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 "X86InstrBuilder.h"
19 #include "X86MachineFunctionInfo.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/ADT/SmallBitVector.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/ADT/StringExtras.h"
26 #include "llvm/ADT/StringSwitch.h"
27 #include "llvm/ADT/VariadicFunction.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/Constants.h"
38 #include "llvm/IR/DerivedTypes.h"
39 #include "llvm/IR/Function.h"
40 #include "llvm/IR/GlobalAlias.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/Intrinsics.h"
44 #include "llvm/MC/MCAsmInfo.h"
45 #include "llvm/MC/MCContext.h"
46 #include "llvm/MC/MCExpr.h"
47 #include "llvm/MC/MCSymbol.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/ErrorHandling.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Target/TargetOptions.h"
53 #include "X86IntrinsicsInfo.h"
59 #define DEBUG_TYPE "x86-isel"
61 STATISTIC(NumTailCalls, "Number of tail calls");
63 static cl::opt<bool> ExperimentalVectorWideningLegalization(
64 "x86-experimental-vector-widening-legalization", cl::init(false),
65 cl::desc("Enable an experimental vector type legalization through widening "
66 "rather than promotion."),
69 static cl::opt<bool> ExperimentalVectorShuffleLowering(
70 "x86-experimental-vector-shuffle-lowering", cl::init(true),
71 cl::desc("Enable an experimental vector shuffle lowering code path."),
74 static cl::opt<bool> ExperimentalVectorShuffleLegality(
75 "x86-experimental-vector-shuffle-legality", cl::init(false),
76 cl::desc("Enable experimental shuffle legality based on the experimental "
77 "shuffle lowering. Should only be used with the experimental "
81 static cl::opt<int> ReciprocalEstimateRefinementSteps(
82 "x86-recip-refinement-steps", cl::init(1),
83 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
84 "result of the hardware reciprocal estimate instruction."),
87 // Forward declarations.
88 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
91 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
92 SelectionDAG &DAG, SDLoc dl,
93 unsigned vectorWidth) {
94 assert((vectorWidth == 128 || vectorWidth == 256) &&
95 "Unsupported vector width");
96 EVT VT = Vec.getValueType();
97 EVT ElVT = VT.getVectorElementType();
98 unsigned Factor = VT.getSizeInBits()/vectorWidth;
99 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
100 VT.getVectorNumElements()/Factor);
102 // Extract from UNDEF is UNDEF.
103 if (Vec.getOpcode() == ISD::UNDEF)
104 return DAG.getUNDEF(ResultVT);
106 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
107 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
109 // This is the index of the first element of the vectorWidth-bit chunk
111 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
114 // If the input is a buildvector just emit a smaller one.
115 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
116 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
117 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
120 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
121 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
124 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
125 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
126 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
127 /// instructions or a simple subregister reference. Idx is an index in the
128 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
129 /// lowering EXTRACT_VECTOR_ELT operations easier.
130 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
131 SelectionDAG &DAG, SDLoc dl) {
132 assert((Vec.getValueType().is256BitVector() ||
133 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
134 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
137 /// Generate a DAG to grab 256-bits from a 512-bit vector.
138 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
139 SelectionDAG &DAG, SDLoc dl) {
140 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
141 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
144 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
145 unsigned IdxVal, SelectionDAG &DAG,
146 SDLoc dl, unsigned vectorWidth) {
147 assert((vectorWidth == 128 || vectorWidth == 256) &&
148 "Unsupported vector width");
149 // Inserting UNDEF is Result
150 if (Vec.getOpcode() == ISD::UNDEF)
152 EVT VT = Vec.getValueType();
153 EVT ElVT = VT.getVectorElementType();
154 EVT ResultVT = Result.getValueType();
156 // Insert the relevant vectorWidth bits.
157 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
159 // This is the index of the first element of the vectorWidth-bit chunk
161 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
164 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
165 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
168 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
169 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
170 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
171 /// simple superregister reference. Idx is an index in the 128 bits
172 /// we want. It need not be aligned to a 128-bit boundary. That makes
173 /// lowering INSERT_VECTOR_ELT operations easier.
174 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
175 SelectionDAG &DAG,SDLoc dl) {
176 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
177 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
180 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
181 SelectionDAG &DAG, SDLoc dl) {
182 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
183 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
186 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
187 /// instructions. This is used because creating CONCAT_VECTOR nodes of
188 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
189 /// large BUILD_VECTORS.
190 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
191 unsigned NumElems, SelectionDAG &DAG,
193 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
194 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
197 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
198 unsigned NumElems, SelectionDAG &DAG,
200 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
201 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
204 // FIXME: This should stop caching the target machine as soon as
205 // we can remove resetOperationActions et al.
206 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM)
207 : TargetLowering(TM) {
208 Subtarget = &TM.getSubtarget<X86Subtarget>();
209 X86ScalarSSEf64 = Subtarget->hasSSE2();
210 X86ScalarSSEf32 = Subtarget->hasSSE1();
211 TD = getDataLayout();
213 resetOperationActions();
216 void X86TargetLowering::resetOperationActions() {
217 const TargetMachine &TM = getTargetMachine();
218 static bool FirstTimeThrough = true;
220 // If none of the target options have changed, then we don't need to reset the
221 // operation actions.
222 if (!FirstTimeThrough && TO == TM.Options) return;
224 if (!FirstTimeThrough) {
225 // Reinitialize the actions.
227 FirstTimeThrough = false;
232 // Set up the TargetLowering object.
233 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
235 // X86 is weird. It always uses i8 for shift amounts and setcc results.
236 setBooleanContents(ZeroOrOneBooleanContent);
237 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
238 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
240 // For 64-bit, since we have so many registers, use the ILP scheduler.
241 // For 32-bit, use the register pressure specific scheduling.
242 // For Atom, always use ILP scheduling.
243 if (Subtarget->isAtom())
244 setSchedulingPreference(Sched::ILP);
245 else if (Subtarget->is64Bit())
246 setSchedulingPreference(Sched::ILP);
248 setSchedulingPreference(Sched::RegPressure);
249 const X86RegisterInfo *RegInfo =
250 TM.getSubtarget<X86Subtarget>().getRegisterInfo();
251 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
253 // Bypass expensive divides on Atom when compiling with O2.
254 if (TM.getOptLevel() >= CodeGenOpt::Default) {
255 if (Subtarget->hasSlowDivide32())
256 addBypassSlowDiv(32, 8);
257 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
258 addBypassSlowDiv(64, 16);
261 if (Subtarget->isTargetKnownWindowsMSVC()) {
262 // Setup Windows compiler runtime calls.
263 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
264 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
265 setLibcallName(RTLIB::SREM_I64, "_allrem");
266 setLibcallName(RTLIB::UREM_I64, "_aullrem");
267 setLibcallName(RTLIB::MUL_I64, "_allmul");
268 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
269 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
270 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
271 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
272 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
274 // The _ftol2 runtime function has an unusual calling conv, which
275 // is modeled by a special pseudo-instruction.
276 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
277 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
278 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
279 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
282 if (Subtarget->isTargetDarwin()) {
283 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
284 setUseUnderscoreSetJmp(false);
285 setUseUnderscoreLongJmp(false);
286 } else if (Subtarget->isTargetWindowsGNU()) {
287 // MS runtime is weird: it exports _setjmp, but longjmp!
288 setUseUnderscoreSetJmp(true);
289 setUseUnderscoreLongJmp(false);
291 setUseUnderscoreSetJmp(true);
292 setUseUnderscoreLongJmp(true);
295 // Set up the register classes.
296 addRegisterClass(MVT::i8, &X86::GR8RegClass);
297 addRegisterClass(MVT::i16, &X86::GR16RegClass);
298 addRegisterClass(MVT::i32, &X86::GR32RegClass);
299 if (Subtarget->is64Bit())
300 addRegisterClass(MVT::i64, &X86::GR64RegClass);
302 for (MVT VT : MVT::integer_valuetypes())
303 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
305 // We don't accept any truncstore of integer registers.
306 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
307 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
308 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
309 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
310 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
311 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
313 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
315 // SETOEQ and SETUNE require checking two conditions.
316 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
317 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
318 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
319 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
320 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
321 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
323 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
325 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
326 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
327 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
329 if (Subtarget->is64Bit()) {
330 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
331 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
332 } else if (!TM.Options.UseSoftFloat) {
333 // We have an algorithm for SSE2->double, and we turn this into a
334 // 64-bit FILD followed by conditional FADD for other targets.
335 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
336 // We have an algorithm for SSE2, and we turn this into a 64-bit
337 // FILD for other targets.
338 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
341 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
343 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
344 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
346 if (!TM.Options.UseSoftFloat) {
347 // SSE has no i16 to fp conversion, only i32
348 if (X86ScalarSSEf32) {
349 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
350 // f32 and f64 cases are Legal, f80 case is not
351 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
353 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
354 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
357 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
358 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
361 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
362 // are Legal, f80 is custom lowered.
363 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
364 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
366 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
368 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
369 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
371 if (X86ScalarSSEf32) {
372 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
373 // f32 and f64 cases are Legal, f80 case is not
374 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
376 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
377 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
380 // Handle FP_TO_UINT by promoting the destination to a larger signed
382 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
383 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
384 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
386 if (Subtarget->is64Bit()) {
387 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
388 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
389 } else if (!TM.Options.UseSoftFloat) {
390 // Since AVX is a superset of SSE3, only check for SSE here.
391 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
392 // Expand FP_TO_UINT into a select.
393 // FIXME: We would like to use a Custom expander here eventually to do
394 // the optimal thing for SSE vs. the default expansion in the legalizer.
395 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
397 // With SSE3 we can use fisttpll to convert to a signed i64; without
398 // SSE, we're stuck with a fistpll.
399 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
402 if (isTargetFTOL()) {
403 // Use the _ftol2 runtime function, which has a pseudo-instruction
404 // to handle its weird calling convention.
405 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
408 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
409 if (!X86ScalarSSEf64) {
410 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
411 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
412 if (Subtarget->is64Bit()) {
413 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
414 // Without SSE, i64->f64 goes through memory.
415 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
419 // Scalar integer divide and remainder are lowered to use operations that
420 // produce two results, to match the available instructions. This exposes
421 // the two-result form to trivial CSE, which is able to combine x/y and x%y
422 // into a single instruction.
424 // Scalar integer multiply-high is also lowered to use two-result
425 // operations, to match the available instructions. However, plain multiply
426 // (low) operations are left as Legal, as there are single-result
427 // instructions for this in x86. Using the two-result multiply instructions
428 // when both high and low results are needed must be arranged by dagcombine.
429 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
431 setOperationAction(ISD::MULHS, VT, Expand);
432 setOperationAction(ISD::MULHU, VT, Expand);
433 setOperationAction(ISD::SDIV, VT, Expand);
434 setOperationAction(ISD::UDIV, VT, Expand);
435 setOperationAction(ISD::SREM, VT, Expand);
436 setOperationAction(ISD::UREM, VT, Expand);
438 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
439 setOperationAction(ISD::ADDC, VT, Custom);
440 setOperationAction(ISD::ADDE, VT, Custom);
441 setOperationAction(ISD::SUBC, VT, Custom);
442 setOperationAction(ISD::SUBE, VT, Custom);
445 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
446 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
447 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
448 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
449 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
450 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
451 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
452 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
453 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
454 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
455 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
456 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
457 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
458 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
459 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
460 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
461 if (Subtarget->is64Bit())
462 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
463 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
464 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
465 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
466 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
467 setOperationAction(ISD::FREM , MVT::f32 , Expand);
468 setOperationAction(ISD::FREM , MVT::f64 , Expand);
469 setOperationAction(ISD::FREM , MVT::f80 , Expand);
470 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
472 // Promote the i8 variants and force them on up to i32 which has a shorter
474 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
475 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
476 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
477 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
478 if (Subtarget->hasBMI()) {
479 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
480 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
481 if (Subtarget->is64Bit())
482 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
484 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
485 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
486 if (Subtarget->is64Bit())
487 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
490 if (Subtarget->hasLZCNT()) {
491 // When promoting the i8 variants, force them to i32 for a shorter
493 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
494 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
495 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
496 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
497 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
498 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
499 if (Subtarget->is64Bit())
500 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
502 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
503 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
504 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
505 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
506 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
507 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
508 if (Subtarget->is64Bit()) {
509 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
510 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
514 // Special handling for half-precision floating point conversions.
515 // If we don't have F16C support, then lower half float conversions
516 // into library calls.
517 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
518 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
519 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
522 // There's never any support for operations beyond MVT::f32.
523 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
524 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
525 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
526 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
528 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
529 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
530 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
531 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
532 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
533 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
535 if (Subtarget->hasPOPCNT()) {
536 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
538 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
539 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
540 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
541 if (Subtarget->is64Bit())
542 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
545 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
547 if (!Subtarget->hasMOVBE())
548 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
550 // These should be promoted to a larger select which is supported.
551 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
552 // X86 wants to expand cmov itself.
553 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
554 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
555 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
556 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
557 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
558 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
559 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
560 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
561 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
562 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
563 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
564 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
565 if (Subtarget->is64Bit()) {
566 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
567 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
569 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
570 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
571 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
572 // support continuation, user-level threading, and etc.. As a result, no
573 // other SjLj exception interfaces are implemented and please don't build
574 // your own exception handling based on them.
575 // LLVM/Clang supports zero-cost DWARF exception handling.
576 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
577 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
580 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
581 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
582 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
583 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
584 if (Subtarget->is64Bit())
585 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
586 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
587 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
588 if (Subtarget->is64Bit()) {
589 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
590 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
591 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
592 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
593 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
595 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
596 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
597 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
598 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
599 if (Subtarget->is64Bit()) {
600 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
601 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
602 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
605 if (Subtarget->hasSSE1())
606 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
608 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
610 // Expand certain atomics
611 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
613 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
614 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
615 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
618 if (Subtarget->hasCmpxchg16b()) {
619 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
622 // FIXME - use subtarget debug flags
623 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
624 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
625 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
628 if (Subtarget->is64Bit()) {
629 setExceptionPointerRegister(X86::RAX);
630 setExceptionSelectorRegister(X86::RDX);
632 setExceptionPointerRegister(X86::EAX);
633 setExceptionSelectorRegister(X86::EDX);
635 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
636 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
638 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
639 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
641 setOperationAction(ISD::TRAP, MVT::Other, Legal);
642 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
644 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
645 setOperationAction(ISD::VASTART , MVT::Other, Custom);
646 setOperationAction(ISD::VAEND , MVT::Other, Expand);
647 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
648 // TargetInfo::X86_64ABIBuiltinVaList
649 setOperationAction(ISD::VAARG , MVT::Other, Custom);
650 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
652 // TargetInfo::CharPtrBuiltinVaList
653 setOperationAction(ISD::VAARG , MVT::Other, Expand);
654 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
657 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
658 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
660 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
662 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
663 // f32 and f64 use SSE.
664 // Set up the FP register classes.
665 addRegisterClass(MVT::f32, &X86::FR32RegClass);
666 addRegisterClass(MVT::f64, &X86::FR64RegClass);
668 // Use ANDPD to simulate FABS.
669 setOperationAction(ISD::FABS , MVT::f64, Custom);
670 setOperationAction(ISD::FABS , MVT::f32, Custom);
672 // Use XORP to simulate FNEG.
673 setOperationAction(ISD::FNEG , MVT::f64, Custom);
674 setOperationAction(ISD::FNEG , MVT::f32, Custom);
676 // Use ANDPD and ORPD to simulate FCOPYSIGN.
677 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
678 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
680 // Lower this to FGETSIGNx86 plus an AND.
681 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
682 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
684 // We don't support sin/cos/fmod
685 setOperationAction(ISD::FSIN , MVT::f64, Expand);
686 setOperationAction(ISD::FCOS , MVT::f64, Expand);
687 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
688 setOperationAction(ISD::FSIN , MVT::f32, Expand);
689 setOperationAction(ISD::FCOS , MVT::f32, Expand);
690 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
692 // Expand FP immediates into loads from the stack, except for the special
694 addLegalFPImmediate(APFloat(+0.0)); // xorpd
695 addLegalFPImmediate(APFloat(+0.0f)); // xorps
696 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
697 // Use SSE for f32, x87 for f64.
698 // Set up the FP register classes.
699 addRegisterClass(MVT::f32, &X86::FR32RegClass);
700 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
702 // Use ANDPS to simulate FABS.
703 setOperationAction(ISD::FABS , MVT::f32, Custom);
705 // Use XORP to simulate FNEG.
706 setOperationAction(ISD::FNEG , MVT::f32, Custom);
708 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
710 // Use ANDPS and ORPS to simulate FCOPYSIGN.
711 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
712 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
714 // We don't support sin/cos/fmod
715 setOperationAction(ISD::FSIN , MVT::f32, Expand);
716 setOperationAction(ISD::FCOS , MVT::f32, Expand);
717 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
719 // Special cases we handle for FP constants.
720 addLegalFPImmediate(APFloat(+0.0f)); // xorps
721 addLegalFPImmediate(APFloat(+0.0)); // FLD0
722 addLegalFPImmediate(APFloat(+1.0)); // FLD1
723 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
724 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
726 if (!TM.Options.UnsafeFPMath) {
727 setOperationAction(ISD::FSIN , MVT::f64, Expand);
728 setOperationAction(ISD::FCOS , MVT::f64, Expand);
729 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
731 } else if (!TM.Options.UseSoftFloat) {
732 // f32 and f64 in x87.
733 // Set up the FP register classes.
734 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
735 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
737 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
738 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
739 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
740 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
742 if (!TM.Options.UnsafeFPMath) {
743 setOperationAction(ISD::FSIN , MVT::f64, Expand);
744 setOperationAction(ISD::FSIN , MVT::f32, Expand);
745 setOperationAction(ISD::FCOS , MVT::f64, Expand);
746 setOperationAction(ISD::FCOS , MVT::f32, Expand);
747 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
748 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
750 addLegalFPImmediate(APFloat(+0.0)); // FLD0
751 addLegalFPImmediate(APFloat(+1.0)); // FLD1
752 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
753 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
754 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
755 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
756 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
757 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
760 // We don't support FMA.
761 setOperationAction(ISD::FMA, MVT::f64, Expand);
762 setOperationAction(ISD::FMA, MVT::f32, Expand);
764 // Long double always uses X87.
765 if (!TM.Options.UseSoftFloat) {
766 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
767 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
768 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
770 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
771 addLegalFPImmediate(TmpFlt); // FLD0
773 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
776 APFloat TmpFlt2(+1.0);
777 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
779 addLegalFPImmediate(TmpFlt2); // FLD1
780 TmpFlt2.changeSign();
781 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
784 if (!TM.Options.UnsafeFPMath) {
785 setOperationAction(ISD::FSIN , MVT::f80, Expand);
786 setOperationAction(ISD::FCOS , MVT::f80, Expand);
787 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
790 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
791 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
792 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
793 setOperationAction(ISD::FRINT, MVT::f80, Expand);
794 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
795 setOperationAction(ISD::FMA, MVT::f80, Expand);
798 // Always use a library call for pow.
799 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
800 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
801 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
803 setOperationAction(ISD::FLOG, MVT::f80, Expand);
804 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
805 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
806 setOperationAction(ISD::FEXP, MVT::f80, Expand);
807 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
808 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
809 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
811 // First set operation action for all vector types to either promote
812 // (for widening) or expand (for scalarization). Then we will selectively
813 // turn on ones that can be effectively codegen'd.
814 for (MVT VT : MVT::vector_valuetypes()) {
815 setOperationAction(ISD::ADD , VT, Expand);
816 setOperationAction(ISD::SUB , VT, Expand);
817 setOperationAction(ISD::FADD, VT, Expand);
818 setOperationAction(ISD::FNEG, VT, Expand);
819 setOperationAction(ISD::FSUB, VT, Expand);
820 setOperationAction(ISD::MUL , VT, Expand);
821 setOperationAction(ISD::FMUL, VT, Expand);
822 setOperationAction(ISD::SDIV, VT, Expand);
823 setOperationAction(ISD::UDIV, VT, Expand);
824 setOperationAction(ISD::FDIV, VT, Expand);
825 setOperationAction(ISD::SREM, VT, Expand);
826 setOperationAction(ISD::UREM, VT, Expand);
827 setOperationAction(ISD::LOAD, VT, Expand);
828 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
829 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
830 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
831 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
832 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
833 setOperationAction(ISD::FABS, VT, Expand);
834 setOperationAction(ISD::FSIN, VT, Expand);
835 setOperationAction(ISD::FSINCOS, VT, Expand);
836 setOperationAction(ISD::FCOS, VT, Expand);
837 setOperationAction(ISD::FSINCOS, VT, Expand);
838 setOperationAction(ISD::FREM, VT, Expand);
839 setOperationAction(ISD::FMA, VT, Expand);
840 setOperationAction(ISD::FPOWI, VT, Expand);
841 setOperationAction(ISD::FSQRT, VT, Expand);
842 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
843 setOperationAction(ISD::FFLOOR, VT, Expand);
844 setOperationAction(ISD::FCEIL, VT, Expand);
845 setOperationAction(ISD::FTRUNC, VT, Expand);
846 setOperationAction(ISD::FRINT, VT, Expand);
847 setOperationAction(ISD::FNEARBYINT, VT, Expand);
848 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
849 setOperationAction(ISD::MULHS, VT, Expand);
850 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
851 setOperationAction(ISD::MULHU, VT, Expand);
852 setOperationAction(ISD::SDIVREM, VT, Expand);
853 setOperationAction(ISD::UDIVREM, VT, Expand);
854 setOperationAction(ISD::FPOW, VT, Expand);
855 setOperationAction(ISD::CTPOP, VT, Expand);
856 setOperationAction(ISD::CTTZ, VT, Expand);
857 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
858 setOperationAction(ISD::CTLZ, VT, Expand);
859 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
860 setOperationAction(ISD::SHL, VT, Expand);
861 setOperationAction(ISD::SRA, VT, Expand);
862 setOperationAction(ISD::SRL, VT, Expand);
863 setOperationAction(ISD::ROTL, VT, Expand);
864 setOperationAction(ISD::ROTR, VT, Expand);
865 setOperationAction(ISD::BSWAP, VT, Expand);
866 setOperationAction(ISD::SETCC, VT, Expand);
867 setOperationAction(ISD::FLOG, VT, Expand);
868 setOperationAction(ISD::FLOG2, VT, Expand);
869 setOperationAction(ISD::FLOG10, VT, Expand);
870 setOperationAction(ISD::FEXP, VT, Expand);
871 setOperationAction(ISD::FEXP2, VT, Expand);
872 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
873 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
874 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
875 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
876 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
877 setOperationAction(ISD::TRUNCATE, VT, Expand);
878 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
879 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
880 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
881 setOperationAction(ISD::VSELECT, VT, Expand);
882 setOperationAction(ISD::SELECT_CC, VT, Expand);
883 for (MVT InnerVT : MVT::vector_valuetypes()) {
884 setTruncStoreAction(InnerVT, VT, Expand);
886 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
887 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
889 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
890 // types, we have to deal with them whether we ask for Expansion or not.
891 // Setting Expand causes its own optimisation problems though, so leave
893 if (VT.getVectorElementType() == MVT::i1)
894 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
898 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
899 // with -msoft-float, disable use of MMX as well.
900 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
901 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
902 // No operations on x86mmx supported, everything uses intrinsics.
905 // MMX-sized vectors (other than x86mmx) are expected to be expanded
906 // into smaller operations.
907 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
908 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
909 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
910 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
911 setOperationAction(ISD::AND, MVT::v8i8, Expand);
912 setOperationAction(ISD::AND, MVT::v4i16, Expand);
913 setOperationAction(ISD::AND, MVT::v2i32, Expand);
914 setOperationAction(ISD::AND, MVT::v1i64, Expand);
915 setOperationAction(ISD::OR, MVT::v8i8, Expand);
916 setOperationAction(ISD::OR, MVT::v4i16, Expand);
917 setOperationAction(ISD::OR, MVT::v2i32, Expand);
918 setOperationAction(ISD::OR, MVT::v1i64, Expand);
919 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
920 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
921 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
922 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
923 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
924 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
925 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
926 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
927 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
928 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
929 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
930 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
931 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
932 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
933 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
934 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
935 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
937 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
938 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
940 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
941 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
942 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
943 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
944 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
945 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
946 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
947 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
948 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
949 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
950 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
951 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
952 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
955 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
956 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
958 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
959 // registers cannot be used even for integer operations.
960 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
961 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
962 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
963 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
965 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
966 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
967 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
968 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
969 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
970 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
971 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
972 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
973 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
974 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
975 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
976 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
977 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
978 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
979 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
980 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
981 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
982 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
983 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
984 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
985 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
986 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
988 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
989 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
990 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
991 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
993 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
994 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
995 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
996 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
997 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
999 // Only provide customized ctpop vector bit twiddling for vector types we
1000 // know to perform better than using the popcnt instructions on each vector
1001 // element. If popcnt isn't supported, always provide the custom version.
1002 if (!Subtarget->hasPOPCNT()) {
1003 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
1004 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
1007 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
1008 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1009 MVT VT = (MVT::SimpleValueType)i;
1010 // Do not attempt to custom lower non-power-of-2 vectors
1011 if (!isPowerOf2_32(VT.getVectorNumElements()))
1013 // Do not attempt to custom lower non-128-bit vectors
1014 if (!VT.is128BitVector())
1016 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1017 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1018 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1021 // We support custom legalizing of sext and anyext loads for specific
1022 // memory vector types which we can load as a scalar (or sequence of
1023 // scalars) and extend in-register to a legal 128-bit vector type. For sext
1024 // loads these must work with a single scalar load.
1025 for (MVT VT : MVT::integer_vector_valuetypes()) {
1026 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
1027 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
1028 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
1029 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
1030 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
1031 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
1032 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
1033 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
1034 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
1037 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1038 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1039 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1040 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1041 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1042 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1044 if (Subtarget->is64Bit()) {
1045 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1046 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1049 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1050 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1051 MVT VT = (MVT::SimpleValueType)i;
1053 // Do not attempt to promote non-128-bit vectors
1054 if (!VT.is128BitVector())
1057 setOperationAction(ISD::AND, VT, Promote);
1058 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1059 setOperationAction(ISD::OR, VT, Promote);
1060 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1061 setOperationAction(ISD::XOR, VT, Promote);
1062 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1063 setOperationAction(ISD::LOAD, VT, Promote);
1064 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1065 setOperationAction(ISD::SELECT, VT, Promote);
1066 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1069 // Custom lower v2i64 and v2f64 selects.
1070 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1071 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1072 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1073 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1075 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1076 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1078 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1079 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1080 // As there is no 64-bit GPR available, we need build a special custom
1081 // sequence to convert from v2i32 to v2f32.
1082 if (!Subtarget->is64Bit())
1083 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1085 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1086 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1088 for (MVT VT : MVT::fp_vector_valuetypes())
1089 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
1091 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1092 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1093 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1096 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1097 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1098 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1099 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1100 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1101 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1102 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1103 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1104 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1105 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1106 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1108 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1109 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1110 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1111 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1112 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1113 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1114 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1115 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1116 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1117 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1119 // FIXME: Do we need to handle scalar-to-vector here?
1120 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1122 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1123 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1124 setOperationAction(ISD::VSELECT, MVT::v4i32, Custom);
1125 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
1126 setOperationAction(ISD::VSELECT, MVT::v8i16, Custom);
1127 // There is no BLENDI for byte vectors. We don't need to custom lower
1128 // some vselects for now.
1129 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1131 // SSE41 brings specific instructions for doing vector sign extend even in
1132 // cases where we don't have SRA.
1133 for (MVT VT : MVT::integer_vector_valuetypes()) {
1134 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
1135 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
1136 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
1139 // i8 and i16 vectors are custom because the source register and source
1140 // source memory operand types are not the same width. f32 vectors are
1141 // custom since the immediate controlling the insert encodes additional
1143 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1144 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1145 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1146 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1148 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1149 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1150 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1151 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1153 // FIXME: these should be Legal, but that's only for the case where
1154 // the index is constant. For now custom expand to deal with that.
1155 if (Subtarget->is64Bit()) {
1156 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1157 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1161 if (Subtarget->hasSSE2()) {
1162 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1163 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1165 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1166 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1168 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1169 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1171 // In the customized shift lowering, the legal cases in AVX2 will be
1173 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1174 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1176 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1177 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1179 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1182 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1183 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1184 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1185 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1186 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1187 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1188 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1190 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1191 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1192 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1194 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1195 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1196 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1197 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1198 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1199 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1200 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1201 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1202 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1203 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1204 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1205 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1207 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1208 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1209 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1210 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1211 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1212 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1213 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1214 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1215 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1216 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1217 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1218 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1220 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1221 // even though v8i16 is a legal type.
1222 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1223 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1224 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1226 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1227 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1228 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1230 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1231 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1233 for (MVT VT : MVT::fp_vector_valuetypes())
1234 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1236 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1237 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1239 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1240 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1242 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1243 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1245 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1246 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1247 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1248 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1250 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1251 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1252 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1254 setOperationAction(ISD::VSELECT, MVT::v4f64, Custom);
1255 setOperationAction(ISD::VSELECT, MVT::v4i64, Custom);
1256 setOperationAction(ISD::VSELECT, MVT::v8i32, Custom);
1257 setOperationAction(ISD::VSELECT, MVT::v8f32, Custom);
1259 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1260 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1261 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1262 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1263 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1264 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1265 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1266 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1267 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1268 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1269 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1270 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1272 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1273 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1274 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1275 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1276 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1277 setOperationAction(ISD::FMA, MVT::f32, Legal);
1278 setOperationAction(ISD::FMA, MVT::f64, Legal);
1281 if (Subtarget->hasInt256()) {
1282 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1283 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1284 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1285 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1287 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1288 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1289 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1290 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1292 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1293 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1294 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1295 // Don't lower v32i8 because there is no 128-bit byte mul
1297 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1298 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1299 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1300 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1302 setOperationAction(ISD::VSELECT, MVT::v16i16, Custom);
1303 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1305 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1306 // when we have a 256bit-wide blend with immediate.
1307 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1309 // Only provide customized ctpop vector bit twiddling for vector types we
1310 // know to perform better than using the popcnt instructions on each
1311 // vector element. If popcnt isn't supported, always provide the custom
1313 if (!Subtarget->hasPOPCNT())
1314 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1316 // Custom CTPOP always performs better on natively supported v8i32
1317 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1319 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1320 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1321 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1322 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1324 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1325 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1326 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1327 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1329 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1330 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1331 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1332 // Don't lower v32i8 because there is no 128-bit byte mul
1335 // In the customized shift lowering, the legal cases in AVX2 will be
1337 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1338 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1340 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1341 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1343 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1345 // Custom lower several nodes for 256-bit types.
1346 for (MVT VT : MVT::vector_valuetypes()) {
1347 if (VT.getScalarSizeInBits() >= 32) {
1348 setOperationAction(ISD::MLOAD, VT, Legal);
1349 setOperationAction(ISD::MSTORE, VT, Legal);
1351 // Extract subvector is special because the value type
1352 // (result) is 128-bit but the source is 256-bit wide.
1353 if (VT.is128BitVector()) {
1354 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1356 // Do not attempt to custom lower other non-256-bit vectors
1357 if (!VT.is256BitVector())
1360 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1361 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1362 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1363 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1364 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1365 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1366 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1369 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1370 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1371 MVT VT = (MVT::SimpleValueType)i;
1373 // Do not attempt to promote non-256-bit vectors
1374 if (!VT.is256BitVector())
1377 setOperationAction(ISD::AND, VT, Promote);
1378 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1379 setOperationAction(ISD::OR, VT, Promote);
1380 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1381 setOperationAction(ISD::XOR, VT, Promote);
1382 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1383 setOperationAction(ISD::LOAD, VT, Promote);
1384 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1385 setOperationAction(ISD::SELECT, VT, Promote);
1386 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1390 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1391 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1392 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1393 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1394 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1396 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1397 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1398 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1400 for (MVT VT : MVT::fp_vector_valuetypes())
1401 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1403 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1404 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1405 setOperationAction(ISD::XOR, MVT::i1, Legal);
1406 setOperationAction(ISD::OR, MVT::i1, Legal);
1407 setOperationAction(ISD::AND, MVT::i1, Legal);
1408 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1409 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1410 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1411 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1412 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1414 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1415 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1416 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1417 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1418 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1419 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1421 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1422 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1423 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1424 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1425 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1426 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1427 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1428 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1430 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1431 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1432 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1433 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1434 if (Subtarget->is64Bit()) {
1435 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1436 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1438 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1440 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1441 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1442 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1443 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1444 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1445 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1446 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1447 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1448 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1449 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1450 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1451 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1452 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1453 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1455 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1456 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1457 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1458 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1459 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1460 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1461 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1462 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1463 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1464 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1465 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1466 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1467 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1472 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1474 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1476 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1477 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1479 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1481 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1482 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1483 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1484 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1485 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1486 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1489 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1491 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1492 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1494 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1495 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1497 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1499 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1500 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1502 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1503 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1505 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1506 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1508 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1509 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1510 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1511 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1512 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1513 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1515 if (Subtarget->hasCDI()) {
1516 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1517 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1520 // Custom lower several nodes.
1521 for (MVT VT : MVT::vector_valuetypes()) {
1522 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1523 // Extract subvector is special because the value type
1524 // (result) is 256/128-bit but the source is 512-bit wide.
1525 if (VT.is128BitVector() || VT.is256BitVector()) {
1526 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1528 if (VT.getVectorElementType() == MVT::i1)
1529 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1531 // Do not attempt to custom lower other non-512-bit vectors
1532 if (!VT.is512BitVector())
1535 if ( EltSize >= 32) {
1536 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1537 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1538 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1539 setOperationAction(ISD::VSELECT, VT, Legal);
1540 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1541 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1542 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1543 setOperationAction(ISD::MLOAD, VT, Legal);
1544 setOperationAction(ISD::MSTORE, VT, Legal);
1547 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1548 MVT VT = (MVT::SimpleValueType)i;
1550 // Do not attempt to promote non-512-bit vectors.
1551 if (!VT.is512BitVector())
1554 setOperationAction(ISD::SELECT, VT, Promote);
1555 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1559 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1560 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1561 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1563 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1564 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1566 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1567 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1568 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1569 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1570 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1571 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1572 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1573 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1574 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1576 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1577 const MVT VT = (MVT::SimpleValueType)i;
1579 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1581 // Do not attempt to promote non-512-bit vectors.
1582 if (!VT.is512BitVector())
1586 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1587 setOperationAction(ISD::VSELECT, VT, Legal);
1592 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1593 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1594 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1596 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1597 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1598 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal);
1600 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1601 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1602 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1603 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1604 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1605 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1608 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1609 // of this type with custom code.
1610 for (MVT VT : MVT::vector_valuetypes())
1611 setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Custom);
1613 // We want to custom lower some of our intrinsics.
1614 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1615 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1616 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1617 if (!Subtarget->is64Bit())
1618 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1620 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1621 // handle type legalization for these operations here.
1623 // FIXME: We really should do custom legalization for addition and
1624 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1625 // than generic legalization for 64-bit multiplication-with-overflow, though.
1626 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1627 // Add/Sub/Mul with overflow operations are custom lowered.
1629 setOperationAction(ISD::SADDO, VT, Custom);
1630 setOperationAction(ISD::UADDO, VT, Custom);
1631 setOperationAction(ISD::SSUBO, VT, Custom);
1632 setOperationAction(ISD::USUBO, VT, Custom);
1633 setOperationAction(ISD::SMULO, VT, Custom);
1634 setOperationAction(ISD::UMULO, VT, Custom);
1638 if (!Subtarget->is64Bit()) {
1639 // These libcalls are not available in 32-bit.
1640 setLibcallName(RTLIB::SHL_I128, nullptr);
1641 setLibcallName(RTLIB::SRL_I128, nullptr);
1642 setLibcallName(RTLIB::SRA_I128, nullptr);
1645 // Combine sin / cos into one node or libcall if possible.
1646 if (Subtarget->hasSinCos()) {
1647 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1648 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1649 if (Subtarget->isTargetDarwin()) {
1650 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1651 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1652 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1653 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1657 if (Subtarget->isTargetWin64()) {
1658 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1659 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1660 setOperationAction(ISD::SREM, MVT::i128, Custom);
1661 setOperationAction(ISD::UREM, MVT::i128, Custom);
1662 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1663 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1666 // We have target-specific dag combine patterns for the following nodes:
1667 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1668 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
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::STORE);
1683 setTargetDAGCombine(ISD::ZERO_EXTEND);
1684 setTargetDAGCombine(ISD::ANY_EXTEND);
1685 setTargetDAGCombine(ISD::SIGN_EXTEND);
1686 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1687 setTargetDAGCombine(ISD::TRUNCATE);
1688 setTargetDAGCombine(ISD::SINT_TO_FP);
1689 setTargetDAGCombine(ISD::SETCC);
1690 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1691 setTargetDAGCombine(ISD::BUILD_VECTOR);
1692 if (Subtarget->is64Bit())
1693 setTargetDAGCombine(ISD::MUL);
1694 setTargetDAGCombine(ISD::XOR);
1696 computeRegisterProperties();
1698 // On Darwin, -Os means optimize for size without hurting performance,
1699 // do not reduce the limit.
1700 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1701 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1702 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1703 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1704 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1705 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1706 setPrefLoopAlignment(4); // 2^4 bytes.
1708 // Predictable cmov don't hurt on atom because it's in-order.
1709 PredictableSelectIsExpensive = !Subtarget->isAtom();
1710 EnableExtLdPromotion = true;
1711 setPrefFunctionAlignment(4); // 2^4 bytes.
1713 verifyIntrinsicTables();
1716 // This has so far only been implemented for 64-bit MachO.
1717 bool X86TargetLowering::useLoadStackGuardNode() const {
1718 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1721 TargetLoweringBase::LegalizeTypeAction
1722 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1723 if (ExperimentalVectorWideningLegalization &&
1724 VT.getVectorNumElements() != 1 &&
1725 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1726 return TypeWidenVector;
1728 return TargetLoweringBase::getPreferredVectorAction(VT);
1731 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1733 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1735 const unsigned NumElts = VT.getVectorNumElements();
1736 const EVT EltVT = VT.getVectorElementType();
1737 if (VT.is512BitVector()) {
1738 if (Subtarget->hasAVX512())
1739 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1740 EltVT == MVT::f32 || EltVT == MVT::f64)
1742 case 8: return MVT::v8i1;
1743 case 16: return MVT::v16i1;
1745 if (Subtarget->hasBWI())
1746 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1748 case 32: return MVT::v32i1;
1749 case 64: return MVT::v64i1;
1753 if (VT.is256BitVector() || VT.is128BitVector()) {
1754 if (Subtarget->hasVLX())
1755 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1756 EltVT == MVT::f32 || EltVT == MVT::f64)
1758 case 2: return MVT::v2i1;
1759 case 4: return MVT::v4i1;
1760 case 8: return MVT::v8i1;
1762 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1763 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1765 case 8: return MVT::v8i1;
1766 case 16: return MVT::v16i1;
1767 case 32: return MVT::v32i1;
1771 return VT.changeVectorElementTypeToInteger();
1774 /// Helper for getByValTypeAlignment to determine
1775 /// the desired ByVal argument alignment.
1776 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1779 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1780 if (VTy->getBitWidth() == 128)
1782 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1783 unsigned EltAlign = 0;
1784 getMaxByValAlign(ATy->getElementType(), EltAlign);
1785 if (EltAlign > MaxAlign)
1786 MaxAlign = EltAlign;
1787 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1788 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1789 unsigned EltAlign = 0;
1790 getMaxByValAlign(STy->getElementType(i), EltAlign);
1791 if (EltAlign > MaxAlign)
1792 MaxAlign = EltAlign;
1799 /// Return the desired alignment for ByVal aggregate
1800 /// function arguments in the caller parameter area. For X86, aggregates
1801 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1802 /// are at 4-byte boundaries.
1803 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1804 if (Subtarget->is64Bit()) {
1805 // Max of 8 and alignment of type.
1806 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1813 if (Subtarget->hasSSE1())
1814 getMaxByValAlign(Ty, Align);
1818 /// Returns the target specific optimal type for load
1819 /// and store operations as a result of memset, memcpy, and memmove
1820 /// lowering. If DstAlign is zero that means it's safe to destination
1821 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1822 /// means there isn't a need to check it against alignment requirement,
1823 /// probably because the source does not need to be loaded. If 'IsMemset' is
1824 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1825 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1826 /// source is constant so it does not need to be loaded.
1827 /// It returns EVT::Other if the type should be determined using generic
1828 /// target-independent logic.
1830 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1831 unsigned DstAlign, unsigned SrcAlign,
1832 bool IsMemset, bool ZeroMemset,
1834 MachineFunction &MF) const {
1835 const Function *F = MF.getFunction();
1836 if ((!IsMemset || ZeroMemset) &&
1837 !F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
1838 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 // FIXME: Why this routine is here? Move to RegInfo!
1934 std::pair<const TargetRegisterClass*, uint8_t>
1935 X86TargetLowering::findRepresentativeClass(MVT VT) const{
1936 const TargetRegisterClass *RRC = nullptr;
1938 switch (VT.SimpleTy) {
1940 return TargetLowering::findRepresentativeClass(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.
2099 if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
2100 (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
2101 MachineFunction &MF = DAG.getMachineFunction();
2102 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2103 unsigned Reg = FuncInfo->getSRetReturnReg();
2105 "SRetReturnReg should have been set in LowerFormalArguments().");
2106 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
2109 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2110 X86::RAX : X86::EAX;
2111 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2112 Flag = Chain.getValue(1);
2114 // RAX/EAX now acts like a return value.
2115 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2118 RetOps[0] = Chain; // Update chain.
2120 // Add the flag if we have it.
2122 RetOps.push_back(Flag);
2124 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2127 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2128 if (N->getNumValues() != 1)
2130 if (!N->hasNUsesOfValue(1, 0))
2133 SDValue TCChain = Chain;
2134 SDNode *Copy = *N->use_begin();
2135 if (Copy->getOpcode() == ISD::CopyToReg) {
2136 // If the copy has a glue operand, we conservatively assume it isn't safe to
2137 // perform a tail call.
2138 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2140 TCChain = Copy->getOperand(0);
2141 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2144 bool HasRet = false;
2145 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2147 if (UI->getOpcode() != X86ISD::RET_FLAG)
2149 // If we are returning more than one value, we can definitely
2150 // not make a tail call see PR19530
2151 if (UI->getNumOperands() > 4)
2153 if (UI->getNumOperands() == 4 &&
2154 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2167 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2168 ISD::NodeType ExtendKind) const {
2170 // TODO: Is this also valid on 32-bit?
2171 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2172 ReturnMVT = MVT::i8;
2174 ReturnMVT = MVT::i32;
2176 EVT MinVT = getRegisterType(Context, ReturnMVT);
2177 return VT.bitsLT(MinVT) ? MinVT : VT;
2180 /// Lower the result values of a call into the
2181 /// appropriate copies out of appropriate physical registers.
2184 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2185 CallingConv::ID CallConv, bool isVarArg,
2186 const SmallVectorImpl<ISD::InputArg> &Ins,
2187 SDLoc dl, SelectionDAG &DAG,
2188 SmallVectorImpl<SDValue> &InVals) const {
2190 // Assign locations to each value returned by this call.
2191 SmallVector<CCValAssign, 16> RVLocs;
2192 bool Is64Bit = Subtarget->is64Bit();
2193 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2195 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2197 // Copy all of the result registers out of their specified physreg.
2198 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2199 CCValAssign &VA = RVLocs[i];
2200 EVT CopyVT = VA.getValVT();
2202 // If this is x86-64, and we disabled SSE, we can't return FP values
2203 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2204 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2205 report_fatal_error("SSE register return with SSE disabled");
2208 // If we prefer to use the value in xmm registers, copy it out as f80 and
2209 // use a truncate to move it from fp stack reg to xmm reg.
2210 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2211 isScalarFPTypeInSSEReg(VA.getValVT()))
2214 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2215 CopyVT, InFlag).getValue(1);
2216 SDValue Val = Chain.getValue(0);
2218 if (CopyVT != VA.getValVT())
2219 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2220 // This truncation won't change the value.
2221 DAG.getIntPtrConstant(1));
2223 InFlag = Chain.getValue(2);
2224 InVals.push_back(Val);
2230 //===----------------------------------------------------------------------===//
2231 // C & StdCall & Fast Calling Convention implementation
2232 //===----------------------------------------------------------------------===//
2233 // StdCall calling convention seems to be standard for many Windows' API
2234 // routines and around. It differs from C calling convention just a little:
2235 // callee should clean up the stack, not caller. Symbols should be also
2236 // decorated in some fancy way :) It doesn't support any vector arguments.
2237 // For info on fast calling convention see Fast Calling Convention (tail call)
2238 // implementation LowerX86_32FastCCCallTo.
2240 /// CallIsStructReturn - Determines whether a call uses struct return
2242 enum StructReturnType {
2247 static StructReturnType
2248 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2250 return NotStructReturn;
2252 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2253 if (!Flags.isSRet())
2254 return NotStructReturn;
2255 if (Flags.isInReg())
2256 return RegStructReturn;
2257 return StackStructReturn;
2260 /// Determines whether a function uses struct return semantics.
2261 static StructReturnType
2262 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2264 return NotStructReturn;
2266 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2267 if (!Flags.isSRet())
2268 return NotStructReturn;
2269 if (Flags.isInReg())
2270 return RegStructReturn;
2271 return StackStructReturn;
2274 /// Make a copy of an aggregate at address specified by "Src" to address
2275 /// "Dst" with size and alignment information specified by the specific
2276 /// parameter attribute. The copy will be passed as a byval function parameter.
2278 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2279 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2281 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2283 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2284 /*isVolatile*/false, /*AlwaysInline=*/true,
2285 MachinePointerInfo(), MachinePointerInfo());
2288 /// Return true if the calling convention is one that
2289 /// supports tail call optimization.
2290 static bool IsTailCallConvention(CallingConv::ID CC) {
2291 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2292 CC == CallingConv::HiPE);
2295 /// \brief Return true if the calling convention is a C calling convention.
2296 static bool IsCCallConvention(CallingConv::ID CC) {
2297 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2298 CC == CallingConv::X86_64_SysV);
2301 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2302 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2306 CallingConv::ID CalleeCC = CS.getCallingConv();
2307 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2313 /// Return true if the function is being made into
2314 /// a tailcall target by changing its ABI.
2315 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2316 bool GuaranteedTailCallOpt) {
2317 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2321 X86TargetLowering::LowerMemArgument(SDValue Chain,
2322 CallingConv::ID CallConv,
2323 const SmallVectorImpl<ISD::InputArg> &Ins,
2324 SDLoc dl, SelectionDAG &DAG,
2325 const CCValAssign &VA,
2326 MachineFrameInfo *MFI,
2328 // Create the nodes corresponding to a load from this parameter slot.
2329 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2330 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2331 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2332 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2335 // If value is passed by pointer we have address passed instead of the value
2337 if (VA.getLocInfo() == CCValAssign::Indirect)
2338 ValVT = VA.getLocVT();
2340 ValVT = VA.getValVT();
2342 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2343 // changed with more analysis.
2344 // In case of tail call optimization mark all arguments mutable. Since they
2345 // could be overwritten by lowering of arguments in case of a tail call.
2346 if (Flags.isByVal()) {
2347 unsigned Bytes = Flags.getByValSize();
2348 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2349 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2350 return DAG.getFrameIndex(FI, getPointerTy());
2352 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2353 VA.getLocMemOffset(), isImmutable);
2354 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2355 return DAG.getLoad(ValVT, dl, Chain, FIN,
2356 MachinePointerInfo::getFixedStack(FI),
2357 false, false, false, 0);
2361 // FIXME: Get this from tablegen.
2362 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2363 const X86Subtarget *Subtarget) {
2364 assert(Subtarget->is64Bit());
2366 if (Subtarget->isCallingConvWin64(CallConv)) {
2367 static const MCPhysReg GPR64ArgRegsWin64[] = {
2368 X86::RCX, X86::RDX, X86::R8, X86::R9
2370 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2373 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2374 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2376 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2379 // FIXME: Get this from tablegen.
2380 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2381 CallingConv::ID CallConv,
2382 const X86Subtarget *Subtarget) {
2383 assert(Subtarget->is64Bit());
2384 if (Subtarget->isCallingConvWin64(CallConv)) {
2385 // The XMM registers which might contain var arg parameters are shadowed
2386 // in their paired GPR. So we only need to save the GPR to their home
2388 // TODO: __vectorcall will change this.
2392 const Function *Fn = MF.getFunction();
2393 bool NoImplicitFloatOps = Fn->getAttributes().
2394 hasAttribute(AttributeSet::FunctionIndex, 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->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
2563 Attribute::NoImplicitFloat)) &&
2564 "SSE register cannot be used when SSE is disabled!");
2566 // 64-bit calling conventions support varargs and register parameters, so we
2567 // have to do extra work to spill them in the prologue.
2568 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2569 // Find the first unallocated argument registers.
2570 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2571 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2572 unsigned NumIntRegs =
2573 CCInfo.getFirstUnallocated(ArgGPRs.data(), ArgGPRs.size());
2574 unsigned NumXMMRegs =
2575 CCInfo.getFirstUnallocated(ArgXMMs.data(), ArgXMMs.size());
2576 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2577 "SSE register cannot be used when SSE is disabled!");
2579 // Gather all the live in physical registers.
2580 SmallVector<SDValue, 6> LiveGPRs;
2581 SmallVector<SDValue, 8> LiveXMMRegs;
2583 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2584 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2586 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2588 if (!ArgXMMs.empty()) {
2589 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2590 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2591 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2592 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2593 LiveXMMRegs.push_back(
2594 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2599 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
2600 // Get to the caller-allocated home save location. Add 8 to account
2601 // for the return address.
2602 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2603 FuncInfo->setRegSaveFrameIndex(
2604 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2605 // Fixup to set vararg frame on shadow area (4 x i64).
2607 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2609 // For X86-64, if there are vararg parameters that are passed via
2610 // registers, then we must store them to their spots on the stack so
2611 // they may be loaded by deferencing the result of va_next.
2612 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2613 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2614 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2615 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2618 // Store the integer parameter registers.
2619 SmallVector<SDValue, 8> MemOps;
2620 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2622 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2623 for (SDValue Val : LiveGPRs) {
2624 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2625 DAG.getIntPtrConstant(Offset));
2627 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2628 MachinePointerInfo::getFixedStack(
2629 FuncInfo->getRegSaveFrameIndex(), Offset),
2631 MemOps.push_back(Store);
2635 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2636 // Now store the XMM (fp + vector) parameter registers.
2637 SmallVector<SDValue, 12> SaveXMMOps;
2638 SaveXMMOps.push_back(Chain);
2639 SaveXMMOps.push_back(ALVal);
2640 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2641 FuncInfo->getRegSaveFrameIndex()));
2642 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2643 FuncInfo->getVarArgsFPOffset()));
2644 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2646 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2647 MVT::Other, SaveXMMOps));
2650 if (!MemOps.empty())
2651 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2654 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2655 // Find the largest legal vector type.
2656 MVT VecVT = MVT::Other;
2657 // FIXME: Only some x86_32 calling conventions support AVX512.
2658 if (Subtarget->hasAVX512() &&
2659 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2660 CallConv == CallingConv::Intel_OCL_BI)))
2661 VecVT = MVT::v16f32;
2662 else if (Subtarget->hasAVX())
2664 else if (Subtarget->hasSSE2())
2667 // We forward some GPRs and some vector types.
2668 SmallVector<MVT, 2> RegParmTypes;
2669 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2670 RegParmTypes.push_back(IntVT);
2671 if (VecVT != MVT::Other)
2672 RegParmTypes.push_back(VecVT);
2674 // Compute the set of forwarded registers. The rest are scratch.
2675 SmallVectorImpl<ForwardedRegister> &Forwards =
2676 FuncInfo->getForwardedMustTailRegParms();
2677 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2679 // Conservatively forward AL on x86_64, since it might be used for varargs.
2680 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2681 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2682 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2685 // Copy all forwards from physical to virtual registers.
2686 for (ForwardedRegister &F : Forwards) {
2687 // FIXME: Can we use a less constrained schedule?
2688 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2689 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2690 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2694 // Some CCs need callee pop.
2695 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2696 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2697 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2699 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2700 // If this is an sret function, the return should pop the hidden pointer.
2701 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2702 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2703 argsAreStructReturn(Ins) == StackStructReturn)
2704 FuncInfo->setBytesToPopOnReturn(4);
2708 // RegSaveFrameIndex is X86-64 only.
2709 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2710 if (CallConv == CallingConv::X86_FastCall ||
2711 CallConv == CallingConv::X86_ThisCall)
2712 // fastcc functions can't have varargs.
2713 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2716 FuncInfo->setArgumentStackSize(StackSize);
2722 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2723 SDValue StackPtr, SDValue Arg,
2724 SDLoc dl, SelectionDAG &DAG,
2725 const CCValAssign &VA,
2726 ISD::ArgFlagsTy Flags) const {
2727 unsigned LocMemOffset = VA.getLocMemOffset();
2728 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2729 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2730 if (Flags.isByVal())
2731 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2733 return DAG.getStore(Chain, dl, Arg, PtrOff,
2734 MachinePointerInfo::getStack(LocMemOffset),
2738 /// Emit a load of return address if tail call
2739 /// optimization is performed and it is required.
2741 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2742 SDValue &OutRetAddr, SDValue Chain,
2743 bool IsTailCall, bool Is64Bit,
2744 int FPDiff, SDLoc dl) const {
2745 // Adjust the Return address stack slot.
2746 EVT VT = getPointerTy();
2747 OutRetAddr = getReturnAddressFrameIndex(DAG);
2749 // Load the "old" Return address.
2750 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2751 false, false, false, 0);
2752 return SDValue(OutRetAddr.getNode(), 1);
2755 /// Emit a store of the return address if tail call
2756 /// optimization is performed and it is required (FPDiff!=0).
2757 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2758 SDValue Chain, SDValue RetAddrFrIdx,
2759 EVT PtrVT, unsigned SlotSize,
2760 int FPDiff, SDLoc dl) {
2761 // Store the return address to the appropriate stack slot.
2762 if (!FPDiff) return Chain;
2763 // Calculate the new stack slot for the return address.
2764 int NewReturnAddrFI =
2765 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2767 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2768 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2769 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2775 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2776 SmallVectorImpl<SDValue> &InVals) const {
2777 SelectionDAG &DAG = CLI.DAG;
2779 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2780 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2781 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2782 SDValue Chain = CLI.Chain;
2783 SDValue Callee = CLI.Callee;
2784 CallingConv::ID CallConv = CLI.CallConv;
2785 bool &isTailCall = CLI.IsTailCall;
2786 bool isVarArg = CLI.IsVarArg;
2788 MachineFunction &MF = DAG.getMachineFunction();
2789 bool Is64Bit = Subtarget->is64Bit();
2790 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2791 StructReturnType SR = callIsStructReturn(Outs);
2792 bool IsSibcall = false;
2793 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2795 if (MF.getTarget().Options.DisableTailCalls)
2798 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2800 // Force this to be a tail call. The verifier rules are enough to ensure
2801 // that we can lower this successfully without moving the return address
2804 } else if (isTailCall) {
2805 // Check if it's really possible to do a tail call.
2806 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2807 isVarArg, SR != NotStructReturn,
2808 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2809 Outs, OutVals, Ins, DAG);
2811 // Sibcalls are automatically detected tailcalls which do not require
2813 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2820 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2821 "Var args not supported with calling convention fastcc, ghc or hipe");
2823 // Analyze operands of the call, assigning locations to each operand.
2824 SmallVector<CCValAssign, 16> ArgLocs;
2825 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2827 // Allocate shadow area for Win64
2829 CCInfo.AllocateStack(32, 8);
2831 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2833 // Get a count of how many bytes are to be pushed on the stack.
2834 unsigned NumBytes = CCInfo.getNextStackOffset();
2836 // This is a sibcall. The memory operands are available in caller's
2837 // own caller's stack.
2839 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2840 IsTailCallConvention(CallConv))
2841 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2844 if (isTailCall && !IsSibcall && !IsMustTail) {
2845 // Lower arguments at fp - stackoffset + fpdiff.
2846 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2848 FPDiff = NumBytesCallerPushed - NumBytes;
2850 // Set the delta of movement of the returnaddr stackslot.
2851 // But only set if delta is greater than previous delta.
2852 if (FPDiff < X86Info->getTCReturnAddrDelta())
2853 X86Info->setTCReturnAddrDelta(FPDiff);
2856 unsigned NumBytesToPush = NumBytes;
2857 unsigned NumBytesToPop = NumBytes;
2859 // If we have an inalloca argument, all stack space has already been allocated
2860 // for us and be right at the top of the stack. We don't support multiple
2861 // arguments passed in memory when using inalloca.
2862 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2864 if (!ArgLocs.back().isMemLoc())
2865 report_fatal_error("cannot use inalloca attribute on a register "
2867 if (ArgLocs.back().getLocMemOffset() != 0)
2868 report_fatal_error("any parameter with the inalloca attribute must be "
2869 "the only memory argument");
2873 Chain = DAG.getCALLSEQ_START(
2874 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2876 SDValue RetAddrFrIdx;
2877 // Load return address for tail calls.
2878 if (isTailCall && FPDiff)
2879 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2880 Is64Bit, FPDiff, dl);
2882 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2883 SmallVector<SDValue, 8> MemOpChains;
2886 // Walk the register/memloc assignments, inserting copies/loads. In the case
2887 // of tail call optimization arguments are handle later.
2888 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
2889 DAG.getSubtarget().getRegisterInfo());
2890 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2891 // Skip inalloca arguments, they have already been written.
2892 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2893 if (Flags.isInAlloca())
2896 CCValAssign &VA = ArgLocs[i];
2897 EVT RegVT = VA.getLocVT();
2898 SDValue Arg = OutVals[i];
2899 bool isByVal = Flags.isByVal();
2901 // Promote the value if needed.
2902 switch (VA.getLocInfo()) {
2903 default: llvm_unreachable("Unknown loc info!");
2904 case CCValAssign::Full: break;
2905 case CCValAssign::SExt:
2906 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2908 case CCValAssign::ZExt:
2909 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2911 case CCValAssign::AExt:
2912 if (RegVT.is128BitVector()) {
2913 // Special case: passing MMX values in XMM registers.
2914 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2915 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2916 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2918 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2920 case CCValAssign::BCvt:
2921 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2923 case CCValAssign::Indirect: {
2924 // Store the argument.
2925 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2926 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2927 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2928 MachinePointerInfo::getFixedStack(FI),
2935 if (VA.isRegLoc()) {
2936 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2937 if (isVarArg && IsWin64) {
2938 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2939 // shadow reg if callee is a varargs function.
2940 unsigned ShadowReg = 0;
2941 switch (VA.getLocReg()) {
2942 case X86::XMM0: ShadowReg = X86::RCX; break;
2943 case X86::XMM1: ShadowReg = X86::RDX; break;
2944 case X86::XMM2: ShadowReg = X86::R8; break;
2945 case X86::XMM3: ShadowReg = X86::R9; break;
2948 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2950 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2951 assert(VA.isMemLoc());
2952 if (!StackPtr.getNode())
2953 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2955 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2956 dl, DAG, VA, Flags));
2960 if (!MemOpChains.empty())
2961 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2963 if (Subtarget->isPICStyleGOT()) {
2964 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2967 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2968 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2970 // If we are tail calling and generating PIC/GOT style code load the
2971 // address of the callee into ECX. The value in ecx is used as target of
2972 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2973 // for tail calls on PIC/GOT architectures. Normally we would just put the
2974 // address of GOT into ebx and then call target@PLT. But for tail calls
2975 // ebx would be restored (since ebx is callee saved) before jumping to the
2978 // Note: The actual moving to ECX is done further down.
2979 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2980 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2981 !G->getGlobal()->hasProtectedVisibility())
2982 Callee = LowerGlobalAddress(Callee, DAG);
2983 else if (isa<ExternalSymbolSDNode>(Callee))
2984 Callee = LowerExternalSymbol(Callee, DAG);
2988 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2989 // From AMD64 ABI document:
2990 // For calls that may call functions that use varargs or stdargs
2991 // (prototype-less calls or calls to functions containing ellipsis (...) in
2992 // the declaration) %al is used as hidden argument to specify the number
2993 // of SSE registers used. The contents of %al do not need to match exactly
2994 // the number of registers, but must be an ubound on the number of SSE
2995 // registers used and is in the range 0 - 8 inclusive.
2997 // Count the number of XMM registers allocated.
2998 static const MCPhysReg XMMArgRegs[] = {
2999 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3000 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3002 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
3003 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3004 && "SSE registers cannot be used when SSE is disabled");
3006 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3007 DAG.getConstant(NumXMMRegs, MVT::i8)));
3010 if (isVarArg && IsMustTail) {
3011 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3012 for (const auto &F : Forwards) {
3013 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3014 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3018 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3019 // don't need this because the eligibility check rejects calls that require
3020 // shuffling arguments passed in memory.
3021 if (!IsSibcall && isTailCall) {
3022 // Force all the incoming stack arguments to be loaded from the stack
3023 // before any new outgoing arguments are stored to the stack, because the
3024 // outgoing stack slots may alias the incoming argument stack slots, and
3025 // the alias isn't otherwise explicit. This is slightly more conservative
3026 // than necessary, because it means that each store effectively depends
3027 // on every argument instead of just those arguments it would clobber.
3028 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3030 SmallVector<SDValue, 8> MemOpChains2;
3033 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3034 CCValAssign &VA = ArgLocs[i];
3037 assert(VA.isMemLoc());
3038 SDValue Arg = OutVals[i];
3039 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3040 // Skip inalloca arguments. They don't require any work.
