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12025 // And if it is bigger, shrink it first.
12026 if (SrcVT.bitsGT(VT)) {
12027 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
12031 // At this point the operands and the result should have the same
12032 // type, and that won't be f80 since that is not custom lowered.
12034 const fltSemantics &Sem =
12035 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12036 const unsigned SizeInBits = VT.getSizeInBits();
12038 SmallVector<Constant *, 4> CV(
12039 VT == MVT::f64 ? 2 : 4,
12040 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12042 // First, clear all bits but the sign bit from the second operand (sign).
12043 CV[0] = ConstantFP::get(*Context,
12044 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12045 Constant *C = ConstantVector::get(CV);
12046 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12047 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12048 MachinePointerInfo::getConstantPool(),
12049 false, false, false, 16);
12050 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12052 // Next, clear the sign bit from the first operand (magnitude).
12053 // If it's a constant, we can clear it here.
12054 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12055 APFloat APF = Op0CN->getValueAPF();
12056 // If the magnitude is a positive zero, the sign bit alone is enough.
12057 if (APF.isPosZero())
12060 CV[0] = ConstantFP::get(*Context, APF);
12062 CV[0] = ConstantFP::get(
12064 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12066 C = ConstantVector::get(CV);
12067 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12068 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12069 MachinePointerInfo::getConstantPool(),
12070 false, false, false, 16);
12071 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12072 if (!isa<ConstantFPSDNode>(Op0))
12073 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12075 // OR the magnitude value with the sign bit.
12076 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12079 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12080 SDValue N0 = Op.getOperand(0);
12082 MVT VT = Op.getSimpleValueType();
12084 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12085 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12086 DAG.getConstant(1, VT));
12087 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
12090 // Check whether an OR'd tree is PTEST-able.
12091 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12092 SelectionDAG &DAG) {
12093 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12095 if (!Subtarget->hasSSE41())
12098 if (!Op->hasOneUse())
12101 SDNode *N = Op.getNode();
12104 SmallVector<SDValue, 8> Opnds;
12105 DenseMap<SDValue, unsigned> VecInMap;
12106 SmallVector<SDValue, 8> VecIns;
12107 EVT VT = MVT::Other;
12109 // Recognize a special case where a vector is casted into wide integer to
12111 Opnds.push_back(N->getOperand(0));
12112 Opnds.push_back(N->getOperand(1));
12114 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12115 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12116 // BFS traverse all OR'd operands.
12117 if (I->getOpcode() == ISD::OR) {
12118 Opnds.push_back(I->getOperand(0));
12119 Opnds.push_back(I->getOperand(1));
12120 // Re-evaluate the number of nodes to be traversed.
12121 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12125 // Quit if a non-EXTRACT_VECTOR_ELT
12126 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12129 // Quit if without a constant index.
12130 SDValue Idx = I->getOperand(1);
12131 if (!isa<ConstantSDNode>(Idx))
12134 SDValue ExtractedFromVec = I->getOperand(0);
12135 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12136 if (M == VecInMap.end()) {
12137 VT = ExtractedFromVec.getValueType();
12138 // Quit if not 128/256-bit vector.
12139 if (!VT.is128BitVector() && !VT.is256BitVector())
12141 // Quit if not the same type.
12142 if (VecInMap.begin() != VecInMap.end() &&
12143 VT != VecInMap.begin()->first.getValueType())
12145 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12146 VecIns.push_back(ExtractedFromVec);
12148 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12151 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12152 "Not extracted from 128-/256-bit vector.");
12154 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12156 for (DenseMap<SDValue, unsigned>::const_iterator
12157 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12158 // Quit if not all elements are used.
12159 if (I->second != FullMask)
12163 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12165 // Cast all vectors into TestVT for PTEST.
12166 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12167 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12169 // If more than one full vectors are evaluated, OR them first before PTEST.
12170 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12171 // Each iteration will OR 2 nodes and append the result until there is only
12172 // 1 node left, i.e. the final OR'd value of all vectors.
12173 SDValue LHS = VecIns[Slot];
12174 SDValue RHS = VecIns[Slot + 1];
12175 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12178 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12179 VecIns.back(), VecIns.back());
12182 /// \brief return true if \c Op has a use that doesn't just read flags.
12183 static bool hasNonFlagsUse(SDValue Op) {
12184 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12186 SDNode *User = *UI;
12187 unsigned UOpNo = UI.getOperandNo();
12188 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12189 // Look pass truncate.
12190 UOpNo = User->use_begin().getOperandNo();
12191 User = *User->use_begin();
12194 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12195 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12201 /// Emit nodes that will be selected as "test Op0,Op0", or something
12203 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12204 SelectionDAG &DAG) const {
12205 if (Op.getValueType() == MVT::i1) {
12206 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12207 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12208 DAG.getConstant(0, MVT::i8));
12210 // CF and OF aren't always set the way we want. Determine which
12211 // of these we need.
12212 bool NeedCF = false;
12213 bool NeedOF = false;
12216 case X86::COND_A: case X86::COND_AE:
12217 case X86::COND_B: case X86::COND_BE:
12220 case X86::COND_G: case X86::COND_GE:
12221 case X86::COND_L: case X86::COND_LE:
12222 case X86::COND_O: case X86::COND_NO: {
12223 // Check if we really need to set the
12224 // Overflow flag. If NoSignedWrap is present
12225 // that is not actually needed.
12226 switch (Op->getOpcode()) {
12231 const BinaryWithFlagsSDNode *BinNode =
12232 cast<BinaryWithFlagsSDNode>(Op.getNode());
12233 if (BinNode->hasNoSignedWrap())
12243 // See if we can use the EFLAGS value from the operand instead of
12244 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12245 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12246 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12247 // Emit a CMP with 0, which is the TEST pattern.
12248 //if (Op.getValueType() == MVT::i1)
12249 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12250 // DAG.getConstant(0, MVT::i1));
12251 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12252 DAG.getConstant(0, Op.getValueType()));
12254 unsigned Opcode = 0;
12255 unsigned NumOperands = 0;
12257 // Truncate operations may prevent the merge of the SETCC instruction
12258 // and the arithmetic instruction before it. Attempt to truncate the operands
12259 // of the arithmetic instruction and use a reduced bit-width instruction.
12260 bool NeedTruncation = false;
12261 SDValue ArithOp = Op;
12262 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12263 SDValue Arith = Op->getOperand(0);
12264 // Both the trunc and the arithmetic op need to have one user each.
12265 if (Arith->hasOneUse())
12266 switch (Arith.getOpcode()) {
12273 NeedTruncation = true;
12279 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12280 // which may be the result of a CAST. We use the variable 'Op', which is the
12281 // non-casted variable when we check for possible users.
12282 switch (ArithOp.getOpcode()) {
12284 // Due to an isel shortcoming, be conservative if this add is likely to be
12285 // selected as part of a load-modify-store instruction. When the root node
12286 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12287 // uses of other nodes in the match, such as the ADD in this case. This
12288 // leads to the ADD being left around and reselected, with the result being
12289 // two adds in the output. Alas, even if none our users are stores, that
12290 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12291 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12292 // climbing the DAG back to the root, and it doesn't seem to be worth the
12294 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12295 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12296 if (UI->getOpcode() != ISD::CopyToReg &&
12297 UI->getOpcode() != ISD::SETCC &&
12298 UI->getOpcode() != ISD::STORE)
12301 if (ConstantSDNode *C =
12302 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12303 // An add of one will be selected as an INC.
12304 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12305 Opcode = X86ISD::INC;
12310 // An add of negative one (subtract of one) will be selected as a DEC.
12311 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12312 Opcode = X86ISD::DEC;
12318 // Otherwise use a regular EFLAGS-setting add.
12319 Opcode = X86ISD::ADD;
12324 // If we have a constant logical shift that's only used in a comparison
12325 // against zero turn it into an equivalent AND. This allows turning it into
12326 // a TEST instruction later.
12327 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12328 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12329 EVT VT = Op.getValueType();
12330 unsigned BitWidth = VT.getSizeInBits();
12331 unsigned ShAmt = Op->getConstantOperandVal(1);
12332 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12334 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12335 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12336 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12337 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12339 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12340 DAG.getConstant(Mask, VT));
12341 DAG.ReplaceAllUsesWith(Op, New);
12347 // If the primary and result isn't used, don't bother using X86ISD::AND,
12348 // because a TEST instruction will be better.
12349 if (!hasNonFlagsUse(Op))
12355 // Due to the ISEL shortcoming noted above, be conservative if this op is
12356 // likely to be selected as part of a load-modify-store instruction.
12357 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12358 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12359 if (UI->getOpcode() == ISD::STORE)
12362 // Otherwise use a regular EFLAGS-setting instruction.
12363 switch (ArithOp.getOpcode()) {
12364 default: llvm_unreachable("unexpected operator!");
12365 case ISD::SUB: Opcode = X86ISD::SUB; break;
12366 case ISD::XOR: Opcode = X86ISD::XOR; break;
12367 case ISD::AND: Opcode = X86ISD::AND; break;
12369 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12370 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12371 if (EFLAGS.getNode())
12374 Opcode = X86ISD::OR;
12388 return SDValue(Op.getNode(), 1);
12394 // If we found that truncation is beneficial, perform the truncation and
12396 if (NeedTruncation) {
12397 EVT VT = Op.getValueType();
12398 SDValue WideVal = Op->getOperand(0);
12399 EVT WideVT = WideVal.getValueType();
12400 unsigned ConvertedOp = 0;
12401 // Use a target machine opcode to prevent further DAGCombine
12402 // optimizations that may separate the arithmetic operations
12403 // from the setcc node.
12404 switch (WideVal.getOpcode()) {
12406 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12407 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12408 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12409 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12410 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12414 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12415 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12416 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12417 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12418 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12424 // Emit a CMP with 0, which is the TEST pattern.
12425 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12426 DAG.getConstant(0, Op.getValueType()));
12428 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12429 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12431 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12432 DAG.ReplaceAllUsesWith(Op, New);
12433 return SDValue(New.getNode(), 1);
12436 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12438 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12439 SDLoc dl, SelectionDAG &DAG) const {
12440 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12441 if (C->getAPIntValue() == 0)
12442 return EmitTest(Op0, X86CC, dl, DAG);
12444 if (Op0.getValueType() == MVT::i1)
12445 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12448 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12449 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12450 // Do the comparison at i32 if it's smaller, besides the Atom case.
12451 // This avoids subregister aliasing issues. Keep the smaller reference
12452 // if we're optimizing for size, however, as that'll allow better folding
12453 // of memory operations.
12454 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12455 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12456 Attribute::MinSize) &&
12457 !Subtarget->isAtom()) {
12458 unsigned ExtendOp =
12459 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12460 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12461 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12463 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12464 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12465 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12467 return SDValue(Sub.getNode(), 1);
12469 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12472 /// Convert a comparison if required by the subtarget.
12473 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12474 SelectionDAG &DAG) const {
12475 // If the subtarget does not support the FUCOMI instruction, floating-point
12476 // comparisons have to be converted.
12477 if (Subtarget->hasCMov() ||
12478 Cmp.getOpcode() != X86ISD::CMP ||
12479 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12480 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12483 // The instruction selector will select an FUCOM instruction instead of
12484 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12485 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12486 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12488 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12489 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12490 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12491 DAG.getConstant(8, MVT::i8));
12492 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12493 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12496 /// The minimum architected relative accuracy is 2^-12. We need one
12497 /// Newton-Raphson step to have a good float result (24 bits of precision).
12498 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12499 DAGCombinerInfo &DCI,
12500 unsigned &RefinementSteps,
12501 bool &UseOneConstNR) const {
12502 // FIXME: We should use instruction latency models to calculate the cost of
12503 // each potential sequence, but this is very hard to do reliably because
12504 // at least Intel's Core* chips have variable timing based on the number of
12505 // significant digits in the divisor and/or sqrt operand.
12506 if (!Subtarget->useSqrtEst())
12509 EVT VT = Op.getValueType();
12511 // SSE1 has rsqrtss and rsqrtps.
12512 // TODO: Add support for AVX512 (v16f32).
12513 // It is likely not profitable to do this for f64 because a double-precision
12514 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12515 // instructions: convert to single, rsqrtss, convert back to double, refine
12516 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12517 // along with FMA, this could be a throughput win.
12518 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12519 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12520 RefinementSteps = 1;
12521 UseOneConstNR = false;
12522 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12527 /// The minimum architected relative accuracy is 2^-12. We need one
12528 /// Newton-Raphson step to have a good float result (24 bits of precision).
12529 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12530 DAGCombinerInfo &DCI,
12531 unsigned &RefinementSteps) const {
12532 // FIXME: We should use instruction latency models to calculate the cost of
12533 // each potential sequence, but this is very hard to do reliably because
12534 // at least Intel's Core* chips have variable timing based on the number of
12535 // significant digits in the divisor.
12536 if (!Subtarget->useReciprocalEst())
12539 EVT VT = Op.getValueType();
12541 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12542 // TODO: Add support for AVX512 (v16f32).
12543 // It is likely not profitable to do this for f64 because a double-precision
12544 // reciprocal estimate with refinement on x86 prior to FMA requires
12545 // 15 instructions: convert to single, rcpss, convert back to double, refine
12546 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12547 // along with FMA, this could be a throughput win.
12548 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12549 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12550 RefinementSteps = ReciprocalEstimateRefinementSteps;
12551 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12556 static bool isAllOnes(SDValue V) {
12557 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12558 return C && C->isAllOnesValue();
12561 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12562 /// if it's possible.
12563 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12564 SDLoc dl, SelectionDAG &DAG) const {
12565 SDValue Op0 = And.getOperand(0);
12566 SDValue Op1 = And.getOperand(1);
12567 if (Op0.getOpcode() == ISD::TRUNCATE)
12568 Op0 = Op0.getOperand(0);
12569 if (Op1.getOpcode() == ISD::TRUNCATE)
12570 Op1 = Op1.getOperand(0);
12573 if (Op1.getOpcode() == ISD::SHL)
12574 std::swap(Op0, Op1);
12575 if (Op0.getOpcode() == ISD::SHL) {
12576 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12577 if (And00C->getZExtValue() == 1) {
12578 // If we looked past a truncate, check that it's only truncating away
12580 unsigned BitWidth = Op0.getValueSizeInBits();
12581 unsigned AndBitWidth = And.getValueSizeInBits();
12582 if (BitWidth > AndBitWidth) {
12584 DAG.computeKnownBits(Op0, Zeros, Ones);
12585 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12589 RHS = Op0.getOperand(1);
12591 } else if (Op1.getOpcode() == ISD::Constant) {
12592 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
12593 uint64_t AndRHSVal = AndRHS->getZExtValue();
12594 SDValue AndLHS = Op0;
12596 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
12597 LHS = AndLHS.getOperand(0);
12598 RHS = AndLHS.getOperand(1);
12601 // Use BT if the immediate can't be encoded in a TEST instruction.
12602 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
12604 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
12608 if (LHS.getNode()) {
12609 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12610 // instruction. Since the shift amount is in-range-or-undefined, we know
12611 // that doing a bittest on the i32 value is ok. We extend to i32 because
12612 // the encoding for the i16 version is larger than the i32 version.
12613 // Also promote i16 to i32 for performance / code size reason.
12614 if (LHS.getValueType() == MVT::i8 ||
12615 LHS.getValueType() == MVT::i16)
12616 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12618 // If the operand types disagree, extend the shift amount to match. Since
12619 // BT ignores high bits (like shifts) we can use anyextend.
12620 if (LHS.getValueType() != RHS.getValueType())
12621 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12623 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12624 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12625 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12626 DAG.getConstant(Cond, MVT::i8), BT);
12632 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12634 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12639 // SSE Condition code mapping:
12648 switch (SetCCOpcode) {
12649 default: llvm_unreachable("Unexpected SETCC condition");
12651 case ISD::SETEQ: SSECC = 0; break;
12653 case ISD::SETGT: Swap = true; // Fallthrough
12655 case ISD::SETOLT: SSECC = 1; break;
12657 case ISD::SETGE: Swap = true; // Fallthrough
12659 case ISD::SETOLE: SSECC = 2; break;
12660 case ISD::SETUO: SSECC = 3; break;
12662 case ISD::SETNE: SSECC = 4; break;
12663 case ISD::SETULE: Swap = true; // Fallthrough
12664 case ISD::SETUGE: SSECC = 5; break;
12665 case ISD::SETULT: Swap = true; // Fallthrough
12666 case ISD::SETUGT: SSECC = 6; break;
12667 case ISD::SETO: SSECC = 7; break;
12669 case ISD::SETONE: SSECC = 8; break;
12672 std::swap(Op0, Op1);
12677 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
12678 // ones, and then concatenate the result back.
12679 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
12680 MVT VT = Op.getSimpleValueType();
12682 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
12683 "Unsupported value type for operation");
12685 unsigned NumElems = VT.getVectorNumElements();
12687 SDValue CC = Op.getOperand(2);
12689 // Extract the LHS vectors
12690 SDValue LHS = Op.getOperand(0);
12691 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12692 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12694 // Extract the RHS vectors
12695 SDValue RHS = Op.getOperand(1);
12696 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
12697 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
12699 // Issue the operation on the smaller types and concatenate the result back
12700 MVT EltVT = VT.getVectorElementType();
12701 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12702 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
12703 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
12704 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
12707 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
12708 const X86Subtarget *Subtarget) {
12709 SDValue Op0 = Op.getOperand(0);
12710 SDValue Op1 = Op.getOperand(1);
12711 SDValue CC = Op.getOperand(2);
12712 MVT VT = Op.getSimpleValueType();
12715 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
12716 Op.getValueType().getScalarType() == MVT::i1 &&
12717 "Cannot set masked compare for this operation");
12719 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12721 bool Unsigned = false;
12724 switch (SetCCOpcode) {
12725 default: llvm_unreachable("Unexpected SETCC condition");
12726 case ISD::SETNE: SSECC = 4; break;
12727 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
12728 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
12729 case ISD::SETLT: Swap = true; //fall-through
12730 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
12731 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
12732 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
12733 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
12734 case ISD::SETULE: Unsigned = true; //fall-through
12735 case ISD::SETLE: SSECC = 2; break;
12739 std::swap(Op0, Op1);
12741 return DAG.getNode(Opc, dl, VT, Op0, Op1);
12742 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
12743 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12744 DAG.getConstant(SSECC, MVT::i8));
12747 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
12748 /// operand \p Op1. If non-trivial (for example because it's not constant)
12749 /// return an empty value.
12750 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
12752 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
12756 MVT VT = Op1.getSimpleValueType();
12757 MVT EVT = VT.getVectorElementType();
12758 unsigned n = VT.getVectorNumElements();
12759 SmallVector<SDValue, 8> ULTOp1;
12761 for (unsigned i = 0; i < n; ++i) {
12762 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
12763 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
12766 // Avoid underflow.
12767 APInt Val = Elt->getAPIntValue();
12771 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
12774 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
12777 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
12778 SelectionDAG &DAG) {
12779 SDValue Op0 = Op.getOperand(0);
12780 SDValue Op1 = Op.getOperand(1);
12781 SDValue CC = Op.getOperand(2);
12782 MVT VT = Op.getSimpleValueType();
12783 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12784 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
12789 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
12790 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
12793 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
12794 unsigned Opc = X86ISD::CMPP;
12795 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
12796 assert(VT.getVectorNumElements() <= 16);
12797 Opc = X86ISD::CMPM;
12799 // In the two special cases we can't handle, emit two comparisons.
12802 unsigned CombineOpc;
12803 if (SetCCOpcode == ISD::SETUEQ) {
12804 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
12806 assert(SetCCOpcode == ISD::SETONE);
12807 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
12810 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12811 DAG.getConstant(CC0, MVT::i8));
12812 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12813 DAG.getConstant(CC1, MVT::i8));
12814 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
12816 // Handle all other FP comparisons here.
12817 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12818 DAG.getConstant(SSECC, MVT::i8));
12821 // Break 256-bit integer vector compare into smaller ones.
12822 if (VT.is256BitVector() && !Subtarget->hasInt256())
12823 return Lower256IntVSETCC(Op, DAG);
12825 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
12826 EVT OpVT = Op1.getValueType();
12827 if (Subtarget->hasAVX512()) {
12828 if (Op1.getValueType().is512BitVector() ||
12829 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
12830 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
12831 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
12833 // In AVX-512 architecture setcc returns mask with i1 elements,
12834 // But there is no compare instruction for i8 and i16 elements in KNL.
12835 // We are not talking about 512-bit operands in this case, these
12836 // types are illegal.
12838 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
12839 OpVT.getVectorElementType().getSizeInBits() >= 8))
12840 return DAG.getNode(ISD::TRUNCATE, dl, VT,
12841 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
12844 // We are handling one of the integer comparisons here. Since SSE only has
12845 // GT and EQ comparisons for integer, swapping operands and multiple
12846 // operations may be required for some comparisons.
12848 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
12849 bool Subus = false;
12851 switch (SetCCOpcode) {
12852 default: llvm_unreachable("Unexpected SETCC condition");
12853 case ISD::SETNE: Invert = true;
12854 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
12855 case ISD::SETLT: Swap = true;
12856 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
12857 case ISD::SETGE: Swap = true;
12858 case ISD::SETLE: Opc = X86ISD::PCMPGT;
12859 Invert = true; break;
12860 case ISD::SETULT: Swap = true;
12861 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
12862 FlipSigns = true; break;
12863 case ISD::SETUGE: Swap = true;
12864 case ISD::SETULE: Opc = X86ISD::PCMPGT;
12865 FlipSigns = true; Invert = true; break;
12868 // Special case: Use min/max operations for SETULE/SETUGE
12869 MVT VET = VT.getVectorElementType();
12871 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
12872 || (Subtarget->hasSSE2() && (VET == MVT::i8));
12875 switch (SetCCOpcode) {
12877 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
12878 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
12881 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
12884 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
12885 if (!MinMax && hasSubus) {
12886 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
12888 // t = psubus Op0, Op1
12889 // pcmpeq t, <0..0>
12890 switch (SetCCOpcode) {
12892 case ISD::SETULT: {
12893 // If the comparison is against a constant we can turn this into a
12894 // setule. With psubus, setule does not require a swap. This is
12895 // beneficial because the constant in the register is no longer
12896 // destructed as the destination so it can be hoisted out of a loop.
12897 // Only do this pre-AVX since vpcmp* is no longer destructive.
12898 if (Subtarget->hasAVX())
12900 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
12901 if (ULEOp1.getNode()) {
12903 Subus = true; Invert = false; Swap = false;
12907 // Psubus is better than flip-sign because it requires no inversion.
12908 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
12909 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
12913 Opc = X86ISD::SUBUS;
12919 std::swap(Op0, Op1);
12921 // Check that the operation in question is available (most are plain SSE2,
12922 // but PCMPGTQ and PCMPEQQ have different requirements).
12923 if (VT == MVT::v2i64) {
12924 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
12925 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
12927 // First cast everything to the right type.
12928 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12929 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12931 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12932 // bits of the inputs before performing those operations. The lower
12933 // compare is always unsigned.
12936 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
12938 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
12939 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
12940 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
12941 Sign, Zero, Sign, Zero);
12943 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
12944 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
12946 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
12947 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
12948 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
12950 // Create masks for only the low parts/high parts of the 64 bit integers.
12951 static const int MaskHi[] = { 1, 1, 3, 3 };
12952 static const int MaskLo[] = { 0, 0, 2, 2 };
12953 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
12954 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
12955 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
12957 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
12958 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
12961 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12963 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12966 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
12967 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
12968 // pcmpeqd + pshufd + pand.
12969 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
12971 // First cast everything to the right type.
12972 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12973 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12976 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
12978 // Make sure the lower and upper halves are both all-ones.
12979 static const int Mask[] = { 1, 0, 3, 2 };
12980 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
12981 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
12984 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12986 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12990 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12991 // bits of the inputs before performing those operations.
12993 EVT EltVT = VT.getVectorElementType();
12994 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
12995 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
12996 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
12999 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13001 // If the logical-not of the result is required, perform that now.
13003 Result = DAG.getNOT(dl, Result, VT);
13006 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13009 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13010 getZeroVector(VT, Subtarget, DAG, dl));
13015 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13017 MVT VT = Op.getSimpleValueType();
13019 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13021 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13022 && "SetCC type must be 8-bit or 1-bit integer");
13023 SDValue Op0 = Op.getOperand(0);
13024 SDValue Op1 = Op.getOperand(1);
13026 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13028 // Optimize to BT if possible.
13029 // Lower (X & (1 << N)) == 0 to BT(X, N).
13030 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13031 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13032 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13033 Op1.getOpcode() == ISD::Constant &&
13034 cast<ConstantSDNode>(Op1)->isNullValue() &&
13035 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13036 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13037 if (NewSetCC.getNode()) {
13039 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13044 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13046 if (Op1.getOpcode() == ISD::Constant &&
13047 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13048 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13049 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13051 // If the input is a setcc, then reuse the input setcc or use a new one with
13052 // the inverted condition.
13053 if (Op0.getOpcode() == X86ISD::SETCC) {
13054 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13055 bool Invert = (CC == ISD::SETNE) ^
13056 cast<ConstantSDNode>(Op1)->isNullValue();
13060 CCode = X86::GetOppositeBranchCondition(CCode);
13061 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13062 DAG.getConstant(CCode, MVT::i8),
13063 Op0.getOperand(1));
13065 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13069 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13070 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13071 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13073 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13074 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
13077 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13078 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
13079 if (X86CC == X86::COND_INVALID)
13082 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13083 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13084 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13085 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
13087 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13091 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13092 static bool isX86LogicalCmp(SDValue Op) {
13093 unsigned Opc = Op.getNode()->getOpcode();
13094 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13095 Opc == X86ISD::SAHF)
13097 if (Op.getResNo() == 1 &&
13098 (Opc == X86ISD::ADD ||
13099 Opc == X86ISD::SUB ||
13100 Opc == X86ISD::ADC ||
13101 Opc == X86ISD::SBB ||
13102 Opc == X86ISD::SMUL ||
13103 Opc == X86ISD::UMUL ||
13104 Opc == X86ISD::INC ||
13105 Opc == X86ISD::DEC ||
13106 Opc == X86ISD::OR ||
13107 Opc == X86ISD::XOR ||
13108 Opc == X86ISD::AND))
13111 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13117 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13118 if (V.getOpcode() != ISD::TRUNCATE)
13121 SDValue VOp0 = V.getOperand(0);
13122 unsigned InBits = VOp0.getValueSizeInBits();
13123 unsigned Bits = V.getValueSizeInBits();
13124 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13127 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13128 bool addTest = true;
13129 SDValue Cond = Op.getOperand(0);
13130 SDValue Op1 = Op.getOperand(1);
13131 SDValue Op2 = Op.getOperand(2);
13133 EVT VT = Op1.getValueType();
13136 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13137 // are available. Otherwise fp cmovs get lowered into a less efficient branch
13138 // sequence later on.
13139 if (Cond.getOpcode() == ISD::SETCC &&
13140 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13141 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13142 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13143 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13144 int SSECC = translateX86FSETCC(
13145 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13148 if (Subtarget->hasAVX512()) {
13149 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13150 DAG.getConstant(SSECC, MVT::i8));
13151 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13153 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13154 DAG.getConstant(SSECC, MVT::i8));
13155 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13156 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13157 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13161 if (Cond.getOpcode() == ISD::SETCC) {
13162 SDValue NewCond = LowerSETCC(Cond, DAG);
13163 if (NewCond.getNode())
13167 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13168 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13169 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13170 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13171 if (Cond.getOpcode() == X86ISD::SETCC &&
13172 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13173 isZero(Cond.getOperand(1).getOperand(1))) {
13174 SDValue Cmp = Cond.getOperand(1);
13176 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13178 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13179 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13180 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13182 SDValue CmpOp0 = Cmp.getOperand(0);
13183 // Apply further optimizations for special cases
13184 // (select (x != 0), -1, 0) -> neg & sbb
13185 // (select (x == 0), 0, -1) -> neg & sbb
13186 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13187 if (YC->isNullValue() &&
13188 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13189 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13190 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13191 DAG.getConstant(0, CmpOp0.getValueType()),
13193 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13194 DAG.getConstant(X86::COND_B, MVT::i8),
13195 SDValue(Neg.getNode(), 1));
13199 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13200 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
13201 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13203 SDValue Res = // Res = 0 or -1.
13204 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13205 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
13207 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13208 Res = DAG.getNOT(DL, Res, Res.getValueType());
13210 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13211 if (!N2C || !N2C->isNullValue())
13212 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13217 // Look past (and (setcc_carry (cmp ...)), 1).
13218 if (Cond.getOpcode() == ISD::AND &&
13219 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13220 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13221 if (C && C->getAPIntValue() == 1)
13222 Cond = Cond.getOperand(0);
13225 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13226 // setting operand in place of the X86ISD::SETCC.
13227 unsigned CondOpcode = Cond.getOpcode();
13228 if (CondOpcode == X86ISD::SETCC ||
13229 CondOpcode == X86ISD::SETCC_CARRY) {
13230 CC = Cond.getOperand(0);
13232 SDValue Cmp = Cond.getOperand(1);
13233 unsigned Opc = Cmp.getOpcode();
13234 MVT VT = Op.getSimpleValueType();
13236 bool IllegalFPCMov = false;
13237 if (VT.isFloatingPoint() && !VT.isVector() &&
13238 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13239 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13241 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13242 Opc == X86ISD::BT) { // FIXME
13246 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13247 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13248 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13249 Cond.getOperand(0).getValueType() != MVT::i8)) {
13250 SDValue LHS = Cond.getOperand(0);
13251 SDValue RHS = Cond.getOperand(1);
13252 unsigned X86Opcode;
13255 switch (CondOpcode) {
13256 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13257 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13258 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13259 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13260 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13261 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13262 default: llvm_unreachable("unexpected overflowing operator");
13264 if (CondOpcode == ISD::UMULO)
13265 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13268 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13270 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13272 if (CondOpcode == ISD::UMULO)
13273 Cond = X86Op.getValue(2);
13275 Cond = X86Op.getValue(1);
13277 CC = DAG.getConstant(X86Cond, MVT::i8);
13282 // Look pass the truncate if the high bits are known zero.
13283 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13284 Cond = Cond.getOperand(0);
13286 // We know the result of AND is compared against zero. Try to match
13288 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13289 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13290 if (NewSetCC.getNode()) {
13291 CC = NewSetCC.getOperand(0);
13292 Cond = NewSetCC.getOperand(1);
13299 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13300 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13303 // a < b ? -1 : 0 -> RES = ~setcc_carry
13304 // a < b ? 0 : -1 -> RES = setcc_carry
13305 // a >= b ? -1 : 0 -> RES = setcc_carry
13306 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13307 if (Cond.getOpcode() == X86ISD::SUB) {
13308 Cond = ConvertCmpIfNecessary(Cond, DAG);
13309 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13311 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13312 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13313 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13314 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
13315 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13316 return DAG.getNOT(DL, Res, Res.getValueType());
13321 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13322 // widen the cmov and push the truncate through. This avoids introducing a new
13323 // branch during isel and doesn't add any extensions.
13324 if (Op.getValueType() == MVT::i8 &&
13325 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13326 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13327 if (T1.getValueType() == T2.getValueType() &&
13328 // Blacklist CopyFromReg to avoid partial register stalls.
13329 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13330 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13331 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13332 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13336 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13337 // condition is true.
13338 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13339 SDValue Ops[] = { Op2, Op1, CC, Cond };
13340 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13343 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13344 SelectionDAG &DAG) {
13345 MVT VT = Op->getSimpleValueType(0);
13346 SDValue In = Op->getOperand(0);
13347 MVT InVT = In.getSimpleValueType();
13348 MVT VTElt = VT.getVectorElementType();
13349 MVT InVTElt = InVT.getVectorElementType();
13353 if ((InVTElt == MVT::i1) &&
13354 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13355 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13357 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13358 VTElt.getSizeInBits() <= 16)) ||
13360 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13361 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13363 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13364 VTElt.getSizeInBits() >= 32))))
13365 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13367 unsigned int NumElts = VT.getVectorNumElements();
13369 if (NumElts != 8 && NumElts != 16)
13372 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13373 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13374 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13375 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13378 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13379 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13381 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
13382 Constant *C = ConstantInt::get(*DAG.getContext(),
13383 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
13385 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
13386 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13387 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
13388 MachinePointerInfo::getConstantPool(),
13389 false, false, false, Alignment);
13390 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
13391 if (VT.is512BitVector())
13393 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
13396 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13397 SelectionDAG &DAG) {
13398 MVT VT = Op->getSimpleValueType(0);
13399 SDValue In = Op->getOperand(0);
13400 MVT InVT = In.getSimpleValueType();
13403 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13404 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13406 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13407 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13408 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13411 if (Subtarget->hasInt256())
13412 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13414 // Optimize vectors in AVX mode
13415 // Sign extend v8i16 to v8i32 and
13418 // Divide input vector into two parts
13419 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13420 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13421 // concat the vectors to original VT
13423 unsigned NumElems = InVT.getVectorNumElements();
13424 SDValue Undef = DAG.getUNDEF(InVT);
13426 SmallVector<int,8> ShufMask1(NumElems, -1);
13427 for (unsigned i = 0; i != NumElems/2; ++i)
13430 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13432 SmallVector<int,8> ShufMask2(NumElems, -1);
13433 for (unsigned i = 0; i != NumElems/2; ++i)
13434 ShufMask2[i] = i + NumElems/2;
13436 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13438 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13439 VT.getVectorNumElements()/2);
13441 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13442 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13444 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13447 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13448 // may emit an illegal shuffle but the expansion is still better than scalar
13449 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13450 // we'll emit a shuffle and a arithmetic shift.
13451 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13452 // TODO: It is possible to support ZExt by zeroing the undef values during
13453 // the shuffle phase or after the shuffle.
13454 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13455 SelectionDAG &DAG) {
13456 MVT RegVT = Op.getSimpleValueType();
13457 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13458 assert(RegVT.isInteger() &&
13459 "We only custom lower integer vector sext loads.");
13461 // Nothing useful we can do without SSE2 shuffles.
13462 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13464 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13466 EVT MemVT = Ld->getMemoryVT();
13467 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13468 unsigned RegSz = RegVT.getSizeInBits();
13470 ISD::LoadExtType Ext = Ld->getExtensionType();
13472 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13473 && "Only anyext and sext are currently implemented.");
13474 assert(MemVT != RegVT && "Cannot extend to the same type");
13475 assert(MemVT.isVector() && "Must load a vector from memory");
13477 unsigned NumElems = RegVT.getVectorNumElements();
13478 unsigned MemSz = MemVT.getSizeInBits();
13479 assert(RegSz > MemSz && "Register size must be greater than the mem size");
13481 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
13482 // The only way in which we have a legal 256-bit vector result but not the
13483 // integer 256-bit operations needed to directly lower a sextload is if we
13484 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
13485 // a 128-bit vector and a normal sign_extend to 256-bits that should get
13486 // correctly legalized. We do this late to allow the canonical form of
13487 // sextload to persist throughout the rest of the DAG combiner -- it wants
13488 // to fold together any extensions it can, and so will fuse a sign_extend
13489 // of an sextload into a sextload targeting a wider value.
13491 if (MemSz == 128) {
13492 // Just switch this to a normal load.
13493 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
13494 "it must be a legal 128-bit vector "
13496 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
13497 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
13498 Ld->isInvariant(), Ld->getAlignment());
13500 assert(MemSz < 128 &&
13501 "Can't extend a type wider than 128 bits to a 256 bit vector!");
13502 // Do an sext load to a 128-bit vector type. We want to use the same
13503 // number of elements, but elements half as wide. This will end up being
13504 // recursively lowered by this routine, but will succeed as we definitely
13505 // have all the necessary features if we're using AVX1.
13507 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
13508 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
13510 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
13511 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
13512 Ld->isNonTemporal(), Ld->isInvariant(),
13513 Ld->getAlignment());
13516 // Replace chain users with the new chain.
13517 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
13518 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
13520 // Finally, do a normal sign-extend to the desired register.
13521 return DAG.getSExtOrTrunc(Load, dl, RegVT);
13524 // All sizes must be a power of two.
13525 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
13526 "Non-power-of-two elements are not custom lowered!");
13528 // Attempt to load the original value using scalar loads.
13529 // Find the largest scalar type that divides the total loaded size.
13530 MVT SclrLoadTy = MVT::i8;
13531 for (MVT Tp : MVT::integer_valuetypes()) {
13532 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
13537 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
13538 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
13540 SclrLoadTy = MVT::f64;
13542 // Calculate the number of scalar loads that we need to perform
13543 // in order to load our vector from memory.
13544 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
13546 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
13547 "Can only lower sext loads with a single scalar load!");
13549 unsigned loadRegZize = RegSz;
13550 if (Ext == ISD::SEXTLOAD && RegSz == 256)
13553 // Represent our vector as a sequence of elements which are the
13554 // largest scalar that we can load.
13555 EVT LoadUnitVecVT = EVT::getVectorVT(
13556 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
13558 // Represent the data using the same element type that is stored in
13559 // memory. In practice, we ''widen'' MemVT.
13561 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
13562 loadRegZize / MemVT.getScalarType().getSizeInBits());
13564 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
13565 "Invalid vector type");
13567 // We can't shuffle using an illegal type.
13568 assert(TLI.isTypeLegal(WideVecVT) &&
13569 "We only lower types that form legal widened vector types");
13571 SmallVector<SDValue, 8> Chains;
13572 SDValue Ptr = Ld->getBasePtr();
13573 SDValue Increment =
13574 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
13575 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
13577 for (unsigned i = 0; i < NumLoads; ++i) {
13578 // Perform a single load.
13579 SDValue ScalarLoad =
13580 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
13581 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
13582 Ld->getAlignment());
13583 Chains.push_back(ScalarLoad.getValue(1));
13584 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
13585 // another round of DAGCombining.
13587 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
13589 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
13590 ScalarLoad, DAG.getIntPtrConstant(i));
13592 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
13595 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
13597 // Bitcast the loaded value to a vector of the original element type, in
13598 // the size of the target vector type.
13599 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
13600 unsigned SizeRatio = RegSz / MemSz;
13602 if (Ext == ISD::SEXTLOAD) {
13603 // If we have SSE4.1, we can directly emit a VSEXT node.
13604 if (Subtarget->hasSSE41()) {
13605 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
13606 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13610 // Otherwise we'll shuffle the small elements in the high bits of the
13611 // larger type and perform an arithmetic shift. If the shift is not legal
13612 // it's better to scalarize.