3041 if (Flags.isInAlloca())
3043 // Create frame index.
3044 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3045 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3046 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3047 FIN = DAG.getFrameIndex(FI, getPointerTy());
3049 if (Flags.isByVal()) {
3050 // Copy relative to framepointer.
3051 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3052 if (!StackPtr.getNode())
3053 StackPtr = DAG.getCopyFromReg(Chain, dl,
3054 RegInfo->getStackRegister(),
3056 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3058 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3062 // Store relative to framepointer.
3063 MemOpChains2.push_back(
3064 DAG.getStore(ArgChain, dl, Arg, FIN,
3065 MachinePointerInfo::getFixedStack(FI),
3070 if (!MemOpChains2.empty())
3071 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3073 // Store the return address to the appropriate stack slot.
3074 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3075 getPointerTy(), RegInfo->getSlotSize(),
3079 // Build a sequence of copy-to-reg nodes chained together with token chain
3080 // and flag operands which copy the outgoing args into registers.
3082 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3083 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3084 RegsToPass[i].second, InFlag);
3085 InFlag = Chain.getValue(1);
3088 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3089 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3090 // In the 64-bit large code model, we have to make all calls
3091 // through a register, since the call instruction's 32-bit
3092 // pc-relative offset may not be large enough to hold the whole
3094 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3095 // If the callee is a GlobalAddress node (quite common, every direct call
3096 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3098 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3100 // We should use extra load for direct calls to dllimported functions in
3102 const GlobalValue *GV = G->getGlobal();
3103 if (!GV->hasDLLImportStorageClass()) {
3104 unsigned char OpFlags = 0;
3105 bool ExtraLoad = false;
3106 unsigned WrapperKind = ISD::DELETED_NODE;
3108 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3109 // external symbols most go through the PLT in PIC mode. If the symbol
3110 // has hidden or protected visibility, or if it is static or local, then
3111 // we don't need to use the PLT - we can directly call it.
3112 if (Subtarget->isTargetELF() &&
3113 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3114 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3115 OpFlags = X86II::MO_PLT;
3116 } else if (Subtarget->isPICStyleStubAny() &&
3117 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3118 (!Subtarget->getTargetTriple().isMacOSX() ||
3119 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3120 // PC-relative references to external symbols should go through $stub,
3121 // unless we're building with the leopard linker or later, which
3122 // automatically synthesizes these stubs.
3123 OpFlags = X86II::MO_DARWIN_STUB;
3124 } else if (Subtarget->isPICStyleRIPRel() &&
3125 isa<Function>(GV) &&
3126 cast<Function>(GV)->getAttributes().
3127 hasAttribute(AttributeSet::FunctionIndex,
3128 Attribute::NonLazyBind)) {
3129 // If the function is marked as non-lazy, generate an indirect call
3130 // which loads from the GOT directly. This avoids runtime overhead
3131 // at the cost of eager binding (and one extra byte of encoding).
3132 OpFlags = X86II::MO_GOTPCREL;
3133 WrapperKind = X86ISD::WrapperRIP;
3137 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3138 G->getOffset(), OpFlags);
3140 // Add a wrapper if needed.
3141 if (WrapperKind != ISD::DELETED_NODE)
3142 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3143 // Add extra indirection if needed.
3145 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3146 MachinePointerInfo::getGOT(),
3147 false, false, false, 0);
3149 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3150 unsigned char OpFlags = 0;
3152 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3153 // external symbols should go through the PLT.
3154 if (Subtarget->isTargetELF() &&
3155 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3156 OpFlags = X86II::MO_PLT;
3157 } else if (Subtarget->isPICStyleStubAny() &&
3158 (!Subtarget->getTargetTriple().isMacOSX() ||
3159 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3160 // PC-relative references to external symbols should go through $stub,
3161 // unless we're building with the leopard linker or later, which
3162 // automatically synthesizes these stubs.
3163 OpFlags = X86II::MO_DARWIN_STUB;
3166 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3168 } else if (Subtarget->isTarget64BitILP32() && Callee->getValueType(0) == MVT::i32) {
3169 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3170 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3173 // Returns a chain & a flag for retval copy to use.
3174 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3175 SmallVector<SDValue, 8> Ops;
3177 if (!IsSibcall && isTailCall) {
3178 Chain = DAG.getCALLSEQ_END(Chain,
3179 DAG.getIntPtrConstant(NumBytesToPop, true),
3180 DAG.getIntPtrConstant(0, true), InFlag, dl);
3181 InFlag = Chain.getValue(1);
3184 Ops.push_back(Chain);
3185 Ops.push_back(Callee);
3188 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3190 // Add argument registers to the end of the list so that they are known live
3192 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3193 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3194 RegsToPass[i].second.getValueType()));
3196 // Add a register mask operand representing the call-preserved registers.
3197 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
3198 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3199 assert(Mask && "Missing call preserved mask for calling convention");
3200 Ops.push_back(DAG.getRegisterMask(Mask));
3202 if (InFlag.getNode())
3203 Ops.push_back(InFlag);
3207 //// If this is the first return lowered for this function, add the regs
3208 //// to the liveout set for the function.
3209 // This isn't right, although it's probably harmless on x86; liveouts
3210 // should be computed from returns not tail calls. Consider a void
3211 // function making a tail call to a function returning int.
3212 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3215 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3216 InFlag = Chain.getValue(1);
3218 // Create the CALLSEQ_END node.
3219 unsigned NumBytesForCalleeToPop;
3220 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3221 DAG.getTarget().Options.GuaranteedTailCallOpt))
3222 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3223 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3224 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3225 SR == StackStructReturn)
3226 // If this is a call to a struct-return function, the callee
3227 // pops the hidden struct pointer, so we have to push it back.
3228 // This is common for Darwin/X86, Linux & Mingw32 targets.
3229 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3230 NumBytesForCalleeToPop = 4;
3232 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3234 // Returns a flag for retval copy to use.
3236 Chain = DAG.getCALLSEQ_END(Chain,
3237 DAG.getIntPtrConstant(NumBytesToPop, true),
3238 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3241 InFlag = Chain.getValue(1);
3244 // Handle result values, copying them out of physregs into vregs that we
3246 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3247 Ins, dl, DAG, InVals);
3250 //===----------------------------------------------------------------------===//
3251 // Fast Calling Convention (tail call) implementation
3252 //===----------------------------------------------------------------------===//
3254 // Like std call, callee cleans arguments, convention except that ECX is
3255 // reserved for storing the tail called function address. Only 2 registers are
3256 // free for argument passing (inreg). Tail call optimization is performed
3258 // * tailcallopt is enabled
3259 // * caller/callee are fastcc
3260 // On X86_64 architecture with GOT-style position independent code only local
3261 // (within module) calls are supported at the moment.
3262 // To keep the stack aligned according to platform abi the function
3263 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3264 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3265 // If a tail called function callee has more arguments than the caller the
3266 // caller needs to make sure that there is room to move the RETADDR to. This is
3267 // achieved by reserving an area the size of the argument delta right after the
3268 // original RETADDR, but before the saved framepointer or the spilled registers
3269 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3281 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3282 /// for a 16 byte align requirement.
3284 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3285 SelectionDAG& DAG) const {
3286 MachineFunction &MF = DAG.getMachineFunction();
3287 const TargetMachine &TM = MF.getTarget();
3288 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3289 TM.getSubtargetImpl()->getRegisterInfo());
3290 const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
3291 unsigned StackAlignment = TFI.getStackAlignment();
3292 uint64_t AlignMask = StackAlignment - 1;
3293 int64_t Offset = StackSize;
3294 unsigned SlotSize = RegInfo->getSlotSize();
3295 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3296 // Number smaller than 12 so just add the difference.
3297 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3299 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3300 Offset = ((~AlignMask) & Offset) + StackAlignment +
3301 (StackAlignment-SlotSize);
3306 /// MatchingStackOffset - Return true if the given stack call argument is
3307 /// already available in the same position (relatively) of the caller's
3308 /// incoming argument stack.
3310 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3311 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3312 const X86InstrInfo *TII) {
3313 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3315 if (Arg.getOpcode() == ISD::CopyFromReg) {
3316 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3317 if (!TargetRegisterInfo::isVirtualRegister(VR))
3319 MachineInstr *Def = MRI->getVRegDef(VR);
3322 if (!Flags.isByVal()) {
3323 if (!TII->isLoadFromStackSlot(Def, FI))
3326 unsigned Opcode = Def->getOpcode();
3327 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
3328 Def->getOperand(1).isFI()) {
3329 FI = Def->getOperand(1).getIndex();
3330 Bytes = Flags.getByValSize();
3334 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3335 if (Flags.isByVal())
3336 // ByVal argument is passed in as a pointer but it's now being
3337 // dereferenced. e.g.
3338 // define @foo(%struct.X* %A) {
3339 // tail call @bar(%struct.X* byval %A)
3342 SDValue Ptr = Ld->getBasePtr();
3343 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3346 FI = FINode->getIndex();
3347 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3348 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3349 FI = FINode->getIndex();
3350 Bytes = Flags.getByValSize();
3354 assert(FI != INT_MAX);
3355 if (!MFI->isFixedObjectIndex(FI))
3357 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3360 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3361 /// for tail call optimization. Targets which want to do tail call
3362 /// optimization should implement this function.
3364 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3365 CallingConv::ID CalleeCC,
3367 bool isCalleeStructRet,
3368 bool isCallerStructRet,
3370 const SmallVectorImpl<ISD::OutputArg> &Outs,
3371 const SmallVectorImpl<SDValue> &OutVals,
3372 const SmallVectorImpl<ISD::InputArg> &Ins,
3373 SelectionDAG &DAG) const {
3374 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3377 // If -tailcallopt is specified, make fastcc functions tail-callable.
3378 const MachineFunction &MF = DAG.getMachineFunction();
3379 const Function *CallerF = MF.getFunction();
3381 // If the function return type is x86_fp80 and the callee return type is not,
3382 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3383 // perform a tailcall optimization here.
3384 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3387 CallingConv::ID CallerCC = CallerF->getCallingConv();
3388 bool CCMatch = CallerCC == CalleeCC;
3389 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3390 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3392 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3393 if (IsTailCallConvention(CalleeCC) && CCMatch)
3398 // Look for obvious safe cases to perform tail call optimization that do not
3399 // require ABI changes. This is what gcc calls sibcall.
3401 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3402 // emit a special epilogue.
3403 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3404 DAG.getSubtarget().getRegisterInfo());
3405 if (RegInfo->needsStackRealignment(MF))
3408 // Also avoid sibcall optimization if either caller or callee uses struct
3409 // return semantics.
3410 if (isCalleeStructRet || isCallerStructRet)
3413 // An stdcall/thiscall caller is expected to clean up its arguments; the
3414 // callee isn't going to do that.
3415 // FIXME: this is more restrictive than needed. We could produce a tailcall
3416 // when the stack adjustment matches. For example, with a thiscall that takes
3417 // only one argument.
3418 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3419 CallerCC == CallingConv::X86_ThisCall))
3422 // Do not sibcall optimize vararg calls unless all arguments are passed via
3424 if (isVarArg && !Outs.empty()) {
3426 // Optimizing for varargs on Win64 is unlikely to be safe without
3427 // additional testing.
3428 if (IsCalleeWin64 || IsCallerWin64)
3431 SmallVector<CCValAssign, 16> ArgLocs;
3432 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3435 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3436 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3437 if (!ArgLocs[i].isRegLoc())
3441 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3442 // stack. Therefore, if it's not used by the call it is not safe to optimize
3443 // this into a sibcall.
3444 bool Unused = false;
3445 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3452 SmallVector<CCValAssign, 16> RVLocs;
3453 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3455 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3456 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3457 CCValAssign &VA = RVLocs[i];
3458 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3463 // If the calling conventions do not match, then we'd better make sure the
3464 // results are returned in the same way as what the caller expects.
3466 SmallVector<CCValAssign, 16> RVLocs1;
3467 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3469 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3471 SmallVector<CCValAssign, 16> RVLocs2;
3472 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3474 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3476 if (RVLocs1.size() != RVLocs2.size())
3478 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3479 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3481 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3483 if (RVLocs1[i].isRegLoc()) {
3484 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3487 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3493 // If the callee takes no arguments then go on to check the results of the
3495 if (!Outs.empty()) {
3496 // Check if stack adjustment is needed. For now, do not do this if any
3497 // argument is passed on the stack.
3498 SmallVector<CCValAssign, 16> ArgLocs;
3499 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3502 // Allocate shadow area for Win64
3504 CCInfo.AllocateStack(32, 8);
3506 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3507 if (CCInfo.getNextStackOffset()) {
3508 MachineFunction &MF = DAG.getMachineFunction();
3509 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3512 // Check if the arguments are already laid out in the right way as
3513 // the caller's fixed stack objects.
3514 MachineFrameInfo *MFI = MF.getFrameInfo();
3515 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3516 const X86InstrInfo *TII =
3517 static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo());
3518 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3519 CCValAssign &VA = ArgLocs[i];
3520 SDValue Arg = OutVals[i];
3521 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3522 if (VA.getLocInfo() == CCValAssign::Indirect)
3524 if (!VA.isRegLoc()) {
3525 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3532 // If the tailcall address may be in a register, then make sure it's
3533 // possible to register allocate for it. In 32-bit, the call address can
3534 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3535 // callee-saved registers are restored. These happen to be the same
3536 // registers used to pass 'inreg' arguments so watch out for those.
3537 if (!Subtarget->is64Bit() &&
3538 ((!isa<GlobalAddressSDNode>(Callee) &&
3539 !isa<ExternalSymbolSDNode>(Callee)) ||
3540 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3541 unsigned NumInRegs = 0;
3542 // In PIC we need an extra register to formulate the address computation
3544 unsigned MaxInRegs =
3545 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3547 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3548 CCValAssign &VA = ArgLocs[i];
3551 unsigned Reg = VA.getLocReg();
3554 case X86::EAX: case X86::EDX: case X86::ECX:
3555 if (++NumInRegs == MaxInRegs)
3567 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3568 const TargetLibraryInfo *libInfo) const {
3569 return X86::createFastISel(funcInfo, libInfo);
3572 //===----------------------------------------------------------------------===//
3573 // Other Lowering Hooks
3574 //===----------------------------------------------------------------------===//
3576 static bool MayFoldLoad(SDValue Op) {
3577 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3580 static bool MayFoldIntoStore(SDValue Op) {
3581 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3584 static bool isTargetShuffle(unsigned Opcode) {
3586 default: return false;
3587 case X86ISD::BLENDI:
3588 case X86ISD::PSHUFB:
3589 case X86ISD::PSHUFD:
3590 case X86ISD::PSHUFHW:
3591 case X86ISD::PSHUFLW:
3593 case X86ISD::PALIGNR:
3594 case X86ISD::MOVLHPS:
3595 case X86ISD::MOVLHPD:
3596 case X86ISD::MOVHLPS:
3597 case X86ISD::MOVLPS:
3598 case X86ISD::MOVLPD:
3599 case X86ISD::MOVSHDUP:
3600 case X86ISD::MOVSLDUP:
3601 case X86ISD::MOVDDUP:
3604 case X86ISD::UNPCKL:
3605 case X86ISD::UNPCKH:
3606 case X86ISD::VPERMILPI:
3607 case X86ISD::VPERM2X128:
3608 case X86ISD::VPERMI:
3613 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3614 SDValue V1, SelectionDAG &DAG) {
3616 default: llvm_unreachable("Unknown x86 shuffle node");
3617 case X86ISD::MOVSHDUP:
3618 case X86ISD::MOVSLDUP:
3619 case X86ISD::MOVDDUP:
3620 return DAG.getNode(Opc, dl, VT, V1);
3624 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3625 SDValue V1, unsigned TargetMask,
3626 SelectionDAG &DAG) {
3628 default: llvm_unreachable("Unknown x86 shuffle node");
3629 case X86ISD::PSHUFD:
3630 case X86ISD::PSHUFHW:
3631 case X86ISD::PSHUFLW:
3632 case X86ISD::VPERMILPI:
3633 case X86ISD::VPERMI:
3634 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3638 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3639 SDValue V1, SDValue V2, unsigned TargetMask,
3640 SelectionDAG &DAG) {
3642 default: llvm_unreachable("Unknown x86 shuffle node");
3643 case X86ISD::PALIGNR:
3644 case X86ISD::VALIGN:
3646 case X86ISD::VPERM2X128:
3647 return DAG.getNode(Opc, dl, VT, V1, V2,
3648 DAG.getConstant(TargetMask, MVT::i8));
3652 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3653 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3655 default: llvm_unreachable("Unknown x86 shuffle node");
3656 case X86ISD::MOVLHPS:
3657 case X86ISD::MOVLHPD:
3658 case X86ISD::MOVHLPS:
3659 case X86ISD::MOVLPS:
3660 case X86ISD::MOVLPD:
3663 case X86ISD::UNPCKL:
3664 case X86ISD::UNPCKH:
3665 return DAG.getNode(Opc, dl, VT, V1, V2);
3669 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3670 MachineFunction &MF = DAG.getMachineFunction();
3671 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
3672 DAG.getSubtarget().getRegisterInfo());
3673 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3674 int ReturnAddrIndex = FuncInfo->getRAIndex();
3676 if (ReturnAddrIndex == 0) {
3677 // Set up a frame object for the return address.
3678 unsigned SlotSize = RegInfo->getSlotSize();
3679 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3682 FuncInfo->setRAIndex(ReturnAddrIndex);
3685 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3688 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3689 bool hasSymbolicDisplacement) {
3690 // Offset should fit into 32 bit immediate field.
3691 if (!isInt<32>(Offset))
3694 // If we don't have a symbolic displacement - we don't have any extra
3696 if (!hasSymbolicDisplacement)
3699 // FIXME: Some tweaks might be needed for medium code model.
3700 if (M != CodeModel::Small && M != CodeModel::Kernel)
3703 // For small code model we assume that latest object is 16MB before end of 31
3704 // bits boundary. We may also accept pretty large negative constants knowing
3705 // that all objects are in the positive half of address space.
3706 if (M == CodeModel::Small && Offset < 16*1024*1024)
3709 // For kernel code model we know that all object resist in the negative half
3710 // of 32bits address space. We may not accept negative offsets, since they may
3711 // be just off and we may accept pretty large positive ones.
3712 if (M == CodeModel::Kernel && Offset >= 0)
3718 /// isCalleePop - Determines whether the callee is required to pop its
3719 /// own arguments. Callee pop is necessary to support tail calls.
3720 bool X86::isCalleePop(CallingConv::ID CallingConv,
3721 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3722 switch (CallingConv) {
3725 case CallingConv::X86_StdCall:
3726 case CallingConv::X86_FastCall:
3727 case CallingConv::X86_ThisCall:
3729 case CallingConv::Fast:
3730 case CallingConv::GHC:
3731 case CallingConv::HiPE:
3738 /// \brief Return true if the condition is an unsigned comparison operation.
3739 static bool isX86CCUnsigned(unsigned X86CC) {
3741 default: llvm_unreachable("Invalid integer condition!");
3742 case X86::COND_E: return true;
3743 case X86::COND_G: return false;
3744 case X86::COND_GE: return false;
3745 case X86::COND_L: return false;
3746 case X86::COND_LE: return false;
3747 case X86::COND_NE: return true;
3748 case X86::COND_B: return true;
3749 case X86::COND_A: return true;
3750 case X86::COND_BE: return true;
3751 case X86::COND_AE: return true;
3753 llvm_unreachable("covered switch fell through?!");
3756 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3757 /// specific condition code, returning the condition code and the LHS/RHS of the
3758 /// comparison to make.
3759 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3760 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3762 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3763 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3764 // X > -1 -> X == 0, jump !sign.
3765 RHS = DAG.getConstant(0, RHS.getValueType());
3766 return X86::COND_NS;
3768 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3769 // X < 0 -> X == 0, jump on sign.
3772 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3774 RHS = DAG.getConstant(0, RHS.getValueType());
3775 return X86::COND_LE;
3779 switch (SetCCOpcode) {
3780 default: llvm_unreachable("Invalid integer condition!");
3781 case ISD::SETEQ: return X86::COND_E;
3782 case ISD::SETGT: return X86::COND_G;
3783 case ISD::SETGE: return X86::COND_GE;
3784 case ISD::SETLT: return X86::COND_L;
3785 case ISD::SETLE: return X86::COND_LE;
3786 case ISD::SETNE: return X86::COND_NE;
3787 case ISD::SETULT: return X86::COND_B;
3788 case ISD::SETUGT: return X86::COND_A;
3789 case ISD::SETULE: return X86::COND_BE;
3790 case ISD::SETUGE: return X86::COND_AE;
3794 // First determine if it is required or is profitable to flip the operands.
3796 // If LHS is a foldable load, but RHS is not, flip the condition.
3797 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3798 !ISD::isNON_EXTLoad(RHS.getNode())) {
3799 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3800 std::swap(LHS, RHS);
3803 switch (SetCCOpcode) {
3809 std::swap(LHS, RHS);
3813 // On a floating point condition, the flags are set as follows:
3815 // 0 | 0 | 0 | X > Y
3816 // 0 | 0 | 1 | X < Y
3817 // 1 | 0 | 0 | X == Y
3818 // 1 | 1 | 1 | unordered
3819 switch (SetCCOpcode) {
3820 default: llvm_unreachable("Condcode should be pre-legalized away");
3822 case ISD::SETEQ: return X86::COND_E;
3823 case ISD::SETOLT: // flipped
3825 case ISD::SETGT: return X86::COND_A;
3826 case ISD::SETOLE: // flipped
3828 case ISD::SETGE: return X86::COND_AE;
3829 case ISD::SETUGT: // flipped
3831 case ISD::SETLT: return X86::COND_B;
3832 case ISD::SETUGE: // flipped
3834 case ISD::SETLE: return X86::COND_BE;
3836 case ISD::SETNE: return X86::COND_NE;
3837 case ISD::SETUO: return X86::COND_P;
3838 case ISD::SETO: return X86::COND_NP;
3840 case ISD::SETUNE: return X86::COND_INVALID;
3844 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3845 /// code. Current x86 isa includes the following FP cmov instructions:
3846 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3847 static bool hasFPCMov(unsigned X86CC) {
3863 /// isFPImmLegal - Returns true if the target can instruction select the
3864 /// specified FP immediate natively. If false, the legalizer will
3865 /// materialize the FP immediate as a load from a constant pool.
3866 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3867 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3868 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3874 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3875 ISD::LoadExtType ExtTy,
3877 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3878 // relocation target a movq or addq instruction: don't let the load shrink.
3879 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3880 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3881 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3882 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3886 /// \brief Returns true if it is beneficial to convert a load of a constant
3887 /// to just the constant itself.
3888 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3890 assert(Ty->isIntegerTy());
3892 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3893 if (BitSize == 0 || BitSize > 64)
3898 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3899 unsigned Index) const {
3900 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3903 return (Index == 0 || Index == ResVT.getVectorNumElements());
3906 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3907 // Speculate cttz only if we can directly use TZCNT.
3908 return Subtarget->hasBMI();
3911 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3912 // Speculate ctlz only if we can directly use LZCNT.
3913 return Subtarget->hasLZCNT();
3916 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3917 /// the specified range (L, H].
3918 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3919 return (Val < 0) || (Val >= Low && Val < Hi);
3922 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3923 /// specified value.
3924 static bool isUndefOrEqual(int Val, int CmpVal) {
3925 return (Val < 0 || Val == CmpVal);
3928 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3929 /// from position Pos and ending in Pos+Size, falls within the specified
3930 /// sequential range (Low, Low+Size]. or is undef.
3931 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3932 unsigned Pos, unsigned Size, int Low) {
3933 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3934 if (!isUndefOrEqual(Mask[i], Low))
3939 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3940 /// is suitable for input to PSHUFD. That is, it doesn't reference the other
3941 /// operand - by default will match for first operand.
3942 static bool isPSHUFDMask(ArrayRef<int> Mask, MVT VT,
3943 bool TestSecondOperand = false) {
3944 if (VT != MVT::v4f32 && VT != MVT::v4i32 &&
3945 VT != MVT::v2f64 && VT != MVT::v2i64)
3948 unsigned NumElems = VT.getVectorNumElements();
3949 unsigned Lo = TestSecondOperand ? NumElems : 0;
3950 unsigned Hi = Lo + NumElems;
3952 for (unsigned i = 0; i < NumElems; ++i)
3953 if (!isUndefOrInRange(Mask[i], (int)Lo, (int)Hi))
3959 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3960 /// is suitable for input to PSHUFHW.
3961 static bool isPSHUFHWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3962 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3965 // Lower quadword copied in order or undef.
3966 if (!isSequentialOrUndefInRange(Mask, 0, 4, 0))
3969 // Upper quadword shuffled.
3970 for (unsigned i = 4; i != 8; ++i)
3971 if (!isUndefOrInRange(Mask[i], 4, 8))
3974 if (VT == MVT::v16i16) {
3975 // Lower quadword copied in order or undef.
3976 if (!isSequentialOrUndefInRange(Mask, 8, 4, 8))
3979 // Upper quadword shuffled.
3980 for (unsigned i = 12; i != 16; ++i)
3981 if (!isUndefOrInRange(Mask[i], 12, 16))
3988 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3989 /// is suitable for input to PSHUFLW.
3990 static bool isPSHUFLWMask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
3991 if (VT != MVT::v8i16 && (!HasInt256 || VT != MVT::v16i16))
3994 // Upper quadword copied in order.
3995 if (!isSequentialOrUndefInRange(Mask, 4, 4, 4))
3998 // Lower quadword shuffled.
3999 for (unsigned i = 0; i != 4; ++i)
4000 if (!isUndefOrInRange(Mask[i], 0, 4))
4003 if (VT == MVT::v16i16) {
4004 // Upper quadword copied in order.
4005 if (!isSequentialOrUndefInRange(Mask, 12, 4, 12))
4008 // Lower quadword shuffled.
4009 for (unsigned i = 8; i != 12; ++i)
4010 if (!isUndefOrInRange(Mask[i], 8, 12))
4017 /// \brief Return true if the mask specifies a shuffle of elements that is
4018 /// suitable for input to intralane (palignr) or interlane (valign) vector
4020 static bool isAlignrMask(ArrayRef<int> Mask, MVT VT, bool InterLane) {
4021 unsigned NumElts = VT.getVectorNumElements();
4022 unsigned NumLanes = InterLane ? 1: VT.getSizeInBits()/128;
4023 unsigned NumLaneElts = NumElts/NumLanes;
4025 // Do not handle 64-bit element shuffles with palignr.
4026 if (NumLaneElts == 2)
4029 for (unsigned l = 0; l != NumElts; l+=NumLaneElts) {
4031 for (i = 0; i != NumLaneElts; ++i) {
4036 // Lane is all undef, go to next lane
4037 if (i == NumLaneElts)
4040 int Start = Mask[i+l];
4042 // Make sure its in this lane in one of the sources
4043 if (!isUndefOrInRange(Start, l, l+NumLaneElts) &&
4044 !isUndefOrInRange(Start, l+NumElts, l+NumElts+NumLaneElts))
4047 // If not lane 0, then we must match lane 0
4048 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Start, Mask[i]+l))
4051 // Correct second source to be contiguous with first source
4052 if (Start >= (int)NumElts)
4053 Start -= NumElts - NumLaneElts;
4055 // Make sure we're shifting in the right direction.
4056 if (Start <= (int)(i+l))
4061 // Check the rest of the elements to see if they are consecutive.
4062 for (++i; i != NumLaneElts; ++i) {
4063 int Idx = Mask[i+l];
4065 // Make sure its in this lane
4066 if (!isUndefOrInRange(Idx, l, l+NumLaneElts) &&
4067 !isUndefOrInRange(Idx, l+NumElts, l+NumElts+NumLaneElts))
4070 // If not lane 0, then we must match lane 0
4071 if (l != 0 && Mask[i] >= 0 && !isUndefOrEqual(Idx, Mask[i]+l))
4074 if (Idx >= (int)NumElts)
4075 Idx -= NumElts - NumLaneElts;
4077 if (!isUndefOrEqual(Idx, Start+i))
4086 /// \brief Return true if the node specifies a shuffle of elements that is
4087 /// suitable for input to PALIGNR.
4088 static bool isPALIGNRMask(ArrayRef<int> Mask, MVT VT,
4089 const X86Subtarget *Subtarget) {
4090 if ((VT.is128BitVector() && !Subtarget->hasSSSE3()) ||
4091 (VT.is256BitVector() && !Subtarget->hasInt256()) ||
4092 VT.is512BitVector())
4093 // FIXME: Add AVX512BW.
4096 return isAlignrMask(Mask, VT, false);
4099 /// \brief Return true if the node specifies a shuffle of elements that is
4100 /// suitable for input to VALIGN.
4101 static bool isVALIGNMask(ArrayRef<int> Mask, MVT VT,
4102 const X86Subtarget *Subtarget) {
4103 // FIXME: Add AVX512VL.
4104 if (!VT.is512BitVector() || !Subtarget->hasAVX512())
4106 return isAlignrMask(Mask, VT, true);
4109 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4110 /// the two vector operands have swapped position.
4111 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
4112 unsigned NumElems) {
4113 for (unsigned i = 0; i != NumElems; ++i) {
4117 else if (idx < (int)NumElems)
4118 Mask[i] = idx + NumElems;
4120 Mask[i] = idx - NumElems;
4124 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
4125 /// specifies a shuffle of elements that is suitable for input to 128/256-bit
4126 /// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
4127 /// reverse of what x86 shuffles want.
4128 static bool isSHUFPMask(ArrayRef<int> Mask, MVT VT, bool Commuted = false) {
4130 unsigned NumElems = VT.getVectorNumElements();
4131 unsigned NumLanes = VT.getSizeInBits()/128;
4132 unsigned NumLaneElems = NumElems/NumLanes;
4134 if (NumLaneElems != 2 && NumLaneElems != 4)
4137 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4138 bool symetricMaskRequired =
4139 (VT.getSizeInBits() >= 256) && (EltSize == 32);
4141 // VSHUFPSY divides the resulting vector into 4 chunks.
4142 // The sources are also splitted into 4 chunks, and each destination
4143 // chunk must come from a different source chunk.
4145 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
4146 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
4148 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
4149 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
4151 // VSHUFPDY divides the resulting vector into 4 chunks.
4152 // The sources are also splitted into 4 chunks, and each destination
4153 // chunk must come from a different source chunk.
4155 // SRC1 => X3 X2 X1 X0
4156 // SRC2 => Y3 Y2 Y1 Y0
4158 // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
4160 SmallVector<int, 4> MaskVal(NumLaneElems, -1);
4161 unsigned HalfLaneElems = NumLaneElems/2;
4162 for (unsigned l = 0; l != NumElems; l += NumLaneElems) {
4163 for (unsigned i = 0; i != NumLaneElems; ++i) {
4164 int Idx = Mask[i+l];
4165 unsigned RngStart = l + ((Commuted == (i<HalfLaneElems)) ? NumElems : 0);
4166 if (!isUndefOrInRange(Idx, RngStart, RngStart+NumLaneElems))
4168 // For VSHUFPSY, the mask of the second half must be the same as the
4169 // first but with the appropriate offsets. This works in the same way as
4170 // VPERMILPS works with masks.
4171 if (!symetricMaskRequired || Idx < 0)
4173 if (MaskVal[i] < 0) {
4174 MaskVal[i] = Idx - l;
4177 if ((signed)(Idx - l) != MaskVal[i])
4185 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
4186 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
4187 static bool isMOVHLPSMask(ArrayRef<int> Mask, MVT VT) {
4188 if (!VT.is128BitVector())
4191 unsigned NumElems = VT.getVectorNumElements();
4196 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
4197 return isUndefOrEqual(Mask[0], 6) &&
4198 isUndefOrEqual(Mask[1], 7) &&
4199 isUndefOrEqual(Mask[2], 2) &&
4200 isUndefOrEqual(Mask[3], 3);
4203 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
4204 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
4206 static bool isMOVHLPS_v_undef_Mask(ArrayRef<int> Mask, MVT VT) {
4207 if (!VT.is128BitVector())
4210 unsigned NumElems = VT.getVectorNumElements();
4215 return isUndefOrEqual(Mask[0], 2) &&
4216 isUndefOrEqual(Mask[1], 3) &&
4217 isUndefOrEqual(Mask[2], 2) &&
4218 isUndefOrEqual(Mask[3], 3);
4221 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
4222 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
4223 static bool isMOVLPMask(ArrayRef<int> Mask, MVT VT) {
4224 if (!VT.is128BitVector())
4227 unsigned NumElems = VT.getVectorNumElements();
4229 if (NumElems != 2 && NumElems != 4)
4232 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4233 if (!isUndefOrEqual(Mask[i], i + NumElems))
4236 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
4237 if (!isUndefOrEqual(Mask[i], i))
4243 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
4244 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
4245 static bool isMOVLHPSMask(ArrayRef<int> Mask, MVT VT) {
4246 if (!VT.is128BitVector())
4249 unsigned NumElems = VT.getVectorNumElements();
4251 if (NumElems != 2 && NumElems != 4)
4254 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4255 if (!isUndefOrEqual(Mask[i], i))
4258 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4259 if (!isUndefOrEqual(Mask[i + e], i + NumElems))
4265 /// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
4266 /// specifies a shuffle of elements that is suitable for input to INSERTPS.
4267 /// i. e: If all but one element come from the same vector.
4268 static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
4269 // TODO: Deal with AVX's VINSERTPS
4270 if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
4273 unsigned CorrectPosV1 = 0;
4274 unsigned CorrectPosV2 = 0;
4275 for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i) {
4276 if (Mask[i] == -1) {
4284 else if (Mask[i] == i + 4)
4288 if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
4289 // We have 3 elements (undefs count as elements from any vector) from one
4290 // vector, and one from another.
4297 // Some special combinations that can be optimized.
4300 SDValue Compact8x32ShuffleNode(ShuffleVectorSDNode *SVOp,
4301 SelectionDAG &DAG) {
4302 MVT VT = SVOp->getSimpleValueType(0);
4305 if (VT != MVT::v8i32 && VT != MVT::v8f32)
4308 ArrayRef<int> Mask = SVOp->getMask();
4310 // These are the special masks that may be optimized.
4311 static const int MaskToOptimizeEven[] = {0, 8, 2, 10, 4, 12, 6, 14};
4312 static const int MaskToOptimizeOdd[] = {1, 9, 3, 11, 5, 13, 7, 15};
4313 bool MatchEvenMask = true;
4314 bool MatchOddMask = true;
4315 for (int i=0; i<8; ++i) {
4316 if (!isUndefOrEqual(Mask[i], MaskToOptimizeEven[i]))
4317 MatchEvenMask = false;
4318 if (!isUndefOrEqual(Mask[i], MaskToOptimizeOdd[i]))
4319 MatchOddMask = false;
4322 if (!MatchEvenMask && !MatchOddMask)
4325 SDValue UndefNode = DAG.getNode(ISD::UNDEF, dl, VT);
4327 SDValue Op0 = SVOp->getOperand(0);
4328 SDValue Op1 = SVOp->getOperand(1);
4330 if (MatchEvenMask) {
4331 // Shift the second operand right to 32 bits.
4332 static const int ShiftRightMask[] = {-1, 0, -1, 2, -1, 4, -1, 6 };
4333 Op1 = DAG.getVectorShuffle(VT, dl, Op1, UndefNode, ShiftRightMask);
4335 // Shift the first operand left to 32 bits.
4336 static const int ShiftLeftMask[] = {1, -1, 3, -1, 5, -1, 7, -1 };
4337 Op0 = DAG.getVectorShuffle(VT, dl, Op0, UndefNode, ShiftLeftMask);
4339 static const int BlendMask[] = {0, 9, 2, 11, 4, 13, 6, 15};
4340 return DAG.getVectorShuffle(VT, dl, Op0, Op1, BlendMask);
4343 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
4344 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
4345 static bool isUNPCKLMask(ArrayRef<int> Mask, MVT VT,
4346 bool HasInt256, bool V2IsSplat = false) {
4348 assert(VT.getSizeInBits() >= 128 &&
4349 "Unsupported vector type for unpckl");
4351 unsigned NumElts = VT.getVectorNumElements();
4352 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4353 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4356 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4357 "Unsupported vector type for unpckh");
4359 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4360 unsigned NumLanes = VT.getSizeInBits()/128;
4361 unsigned NumLaneElts = NumElts/NumLanes;
4363 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4364 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4365 int BitI = Mask[l+i];
4366 int BitI1 = Mask[l+i+1];
4367 if (!isUndefOrEqual(BitI, j))
4370 if (!isUndefOrEqual(BitI1, NumElts))
4373 if (!isUndefOrEqual(BitI1, j + NumElts))
4382 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
4383 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
4384 static bool isUNPCKHMask(ArrayRef<int> Mask, MVT VT,
4385 bool HasInt256, bool V2IsSplat = false) {
4386 assert(VT.getSizeInBits() >= 128 &&
4387 "Unsupported vector type for unpckh");
4389 unsigned NumElts = VT.getVectorNumElements();
4390 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4391 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4394 assert((!VT.is512BitVector() || VT.getScalarType().getSizeInBits() >= 32) &&
4395 "Unsupported vector type for unpckh");
4397 // AVX defines UNPCK* to operate independently on 128-bit lanes.
4398 unsigned NumLanes = VT.getSizeInBits()/128;
4399 unsigned NumLaneElts = NumElts/NumLanes;
4401 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4402 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4403 int BitI = Mask[l+i];
4404 int BitI1 = Mask[l+i+1];
4405 if (!isUndefOrEqual(BitI, j))
4408 if (isUndefOrEqual(BitI1, NumElts))
4411 if (!isUndefOrEqual(BitI1, j+NumElts))
4419 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
4420 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
4422 static bool isUNPCKL_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4423 unsigned NumElts = VT.getVectorNumElements();
4424 bool Is256BitVec = VT.is256BitVector();
4426 if (VT.is512BitVector())
4428 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4429 "Unsupported vector type for unpckh");
4431 if (Is256BitVec && NumElts != 4 && NumElts != 8 &&
4432 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4435 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
4436 // FIXME: Need a better way to get rid of this, there's no latency difference
4437 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
4438 // the former later. We should also remove the "_undef" special mask.
4439 if (NumElts == 4 && Is256BitVec)
4442 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4443 // independently on 128-bit lanes.
4444 unsigned NumLanes = VT.getSizeInBits()/128;
4445 unsigned NumLaneElts = NumElts/NumLanes;
4447 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4448 for (unsigned i = 0, j = l; i != NumLaneElts; i += 2, ++j) {
4449 int BitI = Mask[l+i];
4450 int BitI1 = Mask[l+i+1];
4452 if (!isUndefOrEqual(BitI, j))
4454 if (!isUndefOrEqual(BitI1, j))
4462 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
4463 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
4465 static bool isUNPCKH_v_undef_Mask(ArrayRef<int> Mask, MVT VT, bool HasInt256) {
4466 unsigned NumElts = VT.getVectorNumElements();
4468 if (VT.is512BitVector())
4471 assert((VT.is128BitVector() || VT.is256BitVector()) &&
4472 "Unsupported vector type for unpckh");
4474 if (VT.is256BitVector() && NumElts != 4 && NumElts != 8 &&
4475 (!HasInt256 || (NumElts != 16 && NumElts != 32)))
4478 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
4479 // independently on 128-bit lanes.
4480 unsigned NumLanes = VT.getSizeInBits()/128;
4481 unsigned NumLaneElts = NumElts/NumLanes;
4483 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
4484 for (unsigned i = 0, j = l+NumLaneElts/2; i != NumLaneElts; i += 2, ++j) {
4485 int BitI = Mask[l+i];
4486 int BitI1 = Mask[l+i+1];
4487 if (!isUndefOrEqual(BitI, j))
4489 if (!isUndefOrEqual(BitI1, j))
4496 // Match for INSERTI64x4 INSERTF64x4 instructions (src0[0], src1[0]) or
4497 // (src1[0], src0[1]), manipulation with 256-bit sub-vectors
4498 static bool isINSERT64x4Mask(ArrayRef<int> Mask, MVT VT, unsigned int *Imm) {
4499 if (!VT.is512BitVector())
4502 unsigned NumElts = VT.getVectorNumElements();
4503 unsigned HalfSize = NumElts/2;
4504 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, 0)) {
4505 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, NumElts)) {
4510 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, NumElts)) {
4511 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, HalfSize)) {
4519 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
4520 /// specifies a shuffle of elements that is suitable for input to MOVSS,
4521 /// MOVSD, and MOVD, i.e. setting the lowest element.
4522 static bool isMOVLMask(ArrayRef<int> Mask, EVT VT) {
4523 if (VT.getVectorElementType().getSizeInBits() < 32)
4525 if (!VT.is128BitVector())
4528 unsigned NumElts = VT.getVectorNumElements();
4530 if (!isUndefOrEqual(Mask[0], NumElts))
4533 for (unsigned i = 1; i != NumElts; ++i)
4534 if (!isUndefOrEqual(Mask[i], i))
4540 /// isVPERM2X128Mask - Match 256-bit shuffles where the elements are considered
4541 /// as permutations between 128-bit chunks or halves. As an example: this
4543 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
4544 /// The first half comes from the second half of V1 and the second half from the
4545 /// the second half of V2.
4546 static bool isVPERM2X128Mask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4547 if (!HasFp256 || !VT.is256BitVector())
4550 // The shuffle result is divided into half A and half B. In total the two
4551 // sources have 4 halves, namely: C, D, E, F. The final values of A and
4552 // B must come from C, D, E or F.
4553 unsigned HalfSize = VT.getVectorNumElements()/2;
4554 bool MatchA = false, MatchB = false;
4556 // Check if A comes from one of C, D, E, F.
4557 for (unsigned Half = 0; Half != 4; ++Half) {
4558 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
4564 // Check if B comes from one of C, D, E, F.
4565 for (unsigned Half = 0; Half != 4; ++Half) {
4566 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
4572 return MatchA && MatchB;
4575 /// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
4576 /// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
4577 static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
4578 MVT VT = SVOp->getSimpleValueType(0);
4580 unsigned HalfSize = VT.getVectorNumElements()/2;
4582 unsigned FstHalf = 0, SndHalf = 0;
4583 for (unsigned i = 0; i < HalfSize; ++i) {
4584 if (SVOp->getMaskElt(i) > 0) {
4585 FstHalf = SVOp->getMaskElt(i)/HalfSize;
4589 for (unsigned i = HalfSize; i < HalfSize*2; ++i) {
4590 if (SVOp->getMaskElt(i) > 0) {
4591 SndHalf = SVOp->getMaskElt(i)/HalfSize;
4596 return (FstHalf | (SndHalf << 4));
4599 // Symetric in-lane mask. Each lane has 4 elements (for imm8)
4600 static bool isPermImmMask(ArrayRef<int> Mask, MVT VT, unsigned& Imm8) {
4601 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4605 unsigned NumElts = VT.getVectorNumElements();
4607 if (VT.is128BitVector() || (VT.is256BitVector() && EltSize == 64)) {
4608 for (unsigned i = 0; i != NumElts; ++i) {
4611 Imm8 |= Mask[i] << (i*2);
4616 unsigned LaneSize = 4;
4617 SmallVector<int, 4> MaskVal(LaneSize, -1);
4619 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4620 for (unsigned i = 0; i != LaneSize; ++i) {
4621 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4625 if (MaskVal[i] < 0) {
4626 MaskVal[i] = Mask[i+l] - l;
4627 Imm8 |= MaskVal[i] << (i*2);
4630 if (Mask[i+l] != (signed)(MaskVal[i]+l))
4637 /// isVPERMILPMask - Return true if the specified VECTOR_SHUFFLE operand
4638 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
4639 /// Note that VPERMIL mask matching is different depending whether theunderlying
4640 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
4641 /// to the same elements of the low, but to the higher half of the source.
4642 /// In VPERMILPD the two lanes could be shuffled independently of each other
4643 /// with the same restriction that lanes can't be crossed. Also handles PSHUFDY.
4644 static bool isVPERMILPMask(ArrayRef<int> Mask, MVT VT) {
4645 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4646 if (VT.getSizeInBits() < 256 || EltSize < 32)
4648 bool symetricMaskRequired = (EltSize == 32);
4649 unsigned NumElts = VT.getVectorNumElements();
4651 unsigned NumLanes = VT.getSizeInBits()/128;
4652 unsigned LaneSize = NumElts/NumLanes;
4653 // 2 or 4 elements in one lane
4655 SmallVector<int, 4> ExpectedMaskVal(LaneSize, -1);
4656 for (unsigned l = 0; l != NumElts; l += LaneSize) {
4657 for (unsigned i = 0; i != LaneSize; ++i) {
4658 if (!isUndefOrInRange(Mask[i+l], l, l+LaneSize))
4660 if (symetricMaskRequired) {
4661 if (ExpectedMaskVal[i] < 0 && Mask[i+l] >= 0) {
4662 ExpectedMaskVal[i] = Mask[i+l] - l;
4665 if (!isUndefOrEqual(Mask[i+l], ExpectedMaskVal[i]+l))
4673 /// isCommutedMOVLMask - Returns true if the shuffle mask is except the reverse
4674 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
4675 /// element of vector 2 and the other elements to come from vector 1 in order.
4676 static bool isCommutedMOVLMask(ArrayRef<int> Mask, MVT VT,
4677 bool V2IsSplat = false, bool V2IsUndef = false) {
4678 if (!VT.is128BitVector())
4681 unsigned NumOps = VT.getVectorNumElements();
4682 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
4685 if (!isUndefOrEqual(Mask[0], 0))
4688 for (unsigned i = 1; i != NumOps; ++i)
4689 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
4690 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
4691 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
4697 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4698 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
4699 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
4700 static bool isMOVSHDUPMask(ArrayRef<int> Mask, MVT VT,
4701 const X86Subtarget *Subtarget) {
4702 if (!Subtarget->hasSSE3())
4705 unsigned NumElems = VT.getVectorNumElements();
4707 if ((VT.is128BitVector() && NumElems != 4) ||
4708 (VT.is256BitVector() && NumElems != 8) ||
4709 (VT.is512BitVector() && NumElems != 16))
4712 // "i+1" is the value the indexed mask element must have
4713 for (unsigned i = 0; i != NumElems; i += 2)
4714 if (!isUndefOrEqual(Mask[i], i+1) ||
4715 !isUndefOrEqual(Mask[i+1], i+1))
4721 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4722 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
4723 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
4724 static bool isMOVSLDUPMask(ArrayRef<int> Mask, MVT VT,
4725 const X86Subtarget *Subtarget) {
4726 if (!Subtarget->hasSSE3())
4729 unsigned NumElems = VT.getVectorNumElements();
4731 if ((VT.is128BitVector() && NumElems != 4) ||
4732 (VT.is256BitVector() && NumElems != 8) ||
4733 (VT.is512BitVector() && NumElems != 16))
4736 // "i" is the value the indexed mask element must have
4737 for (unsigned i = 0; i != NumElems; i += 2)
4738 if (!isUndefOrEqual(Mask[i], i) ||
4739 !isUndefOrEqual(Mask[i+1], i))
4745 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
4746 /// specifies a shuffle of elements that is suitable for input to 256-bit
4747 /// version of MOVDDUP.
4748 static bool isMOVDDUPYMask(ArrayRef<int> Mask, MVT VT, bool HasFp256) {
4749 if (!HasFp256 || !VT.is256BitVector())
4752 unsigned NumElts = VT.getVectorNumElements();
4756 for (unsigned i = 0; i != NumElts/2; ++i)
4757 if (!isUndefOrEqual(Mask[i], 0))
4759 for (unsigned i = NumElts/2; i != NumElts; ++i)
4760 if (!isUndefOrEqual(Mask[i], NumElts/2))
4765 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
4766 /// specifies a shuffle of elements that is suitable for input to 128-bit
4767 /// version of MOVDDUP.
4768 static bool isMOVDDUPMask(ArrayRef<int> Mask, MVT VT) {
4769 if (!VT.is128BitVector())
4772 unsigned e = VT.getVectorNumElements() / 2;
4773 for (unsigned i = 0; i != e; ++i)
4774 if (!isUndefOrEqual(Mask[i], i))
4776 for (unsigned i = 0; i != e; ++i)
4777 if (!isUndefOrEqual(Mask[e+i], i))
4782 /// isVEXTRACTIndex - Return true if the specified
4783 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
4784 /// suitable for instruction that extract 128 or 256 bit vectors
4785 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4786 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4787 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4790 // The index should be aligned on a vecWidth-bit boundary.
4792 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4794 MVT VT = N->getSimpleValueType(0);
4795 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4796 bool Result = (Index * ElSize) % vecWidth == 0;
4801 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
4802 /// operand specifies a subvector insert that is suitable for input to
4803 /// insertion of 128 or 256-bit subvectors
4804 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4805 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4806 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4808 // The index should be aligned on a vecWidth-bit boundary.
4810 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4812 MVT VT = N->getSimpleValueType(0);
4813 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4814 bool Result = (Index * ElSize) % vecWidth == 0;
4819 bool X86::isVINSERT128Index(SDNode *N) {
4820 return isVINSERTIndex(N, 128);
4823 bool X86::isVINSERT256Index(SDNode *N) {
4824 return isVINSERTIndex(N, 256);
4827 bool X86::isVEXTRACT128Index(SDNode *N) {
4828 return isVEXTRACTIndex(N, 128);
4831 bool X86::isVEXTRACT256Index(SDNode *N) {
4832 return isVEXTRACTIndex(N, 256);
4835 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4836 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4837 /// Handles 128-bit and 256-bit.
4838 static unsigned getShuffleSHUFImmediate(ShuffleVectorSDNode *N) {
4839 MVT VT = N->getSimpleValueType(0);
4841 assert((VT.getSizeInBits() >= 128) &&
4842 "Unsupported vector type for PSHUF/SHUFP");
4844 // Handle 128 and 256-bit vector lengths. AVX defines PSHUF/SHUFP to operate
4845 // independently on 128-bit lanes.
4846 unsigned NumElts = VT.getVectorNumElements();
4847 unsigned NumLanes = VT.getSizeInBits()/128;
4848 unsigned NumLaneElts = NumElts/NumLanes;
4850 assert((NumLaneElts == 2 || NumLaneElts == 4 || NumLaneElts == 8) &&
4851 "Only supports 2, 4 or 8 elements per lane");
4853 unsigned Shift = (NumLaneElts >= 4) ? 1 : 0;
4855 for (unsigned i = 0; i != NumElts; ++i) {
4856 int Elt = N->getMaskElt(i);
4857 if (Elt < 0) continue;
4858 Elt &= NumLaneElts - 1;
4859 unsigned ShAmt = (i << Shift) % 8;
4860 Mask |= Elt << ShAmt;
4866 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4867 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4868 static unsigned getShufflePSHUFHWImmediate(ShuffleVectorSDNode *N) {
4869 MVT VT = N->getSimpleValueType(0);
4871 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4872 "Unsupported vector type for PSHUFHW");
4874 unsigned NumElts = VT.getVectorNumElements();
4877 for (unsigned l = 0; l != NumElts; l += 8) {
4878 // 8 nodes per lane, but we only care about the last 4.
4879 for (unsigned i = 0; i < 4; ++i) {
4880 int Elt = N->getMaskElt(l+i+4);
4881 if (Elt < 0) continue;
4882 Elt &= 0x3; // only 2-bits.
4883 Mask |= Elt << (i * 2);
4890 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4891 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4892 static unsigned getShufflePSHUFLWImmediate(ShuffleVectorSDNode *N) {
4893 MVT VT = N->getSimpleValueType(0);
4895 assert((VT == MVT::v8i16 || VT == MVT::v16i16) &&
4896 "Unsupported vector type for PSHUFHW");
4898 unsigned NumElts = VT.getVectorNumElements();
4901 for (unsigned l = 0; l != NumElts; l += 8) {
4902 // 8 nodes per lane, but we only care about the first 4.
4903 for (unsigned i = 0; i < 4; ++i) {
4904 int Elt = N->getMaskElt(l+i);
4905 if (Elt < 0) continue;
4906 Elt &= 0x3; // only 2-bits
4907 Mask |= Elt << (i * 2);
4914 /// \brief Return the appropriate immediate to shuffle the specified
4915 /// VECTOR_SHUFFLE mask with the PALIGNR (if InterLane is false) or with
4916 /// VALIGN (if Interlane is true) instructions.
4917 static unsigned getShuffleAlignrImmediate(ShuffleVectorSDNode *SVOp,
4919 MVT VT = SVOp->getSimpleValueType(0);
4920 unsigned EltSize = InterLane ? 1 :
4921 VT.getVectorElementType().getSizeInBits() >> 3;
4923 unsigned NumElts = VT.getVectorNumElements();
4924 unsigned NumLanes = VT.is512BitVector() ? 1 : VT.getSizeInBits()/128;
4925 unsigned NumLaneElts = NumElts/NumLanes;
4929 for (i = 0; i != NumElts; ++i) {
4930 Val = SVOp->getMaskElt(i);
4934 if (Val >= (int)NumElts)
4935 Val -= NumElts - NumLaneElts;
4937 assert(Val - i > 0 && "PALIGNR imm should be positive");
4938 return (Val - i) * EltSize;
4941 /// \brief Return the appropriate immediate to shuffle the specified
4942 /// VECTOR_SHUFFLE mask with the PALIGNR instruction.
4943 static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
4944 return getShuffleAlignrImmediate(SVOp, false);
4947 /// \brief Return the appropriate immediate to shuffle the specified
4948 /// VECTOR_SHUFFLE mask with the VALIGN instruction.
4949 static unsigned getShuffleVALIGNImmediate(ShuffleVectorSDNode *SVOp) {
4950 return getShuffleAlignrImmediate(SVOp, true);
4954 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4955 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4956 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4957 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4960 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4962 MVT VecVT = N->getOperand(0).getSimpleValueType();
4963 MVT ElVT = VecVT.getVectorElementType();
4965 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4966 return Index / NumElemsPerChunk;
4969 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4970 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4971 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4972 llvm_unreachable("Illegal insert subvector for VINSERT");
4975 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4977 MVT VecVT = N->getSimpleValueType(0);
4978 MVT ElVT = VecVT.getVectorElementType();
4980 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4981 return Index / NumElemsPerChunk;
4984 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4985 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4986 /// and VINSERTI128 instructions.
4987 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4988 return getExtractVEXTRACTImmediate(N, 128);
4991 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4992 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4993 /// and VINSERTI64x4 instructions.
4994 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4995 return getExtractVEXTRACTImmediate(N, 256);
4998 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4999 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
5000 /// and VINSERTI128 instructions.
5001 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
5002 return getInsertVINSERTImmediate(N, 128);
5005 /// getInsertVINSERT256Immediate - Return the appropriate immediate
5006 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
5007 /// and VINSERTI64x4 instructions.
5008 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
5009 return getInsertVINSERTImmediate(N, 256);
5012 /// isZero - Returns true if Elt is a constant integer zero
5013 static bool isZero(SDValue V) {
5014 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
5015 return C && C->isNullValue();
5018 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
5020 bool X86::isZeroNode(SDValue Elt) {
5023 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
5024 return CFP->getValueAPF().isPosZero();
5028 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
5029 /// match movhlps. The lower half elements should come from upper half of
5030 /// V1 (and in order), and the upper half elements should come from the upper
5031 /// half of V2 (and in order).
5032 static bool ShouldXformToMOVHLPS(ArrayRef<int> Mask, MVT VT) {
5033 if (!VT.is128BitVector())
5035 if (VT.getVectorNumElements() != 4)
5037 for (unsigned i = 0, e = 2; i != e; ++i)
5038 if (!isUndefOrEqual(Mask[i], i+2))
5040 for (unsigned i = 2; i != 4; ++i)
5041 if (!isUndefOrEqual(Mask[i], i+4))
5046 /// isScalarLoadToVector - Returns true if the node is a scalar load that
5047 /// is promoted to a vector. It also returns the LoadSDNode by reference if
5049 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
5050 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
5052 N = N->getOperand(0).getNode();
5053 if (!ISD::isNON_EXTLoad(N))
5056 *LD = cast<LoadSDNode>(N);
5060 // Test whether the given value is a vector value which will be legalized
5062 static bool WillBeConstantPoolLoad(SDNode *N) {
5063 if (N->getOpcode() != ISD::BUILD_VECTOR)
5066 // Check for any non-constant elements.
5067 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
5068 switch (N->getOperand(i).getNode()->getOpcode()) {
5070 case ISD::ConstantFP:
5077 // Vectors of all-zeros and all-ones are materialized with special
5078 // instructions rather than being loaded.
5079 return !ISD::isBuildVectorAllZeros(N) &&
5080 !ISD::isBuildVectorAllOnes(N);
5083 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
5084 /// match movlp{s|d}. The lower half elements should come from lower half of
5085 /// V1 (and in order), and the upper half elements should come from the upper
5086 /// half of V2 (and in order). And since V1 will become the source of the
5087 /// MOVLP, it must be either a vector load or a scalar load to vector.
5088 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
5089 ArrayRef<int> Mask, MVT VT) {
5090 if (!VT.is128BitVector())
5093 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
5095 // Is V2 is a vector load, don't do this transformation. We will try to use
5096 // load folding shufps op.
5097 if (ISD::isNON_EXTLoad(V2) || WillBeConstantPoolLoad(V2))
5100 unsigned NumElems = VT.getVectorNumElements();
5102 if (NumElems != 2 && NumElems != 4)
5104 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
5105 if (!isUndefOrEqual(Mask[i], i))
5107 for (unsigned i = NumElems/2, e = NumElems; i != e; ++i)
5108 if (!isUndefOrEqual(Mask[i], i+NumElems))
5113 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
5114 /// to an zero vector.
5115 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
5116 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
5117 SDValue V1 = N->getOperand(0);
5118 SDValue V2 = N->getOperand(1);
5119 unsigned NumElems = N->getValueType(0).getVectorNumElements();
5120 for (unsigned i = 0; i != NumElems; ++i) {
5121 int Idx = N->getMaskElt(i);
5122 if (Idx >= (int)NumElems) {
5123 unsigned Opc = V2.getOpcode();
5124 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
5126 if (Opc != ISD::BUILD_VECTOR ||
5127 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
5129 } else if (Idx >= 0) {
5130 unsigned Opc = V1.getOpcode();
5131 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
5133 if (Opc != ISD::BUILD_VECTOR ||
5134 !X86::isZeroNode(V1.getOperand(Idx)))
5141 /// getZeroVector - Returns a vector of specified type with all zero elements.
5143 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
5144 SelectionDAG &DAG, SDLoc dl) {
5145 assert(VT.isVector() && "Expected a vector type");
5147 // Always build SSE zero vectors as <4 x i32> bitcasted
5148 // to their dest type. This ensures they get CSE'd.
5150 if (VT.is128BitVector()) { // SSE
5151 if (Subtarget->hasSSE2()) { // SSE2
5152 SDValue Cst = DAG.getConstant(0, MVT::i32);
5153 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5155 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5156 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
5158 } else if (VT.is256BitVector()) { // AVX
5159 if (Subtarget->hasInt256()) { // AVX2
5160 SDValue Cst = DAG.getConstant(0, MVT::i32);
5161 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5162 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5164 // 256-bit logic and arithmetic instructions in AVX are all
5165 // floating-point, no support for integer ops. Emit fp zeroed vectors.
5166 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
5167 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5168 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
5170 } else if (VT.is512BitVector()) { // AVX-512
5171 SDValue Cst = DAG.getConstant(0, MVT::i32);
5172 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
5173 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5174 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
5175 } else if (VT.getScalarType() == MVT::i1) {
5176 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
5177 SDValue Cst = DAG.getConstant(0, MVT::i1);
5178 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5179 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5181 llvm_unreachable("Unexpected vector type");
5183 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5186 /// getOnesVector - Returns a vector of specified type with all bits set.
5187 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
5188 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
5189 /// Then bitcast to their original type, ensuring they get CSE'd.
5190 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
5192 assert(VT.isVector() && "Expected a vector type");
5194 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
5196 if (VT.is256BitVector()) {
5197 if (HasInt256) { // AVX2
5198 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
5199 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
5201 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5202 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
5204 } else if (VT.is128BitVector()) {
5205 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
5207 llvm_unreachable("Unexpected vector type");
5209 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
5212 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
5213 /// that point to V2 points to its first element.
5214 static void NormalizeMask(SmallVectorImpl<int> &Mask, unsigned NumElems) {
5215 for (unsigned i = 0; i != NumElems; ++i) {
5216 if (Mask[i] > (int)NumElems) {
5222 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
5223 /// operation of specified width.
5224 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
5226 unsigned NumElems = VT.getVectorNumElements();
5227 SmallVector<int, 8> Mask;
5228 Mask.push_back(NumElems);
5229 for (unsigned i = 1; i != NumElems; ++i)
5231 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5234 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
5235 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5237 unsigned NumElems = VT.getVectorNumElements();
5238 SmallVector<int, 8> Mask;
5239 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
5241 Mask.push_back(i + NumElems);
5243 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5246 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
5247 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
5249 unsigned NumElems = VT.getVectorNumElements();
5250 SmallVector<int, 8> Mask;
5251 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
5252 Mask.push_back(i + Half);
5253 Mask.push_back(i + NumElems + Half);
5255 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
5258 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
5259 // a generic shuffle instruction because the target has no such instructions.
5260 // Generate shuffles which repeat i16 and i8 several times until they can be
5261 // represented by v4f32 and then be manipulated by target suported shuffles.
5262 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
5263 MVT VT = V.getSimpleValueType();
5264 int NumElems = VT.getVectorNumElements();
5267 while (NumElems > 4) {
5268 if (EltNo < NumElems/2) {
5269 V = getUnpackl(DAG, dl, VT, V, V);
5271 V = getUnpackh(DAG, dl, VT, V, V);
5272 EltNo -= NumElems/2;
5279 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
5280 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
5281 MVT VT = V.getSimpleValueType();
5284 if (VT.is128BitVector()) {
5285 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
5286 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
5287 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
5289 } else if (VT.is256BitVector()) {
5290 // To use VPERMILPS to splat scalars, the second half of indicies must
5291 // refer to the higher part, which is a duplication of the lower one,
5292 // because VPERMILPS can only handle in-lane permutations.
5293 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
5294 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
5296 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
5297 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
5300 llvm_unreachable("Vector size not supported");
5302 return DAG.getNode(ISD::BITCAST, dl, VT, V);
5305 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
5306 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
5307 MVT SrcVT = SV->getSimpleValueType(0);
5308 SDValue V1 = SV->getOperand(0);
5311 int EltNo = SV->getSplatIndex();
5312 int NumElems = SrcVT.getVectorNumElements();
5313 bool Is256BitVec = SrcVT.is256BitVector();
5315 assert(((SrcVT.is128BitVector() && NumElems > 4) || Is256BitVec) &&
5316 "Unknown how to promote splat for type");
5318 // Extract the 128-bit part containing the splat element and update
5319 // the splat element index when it refers to the higher register.
5321 V1 = Extract128BitVector(V1, EltNo, DAG, dl);
5322 if (EltNo >= NumElems/2)
5323 EltNo -= NumElems/2;
5326 // All i16 and i8 vector types can't be used directly by a generic shuffle
5327 // instruction because the target has no such instruction. Generate shuffles
5328 // which repeat i16 and i8 several times until they fit in i32, and then can
5329 // be manipulated by target suported shuffles.
5330 MVT EltVT = SrcVT.getVectorElementType();
5331 if (EltVT == MVT::i8 || EltVT == MVT::i16)
5332 V1 = PromoteSplati8i16(V1, DAG, EltNo);
5334 // Recreate the 256-bit vector and place the same 128-bit vector
5335 // into the low and high part. This is necessary because we want
5336 // to use VPERM* to shuffle the vectors
5338 V1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, SrcVT, V1, V1);
5341 return getLegalSplat(DAG, V1, EltNo);
5344 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
5345 /// vector of zero or undef vector. This produces a shuffle where the low
5346 /// element of V2 is swizzled into the zero/undef vector, landing at element
5347 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
5348 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
5350 const X86Subtarget *Subtarget,
5351 SelectionDAG &DAG) {
5352 MVT VT = V2.getSimpleValueType();
5354 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
5355 unsigned NumElems = VT.getVectorNumElements();
5356 SmallVector<int, 16> MaskVec;
5357 for (unsigned i = 0; i != NumElems; ++i)
5358 // If this is the insertion idx, put the low elt of V2 here.
5359 MaskVec.push_back(i == Idx ? NumElems : i);
5360 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
5363 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
5364 /// target specific opcode. Returns true if the Mask could be calculated. Sets
5365 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
5366 /// shuffles which use a single input multiple times, and in those cases it will
5367 /// adjust the mask to only have indices within that single input.
5368 static bool getTargetShuffleMask(SDNode *N, MVT VT,
5369 SmallVectorImpl<int> &Mask, bool &IsUnary) {
5370 unsigned NumElems = VT.getVectorNumElements();
5374 bool IsFakeUnary = false;
5375 switch(N->getOpcode()) {
5376 case X86ISD::BLENDI:
5377 ImmN = N->getOperand(N->getNumOperands()-1);
5378 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5381 ImmN = N->getOperand(N->getNumOperands()-1);
5382 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5383 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5385 case X86ISD::UNPCKH:
5386 DecodeUNPCKHMask(VT, Mask);
5387 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5389 case X86ISD::UNPCKL:
5390 DecodeUNPCKLMask(VT, Mask);
5391 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5393 case X86ISD::MOVHLPS:
5394 DecodeMOVHLPSMask(NumElems, Mask);
5395 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5397 case X86ISD::MOVLHPS:
5398 DecodeMOVLHPSMask(NumElems, Mask);
5399 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
5401 case X86ISD::PALIGNR:
5402 ImmN = N->getOperand(N->getNumOperands()-1);
5403 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5405 case X86ISD::PSHUFD:
5406 case X86ISD::VPERMILPI:
5407 ImmN = N->getOperand(N->getNumOperands()-1);
5408 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5411 case X86ISD::PSHUFHW:
5412 ImmN = N->getOperand(N->getNumOperands()-1);
5413 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5416 case X86ISD::PSHUFLW:
5417 ImmN = N->getOperand(N->getNumOperands()-1);
5418 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5421 case X86ISD::PSHUFB: {
5423 SDValue MaskNode = N->getOperand(1);
5424 while (MaskNode->getOpcode() == ISD::BITCAST)
5425 MaskNode = MaskNode->getOperand(0);
5427 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
5428 // If we have a build-vector, then things are easy.
5429 EVT VT = MaskNode.getValueType();
5430 assert(VT.isVector() &&
5431 "Can't produce a non-vector with a build_vector!");
5432 if (!VT.isInteger())
5435 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
5437 SmallVector<uint64_t, 32> RawMask;
5438 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
5439 SDValue Op = MaskNode->getOperand(i);
5440 if (Op->getOpcode() == ISD::UNDEF) {
5441 RawMask.push_back((uint64_t)SM_SentinelUndef);
5444 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
5447 APInt MaskElement = CN->getAPIntValue();
5449 // We now have to decode the element which could be any integer size and
5450 // extract each byte of it.
5451 for (int j = 0; j < NumBytesPerElement; ++j) {
5452 // Note that this is x86 and so always little endian: the low byte is
5453 // the first byte of the mask.
5454 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
5455 MaskElement = MaskElement.lshr(8);
5458 DecodePSHUFBMask(RawMask, Mask);
5462 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
5466 SDValue Ptr = MaskLoad->getBasePtr();
5467 if (Ptr->getOpcode() == X86ISD::Wrapper)
5468 Ptr = Ptr->getOperand(0);
5470 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
5471 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
5474 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
5475 // FIXME: Support AVX-512 here.
5476 Type *Ty = C->getType();
5477 if (!Ty->isVectorTy() || (Ty->getVectorNumElements() != 16 &&
5478 Ty->getVectorNumElements() != 32))
5481 DecodePSHUFBMask(C, Mask);
5487 case X86ISD::VPERMI:
5488 ImmN = N->getOperand(N->getNumOperands()-1);
5489 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5493 case X86ISD::MOVSD: {
5494 // The index 0 always comes from the first element of the second source,
5495 // this is why MOVSS and MOVSD are used in the first place. The other
5496 // elements come from the other positions of the first source vector
5497 Mask.push_back(NumElems);
5498 for (unsigned i = 1; i != NumElems; ++i) {
5503 case X86ISD::VPERM2X128:
5504 ImmN = N->getOperand(N->getNumOperands()-1);
5505 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
5506 if (Mask.empty()) return false;
5508 case X86ISD::MOVSLDUP:
5509 DecodeMOVSLDUPMask(VT, Mask);
5511 case X86ISD::MOVSHDUP:
5512 DecodeMOVSHDUPMask(VT, Mask);
5514 case X86ISD::MOVDDUP:
5515 case X86ISD::MOVLHPD:
5516 case X86ISD::MOVLPD:
5517 case X86ISD::MOVLPS:
5518 // Not yet implemented
5520 default: llvm_unreachable("unknown target shuffle node");
5523 // If we have a fake unary shuffle, the shuffle mask is spread across two
5524 // inputs that are actually the same node. Re-map the mask to always point
5525 // into the first input.
5528 if (M >= (int)Mask.size())
5534 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
5535 /// element of the result of the vector shuffle.
5536 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
5539 return SDValue(); // Limit search depth.
5541 SDValue V = SDValue(N, 0);
5542 EVT VT = V.getValueType();
5543 unsigned Opcode = V.getOpcode();
5545 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
5546 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
5547 int Elt = SV->getMaskElt(Index);
5550 return DAG.getUNDEF(VT.getVectorElementType());
5552 unsigned NumElems = VT.getVectorNumElements();
5553 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
5554 : SV->getOperand(1);
5555 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
5558 // Recurse into target specific vector shuffles to find scalars.
5559 if (isTargetShuffle(Opcode)) {
5560 MVT ShufVT = V.getSimpleValueType();
5561 unsigned NumElems = ShufVT.getVectorNumElements();
5562 SmallVector<int, 16> ShuffleMask;
5565 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
5568 int Elt = ShuffleMask[Index];
5570 return DAG.getUNDEF(ShufVT.getVectorElementType());
5572 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
5574 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
5578 // Actual nodes that may contain scalar elements
5579 if (Opcode == ISD::BITCAST) {
5580 V = V.getOperand(0);
5581 EVT SrcVT = V.getValueType();
5582 unsigned NumElems = VT.getVectorNumElements();
5584 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
5588 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5589 return (Index == 0) ? V.getOperand(0)
5590 : DAG.getUNDEF(VT.getVectorElementType());
5592 if (V.getOpcode() == ISD::BUILD_VECTOR)
5593 return V.getOperand(Index);
5598 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
5599 /// shuffle operation which come from a consecutively from a zero. The
5600 /// search can start in two different directions, from left or right.
5601 /// We count undefs as zeros until PreferredNum is reached.
5602 static unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp,
5603 unsigned NumElems, bool ZerosFromLeft,
5605 unsigned PreferredNum = -1U) {
5606 unsigned NumZeros = 0;
5607 for (unsigned i = 0; i != NumElems; ++i) {
5608 unsigned Index = ZerosFromLeft ? i : NumElems - i - 1;
5609 SDValue Elt = getShuffleScalarElt(SVOp, Index, DAG, 0);
5613 if (X86::isZeroNode(Elt))
5615 else if (Elt.getOpcode() == ISD::UNDEF) // Undef as zero up to PreferredNum.
5616 NumZeros = std::min(NumZeros + 1, PreferredNum);
5624 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies [MaskI, MaskE)
5625 /// correspond consecutively to elements from one of the vector operands,
5626 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
5628 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp,
5629 unsigned MaskI, unsigned MaskE, unsigned OpIdx,
5630 unsigned NumElems, unsigned &OpNum) {
5631 bool SeenV1 = false;
5632 bool SeenV2 = false;
5634 for (unsigned i = MaskI; i != MaskE; ++i, ++OpIdx) {
5635 int Idx = SVOp->getMaskElt(i);
5636 // Ignore undef indicies
5640 if (Idx < (int)NumElems)
5645 // Only accept consecutive elements from the same vector
5646 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
5650 OpNum = SeenV1 ? 0 : 1;
5654 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
5655 /// logical left shift of a vector.
5656 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5657 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5659 SVOp->getSimpleValueType(0).getVectorNumElements();
5660 unsigned NumZeros = getNumOfConsecutiveZeros(
5661 SVOp, NumElems, false /* check zeros from right */, DAG,
5662 SVOp->getMaskElt(0));
5668 // Considering the elements in the mask that are not consecutive zeros,
5669 // check if they consecutively come from only one of the source vectors.
5671 // V1 = {X, A, B, C} 0
5673 // vector_shuffle V1, V2 <1, 2, 3, X>
5675 if (!isShuffleMaskConsecutive(SVOp,
5676 0, // Mask Start Index
5677 NumElems-NumZeros, // Mask End Index(exclusive)
5678 NumZeros, // Where to start looking in the src vector
5679 NumElems, // Number of elements in vector
5680 OpSrc)) // Which source operand ?
5685 ShVal = SVOp->getOperand(OpSrc);
5689 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
5690 /// logical left shift of a vector.
5691 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5692 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5694 SVOp->getSimpleValueType(0).getVectorNumElements();
5695 unsigned NumZeros = getNumOfConsecutiveZeros(
5696 SVOp, NumElems, true /* check zeros from left */, DAG,
5697 NumElems - SVOp->getMaskElt(NumElems - 1) - 1);
5703 // Considering the elements in the mask that are not consecutive zeros,
5704 // check if they consecutively come from only one of the source vectors.
5706 // 0 { A, B, X, X } = V2
5708 // vector_shuffle V1, V2 <X, X, 4, 5>
5710 if (!isShuffleMaskConsecutive(SVOp,
5711 NumZeros, // Mask Start Index
5712 NumElems, // Mask End Index(exclusive)
5713 0, // Where to start looking in the src vector
5714 NumElems, // Number of elements in vector
5715 OpSrc)) // Which source operand ?
5720 ShVal = SVOp->getOperand(OpSrc);
5724 /// isVectorShift - Returns true if the shuffle can be implemented as a
5725 /// logical left or right shift of a vector.
5726 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
5727 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
5728 // Although the logic below support any bitwidth size, there are no
5729 // shift instructions which handle more than 128-bit vectors.
5730 if (!SVOp->getSimpleValueType(0).is128BitVector())
5733 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
5734 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
5740 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
5742 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
5743 unsigned NumNonZero, unsigned NumZero,
5745 const X86Subtarget* Subtarget,
5746 const TargetLowering &TLI) {
5753 for (unsigned i = 0; i < 16; ++i) {
5754 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5755 if (ThisIsNonZero && First) {
5757 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5759 V = DAG.getUNDEF(MVT::v8i16);
5764 SDValue ThisElt, LastElt;
5765 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5766 if (LastIsNonZero) {
5767 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5768 MVT::i16, Op.getOperand(i-1));
5770 if (ThisIsNonZero) {
5771 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5772 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5773 ThisElt, DAG.getConstant(8, MVT::i8));
5775 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5779 if (ThisElt.getNode())
5780 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5781 DAG.getIntPtrConstant(i/2));
5785 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
5788 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
5790 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5791 unsigned NumNonZero, unsigned NumZero,
5793 const X86Subtarget* Subtarget,
5794 const TargetLowering &TLI) {
5801 for (unsigned i = 0; i < 8; ++i) {
5802 bool isNonZero = (NonZeros & (1 << i)) != 0;
5806 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5808 V = DAG.getUNDEF(MVT::v8i16);
5811 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5812 MVT::v8i16, V, Op.getOperand(i),
5813 DAG.getIntPtrConstant(i));
5820 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
5821 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5822 const X86Subtarget *Subtarget,
5823 const TargetLowering &TLI) {
5824 // Find all zeroable elements.
5826 for (int i=0; i < 4; ++i) {
5827 SDValue Elt = Op->getOperand(i);
5828 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5830 assert(std::count_if(&Zeroable[0], &Zeroable[4],
5831 [](bool M) { return !M; }) > 1 &&
5832 "We expect at least two non-zero elements!");
5834 // We only know how to deal with build_vector nodes where elements are either
5835 // zeroable or extract_vector_elt with constant index.
5836 SDValue FirstNonZero;
5837 unsigned FirstNonZeroIdx;
5838 for (unsigned i=0; i < 4; ++i) {
5841 SDValue Elt = Op->getOperand(i);
5842 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5843 !isa<ConstantSDNode>(Elt.getOperand(1)))
5845 // Make sure that this node is extracting from a 128-bit vector.
5846 MVT VT = Elt.getOperand(0).getSimpleValueType();
5847 if (!VT.is128BitVector())
5849 if (!FirstNonZero.getNode()) {
5851 FirstNonZeroIdx = i;
5855 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5856 SDValue V1 = FirstNonZero.getOperand(0);
5857 MVT VT = V1.getSimpleValueType();
5859 // See if this build_vector can be lowered as a blend with zero.
5861 unsigned EltMaskIdx, EltIdx;
5863 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5864 if (Zeroable[EltIdx]) {
5865 // The zero vector will be on the right hand side.
5866 Mask[EltIdx] = EltIdx+4;
5870 Elt = Op->getOperand(EltIdx);
5871 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5872 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5873 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5875 Mask[EltIdx] = EltIdx;
5879 // Let the shuffle legalizer deal with blend operations.
5880 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5881 if (V1.getSimpleValueType() != VT)
5882 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5883 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5886 // See if we can lower this build_vector to a INSERTPS.
5887 if (!Subtarget->hasSSE41())
5890 SDValue V2 = Elt.getOperand(0);
5891 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5894 bool CanFold = true;
5895 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5899 SDValue Current = Op->getOperand(i);
5900 SDValue SrcVector = Current->getOperand(0);
5903 CanFold = SrcVector == V1 &&
5904 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5910 assert(V1.getNode() && "Expected at least two non-zero elements!");
5911 if (V1.getSimpleValueType() != MVT::v4f32)
5912 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5913 if (V2.getSimpleValueType() != MVT::v4f32)
5914 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5916 // Ok, we can emit an INSERTPS instruction.
5918 for (int i = 0; i < 4; ++i)
5922 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5923 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5924 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
5925 DAG.getIntPtrConstant(InsertPSMask));
5926 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
5929 /// getVShift - Return a vector logical shift node.
5931 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5932 unsigned NumBits, SelectionDAG &DAG,
5933 const TargetLowering &TLI, SDLoc dl) {
5934 assert(VT.is128BitVector() && "Unknown type for VShift");
5935 EVT ShVT = MVT::v2i64;
5936 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5937 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
5938 return DAG.getNode(ISD::BITCAST, dl, VT,
5939 DAG.getNode(Opc, dl, ShVT, SrcOp,
5940 DAG.getConstant(NumBits,
5941 TLI.getScalarShiftAmountTy(SrcOp.getValueType()))));
5945 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5947 // Check if the scalar load can be widened into a vector load. And if
5948 // the address is "base + cst" see if the cst can be "absorbed" into
5949 // the shuffle mask.
5950 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5951 SDValue Ptr = LD->getBasePtr();
5952 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5954 EVT PVT = LD->getValueType(0);
5955 if (PVT != MVT::i32 && PVT != MVT::f32)
5960 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5961 FI = FINode->getIndex();
5963 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5964 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5965 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5966 Offset = Ptr.getConstantOperandVal(1);
5967 Ptr = Ptr.getOperand(0);
5972 // FIXME: 256-bit vector instructions don't require a strict alignment,
5973 // improve this code to support it better.
5974 unsigned RequiredAlign = VT.getSizeInBits()/8;
5975 SDValue Chain = LD->getChain();
5976 // Make sure the stack object alignment is at least 16 or 32.
5977 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5978 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5979 if (MFI->isFixedObjectIndex(FI)) {
5980 // Can't change the alignment. FIXME: It's possible to compute
5981 // the exact stack offset and reference FI + adjust offset instead.
5982 // If someone *really* cares about this. That's the way to implement it.
5985 MFI->setObjectAlignment(FI, RequiredAlign);
5989 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5990 // Ptr + (Offset & ~15).
5993 if ((Offset % RequiredAlign) & 3)
5995 int64_t StartOffset = Offset & ~(RequiredAlign-1);
5997 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
5998 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
6000 int EltNo = (Offset - StartOffset) >> 2;
6001 unsigned NumElems = VT.getVectorNumElements();
6003 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
6004 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
6005 LD->getPointerInfo().getWithOffset(StartOffset),
6006 false, false, false, 0);
6008 SmallVector<int, 8> Mask;
6009 for (unsigned i = 0; i != NumElems; ++i)
6010 Mask.push_back(EltNo);
6012 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
6018 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
6019 /// vector of type 'VT', see if the elements can be replaced by a single large
6020 /// load which has the same value as a build_vector whose operands are 'elts'.
6022 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
6024 /// FIXME: we'd also like to handle the case where the last elements are zero
6025 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
6026 /// There's even a handy isZeroNode for that purpose.
6027 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
6028 SDLoc &DL, SelectionDAG &DAG,
6029 bool isAfterLegalize) {
6030 EVT EltVT = VT.getVectorElementType();
6031 unsigned NumElems = Elts.size();
6033 LoadSDNode *LDBase = nullptr;
6034 unsigned LastLoadedElt = -1U;
6036 // For each element in the initializer, see if we've found a load or an undef.
6037 // If we don't find an initial load element, or later load elements are
6038 // non-consecutive, bail out.
6039 for (unsigned i = 0; i < NumElems; ++i) {
6040 SDValue Elt = Elts[i];
6042 if (!Elt.getNode() ||
6043 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
6046 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
6048 LDBase = cast<LoadSDNode>(Elt.getNode());
6052 if (Elt.getOpcode() == ISD::UNDEF)
6055 LoadSDNode *LD = cast<LoadSDNode>(Elt);
6056 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
6061 // If we have found an entire vector of loads and undefs, then return a large
6062 // load of the entire vector width starting at the base pointer. If we found
6063 // consecutive loads for the low half, generate a vzext_load node.
6064 if (LastLoadedElt == NumElems - 1) {
6066 if (isAfterLegalize &&
6067 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
6070 SDValue NewLd = SDValue();
6072 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
6073 LDBase->getPointerInfo(), LDBase->isVolatile(),
6074 LDBase->isNonTemporal(), LDBase->isInvariant(),
6075 LDBase->getAlignment());
6077 if (LDBase->hasAnyUseOfValue(1)) {
6078 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6080 SDValue(NewLd.getNode(), 1));
6081 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6082 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6083 SDValue(NewLd.getNode(), 1));
6089 //TODO: The code below fires only for for loading the low v2i32 / v2f32
6090 //of a v4i32 / v4f32. It's probably worth generalizing.
6091 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
6092 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
6093 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
6094 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
6096 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
6097 LDBase->getPointerInfo(),
6098 LDBase->getAlignment(),
6099 false/*isVolatile*/, true/*ReadMem*/,
6102 // Make sure the newly-created LOAD is in the same position as LDBase in
6103 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
6104 // update uses of LDBase's output chain to use the TokenFactor.
6105 if (LDBase->hasAnyUseOfValue(1)) {
6106 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
6107 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
6108 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
6109 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
6110 SDValue(ResNode.getNode(), 1));
6113 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
6118 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
6119 /// to generate a splat value for the following cases:
6120 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
6121 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
6122 /// a scalar load, or a constant.
6123 /// The VBROADCAST node is returned when a pattern is found,
6124 /// or SDValue() otherwise.
6125 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
6126 SelectionDAG &DAG) {
6127 // VBROADCAST requires AVX.
6128 // TODO: Splats could be generated for non-AVX CPUs using SSE
6129 // instructions, but there's less potential gain for only 128-bit vectors.
6130 if (!Subtarget->hasAVX())
6133 MVT VT = Op.getSimpleValueType();
6136 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
6137 "Unsupported vector type for broadcast.");
6142 switch (Op.getOpcode()) {
6144 // Unknown pattern found.
6147 case ISD::BUILD_VECTOR: {
6148 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
6149 BitVector UndefElements;
6150 SDValue Splat = BVOp->getSplatValue(&UndefElements);
6152 // We need a splat of a single value to use broadcast, and it doesn't
6153 // make any sense if the value is only in one element of the vector.
6154 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
6158 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6159 Ld.getOpcode() == ISD::ConstantFP);
6161 // Make sure that all of the users of a non-constant load are from the
6162 // BUILD_VECTOR node.
6163 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
6168 case ISD::VECTOR_SHUFFLE: {
6169 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6171 // Shuffles must have a splat mask where the first element is
6173 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
6176 SDValue Sc = Op.getOperand(0);
6177 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
6178 Sc.getOpcode() != ISD::BUILD_VECTOR) {
6180 if (!Subtarget->hasInt256())
6183 // Use the register form of the broadcast instruction available on AVX2.
6184 if (VT.getSizeInBits() >= 256)
6185 Sc = Extract128BitVector(Sc, 0, DAG, dl);
6186 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
6189 Ld = Sc.getOperand(0);
6190 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
6191 Ld.getOpcode() == ISD::ConstantFP);
6193 // The scalar_to_vector node and the suspected
6194 // load node must have exactly one user.
6195 // Constants may have multiple users.
6197 // AVX-512 has register version of the broadcast
6198 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
6199 Ld.getValueType().getSizeInBits() >= 32;
6200 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
6207 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
6208 bool IsGE256 = (VT.getSizeInBits() >= 256);
6210 // When optimizing for size, generate up to 5 extra bytes for a broadcast
6211 // instruction to save 8 or more bytes of constant pool data.
6212 // TODO: If multiple splats are generated to load the same constant,
6213 // it may be detrimental to overall size. There needs to be a way to detect
6214 // that condition to know if this is truly a size win.
6215 const Function *F = DAG.getMachineFunction().getFunction();
6216 bool OptForSize = F->getAttributes().
6217 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
6219 // Handle broadcasting a single constant scalar from the constant pool
6221 // On Sandybridge (no AVX2), it is still better to load a constant vector
6222 // from the constant pool and not to broadcast it from a scalar.
6223 // But override that restriction when optimizing for size.
6224 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
6225 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
6226 EVT CVT = Ld.getValueType();
6227 assert(!CVT.isVector() && "Must not broadcast a vector type");
6229 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
6230 // For size optimization, also splat v2f64 and v2i64, and for size opt
6231 // with AVX2, also splat i8 and i16.
6232 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
6233 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6234 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
6235 const Constant *C = nullptr;
6236 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
6237 C = CI->getConstantIntValue();
6238 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
6239 C = CF->getConstantFPValue();
6241 assert(C && "Invalid constant type");
6243 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6244 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
6245 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
6246 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
6247 MachinePointerInfo::getConstantPool(),
6248 false, false, false, Alignment);
6250 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6254 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
6256 // Handle AVX2 in-register broadcasts.
6257 if (!IsLoad && Subtarget->hasInt256() &&
6258 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
6259 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6261 // The scalar source must be a normal load.
6265 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
6266 (Subtarget->hasVLX() && ScalarSize == 64))
6267 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6269 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
6270 // double since there is no vbroadcastsd xmm
6271 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
6272 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
6273 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
6276 // Unsupported broadcast.
6280 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
6281 /// underlying vector and index.
6283 /// Modifies \p ExtractedFromVec to the real vector and returns the real
6285 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
6287 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
6288 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
6291 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
6293 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
6295 // (extract_vector_elt (vector_shuffle<2,u,u,u>
6296 // (extract_subvector (v8f32 %vreg0), Constant<4>),
6299 // In this case the vector is the extract_subvector expression and the index
6300 // is 2, as specified by the shuffle.
6301 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
6302 SDValue ShuffleVec = SVOp->getOperand(0);
6303 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
6304 assert(ShuffleVecVT.getVectorElementType() ==
6305 ExtractedFromVec.getSimpleValueType().getVectorElementType());
6307 int ShuffleIdx = SVOp->getMaskElt(Idx);
6308 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
6309 ExtractedFromVec = ShuffleVec;
6315 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
6316 MVT VT = Op.getSimpleValueType();
6318 // Skip if insert_vec_elt is not supported.
6319 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6320 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
6324 unsigned NumElems = Op.getNumOperands();
6328 SmallVector<unsigned, 4> InsertIndices;
6329 SmallVector<int, 8> Mask(NumElems, -1);
6331 for (unsigned i = 0; i != NumElems; ++i) {
6332 unsigned Opc = Op.getOperand(i).getOpcode();
6334 if (Opc == ISD::UNDEF)
6337 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
6338 // Quit if more than 1 elements need inserting.
6339 if (InsertIndices.size() > 1)
6342 InsertIndices.push_back(i);
6346 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
6347 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
6348 // Quit if non-constant index.
6349 if (!isa<ConstantSDNode>(ExtIdx))
6351 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
6353 // Quit if extracted from vector of different type.
6354 if (ExtractedFromVec.getValueType() != VT)
6357 if (!VecIn1.getNode())
6358 VecIn1 = ExtractedFromVec;
6359 else if (VecIn1 != ExtractedFromVec) {
6360 if (!VecIn2.getNode())
6361 VecIn2 = ExtractedFromVec;
6362 else if (VecIn2 != ExtractedFromVec)
6363 // Quit if more than 2 vectors to shuffle
6367 if (ExtractedFromVec == VecIn1)
6369 else if (ExtractedFromVec == VecIn2)
6370 Mask[i] = Idx + NumElems;
6373 if (!VecIn1.getNode())
6376 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
6377 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
6378 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
6379 unsigned Idx = InsertIndices[i];
6380 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
6381 DAG.getIntPtrConstant(Idx));
6387 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
6389 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
6391 MVT VT = Op.getSimpleValueType();
6392 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
6393 "Unexpected type in LowerBUILD_VECTORvXi1!");
6396 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6397 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
6398 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6399 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6402 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
6403 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
6404 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
6405 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
6408 bool AllContants = true;
6409 uint64_t Immediate = 0;
6410 int NonConstIdx = -1;
6411 bool IsSplat = true;
6412 unsigned NumNonConsts = 0;
6413 unsigned NumConsts = 0;
6414 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
6415 SDValue In = Op.getOperand(idx);
6416 if (In.getOpcode() == ISD::UNDEF)
6418 if (!isa<ConstantSDNode>(In)) {
6419 AllContants = false;
6424 if (cast<ConstantSDNode>(In)->getZExtValue())
6425 Immediate |= (1ULL << idx);
6427 if (In != Op.getOperand(0))
6432 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
6433 DAG.getConstant(Immediate, MVT::i16));
6434 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
6435 DAG.getIntPtrConstant(0));
6438 if (NumNonConsts == 1 && NonConstIdx != 0) {
6441 SDValue VecAsImm = DAG.getConstant(Immediate,
6442 MVT::getIntegerVT(VT.getSizeInBits()));
6443 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
6446 DstVec = DAG.getUNDEF(VT);
6447 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
6448 Op.getOperand(NonConstIdx),
6449 DAG.getIntPtrConstant(NonConstIdx));
6451 if (!IsSplat && (NonConstIdx != 0))
6452 llvm_unreachable("Unsupported BUILD_VECTOR operation");
6453 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
6456 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6457 DAG.getConstant(-1, SelectVT),
6458 DAG.getConstant(0, SelectVT));
6460 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
6461 DAG.getConstant((Immediate | 1), SelectVT),
6462 DAG.getConstant(Immediate, SelectVT));
6463 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
6466 /// \brief Return true if \p N implements a horizontal binop and return the
6467 /// operands for the horizontal binop into V0 and V1.
6469 /// This is a helper function of PerformBUILD_VECTORCombine.
6470 /// This function checks that the build_vector \p N in input implements a
6471 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
6472 /// operation to match.
6473 /// For example, if \p Opcode is equal to ISD::ADD, then this function
6474 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
6475 /// is equal to ISD::SUB, then this function checks if this is a horizontal
6478 /// This function only analyzes elements of \p N whose indices are
6479 /// in range [BaseIdx, LastIdx).
6480 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
6482 unsigned BaseIdx, unsigned LastIdx,
6483 SDValue &V0, SDValue &V1) {
6484 EVT VT = N->getValueType(0);
6486 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
6487 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
6488 "Invalid Vector in input!");
6490 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
6491 bool CanFold = true;
6492 unsigned ExpectedVExtractIdx = BaseIdx;
6493 unsigned NumElts = LastIdx - BaseIdx;
6494 V0 = DAG.getUNDEF(VT);
6495 V1 = DAG.getUNDEF(VT);
6497 // Check if N implements a horizontal binop.
6498 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
6499 SDValue Op = N->getOperand(i + BaseIdx);
6502 if (Op->getOpcode() == ISD::UNDEF) {
6503 // Update the expected vector extract index.
6504 if (i * 2 == NumElts)
6505 ExpectedVExtractIdx = BaseIdx;
6506 ExpectedVExtractIdx += 2;
6510 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
6515 SDValue Op0 = Op.getOperand(0);
6516 SDValue Op1 = Op.getOperand(1);
6518 // Try to match the following pattern:
6519 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
6520 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6521 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6522 Op0.getOperand(0) == Op1.getOperand(0) &&
6523 isa<ConstantSDNode>(Op0.getOperand(1)) &&
6524 isa<ConstantSDNode>(Op1.getOperand(1)));
6528 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6529 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
6531 if (i * 2 < NumElts) {
6532 if (V0.getOpcode() == ISD::UNDEF)
6533 V0 = Op0.getOperand(0);
6535 if (V1.getOpcode() == ISD::UNDEF)
6536 V1 = Op0.getOperand(0);
6537 if (i * 2 == NumElts)
6538 ExpectedVExtractIdx = BaseIdx;
6541 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
6542 if (I0 == ExpectedVExtractIdx)
6543 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
6544 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
6545 // Try to match the following dag sequence:
6546 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
6547 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
6551 ExpectedVExtractIdx += 2;
6557 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
6558 /// a concat_vector.
6560 /// This is a helper function of PerformBUILD_VECTORCombine.
6561 /// This function expects two 256-bit vectors called V0 and V1.
6562 /// At first, each vector is split into two separate 128-bit vectors.
6563 /// Then, the resulting 128-bit vectors are used to implement two
6564 /// horizontal binary operations.
6566 /// The kind of horizontal binary operation is defined by \p X86Opcode.
6568 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
6569 /// the two new horizontal binop.
6570 /// When Mode is set, the first horizontal binop dag node would take as input
6571 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
6572 /// horizontal binop dag node would take as input the lower 128-bit of V1
6573 /// and the upper 128-bit of V1.
6575 /// HADD V0_LO, V0_HI
6576 /// HADD V1_LO, V1_HI
6578 /// Otherwise, the first horizontal binop dag node takes as input the lower
6579 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
6580 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
6582 /// HADD V0_LO, V1_LO
6583 /// HADD V0_HI, V1_HI
6585 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
6586 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
6587 /// the upper 128-bits of the result.
6588 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
6589 SDLoc DL, SelectionDAG &DAG,
6590 unsigned X86Opcode, bool Mode,
6591 bool isUndefLO, bool isUndefHI) {
6592 EVT VT = V0.getValueType();
6593 assert(VT.is256BitVector() && VT == V1.getValueType() &&
6594 "Invalid nodes in input!");
6596 unsigned NumElts = VT.getVectorNumElements();
6597 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
6598 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
6599 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
6600 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
6601 EVT NewVT = V0_LO.getValueType();
6603 SDValue LO = DAG.getUNDEF(NewVT);
6604 SDValue HI = DAG.getUNDEF(NewVT);
6607 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6608 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
6609 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
6610 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
6611 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
6613 // Don't emit a horizontal binop if the result is expected to be UNDEF.
6614 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
6615 V1_LO->getOpcode() != ISD::UNDEF))
6616 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
6618 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
6619 V1_HI->getOpcode() != ISD::UNDEF))
6620 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
6623 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
6626 /// \brief Try to fold a build_vector that performs an 'addsub' into the
6627 /// sequence of 'vadd + vsub + blendi'.
6628 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
6629 const X86Subtarget *Subtarget) {
6631 EVT VT = BV->getValueType(0);
6632 unsigned NumElts = VT.getVectorNumElements();
6633 SDValue InVec0 = DAG.getUNDEF(VT);
6634 SDValue InVec1 = DAG.getUNDEF(VT);
6636 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
6637 VT == MVT::v2f64) && "build_vector with an invalid type found!");
6639 // Odd-numbered elements in the input build vector are obtained from
6640 // adding two integer/float elements.
6641 // Even-numbered elements in the input build vector are obtained from
6642 // subtracting two integer/float elements.
6643 unsigned ExpectedOpcode = ISD::FSUB;
6644 unsigned NextExpectedOpcode = ISD::FADD;
6645 bool AddFound = false;
6646 bool SubFound = false;
6648 for (unsigned i = 0, e = NumElts; i != e; i++) {
6649 SDValue Op = BV->getOperand(i);
6651 // Skip 'undef' values.
6652 unsigned Opcode = Op.getOpcode();
6653 if (Opcode == ISD::UNDEF) {
6654 std::swap(ExpectedOpcode, NextExpectedOpcode);
6658 // Early exit if we found an unexpected opcode.
6659 if (Opcode != ExpectedOpcode)
6662 SDValue Op0 = Op.getOperand(0);
6663 SDValue Op1 = Op.getOperand(1);
6665 // Try to match the following pattern:
6666 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
6667 // Early exit if we cannot match that sequence.
6668 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6669 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
6670 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
6671 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
6672 Op0.getOperand(1) != Op1.getOperand(1))
6675 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
6679 // We found a valid add/sub node. Update the information accordingly.
6685 // Update InVec0 and InVec1.
6686 if (InVec0.getOpcode() == ISD::UNDEF)
6687 InVec0 = Op0.getOperand(0);
6688 if (InVec1.getOpcode() == ISD::UNDEF)
6689 InVec1 = Op1.getOperand(0);
6691 // Make sure that operands in input to each add/sub node always
6692 // come from a same pair of vectors.
6693 if (InVec0 != Op0.getOperand(0)) {
6694 if (ExpectedOpcode == ISD::FSUB)
6697 // FADD is commutable. Try to commute the operands
6698 // and then test again.
6699 std::swap(Op0, Op1);
6700 if (InVec0 != Op0.getOperand(0))
6704 if (InVec1 != Op1.getOperand(0))
6707 // Update the pair of expected opcodes.
6708 std::swap(ExpectedOpcode, NextExpectedOpcode);
6711 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6712 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6713 InVec1.getOpcode() != ISD::UNDEF)
6714 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6719 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
6720 const X86Subtarget *Subtarget) {
6722 EVT VT = N->getValueType(0);
6723 unsigned NumElts = VT.getVectorNumElements();
6724 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
6725 SDValue InVec0, InVec1;
6727 // Try to match an ADDSUB.
6728 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
6729 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
6730 SDValue Value = matchAddSub(BV, DAG, Subtarget);
6731 if (Value.getNode())
6735 // Try to match horizontal ADD/SUB.
6736 unsigned NumUndefsLO = 0;
6737 unsigned NumUndefsHI = 0;
6738 unsigned Half = NumElts/2;
6740 // Count the number of UNDEF operands in the build_vector in input.
6741 for (unsigned i = 0, e = Half; i != e; ++i)
6742 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6745 for (unsigned i = Half, e = NumElts; i != e; ++i)
6746 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6749 // Early exit if this is either a build_vector of all UNDEFs or all the
6750 // operands but one are UNDEF.
6751 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6754 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6755 // Try to match an SSE3 float HADD/HSUB.
6756 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6757 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6759 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6760 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6761 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6762 // Try to match an SSSE3 integer HADD/HSUB.
6763 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6764 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6766 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6767 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6770 if (!Subtarget->hasAVX())
6773 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6774 // Try to match an AVX horizontal add/sub of packed single/double
6775 // precision floating point values from 256-bit vectors.
6776 SDValue InVec2, InVec3;
6777 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6778 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6779 ((InVec0.getOpcode() == ISD::UNDEF ||
6780 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6781 ((InVec1.getOpcode() == ISD::UNDEF ||
6782 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6783 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6785 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6786 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6787 ((InVec0.getOpcode() == ISD::UNDEF ||
6788 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6789 ((InVec1.getOpcode() == ISD::UNDEF ||
6790 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6791 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6792 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6793 // Try to match an AVX2 horizontal add/sub of signed integers.
6794 SDValue InVec2, InVec3;
6796 bool CanFold = true;
6798 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6799 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6800 ((InVec0.getOpcode() == ISD::UNDEF ||
6801 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6802 ((InVec1.getOpcode() == ISD::UNDEF ||
6803 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6804 X86Opcode = X86ISD::HADD;
6805 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6806 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6807 ((InVec0.getOpcode() == ISD::UNDEF ||
6808 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6809 ((InVec1.getOpcode() == ISD::UNDEF ||
6810 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6811 X86Opcode = X86ISD::HSUB;
6816 // Fold this build_vector into a single horizontal add/sub.
6817 // Do this only if the target has AVX2.
6818 if (Subtarget->hasAVX2())
6819 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6821 // Do not try to expand this build_vector into a pair of horizontal
6822 // add/sub if we can emit a pair of scalar add/sub.
6823 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6826 // Convert this build_vector into a pair of horizontal binop followed by
6828 bool isUndefLO = NumUndefsLO == Half;
6829 bool isUndefHI = NumUndefsHI == Half;
6830 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6831 isUndefLO, isUndefHI);
6835 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6836 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6838 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6839 X86Opcode = X86ISD::HADD;
6840 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6841 X86Opcode = X86ISD::HSUB;
6842 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6843 X86Opcode = X86ISD::FHADD;
6844 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6845 X86Opcode = X86ISD::FHSUB;
6849 // Don't try to expand this build_vector into a pair of horizontal add/sub
6850 // if we can simply emit a pair of scalar add/sub.
6851 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6854 // Convert this build_vector into two horizontal add/sub followed by
6856 bool isUndefLO = NumUndefsLO == Half;
6857 bool isUndefHI = NumUndefsHI == Half;
6858 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6859 isUndefLO, isUndefHI);
6866 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6869 MVT VT = Op.getSimpleValueType();
6870 MVT ExtVT = VT.getVectorElementType();
6871 unsigned NumElems = Op.getNumOperands();
6873 // Generate vectors for predicate vectors.
6874 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6875 return LowerBUILD_VECTORvXi1(Op, DAG);
6877 // Vectors containing all zeros can be matched by pxor and xorps later
6878 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6879 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6880 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6881 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6884 return getZeroVector(VT, Subtarget, DAG, dl);
6887 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6888 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6889 // vpcmpeqd on 256-bit vectors.
6890 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6891 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6894 if (!VT.is512BitVector())
6895 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
6898 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
6899 if (Broadcast.getNode())
6902 unsigned EVTBits = ExtVT.getSizeInBits();
6904 unsigned NumZero = 0;
6905 unsigned NumNonZero = 0;
6906 unsigned NonZeros = 0;
6907 bool IsAllConstants = true;
6908 SmallSet<SDValue, 8> Values;
6909 for (unsigned i = 0; i < NumElems; ++i) {
6910 SDValue Elt = Op.getOperand(i);
6911 if (Elt.getOpcode() == ISD::UNDEF)
6914 if (Elt.getOpcode() != ISD::Constant &&
6915 Elt.getOpcode() != ISD::ConstantFP)
6916 IsAllConstants = false;
6917 if (X86::isZeroNode(Elt))
6920 NonZeros |= (1 << i);
6925 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6926 if (NumNonZero == 0)
6927 return DAG.getUNDEF(VT);
6929 // Special case for single non-zero, non-undef, element.
6930 if (NumNonZero == 1) {
6931 unsigned Idx = countTrailingZeros(NonZeros);
6932 SDValue Item = Op.getOperand(Idx);
6934 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6935 // the value are obviously zero, truncate the value to i32 and do the
6936 // insertion that way. Only do this if the value is non-constant or if the
6937 // value is a constant being inserted into element 0. It is cheaper to do
6938 // a constant pool load than it is to do a movd + shuffle.
6939 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6940 (!IsAllConstants || Idx == 0)) {
6941 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6943 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6944 EVT VecVT = MVT::v4i32;
6945 unsigned VecElts = 4;
6947 // Truncate the value (which may itself be a constant) to i32, and
6948 // convert it to a vector with movd (S2V+shuffle to zero extend).
6949 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6950 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6952 // If using the new shuffle lowering, just directly insert this.
6953 if (ExperimentalVectorShuffleLowering)
6955 ISD::BITCAST, dl, VT,
6956 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
6958 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6960 // Now we have our 32-bit value zero extended in the low element of
6961 // a vector. If Idx != 0, swizzle it into place.
6963 SmallVector<int, 4> Mask;
6964 Mask.push_back(Idx);
6965 for (unsigned i = 1; i != VecElts; ++i)
6967 Item = DAG.getVectorShuffle(VecVT, dl, Item, DAG.getUNDEF(VecVT),
6970 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
6974 // If we have a constant or non-constant insertion into the low element of
6975 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6976 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6977 // depending on what the source datatype is.
6980 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6982 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6983 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6984 if (VT.is256BitVector() || VT.is512BitVector()) {
6985 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6986 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6987 Item, DAG.getIntPtrConstant(0));
6989 assert(VT.is128BitVector() && "Expected an SSE value type!");
6990 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6991 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6992 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6995 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6996 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6997 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6998 if (VT.is256BitVector()) {
6999 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
7000 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
7002 assert(VT.is128BitVector() && "Expected an SSE value type!");
7003 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
7005 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
7009 // Is it a vector logical left shift?
7010 if (NumElems == 2 && Idx == 1 &&
7011 X86::isZeroNode(Op.getOperand(0)) &&
7012 !X86::isZeroNode(Op.getOperand(1))) {
7013 unsigned NumBits = VT.getSizeInBits();
7014 return getVShift(true, VT,
7015 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7016 VT, Op.getOperand(1)),
7017 NumBits/2, DAG, *this, dl);
7020 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
7023 // Otherwise, if this is a vector with i32 or f32 elements, and the element
7024 // is a non-constant being inserted into an element other than the low one,
7025 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
7026 // movd/movss) to move this into the low element, then shuffle it into
7028 if (EVTBits == 32) {
7029 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
7031 // If using the new shuffle lowering, just directly insert this.
7032 if (ExperimentalVectorShuffleLowering)
7033 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
7035 // Turn it into a shuffle of zero and zero-extended scalar to vector.
7036 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0, Subtarget, DAG);
7037 SmallVector<int, 8> MaskVec;
7038 for (unsigned i = 0; i != NumElems; ++i)
7039 MaskVec.push_back(i == Idx ? 0 : 1);
7040 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
7044 // Splat is obviously ok. Let legalizer expand it to a shuffle.
7045 if (Values.size() == 1) {
7046 if (EVTBits == 32) {
7047 // Instead of a shuffle like this:
7048 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
7049 // Check if it's possible to issue this instead.
7050 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
7051 unsigned Idx = countTrailingZeros(NonZeros);
7052 SDValue Item = Op.getOperand(Idx);
7053 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
7054 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
7059 // A vector full of immediates; various special cases are already
7060 // handled, so this is best done with a single constant-pool load.
7064 // For AVX-length vectors, see if we can use a vector load to get all of the
7065 // elements, otherwise build the individual 128-bit pieces and use
7066 // shuffles to put them in place.
7067 if (VT.is256BitVector() || VT.is512BitVector()) {
7068 SmallVector<SDValue, 64> V;
7069 for (unsigned i = 0; i != NumElems; ++i)
7070 V.push_back(Op.getOperand(i));
7072 // Check for a build vector of consecutive loads.
7073 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
7076 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
7078 // Build both the lower and upper subvector.
7079 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7080 makeArrayRef(&V[0], NumElems/2));
7081 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
7082 makeArrayRef(&V[NumElems / 2], NumElems/2));
7084 // Recreate the wider vector with the lower and upper part.
7085 if (VT.is256BitVector())
7086 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7087 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
7090 // Let legalizer expand 2-wide build_vectors.
7091 if (EVTBits == 64) {
7092 if (NumNonZero == 1) {
7093 // One half is zero or undef.
7094 unsigned Idx = countTrailingZeros(NonZeros);
7095 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
7096 Op.getOperand(Idx));
7097 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
7102 // If element VT is < 32 bits, convert it to inserts into a zero vector.
7103 if (EVTBits == 8 && NumElems == 16) {
7104 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
7106 if (V.getNode()) return V;
7109 if (EVTBits == 16 && NumElems == 8) {
7110 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
7112 if (V.getNode()) return V;
7115 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
7116 if (EVTBits == 32 && NumElems == 4) {
7117 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
7122 // If element VT is == 32 bits, turn it into a number of shuffles.
7123 SmallVector<SDValue, 8> V(NumElems);
7124 if (NumElems == 4 && NumZero > 0) {
7125 for (unsigned i = 0; i < 4; ++i) {
7126 bool isZero = !(NonZeros & (1 << i));
7128 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
7130 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7133 for (unsigned i = 0; i < 2; ++i) {
7134 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
7137 V[i] = V[i*2]; // Must be a zero vector.
7140 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
7143 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
7146 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
7151 bool Reverse1 = (NonZeros & 0x3) == 2;
7152 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
7156 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
7157 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
7159 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
7162 if (Values.size() > 1 && VT.is128BitVector()) {
7163 // Check for a build vector of consecutive loads.
7164 for (unsigned i = 0; i < NumElems; ++i)
7165 V[i] = Op.getOperand(i);
7167 // Check for elements which are consecutive loads.
7168 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
7172 // Check for a build vector from mostly shuffle plus few inserting.
7173 SDValue Sh = buildFromShuffleMostly(Op, DAG);
7177 // For SSE 4.1, use insertps to put the high elements into the low element.
7178 if (getSubtarget()->hasSSE41()) {
7180 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
7181 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
7183 Result = DAG.getUNDEF(VT);
7185 for (unsigned i = 1; i < NumElems; ++i) {
7186 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
7187 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
7188 Op.getOperand(i), DAG.getIntPtrConstant(i));
7193 // Otherwise, expand into a number of unpckl*, start by extending each of
7194 // our (non-undef) elements to the full vector width with the element in the
7195 // bottom slot of the vector (which generates no code for SSE).
7196 for (unsigned i = 0; i < NumElems; ++i) {
7197 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
7198 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
7200 V[i] = DAG.getUNDEF(VT);
7203 // Next, we iteratively mix elements, e.g. for v4f32:
7204 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
7205 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
7206 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
7207 unsigned EltStride = NumElems >> 1;
7208 while (EltStride != 0) {
7209 for (unsigned i = 0; i < EltStride; ++i) {
7210 // If V[i+EltStride] is undef and this is the first round of mixing,
7211 // then it is safe to just drop this shuffle: V[i] is already in the
7212 // right place, the one element (since it's the first round) being
7213 // inserted as undef can be dropped. This isn't safe for successive
7214 // rounds because they will permute elements within both vectors.
7215 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
7216 EltStride == NumElems/2)
7219 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
7228 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
7229 // to create 256-bit vectors from two other 128-bit ones.
7230 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7232 MVT ResVT = Op.getSimpleValueType();
7234 assert((ResVT.is256BitVector() ||
7235 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
7237 SDValue V1 = Op.getOperand(0);
7238 SDValue V2 = Op.getOperand(1);
7239 unsigned NumElems = ResVT.getVectorNumElements();
7240 if(ResVT.is256BitVector())
7241 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7243 if (Op.getNumOperands() == 4) {
7244 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
7245 ResVT.getVectorNumElements()/2);
7246 SDValue V3 = Op.getOperand(2);
7247 SDValue V4 = Op.getOperand(3);
7248 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
7249 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
7251 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
7254 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
7255 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
7256 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
7257 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
7258 Op.getNumOperands() == 4)));
7260 // AVX can use the vinsertf128 instruction to create 256-bit vectors
7261 // from two other 128-bit ones.
7263 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
7264 return LowerAVXCONCAT_VECTORS(Op, DAG);
7268 //===----------------------------------------------------------------------===//
7269 // Vector shuffle lowering
7271 // This is an experimental code path for lowering vector shuffles on x86. It is
7272 // designed to handle arbitrary vector shuffles and blends, gracefully
7273 // degrading performance as necessary. It works hard to recognize idiomatic
7274 // shuffles and lower them to optimal instruction patterns without leaving
7275 // a framework that allows reasonably efficient handling of all vector shuffle
7277 //===----------------------------------------------------------------------===//
7279 /// \brief Tiny helper function to identify a no-op mask.
7281 /// This is a somewhat boring predicate function. It checks whether the mask
7282 /// array input, which is assumed to be a single-input shuffle mask of the kind
7283 /// used by the X86 shuffle instructions (not a fully general
7284 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
7285 /// in-place shuffle are 'no-op's.
7286 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
7287 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7288 if (Mask[i] != -1 && Mask[i] != i)
7293 /// \brief Helper function to classify a mask as a single-input mask.
7295 /// This isn't a generic single-input test because in the vector shuffle
7296 /// lowering we canonicalize single inputs to be the first input operand. This
7297 /// means we can more quickly test for a single input by only checking whether
7298 /// an input from the second operand exists. We also assume that the size of
7299 /// mask corresponds to the size of the input vectors which isn't true in the
7300 /// fully general case.
7301 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
7303 if (M >= (int)Mask.size())
7308 /// \brief Test whether there are elements crossing 128-bit lanes in this
7311 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
7312 /// and we routinely test for these.
7313 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
7314 int LaneSize = 128 / VT.getScalarSizeInBits();
7315 int Size = Mask.size();
7316 for (int i = 0; i < Size; ++i)
7317 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
7322 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
7324 /// This checks a shuffle mask to see if it is performing the same
7325 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
7326 /// that it is also not lane-crossing. It may however involve a blend from the
7327 /// same lane of a second vector.
7329 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
7330 /// non-trivial to compute in the face of undef lanes. The representation is
7331 /// *not* suitable for use with existing 128-bit shuffles as it will contain
7332 /// entries from both V1 and V2 inputs to the wider mask.
7334 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
7335 SmallVectorImpl<int> &RepeatedMask) {
7336 int LaneSize = 128 / VT.getScalarSizeInBits();
7337 RepeatedMask.resize(LaneSize, -1);
7338 int Size = Mask.size();
7339 for (int i = 0; i < Size; ++i) {
7342 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
7343 // This entry crosses lanes, so there is no way to model this shuffle.
7346 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
7347 if (RepeatedMask[i % LaneSize] == -1)
7348 // This is the first non-undef entry in this slot of a 128-bit lane.
7349 RepeatedMask[i % LaneSize] =
7350 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
7351 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
7352 // Found a mismatch with the repeated mask.
7358 // Hide this symbol with an anonymous namespace instead of 'static' so that MSVC
7359 // 2013 will allow us to use it as a non-type template parameter.
7362 /// \brief Implementation of the \c isShuffleEquivalent variadic functor.
7364 /// See its documentation for details.
7365 bool isShuffleEquivalentImpl(ArrayRef<int> Mask, ArrayRef<const int *> Args) {
7366 if (Mask.size() != Args.size())
7368 for (int i = 0, e = Mask.size(); i < e; ++i) {
7369 assert(*Args[i] >= 0 && "Arguments must be positive integers!");
7370 if (Mask[i] != -1 && Mask[i] != *Args[i])
7378 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
7381 /// This is a fast way to test a shuffle mask against a fixed pattern:
7383 /// if (isShuffleEquivalent(Mask, 3, 2, 1, 0)) { ... }
7385 /// It returns true if the mask is exactly as wide as the argument list, and
7386 /// each element of the mask is either -1 (signifying undef) or the value given
7387 /// in the argument.
7388 static const VariadicFunction1<
7389 bool, ArrayRef<int>, int, isShuffleEquivalentImpl> isShuffleEquivalent = {};
7391 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
7393 /// This helper function produces an 8-bit shuffle immediate corresponding to
7394 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
7395 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
7398 /// NB: We rely heavily on "undef" masks preserving the input lane.
7399 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
7400 SelectionDAG &DAG) {
7401 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
7402 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
7403 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
7404 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
7405 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
7408 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
7409 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
7410 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
7411 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
7412 return DAG.getConstant(Imm, MVT::i8);
7415 /// \brief Try to emit a blend instruction for a shuffle.
7417 /// This doesn't do any checks for the availability of instructions for blending
7418 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
7419 /// be matched in the backend with the type given. What it does check for is
7420 /// that the shuffle mask is in fact a blend.
7421 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
7422 SDValue V2, ArrayRef<int> Mask,
7423 const X86Subtarget *Subtarget,
7424 SelectionDAG &DAG) {
7426 unsigned BlendMask = 0;
7427 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7428 if (Mask[i] >= Size) {
7429 if (Mask[i] != i + Size)
7430 return SDValue(); // Shuffled V2 input!
7431 BlendMask |= 1u << i;
7434 if (Mask[i] >= 0 && Mask[i] != i)
7435 return SDValue(); // Shuffled V1 input!
7437 switch (VT.SimpleTy) {
7442 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
7443 DAG.getConstant(BlendMask, MVT::i8));
7447 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7451 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
7452 // that instruction.
7453 if (Subtarget->hasAVX2()) {
7454 // Scale the blend by the number of 32-bit dwords per element.
7455 int Scale = VT.getScalarSizeInBits() / 32;
7457 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7458 if (Mask[i] >= Size)
7459 for (int j = 0; j < Scale; ++j)
7460 BlendMask |= 1u << (i * Scale + j);
7462 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
7463 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
7464 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
7465 return DAG.getNode(ISD::BITCAST, DL, VT,
7466 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
7467 DAG.getConstant(BlendMask, MVT::i8)));
7471 // For integer shuffles we need to expand the mask and cast the inputs to
7472 // v8i16s prior to blending.
7473 int Scale = 8 / VT.getVectorNumElements();
7475 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7476 if (Mask[i] >= Size)
7477 for (int j = 0; j < Scale; ++j)
7478 BlendMask |= 1u << (i * Scale + j);
7480 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
7481 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
7482 return DAG.getNode(ISD::BITCAST, DL, VT,
7483 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
7484 DAG.getConstant(BlendMask, MVT::i8)));
7488 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7489 SmallVector<int, 8> RepeatedMask;
7490 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
7491 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
7492 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
7494 for (int i = 0; i < 8; ++i)
7495 if (RepeatedMask[i] >= 16)
7496 BlendMask |= 1u << i;
7497 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
7498 DAG.getConstant(BlendMask, MVT::i8));
7503 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
7504 // Scale the blend by the number of bytes per element.
7505 int Scale = VT.getScalarSizeInBits() / 8;
7506 assert(Mask.size() * Scale == 32 && "Not a 256-bit vector!");
7508 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7509 // mix of LLVM's code generator and the x86 backend. We tell the code
7510 // generator that boolean values in the elements of an x86 vector register
7511 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7512 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7513 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7514 // of the element (the remaining are ignored) and 0 in that high bit would
7515 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7516 // the LLVM model for boolean values in vector elements gets the relevant
7517 // bit set, it is set backwards and over constrained relative to x86's
7519 SDValue VSELECTMask[32];
7520 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7521 for (int j = 0; j < Scale; ++j)
7522 VSELECTMask[Scale * i + j] =
7523 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7524 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8);
7526 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1);
7527 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V2);
7529 ISD::BITCAST, DL, VT,
7530 DAG.getNode(ISD::VSELECT, DL, MVT::v32i8,
7531 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, VSELECTMask),
7536 llvm_unreachable("Not a supported integer vector type!");
7540 /// \brief Generic routine to lower a shuffle and blend as a decomposed set of
7541 /// unblended shuffles followed by an unshuffled blend.
7543 /// This matches the extremely common pattern for handling combined
7544 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7546 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7550 SelectionDAG &DAG) {
7551 // Shuffle the input elements into the desired positions in V1 and V2 and
7552 // blend them together.
7553 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7554 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7555 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7556 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7557 if (Mask[i] >= 0 && Mask[i] < Size) {
7558 V1Mask[i] = Mask[i];
7560 } else if (Mask[i] >= Size) {
7561 V2Mask[i] = Mask[i] - Size;
7562 BlendMask[i] = i + Size;
7565 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7566 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7567 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7570 /// \brief Try to lower a vector shuffle as a byte rotation.
7572 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7573 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7574 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7575 /// try to generically lower a vector shuffle through such an pattern. It
7576 /// does not check for the profitability of lowering either as PALIGNR or
7577 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7578 /// This matches shuffle vectors that look like:
7580 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7582 /// Essentially it concatenates V1 and V2, shifts right by some number of
7583 /// elements, and takes the low elements as the result. Note that while this is
7584 /// specified as a *right shift* because x86 is little-endian, it is a *left
7585 /// rotate* of the vector lanes.
7587 /// Note that this only handles 128-bit vector widths currently.
7588 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7591 const X86Subtarget *Subtarget,
7592 SelectionDAG &DAG) {
7593 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7595 // We need to detect various ways of spelling a rotation:
7596 // [11, 12, 13, 14, 15, 0, 1, 2]
7597 // [-1, 12, 13, 14, -1, -1, 1, -1]
7598 // [-1, -1, -1, -1, -1, -1, 1, 2]
7599 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7600 // [-1, 4, 5, 6, -1, -1, 9, -1]
7601 // [-1, 4, 5, 6, -1, -1, -1, -1]
7604 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7607 assert(Mask[i] >= 0 && "Only -1 is a valid negative mask element!");
7609 // Based on the mod-Size value of this mask element determine where
7610 // a rotated vector would have started.
7611 int StartIdx = i - (Mask[i] % Size);
7613 // The identity rotation isn't interesting, stop.
7616 // If we found the tail of a vector the rotation must be the missing
7617 // front. If we found the head of a vector, it must be how much of the head.
7618 int CandidateRotation = StartIdx < 0 ? -StartIdx : Size - StartIdx;
7621 Rotation = CandidateRotation;
7622 else if (Rotation != CandidateRotation)
7623 // The rotations don't match, so we can't match this mask.
7626 // Compute which value this mask is pointing at.
7627 SDValue MaskV = Mask[i] < Size ? V1 : V2;
7629 // Compute which of the two target values this index should be assigned to.
7630 // This reflects whether the high elements are remaining or the low elements
7632 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7634 // Either set up this value if we've not encountered it before, or check
7635 // that it remains consistent.
7638 else if (TargetV != MaskV)
7639 // This may be a rotation, but it pulls from the inputs in some
7640 // unsupported interleaving.
7644 // Check that we successfully analyzed the mask, and normalize the results.
7645 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7646 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7652 assert(VT.getSizeInBits() == 128 &&
7653 "Rotate-based lowering only supports 128-bit lowering!");
7654 assert(Mask.size() <= 16 &&
7655 "Can shuffle at most 16 bytes in a 128-bit vector!");
7657 // The actual rotate instruction rotates bytes, so we need to scale the
7658 // rotation based on how many bytes are in the vector.
7659 int Scale = 16 / Mask.size();
7661 // SSSE3 targets can use the palignr instruction
7662 if (Subtarget->hasSSSE3()) {
7663 // Cast the inputs to v16i8 to match PALIGNR.