13613 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
13614 "We can't implement a sext load without an arithmetic right shift!");
13616 // Redistribute the loaded elements into the different locations.
13617 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13618 for (unsigned i = 0; i != NumElems; ++i)
13619 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
13621 SDValue Shuff = DAG.getVectorShuffle(
13622 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13624 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13626 // Build the arithmetic shift.
13627 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
13628 MemVT.getVectorElementType().getSizeInBits();
13630 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
13632 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13636 // Redistribute the loaded elements into the different locations.
13637 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13638 for (unsigned i = 0; i != NumElems; ++i)
13639 ShuffleVec[i * SizeRatio] = i;
13641 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
13642 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13644 // Bitcast to the requested type.
13645 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13646 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13650 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
13651 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
13652 // from the AND / OR.
13653 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
13654 Opc = Op.getOpcode();
13655 if (Opc != ISD::OR && Opc != ISD::AND)
13657 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13658 Op.getOperand(0).hasOneUse() &&
13659 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
13660 Op.getOperand(1).hasOneUse());
13663 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
13664 // 1 and that the SETCC node has a single use.
13665 static bool isXor1OfSetCC(SDValue Op) {
13666 if (Op.getOpcode() != ISD::XOR)
13668 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
13669 if (N1C && N1C->getAPIntValue() == 1) {
13670 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13671 Op.getOperand(0).hasOneUse();
13676 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
13677 bool addTest = true;
13678 SDValue Chain = Op.getOperand(0);
13679 SDValue Cond = Op.getOperand(1);
13680 SDValue Dest = Op.getOperand(2);
13683 bool Inverted = false;
13685 if (Cond.getOpcode() == ISD::SETCC) {
13686 // Check for setcc([su]{add,sub,mul}o == 0).
13687 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
13688 isa<ConstantSDNode>(Cond.getOperand(1)) &&
13689 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
13690 Cond.getOperand(0).getResNo() == 1 &&
13691 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
13692 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
13693 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
13694 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
13695 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
13696 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
13698 Cond = Cond.getOperand(0);
13700 SDValue NewCond = LowerSETCC(Cond, DAG);
13701 if (NewCond.getNode())
13706 // FIXME: LowerXALUO doesn't handle these!!
13707 else if (Cond.getOpcode() == X86ISD::ADD ||
13708 Cond.getOpcode() == X86ISD::SUB ||
13709 Cond.getOpcode() == X86ISD::SMUL ||
13710 Cond.getOpcode() == X86ISD::UMUL)
13711 Cond = LowerXALUO(Cond, DAG);
13714 // Look pass (and (setcc_carry (cmp ...)), 1).
13715 if (Cond.getOpcode() == ISD::AND &&
13716 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13717 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13718 if (C && C->getAPIntValue() == 1)
13719 Cond = Cond.getOperand(0);
13722 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13723 // setting operand in place of the X86ISD::SETCC.
13724 unsigned CondOpcode = Cond.getOpcode();
13725 if (CondOpcode == X86ISD::SETCC ||
13726 CondOpcode == X86ISD::SETCC_CARRY) {
13727 CC = Cond.getOperand(0);
13729 SDValue Cmp = Cond.getOperand(1);
13730 unsigned Opc = Cmp.getOpcode();
13731 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
13732 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
13736 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
13740 // These can only come from an arithmetic instruction with overflow,
13741 // e.g. SADDO, UADDO.
13742 Cond = Cond.getNode()->getOperand(1);
13748 CondOpcode = Cond.getOpcode();
13749 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13750 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13751 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13752 Cond.getOperand(0).getValueType() != MVT::i8)) {
13753 SDValue LHS = Cond.getOperand(0);
13754 SDValue RHS = Cond.getOperand(1);
13755 unsigned X86Opcode;
13758 // Keep this in sync with LowerXALUO, otherwise we might create redundant
13759 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
13761 switch (CondOpcode) {
13762 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13764 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13766 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
13769 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13770 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13772 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13774 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
13777 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13778 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13779 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13780 default: llvm_unreachable("unexpected overflowing operator");
13783 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
13784 if (CondOpcode == ISD::UMULO)
13785 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13788 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13790 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
13792 if (CondOpcode == ISD::UMULO)
13793 Cond = X86Op.getValue(2);
13795 Cond = X86Op.getValue(1);
13797 CC = DAG.getConstant(X86Cond, MVT::i8);
13801 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
13802 SDValue Cmp = Cond.getOperand(0).getOperand(1);
13803 if (CondOpc == ISD::OR) {
13804 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
13805 // two branches instead of an explicit OR instruction with a
13807 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13808 isX86LogicalCmp(Cmp)) {
13809 CC = Cond.getOperand(0).getOperand(0);
13810 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13811 Chain, Dest, CC, Cmp);
13812 CC = Cond.getOperand(1).getOperand(0);
13816 } else { // ISD::AND
13817 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
13818 // two branches instead of an explicit AND instruction with a
13819 // separate test. However, we only do this if this block doesn't
13820 // have a fall-through edge, because this requires an explicit
13821 // jmp when the condition is false.
13822 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13823 isX86LogicalCmp(Cmp) &&
13824 Op.getNode()->hasOneUse()) {
13825 X86::CondCode CCode =
13826 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13827 CCode = X86::GetOppositeBranchCondition(CCode);
13828 CC = DAG.getConstant(CCode, MVT::i8);
13829 SDNode *User = *Op.getNode()->use_begin();
13830 // Look for an unconditional branch following this conditional branch.
13831 // We need this because we need to reverse the successors in order
13832 // to implement FCMP_OEQ.
13833 if (User->getOpcode() == ISD::BR) {
13834 SDValue FalseBB = User->getOperand(1);
13836 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13837 assert(NewBR == User);
13841 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13842 Chain, Dest, CC, Cmp);
13843 X86::CondCode CCode =
13844 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
13845 CCode = X86::GetOppositeBranchCondition(CCode);
13846 CC = DAG.getConstant(CCode, MVT::i8);
13852 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
13853 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
13854 // It should be transformed during dag combiner except when the condition
13855 // is set by a arithmetics with overflow node.
13856 X86::CondCode CCode =
13857 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13858 CCode = X86::GetOppositeBranchCondition(CCode);
13859 CC = DAG.getConstant(CCode, MVT::i8);
13860 Cond = Cond.getOperand(0).getOperand(1);
13862 } else if (Cond.getOpcode() == ISD::SETCC &&
13863 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
13864 // For FCMP_OEQ, we can emit
13865 // two branches instead of an explicit AND instruction with a
13866 // separate test. However, we only do this if this block doesn't
13867 // have a fall-through edge, because this requires an explicit
13868 // jmp when the condition is false.
13869 if (Op.getNode()->hasOneUse()) {
13870 SDNode *User = *Op.getNode()->use_begin();
13871 // Look for an unconditional branch following this conditional branch.
13872 // We need this because we need to reverse the successors in order
13873 // to implement FCMP_OEQ.
13874 if (User->getOpcode() == ISD::BR) {
13875 SDValue FalseBB = User->getOperand(1);
13877 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13878 assert(NewBR == User);
13882 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13883 Cond.getOperand(0), Cond.getOperand(1));
13884 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13885 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13886 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13887 Chain, Dest, CC, Cmp);
13888 CC = DAG.getConstant(X86::COND_P, MVT::i8);
13893 } else if (Cond.getOpcode() == ISD::SETCC &&
13894 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
13895 // For FCMP_UNE, we can emit
13896 // two branches instead of an explicit AND instruction with a
13897 // separate test. However, we only do this if this block doesn't
13898 // have a fall-through edge, because this requires an explicit
13899 // jmp when the condition is false.
13900 if (Op.getNode()->hasOneUse()) {
13901 SDNode *User = *Op.getNode()->use_begin();
13902 // Look for an unconditional branch following this conditional branch.
13903 // We need this because we need to reverse the successors in order
13904 // to implement FCMP_UNE.
13905 if (User->getOpcode() == ISD::BR) {
13906 SDValue FalseBB = User->getOperand(1);
13908 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13909 assert(NewBR == User);
13912 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13913 Cond.getOperand(0), Cond.getOperand(1));
13914 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13915 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13916 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13917 Chain, Dest, CC, Cmp);
13918 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
13928 // Look pass the truncate if the high bits are known zero.
13929 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13930 Cond = Cond.getOperand(0);
13932 // We know the result of AND is compared against zero. Try to match
13934 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13935 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
13936 if (NewSetCC.getNode()) {
13937 CC = NewSetCC.getOperand(0);
13938 Cond = NewSetCC.getOperand(1);
13945 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
13946 CC = DAG.getConstant(X86Cond, MVT::i8);
13947 Cond = EmitTest(Cond, X86Cond, dl, DAG);
13949 Cond = ConvertCmpIfNecessary(Cond, DAG);
13950 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13951 Chain, Dest, CC, Cond);
13954 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
13955 // Calls to _alloca are needed to probe the stack when allocating more than 4k
13956 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
13957 // that the guard pages used by the OS virtual memory manager are allocated in
13958 // correct sequence.
13960 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
13961 SelectionDAG &DAG) const {
13962 MachineFunction &MF = DAG.getMachineFunction();
13963 bool SplitStack = MF.shouldSplitStack();
13964 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
13969 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13970 SDNode* Node = Op.getNode();
13972 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
13973 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
13974 " not tell us which reg is the stack pointer!");
13975 EVT VT = Node->getValueType(0);
13976 SDValue Tmp1 = SDValue(Node, 0);
13977 SDValue Tmp2 = SDValue(Node, 1);
13978 SDValue Tmp3 = Node->getOperand(2);
13979 SDValue Chain = Tmp1.getOperand(0);
13981 // Chain the dynamic stack allocation so that it doesn't modify the stack
13982 // pointer when other instructions are using the stack.
13983 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
13986 SDValue Size = Tmp2.getOperand(1);
13987 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
13988 Chain = SP.getValue(1);
13989 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
13990 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
13991 unsigned StackAlign = TFI.getStackAlignment();
13992 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
13993 if (Align > StackAlign)
13994 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
13995 DAG.getConstant(-(uint64_t)Align, VT));
13996 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
13998 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
13999 DAG.getIntPtrConstant(0, true), SDValue(),
14002 SDValue Ops[2] = { Tmp1, Tmp2 };
14003 return DAG.getMergeValues(Ops, dl);
14007 SDValue Chain = Op.getOperand(0);
14008 SDValue Size = Op.getOperand(1);
14009 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14010 EVT VT = Op.getNode()->getValueType(0);
14012 bool Is64Bit = Subtarget->is64Bit();
14013 EVT SPTy = getPointerTy();
14016 MachineRegisterInfo &MRI = MF.getRegInfo();
14019 // The 64 bit implementation of segmented stacks needs to clobber both r10
14020 // r11. This makes it impossible to use it along with nested parameters.
14021 const Function *F = MF.getFunction();
14023 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14025 if (I->hasNestAttr())
14026 report_fatal_error("Cannot use segmented stacks with functions that "
14027 "have nested arguments.");
14030 const TargetRegisterClass *AddrRegClass =
14031 getRegClassFor(getPointerTy());
14032 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14033 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14034 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14035 DAG.getRegister(Vreg, SPTy));
14036 SDValue Ops1[2] = { Value, Chain };
14037 return DAG.getMergeValues(Ops1, dl);
14040 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14042 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14043 Flag = Chain.getValue(1);
14044 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14046 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14048 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14049 unsigned SPReg = RegInfo->getStackRegister();
14050 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14051 Chain = SP.getValue(1);
14054 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14055 DAG.getConstant(-(uint64_t)Align, VT));
14056 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14059 SDValue Ops1[2] = { SP, Chain };
14060 return DAG.getMergeValues(Ops1, dl);
14064 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14065 MachineFunction &MF = DAG.getMachineFunction();
14066 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14068 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14071 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14072 // vastart just stores the address of the VarArgsFrameIndex slot into the
14073 // memory location argument.
14074 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14076 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14077 MachinePointerInfo(SV), false, false, 0);
14081 // gp_offset (0 - 6 * 8)
14082 // fp_offset (48 - 48 + 8 * 16)
14083 // overflow_arg_area (point to parameters coming in memory).
14085 SmallVector<SDValue, 8> MemOps;
14086 SDValue FIN = Op.getOperand(1);
14088 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14089 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14091 FIN, MachinePointerInfo(SV), false, false, 0);
14092 MemOps.push_back(Store);
14095 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14096 FIN, DAG.getIntPtrConstant(4));
14097 Store = DAG.getStore(Op.getOperand(0), DL,
14098 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
14100 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14101 MemOps.push_back(Store);
14103 // Store ptr to overflow_arg_area
14104 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14105 FIN, DAG.getIntPtrConstant(4));
14106 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14108 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14109 MachinePointerInfo(SV, 8),
14111 MemOps.push_back(Store);
14113 // Store ptr to reg_save_area.
14114 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14115 FIN, DAG.getIntPtrConstant(8));
14116 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14118 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14119 MachinePointerInfo(SV, 16), false, false, 0);
14120 MemOps.push_back(Store);
14121 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14124 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14125 assert(Subtarget->is64Bit() &&
14126 "LowerVAARG only handles 64-bit va_arg!");
14127 assert((Subtarget->isTargetLinux() ||
14128 Subtarget->isTargetDarwin()) &&
14129 "Unhandled target in LowerVAARG");
14130 assert(Op.getNode()->getNumOperands() == 4);
14131 SDValue Chain = Op.getOperand(0);
14132 SDValue SrcPtr = Op.getOperand(1);
14133 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14134 unsigned Align = Op.getConstantOperandVal(3);
14137 EVT ArgVT = Op.getNode()->getValueType(0);
14138 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14139 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14142 // Decide which area this value should be read from.
14143 // TODO: Implement the AMD64 ABI in its entirety. This simple
14144 // selection mechanism works only for the basic types.
14145 if (ArgVT == MVT::f80) {
14146 llvm_unreachable("va_arg for f80 not yet implemented");
14147 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14148 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14149 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14150 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14152 llvm_unreachable("Unhandled argument type in LowerVAARG");
14155 if (ArgMode == 2) {
14156 // Sanity Check: Make sure using fp_offset makes sense.
14157 assert(!DAG.getTarget().Options.UseSoftFloat &&
14158 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14159 Attribute::NoImplicitFloat)) &&
14160 Subtarget->hasSSE1());
14163 // Insert VAARG_64 node into the DAG
14164 // VAARG_64 returns two values: Variable Argument Address, Chain
14165 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, MVT::i32),
14166 DAG.getConstant(ArgMode, MVT::i8),
14167 DAG.getConstant(Align, MVT::i32)};
14168 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14169 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14170 VTs, InstOps, MVT::i64,
14171 MachinePointerInfo(SV),
14173 /*Volatile=*/false,
14175 /*WriteMem=*/true);
14176 Chain = VAARG.getValue(1);
14178 // Load the next argument and return it
14179 return DAG.getLoad(ArgVT, dl,
14182 MachinePointerInfo(),
14183 false, false, false, 0);
14186 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14187 SelectionDAG &DAG) {
14188 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14189 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14190 SDValue Chain = Op.getOperand(0);
14191 SDValue DstPtr = Op.getOperand(1);
14192 SDValue SrcPtr = Op.getOperand(2);
14193 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14194 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14197 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14198 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
14200 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14203 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14204 // amount is a constant. Takes immediate version of shift as input.
14205 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14206 SDValue SrcOp, uint64_t ShiftAmt,
14207 SelectionDAG &DAG) {
14208 MVT ElementType = VT.getVectorElementType();
14210 // Fold this packed shift into its first operand if ShiftAmt is 0.
14214 // Check for ShiftAmt >= element width
14215 if (ShiftAmt >= ElementType.getSizeInBits()) {
14216 if (Opc == X86ISD::VSRAI)
14217 ShiftAmt = ElementType.getSizeInBits() - 1;
14219 return DAG.getConstant(0, VT);
14222 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14223 && "Unknown target vector shift-by-constant node");
14225 // Fold this packed vector shift into a build vector if SrcOp is a
14226 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14227 if (VT == SrcOp.getSimpleValueType() &&
14228 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14229 SmallVector<SDValue, 8> Elts;
14230 unsigned NumElts = SrcOp->getNumOperands();
14231 ConstantSDNode *ND;
14234 default: llvm_unreachable(nullptr);
14235 case X86ISD::VSHLI:
14236 for (unsigned i=0; i!=NumElts; ++i) {
14237 SDValue CurrentOp = SrcOp->getOperand(i);
14238 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14239 Elts.push_back(CurrentOp);
14242 ND = cast<ConstantSDNode>(CurrentOp);
14243 const APInt &C = ND->getAPIntValue();
14244 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
14247 case X86ISD::VSRLI:
14248 for (unsigned i=0; i!=NumElts; ++i) {
14249 SDValue CurrentOp = SrcOp->getOperand(i);
14250 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14251 Elts.push_back(CurrentOp);
14254 ND = cast<ConstantSDNode>(CurrentOp);
14255 const APInt &C = ND->getAPIntValue();
14256 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
14259 case X86ISD::VSRAI:
14260 for (unsigned i=0; i!=NumElts; ++i) {
14261 SDValue CurrentOp = SrcOp->getOperand(i);
14262 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14263 Elts.push_back(CurrentOp);
14266 ND = cast<ConstantSDNode>(CurrentOp);
14267 const APInt &C = ND->getAPIntValue();
14268 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
14273 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14276 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
14279 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14280 // may or may not be a constant. Takes immediate version of shift as input.
14281 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14282 SDValue SrcOp, SDValue ShAmt,
14283 SelectionDAG &DAG) {
14284 MVT SVT = ShAmt.getSimpleValueType();
14285 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14287 // Catch shift-by-constant.
14288 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14289 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14290 CShAmt->getZExtValue(), DAG);
14292 // Change opcode to non-immediate version
14294 default: llvm_unreachable("Unknown target vector shift node");
14295 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14296 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14297 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14300 const X86Subtarget &Subtarget =
14301 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14302 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14303 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14304 // Let the shuffle legalizer expand this shift amount node.
14305 SDValue Op0 = ShAmt.getOperand(0);
14306 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14307 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14309 // Need to build a vector containing shift amount.
14310 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14311 SmallVector<SDValue, 4> ShOps;
14312 ShOps.push_back(ShAmt);
14313 if (SVT == MVT::i32) {
14314 ShOps.push_back(DAG.getConstant(0, SVT));
14315 ShOps.push_back(DAG.getUNDEF(SVT));
14317 ShOps.push_back(DAG.getUNDEF(SVT));
14319 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14320 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14323 // The return type has to be a 128-bit type with the same element
14324 // type as the input type.
14325 MVT EltVT = VT.getVectorElementType();
14326 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14328 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14329 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14332 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14333 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14334 /// necessary casting for \p Mask when lowering masking intrinsics.
14335 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14336 SDValue PreservedSrc,
14337 const X86Subtarget *Subtarget,
14338 SelectionDAG &DAG) {
14339 EVT VT = Op.getValueType();
14340 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14341 MVT::i1, VT.getVectorNumElements());
14342 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14343 Mask.getValueType().getSizeInBits());
14346 assert(MaskVT.isSimple() && "invalid mask type");
14348 if (isAllOnes(Mask))
14351 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14352 // are extracted by EXTRACT_SUBVECTOR.
14353 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14354 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14355 DAG.getIntPtrConstant(0));
14357 switch (Op.getOpcode()) {
14359 case X86ISD::PCMPEQM:
14360 case X86ISD::PCMPGTM:
14362 case X86ISD::CMPMU:
14363 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14365 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14366 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14367 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14370 /// \brief Creates an SDNode for a predicated scalar operation.
14371 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14372 /// The mask is comming as MVT::i8 and it should be truncated
14373 /// to MVT::i1 while lowering masking intrinsics.
14374 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14375 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14376 /// a scalar instruction.
14377 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14378 SDValue PreservedSrc,
14379 const X86Subtarget *Subtarget,
14380 SelectionDAG &DAG) {
14381 if (isAllOnes(Mask))
14384 EVT VT = Op.getValueType();
14386 // The mask should be of type MVT::i1
14387 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14389 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14390 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14391 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14394 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14395 SelectionDAG &DAG) {
14397 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14398 EVT VT = Op.getValueType();
14399 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14401 switch(IntrData->Type) {
14402 case INTR_TYPE_1OP:
14403 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14404 case INTR_TYPE_2OP:
14405 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14407 case INTR_TYPE_3OP:
14408 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14409 Op.getOperand(2), Op.getOperand(3));
14410 case INTR_TYPE_1OP_MASK_RM: {
14411 SDValue Src = Op.getOperand(1);
14412 SDValue Src0 = Op.getOperand(2);
14413 SDValue Mask = Op.getOperand(3);
14414 SDValue RoundingMode = Op.getOperand(4);
14415 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14417 Mask, Src0, Subtarget, DAG);
14419 case INTR_TYPE_SCALAR_MASK_RM: {
14420 SDValue Src1 = Op.getOperand(1);
14421 SDValue Src2 = Op.getOperand(2);
14422 SDValue Src0 = Op.getOperand(3);
14423 SDValue Mask = Op.getOperand(4);
14424 // There are 2 kinds of intrinsics in this group:
14425 // (1) With supress-all-exceptions (sae) - 6 operands
14426 // (2) With rounding mode and sae - 7 operands.
14427 if (Op.getNumOperands() == 6) {
14428 SDValue Sae = Op.getOperand(5);
14429 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14431 Mask, Src0, Subtarget, DAG);
14433 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
14434 SDValue RoundingMode = Op.getOperand(5);
14435 SDValue Sae = Op.getOperand(6);
14436 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14437 RoundingMode, Sae),
14438 Mask, Src0, Subtarget, DAG);
14440 case INTR_TYPE_2OP_MASK: {
14441 SDValue Src1 = Op.getOperand(1);
14442 SDValue Src2 = Op.getOperand(2);
14443 SDValue PassThru = Op.getOperand(3);
14444 SDValue Mask = Op.getOperand(4);
14445 // We specify 2 possible opcodes for intrinsics with rounding modes.
14446 // First, we check if the intrinsic may have non-default rounding mode,
14447 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14448 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14449 if (IntrWithRoundingModeOpcode != 0) {
14450 SDValue Rnd = Op.getOperand(5);
14451 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14452 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14453 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14454 dl, Op.getValueType(),
14456 Mask, PassThru, Subtarget, DAG);
14459 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14461 Mask, PassThru, Subtarget, DAG);
14463 case FMA_OP_MASK: {
14464 SDValue Src1 = Op.getOperand(1);
14465 SDValue Src2 = Op.getOperand(2);
14466 SDValue Src3 = Op.getOperand(3);
14467 SDValue Mask = Op.getOperand(4);
14468 // We specify 2 possible opcodes for intrinsics with rounding modes.
14469 // First, we check if the intrinsic may have non-default rounding mode,
14470 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14471 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14472 if (IntrWithRoundingModeOpcode != 0) {
14473 SDValue Rnd = Op.getOperand(5);
14474 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14475 X86::STATIC_ROUNDING::CUR_DIRECTION)
14476 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14477 dl, Op.getValueType(),
14478 Src1, Src2, Src3, Rnd),
14479 Mask, Src1, Subtarget, DAG);
14481 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
14482 dl, Op.getValueType(),
14484 Mask, Src1, Subtarget, DAG);
14487 case CMP_MASK_CC: {
14488 // Comparison intrinsics with masks.
14489 // Example of transformation:
14490 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
14491 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
14493 // (v8i1 (insert_subvector undef,
14494 // (v2i1 (and (PCMPEQM %a, %b),
14495 // (extract_subvector
14496 // (v8i1 (bitcast %mask)), 0))), 0))))
14497 EVT VT = Op.getOperand(1).getValueType();
14498 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14499 VT.getVectorNumElements());
14500 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
14501 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14502 Mask.getValueType().getSizeInBits());
14504 if (IntrData->Type == CMP_MASK_CC) {
14505 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14506 Op.getOperand(2), Op.getOperand(3));
14508 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
14509 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14512 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
14513 DAG.getTargetConstant(0, MaskVT),
14515 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
14516 DAG.getUNDEF(BitcastVT), CmpMask,
14517 DAG.getIntPtrConstant(0));
14518 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
14520 case COMI: { // Comparison intrinsics
14521 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
14522 SDValue LHS = Op.getOperand(1);
14523 SDValue RHS = Op.getOperand(2);
14524 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
14525 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
14526 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
14527 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14528 DAG.getConstant(X86CC, MVT::i8), Cond);
14529 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14532 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
14533 Op.getOperand(1), Op.getOperand(2), DAG);
14535 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
14536 Op.getSimpleValueType(),
14538 Op.getOperand(2), DAG),
14539 Op.getOperand(4), Op.getOperand(3), Subtarget,
14541 case COMPRESS_EXPAND_IN_REG: {
14542 SDValue Mask = Op.getOperand(3);
14543 SDValue DataToCompress = Op.getOperand(1);
14544 SDValue PassThru = Op.getOperand(2);
14545 if (isAllOnes(Mask)) // return data as is
14546 return Op.getOperand(1);
14547 EVT VT = Op.getValueType();
14548 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14549 VT.getVectorNumElements());
14550 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14551 Mask.getValueType().getSizeInBits());
14553 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14554 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14555 DAG.getIntPtrConstant(0));
14557 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
14561 SDValue Mask = Op.getOperand(3);
14562 EVT VT = Op.getValueType();
14563 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14564 VT.getVectorNumElements());
14565 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14566 Mask.getValueType().getSizeInBits());
14568 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14569 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14570 DAG.getIntPtrConstant(0));
14571 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
14580 default: return SDValue(); // Don't custom lower most intrinsics.
14582 case Intrinsic::x86_avx512_mask_valign_q_512:
14583 case Intrinsic::x86_avx512_mask_valign_d_512:
14584 // Vector source operands are swapped.
14585 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
14586 Op.getValueType(), Op.getOperand(2),
14589 Op.getOperand(5), Op.getOperand(4),
14592 // ptest and testp intrinsics. The intrinsic these come from are designed to
14593 // return an integer value, not just an instruction so lower it to the ptest
14594 // or testp pattern and a setcc for the result.
14595 case Intrinsic::x86_sse41_ptestz:
14596 case Intrinsic::x86_sse41_ptestc:
14597 case Intrinsic::x86_sse41_ptestnzc:
14598 case Intrinsic::x86_avx_ptestz_256:
14599 case Intrinsic::x86_avx_ptestc_256:
14600 case Intrinsic::x86_avx_ptestnzc_256:
14601 case Intrinsic::x86_avx_vtestz_ps:
14602 case Intrinsic::x86_avx_vtestc_ps:
14603 case Intrinsic::x86_avx_vtestnzc_ps:
14604 case Intrinsic::x86_avx_vtestz_pd:
14605 case Intrinsic::x86_avx_vtestc_pd:
14606 case Intrinsic::x86_avx_vtestnzc_pd:
14607 case Intrinsic::x86_avx_vtestz_ps_256:
14608 case Intrinsic::x86_avx_vtestc_ps_256:
14609 case Intrinsic::x86_avx_vtestnzc_ps_256:
14610 case Intrinsic::x86_avx_vtestz_pd_256:
14611 case Intrinsic::x86_avx_vtestc_pd_256:
14612 case Intrinsic::x86_avx_vtestnzc_pd_256: {
14613 bool IsTestPacked = false;
14616 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
14617 case Intrinsic::x86_avx_vtestz_ps:
14618 case Intrinsic::x86_avx_vtestz_pd:
14619 case Intrinsic::x86_avx_vtestz_ps_256:
14620 case Intrinsic::x86_avx_vtestz_pd_256:
14621 IsTestPacked = true; // Fallthrough
14622 case Intrinsic::x86_sse41_ptestz:
14623 case Intrinsic::x86_avx_ptestz_256:
14625 X86CC = X86::COND_E;
14627 case Intrinsic::x86_avx_vtestc_ps:
14628 case Intrinsic::x86_avx_vtestc_pd:
14629 case Intrinsic::x86_avx_vtestc_ps_256:
14630 case Intrinsic::x86_avx_vtestc_pd_256:
14631 IsTestPacked = true; // Fallthrough
14632 case Intrinsic::x86_sse41_ptestc:
14633 case Intrinsic::x86_avx_ptestc_256:
14635 X86CC = X86::COND_B;
14637 case Intrinsic::x86_avx_vtestnzc_ps:
14638 case Intrinsic::x86_avx_vtestnzc_pd:
14639 case Intrinsic::x86_avx_vtestnzc_ps_256:
14640 case Intrinsic::x86_avx_vtestnzc_pd_256:
14641 IsTestPacked = true; // Fallthrough
14642 case Intrinsic::x86_sse41_ptestnzc:
14643 case Intrinsic::x86_avx_ptestnzc_256:
14645 X86CC = X86::COND_A;
14649 SDValue LHS = Op.getOperand(1);
14650 SDValue RHS = Op.getOperand(2);
14651 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
14652 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
14653 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14654 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
14655 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14657 case Intrinsic::x86_avx512_kortestz_w:
14658 case Intrinsic::x86_avx512_kortestc_w: {
14659 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
14660 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
14661 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
14662 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14663 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
14664 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
14665 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14668 case Intrinsic::x86_sse42_pcmpistria128:
14669 case Intrinsic::x86_sse42_pcmpestria128:
14670 case Intrinsic::x86_sse42_pcmpistric128:
14671 case Intrinsic::x86_sse42_pcmpestric128:
14672 case Intrinsic::x86_sse42_pcmpistrio128:
14673 case Intrinsic::x86_sse42_pcmpestrio128:
14674 case Intrinsic::x86_sse42_pcmpistris128:
14675 case Intrinsic::x86_sse42_pcmpestris128:
14676 case Intrinsic::x86_sse42_pcmpistriz128:
14677 case Intrinsic::x86_sse42_pcmpestriz128: {
14681 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14682 case Intrinsic::x86_sse42_pcmpistria128:
14683 Opcode = X86ISD::PCMPISTRI;
14684 X86CC = X86::COND_A;
14686 case Intrinsic::x86_sse42_pcmpestria128:
14687 Opcode = X86ISD::PCMPESTRI;
14688 X86CC = X86::COND_A;
14690 case Intrinsic::x86_sse42_pcmpistric128:
14691 Opcode = X86ISD::PCMPISTRI;
14692 X86CC = X86::COND_B;
14694 case Intrinsic::x86_sse42_pcmpestric128:
14695 Opcode = X86ISD::PCMPESTRI;
14696 X86CC = X86::COND_B;
14698 case Intrinsic::x86_sse42_pcmpistrio128:
14699 Opcode = X86ISD::PCMPISTRI;
14700 X86CC = X86::COND_O;
14702 case Intrinsic::x86_sse42_pcmpestrio128:
14703 Opcode = X86ISD::PCMPESTRI;
14704 X86CC = X86::COND_O;
14706 case Intrinsic::x86_sse42_pcmpistris128:
14707 Opcode = X86ISD::PCMPISTRI;
14708 X86CC = X86::COND_S;
14710 case Intrinsic::x86_sse42_pcmpestris128:
14711 Opcode = X86ISD::PCMPESTRI;
14712 X86CC = X86::COND_S;
14714 case Intrinsic::x86_sse42_pcmpistriz128:
14715 Opcode = X86ISD::PCMPISTRI;
14716 X86CC = X86::COND_E;
14718 case Intrinsic::x86_sse42_pcmpestriz128:
14719 Opcode = X86ISD::PCMPESTRI;
14720 X86CC = X86::COND_E;
14723 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14724 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14725 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
14726 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14727 DAG.getConstant(X86CC, MVT::i8),
14728 SDValue(PCMP.getNode(), 1));
14729 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14732 case Intrinsic::x86_sse42_pcmpistri128:
14733 case Intrinsic::x86_sse42_pcmpestri128: {
14735 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
14736 Opcode = X86ISD::PCMPISTRI;
14738 Opcode = X86ISD::PCMPESTRI;
14740 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14741 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14742 return DAG.getNode(Opcode, dl, VTs, NewOps);
14747 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14748 SDValue Src, SDValue Mask, SDValue Base,
14749 SDValue Index, SDValue ScaleOp, SDValue Chain,
14750 const X86Subtarget * Subtarget) {
14752 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14753 assert(C && "Invalid scale type");
14754 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14755 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14756 Index.getSimpleValueType().getVectorNumElements());
14758 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14760 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14762 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14763 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
14764 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14765 SDValue Segment = DAG.getRegister(0, MVT::i32);
14766 if (Src.getOpcode() == ISD::UNDEF)
14767 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
14768 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14769 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14770 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
14771 return DAG.getMergeValues(RetOps, dl);
14774 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14775 SDValue Src, SDValue Mask, SDValue Base,
14776 SDValue Index, SDValue ScaleOp, SDValue Chain) {
14778 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14779 assert(C && "Invalid scale type");
14780 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14781 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14782 SDValue Segment = DAG.getRegister(0, MVT::i32);
14783 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14784 Index.getSimpleValueType().getVectorNumElements());
14786 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14788 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14790 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14791 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
14792 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
14793 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14794 return SDValue(Res, 1);
14797 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14798 SDValue Mask, SDValue Base, SDValue Index,
14799 SDValue ScaleOp, SDValue Chain) {
14801 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14802 assert(C && "Invalid scale type");
14803 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14804 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14805 SDValue Segment = DAG.getRegister(0, MVT::i32);
14807 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
14809 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14811 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14813 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14814 //SDVTList VTs = DAG.getVTList(MVT::Other);
14815 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14816 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
14817 return SDValue(Res, 0);
14820 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
14821 // read performance monitor counters (x86_rdpmc).
14822 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
14823 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14824 SmallVectorImpl<SDValue> &Results) {
14825 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14826 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14829 // The ECX register is used to select the index of the performance counter
14831 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
14833 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
14835 // Reads the content of a 64-bit performance counter and returns it in the
14836 // registers EDX:EAX.
14837 if (Subtarget->is64Bit()) {
14838 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14839 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14842 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14843 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14846 Chain = HI.getValue(1);
14848 if (Subtarget->is64Bit()) {
14849 // The EAX register is loaded with the low-order 32 bits. The EDX register
14850 // is loaded with the supported high-order bits of the counter.
14851 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14852 DAG.getConstant(32, MVT::i8));
14853 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14854 Results.push_back(Chain);
14858 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14859 SDValue Ops[] = { LO, HI };
14860 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14861 Results.push_back(Pair);
14862 Results.push_back(Chain);
14865 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
14866 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
14867 // also used to custom lower READCYCLECOUNTER nodes.
14868 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
14869 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14870 SmallVectorImpl<SDValue> &Results) {
14871 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14872 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
14875 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
14876 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
14877 // and the EAX register is loaded with the low-order 32 bits.
14878 if (Subtarget->is64Bit()) {
14879 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14880 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14883 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14884 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14887 SDValue Chain = HI.getValue(1);
14889 if (Opcode == X86ISD::RDTSCP_DAG) {
14890 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14892 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
14893 // the ECX register. Add 'ecx' explicitly to the chain.
14894 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
14896 // Explicitly store the content of ECX at the location passed in input
14897 // to the 'rdtscp' intrinsic.
14898 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
14899 MachinePointerInfo(), false, false, 0);
14902 if (Subtarget->is64Bit()) {
14903 // The EDX register is loaded with the high-order 32 bits of the MSR, and
14904 // the EAX register is loaded with the low-order 32 bits.
14905 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14906 DAG.getConstant(32, MVT::i8));
14907 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14908 Results.push_back(Chain);
14912 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14913 SDValue Ops[] = { LO, HI };
14914 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14915 Results.push_back(Pair);
14916 Results.push_back(Chain);
14919 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
14920 SelectionDAG &DAG) {
14921 SmallVector<SDValue, 2> Results;
14923 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
14925 return DAG.getMergeValues(Results, DL);
14929 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14930 SelectionDAG &DAG) {
14931 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
14933 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
14938 switch(IntrData->Type) {
14940 llvm_unreachable("Unknown Intrinsic Type");
14944 // Emit the node with the right value type.
14945 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
14946 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
14948 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
14949 // Otherwise return the value from Rand, which is always 0, casted to i32.
14950 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
14951 DAG.getConstant(1, Op->getValueType(1)),
14952 DAG.getConstant(X86::COND_B, MVT::i32),
14953 SDValue(Result.getNode(), 1) };
14954 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
14955 DAG.getVTList(Op->getValueType(1), MVT::Glue),
14958 // Return { result, isValid, chain }.
14959 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
14960 SDValue(Result.getNode(), 2));
14963 //gather(v1, mask, index, base, scale);
14964 SDValue Chain = Op.getOperand(0);
14965 SDValue Src = Op.getOperand(2);
14966 SDValue Base = Op.getOperand(3);
14967 SDValue Index = Op.getOperand(4);
14968 SDValue Mask = Op.getOperand(5);
14969 SDValue Scale = Op.getOperand(6);
14970 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
14974 //scatter(base, mask, index, v1, scale);
14975 SDValue Chain = Op.getOperand(0);
14976 SDValue Base = Op.getOperand(2);
14977 SDValue Mask = Op.getOperand(3);
14978 SDValue Index = Op.getOperand(4);
14979 SDValue Src = Op.getOperand(5);
14980 SDValue Scale = Op.getOperand(6);
14981 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
14984 SDValue Hint = Op.getOperand(6);
14986 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
14987 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
14988 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
14989 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
14990 SDValue Chain = Op.getOperand(0);
14991 SDValue Mask = Op.getOperand(2);
14992 SDValue Index = Op.getOperand(3);
14993 SDValue Base = Op.getOperand(4);
14994 SDValue Scale = Op.getOperand(5);
14995 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
14997 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
14999 SmallVector<SDValue, 2> Results;
15000 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
15001 return DAG.getMergeValues(Results, dl);
15003 // Read Performance Monitoring Counters.
15005 SmallVector<SDValue, 2> Results;
15006 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15007 return DAG.getMergeValues(Results, dl);
15009 // XTEST intrinsics.