7664 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Lo);
7665 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Hi);
7667 return DAG.getNode(ISD::BITCAST, DL, VT,
7668 DAG.getNode(X86ISD::PALIGNR, DL, MVT::v16i8, Hi, Lo,
7669 DAG.getConstant(Rotation * Scale, MVT::i8)));
7672 // Default SSE2 implementation
7673 int LoByteShift = 16 - Rotation * Scale;
7674 int HiByteShift = Rotation * Scale;
7676 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7677 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
7678 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
7680 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7681 DAG.getConstant(8 * LoByteShift, MVT::i8));
7682 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7683 DAG.getConstant(8 * HiByteShift, MVT::i8));
7684 return DAG.getNode(ISD::BITCAST, DL, VT,
7685 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7688 /// \brief Compute whether each element of a shuffle is zeroable.
7690 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
7691 /// Either it is an undef element in the shuffle mask, the element of the input
7692 /// referenced is undef, or the element of the input referenced is known to be
7693 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
7694 /// as many lanes with this technique as possible to simplify the remaining
7696 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
7697 SDValue V1, SDValue V2) {
7698 SmallBitVector Zeroable(Mask.size(), false);
7700 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
7701 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
7703 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7705 // Handle the easy cases.
7706 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
7711 // If this is an index into a build_vector node, dig out the input value and
7713 SDValue V = M < Size ? V1 : V2;
7714 if (V.getOpcode() != ISD::BUILD_VECTOR)
7717 SDValue Input = V.getOperand(M % Size);
7718 // The UNDEF opcode check really should be dead code here, but not quite
7719 // worth asserting on (it isn't invalid, just unexpected).
7720 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
7727 /// \brief Try to lower a vector shuffle as a byte shift (shifts in zeros).
7729 /// Attempts to match a shuffle mask against the PSRLDQ and PSLLDQ SSE2
7730 /// byte-shift instructions. The mask must consist of a shifted sequential
7731 /// shuffle from one of the input vectors and zeroable elements for the
7732 /// remaining 'shifted in' elements.
7734 /// Note that this only handles 128-bit vector widths currently.
7735 static SDValue lowerVectorShuffleAsByteShift(SDLoc DL, MVT VT, SDValue V1,
7736 SDValue V2, ArrayRef<int> Mask,
7737 SelectionDAG &DAG) {
7738 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7740 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7742 int Size = Mask.size();
7743 int Scale = 16 / Size;
7745 for (int Shift = 1; Shift < Size; Shift++) {
7746 int ByteShift = Shift * Scale;
7748 // PSRLDQ : (little-endian) right byte shift
7749 // [ 5, 6, 7, zz, zz, zz, zz, zz]
7750 // [ -1, 5, 6, 7, zz, zz, zz, zz]
7751 // [ 1, 2, -1, -1, -1, -1, zz, zz]
7752 bool ZeroableRight = true;
7753 for (int i = Size - Shift; i < Size; i++) {
7754 ZeroableRight &= Zeroable[i];
7757 if (ZeroableRight) {
7758 bool ValidShiftRight1 =
7759 isSequentialOrUndefInRange(Mask, 0, Size - Shift, Shift);
7760 bool ValidShiftRight2 =
7761 isSequentialOrUndefInRange(Mask, 0, Size - Shift, Size + Shift);
7763 if (ValidShiftRight1 || ValidShiftRight2) {
7764 // Cast the inputs to v2i64 to match PSRLDQ.
7765 SDValue &TargetV = ValidShiftRight1 ? V1 : V2;
7766 SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
7767 SDValue Shifted = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, V,
7768 DAG.getConstant(ByteShift * 8, MVT::i8));
7769 return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
7773 // PSLLDQ : (little-endian) left byte shift
7774 // [ zz, 0, 1, 2, 3, 4, 5, 6]
7775 // [ zz, zz, -1, -1, 2, 3, 4, -1]
7776 // [ zz, zz, zz, zz, zz, zz, -1, 1]
7777 bool ZeroableLeft = true;
7778 for (int i = 0; i < Shift; i++) {
7779 ZeroableLeft &= Zeroable[i];
7783 bool ValidShiftLeft1 =
7784 isSequentialOrUndefInRange(Mask, Shift, Size - Shift, 0);
7785 bool ValidShiftLeft2 =
7786 isSequentialOrUndefInRange(Mask, Shift, Size - Shift, Size);
7788 if (ValidShiftLeft1 || ValidShiftLeft2) {
7789 // Cast the inputs to v2i64 to match PSLLDQ.
7790 SDValue &TargetV = ValidShiftLeft1 ? V1 : V2;
7791 SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, TargetV);
7792 SDValue Shifted = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, V,
7793 DAG.getConstant(ByteShift * 8, MVT::i8));
7794 return DAG.getNode(ISD::BITCAST, DL, VT, Shifted);
7802 /// \brief Lower a vector shuffle as a zero or any extension.
7804 /// Given a specific number of elements, element bit width, and extension
7805 /// stride, produce either a zero or any extension based on the available
7806 /// features of the subtarget.
7807 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7808 SDLoc DL, MVT VT, int NumElements, int Scale, bool AnyExt, SDValue InputV,
7809 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7810 assert(Scale > 1 && "Need a scale to extend.");
7811 int EltBits = VT.getSizeInBits() / NumElements;
7812 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7813 "Only 8, 16, and 32 bit elements can be extended.");
7814 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7816 // Found a valid zext mask! Try various lowering strategies based on the
7817 // input type and available ISA extensions.
7818 if (Subtarget->hasSSE41()) {
7819 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7820 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7821 NumElements / Scale);
7822 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
7823 return DAG.getNode(ISD::BITCAST, DL, VT,
7824 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7827 // For any extends we can cheat for larger element sizes and use shuffle
7828 // instructions that can fold with a load and/or copy.
7829 if (AnyExt && EltBits == 32) {
7830 int PSHUFDMask[4] = {0, -1, 1, -1};
7832 ISD::BITCAST, DL, VT,
7833 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7834 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7835 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7837 if (AnyExt && EltBits == 16 && Scale > 2) {
7838 int PSHUFDMask[4] = {0, -1, 0, -1};
7839 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7840 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
7841 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
7842 int PSHUFHWMask[4] = {1, -1, -1, -1};
7844 ISD::BITCAST, DL, VT,
7845 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7846 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
7847 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
7850 // If this would require more than 2 unpack instructions to expand, use
7851 // pshufb when available. We can only use more than 2 unpack instructions
7852 // when zero extending i8 elements which also makes it easier to use pshufb.
7853 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7854 assert(NumElements == 16 && "Unexpected byte vector width!");
7855 SDValue PSHUFBMask[16];
7856 for (int i = 0; i < 16; ++i)
7858 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
7859 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
7860 return DAG.getNode(ISD::BITCAST, DL, VT,
7861 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7862 DAG.getNode(ISD::BUILD_VECTOR, DL,
7863 MVT::v16i8, PSHUFBMask)));
7866 // Otherwise emit a sequence of unpacks.
7868 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7869 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7870 : getZeroVector(InputVT, Subtarget, DAG, DL);
7871 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
7872 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7876 } while (Scale > 1);
7877 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
7880 /// \brief Try to lower a vector shuffle as a zero extension on any micrarch.
7882 /// This routine will try to do everything in its power to cleverly lower
7883 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7884 /// check for the profitability of this lowering, it tries to aggressively
7885 /// match this pattern. It will use all of the micro-architectural details it
7886 /// can to emit an efficient lowering. It handles both blends with all-zero
7887 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7888 /// masking out later).
7890 /// The reason we have dedicated lowering for zext-style shuffles is that they
7891 /// are both incredibly common and often quite performance sensitive.
7892 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7893 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7894 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7895 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7897 int Bits = VT.getSizeInBits();
7898 int NumElements = Mask.size();
7900 // Define a helper function to check a particular ext-scale and lower to it if
7902 auto Lower = [&](int Scale) -> SDValue {
7905 for (int i = 0; i < NumElements; ++i) {
7907 continue; // Valid anywhere but doesn't tell us anything.
7908 if (i % Scale != 0) {
7909 // Each of the extend elements needs to be zeroable.
7913 // We no lorger are in the anyext case.
7918 // Each of the base elements needs to be consecutive indices into the
7919 // same input vector.
7920 SDValue V = Mask[i] < NumElements ? V1 : V2;
7923 else if (InputV != V)
7924 return SDValue(); // Flip-flopping inputs.
7926 if (Mask[i] % NumElements != i / Scale)
7927 return SDValue(); // Non-consecutive strided elemenst.
7930 // If we fail to find an input, we have a zero-shuffle which should always
7931 // have already been handled.
7932 // FIXME: Maybe handle this here in case during blending we end up with one?
7936 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7937 DL, VT, NumElements, Scale, AnyExt, InputV, Subtarget, DAG);
7940 // The widest scale possible for extending is to a 64-bit integer.
7941 assert(Bits % 64 == 0 &&
7942 "The number of bits in a vector must be divisible by 64 on x86!");
7943 int NumExtElements = Bits / 64;
7945 // Each iteration, try extending the elements half as much, but into twice as
7947 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7948 assert(NumElements % NumExtElements == 0 &&
7949 "The input vector size must be divisble by the extended size.");
7950 if (SDValue V = Lower(NumElements / NumExtElements))
7954 // No viable ext lowering found.
7958 /// \brief Try to get a scalar value for a specific element of a vector.
7960 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7961 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7962 SelectionDAG &DAG) {
7963 MVT VT = V.getSimpleValueType();
7964 MVT EltVT = VT.getVectorElementType();
7965 while (V.getOpcode() == ISD::BITCAST)
7966 V = V.getOperand(0);
7967 // If the bitcasts shift the element size, we can't extract an equivalent
7969 MVT NewVT = V.getSimpleValueType();
7970 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7973 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7974 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
7975 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
7980 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7982 /// This is particularly important because the set of instructions varies
7983 /// significantly based on whether the operand is a load or not.
7984 static bool isShuffleFoldableLoad(SDValue V) {
7985 while (V.getOpcode() == ISD::BITCAST)
7986 V = V.getOperand(0);
7988 return ISD::isNON_EXTLoad(V.getNode());
7991 /// \brief Try to lower insertion of a single element into a zero vector.
7993 /// This is a common pattern that we have especially efficient patterns to lower
7994 /// across all subtarget feature sets.
7995 static SDValue lowerVectorShuffleAsElementInsertion(
7996 MVT VT, SDLoc DL, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7997 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7998 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8000 MVT EltVT = VT.getVectorElementType();
8002 int V2Index = std::find_if(Mask.begin(), Mask.end(),
8003 [&Mask](int M) { return M >= (int)Mask.size(); }) -
8005 bool IsV1Zeroable = true;
8006 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8007 if (i != V2Index && !Zeroable[i]) {
8008 IsV1Zeroable = false;
8012 // Check for a single input from a SCALAR_TO_VECTOR node.
8013 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
8014 // all the smarts here sunk into that routine. However, the current
8015 // lowering of BUILD_VECTOR makes that nearly impossible until the old
8016 // vector shuffle lowering is dead.
8017 if (SDValue V2S = getScalarValueForVectorElement(
8018 V2, Mask[V2Index] - Mask.size(), DAG)) {
8019 // We need to zext the scalar if it is smaller than an i32.
8020 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
8021 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
8022 // Using zext to expand a narrow element won't work for non-zero
8027 // Zero-extend directly to i32.
8029 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
8031 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
8032 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
8033 EltVT == MVT::i16) {
8034 // Either not inserting from the low element of the input or the input
8035 // element size is too small to use VZEXT_MOVL to clear the high bits.
8039 if (!IsV1Zeroable) {
8040 // If V1 can't be treated as a zero vector we have fewer options to lower
8041 // this. We can't support integer vectors or non-zero targets cheaply, and
8042 // the V1 elements can't be permuted in any way.
8043 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
8044 if (!VT.isFloatingPoint() || V2Index != 0)
8046 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
8047 V1Mask[V2Index] = -1;
8048 if (!isNoopShuffleMask(V1Mask))
8050 // This is essentially a special case blend operation, but if we have
8051 // general purpose blend operations, they are always faster. Bail and let
8052 // the rest of the lowering handle these as blends.
8053 if (Subtarget->hasSSE41())
8056 // Otherwise, use MOVSD or MOVSS.
8057 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
8058 "Only two types of floating point element types to handle!");
8059 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
8063 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
8065 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8068 // If we have 4 or fewer lanes we can cheaply shuffle the element into
8069 // the desired position. Otherwise it is more efficient to do a vector
8070 // shift left. We know that we can do a vector shift left because all
8071 // the inputs are zero.
8072 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
8073 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
8074 V2Shuffle[V2Index] = 0;
8075 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
8077 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
8079 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
8081 V2Index * EltVT.getSizeInBits(),
8082 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
8083 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
8089 /// \brief Try to lower broadcast of a single element.
8091 /// For convenience, this code also bundles all of the subtarget feature set
8092 /// filtering. While a little annoying to re-dispatch on type here, there isn't
8093 /// a convenient way to factor it out.
8094 static SDValue lowerVectorShuffleAsBroadcast(MVT VT, SDLoc DL, SDValue V,
8096 const X86Subtarget *Subtarget,
8097 SelectionDAG &DAG) {
8098 if (!Subtarget->hasAVX())
8100 if (VT.isInteger() && !Subtarget->hasAVX2())
8103 // Check that the mask is a broadcast.
8104 int BroadcastIdx = -1;
8106 if (M >= 0 && BroadcastIdx == -1)
8108 else if (M >= 0 && M != BroadcastIdx)
8111 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
8112 "a sorted mask where the broadcast "
8115 // Go up the chain of (vector) values to try and find a scalar load that
8116 // we can combine with the broadcast.
8118 switch (V.getOpcode()) {
8119 case ISD::CONCAT_VECTORS: {
8120 int OperandSize = Mask.size() / V.getNumOperands();
8121 V = V.getOperand(BroadcastIdx / OperandSize);
8122 BroadcastIdx %= OperandSize;
8126 case ISD::INSERT_SUBVECTOR: {
8127 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
8128 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
8132 int BeginIdx = (int)ConstantIdx->getZExtValue();
8134 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
8135 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
8136 BroadcastIdx -= BeginIdx;
8147 // Check if this is a broadcast of a scalar. We special case lowering
8148 // for scalars so that we can more effectively fold with loads.
8149 if (V.getOpcode() == ISD::BUILD_VECTOR ||
8150 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
8151 V = V.getOperand(BroadcastIdx);
8153 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
8155 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
8157 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
8158 // We can't broadcast from a vector register w/o AVX2, and we can only
8159 // broadcast from the zero-element of a vector register.
8163 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
8166 // Check for whether we can use INSERTPS to perform the shuffle. We only use
8167 // INSERTPS when the V1 elements are already in the correct locations
8168 // because otherwise we can just always use two SHUFPS instructions which
8169 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
8170 // perform INSERTPS if a single V1 element is out of place and all V2
8171 // elements are zeroable.
8172 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
8174 SelectionDAG &DAG) {
8175 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8176 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8177 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8178 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8180 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8183 int V1DstIndex = -1;
8184 int V2DstIndex = -1;
8185 bool V1UsedInPlace = false;
8187 for (int i = 0; i < 4; i++) {
8188 // Synthesize a zero mask from the zeroable elements (includes undefs).
8194 // Flag if we use any V1 inputs in place.
8196 V1UsedInPlace = true;
8200 // We can only insert a single non-zeroable element.
8201 if (V1DstIndex != -1 || V2DstIndex != -1)
8205 // V1 input out of place for insertion.
8208 // V2 input for insertion.
8213 // Don't bother if we have no (non-zeroable) element for insertion.
8214 if (V1DstIndex == -1 && V2DstIndex == -1)
8217 // Determine element insertion src/dst indices. The src index is from the
8218 // start of the inserted vector, not the start of the concatenated vector.
8219 unsigned V2SrcIndex = 0;
8220 if (V1DstIndex != -1) {
8221 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8222 // and don't use the original V2 at all.
8223 V2SrcIndex = Mask[V1DstIndex];
8224 V2DstIndex = V1DstIndex;
8227 V2SrcIndex = Mask[V2DstIndex] - 4;
8230 // If no V1 inputs are used in place, then the result is created only from
8231 // the zero mask and the V2 insertion - so remove V1 dependency.
8233 V1 = DAG.getUNDEF(MVT::v4f32);
8235 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8236 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8238 // Insert the V2 element into the desired position.
8240 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8241 DAG.getConstant(InsertPSMask, MVT::i8));
8244 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8246 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8247 /// support for floating point shuffles but not integer shuffles. These
8248 /// instructions will incur a domain crossing penalty on some chips though so
8249 /// it is better to avoid lowering through this for integer vectors where
8251 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8252 const X86Subtarget *Subtarget,
8253 SelectionDAG &DAG) {
8255 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8256 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8257 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8258 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8259 ArrayRef<int> Mask = SVOp->getMask();
8260 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8262 if (isSingleInputShuffleMask(Mask)) {
8263 // Straight shuffle of a single input vector. Simulate this by using the
8264 // single input as both of the "inputs" to this instruction..
8265 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8267 if (Subtarget->hasAVX()) {
8268 // If we have AVX, we can use VPERMILPS which will allow folding a load
8269 // into the shuffle.
8270 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8271 DAG.getConstant(SHUFPDMask, MVT::i8));
8274 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
8275 DAG.getConstant(SHUFPDMask, MVT::i8));
8277 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8278 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8280 // Use dedicated unpack instructions for masks that match their pattern.
8281 if (isShuffleEquivalent(Mask, 0, 2))
8282 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8283 if (isShuffleEquivalent(Mask, 1, 3))
8284 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8286 // If we have a single input, insert that into V1 if we can do so cheaply.
8287 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8288 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8289 MVT::v2f64, DL, V1, V2, Mask, Subtarget, DAG))
8291 // Try inverting the insertion since for v2 masks it is easy to do and we
8292 // can't reliably sort the mask one way or the other.
8293 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8294 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8295 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8296 MVT::v2f64, DL, V2, V1, InverseMask, Subtarget, DAG))
8300 // Try to use one of the special instruction patterns to handle two common
8301 // blend patterns if a zero-blend above didn't work.
8302 if (isShuffleEquivalent(Mask, 0, 3) || isShuffleEquivalent(Mask, 1, 3))
8303 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8304 // We can either use a special instruction to load over the low double or
8305 // to move just the low double.
8307 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8309 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8311 if (Subtarget->hasSSE41())
8312 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8316 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8317 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
8318 DAG.getConstant(SHUFPDMask, MVT::i8));
8321 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8323 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8324 /// the integer unit to minimize domain crossing penalties. However, for blends
8325 /// it falls back to the floating point shuffle operation with appropriate bit
8327 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8328 const X86Subtarget *Subtarget,
8329 SelectionDAG &DAG) {
8331 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8332 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8333 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8334 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8335 ArrayRef<int> Mask = SVOp->getMask();
8336 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8338 if (isSingleInputShuffleMask(Mask)) {
8339 // Check for being able to broadcast a single element.
8340 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v2i64, DL, V1,
8341 Mask, Subtarget, DAG))
8344 // Straight shuffle of a single input vector. For everything from SSE2
8345 // onward this has a single fast instruction with no scary immediates.
8346 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8347 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
8348 int WidenedMask[4] = {
8349 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8350 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8352 ISD::BITCAST, DL, MVT::v2i64,
8353 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
8354 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
8357 // Try to use byte shift instructions.
8358 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8359 DL, MVT::v2i64, V1, V2, Mask, DAG))
8362 // If we have a single input from V2 insert that into V1 if we can do so
8364 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8365 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8366 MVT::v2i64, DL, V1, V2, Mask, Subtarget, DAG))
8368 // Try inverting the insertion since for v2 masks it is easy to do and we
8369 // can't reliably sort the mask one way or the other.
8370 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8371 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8372 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8373 MVT::v2i64, DL, V2, V1, InverseMask, Subtarget, DAG))
8377 // Use dedicated unpack instructions for masks that match their pattern.
8378 if (isShuffleEquivalent(Mask, 0, 2))
8379 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8380 if (isShuffleEquivalent(Mask, 1, 3))
8381 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8383 if (Subtarget->hasSSE41())
8384 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8388 // Try to use byte rotation instructions.
8389 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8390 if (Subtarget->hasSSSE3())
8391 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8392 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8395 // We implement this with SHUFPD which is pretty lame because it will likely
8396 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8397 // However, all the alternatives are still more cycles and newer chips don't
8398 // have this problem. It would be really nice if x86 had better shuffles here.
8399 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
8400 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
8401 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
8402 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8405 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8407 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8408 /// It makes no assumptions about whether this is the *best* lowering, it simply
8410 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8411 ArrayRef<int> Mask, SDValue V1,
8412 SDValue V2, SelectionDAG &DAG) {
8413 SDValue LowV = V1, HighV = V2;
8414 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8417 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8419 if (NumV2Elements == 1) {
8421 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8424 // Compute the index adjacent to V2Index and in the same half by toggling
8426 int V2AdjIndex = V2Index ^ 1;
8428 if (Mask[V2AdjIndex] == -1) {
8429 // Handles all the cases where we have a single V2 element and an undef.
8430 // This will only ever happen in the high lanes because we commute the
8431 // vector otherwise.
8433 std::swap(LowV, HighV);
8434 NewMask[V2Index] -= 4;
8436 // Handle the case where the V2 element ends up adjacent to a V1 element.
8437 // To make this work, blend them together as the first step.
8438 int V1Index = V2AdjIndex;
8439 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8440 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8441 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8443 // Now proceed to reconstruct the final blend as we have the necessary
8444 // high or low half formed.
8451 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8452 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8454 } else if (NumV2Elements == 2) {
8455 if (Mask[0] < 4 && Mask[1] < 4) {
8456 // Handle the easy case where we have V1 in the low lanes and V2 in the
8460 } else if (Mask[2] < 4 && Mask[3] < 4) {
8461 // We also handle the reversed case because this utility may get called
8462 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8463 // arrange things in the right direction.
8469 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8470 // trying to place elements directly, just blend them and set up the final
8471 // shuffle to place them.
8473 // The first two blend mask elements are for V1, the second two are for
8475 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8476 Mask[2] < 4 ? Mask[2] : Mask[3],
8477 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8478 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8479 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8480 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
8482 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8485 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8486 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8487 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8488 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8491 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8492 getV4X86ShuffleImm8ForMask(NewMask, DAG));
8495 /// \brief Lower 4-lane 32-bit floating point shuffles.
8497 /// Uses instructions exclusively from the floating point unit to minimize
8498 /// domain crossing penalties, as these are sufficient to implement all v4f32
8500 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8501 const X86Subtarget *Subtarget,
8502 SelectionDAG &DAG) {
8504 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8505 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8506 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8507 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8508 ArrayRef<int> Mask = SVOp->getMask();
8509 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8512 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8514 if (NumV2Elements == 0) {
8515 // Check for being able to broadcast a single element.
8516 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f32, DL, V1,
8517 Mask, Subtarget, DAG))
8520 if (Subtarget->hasAVX()) {
8521 // If we have AVX, we can use VPERMILPS which will allow folding a load
8522 // into the shuffle.
8523 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8524 getV4X86ShuffleImm8ForMask(Mask, DAG));
8527 // Otherwise, use a straight shuffle of a single input vector. We pass the
8528 // input vector to both operands to simulate this with a SHUFPS.
8529 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8530 getV4X86ShuffleImm8ForMask(Mask, DAG));
8533 // Use dedicated unpack instructions for masks that match their pattern.
8534 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8535 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8536 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8537 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8539 // There are special ways we can lower some single-element blends. However, we
8540 // have custom ways we can lower more complex single-element blends below that
8541 // we defer to if both this and BLENDPS fail to match, so restrict this to
8542 // when the V2 input is targeting element 0 of the mask -- that is the fast
8544 if (NumV2Elements == 1 && Mask[0] >= 4)
8545 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4f32, DL, V1, V2,
8546 Mask, Subtarget, DAG))
8549 if (Subtarget->hasSSE41()) {
8550 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8554 // Use INSERTPS if we can complete the shuffle efficiently.
8555 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8559 // Otherwise fall back to a SHUFPS lowering strategy.
8560 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8563 /// \brief Lower 4-lane i32 vector shuffles.
8565 /// We try to handle these with integer-domain shuffles where we can, but for
8566 /// blends we use the floating point domain blend instructions.
8567 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8568 const X86Subtarget *Subtarget,
8569 SelectionDAG &DAG) {
8571 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8572 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8573 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8574 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8575 ArrayRef<int> Mask = SVOp->getMask();
8576 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8578 // Whenever we can lower this as a zext, that instruction is strictly faster
8579 // than any alternative. It also allows us to fold memory operands into the
8580 // shuffle in many cases.
8581 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8582 Mask, Subtarget, DAG))
8586 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8588 if (NumV2Elements == 0) {
8589 // Check for being able to broadcast a single element.
8590 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i32, DL, V1,
8591 Mask, Subtarget, DAG))
8594 // Straight shuffle of a single input vector. For everything from SSE2
8595 // onward this has a single fast instruction with no scary immediates.
8596 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8597 // but we aren't actually going to use the UNPCK instruction because doing
8598 // so prevents folding a load into this instruction or making a copy.
8599 const int UnpackLoMask[] = {0, 0, 1, 1};
8600 const int UnpackHiMask[] = {2, 2, 3, 3};
8601 if (isShuffleEquivalent(Mask, 0, 0, 1, 1))
8602 Mask = UnpackLoMask;
8603 else if (isShuffleEquivalent(Mask, 2, 2, 3, 3))
8604 Mask = UnpackHiMask;
8606 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8607 getV4X86ShuffleImm8ForMask(Mask, DAG));
8610 // Try to use byte shift instructions.
8611 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8612 DL, MVT::v4i32, V1, V2, Mask, DAG))
8615 // There are special ways we can lower some single-element blends.
8616 if (NumV2Elements == 1)
8617 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v4i32, DL, V1, V2,
8618 Mask, Subtarget, DAG))
8621 // Use dedicated unpack instructions for masks that match their pattern.
8622 if (isShuffleEquivalent(Mask, 0, 4, 1, 5))
8623 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8624 if (isShuffleEquivalent(Mask, 2, 6, 3, 7))
8625 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8627 if (Subtarget->hasSSE41())
8628 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8632 // Try to use byte rotation instructions.
8633 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8634 if (Subtarget->hasSSSE3())
8635 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8636 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8639 // We implement this with SHUFPS because it can blend from two vectors.
8640 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8641 // up the inputs, bypassing domain shift penalties that we would encur if we
8642 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8644 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
8645 DAG.getVectorShuffle(
8647 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
8648 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
8651 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8652 /// shuffle lowering, and the most complex part.
8654 /// The lowering strategy is to try to form pairs of input lanes which are
8655 /// targeted at the same half of the final vector, and then use a dword shuffle
8656 /// to place them onto the right half, and finally unpack the paired lanes into
8657 /// their final position.
8659 /// The exact breakdown of how to form these dword pairs and align them on the
8660 /// correct sides is really tricky. See the comments within the function for
8661 /// more of the details.
8662 static SDValue lowerV8I16SingleInputVectorShuffle(
8663 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
8664 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8665 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
8666 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8667 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8669 SmallVector<int, 4> LoInputs;
8670 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8671 [](int M) { return M >= 0; });
8672 std::sort(LoInputs.begin(), LoInputs.end());
8673 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8674 SmallVector<int, 4> HiInputs;
8675 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8676 [](int M) { return M >= 0; });
8677 std::sort(HiInputs.begin(), HiInputs.end());
8678 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8680 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8681 int NumHToL = LoInputs.size() - NumLToL;
8683 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8684 int NumHToH = HiInputs.size() - NumLToH;
8685 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8686 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8687 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8688 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8690 // Check for being able to broadcast a single element.
8691 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i16, DL, V,
8692 Mask, Subtarget, DAG))
8695 // Try to use byte shift instructions.
8696 if (SDValue Shift = lowerVectorShuffleAsByteShift(
8697 DL, MVT::v8i16, V, V, Mask, DAG))
8700 // Use dedicated unpack instructions for masks that match their pattern.
8701 if (isShuffleEquivalent(Mask, 0, 0, 1, 1, 2, 2, 3, 3))
8702 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V, V);
8703 if (isShuffleEquivalent(Mask, 4, 4, 5, 5, 6, 6, 7, 7))
8704 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V, V);
8706 // Try to use byte rotation instructions.
8707 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8708 DL, MVT::v8i16, V, V, Mask, Subtarget, DAG))
8711 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8712 // such inputs we can swap two of the dwords across the half mark and end up
8713 // with <=2 inputs to each half in each half. Once there, we can fall through
8714 // to the generic code below. For example:
8716 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8717 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8719 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8720 // and an existing 2-into-2 on the other half. In this case we may have to
8721 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8722 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8723 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8724 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8725 // half than the one we target for fixing) will be fixed when we re-enter this
8726 // path. We will also combine away any sequence of PSHUFD instructions that
8727 // result into a single instruction. Here is an example of the tricky case:
8729 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8730 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8732 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8734 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8735 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8737 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8738 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8740 // The result is fine to be handled by the generic logic.
8741 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8742 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8743 int AOffset, int BOffset) {
8744 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8745 "Must call this with A having 3 or 1 inputs from the A half.");
8746 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8747 "Must call this with B having 1 or 3 inputs from the B half.");
8748 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8749 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8751 // Compute the index of dword with only one word among the three inputs in
8752 // a half by taking the sum of the half with three inputs and subtracting
8753 // the sum of the actual three inputs. The difference is the remaining
8756 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8757 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8758 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8759 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8760 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8761 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8762 int TripleNonInputIdx =
8763 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8764 TripleDWord = TripleNonInputIdx / 2;
8766 // We use xor with one to compute the adjacent DWord to whichever one the
8768 OneInputDWord = (OneInput / 2) ^ 1;
8770 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8771 // and BToA inputs. If there is also such a problem with the BToB and AToB
8772 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8773 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8774 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8775 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8776 // Compute how many inputs will be flipped by swapping these DWords. We
8778 // to balance this to ensure we don't form a 3-1 shuffle in the other
8780 int NumFlippedAToBInputs =
8781 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8782 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8783 int NumFlippedBToBInputs =
8784 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8785 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8786 if ((NumFlippedAToBInputs == 1 &&
8787 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8788 (NumFlippedBToBInputs == 1 &&
8789 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8790 // We choose whether to fix the A half or B half based on whether that
8791 // half has zero flipped inputs. At zero, we may not be able to fix it
8792 // with that half. We also bias towards fixing the B half because that
8793 // will more commonly be the high half, and we have to bias one way.
8794 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8795 ArrayRef<int> Inputs) {
8796 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8797 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8798 PinnedIdx ^ 1) != Inputs.end();
8799 // Determine whether the free index is in the flipped dword or the
8800 // unflipped dword based on where the pinned index is. We use this bit
8801 // in an xor to conditionally select the adjacent dword.
8802 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8803 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8804 FixFreeIdx) != Inputs.end();
8805 if (IsFixIdxInput == IsFixFreeIdxInput)
8807 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8808 FixFreeIdx) != Inputs.end();
8809 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8810 "We need to be changing the number of flipped inputs!");
8811 int PSHUFHalfMask[] = {0, 1, 2, 3};
8812 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8813 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8815 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
8818 if (M != -1 && M == FixIdx)
8820 else if (M != -1 && M == FixFreeIdx)
8823 if (NumFlippedBToBInputs != 0) {
8825 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8826 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8828 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8830 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8831 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8836 int PSHUFDMask[] = {0, 1, 2, 3};
8837 PSHUFDMask[ADWord] = BDWord;
8838 PSHUFDMask[BDWord] = ADWord;
8839 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8840 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8841 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8842 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8844 // Adjust the mask to match the new locations of A and B.
8846 if (M != -1 && M/2 == ADWord)
8847 M = 2 * BDWord + M % 2;
8848 else if (M != -1 && M/2 == BDWord)
8849 M = 2 * ADWord + M % 2;
8851 // Recurse back into this routine to re-compute state now that this isn't
8852 // a 3 and 1 problem.
8853 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
8856 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8857 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8858 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8859 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8861 // At this point there are at most two inputs to the low and high halves from
8862 // each half. That means the inputs can always be grouped into dwords and
8863 // those dwords can then be moved to the correct half with a dword shuffle.
8864 // We use at most one low and one high word shuffle to collect these paired
8865 // inputs into dwords, and finally a dword shuffle to place them.
8866 int PSHUFLMask[4] = {-1, -1, -1, -1};
8867 int PSHUFHMask[4] = {-1, -1, -1, -1};
8868 int PSHUFDMask[4] = {-1, -1, -1, -1};
8870 // First fix the masks for all the inputs that are staying in their
8871 // original halves. This will then dictate the targets of the cross-half
8873 auto fixInPlaceInputs =
8874 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8875 MutableArrayRef<int> SourceHalfMask,
8876 MutableArrayRef<int> HalfMask, int HalfOffset) {
8877 if (InPlaceInputs.empty())
8879 if (InPlaceInputs.size() == 1) {
8880 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8881 InPlaceInputs[0] - HalfOffset;
8882 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8885 if (IncomingInputs.empty()) {
8886 // Just fix all of the in place inputs.
8887 for (int Input : InPlaceInputs) {
8888 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8889 PSHUFDMask[Input / 2] = Input / 2;
8894 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8895 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8896 InPlaceInputs[0] - HalfOffset;
8897 // Put the second input next to the first so that they are packed into
8898 // a dword. We find the adjacent index by toggling the low bit.
8899 int AdjIndex = InPlaceInputs[0] ^ 1;
8900 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8901 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8902 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8904 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8905 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8907 // Now gather the cross-half inputs and place them into a free dword of
8908 // their target half.
8909 // FIXME: This operation could almost certainly be simplified dramatically to
8910 // look more like the 3-1 fixing operation.
8911 auto moveInputsToRightHalf = [&PSHUFDMask](
8912 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8913 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8914 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8916 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8917 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8919 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8921 int LowWord = Word & ~1;
8922 int HighWord = Word | 1;
8923 return isWordClobbered(SourceHalfMask, LowWord) ||
8924 isWordClobbered(SourceHalfMask, HighWord);
8927 if (IncomingInputs.empty())
8930 if (ExistingInputs.empty()) {
8931 // Map any dwords with inputs from them into the right half.
8932 for (int Input : IncomingInputs) {
8933 // If the source half mask maps over the inputs, turn those into
8934 // swaps and use the swapped lane.
8935 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8936 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8937 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8938 Input - SourceOffset;
8939 // We have to swap the uses in our half mask in one sweep.
8940 for (int &M : HalfMask)
8941 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8943 else if (M == Input)
8944 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8946 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8947 Input - SourceOffset &&
8948 "Previous placement doesn't match!");
8950 // Note that this correctly re-maps both when we do a swap and when
8951 // we observe the other side of the swap above. We rely on that to
8952 // avoid swapping the members of the input list directly.
8953 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8956 // Map the input's dword into the correct half.
8957 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8958 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8960 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8962 "Previous placement doesn't match!");
8965 // And just directly shift any other-half mask elements to be same-half
8966 // as we will have mirrored the dword containing the element into the
8967 // same position within that half.
8968 for (int &M : HalfMask)
8969 if (M >= SourceOffset && M < SourceOffset + 4) {
8970 M = M - SourceOffset + DestOffset;
8971 assert(M >= 0 && "This should never wrap below zero!");
8976 // Ensure we have the input in a viable dword of its current half. This
8977 // is particularly tricky because the original position may be clobbered
8978 // by inputs being moved and *staying* in that half.
8979 if (IncomingInputs.size() == 1) {
8980 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8981 int InputFixed = std::find(std::begin(SourceHalfMask),
8982 std::end(SourceHalfMask), -1) -
8983 std::begin(SourceHalfMask) + SourceOffset;
8984 SourceHalfMask[InputFixed - SourceOffset] =
8985 IncomingInputs[0] - SourceOffset;
8986 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8988 IncomingInputs[0] = InputFixed;
8990 } else if (IncomingInputs.size() == 2) {
8991 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8992 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8993 // We have two non-adjacent or clobbered inputs we need to extract from
8994 // the source half. To do this, we need to map them into some adjacent
8995 // dword slot in the source mask.
8996 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8997 IncomingInputs[1] - SourceOffset};
8999 // If there is a free slot in the source half mask adjacent to one of
9000 // the inputs, place the other input in it. We use (Index XOR 1) to
9001 // compute an adjacent index.
9002 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9003 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9004 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9005 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9006 InputsFixed[1] = InputsFixed[0] ^ 1;
9007 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9008 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9009 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9010 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9011 InputsFixed[0] = InputsFixed[1] ^ 1;
9012 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9013 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9014 // The two inputs are in the same DWord but it is clobbered and the
9015 // adjacent DWord isn't used at all. Move both inputs to the free
9017 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9018 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9019 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9020 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9022 // The only way we hit this point is if there is no clobbering
9023 // (because there are no off-half inputs to this half) and there is no
9024 // free slot adjacent to one of the inputs. In this case, we have to
9025 // swap an input with a non-input.
9026 for (int i = 0; i < 4; ++i)
9027 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9028 "We can't handle any clobbers here!");
9029 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9030 "Cannot have adjacent inputs here!");
9032 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9033 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9035 // We also have to update the final source mask in this case because
9036 // it may need to undo the above swap.
9037 for (int &M : FinalSourceHalfMask)
9038 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9039 M = InputsFixed[1] + SourceOffset;
9040 else if (M == InputsFixed[1] + SourceOffset)
9041 M = (InputsFixed[0] ^ 1) + SourceOffset;
9043 InputsFixed[1] = InputsFixed[0] ^ 1;
9046 // Point everything at the fixed inputs.
9047 for (int &M : HalfMask)
9048 if (M == IncomingInputs[0])
9049 M = InputsFixed[0] + SourceOffset;
9050 else if (M == IncomingInputs[1])
9051 M = InputsFixed[1] + SourceOffset;
9053 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9054 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9057 llvm_unreachable("Unhandled input size!");
9060 // Now hoist the DWord down to the right half.
9061 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9062 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9063 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9064 for (int &M : HalfMask)
9065 for (int Input : IncomingInputs)
9067 M = FreeDWord * 2 + Input % 2;
9069 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9070 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9071 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9072 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9074 // Now enact all the shuffles we've computed to move the inputs into their
9076 if (!isNoopShuffleMask(PSHUFLMask))
9077 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9078 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
9079 if (!isNoopShuffleMask(PSHUFHMask))
9080 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9081 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
9082 if (!isNoopShuffleMask(PSHUFDMask))
9083 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9084 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
9085 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
9086 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9088 // At this point, each half should contain all its inputs, and we can then
9089 // just shuffle them into their final position.
9090 assert(std::count_if(LoMask.begin(), LoMask.end(),
9091 [](int M) { return M >= 4; }) == 0 &&
9092 "Failed to lift all the high half inputs to the low mask!");
9093 assert(std::count_if(HiMask.begin(), HiMask.end(),
9094 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9095 "Failed to lift all the low half inputs to the high mask!");
9097 // Do a half shuffle for the low mask.
9098 if (!isNoopShuffleMask(LoMask))
9099 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
9100 getV4X86ShuffleImm8ForMask(LoMask, DAG));
9102 // Do a half shuffle with the high mask after shifting its values down.
9103 for (int &M : HiMask)
9106 if (!isNoopShuffleMask(HiMask))
9107 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
9108 getV4X86ShuffleImm8ForMask(HiMask, DAG));
9113 /// \brief Detect whether the mask pattern should be lowered through
9116 /// This essentially tests whether viewing the mask as an interleaving of two
9117 /// sub-sequences reduces the cross-input traffic of a blend operation. If so,
9118 /// lowering it through interleaving is a significantly better strategy.
9119 static bool shouldLowerAsInterleaving(ArrayRef<int> Mask) {
9120 int NumEvenInputs[2] = {0, 0};
9121 int NumOddInputs[2] = {0, 0};
9122 int NumLoInputs[2] = {0, 0};
9123 int NumHiInputs[2] = {0, 0};
9124 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
9128 int InputIdx = Mask[i] >= Size;
9131 ++NumLoInputs[InputIdx];
9133 ++NumHiInputs[InputIdx];
9136 ++NumEvenInputs[InputIdx];
9138 ++NumOddInputs[InputIdx];
9141 // The minimum number of cross-input results for both the interleaved and
9142 // split cases. If interleaving results in fewer cross-input results, return
9144 int InterleavedCrosses = std::min(NumEvenInputs[1] + NumOddInputs[0],
9145 NumEvenInputs[0] + NumOddInputs[1]);
9146 int SplitCrosses = std::min(NumLoInputs[1] + NumHiInputs[0],
9147 NumLoInputs[0] + NumHiInputs[1]);
9148 return InterleavedCrosses < SplitCrosses;
9151 /// \brief Blend two v8i16 vectors using a naive unpack strategy.
9153 /// This strategy only works when the inputs from each vector fit into a single
9154 /// half of that vector, and generally there are not so many inputs as to leave
9155 /// the in-place shuffles required highly constrained (and thus expensive). It
9156 /// shifts all the inputs into a single side of both input vectors and then
9157 /// uses an unpack to interleave these inputs in a single vector. At that
9158 /// point, we will fall back on the generic single input shuffle lowering.
9159 static SDValue lowerV8I16BasicBlendVectorShuffle(SDLoc DL, SDValue V1,
9161 MutableArrayRef<int> Mask,
9162 const X86Subtarget *Subtarget,
9163 SelectionDAG &DAG) {
9164 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9165 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
9166 SmallVector<int, 3> LoV1Inputs, HiV1Inputs, LoV2Inputs, HiV2Inputs;
9167 for (int i = 0; i < 8; ++i)
9168 if (Mask[i] >= 0 && Mask[i] < 4)
9169 LoV1Inputs.push_back(i);
9170 else if (Mask[i] >= 4 && Mask[i] < 8)
9171 HiV1Inputs.push_back(i);
9172 else if (Mask[i] >= 8 && Mask[i] < 12)
9173 LoV2Inputs.push_back(i);
9174 else if (Mask[i] >= 12)
9175 HiV2Inputs.push_back(i);
9177 int NumV1Inputs = LoV1Inputs.size() + HiV1Inputs.size();
9178 int NumV2Inputs = LoV2Inputs.size() + HiV2Inputs.size();
9181 assert(NumV1Inputs > 0 && NumV1Inputs <= 3 && "At most 3 inputs supported");
9182 assert(NumV2Inputs > 0 && NumV2Inputs <= 3 && "At most 3 inputs supported");
9183 assert(NumV1Inputs + NumV2Inputs <= 4 && "At most 4 combined inputs");
9185 bool MergeFromLo = LoV1Inputs.size() + LoV2Inputs.size() >=
9186 HiV1Inputs.size() + HiV2Inputs.size();
9188 auto moveInputsToHalf = [&](SDValue V, ArrayRef<int> LoInputs,
9189 ArrayRef<int> HiInputs, bool MoveToLo,
9191 ArrayRef<int> GoodInputs = MoveToLo ? LoInputs : HiInputs;
9192 ArrayRef<int> BadInputs = MoveToLo ? HiInputs : LoInputs;
9193 if (BadInputs.empty())
9196 int MoveMask[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9197 int MoveOffset = MoveToLo ? 0 : 4;
9199 if (GoodInputs.empty()) {
9200 for (int BadInput : BadInputs) {
9201 MoveMask[Mask[BadInput] % 4 + MoveOffset] = Mask[BadInput] - MaskOffset;
9202 Mask[BadInput] = Mask[BadInput] % 4 + MoveOffset + MaskOffset;
9205 if (GoodInputs.size() == 2) {
9206 // If the low inputs are spread across two dwords, pack them into
9208 MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
9209 MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
9210 Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
9211 Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
9213 // Otherwise pin the good inputs.
9214 for (int GoodInput : GoodInputs)
9215 MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
9218 if (BadInputs.size() == 2) {
9219 // If we have two bad inputs then there may be either one or two good
9220 // inputs fixed in place. Find a fixed input, and then find the *other*
9221 // two adjacent indices by using modular arithmetic.
9223 std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
9224 [](int M) { return M >= 0; }) -
9225 std::begin(MoveMask);
9227 ((((GoodMaskIdx - MoveOffset) & ~1) + 2) % 4) + MoveOffset;
9228 assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
9229 assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
9230 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9231 MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
9232 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9233 Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
9235 assert(BadInputs.size() == 1 && "All sizes handled");
9236 int MoveMaskIdx = std::find(std::begin(MoveMask) + MoveOffset,
9237 std::end(MoveMask), -1) -
9238 std::begin(MoveMask);
9239 MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
9240 Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
9244 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
9247 V1 = moveInputsToHalf(V1, LoV1Inputs, HiV1Inputs, MergeFromLo,
9249 V2 = moveInputsToHalf(V2, LoV2Inputs, HiV2Inputs, MergeFromLo,
9252 // FIXME: Select an interleaving of the merge of V1 and V2 that minimizes
9253 // cross-half traffic in the final shuffle.
9255 // Munge the mask to be a single-input mask after the unpack merges the
9259 M = 2 * (M % 4) + (M / 8);
9261 return DAG.getVectorShuffle(
9262 MVT::v8i16, DL, DAG.getNode(MergeFromLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
9263 DL, MVT::v8i16, V1, V2),
9264 DAG.getUNDEF(MVT::v8i16), Mask);
9267 /// \brief Generic lowering of 8-lane i16 shuffles.
9269 /// This handles both single-input shuffles and combined shuffle/blends with
9270 /// two inputs. The single input shuffles are immediately delegated to
9271 /// a dedicated lowering routine.
9273 /// The blends are lowered in one of three fundamental ways. If there are few
9274 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9275 /// of the input is significantly cheaper when lowered as an interleaving of
9276 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9277 /// halves of the inputs separately (making them have relatively few inputs)
9278 /// and then concatenate them.
9279 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9280 const X86Subtarget *Subtarget,
9281 SelectionDAG &DAG) {
9283 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9284 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9285 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9286 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9287 ArrayRef<int> OrigMask = SVOp->getMask();
9288 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9289 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9290 MutableArrayRef<int> Mask(MaskStorage);
9292 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9294 // Whenever we can lower this as a zext, that instruction is strictly faster
9295 // than any alternative.
9296 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9297 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9300 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9301 auto isV2 = [](int M) { return M >= 8; };
9303 int NumV1Inputs = std::count_if(Mask.begin(), Mask.end(), isV1);
9304 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9306 if (NumV2Inputs == 0)
9307 return lowerV8I16SingleInputVectorShuffle(DL, V1, Mask, Subtarget, DAG);
9309 assert(NumV1Inputs > 0 && "All single-input shuffles should be canonicalized "
9310 "to be V1-input shuffles.");
9312 // Try to use byte shift instructions.
9313 if (SDValue Shift = lowerVectorShuffleAsByteShift(
9314 DL, MVT::v8i16, V1, V2, Mask, DAG))
9317 // There are special ways we can lower some single-element blends.
9318 if (NumV2Inputs == 1)
9319 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v8i16, DL, V1, V2,
9320 Mask, Subtarget, DAG))
9323 // Use dedicated unpack instructions for masks that match their pattern.
9324 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 2, 10, 3, 11))
9325 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9326 if (isShuffleEquivalent(Mask, 4, 12, 5, 13, 6, 14, 7, 15))
9327 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9329 if (Subtarget->hasSSE41())
9330 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9334 // Try to use byte rotation instructions.
9335 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9336 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9339 if (NumV1Inputs + NumV2Inputs <= 4)
9340 return lowerV8I16BasicBlendVectorShuffle(DL, V1, V2, Mask, Subtarget, DAG);
9342 // Check whether an interleaving lowering is likely to be more efficient.
9343 // This isn't perfect but it is a strong heuristic that tends to work well on
9344 // the kinds of shuffles that show up in practice.
9346 // FIXME: Handle 1x, 2x, and 4x interleaving.
9347 if (shouldLowerAsInterleaving(Mask)) {
9348 // FIXME: Figure out whether we should pack these into the low or high
9351 int EMask[8], OMask[8];
9352 for (int i = 0; i < 4; ++i) {
9353 EMask[i] = Mask[2*i];
9354 OMask[i] = Mask[2*i + 1];
9359 SDValue Evens = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, EMask);
9360 SDValue Odds = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, OMask);
9362 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, Evens, Odds);
9365 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9366 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9368 for (int i = 0; i < 4; ++i) {
9369 LoBlendMask[i] = Mask[i];
9370 HiBlendMask[i] = Mask[i + 4];
9373 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
9374 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
9375 LoV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, LoV);
9376 HiV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, HiV);
9378 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9379 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
9382 /// \brief Check whether a compaction lowering can be done by dropping even
9383 /// elements and compute how many times even elements must be dropped.
9385 /// This handles shuffles which take every Nth element where N is a power of
9386 /// two. Example shuffle masks:
9388 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9389 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9390 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9391 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9392 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9393 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9395 /// Any of these lanes can of course be undef.
9397 /// This routine only supports N <= 3.
9398 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9401 /// \returns N above, or the number of times even elements must be dropped if
9402 /// there is such a number. Otherwise returns zero.
9403 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9404 // Figure out whether we're looping over two inputs or just one.
9405 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9407 // The modulus for the shuffle vector entries is based on whether this is
9408 // a single input or not.
9409 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9410 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9411 "We should only be called with masks with a power-of-2 size!");
9413 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9415 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9416 // and 2^3 simultaneously. This is because we may have ambiguity with
9417 // partially undef inputs.
9418 bool ViableForN[3] = {true, true, true};
9420 for (int i = 0, e = Mask.size(); i < e; ++i) {
9421 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9426 bool IsAnyViable = false;
9427 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9428 if (ViableForN[j]) {
9431 // The shuffle mask must be equal to (i * 2^N) % M.
9432 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9435 ViableForN[j] = false;
9437 // Early exit if we exhaust the possible powers of two.
9442 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9446 // Return 0 as there is no viable power of two.
9450 /// \brief Generic lowering of v16i8 shuffles.
9452 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9453 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9454 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9455 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9457 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9458 const X86Subtarget *Subtarget,
9459 SelectionDAG &DAG) {
9461 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9462 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9463 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9464 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9465 ArrayRef<int> OrigMask = SVOp->getMask();
9466 assert(OrigMask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9468 // Try to use byte shift instructions.
9469 if (SDValue Shift = lowerVectorShuffleAsByteShift(
9470 DL, MVT::v16i8, V1, V2, OrigMask, DAG))
9473 // Try to use byte rotation instructions.
9474 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9475 DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
9478 // Try to use a zext lowering.
9479 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9480 DL, MVT::v16i8, V1, V2, OrigMask, Subtarget, DAG))
9483 int MaskStorage[16] = {
9484 OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9485 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7],
9486 OrigMask[8], OrigMask[9], OrigMask[10], OrigMask[11],
9487 OrigMask[12], OrigMask[13], OrigMask[14], OrigMask[15]};
9488 MutableArrayRef<int> Mask(MaskStorage);
9489 MutableArrayRef<int> LoMask = Mask.slice(0, 8);
9490 MutableArrayRef<int> HiMask = Mask.slice(8, 8);
9493 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9495 // For single-input shuffles, there are some nicer lowering tricks we can use.
9496 if (NumV2Elements == 0) {
9497 // Check for being able to broadcast a single element.
9498 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i8, DL, V1,
9499 Mask, Subtarget, DAG))
9502 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9503 // Notably, this handles splat and partial-splat shuffles more efficiently.
9504 // However, it only makes sense if the pre-duplication shuffle simplifies
9505 // things significantly. Currently, this means we need to be able to
9506 // express the pre-duplication shuffle as an i16 shuffle.
9508 // FIXME: We should check for other patterns which can be widened into an
9509 // i16 shuffle as well.
9510 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9511 for (int i = 0; i < 16; i += 2)
9512 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9517 auto tryToWidenViaDuplication = [&]() -> SDValue {
9518 if (!canWidenViaDuplication(Mask))
9520 SmallVector<int, 4> LoInputs;
9521 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9522 [](int M) { return M >= 0 && M < 8; });
9523 std::sort(LoInputs.begin(), LoInputs.end());
9524 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9526 SmallVector<int, 4> HiInputs;
9527 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9528 [](int M) { return M >= 8; });
9529 std::sort(HiInputs.begin(), HiInputs.end());
9530 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9533 bool TargetLo = LoInputs.size() >= HiInputs.size();
9534 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9535 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9537 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9538 SmallDenseMap<int, int, 8> LaneMap;
9539 for (int I : InPlaceInputs) {
9540 PreDupI16Shuffle[I/2] = I/2;
9543 int j = TargetLo ? 0 : 4, je = j + 4;
9544 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9545 // Check if j is already a shuffle of this input. This happens when
9546 // there are two adjacent bytes after we move the low one.
9547 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9548 // If we haven't yet mapped the input, search for a slot into which
9550 while (j < je && PreDupI16Shuffle[j] != -1)
9554 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9557 // Map this input with the i16 shuffle.
9558 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9561 // Update the lane map based on the mapping we ended up with.
9562 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9565 ISD::BITCAST, DL, MVT::v16i8,
9566 DAG.getVectorShuffle(MVT::v8i16, DL,
9567 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9568 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9570 // Unpack the bytes to form the i16s that will be shuffled into place.
9571 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9572 MVT::v16i8, V1, V1);
9574 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9575 for (int i = 0; i < 16; ++i)
9576 if (Mask[i] != -1) {
9577 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9578 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9579 if (PostDupI16Shuffle[i / 2] == -1)
9580 PostDupI16Shuffle[i / 2] = MappedMask;
9582 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9583 "Conflicting entrties in the original shuffle!");
9586 ISD::BITCAST, DL, MVT::v16i8,
9587 DAG.getVectorShuffle(MVT::v8i16, DL,
9588 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
9589 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9591 if (SDValue V = tryToWidenViaDuplication())
9595 // Check whether an interleaving lowering is likely to be more efficient.
9596 // This isn't perfect but it is a strong heuristic that tends to work well on
9597 // the kinds of shuffles that show up in practice.
9599 // FIXME: We need to handle other interleaving widths (i16, i32, ...).
9600 if (shouldLowerAsInterleaving(Mask)) {
9601 int NumLoHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
9602 return (M >= 0 && M < 8) || (M >= 16 && M < 24);
9604 int NumHiHalf = std::count_if(Mask.begin(), Mask.end(), [](int M) {
9605 return (M >= 8 && M < 16) || M >= 24;
9607 int EMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
9608 -1, -1, -1, -1, -1, -1, -1, -1};
9609 int OMask[16] = {-1, -1, -1, -1, -1, -1, -1, -1,
9610 -1, -1, -1, -1, -1, -1, -1, -1};
9611 bool UnpackLo = NumLoHalf >= NumHiHalf;
9612 MutableArrayRef<int> TargetEMask(UnpackLo ? EMask : EMask + 8, 8);
9613 MutableArrayRef<int> TargetOMask(UnpackLo ? OMask : OMask + 8, 8);
9614 for (int i = 0; i < 8; ++i) {
9615 TargetEMask[i] = Mask[2 * i];
9616 TargetOMask[i] = Mask[2 * i + 1];
9619 SDValue Evens = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, EMask);
9620 SDValue Odds = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2, OMask);
9622 return DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9623 MVT::v16i8, Evens, Odds);
9626 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9627 // with PSHUFB. It is important to do this before we attempt to generate any
9628 // blends but after all of the single-input lowerings. If the single input
9629 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9630 // want to preserve that and we can DAG combine any longer sequences into
9631 // a PSHUFB in the end. But once we start blending from multiple inputs,
9632 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9633 // and there are *very* few patterns that would actually be faster than the
9634 // PSHUFB approach because of its ability to zero lanes.
9636 // FIXME: The only exceptions to the above are blends which are exact
9637 // interleavings with direct instructions supporting them. We currently don't
9638 // handle those well here.
9639 if (Subtarget->hasSSSE3()) {
9642 bool V1InUse = false;
9643 bool V2InUse = false;
9644 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9646 for (int i = 0; i < 16; ++i) {
9647 if (Mask[i] == -1) {
9648 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9650 const int ZeroMask = 0x80;
9651 int V1Idx = (Mask[i] < 16 ? Mask[i] : ZeroMask);
9652 int V2Idx = (Mask[i] < 16 ? ZeroMask : Mask[i] - 16);
9654 V1Idx = V2Idx = ZeroMask;
9655 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
9656 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
9657 V1InUse |= (ZeroMask != V1Idx);
9658 V2InUse |= (ZeroMask != V2Idx);
9661 assert((V1InUse || V2InUse) && "Shuffling to a zeroable vector");
9664 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1,
9665 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9667 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2,
9668 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9670 // If we need shuffled inputs from both, blend the two.
9671 if (V1InUse && V2InUse)
9672 return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9674 return V1; // Single inputs are easy.
9676 return V2; // Single inputs are easy.
9679 // There are special ways we can lower some single-element blends.
9680 if (NumV2Elements == 1)
9681 if (SDValue V = lowerVectorShuffleAsElementInsertion(MVT::v16i8, DL, V1, V2,
9682 Mask, Subtarget, DAG))
9685 // Check whether a compaction lowering can be done. This handles shuffles
9686 // which take every Nth element for some even N. See the helper function for
9689 // We special case these as they can be particularly efficiently handled with
9690 // the PACKUSB instruction on x86 and they show up in common patterns of
9691 // rearranging bytes to truncate wide elements.
9692 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9693 // NumEvenDrops is the power of two stride of the elements. Another way of
9694 // thinking about it is that we need to drop the even elements this many
9695 // times to get the original input.
9696 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9698 // First we need to zero all the dropped bytes.
9699 assert(NumEvenDrops <= 3 &&
9700 "No support for dropping even elements more than 3 times.");
9701 // We use the mask type to pick which bytes are preserved based on how many
9702 // elements are dropped.
9703 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9704 SDValue ByteClearMask =
9705 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
9706 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
9707 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9709 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9711 // Now pack things back together.
9712 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
9713 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
9714 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9715 for (int i = 1; i < NumEvenDrops; ++i) {
9716 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
9717 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9723 int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9724 int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9725 int V2LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9726 int V2HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9728 auto buildBlendMasks = [](MutableArrayRef<int> HalfMask,
9729 MutableArrayRef<int> V1HalfBlendMask,
9730 MutableArrayRef<int> V2HalfBlendMask) {
9731 for (int i = 0; i < 8; ++i)
9732 if (HalfMask[i] >= 0 && HalfMask[i] < 16) {
9733 V1HalfBlendMask[i] = HalfMask[i];
9735 } else if (HalfMask[i] >= 16) {
9736 V2HalfBlendMask[i] = HalfMask[i] - 16;
9737 HalfMask[i] = i + 8;
9740 buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
9741 buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
9743 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9745 auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
9746 MutableArrayRef<int> HiBlendMask) {
9748 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9749 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9751 if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
9752 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9753 std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
9754 [](int M) { return M >= 0 && M % 2 == 1; })) {
9755 // Use a mask to drop the high bytes.
9756 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
9757 V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
9758 DAG.getConstant(0x00FF, MVT::v8i16));
9760 // This will be a single vector shuffle instead of a blend so nuke V2.
9761 V2 = DAG.getUNDEF(MVT::v8i16);
9763 // Squash the masks to point directly into V1.
9764 for (int &M : LoBlendMask)
9767 for (int &M : HiBlendMask)
9771 // Otherwise just unpack the low half of V into V1 and the high half into
9772 // V2 so that we can blend them as i16s.
9773 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9774 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9775 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
9776 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9779 SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
9780 SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
9781 return std::make_pair(BlendedLo, BlendedHi);
9783 SDValue V1Lo, V1Hi, V2Lo, V2Hi;
9784 std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
9785 std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
9787 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
9788 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
9790 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9793 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9795 /// This routine breaks down the specific type of 128-bit shuffle and
9796 /// dispatches to the lowering routines accordingly.
9797 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9798 MVT VT, const X86Subtarget *Subtarget,
9799 SelectionDAG &DAG) {
9800 switch (VT.SimpleTy) {
9802 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9804 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9806 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9808 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9810 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9812 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9815 llvm_unreachable("Unimplemented!");
9819 /// \brief Helper function to test whether a shuffle mask could be
9820 /// simplified by widening the elements being shuffled.
9822 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9823 /// leaves it in an unspecified state.
9825 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9826 /// shuffle masks. The latter have the special property of a '-2' representing
9827 /// a zero-ed lane of a vector.
9828 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9829 SmallVectorImpl<int> &WidenedMask) {
9830 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9831 // If both elements are undef, its trivial.
9832 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9833 WidenedMask.push_back(SM_SentinelUndef);
9837 // Check for an undef mask and a mask value properly aligned to fit with
9838 // a pair of values. If we find such a case, use the non-undef mask's value.
9839 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9840 WidenedMask.push_back(Mask[i + 1] / 2);
9843 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9844 WidenedMask.push_back(Mask[i] / 2);
9848 // When zeroing, we need to spread the zeroing across both lanes to widen.
9849 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9850 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9851 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9852 WidenedMask.push_back(SM_SentinelZero);
9858 // Finally check if the two mask values are adjacent and aligned with
9860 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9861 WidenedMask.push_back(Mask[i] / 2);
9865 // Otherwise we can't safely widen the elements used in this shuffle.
9868 assert(WidenedMask.size() == Mask.size() / 2 &&
9869 "Incorrect size of mask after widening the elements!");
9874 /// \brief Generic routine to split ector shuffle into half-sized shuffles.
9876 /// This routine just extracts two subvectors, shuffles them independently, and
9877 /// then concatenates them back together. This should work effectively with all
9878 /// AVX vector shuffle types.
9879 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9880 SDValue V2, ArrayRef<int> Mask,
9881 SelectionDAG &DAG) {
9882 assert(VT.getSizeInBits() >= 256 &&
9883 "Only for 256-bit or wider vector shuffles!");
9884 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9885 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9887 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9888 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9890 int NumElements = VT.getVectorNumElements();
9891 int SplitNumElements = NumElements / 2;
9892 MVT ScalarVT = VT.getScalarType();
9893 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9895 SDValue LoV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
9896 DAG.getIntPtrConstant(0));
9897 SDValue HiV1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V1,
9898 DAG.getIntPtrConstant(SplitNumElements));
9899 SDValue LoV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
9900 DAG.getIntPtrConstant(0));
9901 SDValue HiV2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, V2,
9902 DAG.getIntPtrConstant(SplitNumElements));
9904 // Now create two 4-way blends of these half-width vectors.
9905 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9906 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9907 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9908 for (int i = 0; i < SplitNumElements; ++i) {
9909 int M = HalfMask[i];
9910 if (M >= NumElements) {
9911 if (M >= NumElements + SplitNumElements)
9915 V2BlendMask.push_back(M - NumElements);
9916 V1BlendMask.push_back(-1);
9917 BlendMask.push_back(SplitNumElements + i);
9918 } else if (M >= 0) {
9919 if (M >= SplitNumElements)
9923 V2BlendMask.push_back(-1);
9924 V1BlendMask.push_back(M);
9925 BlendMask.push_back(i);
9927 V2BlendMask.push_back(-1);
9928 V1BlendMask.push_back(-1);
9929 BlendMask.push_back(-1);
9933 // Because the lowering happens after all combining takes place, we need to
9934 // manually combine these blend masks as much as possible so that we create
9935 // a minimal number of high-level vector shuffle nodes.
9937 // First try just blending the halves of V1 or V2.
9938 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9939 return DAG.getUNDEF(SplitVT);
9940 if (!UseLoV2 && !UseHiV2)
9941 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9942 if (!UseLoV1 && !UseHiV1)
9943 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9945 SDValue V1Blend, V2Blend;
9946 if (UseLoV1 && UseHiV1) {
9948 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9950 // We only use half of V1 so map the usage down into the final blend mask.
9951 V1Blend = UseLoV1 ? LoV1 : HiV1;
9952 for (int i = 0; i < SplitNumElements; ++i)
9953 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9954 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9956 if (UseLoV2 && UseHiV2) {
9958 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9960 // We only use half of V2 so map the usage down into the final blend mask.
9961 V2Blend = UseLoV2 ? LoV2 : HiV2;
9962 for (int i = 0; i < SplitNumElements; ++i)
9963 if (BlendMask[i] >= SplitNumElements)
9964 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9966 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9968 SDValue Lo = HalfBlend(LoMask);
9969 SDValue Hi = HalfBlend(HiMask);
9970 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9973 /// \brief Either split a vector in halves or decompose the shuffles and the
9976 /// This is provided as a good fallback for many lowerings of non-single-input
9977 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9978 /// between splitting the shuffle into 128-bit components and stitching those
9979 /// back together vs. extracting the single-input shuffles and blending those
9981 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9982 SDValue V2, ArrayRef<int> Mask,
9983 SelectionDAG &DAG) {
9984 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9985 "lower single-input shuffles as it "
9986 "could then recurse on itself.");
9987 int Size = Mask.size();
9989 // If this can be modeled as a broadcast of two elements followed by a blend,
9990 // prefer that lowering. This is especially important because broadcasts can
9991 // often fold with memory operands.
9992 auto DoBothBroadcast = [&] {
9993 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9996 if (V2BroadcastIdx == -1)
9997 V2BroadcastIdx = M - Size;
9998 else if (M - Size != V2BroadcastIdx)
10000 } else if (M >= 0) {
10001 if (V1BroadcastIdx == -1)
10002 V1BroadcastIdx = M;
10003 else if (M != V1BroadcastIdx)
10008 if (DoBothBroadcast())
10009 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
10012 // If the inputs all stem from a single 128-bit lane of each input, then we
10013 // split them rather than blending because the split will decompose to
10014 // unusually few instructions.
10015 int LaneCount = VT.getSizeInBits() / 128;
10016 int LaneSize = Size / LaneCount;
10017 SmallBitVector LaneInputs[2];
10018 LaneInputs[0].resize(LaneCount, false);
10019 LaneInputs[1].resize(LaneCount, false);
10020 for (int i = 0; i < Size; ++i)
10022 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
10023 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
10024 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10026 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
10027 // that the decomposed single-input shuffles don't end up here.
10028 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10031 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
10032 /// a permutation and blend of those lanes.
10034 /// This essentially blends the out-of-lane inputs to each lane into the lane
10035 /// from a permuted copy of the vector. This lowering strategy results in four
10036 /// instructions in the worst case for a single-input cross lane shuffle which
10037 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
10038 /// of. Special cases for each particular shuffle pattern should be handled
10039 /// prior to trying this lowering.
10040 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
10041 SDValue V1, SDValue V2,
10042 ArrayRef<int> Mask,
10043 SelectionDAG &DAG) {
10044 // FIXME: This should probably be generalized for 512-bit vectors as well.
10045 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
10046 int LaneSize = Mask.size() / 2;
10048 // If there are only inputs from one 128-bit lane, splitting will in fact be
10049 // less expensive. The flags track wether the given lane contains an element
10050 // that crosses to another lane.
10051 bool LaneCrossing[2] = {false, false};
10052 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10053 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
10054 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
10055 if (!LaneCrossing[0] || !LaneCrossing[1])
10056 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10058 if (isSingleInputShuffleMask(Mask)) {
10059 SmallVector<int, 32> FlippedBlendMask;
10060 for (int i = 0, Size = Mask.size(); i < Size; ++i)
10061 FlippedBlendMask.push_back(
10062 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
10064 : Mask[i] % LaneSize +
10065 (i / LaneSize) * LaneSize + Size));
10067 // Flip the vector, and blend the results which should now be in-lane. The
10068 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
10069 // 5 for the high source. The value 3 selects the high half of source 2 and
10070 // the value 2 selects the low half of source 2. We only use source 2 to
10071 // allow folding it into a memory operand.
10072 unsigned PERMMask = 3 | 2 << 4;
10073 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
10074 V1, DAG.getConstant(PERMMask, MVT::i8));
10075 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
10078 // This now reduces to two single-input shuffles of V1 and V2 which at worst
10079 // will be handled by the above logic and a blend of the results, much like
10080 // other patterns in AVX.
10081 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
10084 /// \brief Handle lowering 2-lane 128-bit shuffles.
10085 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
10086 SDValue V2, ArrayRef<int> Mask,
10087 const X86Subtarget *Subtarget,
10088 SelectionDAG &DAG) {
10089 // Blends are faster and handle all the non-lane-crossing cases.
10090 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10094 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10095 VT.getVectorNumElements() / 2);
10096 // Check for patterns which can be matched with a single insert of a 128-bit
10098 if (isShuffleEquivalent(Mask, 0, 1, 0, 1) ||
10099 isShuffleEquivalent(Mask, 0, 1, 4, 5)) {
10100 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10101 DAG.getIntPtrConstant(0));
10102 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10103 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
10104 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10106 if (isShuffleEquivalent(Mask, 0, 1, 6, 7)) {
10107 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10108 DAG.getIntPtrConstant(0));
10109 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
10110 DAG.getIntPtrConstant(2));
10111 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10114 // Otherwise form a 128-bit permutation.
10115 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
10116 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
10117 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10118 DAG.getConstant(PermMask, MVT::i8));
10121 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10122 /// shuffling each lane.
10124 /// This will only succeed when the result of fixing the 128-bit lanes results
10125 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10126 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10127 /// the lane crosses early and then use simpler shuffles within each lane.
10129 /// FIXME: It might be worthwhile at some point to support this without
10130 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10131 /// in x86 only floating point has interesting non-repeating shuffles, and even
10132 /// those are still *marginally* more expensive.
10133 static SDValue lowerVectorShuffleByMerging128BitLanes(
10134 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10135 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10136 assert(!isSingleInputShuffleMask(Mask) &&
10137 "This is only useful with multiple inputs.");
10139 int Size = Mask.size();
10140 int LaneSize = 128 / VT.getScalarSizeInBits();
10141 int NumLanes = Size / LaneSize;
10142 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10144 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10145 // check whether the in-128-bit lane shuffles share a repeating pattern.
10146 SmallVector<int, 4> Lanes;
10147 Lanes.resize(NumLanes, -1);
10148 SmallVector<int, 4> InLaneMask;
10149 InLaneMask.resize(LaneSize, -1);
10150 for (int i = 0; i < Size; ++i) {
10154 int j = i / LaneSize;
10156 if (Lanes[j] < 0) {
10157 // First entry we've seen for this lane.
10158 Lanes[j] = Mask[i] / LaneSize;
10159 } else if (Lanes[j] != Mask[i] / LaneSize) {
10160 // This doesn't match the lane selected previously!
10164 // Check that within each lane we have a consistent shuffle mask.
10165 int k = i % LaneSize;
10166 if (InLaneMask[k] < 0) {
10167 InLaneMask[k] = Mask[i] % LaneSize;
10168 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10169 // This doesn't fit a repeating in-lane mask.
10174 // First shuffle the lanes into place.
10175 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10176 VT.getSizeInBits() / 64);
10177 SmallVector<int, 8> LaneMask;
10178 LaneMask.resize(NumLanes * 2, -1);
10179 for (int i = 0; i < NumLanes; ++i)
10180 if (Lanes[i] >= 0) {
10181 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10182 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10185 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
10186 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
10187 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10189 // Cast it back to the type we actually want.
10190 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
10192 // Now do a simple shuffle that isn't lane crossing.
10193 SmallVector<int, 8> NewMask;
10194 NewMask.resize(Size, -1);
10195 for (int i = 0; i < Size; ++i)
10197 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10198 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10199 "Must not introduce lane crosses at this point!");
10201 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10204 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10207 /// This returns true if the elements from a particular input are already in the
10208 /// slot required by the given mask and require no permutation.
10209 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10210 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10211 int Size = Mask.size();
10212 for (int i = 0; i < Size; ++i)
10213 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10219 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10221 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10222 /// isn't available.
10223 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10224 const X86Subtarget *Subtarget,
10225 SelectionDAG &DAG) {
10227 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10228 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10229 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10230 ArrayRef<int> Mask = SVOp->getMask();
10231 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10233 SmallVector<int, 4> WidenedMask;
10234 if (canWidenShuffleElements(Mask, WidenedMask))
10235 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10238 if (isSingleInputShuffleMask(Mask)) {
10239 // Check for being able to broadcast a single element.
10240 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4f64, DL, V1,
10241 Mask, Subtarget, DAG))
10244 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10245 // Non-half-crossing single input shuffles can be lowerid with an
10246 // interleaved permutation.
10247 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10248 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10249 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10250 DAG.getConstant(VPERMILPMask, MVT::i8));
10253 // With AVX2 we have direct support for this permutation.
10254 if (Subtarget->hasAVX2())
10255 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10256 getV4X86ShuffleImm8ForMask(Mask, DAG));
10258 // Otherwise, fall back.
10259 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10263 // X86 has dedicated unpack instructions that can handle specific blend
10264 // operations: UNPCKH and UNPCKL.
10265 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
10266 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10267 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
10268 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10270 // If we have a single input to the zero element, insert that into V1 if we
10271 // can do so cheaply.
10272 int NumV2Elements =
10273 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
10274 if (NumV2Elements == 1 && Mask[0] >= 4)
10275 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10276 MVT::v4f64, DL, V1, V2, Mask, Subtarget, DAG))
10279 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10283 // Check if the blend happens to exactly fit that of SHUFPD.
10284 if ((Mask[0] == -1 || Mask[0] < 2) &&
10285 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
10286 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
10287 (Mask[3] == -1 || Mask[3] >= 6)) {
10288 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
10289 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
10290 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
10291 DAG.getConstant(SHUFPDMask, MVT::i8));
10293 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
10294 (Mask[1] == -1 || Mask[1] < 2) &&
10295 (Mask[2] == -1 || Mask[2] >= 6) &&
10296 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
10297 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
10298 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
10299 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
10300 DAG.getConstant(SHUFPDMask, MVT::i8));
10303 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10304 // shuffle. However, if we have AVX2 and either inputs are already in place,
10305 // we will be able to shuffle even across lanes the other input in a single
10306 // instruction so skip this pattern.
10307 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10308 isShuffleMaskInputInPlace(1, Mask))))
10309 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10310 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10313 // If we have AVX2 then we always want to lower with a blend because an v4 we
10314 // can fully permute the elements.
10315 if (Subtarget->hasAVX2())
10316 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10319 // Otherwise fall back on generic lowering.
10320 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10323 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10325 /// This routine is only called when we have AVX2 and thus a reasonable
10326 /// instruction set for v4i64 shuffling..
10327 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10328 const X86Subtarget *Subtarget,
10329 SelectionDAG &DAG) {
10331 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10332 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10333 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10334 ArrayRef<int> Mask = SVOp->getMask();
10335 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10336 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10338 SmallVector<int, 4> WidenedMask;
10339 if (canWidenShuffleElements(Mask, WidenedMask))
10340 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10343 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10347 // Check for being able to broadcast a single element.
10348 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v4i64, DL, V1,
10349 Mask, Subtarget, DAG))
10352 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10353 // use lower latency instructions that will operate on both 128-bit lanes.
10354 SmallVector<int, 2> RepeatedMask;
10355 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10356 if (isSingleInputShuffleMask(Mask)) {
10357 int PSHUFDMask[] = {-1, -1, -1, -1};
10358 for (int i = 0; i < 2; ++i)
10359 if (RepeatedMask[i] >= 0) {
10360 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10361 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10363 return DAG.getNode(
10364 ISD::BITCAST, DL, MVT::v4i64,
10365 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10366 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
10367 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
10370 // Use dedicated unpack instructions for masks that match their pattern.
10371 if (isShuffleEquivalent(Mask, 0, 4, 2, 6))
10372 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10373 if (isShuffleEquivalent(Mask, 1, 5, 3, 7))
10374 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10377 // AVX2 provides a direct instruction for permuting a single input across
10379 if (isSingleInputShuffleMask(Mask))
10380 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10381 getV4X86ShuffleImm8ForMask(Mask, DAG));
10383 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10384 // shuffle. However, if we have AVX2 and either inputs are already in place,
10385 // we will be able to shuffle even across lanes the other input in a single
10386 // instruction so skip this pattern.
10387 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10388 isShuffleMaskInputInPlace(1, Mask))))
10389 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10390 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10393 // Otherwise fall back on generic blend lowering.
10394 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10398 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10400 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10401 /// isn't available.
10402 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10403 const X86Subtarget *Subtarget,
10404 SelectionDAG &DAG) {
10406 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10407 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10408 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10409 ArrayRef<int> Mask = SVOp->getMask();
10410 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10412 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10416 // Check for being able to broadcast a single element.
10417 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8f32, DL, V1,
10418 Mask, Subtarget, DAG))
10421 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10422 // options to efficiently lower the shuffle.
10423 SmallVector<int, 4> RepeatedMask;
10424 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10425 assert(RepeatedMask.size() == 4 &&
10426 "Repeated masks must be half the mask width!");
10427 if (isSingleInputShuffleMask(Mask))
10428 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10429 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10431 // Use dedicated unpack instructions for masks that match their pattern.
10432 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10433 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10434 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10435 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10437 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10438 // have already handled any direct blends. We also need to squash the
10439 // repeated mask into a simulated v4f32 mask.
10440 for (int i = 0; i < 4; ++i)
10441 if (RepeatedMask[i] >= 8)
10442 RepeatedMask[i] -= 4;
10443 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10446 // If we have a single input shuffle with different shuffle patterns in the
10447 // two 128-bit lanes use the variable mask to VPERMILPS.
10448 if (isSingleInputShuffleMask(Mask)) {
10449 SDValue VPermMask[8];
10450 for (int i = 0; i < 8; ++i)
10451 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10452 : DAG.getConstant(Mask[i], MVT::i32);
10453 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10454 return DAG.getNode(
10455 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10456 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10458 if (Subtarget->hasAVX2())
10459 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
10460 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
10461 DAG.getNode(ISD::BUILD_VECTOR, DL,
10462 MVT::v8i32, VPermMask)),
10465 // Otherwise, fall back.
10466 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10470 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10472 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10473 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10476 // If we have AVX2 then we always want to lower with a blend because at v8 we
10477 // can fully permute the elements.
10478 if (Subtarget->hasAVX2())
10479 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10482 // Otherwise fall back on generic lowering.
10483 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10486 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10488 /// This routine is only called when we have AVX2 and thus a reasonable
10489 /// instruction set for v8i32 shuffling..
10490 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10491 const X86Subtarget *Subtarget,
10492 SelectionDAG &DAG) {
10494 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10495 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10496 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10497 ArrayRef<int> Mask = SVOp->getMask();
10498 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10499 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10501 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10505 // Check for being able to broadcast a single element.
10506 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v8i32, DL, V1,
10507 Mask, Subtarget, DAG))
10510 // If the shuffle mask is repeated in each 128-bit lane we can use more
10511 // efficient instructions that mirror the shuffles across the two 128-bit
10513 SmallVector<int, 4> RepeatedMask;
10514 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10515 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10516 if (isSingleInputShuffleMask(Mask))
10517 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10518 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
10520 // Use dedicated unpack instructions for masks that match their pattern.
10521 if (isShuffleEquivalent(Mask, 0, 8, 1, 9, 4, 12, 5, 13))
10522 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10523 if (isShuffleEquivalent(Mask, 2, 10, 3, 11, 6, 14, 7, 15))
10524 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10527 // If the shuffle patterns aren't repeated but it is a single input, directly
10528 // generate a cross-lane VPERMD instruction.
10529 if (isSingleInputShuffleMask(Mask)) {
10530 SDValue VPermMask[8];
10531 for (int i = 0; i < 8; ++i)
10532 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10533 : DAG.getConstant(Mask[i], MVT::i32);
10534 return DAG.getNode(
10535 X86ISD::VPERMV, DL, MVT::v8i32,
10536 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10539 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10541 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10542 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10545 // Otherwise fall back on generic blend lowering.
10546 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10550 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10552 /// This routine is only called when we have AVX2 and thus a reasonable
10553 /// instruction set for v16i16 shuffling..
10554 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10555 const X86Subtarget *Subtarget,
10556 SelectionDAG &DAG) {
10558 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10559 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10560 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10561 ArrayRef<int> Mask = SVOp->getMask();
10562 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10563 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10565 // Check for being able to broadcast a single element.
10566 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v16i16, DL, V1,
10567 Mask, Subtarget, DAG))
10570 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10574 // Use dedicated unpack instructions for masks that match their pattern.
10575 if (isShuffleEquivalent(Mask,
10576 // First 128-bit lane:
10577 0, 16, 1, 17, 2, 18, 3, 19,
10578 // Second 128-bit lane:
10579 8, 24, 9, 25, 10, 26, 11, 27))
10580 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10581 if (isShuffleEquivalent(Mask,
10582 // First 128-bit lane:
10583 4, 20, 5, 21, 6, 22, 7, 23,
10584 // Second 128-bit lane:
10585 12, 28, 13, 29, 14, 30, 15, 31))
10586 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10588 if (isSingleInputShuffleMask(Mask)) {
10589 // There are no generalized cross-lane shuffle operations available on i16
10591 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10592 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10595 SDValue PSHUFBMask[32];
10596 for (int i = 0; i < 16; ++i) {
10597 if (Mask[i] == -1) {
10598 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10602 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10603 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10604 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
10605 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
10607 return DAG.getNode(
10608 ISD::BITCAST, DL, MVT::v16i16,
10610 X86ISD::PSHUFB, DL, MVT::v32i8,
10611 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
10612 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
10615 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10617 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10618 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10621 // Otherwise fall back on generic lowering.
10622 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10625 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10627 /// This routine is only called when we have AVX2 and thus a reasonable
10628 /// instruction set for v32i8 shuffling..
10629 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10630 const X86Subtarget *Subtarget,
10631 SelectionDAG &DAG) {
10633 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10634 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10635 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10636 ArrayRef<int> Mask = SVOp->getMask();
10637 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10638 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10640 // Check for being able to broadcast a single element.
10641 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(MVT::v32i8, DL, V1,
10642 Mask, Subtarget, DAG))
10645 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10649 // Use dedicated unpack instructions for masks that match their pattern.
10650 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10652 if (isShuffleEquivalent(
10654 // First 128-bit lane:
10655 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10656 // Second 128-bit lane:
10657 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55))
10658 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10659 if (isShuffleEquivalent(
10661 // First 128-bit lane:
10662 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10663 // Second 128-bit lane:
10664 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63))
10665 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10667 if (isSingleInputShuffleMask(Mask)) {
10668 // There are no generalized cross-lane shuffle operations available on i8
10670 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10671 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10674 SDValue PSHUFBMask[32];
10675 for (int i = 0; i < 32; ++i)
10678 ? DAG.getUNDEF(MVT::i8)
10679 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
10681 return DAG.getNode(
10682 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10683 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10686 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10688 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10689 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10692 // Otherwise fall back on generic lowering.
10693 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10696 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10698 /// This routine either breaks down the specific type of a 256-bit x86 vector
10699 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10700 /// together based on the available instructions.
10701 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10702 MVT VT, const X86Subtarget *Subtarget,
10703 SelectionDAG &DAG) {
10705 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10706 ArrayRef<int> Mask = SVOp->getMask();
10708 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10709 // check for those subtargets here and avoid much of the subtarget querying in
10710 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10711 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10712 // floating point types there eventually, just immediately cast everything to
10713 // a float and operate entirely in that domain.
10714 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10715 int ElementBits = VT.getScalarSizeInBits();
10716 if (ElementBits < 32)
10717 // No floating point type available, decompose into 128-bit vectors.
10718 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10720 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10721 VT.getVectorNumElements());
10722 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
10723 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
10724 return DAG.getNode(ISD::BITCAST, DL, VT,
10725 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10728 switch (VT.SimpleTy) {
10730 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10732 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10734 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10736 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10738 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10740 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10743 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10747 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10748 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10749 const X86Subtarget *Subtarget,
10750 SelectionDAG &DAG) {
10752 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10753 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10754 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10755 ArrayRef<int> Mask = SVOp->getMask();
10756 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10758 // FIXME: Implement direct support for this type!
10759 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10762 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10763 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10764 const X86Subtarget *Subtarget,
10765 SelectionDAG &DAG) {
10767 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10768 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10769 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10770 ArrayRef<int> Mask = SVOp->getMask();
10771 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10773 // FIXME: Implement direct support for this type!
10774 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10777 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10778 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10779 const X86Subtarget *Subtarget,
10780 SelectionDAG &DAG) {
10782 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10783 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10784 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10785 ArrayRef<int> Mask = SVOp->getMask();
10786 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10788 // FIXME: Implement direct support for this type!
10789 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10792 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10793 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10794 const X86Subtarget *Subtarget,
10795 SelectionDAG &DAG) {
10797 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10798 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10799 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10800 ArrayRef<int> Mask = SVOp->getMask();
10801 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10803 // FIXME: Implement direct support for this type!
10804 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10807 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10808 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10809 const X86Subtarget *Subtarget,
10810 SelectionDAG &DAG) {
10812 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10813 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10814 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10815 ArrayRef<int> Mask = SVOp->getMask();
10816 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10817 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10819 // FIXME: Implement direct support for this type!
10820 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10823 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10824 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10825 const X86Subtarget *Subtarget,
10826 SelectionDAG &DAG) {
10828 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10829 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10830 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10831 ArrayRef<int> Mask = SVOp->getMask();
10832 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10833 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10835 // FIXME: Implement direct support for this type!
10836 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10839 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10841 /// This routine either breaks down the specific type of a 512-bit x86 vector
10842 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10843 /// together based on the available instructions.
10844 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10845 MVT VT, const X86Subtarget *Subtarget,
10846 SelectionDAG &DAG) {
10848 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10849 ArrayRef<int> Mask = SVOp->getMask();
10850 assert(Subtarget->hasAVX512() &&
10851 "Cannot lower 512-bit vectors w/ basic ISA!");
10853 // Check for being able to broadcast a single element.
10854 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
10855 Mask, Subtarget, DAG))
10858 // Dispatch to each element type for lowering. If we don't have supprot for
10859 // specific element type shuffles at 512 bits, immediately split them and
10860 // lower them. Each lowering routine of a given type is allowed to assume that
10861 // the requisite ISA extensions for that element type are available.
10862 switch (VT.SimpleTy) {
10864 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10866 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10868 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10870 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10872 if (Subtarget->hasBWI())
10873 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10876 if (Subtarget->hasBWI())
10877 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10881 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10884 // Otherwise fall back on splitting.
10885 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10888 /// \brief Top-level lowering for x86 vector shuffles.
10890 /// This handles decomposition, canonicalization, and lowering of all x86
10891 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10892 /// above in helper routines. The canonicalization attempts to widen shuffles
10893 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10894 /// s.t. only one of the two inputs needs to be tested, etc.
10895 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10896 SelectionDAG &DAG) {
10897 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10898 ArrayRef<int> Mask = SVOp->getMask();
10899 SDValue V1 = Op.getOperand(0);
10900 SDValue V2 = Op.getOperand(1);
10901 MVT VT = Op.getSimpleValueType();
10902 int NumElements = VT.getVectorNumElements();
10905 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10907 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10908 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10909 if (V1IsUndef && V2IsUndef)
10910 return DAG.getUNDEF(VT);
10912 // When we create a shuffle node we put the UNDEF node to second operand,
10913 // but in some cases the first operand may be transformed to UNDEF.
10914 // In this case we should just commute the node.
10916 return DAG.getCommutedVectorShuffle(*SVOp);
10918 // Check for non-undef masks pointing at an undef vector and make the masks
10919 // undef as well. This makes it easier to match the shuffle based solely on
10923 if (M >= NumElements) {
10924 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10925 for (int &M : NewMask)
10926 if (M >= NumElements)
10928 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10931 // Try to collapse shuffles into using a vector type with fewer elements but
10932 // wider element types. We cap this to not form integers or floating point
10933 // elements wider than 64 bits, but it might be interesting to form i128
10934 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10935 SmallVector<int, 16> WidenedMask;
10936 if (VT.getScalarSizeInBits() < 64 &&
10937 canWidenShuffleElements(Mask, WidenedMask)) {
10938 MVT NewEltVT = VT.isFloatingPoint()
10939 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10940 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10941 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10942 // Make sure that the new vector type is legal. For example, v2f64 isn't
10944 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10945 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
10946 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
10947 return DAG.getNode(ISD::BITCAST, dl, VT,
10948 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10952 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10953 for (int M : SVOp->getMask())
10955 ++NumUndefElements;
10956 else if (M < NumElements)
10961 // Commute the shuffle as needed such that more elements come from V1 than
10962 // V2. This allows us to match the shuffle pattern strictly on how many
10963 // elements come from V1 without handling the symmetric cases.
10964 if (NumV2Elements > NumV1Elements)
10965 return DAG.getCommutedVectorShuffle(*SVOp);
10967 // When the number of V1 and V2 elements are the same, try to minimize the
10968 // number of uses of V2 in the low half of the vector. When that is tied,
10969 // ensure that the sum of indices for V1 is equal to or lower than the sum
10970 // indices for V2. When those are equal, try to ensure that the number of odd
10971 // indices for V1 is lower than the number of odd indices for V2.
10972 if (NumV1Elements == NumV2Elements) {
10973 int LowV1Elements = 0, LowV2Elements = 0;
10974 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10975 if (M >= NumElements)
10979 if (LowV2Elements > LowV1Elements) {
10980 return DAG.getCommutedVectorShuffle(*SVOp);
10981 } else if (LowV2Elements == LowV1Elements) {
10982 int SumV1Indices = 0, SumV2Indices = 0;
10983 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10984 if (SVOp->getMask()[i] >= NumElements)
10986 else if (SVOp->getMask()[i] >= 0)
10988 if (SumV2Indices < SumV1Indices) {
10989 return DAG.getCommutedVectorShuffle(*SVOp);
10990 } else if (SumV2Indices == SumV1Indices) {
10991 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10992 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10993 if (SVOp->getMask()[i] >= NumElements)
10994 NumV2OddIndices += i % 2;
10995 else if (SVOp->getMask()[i] >= 0)
10996 NumV1OddIndices += i % 2;
10997 if (NumV2OddIndices < NumV1OddIndices)
10998 return DAG.getCommutedVectorShuffle(*SVOp);
11003 // For each vector width, delegate to a specialized lowering routine.
11004 if (VT.getSizeInBits() == 128)
11005 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11007 if (VT.getSizeInBits() == 256)
11008 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11010 // Force AVX-512 vectors to be scalarized for now.
11011 // FIXME: Implement AVX-512 support!
11012 if (VT.getSizeInBits() == 512)
11013 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11015 llvm_unreachable("Unimplemented!");
11019 //===----------------------------------------------------------------------===//
11020 // Legacy vector shuffle lowering
11022 // This code is the legacy code handling vector shuffles until the above
11023 // replaces its functionality and performance.
11024 //===----------------------------------------------------------------------===//
11026 static bool isBlendMask(ArrayRef<int> MaskVals, MVT VT, bool hasSSE41,
11027 bool hasInt256, unsigned *MaskOut = nullptr) {
11028 MVT EltVT = VT.getVectorElementType();
11030 // There is no blend with immediate in AVX-512.
11031 if (VT.is512BitVector())
11034 if (!hasSSE41 || EltVT == MVT::i8)
11036 if (!hasInt256 && VT == MVT::v16i16)
11039 unsigned MaskValue = 0;
11040 unsigned NumElems = VT.getVectorNumElements();
11041 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11042 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11043 unsigned NumElemsInLane = NumElems / NumLanes;
11045 // Blend for v16i16 should be symetric for the both lanes.
11046 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11048 int SndLaneEltIdx = (NumLanes == 2) ? MaskVals[i + NumElemsInLane] : -1;
11049 int EltIdx = MaskVals[i];
11051 if ((EltIdx < 0 || EltIdx == (int)i) &&
11052 (SndLaneEltIdx < 0 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
11055 if (((unsigned)EltIdx == (i + NumElems)) &&
11056 (SndLaneEltIdx < 0 ||
11057 (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
11058 MaskValue |= (1 << i);
11064 *MaskOut = MaskValue;
11068 // Try to lower a shuffle node into a simple blend instruction.
11069 // This function assumes isBlendMask returns true for this
11070 // SuffleVectorSDNode
11071 static SDValue LowerVECTOR_SHUFFLEtoBlend(ShuffleVectorSDNode *SVOp,
11072 unsigned MaskValue,
11073 const X86Subtarget *Subtarget,
11074 SelectionDAG &DAG) {
11075 MVT VT = SVOp->getSimpleValueType(0);
11076 MVT EltVT = VT.getVectorElementType();
11077 assert(isBlendMask(SVOp->getMask(), VT, Subtarget->hasSSE41(),
11078 Subtarget->hasInt256() && "Trying to lower a "
11079 "VECTOR_SHUFFLE to a Blend but "
11080 "with the wrong mask"));
11081 SDValue V1 = SVOp->getOperand(0);
11082 SDValue V2 = SVOp->getOperand(1);
11084 unsigned NumElems = VT.getVectorNumElements();
11086 // Convert i32 vectors to floating point if it is not AVX2.
11087 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
11089 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
11090 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
11092 V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
11093 V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
11096 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
11097 DAG.getConstant(MaskValue, MVT::i32));
11098 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
11101 /// In vector type \p VT, return true if the element at index \p InputIdx
11102 /// falls on a different 128-bit lane than \p OutputIdx.
11103 static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
11104 unsigned OutputIdx) {
11105 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
11106 return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
11109 /// Generate a PSHUFB if possible. Selects elements from \p V1 according to
11110 /// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
11111 /// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
11112 /// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
11114 static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
11115 SelectionDAG &DAG) {
11116 MVT VT = V1.getSimpleValueType();
11117 assert(VT.is128BitVector() || VT.is256BitVector());
11119 MVT EltVT = VT.getVectorElementType();
11120 unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
11121 unsigned NumElts = VT.getVectorNumElements();
11123 SmallVector<SDValue, 32> PshufbMask;
11124 for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
11125 int InputIdx = MaskVals[OutputIdx];
11126 unsigned InputByteIdx;
11128 if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
11129 InputByteIdx = 0x80;
11131 // Cross lane is not allowed.
11132 if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
11134 InputByteIdx = InputIdx * EltSizeInBytes;
11135 // Index is an byte offset within the 128-bit lane.
11136 InputByteIdx &= 0xf;
11139 for (unsigned j = 0; j < EltSizeInBytes; ++j) {
11140 PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
11141 if (InputByteIdx != 0x80)
11146 MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
11148 V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
11149 return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
11150 DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT, PshufbMask));
11153 // v8i16 shuffles - Prefer shuffles in the following order:
11154 // 1. [all] pshuflw, pshufhw, optional move
11155 // 2. [ssse3] 1 x pshufb
11156 // 3. [ssse3] 2 x pshufb + 1 x por
11157 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
11159 LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
11160 SelectionDAG &DAG) {
11161 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11162 SDValue V1 = SVOp->getOperand(0);
11163 SDValue V2 = SVOp->getOperand(1);
11165 SmallVector<int, 8> MaskVals;
11167 // Determine if more than 1 of the words in each of the low and high quadwords
11168 // of the result come from the same quadword of one of the two inputs. Undef
11169 // mask values count as coming from any quadword, for better codegen.
11171 // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
11172 // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
11173 unsigned LoQuad[] = { 0, 0, 0, 0 };
11174 unsigned HiQuad[] = { 0, 0, 0, 0 };
11175 // Indices of quads used.
11176 std::bitset<4> InputQuads;
11177 for (unsigned i = 0; i < 8; ++i) {
11178 unsigned *Quad = i < 4 ? LoQuad : HiQuad;
11179 int EltIdx = SVOp->getMaskElt(i);
11180 MaskVals.push_back(EltIdx);
11188 ++Quad[EltIdx / 4];
11189 InputQuads.set(EltIdx / 4);
11192 int BestLoQuad = -1;
11193 unsigned MaxQuad = 1;
11194 for (unsigned i = 0; i < 4; ++i) {
11195 if (LoQuad[i] > MaxQuad) {
11197 MaxQuad = LoQuad[i];
11201 int BestHiQuad = -1;
11203 for (unsigned i = 0; i < 4; ++i) {
11204 if (HiQuad[i] > MaxQuad) {
11206 MaxQuad = HiQuad[i];
11210 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
11211 // of the two input vectors, shuffle them into one input vector so only a
11212 // single pshufb instruction is necessary. If there are more than 2 input
11213 // quads, disable the next transformation since it does not help SSSE3.
11214 bool V1Used = InputQuads[0] || InputQuads[1];
11215 bool V2Used = InputQuads[2] || InputQuads[3];
11216 if (Subtarget->hasSSSE3()) {
11217 if (InputQuads.count() == 2 && V1Used && V2Used) {
11218 BestLoQuad = InputQuads[0] ? 0 : 1;
11219 BestHiQuad = InputQuads[2] ? 2 : 3;
11221 if (InputQuads.count() > 2) {
11227 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
11228 // the shuffle mask. If a quad is scored as -1, that means that it contains
11229 // words from all 4 input quadwords.
11231 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
11233 BestLoQuad < 0 ? 0 : BestLoQuad,
11234 BestHiQuad < 0 ? 1 : BestHiQuad
11236 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
11237 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
11238 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
11239 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
11241 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
11242 // source words for the shuffle, to aid later transformations.
11243 bool AllWordsInNewV = true;
11244 bool InOrder[2] = { true, true };
11245 for (unsigned i = 0; i != 8; ++i) {
11246 int idx = MaskVals[i];
11248 InOrder[i/4] = false;
11249 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
11251 AllWordsInNewV = false;
11255 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
11256 if (AllWordsInNewV) {
11257 for (int i = 0; i != 8; ++i) {
11258 int idx = MaskVals[i];
11261 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
11262 if ((idx != i) && idx < 4)
11264 if ((idx != i) && idx > 3)
11273 // If we've eliminated the use of V2, and the new mask is a pshuflw or
11274 // pshufhw, that's as cheap as it gets. Return the new shuffle.
11275 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
11276 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
11277 unsigned TargetMask = 0;
11278 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
11279 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
11280 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11281 TargetMask = pshufhw ? getShufflePSHUFHWImmediate(SVOp):
11282 getShufflePSHUFLWImmediate(SVOp);
11283 V1 = NewV.getOperand(0);
11284 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
11288 // Promote splats to a larger type which usually leads to more efficient code.
11289 // FIXME: Is this true if pshufb is available?
11290 if (SVOp->isSplat())
11291 return PromoteSplat(SVOp, DAG);
11293 // If we have SSSE3, and all words of the result are from 1 input vector,
11294 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
11295 // is present, fall back to case 4.
11296 if (Subtarget->hasSSSE3()) {
11297 SmallVector<SDValue,16> pshufbMask;
11299 // If we have elements from both input vectors, set the high bit of the
11300 // shuffle mask element to zero out elements that come from V2 in the V1
11301 // mask, and elements that come from V1 in the V2 mask, so that the two
11302 // results can be OR'd together.
11303 bool TwoInputs = V1Used && V2Used;
11304 V1 = getPSHUFB(MaskVals, V1, dl, DAG);
11306 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11308 // Calculate the shuffle mask for the second input, shuffle it, and
11309 // OR it with the first shuffled input.
11310 CommuteVectorShuffleMask(MaskVals, 8);
11311 V2 = getPSHUFB(MaskVals, V2, dl, DAG);
11312 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
11313 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11316 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
11317 // and update MaskVals with new element order.
11318 std::bitset<8> InOrder;
11319 if (BestLoQuad >= 0) {
11320 int MaskV[] = { -1, -1, -1, -1, 4, 5, 6, 7 };
11321 for (int i = 0; i != 4; ++i) {
11322 int idx = MaskVals[i];
11325 } else if ((idx / 4) == BestLoQuad) {
11326 MaskV[i] = idx & 3;
11330 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11333 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11334 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11335 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
11336 NewV.getOperand(0),
11337 getShufflePSHUFLWImmediate(SVOp), DAG);
11341 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
11342 // and update MaskVals with the new element order.
11343 if (BestHiQuad >= 0) {
11344 int MaskV[] = { 0, 1, 2, 3, -1, -1, -1, -1 };
11345 for (unsigned i = 4; i != 8; ++i) {
11346 int idx = MaskVals[i];
11349 } else if ((idx / 4) == BestHiQuad) {
11350 MaskV[i] = (idx & 3) + 4;
11354 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
11357 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSE2()) {
11358 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(NewV.getNode());
11359 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
11360 NewV.getOperand(0),
11361 getShufflePSHUFHWImmediate(SVOp), DAG);
11365 // In case BestHi & BestLo were both -1, which means each quadword has a word
11366 // from each of the four input quadwords, calculate the InOrder bitvector now
11367 // before falling through to the insert/extract cleanup.
11368 if (BestLoQuad == -1 && BestHiQuad == -1) {
11370 for (int i = 0; i != 8; ++i)
11371 if (MaskVals[i] < 0 || MaskVals[i] == i)
11375 // The other elements are put in the right place using pextrw and pinsrw.
11376 for (unsigned i = 0; i != 8; ++i) {
11379 int EltIdx = MaskVals[i];
11382 SDValue ExtOp = (EltIdx < 8) ?
11383 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
11384 DAG.getIntPtrConstant(EltIdx)) :
11385 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
11386 DAG.getIntPtrConstant(EltIdx - 8));
11387 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
11388 DAG.getIntPtrConstant(i));
11393 /// \brief v16i16 shuffles
11395 /// FIXME: We only support generation of a single pshufb currently. We can
11396 /// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
11397 /// well (e.g 2 x pshufb + 1 x por).
11399 LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
11400 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11401 SDValue V1 = SVOp->getOperand(0);
11402 SDValue V2 = SVOp->getOperand(1);
11405 if (V2.getOpcode() != ISD::UNDEF)
11408 SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
11409 return getPSHUFB(MaskVals, V1, dl, DAG);
11412 // v16i8 shuffles - Prefer shuffles in the following order:
11413 // 1. [ssse3] 1 x pshufb
11414 // 2. [ssse3] 2 x pshufb + 1 x por
11415 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
11416 static SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
11417 const X86Subtarget* Subtarget,
11418 SelectionDAG &DAG) {
11419 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11420 SDValue V1 = SVOp->getOperand(0);
11421 SDValue V2 = SVOp->getOperand(1);
11423 ArrayRef<int> MaskVals = SVOp->getMask();
11425 // Promote splats to a larger type which usually leads to more efficient code.
11426 // FIXME: Is this true if pshufb is available?
11427 if (SVOp->isSplat())
11428 return PromoteSplat(SVOp, DAG);
11430 // If we have SSSE3, case 1 is generated when all result bytes come from
11431 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
11432 // present, fall back to case 3.
11434 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
11435 if (Subtarget->hasSSSE3()) {
11436 SmallVector<SDValue,16> pshufbMask;
11438 // If all result elements are from one input vector, then only translate
11439 // undef mask values to 0x80 (zero out result) in the pshufb mask.
11441 // Otherwise, we have elements from both input vectors, and must zero out
11442 // elements that come from V2 in the first mask, and V1 in the second mask
11443 // so that we can OR them together.
11444 for (unsigned i = 0; i != 16; ++i) {
11445 int EltIdx = MaskVals[i];
11446 if (EltIdx < 0 || EltIdx >= 16)
11448 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
11450 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
11451 DAG.getNode(ISD::BUILD_VECTOR, dl,
11452 MVT::v16i8, pshufbMask));
11454 // As PSHUFB will zero elements with negative indices, it's safe to ignore
11455 // the 2nd operand if it's undefined or zero.
11456 if (V2.getOpcode() == ISD::UNDEF ||
11457 ISD::isBuildVectorAllZeros(V2.getNode()))
11460 // Calculate the shuffle mask for the second input, shuffle it, and
11461 // OR it with the first shuffled input.
11462 pshufbMask.clear();
11463 for (unsigned i = 0; i != 16; ++i) {
11464 int EltIdx = MaskVals[i];
11465 EltIdx = (EltIdx < 16) ? 0x80 : EltIdx - 16;
11466 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
11468 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
11469 DAG.getNode(ISD::BUILD_VECTOR, dl,
11470 MVT::v16i8, pshufbMask));
11471 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
11474 // No SSSE3 - Calculate in place words and then fix all out of place words
11475 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
11476 // the 16 different words that comprise the two doublequadword input vectors.
11477 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
11478 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
11480 for (int i = 0; i != 8; ++i) {
11481 int Elt0 = MaskVals[i*2];
11482 int Elt1 = MaskVals[i*2+1];
11484 // This word of the result is all undef, skip it.
11485 if (Elt0 < 0 && Elt1 < 0)
11488 // This word of the result is already in the correct place, skip it.
11489 if ((Elt0 == i*2) && (Elt1 == i*2+1))
11492 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
11493 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
11496 // If Elt0 and Elt1 are defined, are consecutive, and can be load
11497 // using a single extract together, load it and store it.
11498 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
11499 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
11500 DAG.getIntPtrConstant(Elt1 / 2));
11501 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
11502 DAG.getIntPtrConstant(i));
11506 // If Elt1 is defined, extract it from the appropriate source. If the
11507 // source byte is not also odd, shift the extracted word left 8 bits
11508 // otherwise clear the bottom 8 bits if we need to do an or.
11510 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
11511 DAG.getIntPtrConstant(Elt1 / 2));
11512 if ((Elt1 & 1) == 0)
11513 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
11515 TLI.getShiftAmountTy(InsElt.getValueType())));
11516 else if (Elt0 >= 0)
11517 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
11518 DAG.getConstant(0xFF00, MVT::i16));
11520 // If Elt0 is defined, extract it from the appropriate source. If the
11521 // source byte is not also even, shift the extracted word right 8 bits. If
11522 // Elt1 was also defined, OR the extracted values together before
11523 // inserting them in the result.
11525 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
11526 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
11527 if ((Elt0 & 1) != 0)
11528 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
11530 TLI.getShiftAmountTy(InsElt0.getValueType())));
11531 else if (Elt1 >= 0)
11532 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
11533 DAG.getConstant(0x00FF, MVT::i16));
11534 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
11537 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
11538 DAG.getIntPtrConstant(i));
11540 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
11543 // v32i8 shuffles - Translate to VPSHUFB if possible.
11545 SDValue LowerVECTOR_SHUFFLEv32i8(ShuffleVectorSDNode *SVOp,
11546 const X86Subtarget *Subtarget,
11547 SelectionDAG &DAG) {
11548 MVT VT = SVOp->getSimpleValueType(0);
11549 SDValue V1 = SVOp->getOperand(0);
11550 SDValue V2 = SVOp->getOperand(1);
11552 SmallVector<int, 32> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
11554 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11555 bool V1IsAllZero = ISD::isBuildVectorAllZeros(V1.getNode());
11556 bool V2IsAllZero = ISD::isBuildVectorAllZeros(V2.getNode());
11558 // VPSHUFB may be generated if
11559 // (1) one of input vector is undefined or zeroinitializer.
11560 // The mask value 0x80 puts 0 in the corresponding slot of the vector.
11561 // And (2) the mask indexes don't cross the 128-bit lane.
11562 if (VT != MVT::v32i8 || !Subtarget->hasInt256() ||
11563 (!V2IsUndef && !V2IsAllZero && !V1IsAllZero))
11566 if (V1IsAllZero && !V2IsAllZero) {
11567 CommuteVectorShuffleMask(MaskVals, 32);
11570 return getPSHUFB(MaskVals, V1, dl, DAG);
11573 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
11574 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
11575 /// done when every pair / quad of shuffle mask elements point to elements in
11576 /// the right sequence. e.g.
11577 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
11579 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
11580 SelectionDAG &DAG) {
11581 MVT VT = SVOp->getSimpleValueType(0);
11583 unsigned NumElems = VT.getVectorNumElements();
11586 switch (VT.SimpleTy) {
11587 default: llvm_unreachable("Unexpected!");
11590 return SDValue(SVOp, 0);
11591 case MVT::v4f32: NewVT = MVT::v2f64; Scale = 2; break;
11592 case MVT::v4i32: NewVT = MVT::v2i64; Scale = 2; break;
11593 case MVT::v8i16: NewVT = MVT::v4i32; Scale = 2; break;
11594 case MVT::v16i8: NewVT = MVT::v4i32; Scale = 4; break;
11595 case MVT::v16i16: NewVT = MVT::v8i32; Scale = 2; break;
11596 case MVT::v32i8: NewVT = MVT::v8i32; Scale = 4; break;
11599 SmallVector<int, 8> MaskVec;
11600 for (unsigned i = 0; i != NumElems; i += Scale) {
11602 for (unsigned j = 0; j != Scale; ++j) {
11603 int EltIdx = SVOp->getMaskElt(i+j);
11607 StartIdx = (EltIdx / Scale);
11608 if (EltIdx != (int)(StartIdx*Scale + j))
11611 MaskVec.push_back(StartIdx);
11614 SDValue V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(0));
11615 SDValue V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, SVOp->getOperand(1));
11616 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
11619 /// getVZextMovL - Return a zero-extending vector move low node.
11621 static SDValue getVZextMovL(MVT VT, MVT OpVT,
11622 SDValue SrcOp, SelectionDAG &DAG,
11623 const X86Subtarget *Subtarget, SDLoc dl) {
11624 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
11625 LoadSDNode *LD = nullptr;
11626 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
11627 LD = dyn_cast<LoadSDNode>(SrcOp);
11629 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
11631 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
11632 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
11633 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11634 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
11635 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
11637 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
11638 return DAG.getNode(ISD::BITCAST, dl, VT,
11639 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
11640 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11642 SrcOp.getOperand(0)
11648 return DAG.getNode(ISD::BITCAST, dl, VT,
11649 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
11650 DAG.getNode(ISD::BITCAST, dl,
11654 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
11655 /// which could not be matched by any known target speficic shuffle
11657 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
11659 SDValue NewOp = Compact8x32ShuffleNode(SVOp, DAG);
11660 if (NewOp.getNode())
11663 MVT VT = SVOp->getSimpleValueType(0);
11665 unsigned NumElems = VT.getVectorNumElements();
11666 unsigned NumLaneElems = NumElems / 2;
11669 MVT EltVT = VT.getVectorElementType();
11670 MVT NVT = MVT::getVectorVT(EltVT, NumLaneElems);
11673 SmallVector<int, 16> Mask;
11674 for (unsigned l = 0; l < 2; ++l) {
11675 // Build a shuffle mask for the output, discovering on the fly which
11676 // input vectors to use as shuffle operands (recorded in InputUsed).
11677 // If building a suitable shuffle vector proves too hard, then bail
11678 // out with UseBuildVector set.
11679 bool UseBuildVector = false;
11680 int InputUsed[2] = { -1, -1 }; // Not yet discovered.
11681 unsigned LaneStart = l * NumLaneElems;
11682 for (unsigned i = 0; i != NumLaneElems; ++i) {
11683 // The mask element. This indexes into the input.
11684 int Idx = SVOp->getMaskElt(i+LaneStart);
11686 // the mask element does not index into any input vector.
11687 Mask.push_back(-1);
11691 // The input vector this mask element indexes into.
11692 int Input = Idx / NumLaneElems;
11694 // Turn the index into an offset from the start of the input vector.
11695 Idx -= Input * NumLaneElems;
11697 // Find or create a shuffle vector operand to hold this input.
11699 for (OpNo = 0; OpNo < array_lengthof(InputUsed); ++OpNo) {
11700 if (InputUsed[OpNo] == Input)
11701 // This input vector is already an operand.
11703 if (InputUsed[OpNo] < 0) {
11704 // Create a new operand for this input vector.
11705 InputUsed[OpNo] = Input;
11710 if (OpNo >= array_lengthof(InputUsed)) {
11711 // More than two input vectors used! Give up on trying to create a
11712 // shuffle vector. Insert all elements into a BUILD_VECTOR instead.
11713 UseBuildVector = true;
11717 // Add the mask index for the new shuffle vector.
11718 Mask.push_back(Idx + OpNo * NumLaneElems);
11721 if (UseBuildVector) {
11722 SmallVector<SDValue, 16> SVOps;
11723 for (unsigned i = 0; i != NumLaneElems; ++i) {
11724 // The mask element. This indexes into the input.
11725 int Idx = SVOp->getMaskElt(i+LaneStart);
11727 SVOps.push_back(DAG.getUNDEF(EltVT));
11731 // The input vector this mask element indexes into.
11732 int Input = Idx / NumElems;
11734 // Turn the index into an offset from the start of the input vector.
11735 Idx -= Input * NumElems;
11737 // Extract the vector element by hand.
11738 SVOps.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
11739 SVOp->getOperand(Input),
11740 DAG.getIntPtrConstant(Idx)));
11743 // Construct the output using a BUILD_VECTOR.
11744 Output[l] = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, SVOps);
11745 } else if (InputUsed[0] < 0) {
11746 // No input vectors were used! The result is undefined.
11747 Output[l] = DAG.getUNDEF(NVT);
11749 SDValue Op0 = Extract128BitVector(SVOp->getOperand(InputUsed[0] / 2),
11750 (InputUsed[0] % 2) * NumLaneElems,
11752 // If only one input was used, use an undefined vector for the other.
11753 SDValue Op1 = (InputUsed[1] < 0) ? DAG.getUNDEF(NVT) :
11754 Extract128BitVector(SVOp->getOperand(InputUsed[1] / 2),
11755 (InputUsed[1] % 2) * NumLaneElems, DAG, dl);
11756 // At least one input vector was used. Create a new shuffle vector.
11757 Output[l] = DAG.getVectorShuffle(NVT, dl, Op0, Op1, &Mask[0]);
11763 // Concatenate the result back
11764 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Output[0], Output[1]);
11767 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
11768 /// 4 elements, and match them with several different shuffle types.
11770 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
11771 SDValue V1 = SVOp->getOperand(0);
11772 SDValue V2 = SVOp->getOperand(1);
11774 MVT VT = SVOp->getSimpleValueType(0);
11776 assert(VT.is128BitVector() && "Unsupported vector size");
11778 std::pair<int, int> Locs[4];
11779 int Mask1[] = { -1, -1, -1, -1 };
11780 SmallVector<int, 8> PermMask(SVOp->getMask().begin(), SVOp->getMask().end());
11782 unsigned NumHi = 0;
11783 unsigned NumLo = 0;
11784 for (unsigned i = 0; i != 4; ++i) {
11785 int Idx = PermMask[i];
11787 Locs[i] = std::make_pair(-1, -1);
11789 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
11791 Locs[i] = std::make_pair(0, NumLo);
11792 Mask1[NumLo] = Idx;
11795 Locs[i] = std::make_pair(1, NumHi);
11797 Mask1[2+NumHi] = Idx;
11803 if (NumLo <= 2 && NumHi <= 2) {
11804 // If no more than two elements come from either vector. This can be
11805 // implemented with two shuffles. First shuffle gather the elements.
11806 // The second shuffle, which takes the first shuffle as both of its
11807 // vector operands, put the elements into the right order.
11808 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
11810 int Mask2[] = { -1, -1, -1, -1 };
11812 for (unsigned i = 0; i != 4; ++i)
11813 if (Locs[i].first != -1) {
11814 unsigned Idx = (i < 2) ? 0 : 4;
11815 Idx += Locs[i].first * 2 + Locs[i].second;
11819 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
11822 if (NumLo == 3 || NumHi == 3) {
11823 // Otherwise, we must have three elements from one vector, call it X, and
11824 // one element from the other, call it Y. First, use a shufps to build an
11825 // intermediate vector with the one element from Y and the element from X
11826 // that will be in the same half in the final destination (the indexes don't
11827 // matter). Then, use a shufps to build the final vector, taking the half
11828 // containing the element from Y from the intermediate, and the other half
11831 // Normalize it so the 3 elements come from V1.
11832 CommuteVectorShuffleMask(PermMask, 4);
11836 // Find the element from V2.
11838 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
11839 int Val = PermMask[HiIndex];
11846 Mask1[0] = PermMask[HiIndex];
11848 Mask1[2] = PermMask[HiIndex^1];
11850 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
11852 if (HiIndex >= 2) {
11853 Mask1[0] = PermMask[0];
11854 Mask1[1] = PermMask[1];
11855 Mask1[2] = HiIndex & 1 ? 6 : 4;
11856 Mask1[3] = HiIndex & 1 ? 4 : 6;
11857 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
11860 Mask1[0] = HiIndex & 1 ? 2 : 0;
11861 Mask1[1] = HiIndex & 1 ? 0 : 2;
11862 Mask1[2] = PermMask[2];
11863 Mask1[3] = PermMask[3];
11868 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
11871 // Break it into (shuffle shuffle_hi, shuffle_lo).
11872 int LoMask[] = { -1, -1, -1, -1 };
11873 int HiMask[] = { -1, -1, -1, -1 };
11875 int *MaskPtr = LoMask;
11876 unsigned MaskIdx = 0;
11877 unsigned LoIdx = 0;
11878 unsigned HiIdx = 2;
11879 for (unsigned i = 0; i != 4; ++i) {
11886 int Idx = PermMask[i];
11888 Locs[i] = std::make_pair(-1, -1);
11889 } else if (Idx < 4) {
11890 Locs[i] = std::make_pair(MaskIdx, LoIdx);
11891 MaskPtr[LoIdx] = Idx;
11894 Locs[i] = std::make_pair(MaskIdx, HiIdx);
11895 MaskPtr[HiIdx] = Idx;
11900 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
11901 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
11902 int MaskOps[] = { -1, -1, -1, -1 };
11903 for (unsigned i = 0; i != 4; ++i)
11904 if (Locs[i].first != -1)
11905 MaskOps[i] = Locs[i].first * 4 + Locs[i].second;
11906 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
11909 static bool MayFoldVectorLoad(SDValue V) {
11910 while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
11911 V = V.getOperand(0);
11913 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
11914 V = V.getOperand(0);
11915 if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
11916 V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
11917 // BUILD_VECTOR (load), undef
11918 V = V.getOperand(0);
11920 return MayFoldLoad(V);
11924 SDValue getMOVDDup(SDValue &Op, SDLoc &dl, SDValue V1, SelectionDAG &DAG) {
11925 MVT VT = Op.getSimpleValueType();
11927 // Canonizalize to v2f64.
11928 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
11929 return DAG.getNode(ISD::BITCAST, dl, VT,
11930 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
11935 SDValue getMOVLowToHigh(SDValue &Op, SDLoc &dl, SelectionDAG &DAG,
11937 SDValue V1 = Op.getOperand(0);
11938 SDValue V2 = Op.getOperand(1);
11939 MVT VT = Op.getSimpleValueType();
11941 assert(VT != MVT::v2i64 && "unsupported shuffle type");
11943 if (HasSSE2 && VT == MVT::v2f64)
11944 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
11946 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
11947 return DAG.getNode(ISD::BITCAST, dl, VT,
11948 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
11949 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
11950 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
11954 SDValue getMOVHighToLow(SDValue &Op, SDLoc &dl, SelectionDAG &DAG) {
11955 SDValue V1 = Op.getOperand(0);
11956 SDValue V2 = Op.getOperand(1);
11957 MVT VT = Op.getSimpleValueType();
11959 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
11960 "unsupported shuffle type");
11962 if (V2.getOpcode() == ISD::UNDEF)
11966 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
11970 SDValue getMOVLP(SDValue &Op, SDLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
11971 SDValue V1 = Op.getOperand(0);
11972 SDValue V2 = Op.getOperand(1);
11973 MVT VT = Op.getSimpleValueType();
11974 unsigned NumElems = VT.getVectorNumElements();
11976 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
11977 // operand of these instructions is only memory, so check if there's a
11978 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
11980 bool CanFoldLoad = false;
11982 // Trivial case, when V2 comes from a load.
11983 if (MayFoldVectorLoad(V2))
11984 CanFoldLoad = true;
11986 // When V1 is a load, it can be folded later into a store in isel, example:
11987 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
11989 // (MOVLPSmr addr:$src1, VR128:$src2)
11990 // So, recognize this potential and also use MOVLPS or MOVLPD
11991 else if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
11992 CanFoldLoad = true;
11994 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11996 if (HasSSE2 && NumElems == 2)
11997 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
12000 // If we don't care about the second element, proceed to use movss.
12001 if (SVOp->getMaskElt(1) != -1)
12002 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
12005 // movl and movlp will both match v2i64, but v2i64 is never matched by
12006 // movl earlier because we make it strict to avoid messing with the movlp load
12007 // folding logic (see the code above getMOVLP call). Match it here then,
12008 // this is horrible, but will stay like this until we move all shuffle
12009 // matching to x86 specific nodes. Note that for the 1st condition all
12010 // types are matched with movsd.
12012 // FIXME: isMOVLMask should be checked and matched before getMOVLP,
12013 // as to remove this logic from here, as much as possible
12014 if (NumElems == 2 || !isMOVLMask(SVOp->getMask(), VT))
12015 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12016 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12019 assert(VT != MVT::v4i32 && "unsupported shuffle type");
12021 // Invert the operand order and use SHUFPS to match it.
12022 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V2, V1,
12023 getShuffleSHUFImmediate(SVOp), DAG);
12026 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
12027 SelectionDAG &DAG) {
12029 MVT VT = Load->getSimpleValueType(0);
12030 MVT EVT = VT.getVectorElementType();
12031 SDValue Addr = Load->getOperand(1);
12032 SDValue NewAddr = DAG.getNode(
12033 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
12034 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
12037 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
12038 DAG.getMachineFunction().getMachineMemOperand(
12039 Load->getMemOperand(), 0, EVT.getStoreSize()));
12043 // It is only safe to call this function if isINSERTPSMask is true for
12044 // this shufflevector mask.
12045 static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
12046 SelectionDAG &DAG) {
12047 // Generate an insertps instruction when inserting an f32 from memory onto a
12048 // v4f32 or when copying a member from one v4f32 to another.
12049 // We also use it for transferring i32 from one register to another,
12050 // since it simply copies the same bits.
12051 // If we're transferring an i32 from memory to a specific element in a
12052 // register, we output a generic DAG that will match the PINSRD
12054 MVT VT = SVOp->getSimpleValueType(0);
12055 MVT EVT = VT.getVectorElementType();
12056 SDValue V1 = SVOp->getOperand(0);
12057 SDValue V2 = SVOp->getOperand(1);
12058 auto Mask = SVOp->getMask();
12059 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
12060 "unsupported vector type for insertps/pinsrd");
12062 auto FromV1Predicate = [](const int &i) { return i < 4 && i > -1; };
12063 auto FromV2Predicate = [](const int &i) { return i >= 4; };
12064 int FromV1 = std::count_if(Mask.begin(), Mask.end(), FromV1Predicate);
12068 unsigned DestIndex;
12072 DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
12075 // If we have 1 element from each vector, we have to check if we're
12076 // changing V1's element's place. If so, we're done. Otherwise, we
12077 // should assume we're changing V2's element's place and behave
12079 int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
12080 assert(DestIndex <= INT32_MAX && "truncated destination index");
12081 if (FromV1 == FromV2 &&
12082 static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
12086 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12089 assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
12090 "More than one element from V1 and from V2, or no elements from one "
12091 "of the vectors. This case should not have returned true from "
12096 std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
12099 // Get an index into the source vector in the range [0,4) (the mask is
12100 // in the range [0,8) because it can address V1 and V2)
12101 unsigned SrcIndex = Mask[DestIndex] % 4;
12102 if (MayFoldLoad(From)) {
12103 // Trivial case, when From comes from a load and is only used by the
12104 // shuffle. Make it use insertps from the vector that we need from that
12107 NarrowVectorLoadToElement(cast<LoadSDNode>(From), SrcIndex, DAG);
12108 if (!NewLoad.getNode())
12111 if (EVT == MVT::f32) {
12112 // Create this as a scalar to vector to match the instruction pattern.