15011 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15012 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15013 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15014 DAG.getConstant(X86::COND_NE, MVT::i8),
15016 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15017 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15018 Ret, SDValue(InTrans.getNode(), 1));
15022 SmallVector<SDValue, 2> Results;
15023 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15024 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15025 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15026 DAG.getConstant(-1, MVT::i8));
15027 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15028 Op.getOperand(4), GenCF.getValue(1));
15029 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15030 Op.getOperand(5), MachinePointerInfo(),
15032 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15033 DAG.getConstant(X86::COND_B, MVT::i8),
15035 Results.push_back(SetCC);
15036 Results.push_back(Store);
15037 return DAG.getMergeValues(Results, dl);
15039 case COMPRESS_TO_MEM: {
15041 SDValue Mask = Op.getOperand(4);
15042 SDValue DataToCompress = Op.getOperand(3);
15043 SDValue Addr = Op.getOperand(2);
15044 SDValue Chain = Op.getOperand(0);
15046 if (isAllOnes(Mask)) // return just a store
15047 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15048 MachinePointerInfo(), false, false, 0);
15050 EVT VT = DataToCompress.getValueType();
15051 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15052 VT.getVectorNumElements());
15053 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15054 Mask.getValueType().getSizeInBits());
15055 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15056 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15057 DAG.getIntPtrConstant(0));
15059 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15060 DataToCompress, DAG.getUNDEF(VT));
15061 return DAG.getStore(Chain, dl, Compressed, Addr,
15062 MachinePointerInfo(), false, false, 0);
15064 case EXPAND_FROM_MEM: {
15066 SDValue Mask = Op.getOperand(4);
15067 SDValue PathThru = Op.getOperand(3);
15068 SDValue Addr = Op.getOperand(2);
15069 SDValue Chain = Op.getOperand(0);
15070 EVT VT = Op.getValueType();
15072 if (isAllOnes(Mask)) // return just a load
15073 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15075 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15076 VT.getVectorNumElements());
15077 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15078 Mask.getValueType().getSizeInBits());
15079 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15080 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15081 DAG.getIntPtrConstant(0));
15083 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15084 false, false, false, 0);
15086 SDValue Results[] = {
15087 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15089 return DAG.getMergeValues(Results, dl);
15094 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15095 SelectionDAG &DAG) const {
15096 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15097 MFI->setReturnAddressIsTaken(true);
15099 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15102 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15104 EVT PtrVT = getPointerTy();
15107 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15108 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15109 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
15110 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15111 DAG.getNode(ISD::ADD, dl, PtrVT,
15112 FrameAddr, Offset),
15113 MachinePointerInfo(), false, false, false, 0);
15116 // Just load the return address.
15117 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15118 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15119 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15122 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15123 MachineFunction &MF = DAG.getMachineFunction();
15124 MachineFrameInfo *MFI = MF.getFrameInfo();
15125 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15126 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15127 EVT VT = Op.getValueType();
15129 MFI->setFrameAddressIsTaken(true);
15131 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15132 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15133 // is not possible to crawl up the stack without looking at the unwind codes
15135 int FrameAddrIndex = FuncInfo->getFAIndex();
15136 if (!FrameAddrIndex) {
15137 // Set up a frame object for the return address.
15138 unsigned SlotSize = RegInfo->getSlotSize();
15139 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15140 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
15141 FuncInfo->setFAIndex(FrameAddrIndex);
15143 return DAG.getFrameIndex(FrameAddrIndex, VT);
15146 unsigned FrameReg =
15147 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15148 SDLoc dl(Op); // FIXME probably not meaningful
15149 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15150 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15151 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15152 "Invalid Frame Register!");
15153 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15155 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15156 MachinePointerInfo(),
15157 false, false, false, 0);
15161 // FIXME? Maybe this could be a TableGen attribute on some registers and
15162 // this table could be generated automatically from RegInfo.
15163 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15165 unsigned Reg = StringSwitch<unsigned>(RegName)
15166 .Case("esp", X86::ESP)
15167 .Case("rsp", X86::RSP)
15171 report_fatal_error("Invalid register name global variable");
15174 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15175 SelectionDAG &DAG) const {
15176 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15177 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
15180 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15181 SDValue Chain = Op.getOperand(0);
15182 SDValue Offset = Op.getOperand(1);
15183 SDValue Handler = Op.getOperand(2);
15186 EVT PtrVT = getPointerTy();
15187 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15188 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15189 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15190 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15191 "Invalid Frame Register!");
15192 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15193 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15195 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15196 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
15197 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15198 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15200 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15202 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15203 DAG.getRegister(StoreAddrReg, PtrVT));
15206 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15207 SelectionDAG &DAG) const {
15209 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15210 DAG.getVTList(MVT::i32, MVT::Other),
15211 Op.getOperand(0), Op.getOperand(1));
15214 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15215 SelectionDAG &DAG) const {
15217 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15218 Op.getOperand(0), Op.getOperand(1));
15221 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15222 return Op.getOperand(0);
15225 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15226 SelectionDAG &DAG) const {
15227 SDValue Root = Op.getOperand(0);
15228 SDValue Trmp = Op.getOperand(1); // trampoline
15229 SDValue FPtr = Op.getOperand(2); // nested function
15230 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15233 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15234 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15236 if (Subtarget->is64Bit()) {
15237 SDValue OutChains[6];
15239 // Large code-model.
15240 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15241 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15243 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15244 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15246 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15248 // Load the pointer to the nested function into R11.
15249 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15250 SDValue Addr = Trmp;
15251 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15252 Addr, MachinePointerInfo(TrmpAddr),
15255 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15256 DAG.getConstant(2, MVT::i64));
15257 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15258 MachinePointerInfo(TrmpAddr, 2),
15261 // Load the 'nest' parameter value into R10.
15262 // R10 is specified in X86CallingConv.td
15263 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15265 DAG.getConstant(10, MVT::i64));
15266 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15267 Addr, MachinePointerInfo(TrmpAddr, 10),
15270 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15271 DAG.getConstant(12, MVT::i64));
15272 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15273 MachinePointerInfo(TrmpAddr, 12),
15276 // Jump to the nested function.
15277 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15278 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15279 DAG.getConstant(20, MVT::i64));
15280 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15281 Addr, MachinePointerInfo(TrmpAddr, 20),
15284 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15285 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15286 DAG.getConstant(22, MVT::i64));
15287 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
15288 MachinePointerInfo(TrmpAddr, 22),
15291 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15293 const Function *Func =
15294 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15295 CallingConv::ID CC = Func->getCallingConv();
15300 llvm_unreachable("Unsupported calling convention");
15301 case CallingConv::C:
15302 case CallingConv::X86_StdCall: {
15303 // Pass 'nest' parameter in ECX.
15304 // Must be kept in sync with X86CallingConv.td
15305 NestReg = X86::ECX;
15307 // Check that ECX wasn't needed by an 'inreg' parameter.
15308 FunctionType *FTy = Func->getFunctionType();
15309 const AttributeSet &Attrs = Func->getAttributes();
15311 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15312 unsigned InRegCount = 0;
15315 for (FunctionType::param_iterator I = FTy->param_begin(),
15316 E = FTy->param_end(); I != E; ++I, ++Idx)
15317 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15318 // FIXME: should only count parameters that are lowered to integers.
15319 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15321 if (InRegCount > 2) {
15322 report_fatal_error("Nest register in use - reduce number of inreg"
15328 case CallingConv::X86_FastCall:
15329 case CallingConv::X86_ThisCall:
15330 case CallingConv::Fast:
15331 // Pass 'nest' parameter in EAX.
15332 // Must be kept in sync with X86CallingConv.td
15333 NestReg = X86::EAX;
15337 SDValue OutChains[4];
15338 SDValue Addr, Disp;
15340 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15341 DAG.getConstant(10, MVT::i32));
15342 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15344 // This is storing the opcode for MOV32ri.
15345 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15346 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15347 OutChains[0] = DAG.getStore(Root, dl,
15348 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
15349 Trmp, MachinePointerInfo(TrmpAddr),
15352 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15353 DAG.getConstant(1, MVT::i32));
15354 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15355 MachinePointerInfo(TrmpAddr, 1),
15358 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15359 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15360 DAG.getConstant(5, MVT::i32));
15361 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
15362 MachinePointerInfo(TrmpAddr, 5),
15365 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15366 DAG.getConstant(6, MVT::i32));
15367 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15368 MachinePointerInfo(TrmpAddr, 6),
15371 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15375 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15376 SelectionDAG &DAG) const {
15378 The rounding mode is in bits 11:10 of FPSR, and has the following
15380 00 Round to nearest
15385 FLT_ROUNDS, on the other hand, expects the following:
15392 To perform the conversion, we do:
15393 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15396 MachineFunction &MF = DAG.getMachineFunction();
15397 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15398 unsigned StackAlignment = TFI.getStackAlignment();
15399 MVT VT = Op.getSimpleValueType();
15402 // Save FP Control Word to stack slot
15403 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15404 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15406 MachineMemOperand *MMO =
15407 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15408 MachineMemOperand::MOStore, 2, 2);
15410 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15411 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15412 DAG.getVTList(MVT::Other),
15413 Ops, MVT::i16, MMO);
15415 // Load FP Control Word from stack slot
15416 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15417 MachinePointerInfo(), false, false, false, 0);
15419 // Transform as necessary
15421 DAG.getNode(ISD::SRL, DL, MVT::i16,
15422 DAG.getNode(ISD::AND, DL, MVT::i16,
15423 CWD, DAG.getConstant(0x800, MVT::i16)),
15424 DAG.getConstant(11, MVT::i8));
15426 DAG.getNode(ISD::SRL, DL, MVT::i16,
15427 DAG.getNode(ISD::AND, DL, MVT::i16,
15428 CWD, DAG.getConstant(0x400, MVT::i16)),
15429 DAG.getConstant(9, MVT::i8));
15432 DAG.getNode(ISD::AND, DL, MVT::i16,
15433 DAG.getNode(ISD::ADD, DL, MVT::i16,
15434 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
15435 DAG.getConstant(1, MVT::i16)),
15436 DAG.getConstant(3, MVT::i16));
15438 return DAG.getNode((VT.getSizeInBits() < 16 ?
15439 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
15442 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
15443 MVT VT = Op.getSimpleValueType();
15445 unsigned NumBits = VT.getSizeInBits();
15448 Op = Op.getOperand(0);
15449 if (VT == MVT::i8) {
15450 // Zero extend to i32 since there is not an i8 bsr.
15452 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15455 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
15456 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15457 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15459 // If src is zero (i.e. bsr sets ZF), returns NumBits.
15462 DAG.getConstant(NumBits+NumBits-1, OpVT),
15463 DAG.getConstant(X86::COND_E, MVT::i8),
15466 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
15468 // Finally xor with NumBits-1.
15469 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15472 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15476 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
15477 MVT VT = Op.getSimpleValueType();
15479 unsigned NumBits = VT.getSizeInBits();
15482 Op = Op.getOperand(0);
15483 if (VT == MVT::i8) {
15484 // Zero extend to i32 since there is not an i8 bsr.
15486 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15489 // Issue a bsr (scan bits in reverse).
15490 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15491 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15493 // And xor with NumBits-1.
15494 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15497 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15501 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
15502 MVT VT = Op.getSimpleValueType();
15503 unsigned NumBits = VT.getSizeInBits();
15505 Op = Op.getOperand(0);
15507 // Issue a bsf (scan bits forward) which also sets EFLAGS.
15508 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
15509 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
15511 // If src is zero (i.e. bsf sets ZF), returns NumBits.
15514 DAG.getConstant(NumBits, VT),
15515 DAG.getConstant(X86::COND_E, MVT::i8),
15518 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
15521 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
15522 // ones, and then concatenate the result back.
15523 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
15524 MVT VT = Op.getSimpleValueType();
15526 assert(VT.is256BitVector() && VT.isInteger() &&
15527 "Unsupported value type for operation");
15529 unsigned NumElems = VT.getVectorNumElements();
15532 // Extract the LHS vectors
15533 SDValue LHS = Op.getOperand(0);
15534 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15535 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15537 // Extract the RHS vectors
15538 SDValue RHS = Op.getOperand(1);
15539 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15540 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15542 MVT EltVT = VT.getVectorElementType();
15543 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15545 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15546 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15547 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15550 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15551 assert(Op.getSimpleValueType().is256BitVector() &&
15552 Op.getSimpleValueType().isInteger() &&
15553 "Only handle AVX 256-bit vector integer operation");
15554 return Lower256IntArith(Op, DAG);
15557 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15558 assert(Op.getSimpleValueType().is256BitVector() &&
15559 Op.getSimpleValueType().isInteger() &&
15560 "Only handle AVX 256-bit vector integer operation");
15561 return Lower256IntArith(Op, DAG);
15564 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
15565 SelectionDAG &DAG) {
15567 MVT VT = Op.getSimpleValueType();
15569 // Decompose 256-bit ops into smaller 128-bit ops.
15570 if (VT.is256BitVector() && !Subtarget->hasInt256())
15571 return Lower256IntArith(Op, DAG);
15573 SDValue A = Op.getOperand(0);
15574 SDValue B = Op.getOperand(1);
15576 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
15577 if (VT == MVT::v4i32) {
15578 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
15579 "Should not custom lower when pmuldq is available!");
15581 // Extract the odd parts.
15582 static const int UnpackMask[] = { 1, -1, 3, -1 };
15583 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
15584 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
15586 // Multiply the even parts.
15587 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
15588 // Now multiply odd parts.
15589 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
15591 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
15592 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
15594 // Merge the two vectors back together with a shuffle. This expands into 2
15596 static const int ShufMask[] = { 0, 4, 2, 6 };
15597 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
15600 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
15601 "Only know how to lower V2I64/V4I64/V8I64 multiply");
15603 // Ahi = psrlqi(a, 32);
15604 // Bhi = psrlqi(b, 32);
15606 // AloBlo = pmuludq(a, b);
15607 // AloBhi = pmuludq(a, Bhi);
15608 // AhiBlo = pmuludq(Ahi, b);
15610 // AloBhi = psllqi(AloBhi, 32);
15611 // AhiBlo = psllqi(AhiBlo, 32);
15612 // return AloBlo + AloBhi + AhiBlo;
15614 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
15615 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
15617 // Bit cast to 32-bit vectors for MULUDQ
15618 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
15619 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
15620 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
15621 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
15622 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
15623 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
15625 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
15626 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
15627 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
15629 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
15630 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
15632 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
15633 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
15636 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
15637 assert(Subtarget->isTargetWin64() && "Unexpected target");
15638 EVT VT = Op.getValueType();
15639 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
15640 "Unexpected return type for lowering");
15644 switch (Op->getOpcode()) {
15645 default: llvm_unreachable("Unexpected request for libcall!");
15646 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
15647 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
15648 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
15649 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
15650 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
15651 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
15655 SDValue InChain = DAG.getEntryNode();
15657 TargetLowering::ArgListTy Args;
15658 TargetLowering::ArgListEntry Entry;
15659 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
15660 EVT ArgVT = Op->getOperand(i).getValueType();
15661 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
15662 "Unexpected argument type for lowering");
15663 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
15664 Entry.Node = StackPtr;
15665 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
15667 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15668 Entry.Ty = PointerType::get(ArgTy,0);
15669 Entry.isSExt = false;
15670 Entry.isZExt = false;
15671 Args.push_back(Entry);
15674 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
15677 TargetLowering::CallLoweringInfo CLI(DAG);
15678 CLI.setDebugLoc(dl).setChain(InChain)
15679 .setCallee(getLibcallCallingConv(LC),
15680 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
15681 Callee, std::move(Args), 0)
15682 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
15684 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
15685 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
15688 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
15689 SelectionDAG &DAG) {
15690 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
15691 EVT VT = Op0.getValueType();
15694 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
15695 (VT == MVT::v8i32 && Subtarget->hasInt256()));
15697 // PMULxD operations multiply each even value (starting at 0) of LHS with
15698 // the related value of RHS and produce a widen result.
15699 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15700 // => <2 x i64> <ae|cg>
15702 // In other word, to have all the results, we need to perform two PMULxD:
15703 // 1. one with the even values.
15704 // 2. one with the odd values.
15705 // To achieve #2, with need to place the odd values at an even position.
15707 // Place the odd value at an even position (basically, shift all values 1
15708 // step to the left):
15709 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
15710 // <a|b|c|d> => <b|undef|d|undef>
15711 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
15712 // <e|f|g|h> => <f|undef|h|undef>
15713 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
15715 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
15717 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
15718 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
15720 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
15721 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15722 // => <2 x i64> <ae|cg>
15723 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
15724 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
15725 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
15726 // => <2 x i64> <bf|dh>
15727 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
15728 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
15730 // Shuffle it back into the right order.
15731 SDValue Highs, Lows;
15732 if (VT == MVT::v8i32) {
15733 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
15734 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15735 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
15736 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15738 const int HighMask[] = {1, 5, 3, 7};
15739 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15740 const int LowMask[] = {0, 4, 2, 6};
15741 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15744 // If we have a signed multiply but no PMULDQ fix up the high parts of a
15745 // unsigned multiply.
15746 if (IsSigned && !Subtarget->hasSSE41()) {
15748 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
15749 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
15750 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
15751 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
15752 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
15754 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
15755 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
15758 // The first result of MUL_LOHI is actually the low value, followed by the
15760 SDValue Ops[] = {Lows, Highs};
15761 return DAG.getMergeValues(Ops, dl);
15764 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
15765 const X86Subtarget *Subtarget) {
15766 MVT VT = Op.getSimpleValueType();
15768 SDValue R = Op.getOperand(0);
15769 SDValue Amt = Op.getOperand(1);
15771 // Optimize shl/srl/sra with constant shift amount.
15772 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
15773 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
15774 uint64_t ShiftAmt = ShiftConst->getZExtValue();
15776 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
15777 (Subtarget->hasInt256() &&
15778 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15779 (Subtarget->hasAVX512() &&
15780 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15781 if (Op.getOpcode() == ISD::SHL)
15782 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15784 if (Op.getOpcode() == ISD::SRL)
15785 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15787 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
15788 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15792 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
15793 unsigned NumElts = VT.getVectorNumElements();
15794 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
15796 if (Op.getOpcode() == ISD::SHL) {
15797 // Make a large shift.
15798 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
15800 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
15801 // Zero out the rightmost bits.
15802 SmallVector<SDValue, 32> V(
15803 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), MVT::i8));
15804 return DAG.getNode(ISD::AND, dl, VT, SHL,
15805 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15807 if (Op.getOpcode() == ISD::SRL) {
15808 // Make a large shift.
15809 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
15811 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
15812 // Zero out the leftmost bits.
15813 SmallVector<SDValue, 32> V(
15814 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, MVT::i8));
15815 return DAG.getNode(ISD::AND, dl, VT, SRL,
15816 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15818 if (Op.getOpcode() == ISD::SRA) {
15819 if (ShiftAmt == 7) {
15820 // R s>> 7 === R s< 0
15821 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15822 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
15825 // R s>> a === ((R u>> a) ^ m) - m
15826 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
15827 SmallVector<SDValue, 32> V(NumElts,
15828 DAG.getConstant(128 >> ShiftAmt, MVT::i8));
15829 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
15830 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
15831 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
15834 llvm_unreachable("Unknown shift opcode.");
15839 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15840 if (!Subtarget->is64Bit() &&
15841 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
15842 Amt.getOpcode() == ISD::BITCAST &&
15843 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15844 Amt = Amt.getOperand(0);
15845 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15846 VT.getVectorNumElements();
15847 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
15848 uint64_t ShiftAmt = 0;
15849 for (unsigned i = 0; i != Ratio; ++i) {
15850 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
15854 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
15856 // Check remaining shift amounts.
15857 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15858 uint64_t ShAmt = 0;
15859 for (unsigned j = 0; j != Ratio; ++j) {
15860 ConstantSDNode *C =
15861 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
15865 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
15867 if (ShAmt != ShiftAmt)
15870 switch (Op.getOpcode()) {
15872 llvm_unreachable("Unknown shift opcode!");
15874 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15877 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15880 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15888 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
15889 const X86Subtarget* Subtarget) {
15890 MVT VT = Op.getSimpleValueType();
15892 SDValue R = Op.getOperand(0);
15893 SDValue Amt = Op.getOperand(1);
15895 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
15896 VT == MVT::v4i32 || VT == MVT::v8i16 ||
15897 (Subtarget->hasInt256() &&
15898 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
15899 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15900 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15902 EVT EltVT = VT.getVectorElementType();
15904 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
15905 // Check if this build_vector node is doing a splat.
15906 // If so, then set BaseShAmt equal to the splat value.
15907 BaseShAmt = BV->getSplatValue();
15908 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
15909 BaseShAmt = SDValue();
15911 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
15912 Amt = Amt.getOperand(0);
15914 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
15915 if (SVN && SVN->isSplat()) {
15916 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
15917 SDValue InVec = Amt.getOperand(0);
15918 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
15919 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
15920 "Unexpected shuffle index found!");
15921 BaseShAmt = InVec.getOperand(SplatIdx);
15922 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
15923 if (ConstantSDNode *C =
15924 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
15925 if (C->getZExtValue() == SplatIdx)
15926 BaseShAmt = InVec.getOperand(1);
15931 // Avoid introducing an extract element from a shuffle.
15932 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
15933 DAG.getIntPtrConstant(SplatIdx));
15937 if (BaseShAmt.getNode()) {
15938 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
15939 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
15940 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
15941 else if (EltVT.bitsLT(MVT::i32))
15942 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
15944 switch (Op.getOpcode()) {
15946 llvm_unreachable("Unknown shift opcode!");
15948 switch (VT.SimpleTy) {
15949 default: return SDValue();
15958 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
15961 switch (VT.SimpleTy) {
15962 default: return SDValue();
15969 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
15972 switch (VT.SimpleTy) {
15973 default: return SDValue();
15982 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
15988 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15989 if (!Subtarget->is64Bit() &&
15990 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
15991 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
15992 Amt.getOpcode() == ISD::BITCAST &&
15993 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15994 Amt = Amt.getOperand(0);
15995 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15996 VT.getVectorNumElements();
15997 std::vector<SDValue> Vals(Ratio);
15998 for (unsigned i = 0; i != Ratio; ++i)
15999 Vals[i] = Amt.getOperand(i);
16000 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16001 for (unsigned j = 0; j != Ratio; ++j)
16002 if (Vals[j] != Amt.getOperand(i + j))
16005 switch (Op.getOpcode()) {
16007 llvm_unreachable("Unknown shift opcode!");
16009 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
16011 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
16013 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
16020 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16021 SelectionDAG &DAG) {
16022 MVT VT = Op.getSimpleValueType();
16024 SDValue R = Op.getOperand(0);
16025 SDValue Amt = Op.getOperand(1);
16028 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16029 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16031 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
16035 V = LowerScalarVariableShift(Op, DAG, Subtarget);
16039 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
16041 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
16042 if (Subtarget->hasInt256()) {
16043 if (Op.getOpcode() == ISD::SRL &&
16044 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16045 VT == MVT::v4i64 || VT == MVT::v8i32))
16047 if (Op.getOpcode() == ISD::SHL &&
16048 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16049 VT == MVT::v4i64 || VT == MVT::v8i32))
16051 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
16055 // If possible, lower this packed shift into a vector multiply instead of
16056 // expanding it into a sequence of scalar shifts.
16057 // Do this only if the vector shift count is a constant build_vector.
16058 if (Op.getOpcode() == ISD::SHL &&
16059 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16060 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16061 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16062 SmallVector<SDValue, 8> Elts;
16063 EVT SVT = VT.getScalarType();
16064 unsigned SVTBits = SVT.getSizeInBits();
16065 const APInt &One = APInt(SVTBits, 1);
16066 unsigned NumElems = VT.getVectorNumElements();
16068 for (unsigned i=0; i !=NumElems; ++i) {
16069 SDValue Op = Amt->getOperand(i);
16070 if (Op->getOpcode() == ISD::UNDEF) {
16071 Elts.push_back(Op);
16075 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16076 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16077 uint64_t ShAmt = C.getZExtValue();
16078 if (ShAmt >= SVTBits) {
16079 Elts.push_back(DAG.getUNDEF(SVT));
16082 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
16084 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16085 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16088 // Lower SHL with variable shift amount.
16089 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16090 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
16092 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
16093 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16094 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16095 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16098 // If possible, lower this shift as a sequence of two shifts by
16099 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16101 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16103 // Could be rewritten as:
16104 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16106 // The advantage is that the two shifts from the example would be
16107 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16108 // the vector shift into four scalar shifts plus four pairs of vector
16110 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16111 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16112 unsigned TargetOpcode = X86ISD::MOVSS;
16113 bool CanBeSimplified;
16114 // The splat value for the first packed shift (the 'X' from the example).
16115 SDValue Amt1 = Amt->getOperand(0);
16116 // The splat value for the second packed shift (the 'Y' from the example).
16117 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16118 Amt->getOperand(2);
16120 // See if it is possible to replace this node with a sequence of
16121 // two shifts followed by a MOVSS/MOVSD
16122 if (VT == MVT::v4i32) {
16123 // Check if it is legal to use a MOVSS.
16124 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16125 Amt2 == Amt->getOperand(3);
16126 if (!CanBeSimplified) {
16127 // Otherwise, check if we can still simplify this node using a MOVSD.
16128 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16129 Amt->getOperand(2) == Amt->getOperand(3);
16130 TargetOpcode = X86ISD::MOVSD;
16131 Amt2 = Amt->getOperand(2);
16134 // Do similar checks for the case where the machine value type
16136 CanBeSimplified = Amt1 == Amt->getOperand(1);
16137 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16138 CanBeSimplified = Amt2 == Amt->getOperand(i);
16140 if (!CanBeSimplified) {
16141 TargetOpcode = X86ISD::MOVSD;
16142 CanBeSimplified = true;
16143 Amt2 = Amt->getOperand(4);
16144 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16145 CanBeSimplified = Amt1 == Amt->getOperand(i);
16146 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16147 CanBeSimplified = Amt2 == Amt->getOperand(j);
16151 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16152 isa<ConstantSDNode>(Amt2)) {
16153 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16154 EVT CastVT = MVT::v4i32;
16156 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
16157 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16159 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
16160 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16161 if (TargetOpcode == X86ISD::MOVSD)
16162 CastVT = MVT::v2i64;
16163 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16164 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16165 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16167 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16171 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16172 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
16175 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
16176 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
16178 // Turn 'a' into a mask suitable for VSELECT
16179 SDValue VSelM = DAG.getConstant(0x80, VT);
16180 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16181 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16183 SDValue CM1 = DAG.getConstant(0x0f, VT);
16184 SDValue CM2 = DAG.getConstant(0x3f, VT);
16186 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
16187 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
16188 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
16189 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16190 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16193 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16194 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16195 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16197 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
16198 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
16199 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
16200 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16201 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16204 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16205 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16206 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16208 // return VSELECT(r, r+r, a);
16209 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16210 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16214 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16215 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16216 // solution better.
16217 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16218 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16220 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16221 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16222 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16223 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16224 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16227 // Decompose 256-bit shifts into smaller 128-bit shifts.
16228 if (VT.is256BitVector()) {
16229 unsigned NumElems = VT.getVectorNumElements();
16230 MVT EltVT = VT.getVectorElementType();
16231 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16233 // Extract the two vectors
16234 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16235 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16237 // Recreate the shift amount vectors
16238 SDValue Amt1, Amt2;
16239 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16240 // Constant shift amount
16241 SmallVector<SDValue, 4> Amt1Csts;
16242 SmallVector<SDValue, 4> Amt2Csts;
16243 for (unsigned i = 0; i != NumElems/2; ++i)
16244 Amt1Csts.push_back(Amt->getOperand(i));
16245 for (unsigned i = NumElems/2; i != NumElems; ++i)
16246 Amt2Csts.push_back(Amt->getOperand(i));
16248 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16249 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16251 // Variable shift amount
16252 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16253 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16256 // Issue new vector shifts for the smaller types
16257 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16258 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16260 // Concatenate the result back
16261 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16267 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16268 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16269 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16270 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16271 // has only one use.
16272 SDNode *N = Op.getNode();
16273 SDValue LHS = N->getOperand(0);
16274 SDValue RHS = N->getOperand(1);
16275 unsigned BaseOp = 0;
16278 switch (Op.getOpcode()) {
16279 default: llvm_unreachable("Unknown ovf instruction!");
16281 // A subtract of one will be selected as a INC. Note that INC doesn't
16282 // set CF, so we can't do this for UADDO.
16283 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16285 BaseOp = X86ISD::INC;
16286 Cond = X86::COND_O;
16289 BaseOp = X86ISD::ADD;
16290 Cond = X86::COND_O;
16293 BaseOp = X86ISD::ADD;
16294 Cond = X86::COND_B;
16297 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16298 // set CF, so we can't do this for USUBO.
16299 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16301 BaseOp = X86ISD::DEC;
16302 Cond = X86::COND_O;
16305 BaseOp = X86ISD::SUB;
16306 Cond = X86::COND_O;
16309 BaseOp = X86ISD::SUB;
16310 Cond = X86::COND_B;
16313 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16314 Cond = X86::COND_O;
16316 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16317 if (N->getValueType(0) == MVT::i8) {
16318 BaseOp = X86ISD::UMUL8;
16319 Cond = X86::COND_O;
16322 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16324 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16327 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16328 DAG.getConstant(X86::COND_O, MVT::i32),
16329 SDValue(Sum.getNode(), 2));
16331 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16335 // Also sets EFLAGS.
16336 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16337 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16340 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16341 DAG.getConstant(Cond, MVT::i32),
16342 SDValue(Sum.getNode(), 1));
16344 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16347 /// Returns true if the operand type is exactly twice the native width, and
16348 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16349 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16350 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16351 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16352 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16355 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16356 else if (OpWidth == 128)
16357 return Subtarget->hasCmpxchg16b();
16362 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16363 return needsCmpXchgNb(SI->getValueOperand()->getType());
16366 // Note: this turns large loads into lock cmpxchg8b/16b.
16367 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16368 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16369 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16370 return needsCmpXchgNb(PTy->getElementType());
16373 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16374 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16375 const Type *MemType = AI->getType();
16377 // If the operand is too big, we must see if cmpxchg8/16b is available
16378 // and default to library calls otherwise.
16379 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16380 return needsCmpXchgNb(MemType);
16382 AtomicRMWInst::BinOp Op = AI->getOperation();
16385 llvm_unreachable("Unknown atomic operation");
16386 case AtomicRMWInst::Xchg:
16387 case AtomicRMWInst::Add:
16388 case AtomicRMWInst::Sub:
16389 // It's better to use xadd, xsub or xchg for these in all cases.
16391 case AtomicRMWInst::Or:
16392 case AtomicRMWInst::And:
16393 case AtomicRMWInst::Xor:
16394 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16395 // prefix to a normal instruction for these operations.
16396 return !AI->use_empty();
16397 case AtomicRMWInst::Nand:
16398 case AtomicRMWInst::Max:
16399 case AtomicRMWInst::Min:
16400 case AtomicRMWInst::UMax:
16401 case AtomicRMWInst::UMin:
16402 // These always require a non-trivial set of data operations on x86. We must
16403 // use a cmpxchg loop.
16408 static bool hasMFENCE(const X86Subtarget& Subtarget) {
16409 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
16410 // no-sse2). There isn't any reason to disable it if the target processor
16412 return Subtarget.hasSSE2() || Subtarget.is64Bit();
16416 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
16417 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16418 const Type *MemType = AI->getType();
16419 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
16420 // there is no benefit in turning such RMWs into loads, and it is actually
16421 // harmful as it introduces a mfence.
16422 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16425 auto Builder = IRBuilder<>(AI);
16426 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16427 auto SynchScope = AI->getSynchScope();
16428 // We must restrict the ordering to avoid generating loads with Release or
16429 // ReleaseAcquire orderings.
16430 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
16431 auto Ptr = AI->getPointerOperand();
16433 // Before the load we need a fence. Here is an example lifted from
16434 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
16437 // x.store(1, relaxed);
16438 // r1 = y.fetch_add(0, release);
16440 // y.fetch_add(42, acquire);
16441 // r2 = x.load(relaxed);
16442 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
16443 // lowered to just a load without a fence. A mfence flushes the store buffer,
16444 // making the optimization clearly correct.
16445 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
16446 // otherwise, we might be able to be more agressive on relaxed idempotent
16447 // rmw. In practice, they do not look useful, so we don't try to be
16448 // especially clever.
16449 if (SynchScope == SingleThread) {
16450 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
16451 // the IR level, so we must wrap it in an intrinsic.
16453 } else if (hasMFENCE(*Subtarget)) {
16454 Function *MFence = llvm::Intrinsic::getDeclaration(M,
16455 Intrinsic::x86_sse2_mfence);
16456 Builder.CreateCall(MFence);
16458 // FIXME: it might make sense to use a locked operation here but on a
16459 // different cache-line to prevent cache-line bouncing. In practice it
16460 // is probably a small win, and x86 processors without mfence are rare
16461 // enough that we do not bother.
16465 // Finally we can emit the atomic load.
16466 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
16467 AI->getType()->getPrimitiveSizeInBits());
16468 Loaded->setAtomic(Order, SynchScope);
16469 AI->replaceAllUsesWith(Loaded);
16470 AI->eraseFromParent();
16474 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
16475 SelectionDAG &DAG) {
16477 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
16478 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
16479 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
16480 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
16482 // The only fence that needs an instruction is a sequentially-consistent
16483 // cross-thread fence.
16484 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
16485 if (hasMFENCE(*Subtarget))
16486 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
16488 SDValue Chain = Op.getOperand(0);
16489 SDValue Zero = DAG.getConstant(0, MVT::i32);
16491 DAG.getRegister(X86::ESP, MVT::i32), // Base
16492 DAG.getTargetConstant(1, MVT::i8), // Scale
16493 DAG.getRegister(0, MVT::i32), // Index
16494 DAG.getTargetConstant(0, MVT::i32), // Disp
16495 DAG.getRegister(0, MVT::i32), // Segment.
16499 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
16500 return SDValue(Res, 0);
16503 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
16504 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
16507 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
16508 SelectionDAG &DAG) {
16509 MVT T = Op.getSimpleValueType();
16513 switch(T.SimpleTy) {
16514 default: llvm_unreachable("Invalid value type!");
16515 case MVT::i8: Reg = X86::AL; size = 1; break;
16516 case MVT::i16: Reg = X86::AX; size = 2; break;
16517 case MVT::i32: Reg = X86::EAX; size = 4; break;
16519 assert(Subtarget->is64Bit() && "Node not type legal!");
16520 Reg = X86::RAX; size = 8;
16523 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
16524 Op.getOperand(2), SDValue());
16525 SDValue Ops[] = { cpIn.getValue(0),
16528 DAG.getTargetConstant(size, MVT::i8),
16529 cpIn.getValue(1) };
16530 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16531 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
16532 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
16536 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
16537 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
16538 MVT::i32, cpOut.getValue(2));
16539 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
16540 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
16542 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
16543 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
16544 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
16548 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
16549 SelectionDAG &DAG) {
16550 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
16551 MVT DstVT = Op.getSimpleValueType();
16553 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
16554 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16555 if (DstVT != MVT::f64)
16556 // This conversion needs to be expanded.
16559 SDValue InVec = Op->getOperand(0);
16561 unsigned NumElts = SrcVT.getVectorNumElements();
16562 EVT SVT = SrcVT.getVectorElementType();
16564 // Widen the vector in input in the case of MVT::v2i32.
16565 // Example: from MVT::v2i32 to MVT::v4i32.
16566 SmallVector<SDValue, 16> Elts;
16567 for (unsigned i = 0, e = NumElts; i != e; ++i)
16568 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
16569 DAG.getIntPtrConstant(i)));
16571 // Explicitly mark the extra elements as Undef.
16572 Elts.append(NumElts, DAG.getUNDEF(SVT));
16574 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
16575 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
16576 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
16577 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
16578 DAG.getIntPtrConstant(0));
16581 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
16582 Subtarget->hasMMX() && "Unexpected custom BITCAST");
16583 assert((DstVT == MVT::i64 ||
16584 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
16585 "Unexpected custom BITCAST");
16586 // i64 <=> MMX conversions are Legal.
16587 if (SrcVT==MVT::i64 && DstVT.isVector())
16589 if (DstVT==MVT::i64 && SrcVT.isVector())
16591 // MMX <=> MMX conversions are Legal.
16592 if (SrcVT.isVector() && DstVT.isVector())
16594 // All other conversions need to be expanded.
16598 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
16599 SelectionDAG &DAG) {
16600 SDNode *Node = Op.getNode();
16603 Op = Op.getOperand(0);
16604 EVT VT = Op.getValueType();
16605 assert((VT.is128BitVector() || VT.is256BitVector()) &&
16606 "CTPOP lowering only implemented for 128/256-bit wide vector types");
16608 unsigned NumElts = VT.getVectorNumElements();
16609 EVT EltVT = VT.getVectorElementType();
16610 unsigned Len = EltVT.getSizeInBits();
16612 // This is the vectorized version of the "best" algorithm from
16613 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
16614 // with a minor tweak to use a series of adds + shifts instead of vector
16615 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
16617 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
16618 // v8i32 => Always profitable
16620 // FIXME: There a couple of possible improvements:
16622 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
16623 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
16625 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
16626 "CTPOP not implemented for this vector element type.");
16628 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
16629 // extra legalization.
16630 bool NeedsBitcast = EltVT == MVT::i32;
16631 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
16633 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
16634 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
16635 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
16637 // v = v - ((v >> 1) & 0x55555555...)
16638 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
16639 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
16640 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
16642 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16644 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
16645 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
16647 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
16649 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
16650 if (VT != And.getValueType())
16651 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16652 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
16654 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
16655 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
16656 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
16657 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
16658 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
16660 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
16661 if (NeedsBitcast) {
16662 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16663 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
16664 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
16667 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
16668 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
16669 if (VT != AndRHS.getValueType()) {
16670 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
16671 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
16673 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
16675 // v = (v + (v >> 4)) & 0x0F0F0F0F...
16676 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
16677 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
16678 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
16679 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16681 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
16682 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
16683 if (NeedsBitcast) {
16684 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16685 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
16687 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
16688 if (VT != And.getValueType())
16689 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16691 // The algorithm mentioned above uses:
16692 // v = (v * 0x01010101...) >> (Len - 8)
16694 // Change it to use vector adds + vector shifts which yield faster results on
16695 // Haswell than using vector integer multiplication.
16697 // For i32 elements:
16698 // v = v + (v >> 8)
16699 // v = v + (v >> 16)
16701 // For i64 elements:
16702 // v = v + (v >> 8)
16703 // v = v + (v >> 16)
16704 // v = v + (v >> 32)
16707 SmallVector<SDValue, 8> Csts;
16708 for (unsigned i = 8; i <= Len/2; i *= 2) {
16709 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
16710 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
16711 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
16712 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16716 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
16717 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
16718 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
16719 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
16720 if (NeedsBitcast) {
16721 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16722 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
16724 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
16725 if (VT != And.getValueType())
16726 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16731 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
16732 SDNode *Node = Op.getNode();
16734 EVT T = Node->getValueType(0);
16735 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
16736 DAG.getConstant(0, T), Node->getOperand(2));
16737 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
16738 cast<AtomicSDNode>(Node)->getMemoryVT(),
16739 Node->getOperand(0),
16740 Node->getOperand(1), negOp,
16741 cast<AtomicSDNode>(Node)->getMemOperand(),
16742 cast<AtomicSDNode>(Node)->getOrdering(),
16743 cast<AtomicSDNode>(Node)->getSynchScope());
16746 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
16747 SDNode *Node = Op.getNode();
16749 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
16751 // Convert seq_cst store -> xchg
16752 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
16753 // FIXME: On 32-bit, store -> fist or movq would be more efficient
16754 // (The only way to get a 16-byte store is cmpxchg16b)
16755 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
16756 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
16757 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
16758 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
16759 cast<AtomicSDNode>(Node)->getMemoryVT(),
16760 Node->getOperand(0),
16761 Node->getOperand(1), Node->getOperand(2),
16762 cast<AtomicSDNode>(Node)->getMemOperand(),
16763 cast<AtomicSDNode>(Node)->getOrdering(),
16764 cast<AtomicSDNode>(Node)->getSynchScope());
16765 return Swap.getValue(1);
16767 // Other atomic stores have a simple pattern.
16771 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
16772 EVT VT = Op.getNode()->getSimpleValueType(0);
16774 // Let legalize expand this if it isn't a legal type yet.
16775 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
16778 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16781 bool ExtraOp = false;
16782 switch (Op.getOpcode()) {
16783 default: llvm_unreachable("Invalid code");
16784 case ISD::ADDC: Opc = X86ISD::ADD; break;
16785 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
16786 case ISD::SUBC: Opc = X86ISD::SUB; break;
16787 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
16791 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16793 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16794 Op.getOperand(1), Op.getOperand(2));
16797 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
16798 SelectionDAG &DAG) {
16799 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
16801 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
16802 // which returns the values as { float, float } (in XMM0) or
16803 // { double, double } (which is returned in XMM0, XMM1).
16805 SDValue Arg = Op.getOperand(0);
16806 EVT ArgVT = Arg.getValueType();
16807 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16809 TargetLowering::ArgListTy Args;
16810 TargetLowering::ArgListEntry Entry;
16814 Entry.isSExt = false;
16815 Entry.isZExt = false;
16816 Args.push_back(Entry);
16818 bool isF64 = ArgVT == MVT::f64;
16819 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
16820 // the small struct {f32, f32} is returned in (eax, edx). For f64,
16821 // the results are returned via SRet in memory.
16822 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
16823 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16824 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
16826 Type *RetTy = isF64
16827 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
16828 : (Type*)VectorType::get(ArgTy, 4);
16830 TargetLowering::CallLoweringInfo CLI(DAG);
16831 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
16832 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
16834 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
16837 // Returned in xmm0 and xmm1.
16838 return CallResult.first;
16840 // Returned in bits 0:31 and 32:64 xmm0.
16841 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16842 CallResult.first, DAG.getIntPtrConstant(0));
16843 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16844 CallResult.first, DAG.getIntPtrConstant(1));
16845 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
16846 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
16849 /// LowerOperation - Provide custom lowering hooks for some operations.
16851 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
16852 switch (Op.getOpcode()) {
16853 default: llvm_unreachable("Should not custom lower this!");
16854 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
16855 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
16856 return LowerCMP_SWAP(Op, Subtarget, DAG);
16857 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
16858 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
16859 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
16860 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
16861 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
16862 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
16863 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
16864 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
16865 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
16866 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
16867 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
16868 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
16869 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
16870 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
16871 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
16872 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
16873 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
16874 case ISD::SHL_PARTS:
16875 case ISD::SRA_PARTS:
16876 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
16877 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
16878 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
16879 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
16880 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
16881 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
16882 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
16883 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
16884 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
16885 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
16886 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
16888 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
16889 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
16890 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
16891 case ISD::SETCC: return LowerSETCC(Op, DAG);
16892 case ISD::SELECT: return LowerSELECT(Op, DAG);
16893 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
16894 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
16895 case ISD::VASTART: return LowerVASTART(Op, DAG);
16896 case ISD::VAARG: return LowerVAARG(Op, DAG);
16897 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
16898 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
16899 case ISD::INTRINSIC_VOID:
16900 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
16901 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
16902 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
16903 case ISD::FRAME_TO_ARGS_OFFSET:
16904 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
16905 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
16906 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
16907 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
16908 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
16909 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
16910 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
16911 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
16912 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
16913 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
16914 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
16915 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
16916 case ISD::UMUL_LOHI:
16917 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
16920 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
16926 case ISD::UMULO: return LowerXALUO(Op, DAG);
16927 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
16928 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
16932 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
16933 case ISD::ADD: return LowerADD(Op, DAG);
16934 case ISD::SUB: return LowerSUB(Op, DAG);
16935 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
16939 /// ReplaceNodeResults - Replace a node with an illegal result type
16940 /// with a new node built out of custom code.
16941 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
16942 SmallVectorImpl<SDValue>&Results,
16943 SelectionDAG &DAG) const {
16945 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16946 switch (N->getOpcode()) {
16948 llvm_unreachable("Do not know how to custom type legalize this operation!");
16949 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
16950 case X86ISD::FMINC:
16952 case X86ISD::FMAXC:
16953 case X86ISD::FMAX: {
16954 EVT VT = N->getValueType(0);
16955 if (VT != MVT::v2f32)
16956 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
16957 SDValue UNDEF = DAG.getUNDEF(VT);
16958 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
16959 N->getOperand(0), UNDEF);
16960 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
16961 N->getOperand(1), UNDEF);
16962 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
16965 case ISD::SIGN_EXTEND_INREG:
16970 // We don't want to expand or promote these.
16977 case ISD::UDIVREM: {
16978 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
16979 Results.push_back(V);
16982 case ISD::FP_TO_SINT:
16983 case ISD::FP_TO_UINT: {
16984 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
16986 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
16989 std::pair<SDValue,SDValue> Vals =
16990 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
16991 SDValue FIST = Vals.first, StackSlot = Vals.second;
16992 if (FIST.getNode()) {
16993 EVT VT = N->getValueType(0);
16994 // Return a load from the stack slot.
16995 if (StackSlot.getNode())
16996 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
16997 MachinePointerInfo(),
16998 false, false, false, 0));
17000 Results.push_back(FIST);
17004 case ISD::UINT_TO_FP: {
17005 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17006 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17007 N->getValueType(0) != MVT::v2f32)
17009 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17011 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
17013 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17014 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17015 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17016 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17017 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17018 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17021 case ISD::FP_ROUND: {
17022 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17024 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17025 Results.push_back(V);
17028 case ISD::INTRINSIC_W_CHAIN: {
17029 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17031 default : llvm_unreachable("Do not know how to custom type "
17032 "legalize this intrinsic operation!");
17033 case Intrinsic::x86_rdtsc:
17034 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17036 case Intrinsic::x86_rdtscp:
17037 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17039 case Intrinsic::x86_rdpmc:
17040 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17043 case ISD::READCYCLECOUNTER: {
17044 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17047 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17048 EVT T = N->getValueType(0);
17049 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17050 bool Regs64bit = T == MVT::i128;
17051 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17052 SDValue cpInL, cpInH;
17053 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17054 DAG.getConstant(0, HalfT));
17055 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17056 DAG.getConstant(1, HalfT));
17057 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17058 Regs64bit ? X86::RAX : X86::EAX,
17060 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17061 Regs64bit ? X86::RDX : X86::EDX,
17062 cpInH, cpInL.getValue(1));
17063 SDValue swapInL, swapInH;
17064 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17065 DAG.getConstant(0, HalfT));
17066 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17067 DAG.getConstant(1, HalfT));
17068 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17069 Regs64bit ? X86::RBX : X86::EBX,
17070 swapInL, cpInH.getValue(1));
17071 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17072 Regs64bit ? X86::RCX : X86::ECX,
17073 swapInH, swapInL.getValue(1));
17074 SDValue Ops[] = { swapInH.getValue(0),
17076 swapInH.getValue(1) };
17077 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17078 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17079 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17080 X86ISD::LCMPXCHG8_DAG;
17081 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17082 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17083 Regs64bit ? X86::RAX : X86::EAX,
17084 HalfT, Result.getValue(1));
17085 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17086 Regs64bit ? X86::RDX : X86::EDX,
17087 HalfT, cpOutL.getValue(2));
17088 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17090 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17091 MVT::i32, cpOutH.getValue(2));
17093 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17094 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17095 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17097 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17098 Results.push_back(Success);
17099 Results.push_back(EFLAGS.getValue(1));
17102 case ISD::ATOMIC_SWAP:
17103 case ISD::ATOMIC_LOAD_ADD:
17104 case ISD::ATOMIC_LOAD_SUB:
17105 case ISD::ATOMIC_LOAD_AND:
17106 case ISD::ATOMIC_LOAD_OR:
17107 case ISD::ATOMIC_LOAD_XOR:
17108 case ISD::ATOMIC_LOAD_NAND:
17109 case ISD::ATOMIC_LOAD_MIN:
17110 case ISD::ATOMIC_LOAD_MAX:
17111 case ISD::ATOMIC_LOAD_UMIN:
17112 case ISD::ATOMIC_LOAD_UMAX:
17113 case ISD::ATOMIC_LOAD: {
17114 // Delegate to generic TypeLegalization. Situations we can really handle
17115 // should have already been dealt with by AtomicExpandPass.cpp.
17118 case ISD::BITCAST: {
17119 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17120 EVT DstVT = N->getValueType(0);
17121 EVT SrcVT = N->getOperand(0)->getValueType(0);
17123 if (SrcVT != MVT::f64 ||
17124 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17127 unsigned NumElts = DstVT.getVectorNumElements();
17128 EVT SVT = DstVT.getVectorElementType();
17129 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17130 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17131 MVT::v2f64, N->getOperand(0));
17132 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17134 if (ExperimentalVectorWideningLegalization) {
17135 // If we are legalizing vectors by widening, we already have the desired
17136 // legal vector type, just return it.
17137 Results.push_back(ToVecInt);
17141 SmallVector<SDValue, 8> Elts;
17142 for (unsigned i = 0, e = NumElts; i != e; ++i)
17143 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17144 ToVecInt, DAG.getIntPtrConstant(i)));
17146 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17151 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17153 default: return nullptr;
17154 case X86ISD::BSF: return "X86ISD::BSF";
17155 case X86ISD::BSR: return "X86ISD::BSR";
17156 case X86ISD::SHLD: return "X86ISD::SHLD";
17157 case X86ISD::SHRD: return "X86ISD::SHRD";
17158 case X86ISD::FAND: return "X86ISD::FAND";
17159 case X86ISD::FANDN: return "X86ISD::FANDN";
17160 case X86ISD::FOR: return "X86ISD::FOR";
17161 case X86ISD::FXOR: return "X86ISD::FXOR";
17162 case X86ISD::FSRL: return "X86ISD::FSRL";
17163 case X86ISD::FILD: return "X86ISD::FILD";
17164 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17165 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17166 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17167 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17168 case X86ISD::FLD: return "X86ISD::FLD";
17169 case X86ISD::FST: return "X86ISD::FST";
17170 case X86ISD::CALL: return "X86ISD::CALL";
17171 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17172 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17173 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17174 case X86ISD::BT: return "X86ISD::BT";
17175 case X86ISD::CMP: return "X86ISD::CMP";
17176 case X86ISD::COMI: return "X86ISD::COMI";
17177 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17178 case X86ISD::CMPM: return "X86ISD::CMPM";
17179 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17180 case X86ISD::SETCC: return "X86ISD::SETCC";
17181 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17182 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17183 case X86ISD::CMOV: return "X86ISD::CMOV";
17184 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17185 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17186 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17187 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17188 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17189 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17190 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17191 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17192 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17193 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17194 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17195 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17196 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17197 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17198 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17199 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17200 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17201 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17202 case X86ISD::HADD: return "X86ISD::HADD";
17203 case X86ISD::HSUB: return "X86ISD::HSUB";
17204 case X86ISD::FHADD: return "X86ISD::FHADD";
17205 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17206 case X86ISD::UMAX: return "X86ISD::UMAX";
17207 case X86ISD::UMIN: return "X86ISD::UMIN";
17208 case X86ISD::SMAX: return "X86ISD::SMAX";
17209 case X86ISD::SMIN: return "X86ISD::SMIN";
17210 case X86ISD::FMAX: return "X86ISD::FMAX";
17211 case X86ISD::FMIN: return "X86ISD::FMIN";
17212 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17213 case X86ISD::FMINC: return "X86ISD::FMINC";
17214 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17215 case X86ISD::FRCP: return "X86ISD::FRCP";
17216 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17217 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17218 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17219 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17220 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17221 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17222 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17223 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17224 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17225 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17226 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17227 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17228 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17229 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17230 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17231 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17232 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17233 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17234 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17235 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17236 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17237 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17238 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17239 case X86ISD::VSHL: return "X86ISD::VSHL";
17240 case X86ISD::VSRL: return "X86ISD::VSRL";
17241 case X86ISD::VSRA: return "X86ISD::VSRA";
17242 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17243 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17244 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17245 case X86ISD::CMPP: return "X86ISD::CMPP";
17246 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17247 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17248 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17249 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17250 case X86ISD::ADD: return "X86ISD::ADD";
17251 case X86ISD::SUB: return "X86ISD::SUB";
17252 case X86ISD::ADC: return "X86ISD::ADC";
17253 case X86ISD::SBB: return "X86ISD::SBB";
17254 case X86ISD::SMUL: return "X86ISD::SMUL";
17255 case X86ISD::UMUL: return "X86ISD::UMUL";
17256 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17257 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17258 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17259 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17260 case X86ISD::INC: return "X86ISD::INC";
17261 case X86ISD::DEC: return "X86ISD::DEC";
17262 case X86ISD::OR: return "X86ISD::OR";
17263 case X86ISD::XOR: return "X86ISD::XOR";
17264 case X86ISD::AND: return "X86ISD::AND";
17265 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17266 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17267 case X86ISD::PTEST: return "X86ISD::PTEST";
17268 case X86ISD::TESTP: return "X86ISD::TESTP";
17269 case X86ISD::TESTM: return "X86ISD::TESTM";
17270 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17271 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17272 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17273 case X86ISD::PACKUS: return "X86ISD::PACKUS";
17274 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
17275 case X86ISD::VALIGN: return "X86ISD::VALIGN";
17276 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
17277 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
17278 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
17279 case X86ISD::SHUFP: return "X86ISD::SHUFP";
17280 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
17281 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
17282 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
17283 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
17284 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
17285 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
17286 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
17287 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
17288 case X86ISD::MOVSD: return "X86ISD::MOVSD";
17289 case X86ISD::MOVSS: return "X86ISD::MOVSS";
17290 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
17291 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
17292 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
17293 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
17294 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
17295 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
17296 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
17297 case X86ISD::VPERMV: return "X86ISD::VPERMV";
17298 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
17299 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
17300 case X86ISD::VPERMI: return "X86ISD::VPERMI";
17301 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
17302 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
17303 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
17304 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
17305 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
17306 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
17307 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
17308 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
17309 case X86ISD::SAHF: return "X86ISD::SAHF";
17310 case X86ISD::RDRAND: return "X86ISD::RDRAND";
17311 case X86ISD::RDSEED: return "X86ISD::RDSEED";
17312 case X86ISD::FMADD: return "X86ISD::FMADD";
17313 case X86ISD::FMSUB: return "X86ISD::FMSUB";
17314 case X86ISD::FNMADD: return "X86ISD::FNMADD";
17315 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
17316 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
17317 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
17318 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
17319 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
17320 case X86ISD::XTEST: return "X86ISD::XTEST";
17321 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
17322 case X86ISD::EXPAND: return "X86ISD::EXPAND";
17323 case X86ISD::SELECT: return "X86ISD::SELECT";
17324 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
17325 case X86ISD::RCP28: return "X86ISD::RCP28";
17326 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
17327 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
17328 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
17329 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
17330 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
17334 // isLegalAddressingMode - Return true if the addressing mode represented
17335 // by AM is legal for this target, for a load/store of the specified type.
17336 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
17338 // X86 supports extremely general addressing modes.
17339 CodeModel::Model M = getTargetMachine().getCodeModel();
17340 Reloc::Model R = getTargetMachine().getRelocationModel();
17342 // X86 allows a sign-extended 32-bit immediate field as a displacement.
17343 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
17348 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
17350 // If a reference to this global requires an extra load, we can't fold it.
17351 if (isGlobalStubReference(GVFlags))
17354 // If BaseGV requires a register for the PIC base, we cannot also have a
17355 // BaseReg specified.
17356 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
17359 // If lower 4G is not available, then we must use rip-relative addressing.
17360 if ((M != CodeModel::Small || R != Reloc::Static) &&
17361 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
17365 switch (AM.Scale) {
17371 // These scales always work.
17376 // These scales are formed with basereg+scalereg. Only accept if there is
17381 default: // Other stuff never works.
17388 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
17389 unsigned Bits = Ty->getScalarSizeInBits();
17391 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
17392 // particularly cheaper than those without.
17396 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
17397 // variable shifts just as cheap as scalar ones.
17398 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
17401 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
17402 // fully general vector.
17406 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
17407 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17409 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
17410 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
17411 return NumBits1 > NumBits2;
17414 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
17415 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17418 if (!isTypeLegal(EVT::getEVT(Ty1)))
17421 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
17423 // Assuming the caller doesn't have a zeroext or signext return parameter,
17424 // truncation all the way down to i1 is valid.
17428 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
17429 return isInt<32>(Imm);
17432 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
17433 // Can also use sub to handle negated immediates.
17434 return isInt<32>(Imm);
17437 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
17438 if (!VT1.isInteger() || !VT2.isInteger())
17440 unsigned NumBits1 = VT1.getSizeInBits();
17441 unsigned NumBits2 = VT2.getSizeInBits();
17442 return NumBits1 > NumBits2;
17445 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
17446 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17447 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
17450 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
17451 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17452 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
17455 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
17456 EVT VT1 = Val.getValueType();
17457 if (isZExtFree(VT1, VT2))
17460 if (Val.getOpcode() != ISD::LOAD)
17463 if (!VT1.isSimple() || !VT1.isInteger() ||
17464 !VT2.isSimple() || !VT2.isInteger())
17467 switch (VT1.getSimpleVT().SimpleTy) {
17472 // X86 has 8, 16, and 32-bit zero-extending loads.
17479 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
17482 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
17483 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
17486 VT = VT.getScalarType();
17488 if (!VT.isSimple())
17491 switch (VT.getSimpleVT().SimpleTy) {
17502 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
17503 // i16 instructions are longer (0x66 prefix) and potentially slower.
17504 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
17507 /// isShuffleMaskLegal - Targets can use this to indicate that they only
17508 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
17509 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
17510 /// are assumed to be legal.
17512 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
17514 if (!VT.isSimple())
17517 // Very little shuffling can be done for 64-bit vectors right now.
17518 if (VT.getSizeInBits() == 64)
17521 // We only care that the types being shuffled are legal. The lowering can
17522 // handle any possible shuffle mask that results.
17523 return isTypeLegal(VT.getSimpleVT());
17527 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
17529 // Just delegate to the generic legality, clear masks aren't special.
17530 return isShuffleMaskLegal(Mask, VT);
17533 //===----------------------------------------------------------------------===//
17534 // X86 Scheduler Hooks
17535 //===----------------------------------------------------------------------===//
17537 /// Utility function to emit xbegin specifying the start of an RTM region.
17538 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
17539 const TargetInstrInfo *TII) {
17540 DebugLoc DL = MI->getDebugLoc();
17542 const BasicBlock *BB = MBB->getBasicBlock();
17543 MachineFunction::iterator I = MBB;
17546 // For the v = xbegin(), we generate
17557 MachineBasicBlock *thisMBB = MBB;
17558 MachineFunction *MF = MBB->getParent();
17559 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
17560 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
17561 MF->insert(I, mainMBB);
17562 MF->insert(I, sinkMBB);
17564 // Transfer the remainder of BB and its successor edges to sinkMBB.
17565 sinkMBB->splice(sinkMBB->begin(), MBB,
17566 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17567 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
17571 // # fallthrough to mainMBB
17572 // # abortion to sinkMBB
17573 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
17574 thisMBB->addSuccessor(mainMBB);
17575 thisMBB->addSuccessor(sinkMBB);
17579 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
17580 mainMBB->addSuccessor(sinkMBB);
17583 // EAX is live into the sinkMBB
17584 sinkMBB->addLiveIn(X86::EAX);
17585 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
17586 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17589 MI->eraseFromParent();
17593 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
17594 // or XMM0_V32I8 in AVX all of this code can be replaced with that
17595 // in the .td file.
17596 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
17597 const TargetInstrInfo *TII) {
17599 switch (MI->getOpcode()) {
17600 default: llvm_unreachable("illegal opcode!");
17601 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
17602 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
17603 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
17604 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
17605 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
17606 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
17607 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
17608 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
17611 DebugLoc dl = MI->getDebugLoc();
17612 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17614 unsigned NumArgs = MI->getNumOperands();
17615 for (unsigned i = 1; i < NumArgs; ++i) {
17616 MachineOperand &Op = MI->getOperand(i);
17617 if (!(Op.isReg() && Op.isImplicit()))
17618 MIB.addOperand(Op);
17620 if (MI->hasOneMemOperand())
17621 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17623 BuildMI(*BB, MI, dl,
17624 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17625 .addReg(X86::XMM0);
17627 MI->eraseFromParent();
17631 // FIXME: Custom handling because TableGen doesn't support multiple implicit
17632 // defs in an instruction pattern
17633 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
17634 const TargetInstrInfo *TII) {
17636 switch (MI->getOpcode()) {
17637 default: llvm_unreachable("illegal opcode!");
17638 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
17639 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
17640 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
17641 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
17642 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
17643 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
17644 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
17645 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
17648 DebugLoc dl = MI->getDebugLoc();
17649 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17651 unsigned NumArgs = MI->getNumOperands(); // remove the results
17652 for (unsigned i = 1; i < NumArgs; ++i) {
17653 MachineOperand &Op = MI->getOperand(i);
17654 if (!(Op.isReg() && Op.isImplicit()))
17655 MIB.addOperand(Op);
17657 if (MI->hasOneMemOperand())
17658 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17660 BuildMI(*BB, MI, dl,
17661 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17664 MI->eraseFromParent();
17668 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
17669 const X86Subtarget *Subtarget) {
17670 DebugLoc dl = MI->getDebugLoc();
17671 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17672 // Address into RAX/EAX, other two args into ECX, EDX.
17673 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
17674 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
17675 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
17676 for (int i = 0; i < X86::AddrNumOperands; ++i)
17677 MIB.addOperand(MI->getOperand(i));
17679 unsigned ValOps = X86::AddrNumOperands;
17680 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
17681 .addReg(MI->getOperand(ValOps).getReg());
17682 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
17683 .addReg(MI->getOperand(ValOps+1).getReg());
17685 // The instruction doesn't actually take any operands though.
17686 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
17688 MI->eraseFromParent(); // The pseudo is gone now.
17692 MachineBasicBlock *
17693 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
17694 MachineBasicBlock *MBB) const {
17695 // Emit va_arg instruction on X86-64.
17697 // Operands to this pseudo-instruction:
17698 // 0 ) Output : destination address (reg)
17699 // 1-5) Input : va_list address (addr, i64mem)
17700 // 6 ) ArgSize : Size (in bytes) of vararg type
17701 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
17702 // 8 ) Align : Alignment of type
17703 // 9 ) EFLAGS (implicit-def)
17705 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
17706 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
17708 unsigned DestReg = MI->getOperand(0).getReg();
17709 MachineOperand &Base = MI->getOperand(1);
17710 MachineOperand &Scale = MI->getOperand(2);
17711 MachineOperand &Index = MI->getOperand(3);
17712 MachineOperand &Disp = MI->getOperand(4);
17713 MachineOperand &Segment = MI->getOperand(5);
17714 unsigned ArgSize = MI->getOperand(6).getImm();
17715 unsigned ArgMode = MI->getOperand(7).getImm();
17716 unsigned Align = MI->getOperand(8).getImm();
17718 // Memory Reference
17719 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
17720 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17721 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17723 // Machine Information
17724 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17725 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
17726 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
17727 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
17728 DebugLoc DL = MI->getDebugLoc();
17730 // struct va_list {
17733 // i64 overflow_area (address)
17734 // i64 reg_save_area (address)
17736 // sizeof(va_list) = 24
17737 // alignment(va_list) = 8
17739 unsigned TotalNumIntRegs = 6;
17740 unsigned TotalNumXMMRegs = 8;
17741 bool UseGPOffset = (ArgMode == 1);
17742 bool UseFPOffset = (ArgMode == 2);
17743 unsigned MaxOffset = TotalNumIntRegs * 8 +
17744 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
17746 /* Align ArgSize to a multiple of 8 */
17747 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
17748 bool NeedsAlign = (Align > 8);
17750 MachineBasicBlock *thisMBB = MBB;
17751 MachineBasicBlock *overflowMBB;
17752 MachineBasicBlock *offsetMBB;
17753 MachineBasicBlock *endMBB;
17755 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
17756 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
17757 unsigned OffsetReg = 0;
17759 if (!UseGPOffset && !UseFPOffset) {
17760 // If we only pull from the overflow region, we don't create a branch.
17761 // We don't need to alter control flow.
17762 OffsetDestReg = 0; // unused
17763 OverflowDestReg = DestReg;
17765 offsetMBB = nullptr;
17766 overflowMBB = thisMBB;
17769 // First emit code to check if gp_offset (or fp_offset) is below the bound.
17770 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
17771 // If not, pull from overflow_area. (branch to overflowMBB)
17776 // offsetMBB overflowMBB
17781 // Registers for the PHI in endMBB
17782 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
17783 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
17785 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17786 MachineFunction *MF = MBB->getParent();
17787 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17788 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17789 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17791 MachineFunction::iterator MBBIter = MBB;
17794 // Insert the new basic blocks
17795 MF->insert(MBBIter, offsetMBB);
17796 MF->insert(MBBIter, overflowMBB);
17797 MF->insert(MBBIter, endMBB);
17799 // Transfer the remainder of MBB and its successor edges to endMBB.
17800 endMBB->splice(endMBB->begin(), thisMBB,
17801 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
17802 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
17804 // Make offsetMBB and overflowMBB successors of thisMBB
17805 thisMBB->addSuccessor(offsetMBB);
17806 thisMBB->addSuccessor(overflowMBB);
17808 // endMBB is a successor of both offsetMBB and overflowMBB
17809 offsetMBB->addSuccessor(endMBB);
17810 overflowMBB->addSuccessor(endMBB);
17812 // Load the offset value into a register
17813 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17814 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
17818 .addDisp(Disp, UseFPOffset ? 4 : 0)
17819 .addOperand(Segment)
17820 .setMemRefs(MMOBegin, MMOEnd);
17822 // Check if there is enough room left to pull this argument.
17823 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
17825 .addImm(MaxOffset + 8 - ArgSizeA8);
17827 // Branch to "overflowMBB" if offset >= max
17828 // Fall through to "offsetMBB" otherwise
17829 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
17830 .addMBB(overflowMBB);
17833 // In offsetMBB, emit code to use the reg_save_area.
17835 assert(OffsetReg != 0);
17837 // Read the reg_save_area address.
17838 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
17839 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
17844 .addOperand(Segment)
17845 .setMemRefs(MMOBegin, MMOEnd);
17847 // Zero-extend the offset
17848 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
17849 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
17852 .addImm(X86::sub_32bit);
17854 // Add the offset to the reg_save_area to get the final address.
17855 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
17856 .addReg(OffsetReg64)
17857 .addReg(RegSaveReg);
17859 // Compute the offset for the next argument
17860 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17861 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
17863 .addImm(UseFPOffset ? 16 : 8);
17865 // Store it back into the va_list.
17866 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
17870 .addDisp(Disp, UseFPOffset ? 4 : 0)
17871 .addOperand(Segment)
17872 .addReg(NextOffsetReg)
17873 .setMemRefs(MMOBegin, MMOEnd);
17876 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
17881 // Emit code to use overflow area
17884 // Load the overflow_area address into a register.
17885 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
17886 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
17891 .addOperand(Segment)
17892 .setMemRefs(MMOBegin, MMOEnd);
17894 // If we need to align it, do so. Otherwise, just copy the address
17895 // to OverflowDestReg.
17897 // Align the overflow address
17898 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
17899 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
17901 // aligned_addr = (addr + (align-1)) & ~(align-1)
17902 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
17903 .addReg(OverflowAddrReg)
17906 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
17908 .addImm(~(uint64_t)(Align-1));
17910 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
17911 .addReg(OverflowAddrReg);
17914 // Compute the next overflow address after this argument.
17915 // (the overflow address should be kept 8-byte aligned)
17916 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
17917 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
17918 .addReg(OverflowDestReg)
17919 .addImm(ArgSizeA8);
17921 // Store the new overflow address.
17922 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
17927 .addOperand(Segment)
17928 .addReg(NextAddrReg)
17929 .setMemRefs(MMOBegin, MMOEnd);
17931 // If we branched, emit the PHI to the front of endMBB.
17933 BuildMI(*endMBB, endMBB->begin(), DL,
17934 TII->get(X86::PHI), DestReg)
17935 .addReg(OffsetDestReg).addMBB(offsetMBB)
17936 .addReg(OverflowDestReg).addMBB(overflowMBB);
17939 // Erase the pseudo instruction
17940 MI->eraseFromParent();
17945 MachineBasicBlock *
17946 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
17948 MachineBasicBlock *MBB) const {
17949 // Emit code to save XMM registers to the stack. The ABI says that the
17950 // number of registers to save is given in %al, so it's theoretically
17951 // possible to do an indirect jump trick to avoid saving all of them,
17952 // however this code takes a simpler approach and just executes all
17953 // of the stores if %al is non-zero. It's less code, and it's probably
17954 // easier on the hardware branch predictor, and stores aren't all that
17955 // expensive anyway.
17957 // Create the new basic blocks. One block contains all the XMM stores,
17958 // and one block is the final destination regardless of whether any
17959 // stores were performed.
17960 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17961 MachineFunction *F = MBB->getParent();
17962 MachineFunction::iterator MBBIter = MBB;
17964 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
17965 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
17966 F->insert(MBBIter, XMMSaveMBB);
17967 F->insert(MBBIter, EndMBB);
17969 // Transfer the remainder of MBB and its successor edges to EndMBB.
17970 EndMBB->splice(EndMBB->begin(), MBB,
17971 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17972 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
17974 // The original block will now fall through to the XMM save block.
17975 MBB->addSuccessor(XMMSaveMBB);
17976 // The XMMSaveMBB will fall through to the end block.
17977 XMMSaveMBB->addSuccessor(EndMBB);
17979 // Now add the instructions.
17980 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17981 DebugLoc DL = MI->getDebugLoc();
17983 unsigned CountReg = MI->getOperand(0).getReg();
17984 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
17985 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
17987 if (!Subtarget->isTargetWin64()) {
17988 // If %al is 0, branch around the XMM save block.
17989 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
17990 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
17991 MBB->addSuccessor(EndMBB);
17994 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
17995 // that was just emitted, but clearly shouldn't be "saved".
17996 assert((MI->getNumOperands() <= 3 ||
17997 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
17998 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
17999 && "Expected last argument to be EFLAGS");
18000 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18001 // In the XMM save block, save all the XMM argument registers.
18002 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18003 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18004 MachineMemOperand *MMO =
18005 F->getMachineMemOperand(
18006 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18007 MachineMemOperand::MOStore,
18008 /*Size=*/16, /*Align=*/16);
18009 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18010 .addFrameIndex(RegSaveFrameIndex)
18011 .addImm(/*Scale=*/1)
18012 .addReg(/*IndexReg=*/0)
18013 .addImm(/*Disp=*/Offset)
18014 .addReg(/*Segment=*/0)
18015 .addReg(MI->getOperand(i).getReg())
18016 .addMemOperand(MMO);
18019 MI->eraseFromParent(); // The pseudo instruction is gone now.
18024 // The EFLAGS operand of SelectItr might be missing a kill marker
18025 // because there were multiple uses of EFLAGS, and ISel didn't know
18026 // which to mark. Figure out whether SelectItr should have had a
18027 // kill marker, and set it if it should. Returns the correct kill
18029 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18030 MachineBasicBlock* BB,
18031 const TargetRegisterInfo* TRI) {
18032 // Scan forward through BB for a use/def of EFLAGS.
18033 MachineBasicBlock::iterator miI(std::next(SelectItr));
18034 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18035 const MachineInstr& mi = *miI;
18036 if (mi.readsRegister(X86::EFLAGS))
18038 if (mi.definesRegister(X86::EFLAGS))
18039 break; // Should have kill-flag - update below.
18042 // If we hit the end of the block, check whether EFLAGS is live into a
18044 if (miI == BB->end()) {
18045 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18046 sEnd = BB->succ_end();
18047 sItr != sEnd; ++sItr) {
18048 MachineBasicBlock* succ = *sItr;
18049 if (succ->isLiveIn(X86::EFLAGS))
18054 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18055 // out. SelectMI should have a kill flag on EFLAGS.
18056 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18060 MachineBasicBlock *
18061 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18062 MachineBasicBlock *BB) const {
18063 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18064 DebugLoc DL = MI->getDebugLoc();
18066 // To "insert" a SELECT_CC instruction, we actually have to insert the
18067 // diamond control-flow pattern. The incoming instruction knows the
18068 // destination vreg to set, the condition code register to branch on, the
18069 // true/false values to select between, and a branch opcode to use.
18070 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18071 MachineFunction::iterator It = BB;
18077 // cmpTY ccX, r1, r2
18079 // fallthrough --> copy0MBB
18080 MachineBasicBlock *thisMBB = BB;
18081 MachineFunction *F = BB->getParent();
18082 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18083 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18084 F->insert(It, copy0MBB);
18085 F->insert(It, sinkMBB);
18087 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18088 // live into the sink and copy blocks.
18089 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18090 if (!MI->killsRegister(X86::EFLAGS) &&
18091 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
18092 copy0MBB->addLiveIn(X86::EFLAGS);
18093 sinkMBB->addLiveIn(X86::EFLAGS);
18096 // Transfer the remainder of BB and its successor edges to sinkMBB.
18097 sinkMBB->splice(sinkMBB->begin(), BB,
18098 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18099 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18101 // Add the true and fallthrough blocks as its successors.
18102 BB->addSuccessor(copy0MBB);
18103 BB->addSuccessor(sinkMBB);
18105 // Create the conditional branch instruction.
18107 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18108 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18111 // %FalseValue = ...
18112 // # fallthrough to sinkMBB
18113 copy0MBB->addSuccessor(sinkMBB);
18116 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18118 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18119 TII->get(X86::PHI), MI->getOperand(0).getReg())
18120 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18121 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18123 MI->eraseFromParent(); // The pseudo instruction is gone now.
18127 MachineBasicBlock *
18128 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18129 MachineBasicBlock *BB) const {
18130 MachineFunction *MF = BB->getParent();
18131 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18132 DebugLoc DL = MI->getDebugLoc();
18133 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18135 assert(MF->shouldSplitStack());
18137 const bool Is64Bit = Subtarget->is64Bit();
18138 const bool IsLP64 = Subtarget->isTarget64BitLP64();
18140 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
18141 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
18144 // ... [Till the alloca]
18145 // If stacklet is not large enough, jump to mallocMBB
18148 // Allocate by subtracting from RSP
18149 // Jump to continueMBB
18152 // Allocate by call to runtime
18156 // [rest of original BB]
18159 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18160 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18161 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18163 MachineRegisterInfo &MRI = MF->getRegInfo();
18164 const TargetRegisterClass *AddrRegClass =
18165 getRegClassFor(getPointerTy());
18167 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18168 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18169 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
18170 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
18171 sizeVReg = MI->getOperand(1).getReg(),
18172 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
18174 MachineFunction::iterator MBBIter = BB;
18177 MF->insert(MBBIter, bumpMBB);
18178 MF->insert(MBBIter, mallocMBB);
18179 MF->insert(MBBIter, continueMBB);
18181 continueMBB->splice(continueMBB->begin(), BB,
18182 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18183 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
18185 // Add code to the main basic block to check if the stack limit has been hit,
18186 // and if so, jump to mallocMBB otherwise to bumpMBB.
18187 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
18188 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
18189 .addReg(tmpSPVReg).addReg(sizeVReg);
18190 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
18191 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
18192 .addReg(SPLimitVReg);
18193 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
18195 // bumpMBB simply decreases the stack pointer, since we know the current
18196 // stacklet has enough space.
18197 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
18198 .addReg(SPLimitVReg);
18199 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
18200 .addReg(SPLimitVReg);
18201 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18203 // Calls into a routine in libgcc to allocate more space from the heap.
18204 const uint32_t *RegMask =
18205 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18207 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
18209 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18210 .addExternalSymbol("__morestack_allocate_stack_space")
18211 .addRegMask(RegMask)
18212 .addReg(X86::RDI, RegState::Implicit)
18213 .addReg(X86::RAX, RegState::ImplicitDefine);
18214 } else if (Is64Bit) {
18215 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
18217 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18218 .addExternalSymbol("__morestack_allocate_stack_space")
18219 .addRegMask(RegMask)
18220 .addReg(X86::EDI, RegState::Implicit)
18221 .addReg(X86::EAX, RegState::ImplicitDefine);
18223 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
18225 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
18226 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
18227 .addExternalSymbol("__morestack_allocate_stack_space")
18228 .addRegMask(RegMask)
18229 .addReg(X86::EAX, RegState::ImplicitDefine);
18233 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
18236 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
18237 .addReg(IsLP64 ? X86::RAX : X86::EAX);
18238 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18240 // Set up the CFG correctly.
18241 BB->addSuccessor(bumpMBB);
18242 BB->addSuccessor(mallocMBB);
18243 mallocMBB->addSuccessor(continueMBB);
18244 bumpMBB->addSuccessor(continueMBB);
18246 // Take care of the PHI nodes.
18247 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
18248 MI->getOperand(0).getReg())
18249 .addReg(mallocPtrVReg).addMBB(mallocMBB)
18250 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
18252 // Delete the original pseudo instruction.