12113 SDValue LoadScalarToVector =
12114 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
12115 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
12116 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
12118 } else { // EVT == MVT::i32
12119 // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
12120 // instruction, to match the PINSRD instruction, which loads an i32 to a
12121 // certain vector element.
12122 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
12123 DAG.getConstant(DestIndex, MVT::i32));
12127 // Vector-element-to-vector
12128 SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
12129 return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
12132 // Reduce a vector shuffle to zext.
12133 static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
12134 SelectionDAG &DAG) {
12135 // PMOVZX is only available from SSE41.
12136 if (!Subtarget->hasSSE41())
12139 MVT VT = Op.getSimpleValueType();
12141 // Only AVX2 support 256-bit vector integer extending.
12142 if (!Subtarget->hasInt256() && VT.is256BitVector())
12145 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12147 SDValue V1 = Op.getOperand(0);
12148 SDValue V2 = Op.getOperand(1);
12149 unsigned NumElems = VT.getVectorNumElements();
12151 // Extending is an unary operation and the element type of the source vector
12152 // won't be equal to or larger than i64.
12153 if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
12154 VT.getVectorElementType() == MVT::i64)
12157 // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
12158 unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
12159 while ((1U << Shift) < NumElems) {
12160 if (SVOp->getMaskElt(1U << Shift) == 1)
12163 // The maximal ratio is 8, i.e. from i8 to i64.
12168 // Check the shuffle mask.
12169 unsigned Mask = (1U << Shift) - 1;
12170 for (unsigned i = 0; i != NumElems; ++i) {
12171 int EltIdx = SVOp->getMaskElt(i);
12172 if ((i & Mask) != 0 && EltIdx != -1)
12174 if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
12178 unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
12179 MVT NeVT = MVT::getIntegerVT(NBits);
12180 MVT NVT = MVT::getVectorVT(NeVT, NumElems >> Shift);
12182 if (!DAG.getTargetLoweringInfo().isTypeLegal(NVT))
12185 return DAG.getNode(ISD::BITCAST, DL, VT,
12186 DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
12189 static SDValue NormalizeVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
12190 SelectionDAG &DAG) {
12191 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12192 MVT VT = Op.getSimpleValueType();
12194 SDValue V1 = Op.getOperand(0);
12195 SDValue V2 = Op.getOperand(1);
12197 if (isZeroShuffle(SVOp))
12198 return getZeroVector(VT, Subtarget, DAG, dl);
12200 // Handle splat operations
12201 if (SVOp->isSplat()) {
12202 // Use vbroadcast whenever the splat comes from a foldable load
12203 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
12204 if (Broadcast.getNode())
12208 // Check integer expanding shuffles.
12209 SDValue NewOp = LowerVectorIntExtend(Op, Subtarget, DAG);
12210 if (NewOp.getNode())
12213 // If the shuffle can be profitably rewritten as a narrower shuffle, then
12215 if (VT == MVT::v8i16 || VT == MVT::v16i8 || VT == MVT::v16i16 ||
12216 VT == MVT::v32i8) {
12217 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12218 if (NewOp.getNode())
12219 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
12220 } else if (VT.is128BitVector() && Subtarget->hasSSE2()) {
12221 // FIXME: Figure out a cleaner way to do this.
12222 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
12223 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12224 if (NewOp.getNode()) {
12225 MVT NewVT = NewOp.getSimpleValueType();
12226 if (isCommutedMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(),
12227 NewVT, true, false))
12228 return getVZextMovL(VT, NewVT, NewOp.getOperand(0), DAG, Subtarget,
12231 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
12232 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG);
12233 if (NewOp.getNode()) {
12234 MVT NewVT = NewOp.getSimpleValueType();
12235 if (isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)->getMask(), NewVT))
12236 return getVZextMovL(VT, NewVT, NewOp.getOperand(1), DAG, Subtarget,
12245 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
12246 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
12247 SDValue V1 = Op.getOperand(0);
12248 SDValue V2 = Op.getOperand(1);
12249 MVT VT = Op.getSimpleValueType();
12251 unsigned NumElems = VT.getVectorNumElements();
12252 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
12253 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
12254 bool V1IsSplat = false;
12255 bool V2IsSplat = false;
12256 bool HasSSE2 = Subtarget->hasSSE2();
12257 bool HasFp256 = Subtarget->hasFp256();
12258 bool HasInt256 = Subtarget->hasInt256();
12259 MachineFunction &MF = DAG.getMachineFunction();
12260 bool OptForSize = MF.getFunction()->getAttributes().
12261 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
12263 // Check if we should use the experimental vector shuffle lowering. If so,
12264 // delegate completely to that code path.
12265 if (ExperimentalVectorShuffleLowering)
12266 return lowerVectorShuffle(Op, Subtarget, DAG);
12268 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
12270 if (V1IsUndef && V2IsUndef)
12271 return DAG.getUNDEF(VT);
12273 // When we create a shuffle node we put the UNDEF node to second operand,
12274 // but in some cases the first operand may be transformed to UNDEF.
12275 // In this case we should just commute the node.
12277 return DAG.getCommutedVectorShuffle(*SVOp);
12279 // Vector shuffle lowering takes 3 steps:
12281 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
12282 // narrowing and commutation of operands should be handled.
12283 // 2) Matching of shuffles with known shuffle masks to x86 target specific
12285 // 3) Rewriting of unmatched masks into new generic shuffle operations,
12286 // so the shuffle can be broken into other shuffles and the legalizer can
12287 // try the lowering again.
12289 // The general idea is that no vector_shuffle operation should be left to
12290 // be matched during isel, all of them must be converted to a target specific
12293 // Normalize the input vectors. Here splats, zeroed vectors, profitable
12294 // narrowing and commutation of operands should be handled. The actual code
12295 // doesn't include all of those, work in progress...
12296 SDValue NewOp = NormalizeVectorShuffle(Op, Subtarget, DAG);
12297 if (NewOp.getNode())
12300 SmallVector<int, 8> M(SVOp->getMask().begin(), SVOp->getMask().end());
12302 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
12303 // unpckh_undef). Only use pshufd if speed is more important than size.
12304 if (OptForSize && isUNPCKL_v_undef_Mask(M, VT, HasInt256))
12305 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12306 if (OptForSize && isUNPCKH_v_undef_Mask(M, VT, HasInt256))
12307 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12309 if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
12310 V2IsUndef && MayFoldVectorLoad(V1))
12311 return getMOVDDup(Op, dl, V1, DAG);
12313 if (isMOVHLPS_v_undef_Mask(M, VT))
12314 return getMOVHighToLow(Op, dl, DAG);
12316 // Use to match splats
12317 if (HasSSE2 && isUNPCKHMask(M, VT, HasInt256) && V2IsUndef &&
12318 (VT == MVT::v2f64 || VT == MVT::v2i64))
12319 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12321 if (isPSHUFDMask(M, VT)) {
12322 // The actual implementation will match the mask in the if above and then
12323 // during isel it can match several different instructions, not only pshufd
12324 // as its name says, sad but true, emulate the behavior for now...
12325 if (isMOVDDUPMask(M, VT) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
12326 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
12328 unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
12330 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
12331 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
12333 if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
12334 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1, TargetMask,
12337 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
12341 if (isPALIGNRMask(M, VT, Subtarget))
12342 return getTargetShuffleNode(X86ISD::PALIGNR, dl, VT, V1, V2,
12343 getShufflePALIGNRImmediate(SVOp),
12346 if (isVALIGNMask(M, VT, Subtarget))
12347 return getTargetShuffleNode(X86ISD::VALIGN, dl, VT, V1, V2,
12348 getShuffleVALIGNImmediate(SVOp),
12351 // Check if this can be converted into a logical shift.
12352 bool isLeft = false;
12353 unsigned ShAmt = 0;
12355 bool isShift = HasSSE2 && isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
12356 if (isShift && ShVal.hasOneUse()) {
12357 // If the shifted value has multiple uses, it may be cheaper to use
12358 // v_set0 + movlhps or movhlps, etc.
12359 MVT EltVT = VT.getVectorElementType();
12360 ShAmt *= EltVT.getSizeInBits();
12361 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12364 if (isMOVLMask(M, VT)) {
12365 if (ISD::isBuildVectorAllZeros(V1.getNode()))
12366 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
12367 if (!isMOVLPMask(M, VT)) {
12368 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
12369 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
12371 if (VT == MVT::v4i32 || VT == MVT::v4f32)
12372 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
12376 // FIXME: fold these into legal mask.
12377 if (isMOVLHPSMask(M, VT) && !isUNPCKLMask(M, VT, HasInt256))
12378 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
12380 if (isMOVHLPSMask(M, VT))
12381 return getMOVHighToLow(Op, dl, DAG);
12383 if (V2IsUndef && isMOVSHDUPMask(M, VT, Subtarget))
12384 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
12386 if (V2IsUndef && isMOVSLDUPMask(M, VT, Subtarget))
12387 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
12389 if (isMOVLPMask(M, VT))
12390 return getMOVLP(Op, dl, DAG, HasSSE2);
12392 if (ShouldXformToMOVHLPS(M, VT) ||
12393 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
12394 return DAG.getCommutedVectorShuffle(*SVOp);
12397 // No better options. Use a vshldq / vsrldq.
12398 MVT EltVT = VT.getVectorElementType();
12399 ShAmt *= EltVT.getSizeInBits();
12400 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
12403 bool Commuted = false;
12404 // FIXME: This should also accept a bitcast of a splat? Be careful, not
12405 // 1,1,1,1 -> v8i16 though.
12406 BitVector UndefElements;
12407 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V1.getNode()))
12408 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12410 if (auto *BVOp = dyn_cast<BuildVectorSDNode>(V2.getNode()))
12411 if (BVOp->getConstantSplatNode(&UndefElements) && UndefElements.none())
12414 // Canonicalize the splat or undef, if present, to be on the RHS.
12415 if (!V2IsUndef && V1IsSplat && !V2IsSplat) {
12416 CommuteVectorShuffleMask(M, NumElems);
12418 std::swap(V1IsSplat, V2IsSplat);
12422 if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
12423 // Shuffling low element of v1 into undef, just return v1.
12426 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
12427 // the instruction selector will not match, so get a canonical MOVL with
12428 // swapped operands to undo the commute.
12429 return getMOVL(DAG, dl, VT, V2, V1);
12432 if (isUNPCKLMask(M, VT, HasInt256))
12433 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12435 if (isUNPCKHMask(M, VT, HasInt256))
12436 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12439 // Normalize mask so all entries that point to V2 points to its first
12440 // element then try to match unpck{h|l} again. If match, return a
12441 // new vector_shuffle with the corrected mask.p
12442 SmallVector<int, 8> NewMask(M.begin(), M.end());
12443 NormalizeMask(NewMask, NumElems);
12444 if (isUNPCKLMask(NewMask, VT, HasInt256, true))
12445 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12446 if (isUNPCKHMask(NewMask, VT, HasInt256, true))
12447 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12451 // Commute is back and try unpck* again.
12452 // FIXME: this seems wrong.
12453 CommuteVectorShuffleMask(M, NumElems);
12455 std::swap(V1IsSplat, V2IsSplat);
12457 if (isUNPCKLMask(M, VT, HasInt256))
12458 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
12460 if (isUNPCKHMask(M, VT, HasInt256))
12461 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
12464 // Normalize the node to match x86 shuffle ops if needed
12465 if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
12466 return DAG.getCommutedVectorShuffle(*SVOp);
12468 // The checks below are all present in isShuffleMaskLegal, but they are
12469 // inlined here right now to enable us to directly emit target specific
12470 // nodes, and remove one by one until they don't return Op anymore.
12472 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
12473 SVOp->getSplatIndex() == 0 && V2IsUndef) {
12474 if (VT == MVT::v2f64 || VT == MVT::v2i64)
12475 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12478 if (isPSHUFHWMask(M, VT, HasInt256))
12479 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
12480 getShufflePSHUFHWImmediate(SVOp),
12483 if (isPSHUFLWMask(M, VT, HasInt256))
12484 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
12485 getShufflePSHUFLWImmediate(SVOp),
12488 unsigned MaskValue;
12489 if (isBlendMask(M, VT, Subtarget->hasSSE41(), Subtarget->hasInt256(),
12491 return LowerVECTOR_SHUFFLEtoBlend(SVOp, MaskValue, Subtarget, DAG);
12493 if (isSHUFPMask(M, VT))
12494 return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V2,
12495 getShuffleSHUFImmediate(SVOp), DAG);
12497 if (isUNPCKL_v_undef_Mask(M, VT, HasInt256))
12498 return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
12499 if (isUNPCKH_v_undef_Mask(M, VT, HasInt256))
12500 return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
12502 //===--------------------------------------------------------------------===//
12503 // Generate target specific nodes for 128 or 256-bit shuffles only
12504 // supported in the AVX instruction set.
12507 // Handle VMOVDDUPY permutations
12508 if (V2IsUndef && isMOVDDUPYMask(M, VT, HasFp256))
12509 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
12511 // Handle VPERMILPS/D* permutations
12512 if (isVPERMILPMask(M, VT)) {
12513 if ((HasInt256 && VT == MVT::v8i32) || VT == MVT::v16i32)
12514 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1,
12515 getShuffleSHUFImmediate(SVOp), DAG);
12516 return getTargetShuffleNode(X86ISD::VPERMILPI, dl, VT, V1,
12517 getShuffleSHUFImmediate(SVOp), DAG);
12521 if (VT.is512BitVector() && isINSERT64x4Mask(M, VT, &Idx))
12522 return Insert256BitVector(V1, Extract256BitVector(V2, 0, DAG, dl),
12523 Idx*(NumElems/2), DAG, dl);
12525 // Handle VPERM2F128/VPERM2I128 permutations
12526 if (isVPERM2X128Mask(M, VT, HasFp256))
12527 return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
12528 V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
12530 if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
12531 return getINSERTPS(SVOp, dl, DAG);
12534 if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
12535 return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
12537 if ((V2IsUndef && HasInt256 && VT.is256BitVector() && NumElems == 8) ||
12538 VT.is512BitVector()) {
12539 MVT MaskEltVT = MVT::getIntegerVT(VT.getVectorElementType().getSizeInBits());
12540 MVT MaskVectorVT = MVT::getVectorVT(MaskEltVT, NumElems);
12541 SmallVector<SDValue, 16> permclMask;
12542 for (unsigned i = 0; i != NumElems; ++i) {
12543 permclMask.push_back(DAG.getConstant((M[i]>=0) ? M[i] : 0, MaskEltVT));
12546 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, MaskVectorVT, permclMask);
12548 // Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
12549 return DAG.getNode(X86ISD::VPERMV, dl, VT,
12550 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
12551 return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
12552 DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
12555 //===--------------------------------------------------------------------===//
12556 // Since no target specific shuffle was selected for this generic one,
12557 // lower it into other known shuffles. FIXME: this isn't true yet, but
12558 // this is the plan.
12561 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
12562 if (VT == MVT::v8i16) {
12563 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, Subtarget, DAG);
12564 if (NewOp.getNode())
12568 if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
12569 SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
12570 if (NewOp.getNode())
12574 if (VT == MVT::v16i8) {
12575 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
12576 if (NewOp.getNode())
12580 if (VT == MVT::v32i8) {
12581 SDValue NewOp = LowerVECTOR_SHUFFLEv32i8(SVOp, Subtarget, DAG);
12582 if (NewOp.getNode())
12586 // Handle all 128-bit wide vectors with 4 elements, and match them with
12587 // several different shuffle types.
12588 if (NumElems == 4 && VT.is128BitVector())
12589 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
12591 // Handle general 256-bit shuffles
12592 if (VT.is256BitVector())
12593 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
12598 // This function assumes its argument is a BUILD_VECTOR of constants or
12599 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
12601 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
12602 unsigned &MaskValue) {
12604 unsigned NumElems = BuildVector->getNumOperands();
12605 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
12606 unsigned NumLanes = (NumElems - 1) / 8 + 1;
12607 unsigned NumElemsInLane = NumElems / NumLanes;
12609 // Blend for v16i16 should be symetric for the both lanes.
12610 for (unsigned i = 0; i < NumElemsInLane; ++i) {
12611 SDValue EltCond = BuildVector->getOperand(i);
12612 SDValue SndLaneEltCond =
12613 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
12615 int Lane1Cond = -1, Lane2Cond = -1;
12616 if (isa<ConstantSDNode>(EltCond))
12617 Lane1Cond = !isZero(EltCond);
12618 if (isa<ConstantSDNode>(SndLaneEltCond))
12619 Lane2Cond = !isZero(SndLaneEltCond);
12621 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
12622 // Lane1Cond != 0, means we want the first argument.
12623 // Lane1Cond == 0, means we want the second argument.
12624 // The encoding of this argument is 0 for the first argument, 1
12625 // for the second. Therefore, invert the condition.
12626 MaskValue |= !Lane1Cond << i;
12627 else if (Lane1Cond < 0)
12628 MaskValue |= !Lane2Cond << i;
12635 /// \brief Try to lower a VSELECT instruction to an immediate-controlled blend
12637 static SDValue lowerVSELECTtoBLENDI(SDValue Op, const X86Subtarget *Subtarget,
12638 SelectionDAG &DAG) {
12639 SDValue Cond = Op.getOperand(0);
12640 SDValue LHS = Op.getOperand(1);
12641 SDValue RHS = Op.getOperand(2);
12643 MVT VT = Op.getSimpleValueType();
12644 MVT EltVT = VT.getVectorElementType();
12645 unsigned NumElems = VT.getVectorNumElements();
12647 // There is no blend with immediate in AVX-512.
12648 if (VT.is512BitVector())
12651 if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
12653 if (!Subtarget->hasInt256() && VT == MVT::v16i16)
12656 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
12659 // Check the mask for BLEND and build the value.
12660 unsigned MaskValue = 0;
12661 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
12664 // Convert i32 vectors to floating point if it is not AVX2.
12665 // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
12667 if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
12668 BlendVT = MVT::getVectorVT(MVT::getFloatingPointVT(EltVT.getSizeInBits()),
12670 LHS = DAG.getNode(ISD::BITCAST, dl, VT, LHS);
12671 RHS = DAG.getNode(ISD::BITCAST, dl, VT, RHS);
12674 SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, LHS, RHS,
12675 DAG.getConstant(MaskValue, MVT::i32));
12676 return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
12679 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
12680 // A vselect where all conditions and data are constants can be optimized into
12681 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
12682 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
12683 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
12684 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
12687 SDValue BlendOp = lowerVSELECTtoBLENDI(Op, Subtarget, DAG);
12688 if (BlendOp.getNode())
12691 // Some types for vselect were previously set to Expand, not Legal or
12692 // Custom. Return an empty SDValue so we fall-through to Expand, after
12693 // the Custom lowering phase.
12694 MVT VT = Op.getSimpleValueType();
12695 switch (VT.SimpleTy) {
12700 if (Subtarget->hasBWI() && Subtarget->hasVLX())
12705 // We couldn't create a "Blend with immediate" node.
12706 // This node should still be legal, but we'll have to emit a blendv*
12711 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
12712 MVT VT = Op.getSimpleValueType();
12715 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
12718 if (VT.getSizeInBits() == 8) {
12719 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
12720 Op.getOperand(0), Op.getOperand(1));
12721 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
12722 DAG.getValueType(VT));
12723 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
12726 if (VT.getSizeInBits() == 16) {
12727 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12728 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
12730 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
12731 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
12732 DAG.getNode(ISD::BITCAST, dl,
12735 Op.getOperand(1)));
12736 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
12737 Op.getOperand(0), Op.getOperand(1));
12738 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
12739 DAG.getValueType(VT));
12740 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
12743 if (VT == MVT::f32) {
12744 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
12745 // the result back to FR32 register. It's only worth matching if the
12746 // result has a single use which is a store or a bitcast to i32. And in
12747 // the case of a store, it's not worth it if the index is a constant 0,
12748 // because a MOVSSmr can be used instead, which is smaller and faster.
12749 if (!Op.hasOneUse())
12751 SDNode *User = *Op.getNode()->use_begin();
12752 if ((User->getOpcode() != ISD::STORE ||
12753 (isa<ConstantSDNode>(Op.getOperand(1)) &&
12754 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
12755 (User->getOpcode() != ISD::BITCAST ||
12756 User->getValueType(0) != MVT::i32))
12758 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
12759 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
12762 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
12765 if (VT == MVT::i32 || VT == MVT::i64) {
12766 // ExtractPS/pextrq works with constant index.
12767 if (isa<ConstantSDNode>(Op.getOperand(1)))
12773 /// Extract one bit from mask vector, like v16i1 or v8i1.
12774 /// AVX-512 feature.
12776 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
12777 SDValue Vec = Op.getOperand(0);
12779 MVT VecVT = Vec.getSimpleValueType();
12780 SDValue Idx = Op.getOperand(1);
12781 MVT EltVT = Op.getSimpleValueType();
12783 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
12785 // variable index can't be handled in mask registers,
12786 // extend vector to VR512
12787 if (!isa<ConstantSDNode>(Idx)) {
12788 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
12789 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
12790 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
12791 ExtVT.getVectorElementType(), Ext, Idx);
12792 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
12795 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12796 const TargetRegisterClass* rc = getRegClassFor(VecVT);
12797 unsigned MaxSift = rc->getSize()*8 - 1;
12798 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
12799 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
12800 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
12801 DAG.getConstant(MaxSift, MVT::i8));
12802 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
12803 DAG.getIntPtrConstant(0));
12807 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
12808 SelectionDAG &DAG) const {
12810 SDValue Vec = Op.getOperand(0);
12811 MVT VecVT = Vec.getSimpleValueType();
12812 SDValue Idx = Op.getOperand(1);
12814 if (Op.getSimpleValueType() == MVT::i1)
12815 return ExtractBitFromMaskVector(Op, DAG);
12817 if (!isa<ConstantSDNode>(Idx)) {
12818 if (VecVT.is512BitVector() ||
12819 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
12820 VecVT.getVectorElementType().getSizeInBits() == 32)) {
12823 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
12824 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
12825 MaskEltVT.getSizeInBits());
12827 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
12828 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
12829 getZeroVector(MaskVT, Subtarget, DAG, dl),
12830 Idx, DAG.getConstant(0, getPointerTy()));
12831 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
12832 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
12833 Perm, DAG.getConstant(0, getPointerTy()));
12838 // If this is a 256-bit vector result, first extract the 128-bit vector and
12839 // then extract the element from the 128-bit vector.
12840 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
12842 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12843 // Get the 128-bit vector.
12844 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
12845 MVT EltVT = VecVT.getVectorElementType();
12847 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
12849 //if (IdxVal >= NumElems/2)
12850 // IdxVal -= NumElems/2;
12851 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
12852 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
12853 DAG.getConstant(IdxVal, MVT::i32));
12856 assert(VecVT.is128BitVector() && "Unexpected vector length");
12858 if (Subtarget->hasSSE41()) {
12859 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
12864 MVT VT = Op.getSimpleValueType();
12865 // TODO: handle v16i8.
12866 if (VT.getSizeInBits() == 16) {
12867 SDValue Vec = Op.getOperand(0);
12868 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12870 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
12871 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
12872 DAG.getNode(ISD::BITCAST, dl,
12874 Op.getOperand(1)));
12875 // Transform it so it match pextrw which produces a 32-bit result.
12876 MVT EltVT = MVT::i32;
12877 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
12878 Op.getOperand(0), Op.getOperand(1));
12879 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
12880 DAG.getValueType(VT));
12881 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
12884 if (VT.getSizeInBits() == 32) {
12885 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12889 // SHUFPS the element to the lowest double word, then movss.
12890 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
12891 MVT VVT = Op.getOperand(0).getSimpleValueType();
12892 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
12893 DAG.getUNDEF(VVT), Mask);
12894 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
12895 DAG.getIntPtrConstant(0));
12898 if (VT.getSizeInBits() == 64) {
12899 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
12900 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
12901 // to match extract_elt for f64.
12902 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
12906 // UNPCKHPD the element to the lowest double word, then movsd.
12907 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
12908 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
12909 int Mask[2] = { 1, -1 };
12910 MVT VVT = Op.getOperand(0).getSimpleValueType();
12911 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
12912 DAG.getUNDEF(VVT), Mask);
12913 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
12914 DAG.getIntPtrConstant(0));
12920 /// Insert one bit to mask vector, like v16i1 or v8i1.
12921 /// AVX-512 feature.
12923 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
12925 SDValue Vec = Op.getOperand(0);
12926 SDValue Elt = Op.getOperand(1);
12927 SDValue Idx = Op.getOperand(2);
12928 MVT VecVT = Vec.getSimpleValueType();
12930 if (!isa<ConstantSDNode>(Idx)) {
12931 // Non constant index. Extend source and destination,
12932 // insert element and then truncate the result.
12933 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
12934 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
12935 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
12936 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
12937 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
12938 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
12941 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
12942 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
12943 if (Vec.getOpcode() == ISD::UNDEF)
12944 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
12945 DAG.getConstant(IdxVal, MVT::i8));
12946 const TargetRegisterClass* rc = getRegClassFor(VecVT);
12947 unsigned MaxSift = rc->getSize()*8 - 1;
12948 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
12949 DAG.getConstant(MaxSift, MVT::i8));
12950 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
12951 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
12952 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
12955 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
12956 SelectionDAG &DAG) const {
12957 MVT VT = Op.getSimpleValueType();
12958 MVT EltVT = VT.getVectorElementType();
12960 if (EltVT == MVT::i1)
12961 return InsertBitToMaskVector(Op, DAG);
12964 SDValue N0 = Op.getOperand(0);
12965 SDValue N1 = Op.getOperand(1);
12966 SDValue N2 = Op.getOperand(2);
12967 if (!isa<ConstantSDNode>(N2))
12969 auto *N2C = cast<ConstantSDNode>(N2);
12970 unsigned IdxVal = N2C->getZExtValue();
12972 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
12973 // into that, and then insert the subvector back into the result.
12974 if (VT.is256BitVector() || VT.is512BitVector()) {
12975 // Get the desired 128-bit vector half.
12976 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
12978 // Insert the element into the desired half.
12979 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
12980 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
12982 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
12983 DAG.getConstant(IdxIn128, MVT::i32));
12985 // Insert the changed part back to the 256-bit vector
12986 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
12988 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
12990 if (Subtarget->hasSSE41()) {
12991 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
12993 if (VT == MVT::v8i16) {
12994 Opc = X86ISD::PINSRW;
12996 assert(VT == MVT::v16i8);
12997 Opc = X86ISD::PINSRB;
13000 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
13002 if (N1.getValueType() != MVT::i32)
13003 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13004 if (N2.getValueType() != MVT::i32)
13005 N2 = DAG.getIntPtrConstant(IdxVal);
13006 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
13009 if (EltVT == MVT::f32) {
13010 // Bits [7:6] of the constant are the source select. This will always be
13011 // zero here. The DAG Combiner may combine an extract_elt index into
13013 // bits. For example (insert (extract, 3), 2) could be matched by
13015 // the '3' into bits [7:6] of X86ISD::INSERTPS.
13016 // Bits [5:4] of the constant are the destination select. This is the
13017 // value of the incoming immediate.
13018 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
13019 // combine either bitwise AND or insert of float 0.0 to set these bits.
13020 N2 = DAG.getIntPtrConstant(IdxVal << 4);
13021 // Create this as a scalar to vector..
13022 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
13023 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
13026 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
13027 // PINSR* works with constant index.
13032 if (EltVT == MVT::i8)
13035 if (EltVT.getSizeInBits() == 16) {
13036 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
13037 // as its second argument.
13038 if (N1.getValueType() != MVT::i32)
13039 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
13040 if (N2.getValueType() != MVT::i32)
13041 N2 = DAG.getIntPtrConstant(IdxVal);
13042 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
13047 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
13049 MVT OpVT = Op.getSimpleValueType();
13051 // If this is a 256-bit vector result, first insert into a 128-bit
13052 // vector and then insert into the 256-bit vector.
13053 if (!OpVT.is128BitVector()) {
13054 // Insert into a 128-bit vector.
13055 unsigned SizeFactor = OpVT.getSizeInBits()/128;
13056 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
13057 OpVT.getVectorNumElements() / SizeFactor);
13059 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
13061 // Insert the 128-bit vector.
13062 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
13065 if (OpVT == MVT::v1i64 &&
13066 Op.getOperand(0).getValueType() == MVT::i64)
13067 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
13069 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
13070 assert(OpVT.is128BitVector() && "Expected an SSE type!");
13071 return DAG.getNode(ISD::BITCAST, dl, OpVT,
13072 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
13075 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
13076 // a simple subregister reference or explicit instructions to grab
13077 // upper bits of a vector.
13078 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13079 SelectionDAG &DAG) {
13081 SDValue In = Op.getOperand(0);
13082 SDValue Idx = Op.getOperand(1);
13083 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13084 MVT ResVT = Op.getSimpleValueType();
13085 MVT InVT = In.getSimpleValueType();
13087 if (Subtarget->hasFp256()) {
13088 if (ResVT.is128BitVector() &&
13089 (InVT.is256BitVector() || InVT.is512BitVector()) &&
13090 isa<ConstantSDNode>(Idx)) {
13091 return Extract128BitVector(In, IdxVal, DAG, dl);
13093 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
13094 isa<ConstantSDNode>(Idx)) {
13095 return Extract256BitVector(In, IdxVal, DAG, dl);
13101 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
13102 // simple superregister reference or explicit instructions to insert
13103 // the upper bits of a vector.
13104 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
13105 SelectionDAG &DAG) {
13106 if (Subtarget->hasFp256()) {
13107 SDLoc dl(Op.getNode());
13108 SDValue Vec = Op.getNode()->getOperand(0);
13109 SDValue SubVec = Op.getNode()->getOperand(1);
13110 SDValue Idx = Op.getNode()->getOperand(2);
13112 if ((Op.getNode()->getSimpleValueType(0).is256BitVector() ||
13113 Op.getNode()->getSimpleValueType(0).is512BitVector()) &&
13114 SubVec.getNode()->getSimpleValueType(0).is128BitVector() &&
13115 isa<ConstantSDNode>(Idx)) {
13116 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13117 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
13120 if (Op.getNode()->getSimpleValueType(0).is512BitVector() &&
13121 SubVec.getNode()->getSimpleValueType(0).is256BitVector() &&
13122 isa<ConstantSDNode>(Idx)) {
13123 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
13124 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
13130 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
13131 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
13132 // one of the above mentioned nodes. It has to be wrapped because otherwise
13133 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
13134 // be used to form addressing mode. These wrapped nodes will be selected
13137 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
13138 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
13140 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13141 // global base reg.
13142 unsigned char OpFlag = 0;
13143 unsigned WrapperKind = X86ISD::Wrapper;
13144 CodeModel::Model M = DAG.getTarget().getCodeModel();
13146 if (Subtarget->isPICStyleRIPRel() &&
13147 (M == CodeModel::Small || M == CodeModel::Kernel))
13148 WrapperKind = X86ISD::WrapperRIP;
13149 else if (Subtarget->isPICStyleGOT())
13150 OpFlag = X86II::MO_GOTOFF;
13151 else if (Subtarget->isPICStyleStubPIC())
13152 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13154 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
13155 CP->getAlignment(),
13156 CP->getOffset(), OpFlag);
13158 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13159 // With PIC, the address is actually $g + Offset.
13161 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13162 DAG.getNode(X86ISD::GlobalBaseReg,
13163 SDLoc(), getPointerTy()),
13170 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
13171 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
13173 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13174 // global base reg.
13175 unsigned char OpFlag = 0;
13176 unsigned WrapperKind = X86ISD::Wrapper;
13177 CodeModel::Model M = DAG.getTarget().getCodeModel();
13179 if (Subtarget->isPICStyleRIPRel() &&
13180 (M == CodeModel::Small || M == CodeModel::Kernel))
13181 WrapperKind = X86ISD::WrapperRIP;
13182 else if (Subtarget->isPICStyleGOT())
13183 OpFlag = X86II::MO_GOTOFF;
13184 else if (Subtarget->isPICStyleStubPIC())
13185 OpFlag = X86II::MO_PIC_BASE_OFFSET;
13187 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
13190 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13192 // With PIC, the address is actually $g + Offset.
13194 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13195 DAG.getNode(X86ISD::GlobalBaseReg,
13196 SDLoc(), getPointerTy()),
13203 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
13204 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
13206 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13207 // global base reg.
13208 unsigned char OpFlag = 0;
13209 unsigned WrapperKind = X86ISD::Wrapper;
13210 CodeModel::Model M = DAG.getTarget().getCodeModel();
13212 if (Subtarget->isPICStyleRIPRel() &&
13213 (M == CodeModel::Small || M == CodeModel::Kernel)) {
13214 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
13215 OpFlag = X86II::MO_GOTPCREL;
13216 WrapperKind = X86ISD::WrapperRIP;
13217 } else if (Subtarget->isPICStyleGOT()) {
13218 OpFlag = X86II::MO_GOT;
13219 } else if (Subtarget->isPICStyleStubPIC()) {
13220 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
13221 } else if (Subtarget->isPICStyleStubNoDynamic()) {
13222 OpFlag = X86II::MO_DARWIN_NONLAZY;
13225 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
13228 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13230 // With PIC, the address is actually $g + Offset.
13231 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
13232 !Subtarget->is64Bit()) {
13233 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13234 DAG.getNode(X86ISD::GlobalBaseReg,
13235 SDLoc(), getPointerTy()),
13239 // For symbols that require a load from a stub to get the address, emit the
13241 if (isGlobalStubReference(OpFlag))
13242 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
13243 MachinePointerInfo::getGOT(), false, false, false, 0);
13249 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
13250 // Create the TargetBlockAddressAddress node.
13251 unsigned char OpFlags =
13252 Subtarget->ClassifyBlockAddressReference();
13253 CodeModel::Model M = DAG.getTarget().getCodeModel();
13254 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
13255 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
13257 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
13260 if (Subtarget->isPICStyleRIPRel() &&
13261 (M == CodeModel::Small || M == CodeModel::Kernel))
13262 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13264 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13266 // With PIC, the address is actually $g + Offset.
13267 if (isGlobalRelativeToPICBase(OpFlags)) {
13268 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13269 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13277 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
13278 int64_t Offset, SelectionDAG &DAG) const {
13279 // Create the TargetGlobalAddress node, folding in the constant
13280 // offset if it is legal.
13281 unsigned char OpFlags =
13282 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
13283 CodeModel::Model M = DAG.getTarget().getCodeModel();
13285 if (OpFlags == X86II::MO_NO_FLAG &&
13286 X86::isOffsetSuitableForCodeModel(Offset, M)) {
13287 // A direct static reference to a global.
13288 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
13291 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
13294 if (Subtarget->isPICStyleRIPRel() &&
13295 (M == CodeModel::Small || M == CodeModel::Kernel))
13296 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
13298 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
13300 // With PIC, the address is actually $g + Offset.
13301 if (isGlobalRelativeToPICBase(OpFlags)) {
13302 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
13303 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
13307 // For globals that require a load from a stub to get the address, emit the
13309 if (isGlobalStubReference(OpFlags))
13310 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
13311 MachinePointerInfo::getGOT(), false, false, false, 0);
13313 // If there was a non-zero offset that we didn't fold, create an explicit
13314 // addition for it.
13316 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
13317 DAG.getConstant(Offset, getPointerTy()));
13323 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
13324 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
13325 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
13326 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
13330 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
13331 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
13332 unsigned char OperandFlags, bool LocalDynamic = false) {
13333 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13334 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13336 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13337 GA->getValueType(0),
13341 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
13345 SDValue Ops[] = { Chain, TGA, *InFlag };
13346 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13348 SDValue Ops[] = { Chain, TGA };
13349 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
13352 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
13353 MFI->setAdjustsStack(true);
13354 MFI->setHasCalls(true);
13356 SDValue Flag = Chain.getValue(1);
13357 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
13360 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
13362 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13365 SDLoc dl(GA); // ? function entry point might be better
13366 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13367 DAG.getNode(X86ISD::GlobalBaseReg,
13368 SDLoc(), PtrVT), InFlag);
13369 InFlag = Chain.getValue(1);
13371 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
13374 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
13376 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13378 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
13379 X86::RAX, X86II::MO_TLSGD);
13382 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
13388 // Get the start address of the TLS block for this module.
13389 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
13390 .getInfo<X86MachineFunctionInfo>();
13391 MFI->incNumLocalDynamicTLSAccesses();
13395 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
13396 X86II::MO_TLSLD, /*LocalDynamic=*/true);
13399 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
13400 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
13401 InFlag = Chain.getValue(1);
13402 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
13403 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
13406 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
13410 unsigned char OperandFlags = X86II::MO_DTPOFF;
13411 unsigned WrapperKind = X86ISD::Wrapper;
13412 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13413 GA->getValueType(0),
13414 GA->getOffset(), OperandFlags);
13415 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
13417 // Add x@dtpoff with the base.
13418 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
13421 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
13422 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
13423 const EVT PtrVT, TLSModel::Model model,
13424 bool is64Bit, bool isPIC) {
13427 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
13428 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
13429 is64Bit ? 257 : 256));
13431 SDValue ThreadPointer =
13432 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
13433 MachinePointerInfo(Ptr), false, false, false, 0);
13435 unsigned char OperandFlags = 0;
13436 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
13438 unsigned WrapperKind = X86ISD::Wrapper;
13439 if (model == TLSModel::LocalExec) {
13440 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
13441 } else if (model == TLSModel::InitialExec) {
13443 OperandFlags = X86II::MO_GOTTPOFF;
13444 WrapperKind = X86ISD::WrapperRIP;
13446 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
13449 llvm_unreachable("Unexpected model");
13452 // emit "addl x@ntpoff,%eax" (local exec)
13453 // or "addl x@indntpoff,%eax" (initial exec)
13454 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
13456 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
13457 GA->getOffset(), OperandFlags);
13458 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
13460 if (model == TLSModel::InitialExec) {
13461 if (isPIC && !is64Bit) {
13462 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
13463 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
13467 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
13468 MachinePointerInfo::getGOT(), false, false, false, 0);
13471 // The address of the thread local variable is the add of the thread
13472 // pointer with the offset of the variable.
13473 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
13477 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
13479 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
13480 const GlobalValue *GV = GA->getGlobal();
13482 if (Subtarget->isTargetELF()) {
13483 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
13486 case TLSModel::GeneralDynamic:
13487 if (Subtarget->is64Bit())
13488 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
13489 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
13490 case TLSModel::LocalDynamic:
13491 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
13492 Subtarget->is64Bit());
13493 case TLSModel::InitialExec:
13494 case TLSModel::LocalExec:
13495 return LowerToTLSExecModel(
13496 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
13497 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
13499 llvm_unreachable("Unknown TLS model.");
13502 if (Subtarget->isTargetDarwin()) {
13503 // Darwin only has one model of TLS. Lower to that.
13504 unsigned char OpFlag = 0;
13505 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
13506 X86ISD::WrapperRIP : X86ISD::Wrapper;
13508 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
13509 // global base reg.
13510 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
13511 !Subtarget->is64Bit();
13513 OpFlag = X86II::MO_TLVP_PIC_BASE;
13515 OpFlag = X86II::MO_TLVP;
13517 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
13518 GA->getValueType(0),
13519 GA->getOffset(), OpFlag);
13520 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
13522 // With PIC32, the address is actually $g + Offset.
13524 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
13525 DAG.getNode(X86ISD::GlobalBaseReg,
13526 SDLoc(), getPointerTy()),
13529 // Lowering the machine isd will make sure everything is in the right
13531 SDValue Chain = DAG.getEntryNode();
13532 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
13533 SDValue Args[] = { Chain, Offset };
13534 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
13536 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
13537 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13538 MFI->setAdjustsStack(true);
13540 // And our return value (tls address) is in the standard call return value
13542 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
13543 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
13544 Chain.getValue(1));
13547 if (Subtarget->isTargetKnownWindowsMSVC() ||
13548 Subtarget->isTargetWindowsGNU()) {
13549 // Just use the implicit TLS architecture
13550 // Need to generate someting similar to:
13551 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
13553 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
13554 // mov rcx, qword [rdx+rcx*8]
13555 // mov eax, .tls$:tlsvar
13556 // [rax+rcx] contains the address
13557 // Windows 64bit: gs:0x58
13558 // Windows 32bit: fs:__tls_array
13561 SDValue Chain = DAG.getEntryNode();
13563 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
13564 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
13565 // use its literal value of 0x2C.
13566 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
13567 ? Type::getInt8PtrTy(*DAG.getContext(),
13569 : Type::getInt32PtrTy(*DAG.getContext(),
13573 Subtarget->is64Bit()
13574 ? DAG.getIntPtrConstant(0x58)
13575 : (Subtarget->isTargetWindowsGNU()
13576 ? DAG.getIntPtrConstant(0x2C)
13577 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
13579 SDValue ThreadPointer =
13580 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
13581 MachinePointerInfo(Ptr), false, false, false, 0);
13583 // Load the _tls_index variable
13584 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
13585 if (Subtarget->is64Bit())
13586 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
13587 IDX, MachinePointerInfo(), MVT::i32,
13588 false, false, false, 0);
13590 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
13591 false, false, false, 0);
13593 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
13595 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
13597 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
13598 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
13599 false, false, false, 0);
13601 // Get the offset of start of .tls section
13602 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
13603 GA->getValueType(0),
13604 GA->getOffset(), X86II::MO_SECREL);
13605 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
13607 // The address of the thread local variable is the add of the thread
13608 // pointer with the offset of the variable.
13609 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
13612 llvm_unreachable("TLS not implemented for this target.");
13615 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
13616 /// and take a 2 x i32 value to shift plus a shift amount.
13617 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
13618 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
13619 MVT VT = Op.getSimpleValueType();
13620 unsigned VTBits = VT.getSizeInBits();
13622 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
13623 SDValue ShOpLo = Op.getOperand(0);
13624 SDValue ShOpHi = Op.getOperand(1);
13625 SDValue ShAmt = Op.getOperand(2);
13626 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
13627 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
13629 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
13630 DAG.getConstant(VTBits - 1, MVT::i8));
13631 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
13632 DAG.getConstant(VTBits - 1, MVT::i8))
13633 : DAG.getConstant(0, VT);
13635 SDValue Tmp2, Tmp3;
13636 if (Op.getOpcode() == ISD::SHL_PARTS) {
13637 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
13638 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
13640 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
13641 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
13644 // If the shift amount is larger or equal than the width of a part we can't
13645 // rely on the results of shld/shrd. Insert a test and select the appropriate
13646 // values for large shift amounts.
13647 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
13648 DAG.getConstant(VTBits, MVT::i8));
13649 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13650 AndNode, DAG.getConstant(0, MVT::i8));
13653 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13654 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
13655 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
13657 if (Op.getOpcode() == ISD::SHL_PARTS) {
13658 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
13659 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
13661 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
13662 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
13665 SDValue Ops[2] = { Lo, Hi };
13666 return DAG.getMergeValues(Ops, dl);
13669 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
13670 SelectionDAG &DAG) const {
13671 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
13674 if (SrcVT.isVector()) {
13675 if (SrcVT.getVectorElementType() == MVT::i1) {
13676 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
13677 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
13678 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
13679 Op.getOperand(0)));
13684 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
13685 "Unknown SINT_TO_FP to lower!");
13687 // These are really Legal; return the operand so the caller accepts it as
13689 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
13691 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
13692 Subtarget->is64Bit()) {
13696 unsigned Size = SrcVT.getSizeInBits()/8;
13697 MachineFunction &MF = DAG.getMachineFunction();
13698 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
13699 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
13700 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
13702 MachinePointerInfo::getFixedStack(SSFI),
13704 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
13707 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
13709 SelectionDAG &DAG) const {
13713 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
13715 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
13717 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
13719 unsigned ByteSize = SrcVT.getSizeInBits()/8;
13721 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
13722 MachineMemOperand *MMO;
13724 int SSFI = FI->getIndex();
13726 DAG.getMachineFunction()
13727 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
13728 MachineMemOperand::MOLoad, ByteSize, ByteSize);
13730 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
13731 StackSlot = StackSlot.getOperand(1);
13733 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
13734 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
13736 Tys, Ops, SrcVT, MMO);
13739 Chain = Result.getValue(1);
13740 SDValue InFlag = Result.getValue(2);
13742 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
13743 // shouldn't be necessary except that RFP cannot be live across
13744 // multiple blocks. When stackifier is fixed, they can be uncoupled.
13745 MachineFunction &MF = DAG.getMachineFunction();
13746 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
13747 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
13748 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
13749 Tys = DAG.getVTList(MVT::Other);
13751 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
13753 MachineMemOperand *MMO =
13754 DAG.getMachineFunction()
13755 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
13756 MachineMemOperand::MOStore, SSFISize, SSFISize);
13758 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
13759 Ops, Op.getValueType(), MMO);
13760 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
13761 MachinePointerInfo::getFixedStack(SSFI),
13762 false, false, false, 0);
13768 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
13769 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
13770 SelectionDAG &DAG) const {
13771 // This algorithm is not obvious. Here it is what we're trying to output:
13774 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
13775 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
13777 haddpd %xmm0, %xmm0
13779 pshufd $0x4e, %xmm0, %xmm1
13785 LLVMContext *Context = DAG.getContext();
13787 // Build some magic constants.
13788 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
13789 Constant *C0 = ConstantDataVector::get(*Context, CV0);
13790 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
13792 SmallVector<Constant*,2> CV1;
13794 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
13795 APInt(64, 0x4330000000000000ULL))));
13797 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
13798 APInt(64, 0x4530000000000000ULL))));
13799 Constant *C1 = ConstantVector::get(CV1);
13800 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
13802 // Load the 64-bit value into an XMM register.
13803 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
13805 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
13806 MachinePointerInfo::getConstantPool(),
13807 false, false, false, 16);
13808 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
13809 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
13812 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
13813 MachinePointerInfo::getConstantPool(),
13814 false, false, false, 16);
13815 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
13816 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
13819 if (Subtarget->hasSSE3()) {
13820 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
13821 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
13823 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
13824 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
13826 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
13827 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
13831 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
13832 DAG.getIntPtrConstant(0));
13835 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
13836 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
13837 SelectionDAG &DAG) const {
13839 // FP constant to bias correct the final result.
13840 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
13843 // Load the 32-bit value into an XMM register.
13844 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
13847 // Zero out the upper parts of the register.
13848 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
13850 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
13851 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
13852 DAG.getIntPtrConstant(0));
13854 // Or the load with the bias.
13855 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
13856 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
13857 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
13858 MVT::v2f64, Load)),
13859 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
13860 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
13861 MVT::v2f64, Bias)));
13862 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
13863 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
13864 DAG.getIntPtrConstant(0));
13866 // Subtract the bias.
13867 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
13869 // Handle final rounding.
13870 EVT DestVT = Op.getValueType();
13872 if (DestVT.bitsLT(MVT::f64))
13873 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
13874 DAG.getIntPtrConstant(0));
13875 if (DestVT.bitsGT(MVT::f64))
13876 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
13878 // Handle final rounding.
13882 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
13883 const X86Subtarget &Subtarget) {
13884 // The algorithm is the following:
13885 // #ifdef __SSE4_1__
13886 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
13887 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
13888 // (uint4) 0x53000000, 0xaa);
13890 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
13891 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
13893 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
13894 // return (float4) lo + fhi;
13897 SDValue V = Op->getOperand(0);
13898 EVT VecIntVT = V.getValueType();
13899 bool Is128 = VecIntVT == MVT::v4i32;
13900 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
13901 // If we convert to something else than the supported type, e.g., to v4f64,
13903 if (VecFloatVT != Op->getValueType(0))
13906 unsigned NumElts = VecIntVT.getVectorNumElements();
13907 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
13908 "Unsupported custom type");
13909 assert(NumElts <= 8 && "The size of the constant array must be fixed");
13911 // In the #idef/#else code, we have in common:
13912 // - The vector of constants:
13918 // Create the splat vector for 0x4b000000.
13919 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
13920 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
13921 CstLow, CstLow, CstLow, CstLow};
13922 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
13923 makeArrayRef(&CstLowArray[0], NumElts));
13924 // Create the splat vector for 0x53000000.
13925 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
13926 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
13927 CstHigh, CstHigh, CstHigh, CstHigh};
13928 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
13929 makeArrayRef(&CstHighArray[0], NumElts));
13931 // Create the right shift.
13932 SDValue CstShift = DAG.getConstant(16, MVT::i32);
13933 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
13934 CstShift, CstShift, CstShift, CstShift};
13935 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
13936 makeArrayRef(&CstShiftArray[0], NumElts));
13937 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
13940 if (Subtarget.hasSSE41()) {
13941 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
13942 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
13943 SDValue VecCstLowBitcast =
13944 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
13945 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
13946 // Low will be bitcasted right away, so do not bother bitcasting back to its
13948 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
13949 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
13950 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
13951 // (uint4) 0x53000000, 0xaa);
13952 SDValue VecCstHighBitcast =
13953 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
13954 SDValue VecShiftBitcast =
13955 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
13956 // High will be bitcasted right away, so do not bother bitcasting back to
13957 // its original type.
13958 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
13959 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
13961 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
13962 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
13963 CstMask, CstMask, CstMask);
13964 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
13965 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
13966 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
13968 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
13969 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
13972 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
13973 SDValue CstFAdd = DAG.getConstantFP(
13974 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
13975 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
13976 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
13977 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
13978 makeArrayRef(&CstFAddArray[0], NumElts));
13980 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
13981 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
13983 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
13984 // return (float4) lo + fhi;
13985 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
13986 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
13989 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
13990 SelectionDAG &DAG) const {
13991 SDValue N0 = Op.getOperand(0);
13992 MVT SVT = N0.getSimpleValueType();
13995 switch (SVT.SimpleTy) {
13997 llvm_unreachable("Custom UINT_TO_FP is not supported!");
14002 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
14003 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
14004 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
14008 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
14010 llvm_unreachable(nullptr);
14013 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
14014 SelectionDAG &DAG) const {
14015 SDValue N0 = Op.getOperand(0);
14018 if (Op.getValueType().isVector())
14019 return lowerUINT_TO_FP_vec(Op, DAG);
14021 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
14022 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
14023 // the optimization here.
14024 if (DAG.SignBitIsZero(N0))
14025 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
14027 MVT SrcVT = N0.getSimpleValueType();
14028 MVT DstVT = Op.getSimpleValueType();
14029 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
14030 return LowerUINT_TO_FP_i64(Op, DAG);
14031 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
14032 return LowerUINT_TO_FP_i32(Op, DAG);
14033 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
14036 // Make a 64-bit buffer, and use it to build an FILD.
14037 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
14038 if (SrcVT == MVT::i32) {
14039 SDValue WordOff = DAG.getConstant(4, getPointerTy());
14040 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
14041 getPointerTy(), StackSlot, WordOff);
14042 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14043 StackSlot, MachinePointerInfo(),
14045 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
14046 OffsetSlot, MachinePointerInfo(),
14048 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
14052 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
14053 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
14054 StackSlot, MachinePointerInfo(),
14056 // For i64 source, we need to add the appropriate power of 2 if the input
14057 // was negative. This is the same as the optimization in
14058 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
14059 // we must be careful to do the computation in x87 extended precision, not
14060 // in SSE. (The generic code can't know it's OK to do this, or how to.)
14061 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
14062 MachineMemOperand *MMO =
14063 DAG.getMachineFunction()
14064 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14065 MachineMemOperand::MOLoad, 8, 8);
14067 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
14068 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
14069 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
14072 APInt FF(32, 0x5F800000ULL);
14074 // Check whether the sign bit is set.
14075 SDValue SignSet = DAG.getSetCC(dl,
14076 getSetCCResultType(*DAG.getContext(), MVT::i64),
14077 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
14080 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
14081 SDValue FudgePtr = DAG.getConstantPool(
14082 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
14085 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
14086 SDValue Zero = DAG.getIntPtrConstant(0);
14087 SDValue Four = DAG.getIntPtrConstant(4);
14088 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
14090 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
14092 // Load the value out, extending it from f32 to f80.
14093 // FIXME: Avoid the extend by constructing the right constant pool?
14094 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
14095 FudgePtr, MachinePointerInfo::getConstantPool(),
14096 MVT::f32, false, false, false, 4);
14097 // Extend everything to 80 bits to force it to be done on x87.
14098 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
14099 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
14102 std::pair<SDValue,SDValue>
14103 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
14104 bool IsSigned, bool IsReplace) const {
14107 EVT DstTy = Op.getValueType();
14109 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
14110 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
14114 assert(DstTy.getSimpleVT() <= MVT::i64 &&
14115 DstTy.getSimpleVT() >= MVT::i16 &&
14116 "Unknown FP_TO_INT to lower!");
14118 // These are really Legal.
14119 if (DstTy == MVT::i32 &&
14120 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14121 return std::make_pair(SDValue(), SDValue());
14122 if (Subtarget->is64Bit() &&
14123 DstTy == MVT::i64 &&
14124 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
14125 return std::make_pair(SDValue(), SDValue());
14127 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
14128 // stack slot, or into the FTOL runtime function.
14129 MachineFunction &MF = DAG.getMachineFunction();
14130 unsigned MemSize = DstTy.getSizeInBits()/8;
14131 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14132 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14135 if (!IsSigned && isIntegerTypeFTOL(DstTy))
14136 Opc = X86ISD::WIN_FTOL;
14138 switch (DstTy.getSimpleVT().SimpleTy) {
14139 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
14140 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
14141 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
14142 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
14145 SDValue Chain = DAG.getEntryNode();
14146 SDValue Value = Op.getOperand(0);
14147 EVT TheVT = Op.getOperand(0).getValueType();
14148 // FIXME This causes a redundant load/store if the SSE-class value is already
14149 // in memory, such as if it is on the callstack.
14150 if (isScalarFPTypeInSSEReg(TheVT)) {
14151 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
14152 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
14153 MachinePointerInfo::getFixedStack(SSFI),
14155 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
14157 Chain, StackSlot, DAG.getValueType(TheVT)
14160 MachineMemOperand *MMO =
14161 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14162 MachineMemOperand::MOLoad, MemSize, MemSize);
14163 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
14164 Chain = Value.getValue(1);
14165 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
14166 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
14169 MachineMemOperand *MMO =
14170 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
14171 MachineMemOperand::MOStore, MemSize, MemSize);
14173 if (Opc != X86ISD::WIN_FTOL) {
14174 // Build the FP_TO_INT*_IN_MEM
14175 SDValue Ops[] = { Chain, Value, StackSlot };
14176 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
14178 return std::make_pair(FIST, StackSlot);
14180 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
14181 DAG.getVTList(MVT::Other, MVT::Glue),
14183 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
14184 MVT::i32, ftol.getValue(1));
14185 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
14186 MVT::i32, eax.getValue(2));
14187 SDValue Ops[] = { eax, edx };
14188 SDValue pair = IsReplace
14189 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
14190 : DAG.getMergeValues(Ops, DL);
14191 return std::make_pair(pair, SDValue());
14195 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
14196 const X86Subtarget *Subtarget) {
14197 MVT VT = Op->getSimpleValueType(0);
14198 SDValue In = Op->getOperand(0);
14199 MVT InVT = In.getSimpleValueType();
14202 // Optimize vectors in AVX mode:
14205 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
14206 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
14207 // Concat upper and lower parts.
14210 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
14211 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
14212 // Concat upper and lower parts.
14215 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
14216 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
14217 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
14220 if (Subtarget->hasInt256())
14221 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
14223 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
14224 SDValue Undef = DAG.getUNDEF(InVT);
14225 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
14226 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14227 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
14229 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
14230 VT.getVectorNumElements()/2);
14232 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
14233 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
14235 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14238 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
14239 SelectionDAG &DAG) {
14240 MVT VT = Op->getSimpleValueType(0);
14241 SDValue In = Op->getOperand(0);
14242 MVT InVT = In.getSimpleValueType();
14244 unsigned int NumElts = VT.getVectorNumElements();
14245 if (NumElts != 8 && NumElts != 16)
14248 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
14249 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
14251 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
14252 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14253 // Now we have only mask extension
14254 assert(InVT.getVectorElementType() == MVT::i1);
14255 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
14256 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14257 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
14258 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14259 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14260 MachinePointerInfo::getConstantPool(),
14261 false, false, false, Alignment);
14263 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
14264 if (VT.is512BitVector())
14266 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
14269 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14270 SelectionDAG &DAG) {
14271 if (Subtarget->hasFp256()) {
14272 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14280 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14281 SelectionDAG &DAG) {
14283 MVT VT = Op.getSimpleValueType();
14284 SDValue In = Op.getOperand(0);
14285 MVT SVT = In.getSimpleValueType();
14287 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
14288 return LowerZERO_EXTEND_AVX512(Op, DAG);
14290 if (Subtarget->hasFp256()) {
14291 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
14296 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
14297 VT.getVectorNumElements() != SVT.getVectorNumElements());
14301 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
14303 MVT VT = Op.getSimpleValueType();
14304 SDValue In = Op.getOperand(0);
14305 MVT InVT = In.getSimpleValueType();
14307 if (VT == MVT::i1) {
14308 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
14309 "Invalid scalar TRUNCATE operation");
14310 if (InVT.getSizeInBits() >= 32)
14312 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
14313 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
14315 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
14316 "Invalid TRUNCATE operation");
14318 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
14319 if (VT.getVectorElementType().getSizeInBits() >=8)
14320 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
14322 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14323 unsigned NumElts = InVT.getVectorNumElements();
14324 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
14325 if (InVT.getSizeInBits() < 512) {
14326 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
14327 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
14331 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
14332 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
14333 SDValue CP = DAG.getConstantPool(C, getPointerTy());
14334 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
14335 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
14336 MachinePointerInfo::getConstantPool(),
14337 false, false, false, Alignment);
14338 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
14339 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
14340 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
14343 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
14344 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
14345 if (Subtarget->hasInt256()) {
14346 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
14347 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
14348 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
14350 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
14351 DAG.getIntPtrConstant(0));
14354 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14355 DAG.getIntPtrConstant(0));
14356 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14357 DAG.getIntPtrConstant(2));
14358 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14359 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14360 static const int ShufMask[] = {0, 2, 4, 6};
14361 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
14364 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
14365 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
14366 if (Subtarget->hasInt256()) {
14367 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
14369 SmallVector<SDValue,32> pshufbMask;
14370 for (unsigned i = 0; i < 2; ++i) {
14371 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
14372 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
14373 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
14374 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
14375 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
14376 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
14377 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
14378 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
14379 for (unsigned j = 0; j < 8; ++j)
14380 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
14382 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
14383 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
14384 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
14386 static const int ShufMask[] = {0, 2, -1, -1};
14387 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
14389 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
14390 DAG.getIntPtrConstant(0));
14391 return DAG.getNode(ISD::BITCAST, DL, VT, In);
14394 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14395 DAG.getIntPtrConstant(0));
14397 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
14398 DAG.getIntPtrConstant(4));
14400 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
14401 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
14403 // The PSHUFB mask:
14404 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
14405 -1, -1, -1, -1, -1, -1, -1, -1};
14407 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
14408 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
14409 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
14411 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
14412 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
14414 // The MOVLHPS Mask:
14415 static const int ShufMask2[] = {0, 1, 4, 5};
14416 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
14417 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
14420 // Handle truncation of V256 to V128 using shuffles.
14421 if (!VT.is128BitVector() || !InVT.is256BitVector())
14424 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
14426 unsigned NumElems = VT.getVectorNumElements();
14427 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
14429 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
14430 // Prepare truncation shuffle mask
14431 for (unsigned i = 0; i != NumElems; ++i)
14432 MaskVec[i] = i * 2;
14433 SDValue V = DAG.getVectorShuffle(NVT, DL,
14434 DAG.getNode(ISD::BITCAST, DL, NVT, In),
14435 DAG.getUNDEF(NVT), &MaskVec[0]);
14436 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
14437 DAG.getIntPtrConstant(0));
14440 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
14441 SelectionDAG &DAG) const {
14442 assert(!Op.getSimpleValueType().isVector());
14444 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
14445 /*IsSigned=*/ true, /*IsReplace=*/ false);
14446 SDValue FIST = Vals.first, StackSlot = Vals.second;
14447 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
14448 if (!FIST.getNode()) return Op;
14450 if (StackSlot.getNode())
14451 // Load the result.
14452 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
14453 FIST, StackSlot, MachinePointerInfo(),
14454 false, false, false, 0);
14456 // The node is the result.
14460 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
14461 SelectionDAG &DAG) const {
14462 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
14463 /*IsSigned=*/ false, /*IsReplace=*/ false);
14464 SDValue FIST = Vals.first, StackSlot = Vals.second;
14465 assert(FIST.getNode() && "Unexpected failure");
14467 if (StackSlot.getNode())
14468 // Load the result.
14469 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
14470 FIST, StackSlot, MachinePointerInfo(),
14471 false, false, false, 0);
14473 // The node is the result.
14477 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
14479 MVT VT = Op.getSimpleValueType();
14480 SDValue In = Op.getOperand(0);
14481 MVT SVT = In.getSimpleValueType();
14483 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
14485 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
14486 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
14487 In, DAG.getUNDEF(SVT)));
14490 /// The only differences between FABS and FNEG are the mask and the logic op.
14491 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
14492 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
14493 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
14494 "Wrong opcode for lowering FABS or FNEG.");
14496 bool IsFABS = (Op.getOpcode() == ISD::FABS);
14498 // If this is a FABS and it has an FNEG user, bail out to fold the combination
14499 // into an FNABS. We'll lower the FABS after that if it is still in use.
14501 for (SDNode *User : Op->uses())
14502 if (User->getOpcode() == ISD::FNEG)
14505 SDValue Op0 = Op.getOperand(0);
14506 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
14509 MVT VT = Op.getSimpleValueType();
14510 // Assume scalar op for initialization; update for vector if needed.