18253 MI->eraseFromParent();
18256 return continueMBB;
18259 MachineBasicBlock *
18260 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
18261 MachineBasicBlock *BB) const {
18262 DebugLoc DL = MI->getDebugLoc();
18264 assert(!Subtarget->isTargetMachO());
18266 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
18268 MI->eraseFromParent(); // The pseudo instruction is gone now.
18272 MachineBasicBlock *
18273 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
18274 MachineBasicBlock *BB) const {
18275 // This is pretty easy. We're taking the value that we received from
18276 // our load from the relocation, sticking it in either RDI (x86-64)
18277 // or EAX and doing an indirect call. The return value will then
18278 // be in the normal return register.
18279 MachineFunction *F = BB->getParent();
18280 const X86InstrInfo *TII = Subtarget->getInstrInfo();
18281 DebugLoc DL = MI->getDebugLoc();
18283 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
18284 assert(MI->getOperand(3).isGlobal() && "This should be a global");
18286 // Get a register mask for the lowered call.
18287 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
18288 // proper register mask.
18289 const uint32_t *RegMask =
18290 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18291 if (Subtarget->is64Bit()) {
18292 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18293 TII->get(X86::MOV64rm), X86::RDI)
18295 .addImm(0).addReg(0)
18296 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18297 MI->getOperand(3).getTargetFlags())
18299 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
18300 addDirectMem(MIB, X86::RDI);
18301 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
18302 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
18303 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18304 TII->get(X86::MOV32rm), X86::EAX)
18306 .addImm(0).addReg(0)
18307 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18308 MI->getOperand(3).getTargetFlags())
18310 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18311 addDirectMem(MIB, X86::EAX);
18312 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18314 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18315 TII->get(X86::MOV32rm), X86::EAX)
18316 .addReg(TII->getGlobalBaseReg(F))
18317 .addImm(0).addReg(0)
18318 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18319 MI->getOperand(3).getTargetFlags())
18321 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18322 addDirectMem(MIB, X86::EAX);
18323 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18326 MI->eraseFromParent(); // The pseudo instruction is gone now.
18330 MachineBasicBlock *
18331 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
18332 MachineBasicBlock *MBB) const {
18333 DebugLoc DL = MI->getDebugLoc();
18334 MachineFunction *MF = MBB->getParent();
18335 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18336 MachineRegisterInfo &MRI = MF->getRegInfo();
18338 const BasicBlock *BB = MBB->getBasicBlock();
18339 MachineFunction::iterator I = MBB;
18342 // Memory Reference
18343 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18344 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18347 unsigned MemOpndSlot = 0;
18349 unsigned CurOp = 0;
18351 DstReg = MI->getOperand(CurOp++).getReg();
18352 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
18353 assert(RC->hasType(MVT::i32) && "Invalid destination!");
18354 unsigned mainDstReg = MRI.createVirtualRegister(RC);
18355 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
18357 MemOpndSlot = CurOp;
18359 MVT PVT = getPointerTy();
18360 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18361 "Invalid Pointer Size!");
18363 // For v = setjmp(buf), we generate
18366 // buf[LabelOffset] = restoreMBB
18367 // SjLjSetup restoreMBB
18373 // v = phi(main, restore)
18376 // if base pointer being used, load it from frame
18379 MachineBasicBlock *thisMBB = MBB;
18380 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18381 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18382 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
18383 MF->insert(I, mainMBB);
18384 MF->insert(I, sinkMBB);
18385 MF->push_back(restoreMBB);
18387 MachineInstrBuilder MIB;
18389 // Transfer the remainder of BB and its successor edges to sinkMBB.
18390 sinkMBB->splice(sinkMBB->begin(), MBB,
18391 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18392 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18395 unsigned PtrStoreOpc = 0;
18396 unsigned LabelReg = 0;
18397 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18398 Reloc::Model RM = MF->getTarget().getRelocationModel();
18399 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
18400 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
18402 // Prepare IP either in reg or imm.
18403 if (!UseImmLabel) {
18404 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
18405 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
18406 LabelReg = MRI.createVirtualRegister(PtrRC);
18407 if (Subtarget->is64Bit()) {
18408 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
18412 .addMBB(restoreMBB)
18415 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
18416 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
18417 .addReg(XII->getGlobalBaseReg(MF))
18420 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
18424 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
18426 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
18427 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18428 if (i == X86::AddrDisp)
18429 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
18431 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
18434 MIB.addReg(LabelReg);
18436 MIB.addMBB(restoreMBB);
18437 MIB.setMemRefs(MMOBegin, MMOEnd);
18439 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
18440 .addMBB(restoreMBB);
18442 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18443 MIB.addRegMask(RegInfo->getNoPreservedMask());
18444 thisMBB->addSuccessor(mainMBB);
18445 thisMBB->addSuccessor(restoreMBB);
18449 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
18450 mainMBB->addSuccessor(sinkMBB);
18453 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18454 TII->get(X86::PHI), DstReg)
18455 .addReg(mainDstReg).addMBB(mainMBB)
18456 .addReg(restoreDstReg).addMBB(restoreMBB);
18459 if (RegInfo->hasBasePointer(*MF)) {
18460 const bool Uses64BitFramePtr =
18461 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
18462 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
18463 X86FI->setRestoreBasePointer(MF);
18464 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
18465 unsigned BasePtr = RegInfo->getBaseRegister();
18466 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
18467 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
18468 FramePtr, true, X86FI->getRestoreBasePointerOffset())
18469 .setMIFlag(MachineInstr::FrameSetup);
18471 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
18472 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
18473 restoreMBB->addSuccessor(sinkMBB);
18475 MI->eraseFromParent();
18479 MachineBasicBlock *
18480 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
18481 MachineBasicBlock *MBB) const {
18482 DebugLoc DL = MI->getDebugLoc();
18483 MachineFunction *MF = MBB->getParent();
18484 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18485 MachineRegisterInfo &MRI = MF->getRegInfo();
18487 // Memory Reference
18488 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18489 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18491 MVT PVT = getPointerTy();
18492 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18493 "Invalid Pointer Size!");
18495 const TargetRegisterClass *RC =
18496 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
18497 unsigned Tmp = MRI.createVirtualRegister(RC);
18498 // Since FP is only updated here but NOT referenced, it's treated as GPR.
18499 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18500 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
18501 unsigned SP = RegInfo->getStackRegister();
18503 MachineInstrBuilder MIB;
18505 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18506 const int64_t SPOffset = 2 * PVT.getStoreSize();
18508 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
18509 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
18512 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
18513 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
18514 MIB.addOperand(MI->getOperand(i));
18515 MIB.setMemRefs(MMOBegin, MMOEnd);
18517 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
18518 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18519 if (i == X86::AddrDisp)
18520 MIB.addDisp(MI->getOperand(i), LabelOffset);
18522 MIB.addOperand(MI->getOperand(i));
18524 MIB.setMemRefs(MMOBegin, MMOEnd);
18526 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
18527 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18528 if (i == X86::AddrDisp)
18529 MIB.addDisp(MI->getOperand(i), SPOffset);
18531 MIB.addOperand(MI->getOperand(i));
18533 MIB.setMemRefs(MMOBegin, MMOEnd);
18535 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
18537 MI->eraseFromParent();
18541 // Replace 213-type (isel default) FMA3 instructions with 231-type for
18542 // accumulator loops. Writing back to the accumulator allows the coalescer
18543 // to remove extra copies in the loop.
18544 MachineBasicBlock *
18545 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
18546 MachineBasicBlock *MBB) const {
18547 MachineOperand &AddendOp = MI->getOperand(3);
18549 // Bail out early if the addend isn't a register - we can't switch these.
18550 if (!AddendOp.isReg())
18553 MachineFunction &MF = *MBB->getParent();
18554 MachineRegisterInfo &MRI = MF.getRegInfo();
18556 // Check whether the addend is defined by a PHI:
18557 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
18558 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
18559 if (!AddendDef.isPHI())
18562 // Look for the following pattern:
18564 // %addend = phi [%entry, 0], [%loop, %result]
18566 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
18570 // %addend = phi [%entry, 0], [%loop, %result]
18572 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
18574 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
18575 assert(AddendDef.getOperand(i).isReg());
18576 MachineOperand PHISrcOp = AddendDef.getOperand(i);
18577 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
18578 if (&PHISrcInst == MI) {
18579 // Found a matching instruction.
18580 unsigned NewFMAOpc = 0;
18581 switch (MI->getOpcode()) {
18582 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
18583 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
18584 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
18585 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
18586 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
18587 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
18588 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
18589 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
18590 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
18591 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
18592 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
18593 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
18594 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
18595 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
18596 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
18597 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
18598 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
18599 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
18600 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
18601 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
18603 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
18604 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
18605 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
18606 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
18607 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
18608 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
18609 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
18610 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
18611 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
18612 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
18613 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
18614 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
18615 default: llvm_unreachable("Unrecognized FMA variant.");
18618 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
18619 MachineInstrBuilder MIB =
18620 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
18621 .addOperand(MI->getOperand(0))
18622 .addOperand(MI->getOperand(3))
18623 .addOperand(MI->getOperand(2))
18624 .addOperand(MI->getOperand(1));
18625 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
18626 MI->eraseFromParent();
18633 MachineBasicBlock *
18634 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
18635 MachineBasicBlock *BB) const {
18636 switch (MI->getOpcode()) {
18637 default: llvm_unreachable("Unexpected instr type to insert");
18638 case X86::TAILJMPd64:
18639 case X86::TAILJMPr64:
18640 case X86::TAILJMPm64:
18641 case X86::TAILJMPd64_REX:
18642 case X86::TAILJMPr64_REX:
18643 case X86::TAILJMPm64_REX:
18644 llvm_unreachable("TAILJMP64 would not be touched here.");
18645 case X86::TCRETURNdi64:
18646 case X86::TCRETURNri64:
18647 case X86::TCRETURNmi64:
18649 case X86::WIN_ALLOCA:
18650 return EmitLoweredWinAlloca(MI, BB);
18651 case X86::SEG_ALLOCA_32:
18652 case X86::SEG_ALLOCA_64:
18653 return EmitLoweredSegAlloca(MI, BB);
18654 case X86::TLSCall_32:
18655 case X86::TLSCall_64:
18656 return EmitLoweredTLSCall(MI, BB);
18657 case X86::CMOV_GR8:
18658 case X86::CMOV_FR32:
18659 case X86::CMOV_FR64:
18660 case X86::CMOV_V4F32:
18661 case X86::CMOV_V2F64:
18662 case X86::CMOV_V2I64:
18663 case X86::CMOV_V8F32:
18664 case X86::CMOV_V4F64:
18665 case X86::CMOV_V4I64:
18666 case X86::CMOV_V16F32:
18667 case X86::CMOV_V8F64:
18668 case X86::CMOV_V8I64:
18669 case X86::CMOV_GR16:
18670 case X86::CMOV_GR32:
18671 case X86::CMOV_RFP32:
18672 case X86::CMOV_RFP64:
18673 case X86::CMOV_RFP80:
18674 return EmitLoweredSelect(MI, BB);
18676 case X86::FP32_TO_INT16_IN_MEM:
18677 case X86::FP32_TO_INT32_IN_MEM:
18678 case X86::FP32_TO_INT64_IN_MEM:
18679 case X86::FP64_TO_INT16_IN_MEM:
18680 case X86::FP64_TO_INT32_IN_MEM:
18681 case X86::FP64_TO_INT64_IN_MEM:
18682 case X86::FP80_TO_INT16_IN_MEM:
18683 case X86::FP80_TO_INT32_IN_MEM:
18684 case X86::FP80_TO_INT64_IN_MEM: {
18685 MachineFunction *F = BB->getParent();
18686 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18687 DebugLoc DL = MI->getDebugLoc();
18689 // Change the floating point control register to use "round towards zero"
18690 // mode when truncating to an integer value.
18691 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
18692 addFrameReference(BuildMI(*BB, MI, DL,
18693 TII->get(X86::FNSTCW16m)), CWFrameIdx);
18695 // Load the old value of the high byte of the control word...
18697 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
18698 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
18701 // Set the high part to be round to zero...
18702 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
18705 // Reload the modified control word now...
18706 addFrameReference(BuildMI(*BB, MI, DL,
18707 TII->get(X86::FLDCW16m)), CWFrameIdx);
18709 // Restore the memory image of control word to original value
18710 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
18713 // Get the X86 opcode to use.
18715 switch (MI->getOpcode()) {
18716 default: llvm_unreachable("illegal opcode!");
18717 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
18718 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
18719 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
18720 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
18721 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
18722 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
18723 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
18724 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
18725 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
18729 MachineOperand &Op = MI->getOperand(0);
18731 AM.BaseType = X86AddressMode::RegBase;
18732 AM.Base.Reg = Op.getReg();
18734 AM.BaseType = X86AddressMode::FrameIndexBase;
18735 AM.Base.FrameIndex = Op.getIndex();
18737 Op = MI->getOperand(1);
18739 AM.Scale = Op.getImm();
18740 Op = MI->getOperand(2);
18742 AM.IndexReg = Op.getImm();
18743 Op = MI->getOperand(3);
18744 if (Op.isGlobal()) {
18745 AM.GV = Op.getGlobal();
18747 AM.Disp = Op.getImm();
18749 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
18750 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
18752 // Reload the original control word now.
18753 addFrameReference(BuildMI(*BB, MI, DL,
18754 TII->get(X86::FLDCW16m)), CWFrameIdx);
18756 MI->eraseFromParent(); // The pseudo instruction is gone now.
18759 // String/text processing lowering.
18760 case X86::PCMPISTRM128REG:
18761 case X86::VPCMPISTRM128REG:
18762 case X86::PCMPISTRM128MEM:
18763 case X86::VPCMPISTRM128MEM:
18764 case X86::PCMPESTRM128REG:
18765 case X86::VPCMPESTRM128REG:
18766 case X86::PCMPESTRM128MEM:
18767 case X86::VPCMPESTRM128MEM:
18768 assert(Subtarget->hasSSE42() &&
18769 "Target must have SSE4.2 or AVX features enabled");
18770 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
18772 // String/text processing lowering.
18773 case X86::PCMPISTRIREG:
18774 case X86::VPCMPISTRIREG:
18775 case X86::PCMPISTRIMEM:
18776 case X86::VPCMPISTRIMEM:
18777 case X86::PCMPESTRIREG:
18778 case X86::VPCMPESTRIREG:
18779 case X86::PCMPESTRIMEM:
18780 case X86::VPCMPESTRIMEM:
18781 assert(Subtarget->hasSSE42() &&
18782 "Target must have SSE4.2 or AVX features enabled");
18783 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
18785 // Thread synchronization.
18787 return EmitMonitor(MI, BB, Subtarget);
18791 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
18793 case X86::VASTART_SAVE_XMM_REGS:
18794 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
18796 case X86::VAARG_64:
18797 return EmitVAARG64WithCustomInserter(MI, BB);
18799 case X86::EH_SjLj_SetJmp32:
18800 case X86::EH_SjLj_SetJmp64:
18801 return emitEHSjLjSetJmp(MI, BB);
18803 case X86::EH_SjLj_LongJmp32:
18804 case X86::EH_SjLj_LongJmp64:
18805 return emitEHSjLjLongJmp(MI, BB);
18807 case TargetOpcode::STATEPOINT:
18808 // As an implementation detail, STATEPOINT shares the STACKMAP format at
18809 // this point in the process. We diverge later.
18810 return emitPatchPoint(MI, BB);
18812 case TargetOpcode::STACKMAP:
18813 case TargetOpcode::PATCHPOINT:
18814 return emitPatchPoint(MI, BB);
18816 case X86::VFMADDPDr213r:
18817 case X86::VFMADDPSr213r:
18818 case X86::VFMADDSDr213r:
18819 case X86::VFMADDSSr213r:
18820 case X86::VFMSUBPDr213r:
18821 case X86::VFMSUBPSr213r:
18822 case X86::VFMSUBSDr213r:
18823 case X86::VFMSUBSSr213r:
18824 case X86::VFNMADDPDr213r:
18825 case X86::VFNMADDPSr213r:
18826 case X86::VFNMADDSDr213r:
18827 case X86::VFNMADDSSr213r:
18828 case X86::VFNMSUBPDr213r:
18829 case X86::VFNMSUBPSr213r:
18830 case X86::VFNMSUBSDr213r:
18831 case X86::VFNMSUBSSr213r:
18832 case X86::VFMADDSUBPDr213r:
18833 case X86::VFMADDSUBPSr213r:
18834 case X86::VFMSUBADDPDr213r:
18835 case X86::VFMSUBADDPSr213r:
18836 case X86::VFMADDPDr213rY:
18837 case X86::VFMADDPSr213rY:
18838 case X86::VFMSUBPDr213rY:
18839 case X86::VFMSUBPSr213rY:
18840 case X86::VFNMADDPDr213rY:
18841 case X86::VFNMADDPSr213rY:
18842 case X86::VFNMSUBPDr213rY:
18843 case X86::VFNMSUBPSr213rY:
18844 case X86::VFMADDSUBPDr213rY:
18845 case X86::VFMADDSUBPSr213rY:
18846 case X86::VFMSUBADDPDr213rY:
18847 case X86::VFMSUBADDPSr213rY:
18848 return emitFMA3Instr(MI, BB);
18852 //===----------------------------------------------------------------------===//
18853 // X86 Optimization Hooks
18854 //===----------------------------------------------------------------------===//
18856 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
18859 const SelectionDAG &DAG,
18860 unsigned Depth) const {
18861 unsigned BitWidth = KnownZero.getBitWidth();
18862 unsigned Opc = Op.getOpcode();
18863 assert((Opc >= ISD::BUILTIN_OP_END ||
18864 Opc == ISD::INTRINSIC_WO_CHAIN ||
18865 Opc == ISD::INTRINSIC_W_CHAIN ||
18866 Opc == ISD::INTRINSIC_VOID) &&
18867 "Should use MaskedValueIsZero if you don't know whether Op"
18868 " is a target node!");
18870 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
18884 // These nodes' second result is a boolean.
18885 if (Op.getResNo() == 0)
18888 case X86ISD::SETCC:
18889 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
18891 case ISD::INTRINSIC_WO_CHAIN: {
18892 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18893 unsigned NumLoBits = 0;
18896 case Intrinsic::x86_sse_movmsk_ps:
18897 case Intrinsic::x86_avx_movmsk_ps_256:
18898 case Intrinsic::x86_sse2_movmsk_pd:
18899 case Intrinsic::x86_avx_movmsk_pd_256:
18900 case Intrinsic::x86_mmx_pmovmskb:
18901 case Intrinsic::x86_sse2_pmovmskb_128:
18902 case Intrinsic::x86_avx2_pmovmskb: {
18903 // High bits of movmskp{s|d}, pmovmskb are known zero.
18905 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
18906 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
18907 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
18908 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
18909 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
18910 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
18911 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
18912 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
18914 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
18923 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
18925 const SelectionDAG &,
18926 unsigned Depth) const {
18927 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
18928 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
18929 return Op.getValueType().getScalarType().getSizeInBits();
18935 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
18936 /// node is a GlobalAddress + offset.
18937 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
18938 const GlobalValue* &GA,
18939 int64_t &Offset) const {
18940 if (N->getOpcode() == X86ISD::Wrapper) {
18941 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
18942 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
18943 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
18947 return TargetLowering::isGAPlusOffset(N, GA, Offset);
18950 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
18951 /// same as extracting the high 128-bit part of 256-bit vector and then
18952 /// inserting the result into the low part of a new 256-bit vector
18953 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
18954 EVT VT = SVOp->getValueType(0);
18955 unsigned NumElems = VT.getVectorNumElements();
18957 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
18958 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
18959 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18960 SVOp->getMaskElt(j) >= 0)
18966 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
18967 /// same as extracting the low 128-bit part of 256-bit vector and then
18968 /// inserting the result into the high part of a new 256-bit vector
18969 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
18970 EVT VT = SVOp->getValueType(0);
18971 unsigned NumElems = VT.getVectorNumElements();
18973 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
18974 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
18975 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18976 SVOp->getMaskElt(j) >= 0)
18982 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
18983 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
18984 TargetLowering::DAGCombinerInfo &DCI,
18985 const X86Subtarget* Subtarget) {
18987 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
18988 SDValue V1 = SVOp->getOperand(0);
18989 SDValue V2 = SVOp->getOperand(1);
18990 EVT VT = SVOp->getValueType(0);
18991 unsigned NumElems = VT.getVectorNumElements();
18993 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
18994 V2.getOpcode() == ISD::CONCAT_VECTORS) {
18998 // V UNDEF BUILD_VECTOR UNDEF
19000 // CONCAT_VECTOR CONCAT_VECTOR
19003 // RESULT: V + zero extended
19005 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19006 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19007 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19010 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19013 // To match the shuffle mask, the first half of the mask should
19014 // be exactly the first vector, and all the rest a splat with the
19015 // first element of the second one.
19016 for (unsigned i = 0; i != NumElems/2; ++i)
19017 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19018 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19021 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19022 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19023 if (Ld->hasNUsesOfValue(1, 0)) {
19024 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19025 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19027 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19029 Ld->getPointerInfo(),
19030 Ld->getAlignment(),
19031 false/*isVolatile*/, true/*ReadMem*/,
19032 false/*WriteMem*/);
19034 // Make sure the newly-created LOAD is in the same position as Ld in
19035 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19036 // and update uses of Ld's output chain to use the TokenFactor.
19037 if (Ld->hasAnyUseOfValue(1)) {
19038 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19039 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19040 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19041 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19042 SDValue(ResNode.getNode(), 1));
19045 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19049 // Emit a zeroed vector and insert the desired subvector on its
19051 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19052 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19053 return DCI.CombineTo(N, InsV);
19056 //===--------------------------------------------------------------------===//
19057 // Combine some shuffles into subvector extracts and inserts:
19060 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19061 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19062 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19063 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19064 return DCI.CombineTo(N, InsV);
19067 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19068 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19069 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19070 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19071 return DCI.CombineTo(N, InsV);
19077 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19080 /// This is the leaf of the recursive combinine below. When we have found some
19081 /// chain of single-use x86 shuffle instructions and accumulated the combined
19082 /// shuffle mask represented by them, this will try to pattern match that mask
19083 /// into either a single instruction if there is a special purpose instruction
19084 /// for this operation, or into a PSHUFB instruction which is a fully general
19085 /// instruction but should only be used to replace chains over a certain depth.
19086 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19087 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19088 TargetLowering::DAGCombinerInfo &DCI,
19089 const X86Subtarget *Subtarget) {
19090 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19092 // Find the operand that enters the chain. Note that multiple uses are OK
19093 // here, we're not going to remove the operand we find.
19094 SDValue Input = Op.getOperand(0);
19095 while (Input.getOpcode() == ISD::BITCAST)
19096 Input = Input.getOperand(0);
19098 MVT VT = Input.getSimpleValueType();
19099 MVT RootVT = Root.getSimpleValueType();
19102 // Just remove no-op shuffle masks.
19103 if (Mask.size() == 1) {
19104 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19109 // Use the float domain if the operand type is a floating point type.
19110 bool FloatDomain = VT.isFloatingPoint();
19112 // For floating point shuffles, we don't have free copies in the shuffle
19113 // instructions or the ability to load as part of the instruction, so
19114 // canonicalize their shuffles to UNPCK or MOV variants.
19116 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19117 // vectors because it can have a load folded into it that UNPCK cannot. This
19118 // doesn't preclude something switching to the shorter encoding post-RA.
19120 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
19121 bool Lo = Mask.equals(0, 0);
19124 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19125 // is no slower than UNPCKLPD but has the option to fold the input operand
19126 // into even an unaligned memory load.
19127 if (Lo && Subtarget->hasSSE3()) {
19128 Shuffle = X86ISD::MOVDDUP;
19129 ShuffleVT = MVT::v2f64;
19131 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
19132 // than the UNPCK variants.
19133 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
19134 ShuffleVT = MVT::v4f32;
19136 if (Depth == 1 && Root->getOpcode() == Shuffle)
19137 return false; // Nothing to do!
19138 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19139 DCI.AddToWorklist(Op.getNode());
19140 if (Shuffle == X86ISD::MOVDDUP)
19141 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19143 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19144 DCI.AddToWorklist(Op.getNode());
19145 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19149 if (Subtarget->hasSSE3() &&
19150 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
19151 bool Lo = Mask.equals(0, 0, 2, 2);
19152 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
19153 MVT ShuffleVT = MVT::v4f32;
19154 if (Depth == 1 && Root->getOpcode() == Shuffle)
19155 return false; // Nothing to do!
19156 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19157 DCI.AddToWorklist(Op.getNode());
19158 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19159 DCI.AddToWorklist(Op.getNode());
19160 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19164 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
19165 bool Lo = Mask.equals(0, 0, 1, 1);
19166 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19167 MVT ShuffleVT = MVT::v4f32;
19168 if (Depth == 1 && Root->getOpcode() == Shuffle)
19169 return false; // Nothing to do!
19170 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19171 DCI.AddToWorklist(Op.getNode());
19172 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19173 DCI.AddToWorklist(Op.getNode());
19174 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19180 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
19181 // variants as none of these have single-instruction variants that are
19182 // superior to the UNPCK formulation.
19183 if (!FloatDomain &&
19184 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
19185 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
19186 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
19187 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
19189 bool Lo = Mask[0] == 0;
19190 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19191 if (Depth == 1 && Root->getOpcode() == Shuffle)
19192 return false; // Nothing to do!
19194 switch (Mask.size()) {
19196 ShuffleVT = MVT::v8i16;
19199 ShuffleVT = MVT::v16i8;
19202 llvm_unreachable("Impossible mask size!");
19204 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19205 DCI.AddToWorklist(Op.getNode());
19206 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19207 DCI.AddToWorklist(Op.getNode());
19208 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19213 // Don't try to re-form single instruction chains under any circumstances now
19214 // that we've done encoding canonicalization for them.
19218 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
19219 // can replace them with a single PSHUFB instruction profitably. Intel's
19220 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
19221 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
19222 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
19223 SmallVector<SDValue, 16> PSHUFBMask;
19224 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
19225 int Ratio = 16 / Mask.size();
19226 for (unsigned i = 0; i < 16; ++i) {
19227 if (Mask[i / Ratio] == SM_SentinelUndef) {
19228 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
19231 int M = Mask[i / Ratio] != SM_SentinelZero
19232 ? Ratio * Mask[i / Ratio] + i % Ratio
19234 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
19236 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
19237 DCI.AddToWorklist(Op.getNode());
19238 SDValue PSHUFBMaskOp =
19239 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
19240 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
19241 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
19242 DCI.AddToWorklist(Op.getNode());
19243 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19248 // Failed to find any combines.
19252 /// \brief Fully generic combining of x86 shuffle instructions.
19254 /// This should be the last combine run over the x86 shuffle instructions. Once
19255 /// they have been fully optimized, this will recursively consider all chains
19256 /// of single-use shuffle instructions, build a generic model of the cumulative
19257 /// shuffle operation, and check for simpler instructions which implement this
19258 /// operation. We use this primarily for two purposes:
19260 /// 1) Collapse generic shuffles to specialized single instructions when
19261 /// equivalent. In most cases, this is just an encoding size win, but
19262 /// sometimes we will collapse multiple generic shuffles into a single
19263 /// special-purpose shuffle.
19264 /// 2) Look for sequences of shuffle instructions with 3 or more total
19265 /// instructions, and replace them with the slightly more expensive SSSE3
19266 /// PSHUFB instruction if available. We do this as the last combining step
19267 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
19268 /// a suitable short sequence of other instructions. The PHUFB will either
19269 /// use a register or have to read from memory and so is slightly (but only
19270 /// slightly) more expensive than the other shuffle instructions.
19272 /// Because this is inherently a quadratic operation (for each shuffle in
19273 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
19274 /// This should never be an issue in practice as the shuffle lowering doesn't
19275 /// produce sequences of more than 8 instructions.
19277 /// FIXME: We will currently miss some cases where the redundant shuffling
19278 /// would simplify under the threshold for PSHUFB formation because of
19279 /// combine-ordering. To fix this, we should do the redundant instruction
19280 /// combining in this recursive walk.
19281 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
19282 ArrayRef<int> RootMask,
19283 int Depth, bool HasPSHUFB,
19285 TargetLowering::DAGCombinerInfo &DCI,
19286 const X86Subtarget *Subtarget) {
19287 // Bound the depth of our recursive combine because this is ultimately
19288 // quadratic in nature.
19292 // Directly rip through bitcasts to find the underlying operand.
19293 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
19294 Op = Op.getOperand(0);
19296 MVT VT = Op.getSimpleValueType();
19297 if (!VT.isVector())
19298 return false; // Bail if we hit a non-vector.
19299 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
19300 // version should be added.
19301 if (VT.getSizeInBits() != 128)
19304 assert(Root.getSimpleValueType().isVector() &&
19305 "Shuffles operate on vector types!");
19306 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
19307 "Can only combine shuffles of the same vector register size.");
19309 if (!isTargetShuffle(Op.getOpcode()))
19311 SmallVector<int, 16> OpMask;
19313 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
19314 // We only can combine unary shuffles which we can decode the mask for.
19315 if (!HaveMask || !IsUnary)
19318 assert(VT.getVectorNumElements() == OpMask.size() &&
19319 "Different mask size from vector size!");
19320 assert(((RootMask.size() > OpMask.size() &&
19321 RootMask.size() % OpMask.size() == 0) ||
19322 (OpMask.size() > RootMask.size() &&
19323 OpMask.size() % RootMask.size() == 0) ||
19324 OpMask.size() == RootMask.size()) &&
19325 "The smaller number of elements must divide the larger.");
19326 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
19327 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
19328 assert(((RootRatio == 1 && OpRatio == 1) ||
19329 (RootRatio == 1) != (OpRatio == 1)) &&
19330 "Must not have a ratio for both incoming and op masks!");
19332 SmallVector<int, 16> Mask;
19333 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
19335 // Merge this shuffle operation's mask into our accumulated mask. Note that
19336 // this shuffle's mask will be the first applied to the input, followed by the
19337 // root mask to get us all the way to the root value arrangement. The reason
19338 // for this order is that we are recursing up the operation chain.
19339 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
19340 int RootIdx = i / RootRatio;
19341 if (RootMask[RootIdx] < 0) {
19342 // This is a zero or undef lane, we're done.
19343 Mask.push_back(RootMask[RootIdx]);
19347 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
19348 int OpIdx = RootMaskedIdx / OpRatio;
19349 if (OpMask[OpIdx] < 0) {
19350 // The incoming lanes are zero or undef, it doesn't matter which ones we
19352 Mask.push_back(OpMask[OpIdx]);
19356 // Ok, we have non-zero lanes, map them through.
19357 Mask.push_back(OpMask[OpIdx] * OpRatio +
19358 RootMaskedIdx % OpRatio);
19361 // See if we can recurse into the operand to combine more things.
19362 switch (Op.getOpcode()) {
19363 case X86ISD::PSHUFB:
19365 case X86ISD::PSHUFD:
19366 case X86ISD::PSHUFHW:
19367 case X86ISD::PSHUFLW:
19368 if (Op.getOperand(0).hasOneUse() &&
19369 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19370 HasPSHUFB, DAG, DCI, Subtarget))
19374 case X86ISD::UNPCKL:
19375 case X86ISD::UNPCKH:
19376 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
19377 // We can't check for single use, we have to check that this shuffle is the only user.
19378 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
19379 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19380 HasPSHUFB, DAG, DCI, Subtarget))
19385 // Minor canonicalization of the accumulated shuffle mask to make it easier
19386 // to match below. All this does is detect masks with squential pairs of
19387 // elements, and shrink them to the half-width mask. It does this in a loop
19388 // so it will reduce the size of the mask to the minimal width mask which
19389 // performs an equivalent shuffle.
19390 SmallVector<int, 16> WidenedMask;
19391 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
19392 Mask = std::move(WidenedMask);
19393 WidenedMask.clear();
19396 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
19400 /// \brief Get the PSHUF-style mask from PSHUF node.
19402 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
19403 /// PSHUF-style masks that can be reused with such instructions.
19404 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
19405 SmallVector<int, 4> Mask;
19407 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
19411 switch (N.getOpcode()) {
19412 case X86ISD::PSHUFD:
19414 case X86ISD::PSHUFLW:
19417 case X86ISD::PSHUFHW:
19418 Mask.erase(Mask.begin(), Mask.begin() + 4);
19419 for (int &M : Mask)
19423 llvm_unreachable("No valid shuffle instruction found!");
19427 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
19429 /// We walk up the chain and look for a combinable shuffle, skipping over
19430 /// shuffles that we could hoist this shuffle's transformation past without
19431 /// altering anything.
19433 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
19435 TargetLowering::DAGCombinerInfo &DCI) {
19436 assert(N.getOpcode() == X86ISD::PSHUFD &&
19437 "Called with something other than an x86 128-bit half shuffle!");
19440 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
19441 // of the shuffles in the chain so that we can form a fresh chain to replace
19443 SmallVector<SDValue, 8> Chain;
19444 SDValue V = N.getOperand(0);
19445 for (; V.hasOneUse(); V = V.getOperand(0)) {
19446 switch (V.getOpcode()) {
19448 return SDValue(); // Nothing combined!
19451 // Skip bitcasts as we always know the type for the target specific
19455 case X86ISD::PSHUFD:
19456 // Found another dword shuffle.
19459 case X86ISD::PSHUFLW:
19460 // Check that the low words (being shuffled) are the identity in the
19461 // dword shuffle, and the high words are self-contained.
19462 if (Mask[0] != 0 || Mask[1] != 1 ||
19463 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
19466 Chain.push_back(V);
19469 case X86ISD::PSHUFHW:
19470 // Check that the high words (being shuffled) are the identity in the
19471 // dword shuffle, and the low words are self-contained.
19472 if (Mask[2] != 2 || Mask[3] != 3 ||
19473 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
19476 Chain.push_back(V);
19479 case X86ISD::UNPCKL:
19480 case X86ISD::UNPCKH:
19481 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
19482 // shuffle into a preceding word shuffle.
19483 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
19486 // Search for a half-shuffle which we can combine with.
19487 unsigned CombineOp =
19488 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
19489 if (V.getOperand(0) != V.getOperand(1) ||
19490 !V->isOnlyUserOf(V.getOperand(0).getNode()))
19492 Chain.push_back(V);
19493 V = V.getOperand(0);
19495 switch (V.getOpcode()) {
19497 return SDValue(); // Nothing to combine.
19499 case X86ISD::PSHUFLW:
19500 case X86ISD::PSHUFHW:
19501 if (V.getOpcode() == CombineOp)
19504 Chain.push_back(V);
19508 V = V.getOperand(0);
19512 } while (V.hasOneUse());
19515 // Break out of the loop if we break out of the switch.
19519 if (!V.hasOneUse())
19520 // We fell out of the loop without finding a viable combining instruction.
19523 // Merge this node's mask and our incoming mask.
19524 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19525 for (int &M : Mask)
19527 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
19528 getV4X86ShuffleImm8ForMask(Mask, DAG));
19530 // Rebuild the chain around this new shuffle.
19531 while (!Chain.empty()) {
19532 SDValue W = Chain.pop_back_val();
19534 if (V.getValueType() != W.getOperand(0).getValueType())
19535 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
19537 switch (W.getOpcode()) {
19539 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
19541 case X86ISD::UNPCKL:
19542 case X86ISD::UNPCKH:
19543 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
19546 case X86ISD::PSHUFD:
19547 case X86ISD::PSHUFLW:
19548 case X86ISD::PSHUFHW:
19549 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
19553 if (V.getValueType() != N.getValueType())
19554 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
19556 // Return the new chain to replace N.
19560 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
19562 /// We walk up the chain, skipping shuffles of the other half and looking
19563 /// through shuffles which switch halves trying to find a shuffle of the same
19564 /// pair of dwords.
19565 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
19567 TargetLowering::DAGCombinerInfo &DCI) {
19569 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
19570 "Called with something other than an x86 128-bit half shuffle!");
19572 unsigned CombineOpcode = N.getOpcode();
19574 // Walk up a single-use chain looking for a combinable shuffle.
19575 SDValue V = N.getOperand(0);
19576 for (; V.hasOneUse(); V = V.getOperand(0)) {
19577 switch (V.getOpcode()) {
19579 return false; // Nothing combined!
19582 // Skip bitcasts as we always know the type for the target specific
19586 case X86ISD::PSHUFLW:
19587 case X86ISD::PSHUFHW:
19588 if (V.getOpcode() == CombineOpcode)
19591 // Other-half shuffles are no-ops.
19594 // Break out of the loop if we break out of the switch.
19598 if (!V.hasOneUse())
19599 // We fell out of the loop without finding a viable combining instruction.
19602 // Combine away the bottom node as its shuffle will be accumulated into
19603 // a preceding shuffle.
19604 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19606 // Record the old value.
19609 // Merge this node's mask and our incoming mask (adjusted to account for all
19610 // the pshufd instructions encountered).
19611 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19612 for (int &M : Mask)
19614 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
19615 getV4X86ShuffleImm8ForMask(Mask, DAG));
19617 // Check that the shuffles didn't cancel each other out. If not, we need to
19618 // combine to the new one.
19620 // Replace the combinable shuffle with the combined one, updating all users
19621 // so that we re-evaluate the chain here.
19622 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
19627 /// \brief Try to combine x86 target specific shuffles.
19628 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
19629 TargetLowering::DAGCombinerInfo &DCI,
19630 const X86Subtarget *Subtarget) {
19632 MVT VT = N.getSimpleValueType();
19633 SmallVector<int, 4> Mask;
19635 switch (N.getOpcode()) {
19636 case X86ISD::PSHUFD:
19637 case X86ISD::PSHUFLW:
19638 case X86ISD::PSHUFHW:
19639 Mask = getPSHUFShuffleMask(N);
19640 assert(Mask.size() == 4);
19646 // Nuke no-op shuffles that show up after combining.
19647 if (isNoopShuffleMask(Mask))
19648 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19650 // Look for simplifications involving one or two shuffle instructions.
19651 SDValue V = N.getOperand(0);
19652 switch (N.getOpcode()) {
19655 case X86ISD::PSHUFLW:
19656 case X86ISD::PSHUFHW:
19657 assert(VT == MVT::v8i16);
19660 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
19661 return SDValue(); // We combined away this shuffle, so we're done.
19663 // See if this reduces to a PSHUFD which is no more expensive and can
19664 // combine with more operations. Note that it has to at least flip the
19665 // dwords as otherwise it would have been removed as a no-op.
19666 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
19667 int DMask[] = {0, 1, 2, 3};
19668 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
19669 DMask[DOffset + 0] = DOffset + 1;
19670 DMask[DOffset + 1] = DOffset + 0;
19671 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
19672 DCI.AddToWorklist(V.getNode());
19673 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
19674 getV4X86ShuffleImm8ForMask(DMask, DAG));
19675 DCI.AddToWorklist(V.getNode());
19676 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
19679 // Look for shuffle patterns which can be implemented as a single unpack.
19680 // FIXME: This doesn't handle the location of the PSHUFD generically, and
19681 // only works when we have a PSHUFD followed by two half-shuffles.
19682 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
19683 (V.getOpcode() == X86ISD::PSHUFLW ||
19684 V.getOpcode() == X86ISD::PSHUFHW) &&
19685 V.getOpcode() != N.getOpcode() &&
19687 SDValue D = V.getOperand(0);
19688 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
19689 D = D.getOperand(0);
19690 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
19691 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19692 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
19693 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19694 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19696 for (int i = 0; i < 4; ++i) {
19697 WordMask[i + NOffset] = Mask[i] + NOffset;
19698 WordMask[i + VOffset] = VMask[i] + VOffset;
19700 // Map the word mask through the DWord mask.
19702 for (int i = 0; i < 8; ++i)
19703 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
19704 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
19705 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
19706 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
19707 std::begin(UnpackLoMask)) ||
19708 std::equal(std::begin(MappedMask), std::end(MappedMask),
19709 std::begin(UnpackHiMask))) {
19710 // We can replace all three shuffles with an unpack.
19711 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
19712 DCI.AddToWorklist(V.getNode());
19713 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
19715 DL, MVT::v8i16, V, V);
19722 case X86ISD::PSHUFD:
19723 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
19732 /// \brief Try to combine a shuffle into a target-specific add-sub node.
19734 /// We combine this directly on the abstract vector shuffle nodes so it is
19735 /// easier to generically match. We also insert dummy vector shuffle nodes for
19736 /// the operands which explicitly discard the lanes which are unused by this
19737 /// operation to try to flow through the rest of the combiner the fact that
19738 /// they're unused.
19739 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
19741 EVT VT = N->getValueType(0);
19743 // We only handle target-independent shuffles.
19744 // FIXME: It would be easy and harmless to use the target shuffle mask
19745 // extraction tool to support more.
19746 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
19749 auto *SVN = cast<ShuffleVectorSDNode>(N);
19750 ArrayRef<int> Mask = SVN->getMask();
19751 SDValue V1 = N->getOperand(0);
19752 SDValue V2 = N->getOperand(1);
19754 // We require the first shuffle operand to be the SUB node, and the second to
19755 // be the ADD node.
19756 // FIXME: We should support the commuted patterns.
19757 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
19760 // If there are other uses of these operations we can't fold them.
19761 if (!V1->hasOneUse() || !V2->hasOneUse())
19764 // Ensure that both operations have the same operands. Note that we can
19765 // commute the FADD operands.
19766 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
19767 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
19768 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
19771 // We're looking for blends between FADD and FSUB nodes. We insist on these
19772 // nodes being lined up in a specific expected pattern.
19773 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
19774 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
19775 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
19778 // Only specific types are legal at this point, assert so we notice if and
19779 // when these change.
19780 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
19781 VT == MVT::v4f64) &&
19782 "Unknown vector type encountered!");
19784 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
19787 /// PerformShuffleCombine - Performs several different shuffle combines.
19788 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
19789 TargetLowering::DAGCombinerInfo &DCI,
19790 const X86Subtarget *Subtarget) {
19792 SDValue N0 = N->getOperand(0);
19793 SDValue N1 = N->getOperand(1);
19794 EVT VT = N->getValueType(0);
19796 // Don't create instructions with illegal types after legalize types has run.
19797 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19798 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
19801 // If we have legalized the vector types, look for blends of FADD and FSUB
19802 // nodes that we can fuse into an ADDSUB node.
19803 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
19804 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
19807 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
19808 if (Subtarget->hasFp256() && VT.is256BitVector() &&
19809 N->getOpcode() == ISD::VECTOR_SHUFFLE)
19810 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
19812 // During Type Legalization, when promoting illegal vector types,
19813 // the backend might introduce new shuffle dag nodes and bitcasts.
19815 // This code performs the following transformation:
19816 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
19817 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
19819 // We do this only if both the bitcast and the BINOP dag nodes have
19820 // one use. Also, perform this transformation only if the new binary
19821 // operation is legal. This is to avoid introducing dag nodes that
19822 // potentially need to be further expanded (or custom lowered) into a
19823 // less optimal sequence of dag nodes.
19824 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
19825 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
19826 N0.getOpcode() == ISD::BITCAST) {
19827 SDValue BC0 = N0.getOperand(0);
19828 EVT SVT = BC0.getValueType();
19829 unsigned Opcode = BC0.getOpcode();
19830 unsigned NumElts = VT.getVectorNumElements();
19832 if (BC0.hasOneUse() && SVT.isVector() &&
19833 SVT.getVectorNumElements() * 2 == NumElts &&
19834 TLI.isOperationLegal(Opcode, VT)) {
19835 bool CanFold = false;
19847 unsigned SVTNumElts = SVT.getVectorNumElements();
19848 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19849 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
19850 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
19851 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
19852 CanFold = SVOp->getMaskElt(i) < 0;
19855 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
19856 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
19857 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
19858 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
19863 // Only handle 128 wide vector from here on.
19864 if (!VT.is128BitVector())
19867 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
19868 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
19869 // consecutive, non-overlapping, and in the right order.
19870 SmallVector<SDValue, 16> Elts;
19871 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
19872 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
19874 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
19878 if (isTargetShuffle(N->getOpcode())) {
19880 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
19881 if (Shuffle.getNode())
19884 // Try recursively combining arbitrary sequences of x86 shuffle
19885 // instructions into higher-order shuffles. We do this after combining
19886 // specific PSHUF instruction sequences into their minimal form so that we
19887 // can evaluate how many specialized shuffle instructions are involved in
19888 // a particular chain.
19889 SmallVector<int, 1> NonceMask; // Just a placeholder.
19890 NonceMask.push_back(0);
19891 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
19892 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
19894 return SDValue(); // This routine will use CombineTo to replace N.
19900 /// PerformTruncateCombine - Converts truncate operation to
19901 /// a sequence of vector shuffle operations.
19902 /// It is possible when we truncate 256-bit vector to 128-bit vector
19903 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
19904 TargetLowering::DAGCombinerInfo &DCI,
19905 const X86Subtarget *Subtarget) {
19909 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
19910 /// specific shuffle of a load can be folded into a single element load.
19911 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
19912 /// shuffles have been custom lowered so we need to handle those here.
19913 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
19914 TargetLowering::DAGCombinerInfo &DCI) {
19915 if (DCI.isBeforeLegalizeOps())
19918 SDValue InVec = N->getOperand(0);
19919 SDValue EltNo = N->getOperand(1);
19921 if (!isa<ConstantSDNode>(EltNo))
19924 EVT OriginalVT = InVec.getValueType();
19926 if (InVec.getOpcode() == ISD::BITCAST) {
19927 // Don't duplicate a load with other uses.
19928 if (!InVec.hasOneUse())
19930 EVT BCVT = InVec.getOperand(0).getValueType();
19931 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
19933 InVec = InVec.getOperand(0);
19936 EVT CurrentVT = InVec.getValueType();
19938 if (!isTargetShuffle(InVec.getOpcode()))
19941 // Don't duplicate a load with other uses.
19942 if (!InVec.hasOneUse())
19945 SmallVector<int, 16> ShuffleMask;
19947 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
19948 ShuffleMask, UnaryShuffle))
19951 // Select the input vector, guarding against out of range extract vector.
19952 unsigned NumElems = CurrentVT.getVectorNumElements();
19953 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
19954 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
19955 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
19956 : InVec.getOperand(1);
19958 // If inputs to shuffle are the same for both ops, then allow 2 uses
19959 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
19960 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
19962 if (LdNode.getOpcode() == ISD::BITCAST) {
19963 // Don't duplicate a load with other uses.
19964 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
19967 AllowedUses = 1; // only allow 1 load use if we have a bitcast
19968 LdNode = LdNode.getOperand(0);
19971 if (!ISD::isNormalLoad(LdNode.getNode()))
19974 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
19976 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
19979 EVT EltVT = N->getValueType(0);
19980 // If there's a bitcast before the shuffle, check if the load type and
19981 // alignment is valid.
19982 unsigned Align = LN0->getAlignment();
19983 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19984 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
19985 EltVT.getTypeForEVT(*DAG.getContext()));
19987 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
19990 // All checks match so transform back to vector_shuffle so that DAG combiner
19991 // can finish the job
19994 // Create shuffle node taking into account the case that its a unary shuffle
19995 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
19996 : InVec.getOperand(1);
19997 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
19998 InVec.getOperand(0), Shuffle,
20000 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20001 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20005 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20006 /// special and don't usually play with other vector types, it's better to
20007 /// handle them early to be sure we emit efficient code by avoiding
20008 /// store-load conversions.
20009 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20010 if (N->getValueType(0) != MVT::x86mmx ||
20011 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20012 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20015 SDValue V = N->getOperand(0);
20016 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20017 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20018 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20019 N->getValueType(0), V.getOperand(0));
20024 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20025 /// generation and convert it from being a bunch of shuffles and extracts
20026 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20027 /// storing the value and loading scalars back, while for x64 we should
20028 /// use 64-bit extracts and shifts.
20029 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20030 TargetLowering::DAGCombinerInfo &DCI) {
20031 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20032 if (NewOp.getNode())
20035 SDValue InputVector = N->getOperand(0);
20037 // Detect mmx to i32 conversion through a v2i32 elt extract.
20038 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20039 N->getValueType(0) == MVT::i32 &&
20040 InputVector.getValueType() == MVT::v2i32) {
20042 // The bitcast source is a direct mmx result.
20043 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20044 if (MMXSrc.getValueType() == MVT::x86mmx)
20045 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20046 N->getValueType(0),
20047 InputVector.getNode()->getOperand(0));
20049 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20050 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20051 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20052 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20053 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20054 MMXSrcOp.getValueType() == MVT::v1i64 &&
20055 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20056 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20057 N->getValueType(0),
20058 MMXSrcOp.getOperand(0));
20061 // Only operate on vectors of 4 elements, where the alternative shuffling
20062 // gets to be more expensive.
20063 if (InputVector.getValueType() != MVT::v4i32)
20066 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20067 // single use which is a sign-extend or zero-extend, and all elements are
20069 SmallVector<SDNode *, 4> Uses;
20070 unsigned ExtractedElements = 0;
20071 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20072 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20073 if (UI.getUse().getResNo() != InputVector.getResNo())
20076 SDNode *Extract = *UI;
20077 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20080 if (Extract->getValueType(0) != MVT::i32)
20082 if (!Extract->hasOneUse())
20084 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20085 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20087 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20090 // Record which element was extracted.
20091 ExtractedElements |=
20092 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20094 Uses.push_back(Extract);
20097 // If not all the elements were used, this may not be worthwhile.
20098 if (ExtractedElements != 15)
20101 // Ok, we've now decided to do the transformation.
20102 // If 64-bit shifts are legal, use the extract-shift sequence,
20103 // otherwise bounce the vector off the cache.
20104 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20106 SDLoc dl(InputVector);
20108 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20109 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20110 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20111 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20112 DAG.getConstant(0, VecIdxTy));
20113 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20114 DAG.getConstant(1, VecIdxTy));
20116 SDValue ShAmt = DAG.getConstant(32,
20117 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20118 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20119 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20120 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20121 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20122 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20123 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
20125 // Store the value to a temporary stack slot.
20126 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
20127 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
20128 MachinePointerInfo(), false, false, 0);
20130 EVT ElementType = InputVector.getValueType().getVectorElementType();
20131 unsigned EltSize = ElementType.getSizeInBits() / 8;
20133 // Replace each use (extract) with a load of the appropriate element.
20134 for (unsigned i = 0; i < 4; ++i) {
20135 uint64_t Offset = EltSize * i;
20136 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
20138 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
20139 StackPtr, OffsetVal);
20141 // Load the scalar.
20142 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
20143 ScalarAddr, MachinePointerInfo(),
20144 false, false, false, 0);
20149 // Replace the extracts
20150 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
20151 UE = Uses.end(); UI != UE; ++UI) {
20152 SDNode *Extract = *UI;
20154 SDValue Idx = Extract->getOperand(1);
20155 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
20156 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
20159 // The replacement was made in place; don't return anything.
20163 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
20164 static std::pair<unsigned, bool>
20165 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
20166 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
20167 if (!VT.isVector())
20168 return std::make_pair(0, false);
20170 bool NeedSplit = false;
20171 switch (VT.getSimpleVT().SimpleTy) {
20172 default: return std::make_pair(0, false);
20175 if (!Subtarget->hasVLX())
20176 return std::make_pair(0, false);
20180 if (!Subtarget->hasBWI())
20181 return std::make_pair(0, false);
20185 if (!Subtarget->hasAVX512())
20186 return std::make_pair(0, false);
20191 if (!Subtarget->hasAVX2())
20193 if (!Subtarget->hasAVX())
20194 return std::make_pair(0, false);
20199 if (!Subtarget->hasSSE2())
20200 return std::make_pair(0, false);
20203 // SSE2 has only a small subset of the operations.
20204 bool hasUnsigned = Subtarget->hasSSE41() ||
20205 (Subtarget->hasSSE2() && VT == MVT::v16i8);
20206 bool hasSigned = Subtarget->hasSSE41() ||
20207 (Subtarget->hasSSE2() && VT == MVT::v8i16);
20209 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20212 // Check for x CC y ? x : y.
20213 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20214 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20219 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20222 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20225 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20228 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20230 // Check for x CC y ? y : x -- a min/max with reversed arms.
20231 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20232 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20237 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20240 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20243 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20246 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20250 return std::make_pair(Opc, NeedSplit);
20254 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
20255 const X86Subtarget *Subtarget) {
20257 SDValue Cond = N->getOperand(0);
20258 SDValue LHS = N->getOperand(1);
20259 SDValue RHS = N->getOperand(2);
20261 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
20262 SDValue CondSrc = Cond->getOperand(0);
20263 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
20264 Cond = CondSrc->getOperand(0);
20267 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
20270 // A vselect where all conditions and data are constants can be optimized into
20271 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
20272 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
20273 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
20276 unsigned MaskValue = 0;
20277 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
20280 MVT VT = N->getSimpleValueType(0);
20281 unsigned NumElems = VT.getVectorNumElements();
20282 SmallVector<int, 8> ShuffleMask(NumElems, -1);
20283 for (unsigned i = 0; i < NumElems; ++i) {
20284 // Be sure we emit undef where we can.
20285 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
20286 ShuffleMask[i] = -1;
20288 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
20291 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20292 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
20294 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
20297 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
20299 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
20300 TargetLowering::DAGCombinerInfo &DCI,
20301 const X86Subtarget *Subtarget) {
20303 SDValue Cond = N->getOperand(0);
20304 // Get the LHS/RHS of the select.
20305 SDValue LHS = N->getOperand(1);
20306 SDValue RHS = N->getOperand(2);
20307 EVT VT = LHS.getValueType();
20308 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20310 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
20311 // instructions match the semantics of the common C idiom x<y?x:y but not
20312 // x<=y?x:y, because of how they handle negative zero (which can be
20313 // ignored in unsafe-math mode).
20314 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
20315 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
20316 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
20317 (Subtarget->hasSSE2() ||
20318 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
20319 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20321 unsigned Opcode = 0;
20322 // Check for x CC y ? x : y.
20323 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20324 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20328 // Converting this to a min would handle NaNs incorrectly, and swapping
20329 // the operands would cause it to handle comparisons between positive
20330 // and negative zero incorrectly.
20331 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20332 if (!DAG.getTarget().Options.UnsafeFPMath &&
20333 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20335 std::swap(LHS, RHS);
20337 Opcode = X86ISD::FMIN;
20340 // Converting this to a min would handle comparisons between positive
20341 // and negative zero incorrectly.
20342 if (!DAG.getTarget().Options.UnsafeFPMath &&
20343 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20345 Opcode = X86ISD::FMIN;
20348 // Converting this to a min would handle both negative zeros and NaNs
20349 // incorrectly, but we can swap the operands to fix both.
20350 std::swap(LHS, RHS);
20354 Opcode = X86ISD::FMIN;
20358 // Converting this to a max would handle comparisons between positive
20359 // and negative zero incorrectly.
20360 if (!DAG.getTarget().Options.UnsafeFPMath &&
20361 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20363 Opcode = X86ISD::FMAX;
20366 // Converting this to a max would handle NaNs incorrectly, and swapping
20367 // the operands would cause it to handle comparisons between positive
20368 // and negative zero incorrectly.
20369 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20370 if (!DAG.getTarget().Options.UnsafeFPMath &&
20371 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20373 std::swap(LHS, RHS);
20375 Opcode = X86ISD::FMAX;
20378 // Converting this to a max would handle both negative zeros and NaNs
20379 // incorrectly, but we can swap the operands to fix both.
20380 std::swap(LHS, RHS);
20384 Opcode = X86ISD::FMAX;
20387 // Check for x CC y ? y : x -- a min/max with reversed arms.
20388 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20389 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20393 // Converting this to a min would handle comparisons between positive
20394 // and negative zero incorrectly, and swapping the operands would
20395 // cause it to handle NaNs incorrectly.
20396 if (!DAG.getTarget().Options.UnsafeFPMath &&
20397 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
20398 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20400 std::swap(LHS, RHS);
20402 Opcode = X86ISD::FMIN;
20405 // Converting this to a min would handle NaNs incorrectly.
20406 if (!DAG.getTarget().Options.UnsafeFPMath &&
20407 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
20409 Opcode = X86ISD::FMIN;
20412 // Converting this to a min would handle both negative zeros and NaNs
20413 // incorrectly, but we can swap the operands to fix both.
20414 std::swap(LHS, RHS);
20418 Opcode = X86ISD::FMIN;
20422 // Converting this to a max would handle NaNs incorrectly.
20423 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20425 Opcode = X86ISD::FMAX;
20428 // Converting this to a max would handle comparisons between positive
20429 // and negative zero incorrectly, and swapping the operands would
20430 // cause it to handle NaNs incorrectly.
20431 if (!DAG.getTarget().Options.UnsafeFPMath &&
20432 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
20433 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20435 std::swap(LHS, RHS);
20437 Opcode = X86ISD::FMAX;
20440 // Converting this to a max would handle both negative zeros and NaNs
20441 // incorrectly, but we can swap the operands to fix both.
20442 std::swap(LHS, RHS);
20446 Opcode = X86ISD::FMAX;
20452 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
20455 EVT CondVT = Cond.getValueType();
20456 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
20457 CondVT.getVectorElementType() == MVT::i1) {
20458 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
20459 // lowering on KNL. In this case we convert it to
20460 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
20461 // The same situation for all 128 and 256-bit vectors of i8 and i16.
20462 // Since SKX these selects have a proper lowering.
20463 EVT OpVT = LHS.getValueType();
20464 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
20465 (OpVT.getVectorElementType() == MVT::i8 ||
20466 OpVT.getVectorElementType() == MVT::i16) &&
20467 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
20468 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
20469 DCI.AddToWorklist(Cond.getNode());
20470 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
20473 // If this is a select between two integer constants, try to do some
20475 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
20476 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
20477 // Don't do this for crazy integer types.
20478 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
20479 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
20480 // so that TrueC (the true value) is larger than FalseC.
20481 bool NeedsCondInvert = false;
20483 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
20484 // Efficiently invertible.
20485 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
20486 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
20487 isa<ConstantSDNode>(Cond.getOperand(1))))) {
20488 NeedsCondInvert = true;
20489 std::swap(TrueC, FalseC);
20492 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
20493 if (FalseC->getAPIntValue() == 0 &&
20494 TrueC->getAPIntValue().isPowerOf2()) {
20495 if (NeedsCondInvert) // Invert the condition if needed.
20496 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20497 DAG.getConstant(1, Cond.getValueType()));
20499 // Zero extend the condition if needed.
20500 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
20502 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
20503 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
20504 DAG.getConstant(ShAmt, MVT::i8));
20507 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
20508 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
20509 if (NeedsCondInvert) // Invert the condition if needed.
20510 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20511 DAG.getConstant(1, Cond.getValueType()));
20513 // Zero extend the condition if needed.
20514 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
20515 FalseC->getValueType(0), Cond);
20516 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20517 SDValue(FalseC, 0));
20520 // Optimize cases that will turn into an LEA instruction. This requires
20521 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
20522 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
20523 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
20524 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
20526 bool isFastMultiplier = false;
20528 switch ((unsigned char)Diff) {
20530 case 1: // result = add base, cond
20531 case 2: // result = lea base( , cond*2)
20532 case 3: // result = lea base(cond, cond*2)
20533 case 4: // result = lea base( , cond*4)
20534 case 5: // result = lea base(cond, cond*4)
20535 case 8: // result = lea base( , cond*8)
20536 case 9: // result = lea base(cond, cond*8)
20537 isFastMultiplier = true;
20542 if (isFastMultiplier) {
20543 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
20544 if (NeedsCondInvert) // Invert the condition if needed.
20545 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20546 DAG.getConstant(1, Cond.getValueType()));
20548 // Zero extend the condition if needed.
20549 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
20551 // Scale the condition by the difference.
20553 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
20554 DAG.getConstant(Diff, Cond.getValueType()));
20556 // Add the base if non-zero.
20557 if (FalseC->getAPIntValue() != 0)
20558 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20559 SDValue(FalseC, 0));
20566 // Canonicalize max and min:
20567 // (x > y) ? x : y -> (x >= y) ? x : y
20568 // (x < y) ? x : y -> (x <= y) ? x : y
20569 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
20570 // the need for an extra compare
20571 // against zero. e.g.
20572 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
20574 // testl %edi, %edi
20576 // cmovgl %edi, %eax
20580 // cmovsl %eax, %edi
20581 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
20582 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20583 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20584 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20589 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
20590 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
20591 Cond.getOperand(0), Cond.getOperand(1), NewCC);
20592 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
20597 // Early exit check
20598 if (!TLI.isTypeLegal(VT))
20601 // Match VSELECTs into subs with unsigned saturation.
20602 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
20603 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
20604 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
20605 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
20606 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20608 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
20609 // left side invert the predicate to simplify logic below.
20611 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
20613 CC = ISD::getSetCCInverse(CC, true);
20614 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
20618 if (Other.getNode() && Other->getNumOperands() == 2 &&
20619 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
20620 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
20621 SDValue CondRHS = Cond->getOperand(1);
20623 // Look for a general sub with unsigned saturation first.
20624 // x >= y ? x-y : 0 --> subus x, y
20625 // x > y ? x-y : 0 --> subus x, y
20626 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
20627 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
20628 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
20630 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
20631 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
20632 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
20633 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
20634 // If the RHS is a constant we have to reverse the const
20635 // canonicalization.
20636 // x > C-1 ? x+-C : 0 --> subus x, C
20637 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
20638 CondRHSConst->getAPIntValue() ==
20639 (-OpRHSConst->getAPIntValue() - 1))
20640 return DAG.getNode(
20641 X86ISD::SUBUS, DL, VT, OpLHS,
20642 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
20644 // Another special case: If C was a sign bit, the sub has been
20645 // canonicalized into a xor.
20646 // FIXME: Would it be better to use computeKnownBits to determine
20647 // whether it's safe to decanonicalize the xor?
20648 // x s< 0 ? x^C : 0 --> subus x, C
20649 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
20650 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
20651 OpRHSConst->getAPIntValue().isSignBit())
20652 // Note that we have to rebuild the RHS constant here to ensure we
20653 // don't rely on particular values of undef lanes.
20654 return DAG.getNode(
20655 X86ISD::SUBUS, DL, VT, OpLHS,
20656 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
20661 // Try to match a min/max vector operation.
20662 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
20663 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
20664 unsigned Opc = ret.first;
20665 bool NeedSplit = ret.second;
20667 if (Opc && NeedSplit) {
20668 unsigned NumElems = VT.getVectorNumElements();
20669 // Extract the LHS vectors
20670 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
20671 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
20673 // Extract the RHS vectors
20674 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
20675 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
20677 // Create min/max for each subvector
20678 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
20679 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
20681 // Merge the result
20682 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
20684 return DAG.getNode(Opc, DL, VT, LHS, RHS);
20687 // Simplify vector selection if condition value type matches vselect
20689 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
20690 assert(Cond.getValueType().isVector() &&
20691 "vector select expects a vector selector!");
20693 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
20694 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
20696 // Try invert the condition if true value is not all 1s and false value
20698 if (!TValIsAllOnes && !FValIsAllZeros &&
20699 // Check if the selector will be produced by CMPP*/PCMP*
20700 Cond.getOpcode() == ISD::SETCC &&
20701 // Check if SETCC has already been promoted
20702 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
20703 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
20704 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
20706 if (TValIsAllZeros || FValIsAllOnes) {
20707 SDValue CC = Cond.getOperand(2);
20708 ISD::CondCode NewCC =
20709 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
20710 Cond.getOperand(0).getValueType().isInteger());
20711 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
20712 std::swap(LHS, RHS);
20713 TValIsAllOnes = FValIsAllOnes;
20714 FValIsAllZeros = TValIsAllZeros;
20718 if (TValIsAllOnes || FValIsAllZeros) {
20721 if (TValIsAllOnes && FValIsAllZeros)
20723 else if (TValIsAllOnes)
20724 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
20725 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
20726 else if (FValIsAllZeros)
20727 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
20728 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
20730 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
20734 // If we know that this node is legal then we know that it is going to be
20735 // matched by one of the SSE/AVX BLEND instructions. These instructions only
20736 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
20737 // to simplify previous instructions.
20738 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
20739 !DCI.isBeforeLegalize() &&
20740 // We explicitly check against SSE4.1, v8i16 and v16i16 because, although
20741 // vselect nodes may be marked as Custom, they might only be legal when
20742 // Cond is a build_vector of constants. This will be taken care in
20743 // a later condition.
20744 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) &&
20745 Subtarget->hasSSE41() && VT != MVT::v16i16 && VT != MVT::v8i16) &&
20746 // Don't optimize vector of constants. Those are handled by
20747 // the generic code and all the bits must be properly set for
20748 // the generic optimizer.
20749 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
20750 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
20752 // Don't optimize vector selects that map to mask-registers.
20756 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
20757 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
20759 APInt KnownZero, KnownOne;
20760 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
20761 DCI.isBeforeLegalizeOps());
20762 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
20763 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
20765 // If we changed the computation somewhere in the DAG, this change
20766 // will affect all users of Cond.
20767 // Make sure it is fine and update all the nodes so that we do not
20768 // use the generic VSELECT anymore. Otherwise, we may perform
20769 // wrong optimizations as we messed up with the actual expectation
20770 // for the vector boolean values.
20771 if (Cond != TLO.Old) {
20772 // Check all uses of that condition operand to check whether it will be
20773 // consumed by non-BLEND instructions, which may depend on all bits are
20775 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20777 if (I->getOpcode() != ISD::VSELECT)
20778 // TODO: Add other opcodes eventually lowered into BLEND.
20781 // Update all the users of the condition, before committing the change,
20782 // so that the VSELECT optimizations that expect the correct vector
20783 // boolean value will not be triggered.
20784 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20786 DAG.ReplaceAllUsesOfValueWith(
20788 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
20789 Cond, I->getOperand(1), I->getOperand(2)));
20790 DCI.CommitTargetLoweringOpt(TLO);
20793 // At this point, only Cond is changed. Change the condition
20794 // just for N to keep the opportunity to optimize all other
20795 // users their own way.
20796 DAG.ReplaceAllUsesOfValueWith(
20798 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
20799 TLO.New, N->getOperand(1), N->getOperand(2)));
20804 // We should generate an X86ISD::BLENDI from a vselect if its argument
20805 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
20806 // constants. This specific pattern gets generated when we split a
20807 // selector for a 512 bit vector in a machine without AVX512 (but with
20808 // 256-bit vectors), during legalization:
20810 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
20812 // Iff we find this pattern and the build_vectors are built from
20813 // constants, we translate the vselect into a shuffle_vector that we
20814 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
20815 if ((N->getOpcode() == ISD::VSELECT ||
20816 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
20817 !DCI.isBeforeLegalize()) {
20818 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
20819 if (Shuffle.getNode())
20826 // Check whether a boolean test is testing a boolean value generated by
20827 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
20830 // Simplify the following patterns:
20831 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
20832 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
20833 // to (Op EFLAGS Cond)
20835 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
20836 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
20837 // to (Op EFLAGS !Cond)
20839 // where Op could be BRCOND or CMOV.
20841 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
20842 // Quit if not CMP and SUB with its value result used.
20843 if (Cmp.getOpcode() != X86ISD::CMP &&
20844 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
20847 // Quit if not used as a boolean value.
20848 if (CC != X86::COND_E && CC != X86::COND_NE)
20851 // Check CMP operands. One of them should be 0 or 1 and the other should be
20852 // an SetCC or extended from it.
20853 SDValue Op1 = Cmp.getOperand(0);
20854 SDValue Op2 = Cmp.getOperand(1);
20857 const ConstantSDNode* C = nullptr;
20858 bool needOppositeCond = (CC == X86::COND_E);
20859 bool checkAgainstTrue = false; // Is it a comparison against 1?
20861 if ((C = dyn_cast<ConstantSDNode>(Op1)))
20863 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
20865 else // Quit if all operands are not constants.
20868 if (C->getZExtValue() == 1) {
20869 needOppositeCond = !needOppositeCond;
20870 checkAgainstTrue = true;
20871 } else if (C->getZExtValue() != 0)
20872 // Quit if the constant is neither 0 or 1.
20875 bool truncatedToBoolWithAnd = false;
20876 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
20877 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
20878 SetCC.getOpcode() == ISD::TRUNCATE ||
20879 SetCC.getOpcode() == ISD::AND) {
20880 if (SetCC.getOpcode() == ISD::AND) {
20882 ConstantSDNode *CS;
20883 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
20884 CS->getZExtValue() == 1)
20886 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
20887 CS->getZExtValue() == 1)
20891 SetCC = SetCC.getOperand(OpIdx);
20892 truncatedToBoolWithAnd = true;
20894 SetCC = SetCC.getOperand(0);
20897 switch (SetCC.getOpcode()) {
20898 case X86ISD::SETCC_CARRY:
20899 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
20900 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
20901 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
20902 // truncated to i1 using 'and'.
20903 if (checkAgainstTrue && !truncatedToBoolWithAnd)
20905 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
20906 "Invalid use of SETCC_CARRY!");
20908 case X86ISD::SETCC:
20909 // Set the condition code or opposite one if necessary.
20910 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
20911 if (needOppositeCond)
20912 CC = X86::GetOppositeBranchCondition(CC);
20913 return SetCC.getOperand(1);
20914 case X86ISD::CMOV: {
20915 // Check whether false/true value has canonical one, i.e. 0 or 1.
20916 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
20917 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
20918 // Quit if true value is not a constant.
20921 // Quit if false value is not a constant.
20923 SDValue Op = SetCC.getOperand(0);
20924 // Skip 'zext' or 'trunc' node.
20925 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
20926 Op.getOpcode() == ISD::TRUNCATE)
20927 Op = Op.getOperand(0);
20928 // A special case for rdrand/rdseed, where 0 is set if false cond is
20930 if ((Op.getOpcode() != X86ISD::RDRAND &&
20931 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
20934 // Quit if false value is not the constant 0 or 1.
20935 bool FValIsFalse = true;
20936 if (FVal && FVal->getZExtValue() != 0) {
20937 if (FVal->getZExtValue() != 1)
20939 // If FVal is 1, opposite cond is needed.
20940 needOppositeCond = !needOppositeCond;
20941 FValIsFalse = false;
20943 // Quit if TVal is not the constant opposite of FVal.
20944 if (FValIsFalse && TVal->getZExtValue() != 1)
20946 if (!FValIsFalse && TVal->getZExtValue() != 0)
20948 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
20949 if (needOppositeCond)
20950 CC = X86::GetOppositeBranchCondition(CC);
20951 return SetCC.getOperand(3);
20958 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
20959 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
20960 TargetLowering::DAGCombinerInfo &DCI,
20961 const X86Subtarget *Subtarget) {
20964 // If the flag operand isn't dead, don't touch this CMOV.
20965 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
20968 SDValue FalseOp = N->getOperand(0);
20969 SDValue TrueOp = N->getOperand(1);
20970 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
20971 SDValue Cond = N->getOperand(3);
20973 if (CC == X86::COND_E || CC == X86::COND_NE) {
20974 switch (Cond.getOpcode()) {
20978 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
20979 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
20980 return (CC == X86::COND_E) ? FalseOp : TrueOp;
20986 Flags = checkBoolTestSetCCCombine(Cond, CC);
20987 if (Flags.getNode() &&
20988 // Extra check as FCMOV only supports a subset of X86 cond.
20989 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
20990 SDValue Ops[] = { FalseOp, TrueOp,
20991 DAG.getConstant(CC, MVT::i8), Flags };
20992 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
20995 // If this is a select between two integer constants, try to do some
20996 // optimizations. Note that the operands are ordered the opposite of SELECT
20998 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
20999 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21000 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21001 // larger than FalseC (the false value).
21002 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21003 CC = X86::GetOppositeBranchCondition(CC);
21004 std::swap(TrueC, FalseC);
21005 std::swap(TrueOp, FalseOp);
21008 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21009 // This is efficient for any integer data type (including i8/i16) and
21011 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21012 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21013 DAG.getConstant(CC, MVT::i8), Cond);
21015 // Zero extend the condition if needed.
21016 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21018 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21019 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21020 DAG.getConstant(ShAmt, MVT::i8));
21021 if (N->getNumValues() == 2) // Dead flag value?
21022 return DCI.CombineTo(N, Cond, SDValue());
21026 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21027 // for any integer data type, including i8/i16.
21028 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21029 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21030 DAG.getConstant(CC, MVT::i8), Cond);
21032 // Zero extend the condition if needed.
21033 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21034 FalseC->getValueType(0), Cond);
21035 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21036 SDValue(FalseC, 0));
21038 if (N->getNumValues() == 2) // Dead flag value?
21039 return DCI.CombineTo(N, Cond, SDValue());
21043 // Optimize cases that will turn into an LEA instruction. This requires
21044 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21045 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21046 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21047 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21049 bool isFastMultiplier = false;
21051 switch ((unsigned char)Diff) {
21053 case 1: // result = add base, cond
21054 case 2: // result = lea base( , cond*2)
21055 case 3: // result = lea base(cond, cond*2)
21056 case 4: // result = lea base( , cond*4)
21057 case 5: // result = lea base(cond, cond*4)
21058 case 8: // result = lea base( , cond*8)
21059 case 9: // result = lea base(cond, cond*8)
21060 isFastMultiplier = true;
21065 if (isFastMultiplier) {
21066 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21067 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21068 DAG.getConstant(CC, MVT::i8), Cond);
21069 // Zero extend the condition if needed.
21070 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21072 // Scale the condition by the difference.
21074 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21075 DAG.getConstant(Diff, Cond.getValueType()));
21077 // Add the base if non-zero.
21078 if (FalseC->getAPIntValue() != 0)
21079 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21080 SDValue(FalseC, 0));
21081 if (N->getNumValues() == 2) // Dead flag value?
21082 return DCI.CombineTo(N, Cond, SDValue());
21089 // Handle these cases:
21090 // (select (x != c), e, c) -> select (x != c), e, x),
21091 // (select (x == c), c, e) -> select (x == c), x, e)
21092 // where the c is an integer constant, and the "select" is the combination
21093 // of CMOV and CMP.
21095 // The rationale for this change is that the conditional-move from a constant
21096 // needs two instructions, however, conditional-move from a register needs
21097 // only one instruction.
21099 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
21100 // some instruction-combining opportunities. This opt needs to be
21101 // postponed as late as possible.
21103 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
21104 // the DCI.xxxx conditions are provided to postpone the optimization as
21105 // late as possible.
21107 ConstantSDNode *CmpAgainst = nullptr;
21108 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
21109 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
21110 !isa<ConstantSDNode>(Cond.getOperand(0))) {
21112 if (CC == X86::COND_NE &&
21113 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
21114 CC = X86::GetOppositeBranchCondition(CC);
21115 std::swap(TrueOp, FalseOp);
21118 if (CC == X86::COND_E &&
21119 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
21120 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
21121 DAG.getConstant(CC, MVT::i8), Cond };
21122 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
21130 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
21131 const X86Subtarget *Subtarget) {
21132 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
21134 default: return SDValue();
21135 // SSE/AVX/AVX2 blend intrinsics.
21136 case Intrinsic::x86_avx2_pblendvb:
21137 case Intrinsic::x86_avx2_pblendw:
21138 case Intrinsic::x86_avx2_pblendd_128:
21139 case Intrinsic::x86_avx2_pblendd_256:
21140 // Don't try to simplify this intrinsic if we don't have AVX2.
21141 if (!Subtarget->hasAVX2())
21144 case Intrinsic::x86_avx_blend_pd_256:
21145 case Intrinsic::x86_avx_blend_ps_256:
21146 case Intrinsic::x86_avx_blendv_pd_256:
21147 case Intrinsic::x86_avx_blendv_ps_256:
21148 // Don't try to simplify this intrinsic if we don't have AVX.
21149 if (!Subtarget->hasAVX())
21152 case Intrinsic::x86_sse41_pblendw:
21153 case Intrinsic::x86_sse41_blendpd:
21154 case Intrinsic::x86_sse41_blendps:
21155 case Intrinsic::x86_sse41_blendvps:
21156 case Intrinsic::x86_sse41_blendvpd:
21157 case Intrinsic::x86_sse41_pblendvb: {
21158 SDValue Op0 = N->getOperand(1);
21159 SDValue Op1 = N->getOperand(2);
21160 SDValue Mask = N->getOperand(3);
21162 // Don't try to simplify this intrinsic if we don't have SSE4.1.
21163 if (!Subtarget->hasSSE41())
21166 // fold (blend A, A, Mask) -> A
21169 // fold (blend A, B, allZeros) -> A
21170 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
21172 // fold (blend A, B, allOnes) -> B
21173 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
21176 // Simplify the case where the mask is a constant i32 value.