14511 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
14512 // generate a 16-byte vector constant and logic op even for the scalar case.
14513 // Using a 16-byte mask allows folding the load of the mask with
14514 // the logic op, so it can save (~4 bytes) on code size.
14516 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
14517 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
14518 // decide if we should generate a 16-byte constant mask when we only need 4 or
14519 // 8 bytes for the scalar case.
14520 if (VT.isVector()) {
14521 EltVT = VT.getVectorElementType();
14522 NumElts = VT.getVectorNumElements();
14525 unsigned EltBits = EltVT.getSizeInBits();
14526 LLVMContext *Context = DAG.getContext();
14527 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
14529 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
14530 Constant *C = ConstantInt::get(*Context, MaskElt);
14531 C = ConstantVector::getSplat(NumElts, C);
14532 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14533 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
14534 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
14535 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
14536 MachinePointerInfo::getConstantPool(),
14537 false, false, false, Alignment);
14539 if (VT.isVector()) {
14540 // For a vector, cast operands to a vector type, perform the logic op,
14541 // and cast the result back to the original value type.
14542 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
14543 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
14544 SDValue Operand = IsFNABS ?
14545 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
14546 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
14547 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
14548 return DAG.getNode(ISD::BITCAST, dl, VT,
14549 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
14552 // If not vector, then scalar.
14553 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
14554 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
14555 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
14558 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
14559 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14560 LLVMContext *Context = DAG.getContext();
14561 SDValue Op0 = Op.getOperand(0);
14562 SDValue Op1 = Op.getOperand(1);
14564 MVT VT = Op.getSimpleValueType();
14565 MVT SrcVT = Op1.getSimpleValueType();
14567 // If second operand is smaller, extend it first.
14568 if (SrcVT.bitsLT(VT)) {
14569 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
14572 // And if it is bigger, shrink it first.
14573 if (SrcVT.bitsGT(VT)) {
14574 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
14578 // At this point the operands and the result should have the same
14579 // type, and that won't be f80 since that is not custom lowered.
14581 const fltSemantics &Sem =
14582 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
14583 const unsigned SizeInBits = VT.getSizeInBits();
14585 SmallVector<Constant *, 4> CV(
14586 VT == MVT::f64 ? 2 : 4,
14587 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
14589 // First, clear all bits but the sign bit from the second operand (sign).
14590 CV[0] = ConstantFP::get(*Context,
14591 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
14592 Constant *C = ConstantVector::get(CV);
14593 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
14594 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
14595 MachinePointerInfo::getConstantPool(),
14596 false, false, false, 16);
14597 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
14599 // Next, clear the sign bit from the first operand (magnitude).
14600 // If it's a constant, we can clear it here.
14601 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
14602 APFloat APF = Op0CN->getValueAPF();
14603 // If the magnitude is a positive zero, the sign bit alone is enough.
14604 if (APF.isPosZero())
14607 CV[0] = ConstantFP::get(*Context, APF);
14609 CV[0] = ConstantFP::get(
14611 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
14613 C = ConstantVector::get(CV);
14614 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
14615 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
14616 MachinePointerInfo::getConstantPool(),
14617 false, false, false, 16);
14618 // If the magnitude operand wasn't a constant, we need to AND out the sign.
14619 if (!isa<ConstantFPSDNode>(Op0))
14620 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
14622 // OR the magnitude value with the sign bit.
14623 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
14626 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
14627 SDValue N0 = Op.getOperand(0);
14629 MVT VT = Op.getSimpleValueType();
14631 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
14632 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
14633 DAG.getConstant(1, VT));
14634 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
14637 // Check whether an OR'd tree is PTEST-able.
14638 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
14639 SelectionDAG &DAG) {
14640 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
14642 if (!Subtarget->hasSSE41())
14645 if (!Op->hasOneUse())
14648 SDNode *N = Op.getNode();
14651 SmallVector<SDValue, 8> Opnds;
14652 DenseMap<SDValue, unsigned> VecInMap;
14653 SmallVector<SDValue, 8> VecIns;
14654 EVT VT = MVT::Other;
14656 // Recognize a special case where a vector is casted into wide integer to
14658 Opnds.push_back(N->getOperand(0));
14659 Opnds.push_back(N->getOperand(1));
14661 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
14662 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
14663 // BFS traverse all OR'd operands.
14664 if (I->getOpcode() == ISD::OR) {
14665 Opnds.push_back(I->getOperand(0));
14666 Opnds.push_back(I->getOperand(1));
14667 // Re-evaluate the number of nodes to be traversed.
14668 e += 2; // 2 more nodes (LHS and RHS) are pushed.
14672 // Quit if a non-EXTRACT_VECTOR_ELT
14673 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
14676 // Quit if without a constant index.
14677 SDValue Idx = I->getOperand(1);
14678 if (!isa<ConstantSDNode>(Idx))
14681 SDValue ExtractedFromVec = I->getOperand(0);
14682 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
14683 if (M == VecInMap.end()) {
14684 VT = ExtractedFromVec.getValueType();
14685 // Quit if not 128/256-bit vector.
14686 if (!VT.is128BitVector() && !VT.is256BitVector())
14688 // Quit if not the same type.
14689 if (VecInMap.begin() != VecInMap.end() &&
14690 VT != VecInMap.begin()->first.getValueType())
14692 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
14693 VecIns.push_back(ExtractedFromVec);
14695 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
14698 assert((VT.is128BitVector() || VT.is256BitVector()) &&
14699 "Not extracted from 128-/256-bit vector.");
14701 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
14703 for (DenseMap<SDValue, unsigned>::const_iterator
14704 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
14705 // Quit if not all elements are used.
14706 if (I->second != FullMask)
14710 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
14712 // Cast all vectors into TestVT for PTEST.
14713 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
14714 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
14716 // If more than one full vectors are evaluated, OR them first before PTEST.
14717 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
14718 // Each iteration will OR 2 nodes and append the result until there is only
14719 // 1 node left, i.e. the final OR'd value of all vectors.
14720 SDValue LHS = VecIns[Slot];
14721 SDValue RHS = VecIns[Slot + 1];
14722 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
14725 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
14726 VecIns.back(), VecIns.back());
14729 /// \brief return true if \c Op has a use that doesn't just read flags.
14730 static bool hasNonFlagsUse(SDValue Op) {
14731 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
14733 SDNode *User = *UI;
14734 unsigned UOpNo = UI.getOperandNo();
14735 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
14736 // Look pass truncate.
14737 UOpNo = User->use_begin().getOperandNo();
14738 User = *User->use_begin();
14741 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
14742 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
14748 /// Emit nodes that will be selected as "test Op0,Op0", or something
14750 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
14751 SelectionDAG &DAG) const {
14752 if (Op.getValueType() == MVT::i1)
14753 // KORTEST instruction should be selected
14754 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14755 DAG.getConstant(0, Op.getValueType()));
14757 // CF and OF aren't always set the way we want. Determine which
14758 // of these we need.
14759 bool NeedCF = false;
14760 bool NeedOF = false;
14763 case X86::COND_A: case X86::COND_AE:
14764 case X86::COND_B: case X86::COND_BE:
14767 case X86::COND_G: case X86::COND_GE:
14768 case X86::COND_L: case X86::COND_LE:
14769 case X86::COND_O: case X86::COND_NO: {
14770 // Check if we really need to set the
14771 // Overflow flag. If NoSignedWrap is present
14772 // that is not actually needed.
14773 switch (Op->getOpcode()) {
14778 const BinaryWithFlagsSDNode *BinNode =
14779 cast<BinaryWithFlagsSDNode>(Op.getNode());
14780 if (BinNode->hasNoSignedWrap())
14790 // See if we can use the EFLAGS value from the operand instead of
14791 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
14792 // we prove that the arithmetic won't overflow, we can't use OF or CF.
14793 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
14794 // Emit a CMP with 0, which is the TEST pattern.
14795 //if (Op.getValueType() == MVT::i1)
14796 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
14797 // DAG.getConstant(0, MVT::i1));
14798 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14799 DAG.getConstant(0, Op.getValueType()));
14801 unsigned Opcode = 0;
14802 unsigned NumOperands = 0;
14804 // Truncate operations may prevent the merge of the SETCC instruction
14805 // and the arithmetic instruction before it. Attempt to truncate the operands
14806 // of the arithmetic instruction and use a reduced bit-width instruction.
14807 bool NeedTruncation = false;
14808 SDValue ArithOp = Op;
14809 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
14810 SDValue Arith = Op->getOperand(0);
14811 // Both the trunc and the arithmetic op need to have one user each.
14812 if (Arith->hasOneUse())
14813 switch (Arith.getOpcode()) {
14820 NeedTruncation = true;
14826 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
14827 // which may be the result of a CAST. We use the variable 'Op', which is the
14828 // non-casted variable when we check for possible users.
14829 switch (ArithOp.getOpcode()) {
14831 // Due to an isel shortcoming, be conservative if this add is likely to be
14832 // selected as part of a load-modify-store instruction. When the root node
14833 // in a match is a store, isel doesn't know how to remap non-chain non-flag
14834 // uses of other nodes in the match, such as the ADD in this case. This
14835 // leads to the ADD being left around and reselected, with the result being
14836 // two adds in the output. Alas, even if none our users are stores, that
14837 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
14838 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
14839 // climbing the DAG back to the root, and it doesn't seem to be worth the
14841 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14842 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14843 if (UI->getOpcode() != ISD::CopyToReg &&
14844 UI->getOpcode() != ISD::SETCC &&
14845 UI->getOpcode() != ISD::STORE)
14848 if (ConstantSDNode *C =
14849 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
14850 // An add of one will be selected as an INC.
14851 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
14852 Opcode = X86ISD::INC;
14857 // An add of negative one (subtract of one) will be selected as a DEC.
14858 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
14859 Opcode = X86ISD::DEC;
14865 // Otherwise use a regular EFLAGS-setting add.
14866 Opcode = X86ISD::ADD;
14871 // If we have a constant logical shift that's only used in a comparison
14872 // against zero turn it into an equivalent AND. This allows turning it into
14873 // a TEST instruction later.
14874 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
14875 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
14876 EVT VT = Op.getValueType();
14877 unsigned BitWidth = VT.getSizeInBits();
14878 unsigned ShAmt = Op->getConstantOperandVal(1);
14879 if (ShAmt >= BitWidth) // Avoid undefined shifts.
14881 APInt Mask = ArithOp.getOpcode() == ISD::SRL
14882 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
14883 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
14884 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
14886 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
14887 DAG.getConstant(Mask, VT));
14888 DAG.ReplaceAllUsesWith(Op, New);
14894 // If the primary and result isn't used, don't bother using X86ISD::AND,
14895 // because a TEST instruction will be better.
14896 if (!hasNonFlagsUse(Op))
14902 // Due to the ISEL shortcoming noted above, be conservative if this op is
14903 // likely to be selected as part of a load-modify-store instruction.
14904 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
14905 UE = Op.getNode()->use_end(); UI != UE; ++UI)
14906 if (UI->getOpcode() == ISD::STORE)
14909 // Otherwise use a regular EFLAGS-setting instruction.
14910 switch (ArithOp.getOpcode()) {
14911 default: llvm_unreachable("unexpected operator!");
14912 case ISD::SUB: Opcode = X86ISD::SUB; break;
14913 case ISD::XOR: Opcode = X86ISD::XOR; break;
14914 case ISD::AND: Opcode = X86ISD::AND; break;
14916 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
14917 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
14918 if (EFLAGS.getNode())
14921 Opcode = X86ISD::OR;
14935 return SDValue(Op.getNode(), 1);
14941 // If we found that truncation is beneficial, perform the truncation and
14943 if (NeedTruncation) {
14944 EVT VT = Op.getValueType();
14945 SDValue WideVal = Op->getOperand(0);
14946 EVT WideVT = WideVal.getValueType();
14947 unsigned ConvertedOp = 0;
14948 // Use a target machine opcode to prevent further DAGCombine
14949 // optimizations that may separate the arithmetic operations
14950 // from the setcc node.
14951 switch (WideVal.getOpcode()) {
14953 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
14954 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
14955 case ISD::AND: ConvertedOp = X86ISD::AND; break;
14956 case ISD::OR: ConvertedOp = X86ISD::OR; break;
14957 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
14961 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14962 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
14963 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
14964 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
14965 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
14971 // Emit a CMP with 0, which is the TEST pattern.
14972 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
14973 DAG.getConstant(0, Op.getValueType()));
14975 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14976 SmallVector<SDValue, 4> Ops;
14977 for (unsigned i = 0; i != NumOperands; ++i)
14978 Ops.push_back(Op.getOperand(i));
14980 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
14981 DAG.ReplaceAllUsesWith(Op, New);
14982 return SDValue(New.getNode(), 1);
14985 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
14987 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
14988 SDLoc dl, SelectionDAG &DAG) const {
14989 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
14990 if (C->getAPIntValue() == 0)
14991 return EmitTest(Op0, X86CC, dl, DAG);
14993 if (Op0.getValueType() == MVT::i1)
14994 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
14997 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
14998 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
14999 // Do the comparison at i32 if it's smaller, besides the Atom case.
15000 // This avoids subregister aliasing issues. Keep the smaller reference
15001 // if we're optimizing for size, however, as that'll allow better folding
15002 // of memory operations.
15003 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
15004 !DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
15005 AttributeSet::FunctionIndex, Attribute::MinSize) &&
15006 !Subtarget->isAtom()) {
15007 unsigned ExtendOp =
15008 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
15009 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
15010 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
15012 // Use SUB instead of CMP to enable CSE between SUB and CMP.
15013 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
15014 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
15016 return SDValue(Sub.getNode(), 1);
15018 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
15021 /// Convert a comparison if required by the subtarget.
15022 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
15023 SelectionDAG &DAG) const {
15024 // If the subtarget does not support the FUCOMI instruction, floating-point
15025 // comparisons have to be converted.
15026 if (Subtarget->hasCMov() ||
15027 Cmp.getOpcode() != X86ISD::CMP ||
15028 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
15029 !Cmp.getOperand(1).getValueType().isFloatingPoint())
15032 // The instruction selector will select an FUCOM instruction instead of
15033 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
15034 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
15035 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
15037 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
15038 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
15039 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
15040 DAG.getConstant(8, MVT::i8));
15041 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
15042 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
15045 /// The minimum architected relative accuracy is 2^-12. We need one
15046 /// Newton-Raphson step to have a good float result (24 bits of precision).
15047 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
15048 DAGCombinerInfo &DCI,
15049 unsigned &RefinementSteps,
15050 bool &UseOneConstNR) const {
15051 // FIXME: We should use instruction latency models to calculate the cost of
15052 // each potential sequence, but this is very hard to do reliably because
15053 // at least Intel's Core* chips have variable timing based on the number of
15054 // significant digits in the divisor and/or sqrt operand.
15055 if (!Subtarget->useSqrtEst())
15058 EVT VT = Op.getValueType();
15060 // SSE1 has rsqrtss and rsqrtps.
15061 // TODO: Add support for AVX512 (v16f32).
15062 // It is likely not profitable to do this for f64 because a double-precision
15063 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
15064 // instructions: convert to single, rsqrtss, convert back to double, refine
15065 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
15066 // along with FMA, this could be a throughput win.
15067 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15068 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15069 RefinementSteps = 1;
15070 UseOneConstNR = false;
15071 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
15076 /// The minimum architected relative accuracy is 2^-12. We need one
15077 /// Newton-Raphson step to have a good float result (24 bits of precision).
15078 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
15079 DAGCombinerInfo &DCI,
15080 unsigned &RefinementSteps) const {
15081 // FIXME: We should use instruction latency models to calculate the cost of
15082 // each potential sequence, but this is very hard to do reliably because
15083 // at least Intel's Core* chips have variable timing based on the number of
15084 // significant digits in the divisor.
15085 if (!Subtarget->useReciprocalEst())
15088 EVT VT = Op.getValueType();
15090 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
15091 // TODO: Add support for AVX512 (v16f32).
15092 // It is likely not profitable to do this for f64 because a double-precision
15093 // reciprocal estimate with refinement on x86 prior to FMA requires
15094 // 15 instructions: convert to single, rcpss, convert back to double, refine
15095 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
15096 // along with FMA, this could be a throughput win.
15097 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
15098 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
15099 RefinementSteps = ReciprocalEstimateRefinementSteps;
15100 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
15105 static bool isAllOnes(SDValue V) {
15106 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
15107 return C && C->isAllOnesValue();
15110 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
15111 /// if it's possible.
15112 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
15113 SDLoc dl, SelectionDAG &DAG) const {
15114 SDValue Op0 = And.getOperand(0);
15115 SDValue Op1 = And.getOperand(1);
15116 if (Op0.getOpcode() == ISD::TRUNCATE)
15117 Op0 = Op0.getOperand(0);
15118 if (Op1.getOpcode() == ISD::TRUNCATE)
15119 Op1 = Op1.getOperand(0);
15122 if (Op1.getOpcode() == ISD::SHL)
15123 std::swap(Op0, Op1);
15124 if (Op0.getOpcode() == ISD::SHL) {
15125 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
15126 if (And00C->getZExtValue() == 1) {
15127 // If we looked past a truncate, check that it's only truncating away
15129 unsigned BitWidth = Op0.getValueSizeInBits();
15130 unsigned AndBitWidth = And.getValueSizeInBits();
15131 if (BitWidth > AndBitWidth) {
15133 DAG.computeKnownBits(Op0, Zeros, Ones);
15134 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
15138 RHS = Op0.getOperand(1);
15140 } else if (Op1.getOpcode() == ISD::Constant) {
15141 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
15142 uint64_t AndRHSVal = AndRHS->getZExtValue();
15143 SDValue AndLHS = Op0;
15145 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
15146 LHS = AndLHS.getOperand(0);
15147 RHS = AndLHS.getOperand(1);
15150 // Use BT if the immediate can't be encoded in a TEST instruction.
15151 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
15153 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
15157 if (LHS.getNode()) {
15158 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
15159 // instruction. Since the shift amount is in-range-or-undefined, we know
15160 // that doing a bittest on the i32 value is ok. We extend to i32 because
15161 // the encoding for the i16 version is larger than the i32 version.
15162 // Also promote i16 to i32 for performance / code size reason.
15163 if (LHS.getValueType() == MVT::i8 ||
15164 LHS.getValueType() == MVT::i16)
15165 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
15167 // If the operand types disagree, extend the shift amount to match. Since
15168 // BT ignores high bits (like shifts) we can use anyextend.
15169 if (LHS.getValueType() != RHS.getValueType())
15170 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
15172 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
15173 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
15174 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15175 DAG.getConstant(Cond, MVT::i8), BT);
15181 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
15183 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
15188 // SSE Condition code mapping:
15197 switch (SetCCOpcode) {
15198 default: llvm_unreachable("Unexpected SETCC condition");
15200 case ISD::SETEQ: SSECC = 0; break;
15202 case ISD::SETGT: Swap = true; // Fallthrough
15204 case ISD::SETOLT: SSECC = 1; break;
15206 case ISD::SETGE: Swap = true; // Fallthrough
15208 case ISD::SETOLE: SSECC = 2; break;
15209 case ISD::SETUO: SSECC = 3; break;
15211 case ISD::SETNE: SSECC = 4; break;
15212 case ISD::SETULE: Swap = true; // Fallthrough
15213 case ISD::SETUGE: SSECC = 5; break;
15214 case ISD::SETULT: Swap = true; // Fallthrough
15215 case ISD::SETUGT: SSECC = 6; break;
15216 case ISD::SETO: SSECC = 7; break;
15218 case ISD::SETONE: SSECC = 8; break;
15221 std::swap(Op0, Op1);
15226 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
15227 // ones, and then concatenate the result back.
15228 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
15229 MVT VT = Op.getSimpleValueType();
15231 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
15232 "Unsupported value type for operation");
15234 unsigned NumElems = VT.getVectorNumElements();
15236 SDValue CC = Op.getOperand(2);
15238 // Extract the LHS vectors
15239 SDValue LHS = Op.getOperand(0);
15240 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15241 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15243 // Extract the RHS vectors
15244 SDValue RHS = Op.getOperand(1);
15245 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15246 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15248 // Issue the operation on the smaller types and concatenate the result back
15249 MVT EltVT = VT.getVectorElementType();
15250 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15251 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15252 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
15253 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
15256 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
15257 const X86Subtarget *Subtarget) {
15258 SDValue Op0 = Op.getOperand(0);
15259 SDValue Op1 = Op.getOperand(1);
15260 SDValue CC = Op.getOperand(2);
15261 MVT VT = Op.getSimpleValueType();
15264 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
15265 Op.getValueType().getScalarType() == MVT::i1 &&
15266 "Cannot set masked compare for this operation");
15268 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15270 bool Unsigned = false;
15273 switch (SetCCOpcode) {
15274 default: llvm_unreachable("Unexpected SETCC condition");
15275 case ISD::SETNE: SSECC = 4; break;
15276 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
15277 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
15278 case ISD::SETLT: Swap = true; //fall-through
15279 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
15280 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
15281 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
15282 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
15283 case ISD::SETULE: Unsigned = true; //fall-through
15284 case ISD::SETLE: SSECC = 2; break;
15288 std::swap(Op0, Op1);
15290 return DAG.getNode(Opc, dl, VT, Op0, Op1);
15291 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
15292 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15293 DAG.getConstant(SSECC, MVT::i8));
15296 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
15297 /// operand \p Op1. If non-trivial (for example because it's not constant)
15298 /// return an empty value.
15299 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
15301 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
15305 MVT VT = Op1.getSimpleValueType();
15306 MVT EVT = VT.getVectorElementType();
15307 unsigned n = VT.getVectorNumElements();
15308 SmallVector<SDValue, 8> ULTOp1;
15310 for (unsigned i = 0; i < n; ++i) {
15311 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
15312 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
15315 // Avoid underflow.
15316 APInt Val = Elt->getAPIntValue();
15320 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
15323 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
15326 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
15327 SelectionDAG &DAG) {
15328 SDValue Op0 = Op.getOperand(0);
15329 SDValue Op1 = Op.getOperand(1);
15330 SDValue CC = Op.getOperand(2);
15331 MVT VT = Op.getSimpleValueType();
15332 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
15333 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
15338 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
15339 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
15342 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
15343 unsigned Opc = X86ISD::CMPP;
15344 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
15345 assert(VT.getVectorNumElements() <= 16);
15346 Opc = X86ISD::CMPM;
15348 // In the two special cases we can't handle, emit two comparisons.
15351 unsigned CombineOpc;
15352 if (SetCCOpcode == ISD::SETUEQ) {
15353 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
15355 assert(SetCCOpcode == ISD::SETONE);
15356 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
15359 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15360 DAG.getConstant(CC0, MVT::i8));
15361 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
15362 DAG.getConstant(CC1, MVT::i8));
15363 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
15365 // Handle all other FP comparisons here.
15366 return DAG.getNode(Opc, dl, VT, Op0, Op1,
15367 DAG.getConstant(SSECC, MVT::i8));
15370 // Break 256-bit integer vector compare into smaller ones.
15371 if (VT.is256BitVector() && !Subtarget->hasInt256())
15372 return Lower256IntVSETCC(Op, DAG);
15374 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
15375 EVT OpVT = Op1.getValueType();
15376 if (Subtarget->hasAVX512()) {
15377 if (Op1.getValueType().is512BitVector() ||
15378 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
15379 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
15380 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
15382 // In AVX-512 architecture setcc returns mask with i1 elements,
15383 // But there is no compare instruction for i8 and i16 elements in KNL.
15384 // We are not talking about 512-bit operands in this case, these
15385 // types are illegal.
15387 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
15388 OpVT.getVectorElementType().getSizeInBits() >= 8))
15389 return DAG.getNode(ISD::TRUNCATE, dl, VT,
15390 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
15393 // We are handling one of the integer comparisons here. Since SSE only has
15394 // GT and EQ comparisons for integer, swapping operands and multiple
15395 // operations may be required for some comparisons.
15397 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
15398 bool Subus = false;
15400 switch (SetCCOpcode) {
15401 default: llvm_unreachable("Unexpected SETCC condition");
15402 case ISD::SETNE: Invert = true;
15403 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
15404 case ISD::SETLT: Swap = true;
15405 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
15406 case ISD::SETGE: Swap = true;
15407 case ISD::SETLE: Opc = X86ISD::PCMPGT;
15408 Invert = true; break;
15409 case ISD::SETULT: Swap = true;
15410 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
15411 FlipSigns = true; break;
15412 case ISD::SETUGE: Swap = true;
15413 case ISD::SETULE: Opc = X86ISD::PCMPGT;
15414 FlipSigns = true; Invert = true; break;
15417 // Special case: Use min/max operations for SETULE/SETUGE
15418 MVT VET = VT.getVectorElementType();
15420 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
15421 || (Subtarget->hasSSE2() && (VET == MVT::i8));
15424 switch (SetCCOpcode) {
15426 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
15427 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
15430 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
15433 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
15434 if (!MinMax && hasSubus) {
15435 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
15437 // t = psubus Op0, Op1
15438 // pcmpeq t, <0..0>
15439 switch (SetCCOpcode) {
15441 case ISD::SETULT: {
15442 // If the comparison is against a constant we can turn this into a
15443 // setule. With psubus, setule does not require a swap. This is
15444 // beneficial because the constant in the register is no longer
15445 // destructed as the destination so it can be hoisted out of a loop.
15446 // Only do this pre-AVX since vpcmp* is no longer destructive.
15447 if (Subtarget->hasAVX())
15449 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
15450 if (ULEOp1.getNode()) {
15452 Subus = true; Invert = false; Swap = false;
15456 // Psubus is better than flip-sign because it requires no inversion.
15457 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
15458 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
15462 Opc = X86ISD::SUBUS;
15468 std::swap(Op0, Op1);
15470 // Check that the operation in question is available (most are plain SSE2,
15471 // but PCMPGTQ and PCMPEQQ have different requirements).
15472 if (VT == MVT::v2i64) {
15473 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
15474 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
15476 // First cast everything to the right type.
15477 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
15478 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
15480 // Since SSE has no unsigned integer comparisons, we need to flip the sign
15481 // bits of the inputs before performing those operations. The lower
15482 // compare is always unsigned.
15485 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
15487 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
15488 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
15489 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
15490 Sign, Zero, Sign, Zero);
15492 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
15493 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
15495 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
15496 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
15497 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
15499 // Create masks for only the low parts/high parts of the 64 bit integers.
15500 static const int MaskHi[] = { 1, 1, 3, 3 };
15501 static const int MaskLo[] = { 0, 0, 2, 2 };
15502 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
15503 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
15504 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
15506 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
15507 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
15510 Result = DAG.getNOT(dl, Result, MVT::v4i32);
15512 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
15515 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
15516 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
15517 // pcmpeqd + pshufd + pand.
15518 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
15520 // First cast everything to the right type.
15521 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
15522 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
15525 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
15527 // Make sure the lower and upper halves are both all-ones.
15528 static const int Mask[] = { 1, 0, 3, 2 };
15529 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
15530 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
15533 Result = DAG.getNOT(dl, Result, MVT::v4i32);
15535 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
15539 // Since SSE has no unsigned integer comparisons, we need to flip the sign
15540 // bits of the inputs before performing those operations.
15542 EVT EltVT = VT.getVectorElementType();
15543 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
15544 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
15545 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
15548 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
15550 // If the logical-not of the result is required, perform that now.
15552 Result = DAG.getNOT(dl, Result, VT);
15555 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
15558 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
15559 getZeroVector(VT, Subtarget, DAG, dl));
15564 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
15566 MVT VT = Op.getSimpleValueType();
15568 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
15570 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
15571 && "SetCC type must be 8-bit or 1-bit integer");
15572 SDValue Op0 = Op.getOperand(0);
15573 SDValue Op1 = Op.getOperand(1);
15575 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
15577 // Optimize to BT if possible.
15578 // Lower (X & (1 << N)) == 0 to BT(X, N).
15579 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
15580 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
15581 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
15582 Op1.getOpcode() == ISD::Constant &&
15583 cast<ConstantSDNode>(Op1)->isNullValue() &&
15584 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15585 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
15586 if (NewSetCC.getNode()) {
15588 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
15593 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
15595 if (Op1.getOpcode() == ISD::Constant &&
15596 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
15597 cast<ConstantSDNode>(Op1)->isNullValue()) &&
15598 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15600 // If the input is a setcc, then reuse the input setcc or use a new one with
15601 // the inverted condition.
15602 if (Op0.getOpcode() == X86ISD::SETCC) {
15603 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
15604 bool Invert = (CC == ISD::SETNE) ^
15605 cast<ConstantSDNode>(Op1)->isNullValue();
15609 CCode = X86::GetOppositeBranchCondition(CCode);
15610 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15611 DAG.getConstant(CCode, MVT::i8),
15612 Op0.getOperand(1));
15614 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
15618 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
15619 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
15620 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
15622 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
15623 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
15626 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
15627 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
15628 if (X86CC == X86::COND_INVALID)
15631 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
15632 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
15633 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15634 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
15636 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
15640 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
15641 static bool isX86LogicalCmp(SDValue Op) {
15642 unsigned Opc = Op.getNode()->getOpcode();
15643 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
15644 Opc == X86ISD::SAHF)
15646 if (Op.getResNo() == 1 &&
15647 (Opc == X86ISD::ADD ||
15648 Opc == X86ISD::SUB ||
15649 Opc == X86ISD::ADC ||
15650 Opc == X86ISD::SBB ||
15651 Opc == X86ISD::SMUL ||
15652 Opc == X86ISD::UMUL ||
15653 Opc == X86ISD::INC ||
15654 Opc == X86ISD::DEC ||
15655 Opc == X86ISD::OR ||
15656 Opc == X86ISD::XOR ||
15657 Opc == X86ISD::AND))
15660 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
15666 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
15667 if (V.getOpcode() != ISD::TRUNCATE)
15670 SDValue VOp0 = V.getOperand(0);
15671 unsigned InBits = VOp0.getValueSizeInBits();
15672 unsigned Bits = V.getValueSizeInBits();
15673 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
15676 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
15677 bool addTest = true;
15678 SDValue Cond = Op.getOperand(0);
15679 SDValue Op1 = Op.getOperand(1);
15680 SDValue Op2 = Op.getOperand(2);
15682 EVT VT = Op1.getValueType();
15685 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
15686 // are available. Otherwise fp cmovs get lowered into a less efficient branch
15687 // sequence later on.
15688 if (Cond.getOpcode() == ISD::SETCC &&
15689 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
15690 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
15691 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
15692 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
15693 int SSECC = translateX86FSETCC(
15694 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
15697 if (Subtarget->hasAVX512()) {
15698 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
15699 DAG.getConstant(SSECC, MVT::i8));
15700 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
15702 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
15703 DAG.getConstant(SSECC, MVT::i8));
15704 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
15705 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
15706 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
15710 if (Cond.getOpcode() == ISD::SETCC) {
15711 SDValue NewCond = LowerSETCC(Cond, DAG);
15712 if (NewCond.getNode())
15716 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
15717 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
15718 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
15719 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
15720 if (Cond.getOpcode() == X86ISD::SETCC &&
15721 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
15722 isZero(Cond.getOperand(1).getOperand(1))) {
15723 SDValue Cmp = Cond.getOperand(1);
15725 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
15727 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
15728 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
15729 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
15731 SDValue CmpOp0 = Cmp.getOperand(0);
15732 // Apply further optimizations for special cases
15733 // (select (x != 0), -1, 0) -> neg & sbb
15734 // (select (x == 0), 0, -1) -> neg & sbb
15735 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
15736 if (YC->isNullValue() &&
15737 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
15738 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
15739 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
15740 DAG.getConstant(0, CmpOp0.getValueType()),
15742 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15743 DAG.getConstant(X86::COND_B, MVT::i8),
15744 SDValue(Neg.getNode(), 1));
15748 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
15749 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
15750 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15752 SDValue Res = // Res = 0 or -1.
15753 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15754 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
15756 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
15757 Res = DAG.getNOT(DL, Res, Res.getValueType());
15759 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
15760 if (!N2C || !N2C->isNullValue())
15761 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
15766 // Look past (and (setcc_carry (cmp ...)), 1).
15767 if (Cond.getOpcode() == ISD::AND &&
15768 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15769 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15770 if (C && C->getAPIntValue() == 1)
15771 Cond = Cond.getOperand(0);
15774 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15775 // setting operand in place of the X86ISD::SETCC.
15776 unsigned CondOpcode = Cond.getOpcode();
15777 if (CondOpcode == X86ISD::SETCC ||
15778 CondOpcode == X86ISD::SETCC_CARRY) {
15779 CC = Cond.getOperand(0);
15781 SDValue Cmp = Cond.getOperand(1);
15782 unsigned Opc = Cmp.getOpcode();
15783 MVT VT = Op.getSimpleValueType();
15785 bool IllegalFPCMov = false;
15786 if (VT.isFloatingPoint() && !VT.isVector() &&
15787 !isScalarFPTypeInSSEReg(VT)) // FPStack?
15788 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
15790 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
15791 Opc == X86ISD::BT) { // FIXME
15795 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15796 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15797 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15798 Cond.getOperand(0).getValueType() != MVT::i8)) {
15799 SDValue LHS = Cond.getOperand(0);
15800 SDValue RHS = Cond.getOperand(1);
15801 unsigned X86Opcode;
15804 switch (CondOpcode) {
15805 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15806 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15807 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15808 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15809 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15810 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15811 default: llvm_unreachable("unexpected overflowing operator");
15813 if (CondOpcode == ISD::UMULO)
15814 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15817 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15819 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
15821 if (CondOpcode == ISD::UMULO)
15822 Cond = X86Op.getValue(2);
15824 Cond = X86Op.getValue(1);
15826 CC = DAG.getConstant(X86Cond, MVT::i8);
15831 // Look pass the truncate if the high bits are known zero.
15832 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15833 Cond = Cond.getOperand(0);
15835 // We know the result of AND is compared against zero. Try to match
15837 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15838 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
15839 if (NewSetCC.getNode()) {
15840 CC = NewSetCC.getOperand(0);
15841 Cond = NewSetCC.getOperand(1);
15848 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
15849 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
15852 // a < b ? -1 : 0 -> RES = ~setcc_carry
15853 // a < b ? 0 : -1 -> RES = setcc_carry
15854 // a >= b ? -1 : 0 -> RES = setcc_carry
15855 // a >= b ? 0 : -1 -> RES = ~setcc_carry
15856 if (Cond.getOpcode() == X86ISD::SUB) {
15857 Cond = ConvertCmpIfNecessary(Cond, DAG);
15858 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
15860 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
15861 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
15862 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
15863 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
15864 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
15865 return DAG.getNOT(DL, Res, Res.getValueType());
15870 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
15871 // widen the cmov and push the truncate through. This avoids introducing a new
15872 // branch during isel and doesn't add any extensions.
15873 if (Op.getValueType() == MVT::i8 &&
15874 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
15875 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
15876 if (T1.getValueType() == T2.getValueType() &&
15877 // Blacklist CopyFromReg to avoid partial register stalls.
15878 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
15879 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
15880 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
15881 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
15885 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
15886 // condition is true.
15887 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
15888 SDValue Ops[] = { Op2, Op1, CC, Cond };
15889 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
15892 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
15893 SelectionDAG &DAG) {
15894 MVT VT = Op->getSimpleValueType(0);
15895 SDValue In = Op->getOperand(0);
15896 MVT InVT = In.getSimpleValueType();
15897 MVT VTElt = VT.getVectorElementType();
15898 MVT InVTElt = InVT.getVectorElementType();
15902 if ((InVTElt == MVT::i1) &&
15903 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
15904 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
15906 ((Subtarget->hasBWI() && VT.is512BitVector() &&
15907 VTElt.getSizeInBits() <= 16)) ||
15909 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
15910 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
15912 ((Subtarget->hasDQI() && VT.is512BitVector() &&
15913 VTElt.getSizeInBits() >= 32))))
15914 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15916 unsigned int NumElts = VT.getVectorNumElements();
15918 if (NumElts != 8 && NumElts != 16)
15921 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
15922 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
15923 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
15924 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15927 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15928 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
15930 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
15931 Constant *C = ConstantInt::get(*DAG.getContext(),
15932 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
15934 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
15935 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
15936 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
15937 MachinePointerInfo::getConstantPool(),
15938 false, false, false, Alignment);
15939 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
15940 if (VT.is512BitVector())
15942 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
15945 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15946 SelectionDAG &DAG) {
15947 MVT VT = Op->getSimpleValueType(0);
15948 SDValue In = Op->getOperand(0);
15949 MVT InVT = In.getSimpleValueType();
15952 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15953 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15955 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15956 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15957 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15960 if (Subtarget->hasInt256())
15961 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15963 // Optimize vectors in AVX mode
15964 // Sign extend v8i16 to v8i32 and
15967 // Divide input vector into two parts
15968 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15969 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15970 // concat the vectors to original VT
15972 unsigned NumElems = InVT.getVectorNumElements();
15973 SDValue Undef = DAG.getUNDEF(InVT);
15975 SmallVector<int,8> ShufMask1(NumElems, -1);
15976 for (unsigned i = 0; i != NumElems/2; ++i)
15979 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15981 SmallVector<int,8> ShufMask2(NumElems, -1);
15982 for (unsigned i = 0; i != NumElems/2; ++i)
15983 ShufMask2[i] = i + NumElems/2;
15985 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15987 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
15988 VT.getVectorNumElements()/2);
15990 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15991 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15993 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15996 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15997 // may emit an illegal shuffle but the expansion is still better than scalar
15998 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15999 // we'll emit a shuffle and a arithmetic shift.
16000 // TODO: It is possible to support ZExt by zeroing the undef values during
16001 // the shuffle phase or after the shuffle.
16002 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
16003 SelectionDAG &DAG) {
16004 MVT RegVT = Op.getSimpleValueType();
16005 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
16006 assert(RegVT.isInteger() &&
16007 "We only custom lower integer vector sext loads.");
16009 // Nothing useful we can do without SSE2 shuffles.
16010 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
16012 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
16014 EVT MemVT = Ld->getMemoryVT();
16015 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16016 unsigned RegSz = RegVT.getSizeInBits();
16018 ISD::LoadExtType Ext = Ld->getExtensionType();
16020 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
16021 && "Only anyext and sext are currently implemented.");
16022 assert(MemVT != RegVT && "Cannot extend to the same type");
16023 assert(MemVT.isVector() && "Must load a vector from memory");
16025 unsigned NumElems = RegVT.getVectorNumElements();
16026 unsigned MemSz = MemVT.getSizeInBits();
16027 assert(RegSz > MemSz && "Register size must be greater than the mem size");
16029 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
16030 // The only way in which we have a legal 256-bit vector result but not the
16031 // integer 256-bit operations needed to directly lower a sextload is if we
16032 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
16033 // a 128-bit vector and a normal sign_extend to 256-bits that should get
16034 // correctly legalized. We do this late to allow the canonical form of
16035 // sextload to persist throughout the rest of the DAG combiner -- it wants
16036 // to fold together any extensions it can, and so will fuse a sign_extend
16037 // of an sextload into a sextload targeting a wider value.
16039 if (MemSz == 128) {
16040 // Just switch this to a normal load.
16041 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
16042 "it must be a legal 128-bit vector "
16044 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
16045 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
16046 Ld->isInvariant(), Ld->getAlignment());
16048 assert(MemSz < 128 &&
16049 "Can't extend a type wider than 128 bits to a 256 bit vector!");
16050 // Do an sext load to a 128-bit vector type. We want to use the same
16051 // number of elements, but elements half as wide. This will end up being
16052 // recursively lowered by this routine, but will succeed as we definitely
16053 // have all the necessary features if we're using AVX1.
16055 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
16056 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
16058 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
16059 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
16060 Ld->isNonTemporal(), Ld->isInvariant(),
16061 Ld->getAlignment());
16064 // Replace chain users with the new chain.
16065 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
16066 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
16068 // Finally, do a normal sign-extend to the desired register.
16069 return DAG.getSExtOrTrunc(Load, dl, RegVT);
16072 // All sizes must be a power of two.
16073 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
16074 "Non-power-of-two elements are not custom lowered!");
16076 // Attempt to load the original value using scalar loads.
16077 // Find the largest scalar type that divides the total loaded size.
16078 MVT SclrLoadTy = MVT::i8;
16079 for (MVT Tp : MVT::integer_valuetypes()) {
16080 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
16085 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
16086 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
16088 SclrLoadTy = MVT::f64;
16090 // Calculate the number of scalar loads that we need to perform
16091 // in order to load our vector from memory.
16092 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
16094 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
16095 "Can only lower sext loads with a single scalar load!");
16097 unsigned loadRegZize = RegSz;
16098 if (Ext == ISD::SEXTLOAD && RegSz == 256)
16101 // Represent our vector as a sequence of elements which are the
16102 // largest scalar that we can load.
16103 EVT LoadUnitVecVT = EVT::getVectorVT(
16104 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
16106 // Represent the data using the same element type that is stored in
16107 // memory. In practice, we ''widen'' MemVT.
16109 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
16110 loadRegZize / MemVT.getScalarType().getSizeInBits());
16112 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
16113 "Invalid vector type");
16115 // We can't shuffle using an illegal type.
16116 assert(TLI.isTypeLegal(WideVecVT) &&
16117 "We only lower types that form legal widened vector types");
16119 SmallVector<SDValue, 8> Chains;
16120 SDValue Ptr = Ld->getBasePtr();
16121 SDValue Increment =
16122 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
16123 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
16125 for (unsigned i = 0; i < NumLoads; ++i) {
16126 // Perform a single load.
16127 SDValue ScalarLoad =
16128 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
16129 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
16130 Ld->getAlignment());
16131 Chains.push_back(ScalarLoad.getValue(1));
16132 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
16133 // another round of DAGCombining.
16135 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
16137 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
16138 ScalarLoad, DAG.getIntPtrConstant(i));
16140 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
16143 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
16145 // Bitcast the loaded value to a vector of the original element type, in
16146 // the size of the target vector type.
16147 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
16148 unsigned SizeRatio = RegSz / MemSz;
16150 if (Ext == ISD::SEXTLOAD) {
16151 // If we have SSE4.1, we can directly emit a VSEXT node.
16152 if (Subtarget->hasSSE41()) {
16153 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
16154 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16158 // Otherwise we'll shuffle the small elements in the high bits of the
16159 // larger type and perform an arithmetic shift. If the shift is not legal
16160 // it's better to scalarize.
16161 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
16162 "We can't implement a sext load without an arithmetic right shift!");
16164 // Redistribute the loaded elements into the different locations.
16165 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16166 for (unsigned i = 0; i != NumElems; ++i)
16167 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
16169 SDValue Shuff = DAG.getVectorShuffle(
16170 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16172 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16174 // Build the arithmetic shift.
16175 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
16176 MemVT.getVectorElementType().getSizeInBits();
16178 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
16180 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16184 // Redistribute the loaded elements into the different locations.
16185 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
16186 for (unsigned i = 0; i != NumElems; ++i)
16187 ShuffleVec[i * SizeRatio] = i;
16189 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
16190 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
16192 // Bitcast to the requested type.
16193 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
16194 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
16198 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
16199 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
16200 // from the AND / OR.
16201 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
16202 Opc = Op.getOpcode();
16203 if (Opc != ISD::OR && Opc != ISD::AND)
16205 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16206 Op.getOperand(0).hasOneUse() &&
16207 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
16208 Op.getOperand(1).hasOneUse());
16211 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
16212 // 1 and that the SETCC node has a single use.
16213 static bool isXor1OfSetCC(SDValue Op) {
16214 if (Op.getOpcode() != ISD::XOR)
16216 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
16217 if (N1C && N1C->getAPIntValue() == 1) {
16218 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
16219 Op.getOperand(0).hasOneUse();
16224 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
16225 bool addTest = true;
16226 SDValue Chain = Op.getOperand(0);
16227 SDValue Cond = Op.getOperand(1);
16228 SDValue Dest = Op.getOperand(2);
16231 bool Inverted = false;
16233 if (Cond.getOpcode() == ISD::SETCC) {
16234 // Check for setcc([su]{add,sub,mul}o == 0).
16235 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
16236 isa<ConstantSDNode>(Cond.getOperand(1)) &&
16237 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
16238 Cond.getOperand(0).getResNo() == 1 &&
16239 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
16240 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
16241 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
16242 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
16243 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
16244 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
16246 Cond = Cond.getOperand(0);
16248 SDValue NewCond = LowerSETCC(Cond, DAG);
16249 if (NewCond.getNode())
16254 // FIXME: LowerXALUO doesn't handle these!!
16255 else if (Cond.getOpcode() == X86ISD::ADD ||
16256 Cond.getOpcode() == X86ISD::SUB ||
16257 Cond.getOpcode() == X86ISD::SMUL ||
16258 Cond.getOpcode() == X86ISD::UMUL)
16259 Cond = LowerXALUO(Cond, DAG);
16262 // Look pass (and (setcc_carry (cmp ...)), 1).
16263 if (Cond.getOpcode() == ISD::AND &&
16264 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
16265 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
16266 if (C && C->getAPIntValue() == 1)
16267 Cond = Cond.getOperand(0);
16270 // If condition flag is set by a X86ISD::CMP, then use it as the condition
16271 // setting operand in place of the X86ISD::SETCC.
16272 unsigned CondOpcode = Cond.getOpcode();
16273 if (CondOpcode == X86ISD::SETCC ||
16274 CondOpcode == X86ISD::SETCC_CARRY) {
16275 CC = Cond.getOperand(0);
16277 SDValue Cmp = Cond.getOperand(1);
16278 unsigned Opc = Cmp.getOpcode();
16279 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
16280 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
16284 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
16288 // These can only come from an arithmetic instruction with overflow,
16289 // e.g. SADDO, UADDO.
16290 Cond = Cond.getNode()->getOperand(1);
16296 CondOpcode = Cond.getOpcode();
16297 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
16298 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
16299 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
16300 Cond.getOperand(0).getValueType() != MVT::i8)) {
16301 SDValue LHS = Cond.getOperand(0);
16302 SDValue RHS = Cond.getOperand(1);
16303 unsigned X86Opcode;
16306 // Keep this in sync with LowerXALUO, otherwise we might create redundant
16307 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
16309 switch (CondOpcode) {
16310 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
16312 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16314 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
16317 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
16318 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
16320 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16322 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
16325 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
16326 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
16327 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
16328 default: llvm_unreachable("unexpected overflowing operator");
16331 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
16332 if (CondOpcode == ISD::UMULO)
16333 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
16336 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
16338 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
16340 if (CondOpcode == ISD::UMULO)
16341 Cond = X86Op.getValue(2);
16343 Cond = X86Op.getValue(1);
16345 CC = DAG.getConstant(X86Cond, MVT::i8);
16349 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
16350 SDValue Cmp = Cond.getOperand(0).getOperand(1);
16351 if (CondOpc == ISD::OR) {
16352 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
16353 // two branches instead of an explicit OR instruction with a
16355 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16356 isX86LogicalCmp(Cmp)) {
16357 CC = Cond.getOperand(0).getOperand(0);
16358 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16359 Chain, Dest, CC, Cmp);
16360 CC = Cond.getOperand(1).getOperand(0);
16364 } else { // ISD::AND
16365 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
16366 // two branches instead of an explicit AND instruction with a
16367 // separate test. However, we only do this if this block doesn't
16368 // have a fall-through edge, because this requires an explicit
16369 // jmp when the condition is false.
16370 if (Cmp == Cond.getOperand(1).getOperand(1) &&
16371 isX86LogicalCmp(Cmp) &&
16372 Op.getNode()->hasOneUse()) {
16373 X86::CondCode CCode =
16374 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16375 CCode = X86::GetOppositeBranchCondition(CCode);
16376 CC = DAG.getConstant(CCode, MVT::i8);
16377 SDNode *User = *Op.getNode()->use_begin();
16378 // Look for an unconditional branch following this conditional branch.
16379 // We need this because we need to reverse the successors in order
16380 // to implement FCMP_OEQ.
16381 if (User->getOpcode() == ISD::BR) {
16382 SDValue FalseBB = User->getOperand(1);
16384 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16385 assert(NewBR == User);
16389 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16390 Chain, Dest, CC, Cmp);
16391 X86::CondCode CCode =
16392 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
16393 CCode = X86::GetOppositeBranchCondition(CCode);
16394 CC = DAG.getConstant(CCode, MVT::i8);
16400 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
16401 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
16402 // It should be transformed during dag combiner except when the condition
16403 // is set by a arithmetics with overflow node.
16404 X86::CondCode CCode =
16405 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
16406 CCode = X86::GetOppositeBranchCondition(CCode);
16407 CC = DAG.getConstant(CCode, MVT::i8);
16408 Cond = Cond.getOperand(0).getOperand(1);
16410 } else if (Cond.getOpcode() == ISD::SETCC &&
16411 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
16412 // For FCMP_OEQ, we can emit
16413 // two branches instead of an explicit AND instruction with a
16414 // separate test. However, we only do this if this block doesn't
16415 // have a fall-through edge, because this requires an explicit
16416 // jmp when the condition is false.
16417 if (Op.getNode()->hasOneUse()) {
16418 SDNode *User = *Op.getNode()->use_begin();
16419 // Look for an unconditional branch following this conditional branch.
16420 // We need this because we need to reverse the successors in order
16421 // to implement FCMP_OEQ.
16422 if (User->getOpcode() == ISD::BR) {
16423 SDValue FalseBB = User->getOperand(1);
16425 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16426 assert(NewBR == User);
16430 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
16431 Cond.getOperand(0), Cond.getOperand(1));
16432 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16433 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16434 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16435 Chain, Dest, CC, Cmp);
16436 CC = DAG.getConstant(X86::COND_P, MVT::i8);
16441 } else if (Cond.getOpcode() == ISD::SETCC &&
16442 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
16443 // For FCMP_UNE, we can emit
16444 // two branches instead of an explicit AND instruction with a
16445 // separate test. However, we only do this if this block doesn't
16446 // have a fall-through edge, because this requires an explicit
16447 // jmp when the condition is false.
16448 if (Op.getNode()->hasOneUse()) {
16449 SDNode *User = *Op.getNode()->use_begin();
16450 // Look for an unconditional branch following this conditional branch.
16451 // We need this because we need to reverse the successors in order
16452 // to implement FCMP_UNE.
16453 if (User->getOpcode() == ISD::BR) {
16454 SDValue FalseBB = User->getOperand(1);
16456 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
16457 assert(NewBR == User);
16460 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
16461 Cond.getOperand(0), Cond.getOperand(1));
16462 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
16463 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
16464 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16465 Chain, Dest, CC, Cmp);
16466 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
16476 // Look pass the truncate if the high bits are known zero.
16477 if (isTruncWithZeroHighBitsInput(Cond, DAG))
16478 Cond = Cond.getOperand(0);
16480 // We know the result of AND is compared against zero. Try to match
16482 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
16483 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
16484 if (NewSetCC.getNode()) {
16485 CC = NewSetCC.getOperand(0);
16486 Cond = NewSetCC.getOperand(1);
16493 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
16494 CC = DAG.getConstant(X86Cond, MVT::i8);
16495 Cond = EmitTest(Cond, X86Cond, dl, DAG);
16497 Cond = ConvertCmpIfNecessary(Cond, DAG);
16498 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
16499 Chain, Dest, CC, Cond);
16502 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
16503 // Calls to _alloca are needed to probe the stack when allocating more than 4k
16504 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
16505 // that the guard pages used by the OS virtual memory manager are allocated in
16506 // correct sequence.
16508 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
16509 SelectionDAG &DAG) const {
16510 MachineFunction &MF = DAG.getMachineFunction();
16511 bool SplitStack = MF.shouldSplitStack();
16512 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
16517 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16518 SDNode* Node = Op.getNode();
16520 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
16521 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
16522 " not tell us which reg is the stack pointer!");
16523 EVT VT = Node->getValueType(0);
16524 SDValue Tmp1 = SDValue(Node, 0);
16525 SDValue Tmp2 = SDValue(Node, 1);
16526 SDValue Tmp3 = Node->getOperand(2);
16527 SDValue Chain = Tmp1.getOperand(0);
16529 // Chain the dynamic stack allocation so that it doesn't modify the stack
16530 // pointer when other instructions are using the stack.
16531 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
16534 SDValue Size = Tmp2.getOperand(1);
16535 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
16536 Chain = SP.getValue(1);
16537 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
16538 const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering();
16539 unsigned StackAlign = TFI.getStackAlignment();
16540 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
16541 if (Align > StackAlign)
16542 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
16543 DAG.getConstant(-(uint64_t)Align, VT));
16544 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
16546 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
16547 DAG.getIntPtrConstant(0, true), SDValue(),
16550 SDValue Ops[2] = { Tmp1, Tmp2 };
16551 return DAG.getMergeValues(Ops, dl);
16555 SDValue Chain = Op.getOperand(0);
16556 SDValue Size = Op.getOperand(1);
16557 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
16558 EVT VT = Op.getNode()->getValueType(0);
16560 bool Is64Bit = Subtarget->is64Bit();
16561 EVT SPTy = getPointerTy();
16564 MachineRegisterInfo &MRI = MF.getRegInfo();
16567 // The 64 bit implementation of segmented stacks needs to clobber both r10
16568 // r11. This makes it impossible to use it along with nested parameters.
16569 const Function *F = MF.getFunction();
16571 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
16573 if (I->hasNestAttr())
16574 report_fatal_error("Cannot use segmented stacks with functions that "
16575 "have nested arguments.");
16578 const TargetRegisterClass *AddrRegClass =
16579 getRegClassFor(getPointerTy());
16580 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
16581 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
16582 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
16583 DAG.getRegister(Vreg, SPTy));
16584 SDValue Ops1[2] = { Value, Chain };
16585 return DAG.getMergeValues(Ops1, dl);
16588 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
16590 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
16591 Flag = Chain.getValue(1);
16592 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
16594 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
16596 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
16597 DAG.getSubtarget().getRegisterInfo());
16598 unsigned SPReg = RegInfo->getStackRegister();
16599 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
16600 Chain = SP.getValue(1);
16603 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
16604 DAG.getConstant(-(uint64_t)Align, VT));
16605 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
16608 SDValue Ops1[2] = { SP, Chain };
16609 return DAG.getMergeValues(Ops1, dl);
16613 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
16614 MachineFunction &MF = DAG.getMachineFunction();
16615 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16617 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16620 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
16621 // vastart just stores the address of the VarArgsFrameIndex slot into the
16622 // memory location argument.
16623 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
16625 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
16626 MachinePointerInfo(SV), false, false, 0);
16630 // gp_offset (0 - 6 * 8)
16631 // fp_offset (48 - 48 + 8 * 16)
16632 // overflow_arg_area (point to parameters coming in memory).
16634 SmallVector<SDValue, 8> MemOps;
16635 SDValue FIN = Op.getOperand(1);
16637 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
16638 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
16640 FIN, MachinePointerInfo(SV), false, false, 0);
16641 MemOps.push_back(Store);
16644 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16645 FIN, DAG.getIntPtrConstant(4));
16646 Store = DAG.getStore(Op.getOperand(0), DL,
16647 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
16649 FIN, MachinePointerInfo(SV, 4), false, false, 0);
16650 MemOps.push_back(Store);
16652 // Store ptr to overflow_arg_area
16653 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16654 FIN, DAG.getIntPtrConstant(4));
16655 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
16657 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
16658 MachinePointerInfo(SV, 8),
16660 MemOps.push_back(Store);
16662 // Store ptr to reg_save_area.
16663 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
16664 FIN, DAG.getIntPtrConstant(8));
16665 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
16667 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
16668 MachinePointerInfo(SV, 16), false, false, 0);
16669 MemOps.push_back(Store);
16670 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
16673 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
16674 assert(Subtarget->is64Bit() &&
16675 "LowerVAARG only handles 64-bit va_arg!");
16676 assert((Subtarget->isTargetLinux() ||
16677 Subtarget->isTargetDarwin()) &&
16678 "Unhandled target in LowerVAARG");
16679 assert(Op.getNode()->getNumOperands() == 4);
16680 SDValue Chain = Op.getOperand(0);
16681 SDValue SrcPtr = Op.getOperand(1);
16682 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
16683 unsigned Align = Op.getConstantOperandVal(3);
16686 EVT ArgVT = Op.getNode()->getValueType(0);
16687 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16688 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
16691 // Decide which area this value should be read from.
16692 // TODO: Implement the AMD64 ABI in its entirety. This simple
16693 // selection mechanism works only for the basic types.
16694 if (ArgVT == MVT::f80) {
16695 llvm_unreachable("va_arg for f80 not yet implemented");
16696 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
16697 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
16698 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
16699 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
16701 llvm_unreachable("Unhandled argument type in LowerVAARG");
16704 if (ArgMode == 2) {
16705 // Sanity Check: Make sure using fp_offset makes sense.
16706 assert(!DAG.getTarget().Options.UseSoftFloat &&
16707 !(DAG.getMachineFunction()
16708 .getFunction()->getAttributes()
16709 .hasAttribute(AttributeSet::FunctionIndex,
16710 Attribute::NoImplicitFloat)) &&
16711 Subtarget->hasSSE1());
16714 // Insert VAARG_64 node into the DAG
16715 // VAARG_64 returns two values: Variable Argument Address, Chain
16716 SmallVector<SDValue, 11> InstOps;
16717 InstOps.push_back(Chain);
16718 InstOps.push_back(SrcPtr);
16719 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
16720 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
16721 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
16722 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
16723 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
16724 VTs, InstOps, MVT::i64,
16725 MachinePointerInfo(SV),
16727 /*Volatile=*/false,
16729 /*WriteMem=*/true);
16730 Chain = VAARG.getValue(1);
16732 // Load the next argument and return it
16733 return DAG.getLoad(ArgVT, dl,
16736 MachinePointerInfo(),
16737 false, false, false, 0);
16740 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
16741 SelectionDAG &DAG) {
16742 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
16743 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
16744 SDValue Chain = Op.getOperand(0);
16745 SDValue DstPtr = Op.getOperand(1);
16746 SDValue SrcPtr = Op.getOperand(2);
16747 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
16748 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16751 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
16752 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
16754 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
16757 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
16758 // amount is a constant. Takes immediate version of shift as input.
16759 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
16760 SDValue SrcOp, uint64_t ShiftAmt,
16761 SelectionDAG &DAG) {
16762 MVT ElementType = VT.getVectorElementType();
16764 // Fold this packed shift into its first operand if ShiftAmt is 0.
16768 // Check for ShiftAmt >= element width
16769 if (ShiftAmt >= ElementType.getSizeInBits()) {
16770 if (Opc == X86ISD::VSRAI)
16771 ShiftAmt = ElementType.getSizeInBits() - 1;
16773 return DAG.getConstant(0, VT);
16776 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
16777 && "Unknown target vector shift-by-constant node");
16779 // Fold this packed vector shift into a build vector if SrcOp is a
16780 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
16781 if (VT == SrcOp.getSimpleValueType() &&
16782 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
16783 SmallVector<SDValue, 8> Elts;
16784 unsigned NumElts = SrcOp->getNumOperands();
16785 ConstantSDNode *ND;
16788 default: llvm_unreachable(nullptr);
16789 case X86ISD::VSHLI:
16790 for (unsigned i=0; i!=NumElts; ++i) {
16791 SDValue CurrentOp = SrcOp->getOperand(i);
16792 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16793 Elts.push_back(CurrentOp);
16796 ND = cast<ConstantSDNode>(CurrentOp);
16797 const APInt &C = ND->getAPIntValue();
16798 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
16801 case X86ISD::VSRLI:
16802 for (unsigned i=0; i!=NumElts; ++i) {
16803 SDValue CurrentOp = SrcOp->getOperand(i);
16804 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16805 Elts.push_back(CurrentOp);
16808 ND = cast<ConstantSDNode>(CurrentOp);
16809 const APInt &C = ND->getAPIntValue();
16810 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
16813 case X86ISD::VSRAI:
16814 for (unsigned i=0; i!=NumElts; ++i) {
16815 SDValue CurrentOp = SrcOp->getOperand(i);
16816 if (CurrentOp->getOpcode() == ISD::UNDEF) {
16817 Elts.push_back(CurrentOp);
16820 ND = cast<ConstantSDNode>(CurrentOp);
16821 const APInt &C = ND->getAPIntValue();
16822 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
16827 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16830 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
16833 // getTargetVShiftNode - Handle vector element shifts where the shift amount
16834 // may or may not be a constant. Takes immediate version of shift as input.
16835 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
16836 SDValue SrcOp, SDValue ShAmt,
16837 SelectionDAG &DAG) {
16838 MVT SVT = ShAmt.getSimpleValueType();
16839 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
16841 // Catch shift-by-constant.
16842 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
16843 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
16844 CShAmt->getZExtValue(), DAG);
16846 // Change opcode to non-immediate version
16848 default: llvm_unreachable("Unknown target vector shift node");
16849 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
16850 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
16851 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
16854 const X86Subtarget &Subtarget =
16855 DAG.getTarget().getSubtarget<X86Subtarget>();
16856 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
16857 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
16858 // Let the shuffle legalizer expand this shift amount node.
16859 SDValue Op0 = ShAmt.getOperand(0);
16860 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
16861 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
16863 // Need to build a vector containing shift amount.
16864 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
16865 SmallVector<SDValue, 4> ShOps;
16866 ShOps.push_back(ShAmt);
16867 if (SVT == MVT::i32) {
16868 ShOps.push_back(DAG.getConstant(0, SVT));
16869 ShOps.push_back(DAG.getUNDEF(SVT));
16871 ShOps.push_back(DAG.getUNDEF(SVT));
16873 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
16874 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
16877 // The return type has to be a 128-bit type with the same element
16878 // type as the input type.