21177 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
21178 if (C->isNullValue())
21180 if (C->isAllOnesValue())
21187 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
21188 case Intrinsic::x86_sse2_psrai_w:
21189 case Intrinsic::x86_sse2_psrai_d:
21190 case Intrinsic::x86_avx2_psrai_w:
21191 case Intrinsic::x86_avx2_psrai_d:
21192 case Intrinsic::x86_sse2_psra_w:
21193 case Intrinsic::x86_sse2_psra_d:
21194 case Intrinsic::x86_avx2_psra_w:
21195 case Intrinsic::x86_avx2_psra_d: {
21196 SDValue Op0 = N->getOperand(1);
21197 SDValue Op1 = N->getOperand(2);
21198 EVT VT = Op0.getValueType();
21199 assert(VT.isVector() && "Expected a vector type!");
21201 if (isa<BuildVectorSDNode>(Op1))
21202 Op1 = Op1.getOperand(0);
21204 if (!isa<ConstantSDNode>(Op1))
21207 EVT SVT = VT.getVectorElementType();
21208 unsigned SVTBits = SVT.getSizeInBits();
21210 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
21211 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
21212 uint64_t ShAmt = C.getZExtValue();
21214 // Don't try to convert this shift into a ISD::SRA if the shift
21215 // count is bigger than or equal to the element size.
21216 if (ShAmt >= SVTBits)
21219 // Trivial case: if the shift count is zero, then fold this
21220 // into the first operand.
21224 // Replace this packed shift intrinsic with a target independent
21226 SDValue Splat = DAG.getConstant(C, VT);
21227 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
21232 /// PerformMulCombine - Optimize a single multiply with constant into two
21233 /// in order to implement it with two cheaper instructions, e.g.
21234 /// LEA + SHL, LEA + LEA.
21235 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
21236 TargetLowering::DAGCombinerInfo &DCI) {
21237 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
21240 EVT VT = N->getValueType(0);
21241 if (VT != MVT::i64 && VT != MVT::i32)
21244 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
21247 uint64_t MulAmt = C->getZExtValue();
21248 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
21251 uint64_t MulAmt1 = 0;
21252 uint64_t MulAmt2 = 0;
21253 if ((MulAmt % 9) == 0) {
21255 MulAmt2 = MulAmt / 9;
21256 } else if ((MulAmt % 5) == 0) {
21258 MulAmt2 = MulAmt / 5;
21259 } else if ((MulAmt % 3) == 0) {
21261 MulAmt2 = MulAmt / 3;
21264 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
21267 if (isPowerOf2_64(MulAmt2) &&
21268 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
21269 // If second multiplifer is pow2, issue it first. We want the multiply by
21270 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
21272 std::swap(MulAmt1, MulAmt2);
21275 if (isPowerOf2_64(MulAmt1))
21276 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
21277 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
21279 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
21280 DAG.getConstant(MulAmt1, VT));
21282 if (isPowerOf2_64(MulAmt2))
21283 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
21284 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
21286 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
21287 DAG.getConstant(MulAmt2, VT));
21289 // Do not add new nodes to DAG combiner worklist.
21290 DCI.CombineTo(N, NewMul, false);
21295 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
21296 SDValue N0 = N->getOperand(0);
21297 SDValue N1 = N->getOperand(1);
21298 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
21299 EVT VT = N0.getValueType();
21301 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
21302 // since the result of setcc_c is all zero's or all ones.
21303 if (VT.isInteger() && !VT.isVector() &&
21304 N1C && N0.getOpcode() == ISD::AND &&
21305 N0.getOperand(1).getOpcode() == ISD::Constant) {
21306 SDValue N00 = N0.getOperand(0);
21307 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
21308 ((N00.getOpcode() == ISD::ANY_EXTEND ||
21309 N00.getOpcode() == ISD::ZERO_EXTEND) &&
21310 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
21311 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
21312 APInt ShAmt = N1C->getAPIntValue();
21313 Mask = Mask.shl(ShAmt);
21315 return DAG.getNode(ISD::AND, SDLoc(N), VT,
21316 N00, DAG.getConstant(Mask, VT));
21320 // Hardware support for vector shifts is sparse which makes us scalarize the
21321 // vector operations in many cases. Also, on sandybridge ADD is faster than
21323 // (shl V, 1) -> add V,V
21324 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
21325 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
21326 assert(N0.getValueType().isVector() && "Invalid vector shift type");
21327 // We shift all of the values by one. In many cases we do not have
21328 // hardware support for this operation. This is better expressed as an ADD
21330 if (N1SplatC->getZExtValue() == 1)
21331 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
21337 /// \brief Returns a vector of 0s if the node in input is a vector logical
21338 /// shift by a constant amount which is known to be bigger than or equal
21339 /// to the vector element size in bits.
21340 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
21341 const X86Subtarget *Subtarget) {
21342 EVT VT = N->getValueType(0);
21344 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
21345 (!Subtarget->hasInt256() ||
21346 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
21349 SDValue Amt = N->getOperand(1);
21351 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
21352 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
21353 APInt ShiftAmt = AmtSplat->getAPIntValue();
21354 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
21356 // SSE2/AVX2 logical shifts always return a vector of 0s
21357 // if the shift amount is bigger than or equal to
21358 // the element size. The constant shift amount will be
21359 // encoded as a 8-bit immediate.
21360 if (ShiftAmt.trunc(8).uge(MaxAmount))
21361 return getZeroVector(VT, Subtarget, DAG, DL);
21367 /// PerformShiftCombine - Combine shifts.
21368 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
21369 TargetLowering::DAGCombinerInfo &DCI,
21370 const X86Subtarget *Subtarget) {
21371 if (N->getOpcode() == ISD::SHL) {
21372 SDValue V = PerformSHLCombine(N, DAG);
21373 if (V.getNode()) return V;
21376 if (N->getOpcode() != ISD::SRA) {
21377 // Try to fold this logical shift into a zero vector.
21378 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
21379 if (V.getNode()) return V;
21385 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
21386 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
21387 // and friends. Likewise for OR -> CMPNEQSS.
21388 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
21389 TargetLowering::DAGCombinerInfo &DCI,
21390 const X86Subtarget *Subtarget) {
21393 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
21394 // we're requiring SSE2 for both.
21395 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
21396 SDValue N0 = N->getOperand(0);
21397 SDValue N1 = N->getOperand(1);
21398 SDValue CMP0 = N0->getOperand(1);
21399 SDValue CMP1 = N1->getOperand(1);
21402 // The SETCCs should both refer to the same CMP.
21403 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
21406 SDValue CMP00 = CMP0->getOperand(0);
21407 SDValue CMP01 = CMP0->getOperand(1);
21408 EVT VT = CMP00.getValueType();
21410 if (VT == MVT::f32 || VT == MVT::f64) {
21411 bool ExpectingFlags = false;
21412 // Check for any users that want flags:
21413 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
21414 !ExpectingFlags && UI != UE; ++UI)
21415 switch (UI->getOpcode()) {
21420 ExpectingFlags = true;
21422 case ISD::CopyToReg:
21423 case ISD::SIGN_EXTEND:
21424 case ISD::ZERO_EXTEND:
21425 case ISD::ANY_EXTEND:
21429 if (!ExpectingFlags) {
21430 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
21431 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
21433 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
21434 X86::CondCode tmp = cc0;
21439 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
21440 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
21441 // FIXME: need symbolic constants for these magic numbers.
21442 // See X86ATTInstPrinter.cpp:printSSECC().
21443 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
21444 if (Subtarget->hasAVX512()) {
21445 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
21446 CMP01, DAG.getConstant(x86cc, MVT::i8));
21447 if (N->getValueType(0) != MVT::i1)
21448 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
21452 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
21453 CMP00.getValueType(), CMP00, CMP01,
21454 DAG.getConstant(x86cc, MVT::i8));
21456 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
21457 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
21459 if (is64BitFP && !Subtarget->is64Bit()) {
21460 // On a 32-bit target, we cannot bitcast the 64-bit float to a
21461 // 64-bit integer, since that's not a legal type. Since
21462 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
21463 // bits, but can do this little dance to extract the lowest 32 bits
21464 // and work with those going forward.
21465 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
21467 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
21469 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
21470 Vector32, DAG.getIntPtrConstant(0));
21474 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
21475 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
21476 DAG.getConstant(1, IntVT));
21477 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
21478 return OneBitOfTruth;
21486 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
21487 /// so it can be folded inside ANDNP.
21488 static bool CanFoldXORWithAllOnes(const SDNode *N) {
21489 EVT VT = N->getValueType(0);
21491 // Match direct AllOnes for 128 and 256-bit vectors
21492 if (ISD::isBuildVectorAllOnes(N))
21495 // Look through a bit convert.
21496 if (N->getOpcode() == ISD::BITCAST)
21497 N = N->getOperand(0).getNode();
21499 // Sometimes the operand may come from a insert_subvector building a 256-bit
21501 if (VT.is256BitVector() &&
21502 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
21503 SDValue V1 = N->getOperand(0);
21504 SDValue V2 = N->getOperand(1);
21506 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
21507 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
21508 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
21509 ISD::isBuildVectorAllOnes(V2.getNode()))
21516 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
21517 // register. In most cases we actually compare or select YMM-sized registers
21518 // and mixing the two types creates horrible code. This method optimizes
21519 // some of the transition sequences.
21520 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
21521 TargetLowering::DAGCombinerInfo &DCI,
21522 const X86Subtarget *Subtarget) {
21523 EVT VT = N->getValueType(0);
21524 if (!VT.is256BitVector())
21527 assert((N->getOpcode() == ISD::ANY_EXTEND ||
21528 N->getOpcode() == ISD::ZERO_EXTEND ||
21529 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
21531 SDValue Narrow = N->getOperand(0);
21532 EVT NarrowVT = Narrow->getValueType(0);
21533 if (!NarrowVT.is128BitVector())
21536 if (Narrow->getOpcode() != ISD::XOR &&
21537 Narrow->getOpcode() != ISD::AND &&
21538 Narrow->getOpcode() != ISD::OR)
21541 SDValue N0 = Narrow->getOperand(0);
21542 SDValue N1 = Narrow->getOperand(1);
21545 // The Left side has to be a trunc.
21546 if (N0.getOpcode() != ISD::TRUNCATE)
21549 // The type of the truncated inputs.
21550 EVT WideVT = N0->getOperand(0)->getValueType(0);
21554 // The right side has to be a 'trunc' or a constant vector.
21555 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
21556 ConstantSDNode *RHSConstSplat = nullptr;
21557 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
21558 RHSConstSplat = RHSBV->getConstantSplatNode();
21559 if (!RHSTrunc && !RHSConstSplat)
21562 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21564 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
21567 // Set N0 and N1 to hold the inputs to the new wide operation.
21568 N0 = N0->getOperand(0);
21569 if (RHSConstSplat) {
21570 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
21571 SDValue(RHSConstSplat, 0));
21572 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
21573 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
21574 } else if (RHSTrunc) {
21575 N1 = N1->getOperand(0);
21578 // Generate the wide operation.
21579 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
21580 unsigned Opcode = N->getOpcode();
21582 case ISD::ANY_EXTEND:
21584 case ISD::ZERO_EXTEND: {
21585 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
21586 APInt Mask = APInt::getAllOnesValue(InBits);
21587 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
21588 return DAG.getNode(ISD::AND, DL, VT,
21589 Op, DAG.getConstant(Mask, VT));
21591 case ISD::SIGN_EXTEND:
21592 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
21593 Op, DAG.getValueType(NarrowVT));
21595 llvm_unreachable("Unexpected opcode");
21599 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
21600 TargetLowering::DAGCombinerInfo &DCI,
21601 const X86Subtarget *Subtarget) {
21602 SDValue N0 = N->getOperand(0);
21603 SDValue N1 = N->getOperand(1);
21606 // A vector zext_in_reg may be represented as a shuffle,
21607 // feeding into a bitcast (this represents anyext) feeding into
21608 // an and with a mask.
21609 // We'd like to try to combine that into a shuffle with zero
21610 // plus a bitcast, removing the and.
21611 if (N0.getOpcode() != ISD::BITCAST ||
21612 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
21615 // The other side of the AND should be a splat of 2^C, where C
21616 // is the number of bits in the source type.
21617 if (N1.getOpcode() == ISD::BITCAST)
21618 N1 = N1.getOperand(0);
21619 if (N1.getOpcode() != ISD::BUILD_VECTOR)
21621 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
21623 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
21624 EVT SrcType = Shuffle->getValueType(0);
21626 // We expect a single-source shuffle
21627 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
21630 unsigned SrcSize = SrcType.getScalarSizeInBits();
21632 APInt SplatValue, SplatUndef;
21633 unsigned SplatBitSize;
21635 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
21636 SplatBitSize, HasAnyUndefs))
21639 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
21640 // Make sure the splat matches the mask we expect
21641 if (SplatBitSize > ResSize ||
21642 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
21645 // Make sure the input and output size make sense
21646 if (SrcSize >= ResSize || ResSize % SrcSize)
21649 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
21650 // The number of u's between each two values depends on the ratio between
21651 // the source and dest type.
21652 unsigned ZextRatio = ResSize / SrcSize;
21653 bool IsZext = true;
21654 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
21655 if (i % ZextRatio) {
21656 if (Shuffle->getMaskElt(i) > 0) {
21662 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
21663 // Expected element number
21673 // Ok, perform the transformation - replace the shuffle with
21674 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
21675 // (instead of undef) where the k elements come from the zero vector.
21676 SmallVector<int, 8> Mask;
21677 unsigned NumElems = SrcType.getVectorNumElements();
21678 for (unsigned i = 0; i < NumElems; ++i)
21680 Mask.push_back(NumElems);
21682 Mask.push_back(i / ZextRatio);
21684 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
21685 Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
21686 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
21689 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
21690 TargetLowering::DAGCombinerInfo &DCI,
21691 const X86Subtarget *Subtarget) {
21692 if (DCI.isBeforeLegalizeOps())
21695 SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget);
21696 if (Zext.getNode())
21699 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21703 EVT VT = N->getValueType(0);
21704 SDValue N0 = N->getOperand(0);
21705 SDValue N1 = N->getOperand(1);
21708 // Create BEXTR instructions
21709 // BEXTR is ((X >> imm) & (2**size-1))
21710 if (VT == MVT::i32 || VT == MVT::i64) {
21711 // Check for BEXTR.
21712 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
21713 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
21714 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
21715 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
21716 if (MaskNode && ShiftNode) {
21717 uint64_t Mask = MaskNode->getZExtValue();
21718 uint64_t Shift = ShiftNode->getZExtValue();
21719 if (isMask_64(Mask)) {
21720 uint64_t MaskSize = countPopulation(Mask);
21721 if (Shift + MaskSize <= VT.getSizeInBits())
21722 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
21723 DAG.getConstant(Shift | (MaskSize << 8), VT));
21731 // Want to form ANDNP nodes:
21732 // 1) In the hopes of then easily combining them with OR and AND nodes
21733 // to form PBLEND/PSIGN.
21734 // 2) To match ANDN packed intrinsics
21735 if (VT != MVT::v2i64 && VT != MVT::v4i64)
21738 // Check LHS for vnot
21739 if (N0.getOpcode() == ISD::XOR &&
21740 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
21741 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
21742 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
21744 // Check RHS for vnot
21745 if (N1.getOpcode() == ISD::XOR &&
21746 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
21747 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
21748 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
21753 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
21754 TargetLowering::DAGCombinerInfo &DCI,
21755 const X86Subtarget *Subtarget) {
21756 if (DCI.isBeforeLegalizeOps())
21759 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21763 SDValue N0 = N->getOperand(0);
21764 SDValue N1 = N->getOperand(1);
21765 EVT VT = N->getValueType(0);
21767 // look for psign/blend
21768 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
21769 if (!Subtarget->hasSSSE3() ||
21770 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
21773 // Canonicalize pandn to RHS
21774 if (N0.getOpcode() == X86ISD::ANDNP)
21776 // or (and (m, y), (pandn m, x))
21777 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
21778 SDValue Mask = N1.getOperand(0);
21779 SDValue X = N1.getOperand(1);
21781 if (N0.getOperand(0) == Mask)
21782 Y = N0.getOperand(1);
21783 if (N0.getOperand(1) == Mask)
21784 Y = N0.getOperand(0);
21786 // Check to see if the mask appeared in both the AND and ANDNP and
21790 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
21791 // Look through mask bitcast.
21792 if (Mask.getOpcode() == ISD::BITCAST)
21793 Mask = Mask.getOperand(0);
21794 if (X.getOpcode() == ISD::BITCAST)
21795 X = X.getOperand(0);
21796 if (Y.getOpcode() == ISD::BITCAST)
21797 Y = Y.getOperand(0);
21799 EVT MaskVT = Mask.getValueType();
21801 // Validate that the Mask operand is a vector sra node.
21802 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
21803 // there is no psrai.b
21804 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
21805 unsigned SraAmt = ~0;
21806 if (Mask.getOpcode() == ISD::SRA) {
21807 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
21808 if (auto *AmtConst = AmtBV->getConstantSplatNode())
21809 SraAmt = AmtConst->getZExtValue();
21810 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
21811 SDValue SraC = Mask.getOperand(1);
21812 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
21814 if ((SraAmt + 1) != EltBits)
21819 // Now we know we at least have a plendvb with the mask val. See if
21820 // we can form a psignb/w/d.
21821 // psign = x.type == y.type == mask.type && y = sub(0, x);
21822 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
21823 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
21824 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
21825 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
21826 "Unsupported VT for PSIGN");
21827 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
21828 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21830 // PBLENDVB only available on SSE 4.1
21831 if (!Subtarget->hasSSE41())
21834 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
21836 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
21837 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
21838 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
21839 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
21840 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21844 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
21847 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
21848 MachineFunction &MF = DAG.getMachineFunction();
21850 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
21852 // SHLD/SHRD instructions have lower register pressure, but on some
21853 // platforms they have higher latency than the equivalent
21854 // series of shifts/or that would otherwise be generated.
21855 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
21856 // have higher latencies and we are not optimizing for size.
21857 if (!OptForSize && Subtarget->isSHLDSlow())
21860 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
21862 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
21864 if (!N0.hasOneUse() || !N1.hasOneUse())
21867 SDValue ShAmt0 = N0.getOperand(1);
21868 if (ShAmt0.getValueType() != MVT::i8)
21870 SDValue ShAmt1 = N1.getOperand(1);
21871 if (ShAmt1.getValueType() != MVT::i8)
21873 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
21874 ShAmt0 = ShAmt0.getOperand(0);
21875 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
21876 ShAmt1 = ShAmt1.getOperand(0);
21879 unsigned Opc = X86ISD::SHLD;
21880 SDValue Op0 = N0.getOperand(0);
21881 SDValue Op1 = N1.getOperand(0);
21882 if (ShAmt0.getOpcode() == ISD::SUB) {
21883 Opc = X86ISD::SHRD;
21884 std::swap(Op0, Op1);
21885 std::swap(ShAmt0, ShAmt1);
21888 unsigned Bits = VT.getSizeInBits();
21889 if (ShAmt1.getOpcode() == ISD::SUB) {
21890 SDValue Sum = ShAmt1.getOperand(0);
21891 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
21892 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
21893 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
21894 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
21895 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
21896 return DAG.getNode(Opc, DL, VT,
21898 DAG.getNode(ISD::TRUNCATE, DL,
21901 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
21902 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
21904 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
21905 return DAG.getNode(Opc, DL, VT,
21906 N0.getOperand(0), N1.getOperand(0),
21907 DAG.getNode(ISD::TRUNCATE, DL,
21914 // Generate NEG and CMOV for integer abs.
21915 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
21916 EVT VT = N->getValueType(0);
21918 // Since X86 does not have CMOV for 8-bit integer, we don't convert
21919 // 8-bit integer abs to NEG and CMOV.
21920 if (VT.isInteger() && VT.getSizeInBits() == 8)
21923 SDValue N0 = N->getOperand(0);
21924 SDValue N1 = N->getOperand(1);
21927 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
21928 // and change it to SUB and CMOV.
21929 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
21930 N0.getOpcode() == ISD::ADD &&
21931 N0.getOperand(1) == N1 &&
21932 N1.getOpcode() == ISD::SRA &&
21933 N1.getOperand(0) == N0.getOperand(0))
21934 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
21935 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
21936 // Generate SUB & CMOV.
21937 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
21938 DAG.getConstant(0, VT), N0.getOperand(0));
21940 SDValue Ops[] = { N0.getOperand(0), Neg,
21941 DAG.getConstant(X86::COND_GE, MVT::i8),
21942 SDValue(Neg.getNode(), 1) };
21943 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
21948 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
21949 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
21950 TargetLowering::DAGCombinerInfo &DCI,
21951 const X86Subtarget *Subtarget) {
21952 if (DCI.isBeforeLegalizeOps())
21955 if (Subtarget->hasCMov()) {
21956 SDValue RV = performIntegerAbsCombine(N, DAG);
21964 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
21965 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
21966 TargetLowering::DAGCombinerInfo &DCI,
21967 const X86Subtarget *Subtarget) {
21968 LoadSDNode *Ld = cast<LoadSDNode>(N);
21969 EVT RegVT = Ld->getValueType(0);
21970 EVT MemVT = Ld->getMemoryVT();
21972 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21974 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
21975 // into two 16-byte operations.
21976 ISD::LoadExtType Ext = Ld->getExtensionType();
21977 unsigned Alignment = Ld->getAlignment();
21978 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
21979 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
21980 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
21981 unsigned NumElems = RegVT.getVectorNumElements();
21985 SDValue Ptr = Ld->getBasePtr();
21986 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
21988 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
21990 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
21991 Ld->getPointerInfo(), Ld->isVolatile(),
21992 Ld->isNonTemporal(), Ld->isInvariant(),
21994 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
21995 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
21996 Ld->getPointerInfo(), Ld->isVolatile(),
21997 Ld->isNonTemporal(), Ld->isInvariant(),
21998 std::min(16U, Alignment));
21999 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22001 Load2.getValue(1));
22003 SDValue NewVec = DAG.getUNDEF(RegVT);
22004 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22005 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22006 return DCI.CombineTo(N, NewVec, TF, true);
22012 /// PerformMLOADCombine - Resolve extending loads
22013 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22014 TargetLowering::DAGCombinerInfo &DCI,
22015 const X86Subtarget *Subtarget) {
22016 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22017 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22020 EVT VT = Mld->getValueType(0);
22021 unsigned NumElems = VT.getVectorNumElements();
22022 EVT LdVT = Mld->getMemoryVT();
22025 assert(LdVT != VT && "Cannot extend to the same type");
22026 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
22027 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
22028 // From, To sizes and ElemCount must be pow of two
22029 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22030 "Unexpected size for extending masked load");
22032 unsigned SizeRatio = ToSz / FromSz;
22033 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
22035 // Create a type on which we perform the shuffle
22036 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22037 LdVT.getScalarType(), NumElems*SizeRatio);
22038 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22040 // Convert Src0 value
22041 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
22042 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
22043 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22044 for (unsigned i = 0; i != NumElems; ++i)
22045 ShuffleVec[i] = i * SizeRatio;
22047 // Can't shuffle using an illegal type.
22048 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22049 && "WideVecVT should be legal");
22050 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
22051 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
22053 // Prepare the new mask
22055 SDValue Mask = Mld->getMask();
22056 if (Mask.getValueType() == VT) {
22057 // Mask and original value have the same type
22058 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22059 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22060 for (unsigned i = 0; i != NumElems; ++i)
22061 ShuffleVec[i] = i * SizeRatio;
22062 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22063 ShuffleVec[i] = NumElems*SizeRatio;
22064 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22065 DAG.getConstant(0, WideVecVT),
22069 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22070 unsigned WidenNumElts = NumElems*SizeRatio;
22071 unsigned MaskNumElts = VT.getVectorNumElements();
22072 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22075 unsigned NumConcat = WidenNumElts / MaskNumElts;
22076 SmallVector<SDValue, 16> Ops(NumConcat);
22077 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22079 for (unsigned i = 1; i != NumConcat; ++i)
22082 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22085 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
22086 Mld->getBasePtr(), NewMask, WideSrc0,
22087 Mld->getMemoryVT(), Mld->getMemOperand(),
22089 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
22090 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
22093 /// PerformMSTORECombine - Resolve truncating stores
22094 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
22095 const X86Subtarget *Subtarget) {
22096 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
22097 if (!Mst->isTruncatingStore())
22100 EVT VT = Mst->getValue().getValueType();
22101 unsigned NumElems = VT.getVectorNumElements();
22102 EVT StVT = Mst->getMemoryVT();
22105 assert(StVT != VT && "Cannot truncate to the same type");
22106 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22107 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22109 // From, To sizes and ElemCount must be pow of two
22110 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22111 "Unexpected size for truncating masked store");
22112 // We are going to use the original vector elt for storing.
22113 // Accumulated smaller vector elements must be a multiple of the store size.
22114 assert (((NumElems * FromSz) % ToSz) == 0 &&
22115 "Unexpected ratio for truncating masked store");
22117 unsigned SizeRatio = FromSz / ToSz;
22118 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22120 // Create a type on which we perform the shuffle
22121 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22122 StVT.getScalarType(), NumElems*SizeRatio);
22124 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22126 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
22127 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22128 for (unsigned i = 0; i != NumElems; ++i)
22129 ShuffleVec[i] = i * SizeRatio;
22131 // Can't shuffle using an illegal type.
22132 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22133 && "WideVecVT should be legal");
22135 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22136 DAG.getUNDEF(WideVecVT),
22140 SDValue Mask = Mst->getMask();
22141 if (Mask.getValueType() == VT) {
22142 // Mask and original value have the same type
22143 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22144 for (unsigned i = 0; i != NumElems; ++i)
22145 ShuffleVec[i] = i * SizeRatio;
22146 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22147 ShuffleVec[i] = NumElems*SizeRatio;
22148 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22149 DAG.getConstant(0, WideVecVT),
22153 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22154 unsigned WidenNumElts = NumElems*SizeRatio;
22155 unsigned MaskNumElts = VT.getVectorNumElements();
22156 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22159 unsigned NumConcat = WidenNumElts / MaskNumElts;
22160 SmallVector<SDValue, 16> Ops(NumConcat);
22161 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22163 for (unsigned i = 1; i != NumConcat; ++i)
22166 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22169 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
22170 NewMask, StVT, Mst->getMemOperand(), false);
22172 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
22173 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
22174 const X86Subtarget *Subtarget) {
22175 StoreSDNode *St = cast<StoreSDNode>(N);
22176 EVT VT = St->getValue().getValueType();
22177 EVT StVT = St->getMemoryVT();
22179 SDValue StoredVal = St->getOperand(1);
22180 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22182 // If we are saving a concatenation of two XMM registers and 32-byte stores
22183 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
22184 unsigned Alignment = St->getAlignment();
22185 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
22186 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22187 StVT == VT && !IsAligned) {
22188 unsigned NumElems = VT.getVectorNumElements();
22192 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
22193 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
22195 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
22196 SDValue Ptr0 = St->getBasePtr();
22197 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
22199 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
22200 St->getPointerInfo(), St->isVolatile(),
22201 St->isNonTemporal(), Alignment);
22202 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
22203 St->getPointerInfo(), St->isVolatile(),
22204 St->isNonTemporal(),
22205 std::min(16U, Alignment));
22206 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
22209 // Optimize trunc store (of multiple scalars) to shuffle and store.
22210 // First, pack all of the elements in one place. Next, store to memory
22211 // in fewer chunks.
22212 if (St->isTruncatingStore() && VT.isVector()) {
22213 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22214 unsigned NumElems = VT.getVectorNumElements();
22215 assert(StVT != VT && "Cannot truncate to the same type");
22216 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22217 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22219 // From, To sizes and ElemCount must be pow of two
22220 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
22221 // We are going to use the original vector elt for storing.
22222 // Accumulated smaller vector elements must be a multiple of the store size.
22223 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
22225 unsigned SizeRatio = FromSz / ToSz;
22227 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22229 // Create a type on which we perform the shuffle
22230 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22231 StVT.getScalarType(), NumElems*SizeRatio);
22233 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22235 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
22236 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
22237 for (unsigned i = 0; i != NumElems; ++i)
22238 ShuffleVec[i] = i * SizeRatio;
22240 // Can't shuffle using an illegal type.
22241 if (!TLI.isTypeLegal(WideVecVT))
22244 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22245 DAG.getUNDEF(WideVecVT),
22247 // At this point all of the data is stored at the bottom of the
22248 // register. We now need to save it to mem.
22250 // Find the largest store unit
22251 MVT StoreType = MVT::i8;
22252 for (MVT Tp : MVT::integer_valuetypes()) {
22253 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
22257 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
22258 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
22259 (64 <= NumElems * ToSz))
22260 StoreType = MVT::f64;
22262 // Bitcast the original vector into a vector of store-size units
22263 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
22264 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
22265 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
22266 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
22267 SmallVector<SDValue, 8> Chains;
22268 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
22269 TLI.getPointerTy());
22270 SDValue Ptr = St->getBasePtr();
22272 // Perform one or more big stores into memory.
22273 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
22274 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
22275 StoreType, ShuffWide,
22276 DAG.getIntPtrConstant(i));
22277 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
22278 St->getPointerInfo(), St->isVolatile(),
22279 St->isNonTemporal(), St->getAlignment());
22280 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22281 Chains.push_back(Ch);
22284 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
22287 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
22288 // the FP state in cases where an emms may be missing.
22289 // A preferable solution to the general problem is to figure out the right
22290 // places to insert EMMS. This qualifies as a quick hack.
22292 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
22293 if (VT.getSizeInBits() != 64)
22296 const Function *F = DAG.getMachineFunction().getFunction();
22297 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
22298 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
22299 && Subtarget->hasSSE2();
22300 if ((VT.isVector() ||
22301 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
22302 isa<LoadSDNode>(St->getValue()) &&
22303 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
22304 St->getChain().hasOneUse() && !St->isVolatile()) {
22305 SDNode* LdVal = St->getValue().getNode();
22306 LoadSDNode *Ld = nullptr;
22307 int TokenFactorIndex = -1;
22308 SmallVector<SDValue, 8> Ops;
22309 SDNode* ChainVal = St->getChain().getNode();
22310 // Must be a store of a load. We currently handle two cases: the load
22311 // is a direct child, and it's under an intervening TokenFactor. It is
22312 // possible to dig deeper under nested TokenFactors.
22313 if (ChainVal == LdVal)
22314 Ld = cast<LoadSDNode>(St->getChain());
22315 else if (St->getValue().hasOneUse() &&
22316 ChainVal->getOpcode() == ISD::TokenFactor) {
22317 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
22318 if (ChainVal->getOperand(i).getNode() == LdVal) {
22319 TokenFactorIndex = i;
22320 Ld = cast<LoadSDNode>(St->getValue());
22322 Ops.push_back(ChainVal->getOperand(i));
22326 if (!Ld || !ISD::isNormalLoad(Ld))
22329 // If this is not the MMX case, i.e. we are just turning i64 load/store
22330 // into f64 load/store, avoid the transformation if there are multiple
22331 // uses of the loaded value.
22332 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
22337 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
22338 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
22340 if (Subtarget->is64Bit() || F64IsLegal) {
22341 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
22342 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
22343 Ld->getPointerInfo(), Ld->isVolatile(),
22344 Ld->isNonTemporal(), Ld->isInvariant(),
22345 Ld->getAlignment());
22346 SDValue NewChain = NewLd.getValue(1);
22347 if (TokenFactorIndex != -1) {
22348 Ops.push_back(NewChain);
22349 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22351 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
22352 St->getPointerInfo(),
22353 St->isVolatile(), St->isNonTemporal(),
22354 St->getAlignment());
22357 // Otherwise, lower to two pairs of 32-bit loads / stores.
22358 SDValue LoAddr = Ld->getBasePtr();
22359 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
22360 DAG.getConstant(4, MVT::i32));
22362 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
22363 Ld->getPointerInfo(),
22364 Ld->isVolatile(), Ld->isNonTemporal(),
22365 Ld->isInvariant(), Ld->getAlignment());
22366 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
22367 Ld->getPointerInfo().getWithOffset(4),
22368 Ld->isVolatile(), Ld->isNonTemporal(),
22370 MinAlign(Ld->getAlignment(), 4));
22372 SDValue NewChain = LoLd.getValue(1);
22373 if (TokenFactorIndex != -1) {
22374 Ops.push_back(LoLd);
22375 Ops.push_back(HiLd);
22376 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22379 LoAddr = St->getBasePtr();
22380 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
22381 DAG.getConstant(4, MVT::i32));
22383 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
22384 St->getPointerInfo(),
22385 St->isVolatile(), St->isNonTemporal(),
22386 St->getAlignment());
22387 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
22388 St->getPointerInfo().getWithOffset(4),
22390 St->isNonTemporal(),
22391 MinAlign(St->getAlignment(), 4));
22392 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
22397 /// Return 'true' if this vector operation is "horizontal"
22398 /// and return the operands for the horizontal operation in LHS and RHS. A
22399 /// horizontal operation performs the binary operation on successive elements
22400 /// of its first operand, then on successive elements of its second operand,
22401 /// returning the resulting values in a vector. For example, if
22402 /// A = < float a0, float a1, float a2, float a3 >
22404 /// B = < float b0, float b1, float b2, float b3 >
22405 /// then the result of doing a horizontal operation on A and B is
22406 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
22407 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
22408 /// A horizontal-op B, for some already available A and B, and if so then LHS is
22409 /// set to A, RHS to B, and the routine returns 'true'.
22410 /// Note that the binary operation should have the property that if one of the
22411 /// operands is UNDEF then the result is UNDEF.
22412 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
22413 // Look for the following pattern: if
22414 // A = < float a0, float a1, float a2, float a3 >
22415 // B = < float b0, float b1, float b2, float b3 >
22417 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
22418 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
22419 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
22420 // which is A horizontal-op B.
22422 // At least one of the operands should be a vector shuffle.
22423 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
22424 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
22427 MVT VT = LHS.getSimpleValueType();
22429 assert((VT.is128BitVector() || VT.is256BitVector()) &&
22430 "Unsupported vector type for horizontal add/sub");
22432 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
22433 // operate independently on 128-bit lanes.
22434 unsigned NumElts = VT.getVectorNumElements();
22435 unsigned NumLanes = VT.getSizeInBits()/128;
22436 unsigned NumLaneElts = NumElts / NumLanes;
22437 assert((NumLaneElts % 2 == 0) &&
22438 "Vector type should have an even number of elements in each lane");
22439 unsigned HalfLaneElts = NumLaneElts/2;
22441 // View LHS in the form
22442 // LHS = VECTOR_SHUFFLE A, B, LMask
22443 // If LHS is not a shuffle then pretend it is the shuffle
22444 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
22445 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
22448 SmallVector<int, 16> LMask(NumElts);
22449 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22450 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
22451 A = LHS.getOperand(0);
22452 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
22453 B = LHS.getOperand(1);
22454 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
22455 std::copy(Mask.begin(), Mask.end(), LMask.begin());
22457 if (LHS.getOpcode() != ISD::UNDEF)
22459 for (unsigned i = 0; i != NumElts; ++i)
22463 // Likewise, view RHS in the form
22464 // RHS = VECTOR_SHUFFLE C, D, RMask
22466 SmallVector<int, 16> RMask(NumElts);
22467 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22468 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
22469 C = RHS.getOperand(0);
22470 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
22471 D = RHS.getOperand(1);
22472 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
22473 std::copy(Mask.begin(), Mask.end(), RMask.begin());
22475 if (RHS.getOpcode() != ISD::UNDEF)
22477 for (unsigned i = 0; i != NumElts; ++i)
22481 // Check that the shuffles are both shuffling the same vectors.
22482 if (!(A == C && B == D) && !(A == D && B == C))
22485 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
22486 if (!A.getNode() && !B.getNode())
22489 // If A and B occur in reverse order in RHS, then "swap" them (which means
22490 // rewriting the mask).
22492 CommuteVectorShuffleMask(RMask, NumElts);
22494 // At this point LHS and RHS are equivalent to
22495 // LHS = VECTOR_SHUFFLE A, B, LMask
22496 // RHS = VECTOR_SHUFFLE A, B, RMask
22497 // Check that the masks correspond to performing a horizontal operation.
22498 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
22499 for (unsigned i = 0; i != NumLaneElts; ++i) {
22500 int LIdx = LMask[i+l], RIdx = RMask[i+l];
22502 // Ignore any UNDEF components.
22503 if (LIdx < 0 || RIdx < 0 ||
22504 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
22505 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
22508 // Check that successive elements are being operated on. If not, this is
22509 // not a horizontal operation.
22510 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
22511 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
22512 if (!(LIdx == Index && RIdx == Index + 1) &&
22513 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
22518 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
22519 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
22523 /// Do target-specific dag combines on floating point adds.
22524 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
22525 const X86Subtarget *Subtarget) {
22526 EVT VT = N->getValueType(0);
22527 SDValue LHS = N->getOperand(0);
22528 SDValue RHS = N->getOperand(1);
22530 // Try to synthesize horizontal adds from adds of shuffles.
22531 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22532 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22533 isHorizontalBinOp(LHS, RHS, true))
22534 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
22538 /// Do target-specific dag combines on floating point subs.
22539 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
22540 const X86Subtarget *Subtarget) {
22541 EVT VT = N->getValueType(0);
22542 SDValue LHS = N->getOperand(0);
22543 SDValue RHS = N->getOperand(1);
22545 // Try to synthesize horizontal subs from subs of shuffles.
22546 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22547 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22548 isHorizontalBinOp(LHS, RHS, false))
22549 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
22553 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
22554 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
22555 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
22557 // F[X]OR(0.0, x) -> x
22558 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22559 if (C->getValueAPF().isPosZero())
22560 return N->getOperand(1);
22562 // F[X]OR(x, 0.0) -> x
22563 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22564 if (C->getValueAPF().isPosZero())
22565 return N->getOperand(0);
22569 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
22570 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
22571 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
22573 // Only perform optimizations if UnsafeMath is used.
22574 if (!DAG.getTarget().Options.UnsafeFPMath)
22577 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
22578 // into FMINC and FMAXC, which are Commutative operations.
22579 unsigned NewOp = 0;
22580 switch (N->getOpcode()) {
22581 default: llvm_unreachable("unknown opcode");
22582 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
22583 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
22586 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
22587 N->getOperand(0), N->getOperand(1));
22590 /// Do target-specific dag combines on X86ISD::FAND nodes.
22591 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
22592 // FAND(0.0, x) -> 0.0
22593 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22594 if (C->getValueAPF().isPosZero())
22595 return N->getOperand(0);
22597 // FAND(x, 0.0) -> 0.0
22598 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22599 if (C->getValueAPF().isPosZero())
22600 return N->getOperand(1);
22605 /// Do target-specific dag combines on X86ISD::FANDN nodes
22606 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
22607 // FANDN(0.0, x) -> x
22608 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22609 if (C->getValueAPF().isPosZero())
22610 return N->getOperand(1);
22612 // FANDN(x, 0.0) -> 0.0
22613 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22614 if (C->getValueAPF().isPosZero())
22615 return N->getOperand(1);
22620 static SDValue PerformBTCombine(SDNode *N,
22622 TargetLowering::DAGCombinerInfo &DCI) {
22623 // BT ignores high bits in the bit index operand.
22624 SDValue Op1 = N->getOperand(1);
22625 if (Op1.hasOneUse()) {
22626 unsigned BitWidth = Op1.getValueSizeInBits();
22627 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
22628 APInt KnownZero, KnownOne;
22629 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
22630 !DCI.isBeforeLegalizeOps());
22631 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22632 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
22633 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
22634 DCI.CommitTargetLoweringOpt(TLO);
22639 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
22640 SDValue Op = N->getOperand(0);
22641 if (Op.getOpcode() == ISD::BITCAST)
22642 Op = Op.getOperand(0);
22643 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
22644 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
22645 VT.getVectorElementType().getSizeInBits() ==
22646 OpVT.getVectorElementType().getSizeInBits()) {
22647 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
22652 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
22653 const X86Subtarget *Subtarget) {
22654 EVT VT = N->getValueType(0);
22655 if (!VT.isVector())
22658 SDValue N0 = N->getOperand(0);
22659 SDValue N1 = N->getOperand(1);
22660 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
22663 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
22664 // both SSE and AVX2 since there is no sign-extended shift right
22665 // operation on a vector with 64-bit elements.
22666 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
22667 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
22668 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
22669 N0.getOpcode() == ISD::SIGN_EXTEND)) {
22670 SDValue N00 = N0.getOperand(0);
22672 // EXTLOAD has a better solution on AVX2,
22673 // it may be replaced with X86ISD::VSEXT node.
22674 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
22675 if (!ISD::isNormalLoad(N00.getNode()))
22678 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
22679 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
22681 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
22687 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
22688 TargetLowering::DAGCombinerInfo &DCI,
22689 const X86Subtarget *Subtarget) {
22690 SDValue N0 = N->getOperand(0);
22691 EVT VT = N->getValueType(0);
22693 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
22694 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
22695 // This exposes the sext to the sdivrem lowering, so that it directly extends
22696 // from AH (which we otherwise need to do contortions to access).
22697 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
22698 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
22700 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22701 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
22702 N0.getOperand(0), N0.getOperand(1));
22703 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22704 return R.getValue(1);
22707 if (!DCI.isBeforeLegalizeOps())
22710 if (!Subtarget->hasFp256())
22713 if (VT.isVector() && VT.getSizeInBits() == 256) {
22714 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22722 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
22723 const X86Subtarget* Subtarget) {
22725 EVT VT = N->getValueType(0);
22727 // Let legalize expand this if it isn't a legal type yet.
22728 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
22731 EVT ScalarVT = VT.getScalarType();
22732 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
22733 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
22736 SDValue A = N->getOperand(0);
22737 SDValue B = N->getOperand(1);
22738 SDValue C = N->getOperand(2);
22740 bool NegA = (A.getOpcode() == ISD::FNEG);
22741 bool NegB = (B.getOpcode() == ISD::FNEG);
22742 bool NegC = (C.getOpcode() == ISD::FNEG);
22744 // Negative multiplication when NegA xor NegB
22745 bool NegMul = (NegA != NegB);
22747 A = A.getOperand(0);
22749 B = B.getOperand(0);
22751 C = C.getOperand(0);
22755 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
22757 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
22759 return DAG.getNode(Opcode, dl, VT, A, B, C);
22762 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
22763 TargetLowering::DAGCombinerInfo &DCI,
22764 const X86Subtarget *Subtarget) {
22765 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
22766 // (and (i32 x86isd::setcc_carry), 1)
22767 // This eliminates the zext. This transformation is necessary because
22768 // ISD::SETCC is always legalized to i8.
22770 SDValue N0 = N->getOperand(0);
22771 EVT VT = N->getValueType(0);
22773 if (N0.getOpcode() == ISD::AND &&
22775 N0.getOperand(0).hasOneUse()) {
22776 SDValue N00 = N0.getOperand(0);
22777 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22778 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22779 if (!C || C->getZExtValue() != 1)
22781 return DAG.getNode(ISD::AND, dl, VT,
22782 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22783 N00.getOperand(0), N00.getOperand(1)),
22784 DAG.getConstant(1, VT));
22788 if (N0.getOpcode() == ISD::TRUNCATE &&
22790 N0.getOperand(0).hasOneUse()) {
22791 SDValue N00 = N0.getOperand(0);
22792 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22793 return DAG.getNode(ISD::AND, dl, VT,
22794 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22795 N00.getOperand(0), N00.getOperand(1)),
22796 DAG.getConstant(1, VT));
22799 if (VT.is256BitVector()) {
22800 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22805 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
22806 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
22807 // This exposes the zext to the udivrem lowering, so that it directly extends
22808 // from AH (which we otherwise need to do contortions to access).
22809 if (N0.getOpcode() == ISD::UDIVREM &&
22810 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
22811 (VT == MVT::i32 || VT == MVT::i64)) {
22812 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22813 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
22814 N0.getOperand(0), N0.getOperand(1));
22815 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22816 return R.getValue(1);
22822 // Optimize x == -y --> x+y == 0
22823 // x != -y --> x+y != 0
22824 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
22825 const X86Subtarget* Subtarget) {
22826 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
22827 SDValue LHS = N->getOperand(0);
22828 SDValue RHS = N->getOperand(1);
22829 EVT VT = N->getValueType(0);
22832 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
22833 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
22834 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
22835 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22836 LHS.getValueType(), RHS, LHS.getOperand(1));
22837 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22838 addV, DAG.getConstant(0, addV.getValueType()), CC);
22840 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
22841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
22842 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
22843 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22844 RHS.getValueType(), LHS, RHS.getOperand(1));
22845 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22846 addV, DAG.getConstant(0, addV.getValueType()), CC);
22849 if (VT.getScalarType() == MVT::i1) {
22850 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
22851 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22852 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
22853 if (!IsSEXT0 && !IsVZero0)
22855 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
22856 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22857 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
22859 if (!IsSEXT1 && !IsVZero1)
22862 if (IsSEXT0 && IsVZero1) {
22863 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
22864 if (CC == ISD::SETEQ)
22865 return DAG.getNOT(DL, LHS.getOperand(0), VT);
22866 return LHS.getOperand(0);
22868 if (IsSEXT1 && IsVZero0) {
22869 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
22870 if (CC == ISD::SETEQ)
22871 return DAG.getNOT(DL, RHS.getOperand(0), VT);
22872 return RHS.getOperand(0);
22879 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
22880 SelectionDAG &DAG) {
22882 MVT VT = Load->getSimpleValueType(0);
22883 MVT EVT = VT.getVectorElementType();
22884 SDValue Addr = Load->getOperand(1);
22885 SDValue NewAddr = DAG.getNode(
22886 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
22887 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
22890 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
22891 DAG.getMachineFunction().getMachineMemOperand(
22892 Load->getMemOperand(), 0, EVT.getStoreSize()));
22896 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
22897 const X86Subtarget *Subtarget) {
22899 MVT VT = N->getOperand(1)->getSimpleValueType(0);
22900 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
22901 "X86insertps is only defined for v4x32");
22903 SDValue Ld = N->getOperand(1);
22904 if (MayFoldLoad(Ld)) {
22905 // Extract the countS bits from the immediate so we can get the proper
22906 // address when narrowing the vector load to a specific element.
22907 // When the second source op is a memory address, insertps doesn't use
22908 // countS and just gets an f32 from that address.
22909 unsigned DestIndex =
22910 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
22912 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
22914 // Create this as a scalar to vector to match the instruction pattern.
22915 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
22916 // countS bits are ignored when loading from memory on insertps, which
22917 // means we don't need to explicitly set them to 0.
22918 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
22919 LoadScalarToVector, N->getOperand(2));
22924 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
22925 SDValue V0 = N->getOperand(0);
22926 SDValue V1 = N->getOperand(1);
22928 EVT VT = N->getValueType(0);
22930 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
22931 // operands and changing the mask to 1. This saves us a bunch of
22932 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
22933 // x86InstrInfo knows how to commute this back after instruction selection
22934 // if it would help register allocation.
22936 // TODO: If optimizing for size or a processor that doesn't suffer from
22937 // partial register update stalls, this should be transformed into a MOVSD
22938 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
22940 if (VT == MVT::v2f64)
22941 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
22942 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
22943 SDValue NewMask = DAG.getConstant(1, MVT::i8);
22944 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
22950 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
22951 // as "sbb reg,reg", since it can be extended without zext and produces
22952 // an all-ones bit which is more useful than 0/1 in some cases.
22953 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
22956 return DAG.getNode(ISD::AND, DL, VT,
22957 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
22958 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
22959 DAG.getConstant(1, VT));
22960 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
22961 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
22962 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
22963 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
22966 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
22967 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
22968 TargetLowering::DAGCombinerInfo &DCI,
22969 const X86Subtarget *Subtarget) {
22971 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
22972 SDValue EFLAGS = N->getOperand(1);
22974 if (CC == X86::COND_A) {
22975 // Try to convert COND_A into COND_B in an attempt to facilitate
22976 // materializing "setb reg".
22978 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
22979 // cannot take an immediate as its first operand.
22981 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
22982 EFLAGS.getValueType().isInteger() &&
22983 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
22984 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
22985 EFLAGS.getNode()->getVTList(),
22986 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
22987 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
22988 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
22992 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
22993 // a zext and produces an all-ones bit which is more useful than 0/1 in some
22995 if (CC == X86::COND_B)
22996 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23000 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23001 if (Flags.getNode()) {
23002 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23003 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23009 // Optimize branch condition evaluation.
23011 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23012 TargetLowering::DAGCombinerInfo &DCI,
23013 const X86Subtarget *Subtarget) {
23015 SDValue Chain = N->getOperand(0);
23016 SDValue Dest = N->getOperand(1);
23017 SDValue EFLAGS = N->getOperand(3);
23018 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23022 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23023 if (Flags.getNode()) {
23024 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23025 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23032 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23033 SelectionDAG &DAG) {
23034 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23035 // optimize away operation when it's from a constant.
23037 // The general transformation is:
23038 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23039 // AND(VECTOR_CMP(x,y), constant2)
23040 // constant2 = UNARYOP(constant)
23042 // Early exit if this isn't a vector operation, the operand of the
23043 // unary operation isn't a bitwise AND, or if the sizes of the operations
23044 // aren't the same.
23045 EVT VT = N->getValueType(0);
23046 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23047 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23048 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23051 // Now check that the other operand of the AND is a constant. We could
23052 // make the transformation for non-constant splats as well, but it's unclear
23053 // that would be a benefit as it would not eliminate any operations, just
23054 // perform one more step in scalar code before moving to the vector unit.
23055 if (BuildVectorSDNode *BV =
23056 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23057 // Bail out if the vector isn't a constant.
23058 if (!BV->isConstant())
23061 // Everything checks out. Build up the new and improved node.
23063 EVT IntVT = BV->getValueType(0);
23064 // Create a new constant of the appropriate type for the transformed
23066 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23067 // The AND node needs bitcasts to/from an integer vector type around it.
23068 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
23069 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
23070 N->getOperand(0)->getOperand(0), MaskConst);
23071 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
23078 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
23079 const X86Subtarget *Subtarget) {
23080 // First try to optimize away the conversion entirely when it's
23081 // conditionally from a constant. Vectors only.
23082 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
23083 if (Res != SDValue())
23086 // Now move on to more general possibilities.
23087 SDValue Op0 = N->getOperand(0);
23088 EVT InVT = Op0->getValueType(0);
23090 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
23091 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
23093 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
23094 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
23095 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
23098 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
23099 // a 32-bit target where SSE doesn't support i64->FP operations.
23100 if (Op0.getOpcode() == ISD::LOAD) {
23101 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
23102 EVT VT = Ld->getValueType(0);
23103 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
23104 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
23105 !Subtarget->is64Bit() && VT == MVT::i64) {
23106 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
23107 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
23108 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
23115 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
23116 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
23117 X86TargetLowering::DAGCombinerInfo &DCI) {
23118 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
23119 // the result is either zero or one (depending on the input carry bit).
23120 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
23121 if (X86::isZeroNode(N->getOperand(0)) &&
23122 X86::isZeroNode(N->getOperand(1)) &&
23123 // We don't have a good way to replace an EFLAGS use, so only do this when
23125 SDValue(N, 1).use_empty()) {
23127 EVT VT = N->getValueType(0);
23128 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
23129 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
23130 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
23131 DAG.getConstant(X86::COND_B,MVT::i8),
23133 DAG.getConstant(1, VT));
23134 return DCI.CombineTo(N, Res1, CarryOut);
23140 // fold (add Y, (sete X, 0)) -> adc 0, Y
23141 // (add Y, (setne X, 0)) -> sbb -1, Y
23142 // (sub (sete X, 0), Y) -> sbb 0, Y
23143 // (sub (setne X, 0), Y) -> adc -1, Y
23144 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
23147 // Look through ZExts.
23148 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
23149 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
23152 SDValue SetCC = Ext.getOperand(0);
23153 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
23156 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
23157 if (CC != X86::COND_E && CC != X86::COND_NE)
23160 SDValue Cmp = SetCC.getOperand(1);
23161 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
23162 !X86::isZeroNode(Cmp.getOperand(1)) ||
23163 !Cmp.getOperand(0).getValueType().isInteger())
23166 SDValue CmpOp0 = Cmp.getOperand(0);
23167 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
23168 DAG.getConstant(1, CmpOp0.getValueType()));
23170 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
23171 if (CC == X86::COND_NE)
23172 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
23173 DL, OtherVal.getValueType(), OtherVal,
23174 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
23175 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
23176 DL, OtherVal.getValueType(), OtherVal,
23177 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
23180 /// PerformADDCombine - Do target-specific dag combines on integer adds.
23181 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
23182 const X86Subtarget *Subtarget) {
23183 EVT VT = N->getValueType(0);
23184 SDValue Op0 = N->getOperand(0);
23185 SDValue Op1 = N->getOperand(1);
23187 // Try to synthesize horizontal adds from adds of shuffles.
23188 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23189 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23190 isHorizontalBinOp(Op0, Op1, true))
23191 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
23193 return OptimizeConditionalInDecrement(N, DAG);
23196 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
23197 const X86Subtarget *Subtarget) {
23198 SDValue Op0 = N->getOperand(0);
23199 SDValue Op1 = N->getOperand(1);
23201 // X86 can't encode an immediate LHS of a sub. See if we can push the
23202 // negation into a preceding instruction.
23203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
23204 // If the RHS of the sub is a XOR with one use and a constant, invert the
23205 // immediate. Then add one to the LHS of the sub so we can turn
23206 // X-Y -> X+~Y+1, saving one register.
23207 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
23208 isa<ConstantSDNode>(Op1.getOperand(1))) {
23209 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
23210 EVT VT = Op0.getValueType();
23211 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
23213 DAG.getConstant(~XorC, VT));
23214 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
23215 DAG.getConstant(C->getAPIntValue()+1, VT));
23219 // Try to synthesize horizontal adds from adds of shuffles.
23220 EVT VT = N->getValueType(0);
23221 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23222 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23223 isHorizontalBinOp(Op0, Op1, true))
23224 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
23226 return OptimizeConditionalInDecrement(N, DAG);
23229 /// performVZEXTCombine - Performs build vector combines
23230 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
23231 TargetLowering::DAGCombinerInfo &DCI,
23232 const X86Subtarget *Subtarget) {
23234 MVT VT = N->getSimpleValueType(0);
23235 SDValue Op = N->getOperand(0);
23236 MVT OpVT = Op.getSimpleValueType();
23237 MVT OpEltVT = OpVT.getVectorElementType();
23238 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
23240 // (vzext (bitcast (vzext (x)) -> (vzext x)
23242 while (V.getOpcode() == ISD::BITCAST)
23243 V = V.getOperand(0);
23245 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
23246 MVT InnerVT = V.getSimpleValueType();
23247 MVT InnerEltVT = InnerVT.getVectorElementType();
23249 // If the element sizes match exactly, we can just do one larger vzext. This
23250 // is always an exact type match as vzext operates on integer types.
23251 if (OpEltVT == InnerEltVT) {
23252 assert(OpVT == InnerVT && "Types must match for vzext!");
23253 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
23256 // The only other way we can combine them is if only a single element of the
23257 // inner vzext is used in the input to the outer vzext.
23258 if (InnerEltVT.getSizeInBits() < InputBits)
23261 // In this case, the inner vzext is completely dead because we're going to
23262 // only look at bits inside of the low element. Just do the outer vzext on
23263 // a bitcast of the input to the inner.
23264 return DAG.getNode(X86ISD::VZEXT, DL, VT,
23265 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
23268 // Check if we can bypass extracting and re-inserting an element of an input
23269 // vector. Essentialy:
23270 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
23271 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
23272 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
23273 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
23274 SDValue ExtractedV = V.getOperand(0);
23275 SDValue OrigV = ExtractedV.getOperand(0);
23276 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
23277 if (ExtractIdx->getZExtValue() == 0) {
23278 MVT OrigVT = OrigV.getSimpleValueType();
23279 // Extract a subvector if necessary...
23280 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
23281 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
23282 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
23283 OrigVT.getVectorNumElements() / Ratio);
23284 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
23285 DAG.getIntPtrConstant(0));
23287 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
23288 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
23295 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
23296 DAGCombinerInfo &DCI) const {
23297 SelectionDAG &DAG = DCI.DAG;
23298 switch (N->getOpcode()) {
23300 case ISD::EXTRACT_VECTOR_ELT:
23301 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
23304 case X86ISD::SHRUNKBLEND:
23305 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
23306 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
23307 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
23308 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
23309 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
23310 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
23311 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
23314 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
23315 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
23316 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
23317 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
23318 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
23319 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
23320 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
23321 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
23322 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
23323 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
23324 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
23326 case X86ISD::FOR: return PerformFORCombine(N, DAG);
23328 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
23329 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
23330 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
23331 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
23332 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
23333 case ISD::ANY_EXTEND:
23334 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
23335 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
23336 case ISD::SIGN_EXTEND_INREG:
23337 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
23338 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
23339 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
23340 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
23341 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
23342 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
23343 case X86ISD::SHUFP: // Handle all target specific shuffles
23344 case X86ISD::PALIGNR:
23345 case X86ISD::UNPCKH:
23346 case X86ISD::UNPCKL:
23347 case X86ISD::MOVHLPS:
23348 case X86ISD::MOVLHPS:
23349 case X86ISD::PSHUFB:
23350 case X86ISD::PSHUFD:
23351 case X86ISD::PSHUFHW:
23352 case X86ISD::PSHUFLW:
23353 case X86ISD::MOVSS:
23354 case X86ISD::MOVSD:
23355 case X86ISD::VPERMILPI:
23356 case X86ISD::VPERM2X128:
23357 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
23358 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
23359 case ISD::INTRINSIC_WO_CHAIN:
23360 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
23361 case X86ISD::INSERTPS: {
23362 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
23363 return PerformINSERTPSCombine(N, DAG, Subtarget);
23366 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
23367 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
23373 /// isTypeDesirableForOp - Return true if the target has native support for
23374 /// the specified value type and it is 'desirable' to use the type for the
23375 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
23376 /// instruction encodings are longer and some i16 instructions are slow.
23377 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
23378 if (!isTypeLegal(VT))
23380 if (VT != MVT::i16)
23387 case ISD::SIGN_EXTEND:
23388 case ISD::ZERO_EXTEND:
23389 case ISD::ANY_EXTEND:
23402 /// IsDesirableToPromoteOp - This method query the target whether it is
23403 /// beneficial for dag combiner to promote the specified node. If true, it
23404 /// should return the desired promotion type by reference.
23405 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
23406 EVT VT = Op.getValueType();
23407 if (VT != MVT::i16)
23410 bool Promote = false;
23411 bool Commute = false;
23412 switch (Op.getOpcode()) {
23415 LoadSDNode *LD = cast<LoadSDNode>(Op);
23416 // If the non-extending load has a single use and it's not live out, then it
23417 // might be folded.
23418 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
23419 Op.hasOneUse()*/) {
23420 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
23421 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
23422 // The only case where we'd want to promote LOAD (rather then it being
23423 // promoted as an operand is when it's only use is liveout.
23424 if (UI->getOpcode() != ISD::CopyToReg)
23431 case ISD::SIGN_EXTEND:
23432 case ISD::ZERO_EXTEND:
23433 case ISD::ANY_EXTEND:
23438 SDValue N0 = Op.getOperand(0);
23439 // Look out for (store (shl (load), x)).
23440 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
23453 SDValue N0 = Op.getOperand(0);
23454 SDValue N1 = Op.getOperand(1);
23455 if (!Commute && MayFoldLoad(N1))
23457 // Avoid disabling potential load folding opportunities.
23458 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
23460 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
23470 //===----------------------------------------------------------------------===//
23471 // X86 Inline Assembly Support
23472 //===----------------------------------------------------------------------===//
23475 // Helper to match a string separated by whitespace.
23476 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
23477 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
23479 for (unsigned i = 0, e = args.size(); i != e; ++i) {
23480 StringRef piece(*args[i]);
23481 if (!s.startswith(piece)) // Check if the piece matches.
23484 s = s.substr(piece.size());
23485 StringRef::size_type pos = s.find_first_not_of(" \t");
23486 if (pos == 0) // We matched a prefix.
23494 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
23497 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
23499 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
23500 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
23501 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
23502 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
23504 if (AsmPieces.size() == 3)
23506 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
23513 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
23514 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
23516 std::string AsmStr = IA->getAsmString();
23518 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
23519 if (!Ty || Ty->getBitWidth() % 16 != 0)
23522 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
23523 SmallVector<StringRef, 4> AsmPieces;
23524 SplitString(AsmStr, AsmPieces, ";\n");
23526 switch (AsmPieces.size()) {
23527 default: return false;
23529 // FIXME: this should verify that we are targeting a 486 or better. If not,
23530 // we will turn this bswap into something that will be lowered to logical
23531 // ops instead of emitting the bswap asm. For now, we don't support 486 or
23532 // lower so don't worry about this.
23534 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
23535 matchAsm(AsmPieces[0], "bswapl", "$0") ||
23536 matchAsm(AsmPieces[0], "bswapq", "$0") ||
23537 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
23538 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
23539 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
23540 // No need to check constraints, nothing other than the equivalent of
23541 // "=r,0" would be valid here.
23542 return IntrinsicLowering::LowerToByteSwap(CI);
23545 // rorw $$8, ${0:w} --> llvm.bswap.i16
23546 if (CI->getType()->isIntegerTy(16) &&
23547 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23548 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
23549 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
23551 const std::string &ConstraintsStr = IA->getConstraintString();
23552 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23553 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23554 if (clobbersFlagRegisters(AsmPieces))
23555 return IntrinsicLowering::LowerToByteSwap(CI);
23559 if (CI->getType()->isIntegerTy(32) &&
23560 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23561 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
23562 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
23563 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
23565 const std::string &ConstraintsStr = IA->getConstraintString();
23566 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23567 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23568 if (clobbersFlagRegisters(AsmPieces))
23569 return IntrinsicLowering::LowerToByteSwap(CI);
23572 if (CI->getType()->isIntegerTy(64)) {
23573 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
23574 if (Constraints.size() >= 2 &&
23575 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
23576 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
23577 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
23578 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
23579 matchAsm(AsmPieces[1], "bswap", "%edx") &&
23580 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
23581 return IntrinsicLowering::LowerToByteSwap(CI);
23589 /// getConstraintType - Given a constraint letter, return the type of
23590 /// constraint it is for this target.
23591 X86TargetLowering::ConstraintType
23592 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
23593 if (Constraint.size() == 1) {
23594 switch (Constraint[0]) {
23605 return C_RegisterClass;
23629 return TargetLowering::getConstraintType(Constraint);
23632 /// Examine constraint type and operand type and determine a weight value.
23633 /// This object must already have been set up with the operand type
23634 /// and the current alternative constraint selected.
23635 TargetLowering::ConstraintWeight
23636 X86TargetLowering::getSingleConstraintMatchWeight(
23637 AsmOperandInfo &info, const char *constraint) const {
23638 ConstraintWeight weight = CW_Invalid;
23639 Value *CallOperandVal = info.CallOperandVal;
23640 // If we don't have a value, we can't do a match,
23641 // but allow it at the lowest weight.
23642 if (!CallOperandVal)
23644 Type *type = CallOperandVal->getType();
23645 // Look at the constraint type.
23646 switch (*constraint) {
23648 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
23659 if (CallOperandVal->getType()->isIntegerTy())
23660 weight = CW_SpecificReg;
23665 if (type->isFloatingPointTy())
23666 weight = CW_SpecificReg;
23669 if (type->isX86_MMXTy() && Subtarget->hasMMX())
23670 weight = CW_SpecificReg;
23674 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
23675 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
23676 weight = CW_Register;
23679 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
23680 if (C->getZExtValue() <= 31)
23681 weight = CW_Constant;
23685 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23686 if (C->getZExtValue() <= 63)
23687 weight = CW_Constant;
23691 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23692 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
23693 weight = CW_Constant;
23697 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23698 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
23699 weight = CW_Constant;
23703 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23704 if (C->getZExtValue() <= 3)
23705 weight = CW_Constant;
23709 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23710 if (C->getZExtValue() <= 0xff)
23711 weight = CW_Constant;
23716 if (dyn_cast<ConstantFP>(CallOperandVal)) {
23717 weight = CW_Constant;
23721 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23722 if ((C->getSExtValue() >= -0x80000000LL) &&
23723 (C->getSExtValue() <= 0x7fffffffLL))
23724 weight = CW_Constant;
23728 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23729 if (C->getZExtValue() <= 0xffffffff)
23730 weight = CW_Constant;
23737 /// LowerXConstraint - try to replace an X constraint, which matches anything,
23738 /// with another that has more specific requirements based on the type of the
23739 /// corresponding operand.
23740 const char *X86TargetLowering::
23741 LowerXConstraint(EVT ConstraintVT) const {
23742 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
23743 // 'f' like normal targets.
23744 if (ConstraintVT.isFloatingPoint()) {
23745 if (Subtarget->hasSSE2())
23747 if (Subtarget->hasSSE1())
23751 return TargetLowering::LowerXConstraint(ConstraintVT);
23754 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
23755 /// vector. If it is invalid, don't add anything to Ops.
23756 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
23757 std::string &Constraint,
23758 std::vector<SDValue>&Ops,
23759 SelectionDAG &DAG) const {
23762 // Only support length 1 constraints for now.
23763 if (Constraint.length() > 1) return;
23765 char ConstraintLetter = Constraint[0];
23766 switch (ConstraintLetter) {
23769 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23770 if (C->getZExtValue() <= 31) {
23771 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23777 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23778 if (C->getZExtValue() <= 63) {
23779 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23785 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23786 if (isInt<8>(C->getSExtValue())) {
23787 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23793 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23794 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
23795 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
23796 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
23802 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23803 if (C->getZExtValue() <= 3) {
23804 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23810 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23811 if (C->getZExtValue() <= 255) {
23812 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23818 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23819 if (C->getZExtValue() <= 127) {
23820 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23826 // 32-bit signed value
23827 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23828 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23829 C->getSExtValue())) {
23830 // Widen to 64 bits here to get it sign extended.
23831 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
23834 // FIXME gcc accepts some relocatable values here too, but only in certain
23835 // memory models; it's complicated.
23840 // 32-bit unsigned value
23841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23842 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23843 C->getZExtValue())) {
23844 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23848 // FIXME gcc accepts some relocatable values here too, but only in certain
23849 // memory models; it's complicated.
23853 // Literal immediates are always ok.
23854 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
23855 // Widen to 64 bits here to get it sign extended.
23856 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
23860 // In any sort of PIC mode addresses need to be computed at runtime by
23861 // adding in a register or some sort of table lookup. These can't
23862 // be used as immediates.
23863 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
23866 // If we are in non-pic codegen mode, we allow the address of a global (with
23867 // an optional displacement) to be used with 'i'.
23868 GlobalAddressSDNode *GA = nullptr;
23869 int64_t Offset = 0;
23871 // Match either (GA), (GA+C), (GA+C1+C2), etc.
23873 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
23874 Offset += GA->getOffset();
23876 } else if (Op.getOpcode() == ISD::ADD) {
23877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23878 Offset += C->getZExtValue();
23879 Op = Op.getOperand(0);
23882 } else if (Op.getOpcode() == ISD::SUB) {
23883 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23884 Offset += -C->getZExtValue();
23885 Op = Op.getOperand(0);
23890 // Otherwise, this isn't something we can handle, reject it.
23894 const GlobalValue *GV = GA->getGlobal();
23895 // If we require an extra load to get this address, as in PIC mode, we
23896 // can't accept it.
23897 if (isGlobalStubReference(
23898 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
23901 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
23902 GA->getValueType(0), Offset);
23907 if (Result.getNode()) {
23908 Ops.push_back(Result);
23911 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
23914 std::pair<unsigned, const TargetRegisterClass*>
23915 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
23917 // First, see if this is a constraint that directly corresponds to an LLVM
23919 if (Constraint.size() == 1) {
23920 // GCC Constraint Letters
23921 switch (Constraint[0]) {
23923 // TODO: Slight differences here in allocation order and leaving
23924 // RIP in the class. Do they matter any more here than they do
23925 // in the normal allocation?
23926 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
23927 if (Subtarget->is64Bit()) {
23928 if (VT == MVT::i32 || VT == MVT::f32)
23929 return std::make_pair(0U, &X86::GR32RegClass);
23930 if (VT == MVT::i16)
23931 return std::make_pair(0U, &X86::GR16RegClass);
23932 if (VT == MVT::i8 || VT == MVT::i1)
23933 return std::make_pair(0U, &X86::GR8RegClass);
23934 if (VT == MVT::i64 || VT == MVT::f64)
23935 return std::make_pair(0U, &X86::GR64RegClass);
23938 // 32-bit fallthrough
23939 case 'Q': // Q_REGS
23940 if (VT == MVT::i32 || VT == MVT::f32)
23941 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
23942 if (VT == MVT::i16)
23943 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
23944 if (VT == MVT::i8 || VT == MVT::i1)
23945 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
23946 if (VT == MVT::i64)
23947 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
23949 case 'r': // GENERAL_REGS
23950 case 'l': // INDEX_REGS
23951 if (VT == MVT::i8 || VT == MVT::i1)
23952 return std::make_pair(0U, &X86::GR8RegClass);
23953 if (VT == MVT::i16)
23954 return std::make_pair(0U, &X86::GR16RegClass);
23955 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
23956 return std::make_pair(0U, &X86::GR32RegClass);
23957 return std::make_pair(0U, &X86::GR64RegClass);
23958 case 'R': // LEGACY_REGS
23959 if (VT == MVT::i8 || VT == MVT::i1)
23960 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
23961 if (VT == MVT::i16)
23962 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
23963 if (VT == MVT::i32 || !Subtarget->is64Bit())
23964 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
23965 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
23966 case 'f': // FP Stack registers.
23967 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
23968 // value to the correct fpstack register class.
23969 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
23970 return std::make_pair(0U, &X86::RFP32RegClass);
23971 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
23972 return std::make_pair(0U, &X86::RFP64RegClass);
23973 return std::make_pair(0U, &X86::RFP80RegClass);
23974 case 'y': // MMX_REGS if MMX allowed.
23975 if (!Subtarget->hasMMX()) break;
23976 return std::make_pair(0U, &X86::VR64RegClass);
23977 case 'Y': // SSE_REGS if SSE2 allowed
23978 if (!Subtarget->hasSSE2()) break;
23980 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
23981 if (!Subtarget->hasSSE1()) break;
23983 switch (VT.SimpleTy) {
23985 // Scalar SSE types.
23988 return std::make_pair(0U, &X86::FR32RegClass);
23991 return std::make_pair(0U, &X86::FR64RegClass);
23999 return std::make_pair(0U, &X86::VR128RegClass);
24007 return std::make_pair(0U, &X86::VR256RegClass);
24012 return std::make_pair(0U, &X86::VR512RegClass);
24018 // Use the default implementation in TargetLowering to convert the register
24019 // constraint into a member of a register class.
24020 std::pair<unsigned, const TargetRegisterClass*> Res;
24021 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
24023 // Not found as a standard register?
24025 // Map st(0) -> st(7) -> ST0
24026 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24027 tolower(Constraint[1]) == 's' &&
24028 tolower(Constraint[2]) == 't' &&
24029 Constraint[3] == '(' &&
24030 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24031 Constraint[5] == ')' &&
24032 Constraint[6] == '}') {
24034 Res.first = X86::FP0+Constraint[4]-'0';
24035 Res.second = &X86::RFP80RegClass;
24039 // GCC allows "st(0)" to be called just plain "st".
24040 if (StringRef("{st}").equals_lower(Constraint)) {
24041 Res.first = X86::FP0;
24042 Res.second = &X86::RFP80RegClass;
24047 if (StringRef("{flags}").equals_lower(Constraint)) {
24048 Res.first = X86::EFLAGS;
24049 Res.second = &X86::CCRRegClass;
24053 // 'A' means EAX + EDX.
24054 if (Constraint == "A") {
24055 Res.first = X86::EAX;
24056 Res.second = &X86::GR32_ADRegClass;
24062 // Otherwise, check to see if this is a register class of the wrong value
24063 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24064 // turn into {ax},{dx}.
24065 if (Res.second->hasType(VT))
24066 return Res; // Correct type already, nothing to do.
24068 // All of the single-register GCC register classes map their values onto
24069 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24070 // really want an 8-bit or 32-bit register, map to the appropriate register
24071 // class and return the appropriate register.
24072 if (Res.second == &X86::GR16RegClass) {
24073 if (VT == MVT::i8 || VT == MVT::i1) {
24074 unsigned DestReg = 0;
24075 switch (Res.first) {
24077 case X86::AX: DestReg = X86::AL; break;
24078 case X86::DX: DestReg = X86::DL; break;
24079 case X86::CX: DestReg = X86::CL; break;
24080 case X86::BX: DestReg = X86::BL; break;
24083 Res.first = DestReg;
24084 Res.second = &X86::GR8RegClass;
24086 } else if (VT == MVT::i32 || VT == MVT::f32) {
24087 unsigned DestReg = 0;
24088 switch (Res.first) {
24090 case X86::AX: DestReg = X86::EAX; break;
24091 case X86::DX: DestReg = X86::EDX; break;
24092 case X86::CX: DestReg = X86::ECX; break;
24093 case X86::BX: DestReg = X86::EBX; break;
24094 case X86::SI: DestReg = X86::ESI; break;
24095 case X86::DI: DestReg = X86::EDI; break;
24096 case X86::BP: DestReg = X86::EBP; break;
24097 case X86::SP: DestReg = X86::ESP; break;
24100 Res.first = DestReg;
24101 Res.second = &X86::GR32RegClass;
24103 } else if (VT == MVT::i64 || VT == MVT::f64) {
24104 unsigned DestReg = 0;
24105 switch (Res.first) {
24107 case X86::AX: DestReg = X86::RAX; break;
24108 case X86::DX: DestReg = X86::RDX; break;
24109 case X86::CX: DestReg = X86::RCX; break;
24110 case X86::BX: DestReg = X86::RBX; break;
24111 case X86::SI: DestReg = X86::RSI; break;
24112 case X86::DI: DestReg = X86::RDI; break;
24113 case X86::BP: DestReg = X86::RBP; break;
24114 case X86::SP: DestReg = X86::RSP; break;
24117 Res.first = DestReg;
24118 Res.second = &X86::GR64RegClass;
24121 } else if (Res.second == &X86::FR32RegClass ||
24122 Res.second == &X86::FR64RegClass ||
24123 Res.second == &X86::VR128RegClass ||
24124 Res.second == &X86::VR256RegClass ||
24125 Res.second == &X86::FR32XRegClass ||
24126 Res.second == &X86::FR64XRegClass ||
24127 Res.second == &X86::VR128XRegClass ||
24128 Res.second == &X86::VR256XRegClass ||
24129 Res.second == &X86::VR512RegClass) {
24130 // Handle references to XMM physical registers that got mapped into the
24131 // wrong class. This can happen with constraints like {xmm0} where the
24132 // target independent register mapper will just pick the first match it can
24133 // find, ignoring the required type.
24135 if (VT == MVT::f32 || VT == MVT::i32)
24136 Res.second = &X86::FR32RegClass;
24137 else if (VT == MVT::f64 || VT == MVT::i64)
24138 Res.second = &X86::FR64RegClass;
24139 else if (X86::VR128RegClass.hasType(VT))
24140 Res.second = &X86::VR128RegClass;
24141 else if (X86::VR256RegClass.hasType(VT))
24142 Res.second = &X86::VR256RegClass;
24143 else if (X86::VR512RegClass.hasType(VT))
24144 Res.second = &X86::VR512RegClass;
24150 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
24152 // Scaling factors are not free at all.
24153 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
24154 // will take 2 allocations in the out of order engine instead of 1
24155 // for plain addressing mode, i.e. inst (reg1).
24157 // vaddps (%rsi,%drx), %ymm0, %ymm1
24158 // Requires two allocations (one for the load, one for the computation)
24160 // vaddps (%rsi), %ymm0, %ymm1
24161 // Requires just 1 allocation, i.e., freeing allocations for other operations
24162 // and having less micro operations to execute.
24164 // For some X86 architectures, this is even worse because for instance for
24165 // stores, the complex addressing mode forces the instruction to use the
24166 // "load" ports instead of the dedicated "store" port.
24167 // E.g., on Haswell:
24168 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
24169 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
24170 if (isLegalAddressingMode(AM, Ty))
24171 // Scale represents reg2 * scale, thus account for 1
24172 // as soon as we use a second register.
24173 return AM.Scale != 0;
24177 bool X86TargetLowering::isTargetFTOL() const {
24178 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();