16879 MVT EltVT = VT.getVectorElementType();
16880 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
16882 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
16883 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
16886 /// \brief Return (and \p Op, \p Mask) for compare instructions or
16887 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
16888 /// necessary casting for \p Mask when lowering masking intrinsics.
16889 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
16890 SDValue PreservedSrc,
16891 const X86Subtarget *Subtarget,
16892 SelectionDAG &DAG) {
16893 EVT VT = Op.getValueType();
16894 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16895 MVT::i1, VT.getVectorNumElements());
16896 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16897 Mask.getValueType().getSizeInBits());
16900 assert(MaskVT.isSimple() && "invalid mask type");
16902 if (isAllOnes(Mask))
16905 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16906 // are extracted by EXTRACT_SUBVECTOR.
16907 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16908 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
16909 DAG.getIntPtrConstant(0));
16911 switch (Op.getOpcode()) {
16913 case X86ISD::PCMPEQM:
16914 case X86ISD::PCMPGTM:
16916 case X86ISD::CMPMU:
16917 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16919 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16920 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16921 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
16924 /// \brief Creates an SDNode for a predicated scalar operation.
16925 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16926 /// The mask is comming as MVT::i8 and it should be truncated
16927 /// to MVT::i1 while lowering masking intrinsics.
16928 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16929 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
16930 /// a scalar instruction.
16931 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16932 SDValue PreservedSrc,
16933 const X86Subtarget *Subtarget,
16934 SelectionDAG &DAG) {
16935 if (isAllOnes(Mask))
16938 EVT VT = Op.getValueType();
16940 // The mask should be of type MVT::i1
16941 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16943 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16944 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16945 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16948 static unsigned getOpcodeForFMAIntrinsic(unsigned IntNo) {
16950 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16951 case Intrinsic::x86_fma_vfmadd_ps:
16952 case Intrinsic::x86_fma_vfmadd_pd:
16953 case Intrinsic::x86_fma_vfmadd_ps_256:
16954 case Intrinsic::x86_fma_vfmadd_pd_256:
16955 case Intrinsic::x86_fma_mask_vfmadd_ps_512:
16956 case Intrinsic::x86_fma_mask_vfmadd_pd_512:
16957 return X86ISD::FMADD;
16958 case Intrinsic::x86_fma_vfmsub_ps:
16959 case Intrinsic::x86_fma_vfmsub_pd:
16960 case Intrinsic::x86_fma_vfmsub_ps_256:
16961 case Intrinsic::x86_fma_vfmsub_pd_256:
16962 case Intrinsic::x86_fma_mask_vfmsub_ps_512:
16963 case Intrinsic::x86_fma_mask_vfmsub_pd_512:
16964 return X86ISD::FMSUB;
16965 case Intrinsic::x86_fma_vfnmadd_ps:
16966 case Intrinsic::x86_fma_vfnmadd_pd:
16967 case Intrinsic::x86_fma_vfnmadd_ps_256:
16968 case Intrinsic::x86_fma_vfnmadd_pd_256:
16969 case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
16970 case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
16971 return X86ISD::FNMADD;
16972 case Intrinsic::x86_fma_vfnmsub_ps:
16973 case Intrinsic::x86_fma_vfnmsub_pd:
16974 case Intrinsic::x86_fma_vfnmsub_ps_256:
16975 case Intrinsic::x86_fma_vfnmsub_pd_256:
16976 case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
16977 case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
16978 return X86ISD::FNMSUB;
16979 case Intrinsic::x86_fma_vfmaddsub_ps:
16980 case Intrinsic::x86_fma_vfmaddsub_pd:
16981 case Intrinsic::x86_fma_vfmaddsub_ps_256:
16982 case Intrinsic::x86_fma_vfmaddsub_pd_256:
16983 case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
16984 case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
16985 return X86ISD::FMADDSUB;
16986 case Intrinsic::x86_fma_vfmsubadd_ps:
16987 case Intrinsic::x86_fma_vfmsubadd_pd:
16988 case Intrinsic::x86_fma_vfmsubadd_ps_256:
16989 case Intrinsic::x86_fma_vfmsubadd_pd_256:
16990 case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
16991 case Intrinsic::x86_fma_mask_vfmsubadd_pd_512:
16992 return X86ISD::FMSUBADD;
16996 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16997 SelectionDAG &DAG) {
16999 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17000 EVT VT = Op.getValueType();
17001 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
17003 switch(IntrData->Type) {
17004 case INTR_TYPE_1OP:
17005 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
17006 case INTR_TYPE_2OP:
17007 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17009 case INTR_TYPE_3OP:
17010 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
17011 Op.getOperand(2), Op.getOperand(3));
17012 case INTR_TYPE_1OP_MASK_RM: {
17013 SDValue Src = Op.getOperand(1);
17014 SDValue Src0 = Op.getOperand(2);
17015 SDValue Mask = Op.getOperand(3);
17016 SDValue RoundingMode = Op.getOperand(4);
17017 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
17019 Mask, Src0, Subtarget, DAG);
17021 case INTR_TYPE_SCALAR_MASK_RM: {
17022 SDValue Src1 = Op.getOperand(1);
17023 SDValue Src2 = Op.getOperand(2);
17024 SDValue Src0 = Op.getOperand(3);
17025 SDValue Mask = Op.getOperand(4);
17026 SDValue RoundingMode = Op.getOperand(5);
17027 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
17029 Mask, Src0, Subtarget, DAG);
17031 case INTR_TYPE_2OP_MASK: {
17032 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Op.getOperand(1),
17034 Op.getOperand(4), Op.getOperand(3), Subtarget, DAG);
17037 case CMP_MASK_CC: {
17038 // Comparison intrinsics with masks.
17039 // Example of transformation:
17040 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
17041 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
17043 // (v8i1 (insert_subvector undef,
17044 // (v2i1 (and (PCMPEQM %a, %b),
17045 // (extract_subvector
17046 // (v8i1 (bitcast %mask)), 0))), 0))))
17047 EVT VT = Op.getOperand(1).getValueType();
17048 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17049 VT.getVectorNumElements());
17050 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
17051 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17052 Mask.getValueType().getSizeInBits());
17054 if (IntrData->Type == CMP_MASK_CC) {
17055 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17056 Op.getOperand(2), Op.getOperand(3));
17058 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
17059 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
17062 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
17063 DAG.getTargetConstant(0, MaskVT),
17065 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
17066 DAG.getUNDEF(BitcastVT), CmpMask,
17067 DAG.getIntPtrConstant(0));
17068 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
17070 case COMI: { // Comparison intrinsics
17071 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
17072 SDValue LHS = Op.getOperand(1);
17073 SDValue RHS = Op.getOperand(2);
17074 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
17075 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
17076 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
17077 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17078 DAG.getConstant(X86CC, MVT::i8), Cond);
17079 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17082 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
17083 Op.getOperand(1), Op.getOperand(2), DAG);
17085 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
17086 Op.getSimpleValueType(),
17088 Op.getOperand(2), DAG),
17089 Op.getOperand(4), Op.getOperand(3), Subtarget,
17091 case COMPRESS_EXPAND_IN_REG: {
17092 SDValue Mask = Op.getOperand(3);
17093 SDValue DataToCompress = Op.getOperand(1);
17094 SDValue PassThru = Op.getOperand(2);
17095 if (isAllOnes(Mask)) // return data as is
17096 return Op.getOperand(1);
17097 EVT VT = Op.getValueType();
17098 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17099 VT.getVectorNumElements());
17100 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17101 Mask.getValueType().getSizeInBits());
17103 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17104 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17105 DAG.getIntPtrConstant(0));
17107 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
17111 SDValue Mask = Op.getOperand(3);
17112 EVT VT = Op.getValueType();
17113 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17114 VT.getVectorNumElements());
17115 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17116 Mask.getValueType().getSizeInBits());
17118 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17119 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17120 DAG.getIntPtrConstant(0));
17121 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
17126 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
17127 dl, Op.getValueType(),
17131 Op.getOperand(4), Op.getOperand(1),
17140 default: return SDValue(); // Don't custom lower most intrinsics.
17142 case Intrinsic::x86_avx512_mask_valign_q_512:
17143 case Intrinsic::x86_avx512_mask_valign_d_512:
17144 // Vector source operands are swapped.
17145 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
17146 Op.getValueType(), Op.getOperand(2),
17149 Op.getOperand(5), Op.getOperand(4),
17152 // ptest and testp intrinsics. The intrinsic these come from are designed to
17153 // return an integer value, not just an instruction so lower it to the ptest
17154 // or testp pattern and a setcc for the result.
17155 case Intrinsic::x86_sse41_ptestz:
17156 case Intrinsic::x86_sse41_ptestc:
17157 case Intrinsic::x86_sse41_ptestnzc:
17158 case Intrinsic::x86_avx_ptestz_256:
17159 case Intrinsic::x86_avx_ptestc_256:
17160 case Intrinsic::x86_avx_ptestnzc_256:
17161 case Intrinsic::x86_avx_vtestz_ps:
17162 case Intrinsic::x86_avx_vtestc_ps:
17163 case Intrinsic::x86_avx_vtestnzc_ps:
17164 case Intrinsic::x86_avx_vtestz_pd:
17165 case Intrinsic::x86_avx_vtestc_pd:
17166 case Intrinsic::x86_avx_vtestnzc_pd:
17167 case Intrinsic::x86_avx_vtestz_ps_256:
17168 case Intrinsic::x86_avx_vtestc_ps_256:
17169 case Intrinsic::x86_avx_vtestnzc_ps_256:
17170 case Intrinsic::x86_avx_vtestz_pd_256:
17171 case Intrinsic::x86_avx_vtestc_pd_256:
17172 case Intrinsic::x86_avx_vtestnzc_pd_256: {
17173 bool IsTestPacked = false;
17176 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
17177 case Intrinsic::x86_avx_vtestz_ps:
17178 case Intrinsic::x86_avx_vtestz_pd:
17179 case Intrinsic::x86_avx_vtestz_ps_256:
17180 case Intrinsic::x86_avx_vtestz_pd_256:
17181 IsTestPacked = true; // Fallthrough
17182 case Intrinsic::x86_sse41_ptestz:
17183 case Intrinsic::x86_avx_ptestz_256:
17185 X86CC = X86::COND_E;
17187 case Intrinsic::x86_avx_vtestc_ps:
17188 case Intrinsic::x86_avx_vtestc_pd:
17189 case Intrinsic::x86_avx_vtestc_ps_256:
17190 case Intrinsic::x86_avx_vtestc_pd_256:
17191 IsTestPacked = true; // Fallthrough
17192 case Intrinsic::x86_sse41_ptestc:
17193 case Intrinsic::x86_avx_ptestc_256:
17195 X86CC = X86::COND_B;
17197 case Intrinsic::x86_avx_vtestnzc_ps:
17198 case Intrinsic::x86_avx_vtestnzc_pd:
17199 case Intrinsic::x86_avx_vtestnzc_ps_256:
17200 case Intrinsic::x86_avx_vtestnzc_pd_256:
17201 IsTestPacked = true; // Fallthrough
17202 case Intrinsic::x86_sse41_ptestnzc:
17203 case Intrinsic::x86_avx_ptestnzc_256:
17205 X86CC = X86::COND_A;
17209 SDValue LHS = Op.getOperand(1);
17210 SDValue RHS = Op.getOperand(2);
17211 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
17212 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
17213 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17214 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
17215 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17217 case Intrinsic::x86_avx512_kortestz_w:
17218 case Intrinsic::x86_avx512_kortestc_w: {
17219 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
17220 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
17221 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
17222 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
17223 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
17224 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
17225 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17228 case Intrinsic::x86_sse42_pcmpistria128:
17229 case Intrinsic::x86_sse42_pcmpestria128:
17230 case Intrinsic::x86_sse42_pcmpistric128:
17231 case Intrinsic::x86_sse42_pcmpestric128:
17232 case Intrinsic::x86_sse42_pcmpistrio128:
17233 case Intrinsic::x86_sse42_pcmpestrio128:
17234 case Intrinsic::x86_sse42_pcmpistris128:
17235 case Intrinsic::x86_sse42_pcmpestris128:
17236 case Intrinsic::x86_sse42_pcmpistriz128:
17237 case Intrinsic::x86_sse42_pcmpestriz128: {
17241 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
17242 case Intrinsic::x86_sse42_pcmpistria128:
17243 Opcode = X86ISD::PCMPISTRI;
17244 X86CC = X86::COND_A;
17246 case Intrinsic::x86_sse42_pcmpestria128:
17247 Opcode = X86ISD::PCMPESTRI;
17248 X86CC = X86::COND_A;
17250 case Intrinsic::x86_sse42_pcmpistric128:
17251 Opcode = X86ISD::PCMPISTRI;
17252 X86CC = X86::COND_B;
17254 case Intrinsic::x86_sse42_pcmpestric128:
17255 Opcode = X86ISD::PCMPESTRI;
17256 X86CC = X86::COND_B;
17258 case Intrinsic::x86_sse42_pcmpistrio128:
17259 Opcode = X86ISD::PCMPISTRI;
17260 X86CC = X86::COND_O;
17262 case Intrinsic::x86_sse42_pcmpestrio128:
17263 Opcode = X86ISD::PCMPESTRI;
17264 X86CC = X86::COND_O;
17266 case Intrinsic::x86_sse42_pcmpistris128:
17267 Opcode = X86ISD::PCMPISTRI;
17268 X86CC = X86::COND_S;
17270 case Intrinsic::x86_sse42_pcmpestris128:
17271 Opcode = X86ISD::PCMPESTRI;
17272 X86CC = X86::COND_S;
17274 case Intrinsic::x86_sse42_pcmpistriz128:
17275 Opcode = X86ISD::PCMPISTRI;
17276 X86CC = X86::COND_E;
17278 case Intrinsic::x86_sse42_pcmpestriz128:
17279 Opcode = X86ISD::PCMPESTRI;
17280 X86CC = X86::COND_E;
17283 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17284 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17285 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
17286 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17287 DAG.getConstant(X86CC, MVT::i8),
17288 SDValue(PCMP.getNode(), 1));
17289 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
17292 case Intrinsic::x86_sse42_pcmpistri128:
17293 case Intrinsic::x86_sse42_pcmpestri128: {
17295 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
17296 Opcode = X86ISD::PCMPISTRI;
17298 Opcode = X86ISD::PCMPESTRI;
17300 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
17301 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
17302 return DAG.getNode(Opcode, dl, VTs, NewOps);
17305 case Intrinsic::x86_fma_mask_vfmadd_ps_512:
17306 case Intrinsic::x86_fma_mask_vfmadd_pd_512:
17307 case Intrinsic::x86_fma_mask_vfmsub_ps_512:
17308 case Intrinsic::x86_fma_mask_vfmsub_pd_512:
17309 case Intrinsic::x86_fma_mask_vfnmadd_ps_512:
17310 case Intrinsic::x86_fma_mask_vfnmadd_pd_512:
17311 case Intrinsic::x86_fma_mask_vfnmsub_ps_512:
17312 case Intrinsic::x86_fma_mask_vfnmsub_pd_512:
17313 case Intrinsic::x86_fma_mask_vfmaddsub_ps_512:
17314 case Intrinsic::x86_fma_mask_vfmaddsub_pd_512:
17315 case Intrinsic::x86_fma_mask_vfmsubadd_ps_512:
17316 case Intrinsic::x86_fma_mask_vfmsubadd_pd_512: {
17317 auto *SAE = cast<ConstantSDNode>(Op.getOperand(5));
17318 if (SAE->getZExtValue() == X86::STATIC_ROUNDING::CUR_DIRECTION)
17319 return getVectorMaskingNode(DAG.getNode(getOpcodeForFMAIntrinsic(IntNo),
17320 dl, Op.getValueType(),
17324 Op.getOperand(4), Op.getOperand(1),
17330 case Intrinsic::x86_fma_vfmadd_ps:
17331 case Intrinsic::x86_fma_vfmadd_pd:
17332 case Intrinsic::x86_fma_vfmsub_ps:
17333 case Intrinsic::x86_fma_vfmsub_pd:
17334 case Intrinsic::x86_fma_vfnmadd_ps:
17335 case Intrinsic::x86_fma_vfnmadd_pd:
17336 case Intrinsic::x86_fma_vfnmsub_ps:
17337 case Intrinsic::x86_fma_vfnmsub_pd:
17338 case Intrinsic::x86_fma_vfmaddsub_ps:
17339 case Intrinsic::x86_fma_vfmaddsub_pd:
17340 case Intrinsic::x86_fma_vfmsubadd_ps:
17341 case Intrinsic::x86_fma_vfmsubadd_pd:
17342 case Intrinsic::x86_fma_vfmadd_ps_256:
17343 case Intrinsic::x86_fma_vfmadd_pd_256:
17344 case Intrinsic::x86_fma_vfmsub_ps_256:
17345 case Intrinsic::x86_fma_vfmsub_pd_256:
17346 case Intrinsic::x86_fma_vfnmadd_ps_256:
17347 case Intrinsic::x86_fma_vfnmadd_pd_256:
17348 case Intrinsic::x86_fma_vfnmsub_ps_256:
17349 case Intrinsic::x86_fma_vfnmsub_pd_256:
17350 case Intrinsic::x86_fma_vfmaddsub_ps_256:
17351 case Intrinsic::x86_fma_vfmaddsub_pd_256:
17352 case Intrinsic::x86_fma_vfmsubadd_ps_256:
17353 case Intrinsic::x86_fma_vfmsubadd_pd_256:
17354 return DAG.getNode(getOpcodeForFMAIntrinsic(IntNo), dl, Op.getValueType(),
17355 Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
17359 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17360 SDValue Src, SDValue Mask, SDValue Base,
17361 SDValue Index, SDValue ScaleOp, SDValue Chain,
17362 const X86Subtarget * Subtarget) {
17364 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17365 assert(C && "Invalid scale type");
17366 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17367 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17368 Index.getSimpleValueType().getVectorNumElements());
17370 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17372 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17374 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17375 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
17376 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17377 SDValue Segment = DAG.getRegister(0, MVT::i32);
17378 if (Src.getOpcode() == ISD::UNDEF)
17379 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
17380 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17381 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17382 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
17383 return DAG.getMergeValues(RetOps, dl);
17386 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17387 SDValue Src, SDValue Mask, SDValue Base,
17388 SDValue Index, SDValue ScaleOp, SDValue Chain) {
17390 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17391 assert(C && "Invalid scale type");
17392 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17393 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17394 SDValue Segment = DAG.getRegister(0, MVT::i32);
17395 EVT MaskVT = MVT::getVectorVT(MVT::i1,
17396 Index.getSimpleValueType().getVectorNumElements());
17398 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17400 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17402 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17403 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
17404 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
17405 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
17406 return SDValue(Res, 1);
17409 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
17410 SDValue Mask, SDValue Base, SDValue Index,
17411 SDValue ScaleOp, SDValue Chain) {
17413 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
17414 assert(C && "Invalid scale type");
17415 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
17416 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
17417 SDValue Segment = DAG.getRegister(0, MVT::i32);
17419 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
17421 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
17423 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
17425 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
17426 //SDVTList VTs = DAG.getVTList(MVT::Other);
17427 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
17428 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
17429 return SDValue(Res, 0);
17432 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
17433 // read performance monitor counters (x86_rdpmc).
17434 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
17435 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17436 SmallVectorImpl<SDValue> &Results) {
17437 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17438 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17441 // The ECX register is used to select the index of the performance counter
17443 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
17445 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
17447 // Reads the content of a 64-bit performance counter and returns it in the
17448 // registers EDX:EAX.
17449 if (Subtarget->is64Bit()) {
17450 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17451 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17454 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17455 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17458 Chain = HI.getValue(1);
17460 if (Subtarget->is64Bit()) {
17461 // The EAX register is loaded with the low-order 32 bits. The EDX register
17462 // is loaded with the supported high-order bits of the counter.
17463 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17464 DAG.getConstant(32, MVT::i8));
17465 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17466 Results.push_back(Chain);
17470 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17471 SDValue Ops[] = { LO, HI };
17472 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17473 Results.push_back(Pair);
17474 Results.push_back(Chain);
17477 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
17478 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
17479 // also used to custom lower READCYCLECOUNTER nodes.
17480 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
17481 SelectionDAG &DAG, const X86Subtarget *Subtarget,
17482 SmallVectorImpl<SDValue> &Results) {
17483 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17484 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
17487 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
17488 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
17489 // and the EAX register is loaded with the low-order 32 bits.
17490 if (Subtarget->is64Bit()) {
17491 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
17492 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
17495 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
17496 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
17499 SDValue Chain = HI.getValue(1);
17501 if (Opcode == X86ISD::RDTSCP_DAG) {
17502 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
17504 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
17505 // the ECX register. Add 'ecx' explicitly to the chain.
17506 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
17508 // Explicitly store the content of ECX at the location passed in input
17509 // to the 'rdtscp' intrinsic.
17510 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
17511 MachinePointerInfo(), false, false, 0);
17514 if (Subtarget->is64Bit()) {
17515 // The EDX register is loaded with the high-order 32 bits of the MSR, and
17516 // the EAX register is loaded with the low-order 32 bits.
17517 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
17518 DAG.getConstant(32, MVT::i8));
17519 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
17520 Results.push_back(Chain);
17524 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
17525 SDValue Ops[] = { LO, HI };
17526 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
17527 Results.push_back(Pair);
17528 Results.push_back(Chain);
17531 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
17532 SelectionDAG &DAG) {
17533 SmallVector<SDValue, 2> Results;
17535 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
17537 return DAG.getMergeValues(Results, DL);
17541 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17542 SelectionDAG &DAG) {
17543 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17545 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17550 switch(IntrData->Type) {
17552 llvm_unreachable("Unknown Intrinsic Type");
17556 // Emit the node with the right value type.
17557 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17558 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17560 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17561 // Otherwise return the value from Rand, which is always 0, casted to i32.
17562 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17563 DAG.getConstant(1, Op->getValueType(1)),
17564 DAG.getConstant(X86::COND_B, MVT::i32),
17565 SDValue(Result.getNode(), 1) };
17566 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17567 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17570 // Return { result, isValid, chain }.
17571 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17572 SDValue(Result.getNode(), 2));
17575 //gather(v1, mask, index, base, scale);
17576 SDValue Chain = Op.getOperand(0);
17577 SDValue Src = Op.getOperand(2);
17578 SDValue Base = Op.getOperand(3);
17579 SDValue Index = Op.getOperand(4);
17580 SDValue Mask = Op.getOperand(5);
17581 SDValue Scale = Op.getOperand(6);
17582 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
17586 //scatter(base, mask, index, v1, scale);
17587 SDValue Chain = Op.getOperand(0);
17588 SDValue Base = Op.getOperand(2);
17589 SDValue Mask = Op.getOperand(3);
17590 SDValue Index = Op.getOperand(4);
17591 SDValue Src = Op.getOperand(5);
17592 SDValue Scale = Op.getOperand(6);
17593 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
17596 SDValue Hint = Op.getOperand(6);
17598 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
17599 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
17600 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
17601 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17602 SDValue Chain = Op.getOperand(0);
17603 SDValue Mask = Op.getOperand(2);
17604 SDValue Index = Op.getOperand(3);
17605 SDValue Base = Op.getOperand(4);
17606 SDValue Scale = Op.getOperand(5);
17607 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17609 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17611 SmallVector<SDValue, 2> Results;
17612 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
17613 return DAG.getMergeValues(Results, dl);
17615 // Read Performance Monitoring Counters.
17617 SmallVector<SDValue, 2> Results;
17618 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17619 return DAG.getMergeValues(Results, dl);
17621 // XTEST intrinsics.
17623 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17624 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17625 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17626 DAG.getConstant(X86::COND_NE, MVT::i8),
17628 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17629 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17630 Ret, SDValue(InTrans.getNode(), 1));
17634 SmallVector<SDValue, 2> Results;
17635 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17636 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17637 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17638 DAG.getConstant(-1, MVT::i8));
17639 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17640 Op.getOperand(4), GenCF.getValue(1));
17641 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17642 Op.getOperand(5), MachinePointerInfo(),
17644 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17645 DAG.getConstant(X86::COND_B, MVT::i8),
17647 Results.push_back(SetCC);
17648 Results.push_back(Store);
17649 return DAG.getMergeValues(Results, dl);
17651 case COMPRESS_TO_MEM: {
17653 SDValue Mask = Op.getOperand(4);
17654 SDValue DataToCompress = Op.getOperand(3);
17655 SDValue Addr = Op.getOperand(2);
17656 SDValue Chain = Op.getOperand(0);
17658 if (isAllOnes(Mask)) // return just a store
17659 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17660 MachinePointerInfo(), false, false, 0);
17662 EVT VT = DataToCompress.getValueType();
17663 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17664 VT.getVectorNumElements());
17665 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17666 Mask.getValueType().getSizeInBits());
17667 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17668 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17669 DAG.getIntPtrConstant(0));
17671 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
17672 DataToCompress, DAG.getUNDEF(VT));
17673 return DAG.getStore(Chain, dl, Compressed, Addr,
17674 MachinePointerInfo(), false, false, 0);
17676 case EXPAND_FROM_MEM: {
17678 SDValue Mask = Op.getOperand(4);
17679 SDValue PathThru = Op.getOperand(3);
17680 SDValue Addr = Op.getOperand(2);
17681 SDValue Chain = Op.getOperand(0);
17682 EVT VT = Op.getValueType();
17684 if (isAllOnes(Mask)) // return just a load
17685 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17687 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17688 VT.getVectorNumElements());
17689 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17690 Mask.getValueType().getSizeInBits());
17691 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17692 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
17693 DAG.getIntPtrConstant(0));
17695 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17696 false, false, false, 0);
17698 SmallVector<SDValue, 2> Results;
17699 Results.push_back(DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand,
17701 Results.push_back(Chain);
17702 return DAG.getMergeValues(Results, dl);
17707 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17708 SelectionDAG &DAG) const {
17709 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17710 MFI->setReturnAddressIsTaken(true);
17712 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17715 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17717 EVT PtrVT = getPointerTy();
17720 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17721 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
17722 DAG.getSubtarget().getRegisterInfo());
17723 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
17724 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17725 DAG.getNode(ISD::ADD, dl, PtrVT,
17726 FrameAddr, Offset),
17727 MachinePointerInfo(), false, false, false, 0);
17730 // Just load the return address.
17731 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17732 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17733 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17736 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17737 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17738 MFI->setFrameAddressIsTaken(true);
17740 EVT VT = Op.getValueType();
17741 SDLoc dl(Op); // FIXME probably not meaningful
17742 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17743 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
17744 DAG.getSubtarget().getRegisterInfo());
17745 unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(
17746 DAG.getMachineFunction());
17747 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17748 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17749 "Invalid Frame Register!");
17750 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17752 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17753 MachinePointerInfo(),
17754 false, false, false, 0);
17758 // FIXME? Maybe this could be a TableGen attribute on some registers and
17759 // this table could be generated automatically from RegInfo.
17760 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
17762 unsigned Reg = StringSwitch<unsigned>(RegName)
17763 .Case("esp", X86::ESP)
17764 .Case("rsp", X86::RSP)
17768 report_fatal_error("Invalid register name global variable");
17771 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17772 SelectionDAG &DAG) const {
17773 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
17774 DAG.getSubtarget().getRegisterInfo());
17775 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
17778 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17779 SDValue Chain = Op.getOperand(0);
17780 SDValue Offset = Op.getOperand(1);
17781 SDValue Handler = Op.getOperand(2);
17784 EVT PtrVT = getPointerTy();
17785 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
17786 DAG.getSubtarget().getRegisterInfo());
17787 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17788 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17789 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17790 "Invalid Frame Register!");
17791 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17792 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17794 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17795 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
17796 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17797 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17799 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17801 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17802 DAG.getRegister(StoreAddrReg, PtrVT));
17805 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17806 SelectionDAG &DAG) const {
17808 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17809 DAG.getVTList(MVT::i32, MVT::Other),
17810 Op.getOperand(0), Op.getOperand(1));
17813 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17814 SelectionDAG &DAG) const {
17816 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17817 Op.getOperand(0), Op.getOperand(1));
17820 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17821 return Op.getOperand(0);
17824 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17825 SelectionDAG &DAG) const {
17826 SDValue Root = Op.getOperand(0);
17827 SDValue Trmp = Op.getOperand(1); // trampoline
17828 SDValue FPtr = Op.getOperand(2); // nested function
17829 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17832 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17833 const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
17835 if (Subtarget->is64Bit()) {
17836 SDValue OutChains[6];
17838 // Large code-model.
17839 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17840 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17842 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17843 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17845 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17847 // Load the pointer to the nested function into R11.
17848 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17849 SDValue Addr = Trmp;
17850 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
17851 Addr, MachinePointerInfo(TrmpAddr),
17854 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17855 DAG.getConstant(2, MVT::i64));
17856 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17857 MachinePointerInfo(TrmpAddr, 2),
17860 // Load the 'nest' parameter value into R10.
17861 // R10 is specified in X86CallingConv.td
17862 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17863 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17864 DAG.getConstant(10, MVT::i64));
17865 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
17866 Addr, MachinePointerInfo(TrmpAddr, 10),
17869 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17870 DAG.getConstant(12, MVT::i64));
17871 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17872 MachinePointerInfo(TrmpAddr, 12),
17875 // Jump to the nested function.
17876 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17877 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17878 DAG.getConstant(20, MVT::i64));
17879 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
17880 Addr, MachinePointerInfo(TrmpAddr, 20),
17883 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17884 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17885 DAG.getConstant(22, MVT::i64));
17886 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
17887 MachinePointerInfo(TrmpAddr, 22),
17890 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17892 const Function *Func =
17893 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17894 CallingConv::ID CC = Func->getCallingConv();
17899 llvm_unreachable("Unsupported calling convention");
17900 case CallingConv::C:
17901 case CallingConv::X86_StdCall: {
17902 // Pass 'nest' parameter in ECX.
17903 // Must be kept in sync with X86CallingConv.td
17904 NestReg = X86::ECX;
17906 // Check that ECX wasn't needed by an 'inreg' parameter.
17907 FunctionType *FTy = Func->getFunctionType();
17908 const AttributeSet &Attrs = Func->getAttributes();
17910 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17911 unsigned InRegCount = 0;
17914 for (FunctionType::param_iterator I = FTy->param_begin(),
17915 E = FTy->param_end(); I != E; ++I, ++Idx)
17916 if (Attrs.hasAttribute(Idx, Attribute::InReg))
17917 // FIXME: should only count parameters that are lowered to integers.
17918 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
17920 if (InRegCount > 2) {
17921 report_fatal_error("Nest register in use - reduce number of inreg"
17927 case CallingConv::X86_FastCall:
17928 case CallingConv::X86_ThisCall:
17929 case CallingConv::Fast:
17930 // Pass 'nest' parameter in EAX.
17931 // Must be kept in sync with X86CallingConv.td
17932 NestReg = X86::EAX;
17936 SDValue OutChains[4];
17937 SDValue Addr, Disp;
17939 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17940 DAG.getConstant(10, MVT::i32));
17941 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17943 // This is storing the opcode for MOV32ri.
17944 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17945 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17946 OutChains[0] = DAG.getStore(Root, dl,
17947 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
17948 Trmp, MachinePointerInfo(TrmpAddr),
17951 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17952 DAG.getConstant(1, MVT::i32));
17953 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17954 MachinePointerInfo(TrmpAddr, 1),
17957 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17958 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17959 DAG.getConstant(5, MVT::i32));
17960 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
17961 MachinePointerInfo(TrmpAddr, 5),
17964 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17965 DAG.getConstant(6, MVT::i32));
17966 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17967 MachinePointerInfo(TrmpAddr, 6),
17970 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17974 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17975 SelectionDAG &DAG) const {
17977 The rounding mode is in bits 11:10 of FPSR, and has the following
17979 00 Round to nearest
17984 FLT_ROUNDS, on the other hand, expects the following:
17991 To perform the conversion, we do:
17992 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17995 MachineFunction &MF = DAG.getMachineFunction();
17996 const TargetMachine &TM = MF.getTarget();
17997 const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
17998 unsigned StackAlignment = TFI.getStackAlignment();
17999 MVT VT = Op.getSimpleValueType();
18002 // Save FP Control Word to stack slot
18003 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
18004 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
18006 MachineMemOperand *MMO =
18007 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
18008 MachineMemOperand::MOStore, 2, 2);
18010 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
18011 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
18012 DAG.getVTList(MVT::Other),
18013 Ops, MVT::i16, MMO);
18015 // Load FP Control Word from stack slot
18016 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
18017 MachinePointerInfo(), false, false, false, 0);
18019 // Transform as necessary
18021 DAG.getNode(ISD::SRL, DL, MVT::i16,
18022 DAG.getNode(ISD::AND, DL, MVT::i16,
18023 CWD, DAG.getConstant(0x800, MVT::i16)),
18024 DAG.getConstant(11, MVT::i8));
18026 DAG.getNode(ISD::SRL, DL, MVT::i16,
18027 DAG.getNode(ISD::AND, DL, MVT::i16,
18028 CWD, DAG.getConstant(0x400, MVT::i16)),
18029 DAG.getConstant(9, MVT::i8));
18032 DAG.getNode(ISD::AND, DL, MVT::i16,
18033 DAG.getNode(ISD::ADD, DL, MVT::i16,
18034 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
18035 DAG.getConstant(1, MVT::i16)),
18036 DAG.getConstant(3, MVT::i16));
18038 return DAG.getNode((VT.getSizeInBits() < 16 ?
18039 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
18042 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
18043 MVT VT = Op.getSimpleValueType();
18045 unsigned NumBits = VT.getSizeInBits();
18048 Op = Op.getOperand(0);
18049 if (VT == MVT::i8) {
18050 // Zero extend to i32 since there is not an i8 bsr.
18052 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18055 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
18056 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18057 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18059 // If src is zero (i.e. bsr sets ZF), returns NumBits.
18062 DAG.getConstant(NumBits+NumBits-1, OpVT),
18063 DAG.getConstant(X86::COND_E, MVT::i8),
18066 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
18068 // Finally xor with NumBits-1.
18069 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18072 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18076 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
18077 MVT VT = Op.getSimpleValueType();
18079 unsigned NumBits = VT.getSizeInBits();
18082 Op = Op.getOperand(0);
18083 if (VT == MVT::i8) {
18084 // Zero extend to i32 since there is not an i8 bsr.
18086 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
18089 // Issue a bsr (scan bits in reverse).
18090 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
18091 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
18093 // And xor with NumBits-1.
18094 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
18097 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
18101 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
18102 MVT VT = Op.getSimpleValueType();
18103 unsigned NumBits = VT.getSizeInBits();
18105 Op = Op.getOperand(0);
18107 // Issue a bsf (scan bits forward) which also sets EFLAGS.
18108 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18109 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
18111 // If src is zero (i.e. bsf sets ZF), returns NumBits.
18114 DAG.getConstant(NumBits, VT),
18115 DAG.getConstant(X86::COND_E, MVT::i8),
18118 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
18121 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
18122 // ones, and then concatenate the result back.
18123 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
18124 MVT VT = Op.getSimpleValueType();
18126 assert(VT.is256BitVector() && VT.isInteger() &&
18127 "Unsupported value type for operation");
18129 unsigned NumElems = VT.getVectorNumElements();
18132 // Extract the LHS vectors
18133 SDValue LHS = Op.getOperand(0);
18134 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
18135 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
18137 // Extract the RHS vectors
18138 SDValue RHS = Op.getOperand(1);
18139 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
18140 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
18142 MVT EltVT = VT.getVectorElementType();
18143 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18145 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
18146 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
18147 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
18150 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
18151 assert(Op.getSimpleValueType().is256BitVector() &&
18152 Op.getSimpleValueType().isInteger() &&
18153 "Only handle AVX 256-bit vector integer operation");
18154 return Lower256IntArith(Op, DAG);
18157 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
18158 assert(Op.getSimpleValueType().is256BitVector() &&
18159 Op.getSimpleValueType().isInteger() &&
18160 "Only handle AVX 256-bit vector integer operation");
18161 return Lower256IntArith(Op, DAG);
18164 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
18165 SelectionDAG &DAG) {
18167 MVT VT = Op.getSimpleValueType();
18169 // Decompose 256-bit ops into smaller 128-bit ops.
18170 if (VT.is256BitVector() && !Subtarget->hasInt256())
18171 return Lower256IntArith(Op, DAG);
18173 SDValue A = Op.getOperand(0);
18174 SDValue B = Op.getOperand(1);
18176 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
18177 if (VT == MVT::v4i32) {
18178 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
18179 "Should not custom lower when pmuldq is available!");
18181 // Extract the odd parts.
18182 static const int UnpackMask[] = { 1, -1, 3, -1 };
18183 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
18184 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
18186 // Multiply the even parts.
18187 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
18188 // Now multiply odd parts.
18189 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
18191 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
18192 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
18194 // Merge the two vectors back together with a shuffle. This expands into 2
18196 static const int ShufMask[] = { 0, 4, 2, 6 };
18197 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
18200 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
18201 "Only know how to lower V2I64/V4I64/V8I64 multiply");
18203 // Ahi = psrlqi(a, 32);
18204 // Bhi = psrlqi(b, 32);
18206 // AloBlo = pmuludq(a, b);
18207 // AloBhi = pmuludq(a, Bhi);
18208 // AhiBlo = pmuludq(Ahi, b);
18210 // AloBhi = psllqi(AloBhi, 32);
18211 // AhiBlo = psllqi(AhiBlo, 32);
18212 // return AloBlo + AloBhi + AhiBlo;
18214 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
18215 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
18217 // Bit cast to 32-bit vectors for MULUDQ
18218 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
18219 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
18220 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
18221 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
18222 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
18223 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
18225 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
18226 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
18227 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
18229 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
18230 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
18232 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
18233 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
18236 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
18237 assert(Subtarget->isTargetWin64() && "Unexpected target");
18238 EVT VT = Op.getValueType();
18239 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
18240 "Unexpected return type for lowering");
18244 switch (Op->getOpcode()) {
18245 default: llvm_unreachable("Unexpected request for libcall!");
18246 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
18247 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
18248 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
18249 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
18250 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
18251 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
18255 SDValue InChain = DAG.getEntryNode();
18257 TargetLowering::ArgListTy Args;
18258 TargetLowering::ArgListEntry Entry;
18259 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
18260 EVT ArgVT = Op->getOperand(i).getValueType();
18261 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
18262 "Unexpected argument type for lowering");
18263 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
18264 Entry.Node = StackPtr;
18265 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
18267 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18268 Entry.Ty = PointerType::get(ArgTy,0);
18269 Entry.isSExt = false;
18270 Entry.isZExt = false;
18271 Args.push_back(Entry);
18274 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
18277 TargetLowering::CallLoweringInfo CLI(DAG);
18278 CLI.setDebugLoc(dl).setChain(InChain)
18279 .setCallee(getLibcallCallingConv(LC),
18280 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18281 Callee, std::move(Args), 0)
18282 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18284 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18285 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
18288 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18289 SelectionDAG &DAG) {
18290 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18291 EVT VT = Op0.getValueType();
18294 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18295 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18297 // PMULxD operations multiply each even value (starting at 0) of LHS with
18298 // the related value of RHS and produce a widen result.
18299 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18300 // => <2 x i64> <ae|cg>
18302 // In other word, to have all the results, we need to perform two PMULxD:
18303 // 1. one with the even values.
18304 // 2. one with the odd values.
18305 // To achieve #2, with need to place the odd values at an even position.
18307 // Place the odd value at an even position (basically, shift all values 1
18308 // step to the left):
18309 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18310 // <a|b|c|d> => <b|undef|d|undef>
18311 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18312 // <e|f|g|h> => <f|undef|h|undef>
18313 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18315 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18317 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18318 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18320 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18321 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18322 // => <2 x i64> <ae|cg>
18323 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
18324 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18325 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18326 // => <2 x i64> <bf|dh>
18327 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
18328 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18330 // Shuffle it back into the right order.
18331 SDValue Highs, Lows;
18332 if (VT == MVT::v8i32) {
18333 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18334 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18335 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18336 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18338 const int HighMask[] = {1, 5, 3, 7};
18339 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18340 const int LowMask[] = {0, 4, 2, 6};
18341 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18344 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18345 // unsigned multiply.
18346 if (IsSigned && !Subtarget->hasSSE41()) {
18348 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
18349 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18350 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18351 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18352 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18354 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18355 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18358 // The first result of MUL_LOHI is actually the low value, followed by the
18360 SDValue Ops[] = {Lows, Highs};
18361 return DAG.getMergeValues(Ops, dl);
18364 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18365 const X86Subtarget *Subtarget) {
18366 MVT VT = Op.getSimpleValueType();
18368 SDValue R = Op.getOperand(0);
18369 SDValue Amt = Op.getOperand(1);
18371 // Optimize shl/srl/sra with constant shift amount.
18372 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18373 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18374 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18376 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
18377 (Subtarget->hasInt256() &&
18378 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
18379 (Subtarget->hasAVX512() &&
18380 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
18381 if (Op.getOpcode() == ISD::SHL)
18382 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
18384 if (Op.getOpcode() == ISD::SRL)
18385 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
18387 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
18388 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
18392 if (VT == MVT::v16i8) {
18393 if (Op.getOpcode() == ISD::SHL) {
18394 // Make a large shift.
18395 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18396 MVT::v8i16, R, ShiftAmt,
18398 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18399 // Zero out the rightmost bits.
18400 SmallVector<SDValue, 16> V(16,
18401 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18403 return DAG.getNode(ISD::AND, dl, VT, SHL,
18404 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18406 if (Op.getOpcode() == ISD::SRL) {
18407 // Make a large shift.
18408 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18409 MVT::v8i16, R, ShiftAmt,
18411 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18412 // Zero out the leftmost bits.
18413 SmallVector<SDValue, 16> V(16,
18414 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18416 return DAG.getNode(ISD::AND, dl, VT, SRL,
18417 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18419 if (Op.getOpcode() == ISD::SRA) {
18420 if (ShiftAmt == 7) {
18421 // R s>> 7 === R s< 0
18422 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18423 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18426 // R s>> a === ((R u>> a) ^ m) - m
18427 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18428 SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
18430 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18431 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18432 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18435 llvm_unreachable("Unknown shift opcode.");
18438 if (Subtarget->hasInt256() && VT == MVT::v32i8) {
18439 if (Op.getOpcode() == ISD::SHL) {
18440 // Make a large shift.
18441 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl,
18442 MVT::v16i16, R, ShiftAmt,
18444 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
18445 // Zero out the rightmost bits.
18446 SmallVector<SDValue, 32> V(32,
18447 DAG.getConstant(uint8_t(-1U << ShiftAmt),
18449 return DAG.getNode(ISD::AND, dl, VT, SHL,
18450 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18452 if (Op.getOpcode() == ISD::SRL) {
18453 // Make a large shift.
18454 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl,
18455 MVT::v16i16, R, ShiftAmt,
18457 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
18458 // Zero out the leftmost bits.
18459 SmallVector<SDValue, 32> V(32,
18460 DAG.getConstant(uint8_t(-1U) >> ShiftAmt,
18462 return DAG.getNode(ISD::AND, dl, VT, SRL,
18463 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18465 if (Op.getOpcode() == ISD::SRA) {
18466 if (ShiftAmt == 7) {
18467 // R s>> 7 === R s< 0
18468 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18469 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18472 // R s>> a === ((R u>> a) ^ m) - m
18473 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18474 SmallVector<SDValue, 32> V(32, DAG.getConstant(128 >> ShiftAmt,
18476 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18477 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18478 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18481 llvm_unreachable("Unknown shift opcode.");
18486 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18487 if (!Subtarget->is64Bit() &&
18488 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18489 Amt.getOpcode() == ISD::BITCAST &&
18490 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18491 Amt = Amt.getOperand(0);
18492 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18493 VT.getVectorNumElements();
18494 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18495 uint64_t ShiftAmt = 0;
18496 for (unsigned i = 0; i != Ratio; ++i) {
18497 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
18501 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18503 // Check remaining shift amounts.
18504 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18505 uint64_t ShAmt = 0;
18506 for (unsigned j = 0; j != Ratio; ++j) {
18507 ConstantSDNode *C =
18508 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18512 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18514 if (ShAmt != ShiftAmt)
18517 switch (Op.getOpcode()) {
18519 llvm_unreachable("Unknown shift opcode!");
18521 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
18524 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
18527 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
18535 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18536 const X86Subtarget* Subtarget) {
18537 MVT VT = Op.getSimpleValueType();
18539 SDValue R = Op.getOperand(0);
18540 SDValue Amt = Op.getOperand(1);
18542 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
18543 VT == MVT::v4i32 || VT == MVT::v8i16 ||
18544 (Subtarget->hasInt256() &&
18545 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
18546 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
18547 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
18549 EVT EltVT = VT.getVectorElementType();
18551 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18552 // Check if this build_vector node is doing a splat.
18553 // If so, then set BaseShAmt equal to the splat value.
18554 BaseShAmt = BV->getSplatValue();
18555 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18556 BaseShAmt = SDValue();
18558 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18559 Amt = Amt.getOperand(0);
18561 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18562 if (SVN && SVN->isSplat()) {
18563 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18564 SDValue InVec = Amt.getOperand(0);
18565 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18566 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18567 "Unexpected shuffle index found!");
18568 BaseShAmt = InVec.getOperand(SplatIdx);
18569 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18570 if (ConstantSDNode *C =
18571 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18572 if (C->getZExtValue() == SplatIdx)
18573 BaseShAmt = InVec.getOperand(1);
18578 // Avoid introducing an extract element from a shuffle.
18579 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18580 DAG.getIntPtrConstant(SplatIdx));
18584 if (BaseShAmt.getNode()) {
18585 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18586 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18587 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18588 else if (EltVT.bitsLT(MVT::i32))
18589 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18591 switch (Op.getOpcode()) {
18593 llvm_unreachable("Unknown shift opcode!");
18595 switch (VT.SimpleTy) {
18596 default: return SDValue();
18605 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
18608 switch (VT.SimpleTy) {
18609 default: return SDValue();
18616 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
18619 switch (VT.SimpleTy) {
18620 default: return SDValue();
18629 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
18635 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18636 if (!Subtarget->is64Bit() &&
18637 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
18638 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
18639 Amt.getOpcode() == ISD::BITCAST &&
18640 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18641 Amt = Amt.getOperand(0);
18642 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18643 VT.getVectorNumElements();
18644 std::vector<SDValue> Vals(Ratio);
18645 for (unsigned i = 0; i != Ratio; ++i)
18646 Vals[i] = Amt.getOperand(i);
18647 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18648 for (unsigned j = 0; j != Ratio; ++j)
18649 if (Vals[j] != Amt.getOperand(i + j))
18652 switch (Op.getOpcode()) {
18654 llvm_unreachable("Unknown shift opcode!");
18656 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
18658 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
18660 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
18667 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18668 SelectionDAG &DAG) {
18669 MVT VT = Op.getSimpleValueType();
18671 SDValue R = Op.getOperand(0);
18672 SDValue Amt = Op.getOperand(1);
18675 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18676 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18678 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
18682 V = LowerScalarVariableShift(Op, DAG, Subtarget);
18686 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
18688 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
18689 if (Subtarget->hasInt256()) {
18690 if (Op.getOpcode() == ISD::SRL &&
18691 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18692 VT == MVT::v4i64 || VT == MVT::v8i32))
18694 if (Op.getOpcode() == ISD::SHL &&
18695 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18696 VT == MVT::v4i64 || VT == MVT::v8i32))
18698 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
18702 // If possible, lower this packed shift into a vector multiply instead of
18703 // expanding it into a sequence of scalar shifts.
18704 // Do this only if the vector shift count is a constant build_vector.
18705 if (Op.getOpcode() == ISD::SHL &&
18706 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18707 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18708 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18709 SmallVector<SDValue, 8> Elts;
18710 EVT SVT = VT.getScalarType();
18711 unsigned SVTBits = SVT.getSizeInBits();
18712 const APInt &One = APInt(SVTBits, 1);
18713 unsigned NumElems = VT.getVectorNumElements();
18715 for (unsigned i=0; i !=NumElems; ++i) {
18716 SDValue Op = Amt->getOperand(i);
18717 if (Op->getOpcode() == ISD::UNDEF) {
18718 Elts.push_back(Op);
18722 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18723 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18724 uint64_t ShAmt = C.getZExtValue();
18725 if (ShAmt >= SVTBits) {
18726 Elts.push_back(DAG.getUNDEF(SVT));
18729 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
18731 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18732 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18735 // Lower SHL with variable shift amount.
18736 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18737 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
18739 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
18740 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
18741 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18742 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18745 // If possible, lower this shift as a sequence of two shifts by
18746 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18748 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18750 // Could be rewritten as:
18751 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18753 // The advantage is that the two shifts from the example would be
18754 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18755 // the vector shift into four scalar shifts plus four pairs of vector
18757 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18758 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18759 unsigned TargetOpcode = X86ISD::MOVSS;
18760 bool CanBeSimplified;
18761 // The splat value for the first packed shift (the 'X' from the example).
18762 SDValue Amt1 = Amt->getOperand(0);
18763 // The splat value for the second packed shift (the 'Y' from the example).
18764 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18765 Amt->getOperand(2);
18767 // See if it is possible to replace this node with a sequence of
18768 // two shifts followed by a MOVSS/MOVSD
18769 if (VT == MVT::v4i32) {
18770 // Check if it is legal to use a MOVSS.
18771 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18772 Amt2 == Amt->getOperand(3);
18773 if (!CanBeSimplified) {
18774 // Otherwise, check if we can still simplify this node using a MOVSD.
18775 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18776 Amt->getOperand(2) == Amt->getOperand(3);
18777 TargetOpcode = X86ISD::MOVSD;
18778 Amt2 = Amt->getOperand(2);
18781 // Do similar checks for the case where the machine value type
18783 CanBeSimplified = Amt1 == Amt->getOperand(1);
18784 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18785 CanBeSimplified = Amt2 == Amt->getOperand(i);
18787 if (!CanBeSimplified) {
18788 TargetOpcode = X86ISD::MOVSD;
18789 CanBeSimplified = true;
18790 Amt2 = Amt->getOperand(4);
18791 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18792 CanBeSimplified = Amt1 == Amt->getOperand(i);
18793 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18794 CanBeSimplified = Amt2 == Amt->getOperand(j);
18798 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18799 isa<ConstantSDNode>(Amt2)) {
18800 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18801 EVT CastVT = MVT::v4i32;
18803 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
18804 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18806 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
18807 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18808 if (TargetOpcode == X86ISD::MOVSD)
18809 CastVT = MVT::v2i64;
18810 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
18811 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
18812 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18814 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
18818 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
18819 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
18822 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
18823 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
18825 // Turn 'a' into a mask suitable for VSELECT
18826 SDValue VSelM = DAG.getConstant(0x80, VT);
18827 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
18828 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
18830 SDValue CM1 = DAG.getConstant(0x0f, VT);
18831 SDValue CM2 = DAG.getConstant(0x3f, VT);
18833 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
18834 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
18835 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
18836 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
18837 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
18840 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
18841 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
18842 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
18844 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
18845 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
18846 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
18847 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
18848 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
18851 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
18852 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
18853 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
18855 // return VSELECT(r, r+r, a);
18856 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
18857 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
18861 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18862 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18863 // solution better.
18864 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18865 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
18867 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18868 R = DAG.getNode(ExtOpc, dl, NewVT, R);
18869 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
18870 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18871 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
18874 // Decompose 256-bit shifts into smaller 128-bit shifts.
18875 if (VT.is256BitVector()) {
18876 unsigned NumElems = VT.getVectorNumElements();
18877 MVT EltVT = VT.getVectorElementType();
18878 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18880 // Extract the two vectors
18881 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18882 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18884 // Recreate the shift amount vectors
18885 SDValue Amt1, Amt2;
18886 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18887 // Constant shift amount
18888 SmallVector<SDValue, 4> Amt1Csts;
18889 SmallVector<SDValue, 4> Amt2Csts;
18890 for (unsigned i = 0; i != NumElems/2; ++i)
18891 Amt1Csts.push_back(Amt->getOperand(i));
18892 for (unsigned i = NumElems/2; i != NumElems; ++i)
18893 Amt2Csts.push_back(Amt->getOperand(i));
18895 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18896 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18898 // Variable shift amount
18899 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18900 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18903 // Issue new vector shifts for the smaller types
18904 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18905 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18907 // Concatenate the result back
18908 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18914 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18915 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18916 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18917 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18918 // has only one use.
18919 SDNode *N = Op.getNode();
18920 SDValue LHS = N->getOperand(0);
18921 SDValue RHS = N->getOperand(1);
18922 unsigned BaseOp = 0;
18925 switch (Op.getOpcode()) {
18926 default: llvm_unreachable("Unknown ovf instruction!");
18928 // A subtract of one will be selected as a INC. Note that INC doesn't
18929 // set CF, so we can't do this for UADDO.
18930 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18932 BaseOp = X86ISD::INC;
18933 Cond = X86::COND_O;
18936 BaseOp = X86ISD::ADD;
18937 Cond = X86::COND_O;
18940 BaseOp = X86ISD::ADD;
18941 Cond = X86::COND_B;
18944 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18945 // set CF, so we can't do this for USUBO.
18946 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18948 BaseOp = X86ISD::DEC;
18949 Cond = X86::COND_O;
18952 BaseOp = X86ISD::SUB;
18953 Cond = X86::COND_O;
18956 BaseOp = X86ISD::SUB;
18957 Cond = X86::COND_B;
18960 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18961 Cond = X86::COND_O;
18963 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18964 if (N->getValueType(0) == MVT::i8) {
18965 BaseOp = X86ISD::UMUL8;
18966 Cond = X86::COND_O;
18969 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18971 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18974 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18975 DAG.getConstant(X86::COND_O, MVT::i32),
18976 SDValue(Sum.getNode(), 2));
18978 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18982 // Also sets EFLAGS.
18983 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18984 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18987 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18988 DAG.getConstant(Cond, MVT::i32),
18989 SDValue(Sum.getNode(), 1));
18991 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18994 // Sign extension of the low part of vector elements. This may be used either
18995 // when sign extend instructions are not available or if the vector element
18996 // sizes already match the sign-extended size. If the vector elements are in
18997 // their pre-extended size and sign extend instructions are available, that will
18998 // be handled by LowerSIGN_EXTEND.
18999 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
19000 SelectionDAG &DAG) const {
19002 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
19003 MVT VT = Op.getSimpleValueType();
19005 if (!Subtarget->hasSSE2() || !VT.isVector())
19008 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
19009 ExtraVT.getScalarType().getSizeInBits();
19011 switch (VT.SimpleTy) {
19012 default: return SDValue();
19015 if (!Subtarget->hasFp256())
19017 if (!Subtarget->hasInt256()) {
19018 // needs to be split
19019 unsigned NumElems = VT.getVectorNumElements();
19021 // Extract the LHS vectors
19022 SDValue LHS = Op.getOperand(0);
19023 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
19024 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
19026 MVT EltVT = VT.getVectorElementType();
19027 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
19029 EVT ExtraEltVT = ExtraVT.getVectorElementType();
19030 unsigned ExtraNumElems = ExtraVT.getVectorNumElements();
19031 ExtraVT = EVT::getVectorVT(*DAG.getContext(), ExtraEltVT,
19033 SDValue Extra = DAG.getValueType(ExtraVT);
19035 LHS1 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, Extra);
19036 LHS2 = DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, Extra);
19038 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, LHS1, LHS2);
19043 SDValue Op0 = Op.getOperand(0);
19045 // This is a sign extension of some low part of vector elements without
19046 // changing the size of the vector elements themselves:
19047 // Shift-Left + Shift-Right-Algebraic.
19048 SDValue Shl = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, Op0,
19050 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, Shl, BitsDiff,
19056 /// Returns true if the operand type is exactly twice the native width, and
19057 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
19058 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
19059 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
19060 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
19061 const X86Subtarget &Subtarget =
19062 getTargetMachine().getSubtarget<X86Subtarget>();
19063 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
19066 return !Subtarget.is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
19067 else if (OpWidth == 128)
19068 return Subtarget.hasCmpxchg16b();
19073 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
19074 return needsCmpXchgNb(SI->getValueOperand()->getType());
19077 // Note: this turns large loads into lock cmpxchg8b/16b.
19078 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
19079 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
19080 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
19081 return needsCmpXchgNb(PTy->getElementType());
19084 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
19085 const X86Subtarget &Subtarget =
19086 getTargetMachine().getSubtarget<X86Subtarget>();
19087 unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
19088 const Type *MemType = AI->getType();
19090 // If the operand is too big, we must see if cmpxchg8/16b is available
19091 // and default to library calls otherwise.
19092 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19093 return needsCmpXchgNb(MemType);
19095 AtomicRMWInst::BinOp Op = AI->getOperation();
19098 llvm_unreachable("Unknown atomic operation");
19099 case AtomicRMWInst::Xchg:
19100 case AtomicRMWInst::Add:
19101 case AtomicRMWInst::Sub:
19102 // It's better to use xadd, xsub or xchg for these in all cases.
19104 case AtomicRMWInst::Or:
19105 case AtomicRMWInst::And:
19106 case AtomicRMWInst::Xor:
19107 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19108 // prefix to a normal instruction for these operations.
19109 return !AI->use_empty();
19110 case AtomicRMWInst::Nand:
19111 case AtomicRMWInst::Max:
19112 case AtomicRMWInst::Min:
19113 case AtomicRMWInst::UMax:
19114 case AtomicRMWInst::UMin:
19115 // These always require a non-trivial set of data operations on x86. We must
19116 // use a cmpxchg loop.
19121 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19122 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19123 // no-sse2). There isn't any reason to disable it if the target processor
19125 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19129 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19130 const X86Subtarget &Subtarget =
19131 getTargetMachine().getSubtarget<X86Subtarget>();
19132 unsigned NativeWidth = Subtarget.is64Bit() ? 64 : 32;
19133 const Type *MemType = AI->getType();
19134 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19135 // there is no benefit in turning such RMWs into loads, and it is actually
19136 // harmful as it introduces a mfence.
19137 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19140 auto Builder = IRBuilder<>(AI);
19141 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19142 auto SynchScope = AI->getSynchScope();
19143 // We must restrict the ordering to avoid generating loads with Release or
19144 // ReleaseAcquire orderings.
19145 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19146 auto Ptr = AI->getPointerOperand();
19148 // Before the load we need a fence. Here is an example lifted from
19149 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19152 // x.store(1, relaxed);
19153 // r1 = y.fetch_add(0, release);
19155 // y.fetch_add(42, acquire);
19156 // r2 = x.load(relaxed);
19157 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19158 // lowered to just a load without a fence. A mfence flushes the store buffer,
19159 // making the optimization clearly correct.
19160 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19161 // otherwise, we might be able to be more agressive on relaxed idempotent
19162 // rmw. In practice, they do not look useful, so we don't try to be
19163 // especially clever.
19164 if (SynchScope == SingleThread) {
19165 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19166 // the IR level, so we must wrap it in an intrinsic.
19168 } else if (hasMFENCE(Subtarget)) {
19169 Function *MFence = llvm::Intrinsic::getDeclaration(M,
19170 Intrinsic::x86_sse2_mfence);
19171 Builder.CreateCall(MFence);
19173 // FIXME: it might make sense to use a locked operation here but on a
19174 // different cache-line to prevent cache-line bouncing. In practice it
19175 // is probably a small win, and x86 processors without mfence are rare
19176 // enough that we do not bother.
19180 // Finally we can emit the atomic load.
19181 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19182 AI->getType()->getPrimitiveSizeInBits());
19183 Loaded->setAtomic(Order, SynchScope);
19184 AI->replaceAllUsesWith(Loaded);
19185 AI->eraseFromParent();
19189 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19190 SelectionDAG &DAG) {
19192 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19193 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19194 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19195 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19197 // The only fence that needs an instruction is a sequentially-consistent
19198 // cross-thread fence.
19199 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19200 if (hasMFENCE(*Subtarget))
19201 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19203 SDValue Chain = Op.getOperand(0);
19204 SDValue Zero = DAG.getConstant(0, MVT::i32);
19206 DAG.getRegister(X86::ESP, MVT::i32), // Base
19207 DAG.getTargetConstant(1, MVT::i8), // Scale
19208 DAG.getRegister(0, MVT::i32), // Index
19209 DAG.getTargetConstant(0, MVT::i32), // Disp
19210 DAG.getRegister(0, MVT::i32), // Segment.
19214 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19215 return SDValue(Res, 0);
19218 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19219 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19222 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19223 SelectionDAG &DAG) {
19224 MVT T = Op.getSimpleValueType();
19228 switch(T.SimpleTy) {
19229 default: llvm_unreachable("Invalid value type!");
19230 case MVT::i8: Reg = X86::AL; size = 1; break;
19231 case MVT::i16: Reg = X86::AX; size = 2; break;
19232 case MVT::i32: Reg = X86::EAX; size = 4; break;
19234 assert(Subtarget->is64Bit() && "Node not type legal!");
19235 Reg = X86::RAX; size = 8;
19238 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19239 Op.getOperand(2), SDValue());
19240 SDValue Ops[] = { cpIn.getValue(0),
19243 DAG.getTargetConstant(size, MVT::i8),
19244 cpIn.getValue(1) };
19245 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19246 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19247 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19251 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19252 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19253 MVT::i32, cpOut.getValue(2));
19254 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19255 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
19257 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19258 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19259 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19263 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19264 SelectionDAG &DAG) {
19265 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19266 MVT DstVT = Op.getSimpleValueType();
19268 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19269 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19270 if (DstVT != MVT::f64)
19271 // This conversion needs to be expanded.
19274 SDValue InVec = Op->getOperand(0);
19276 unsigned NumElts = SrcVT.getVectorNumElements();
19277 EVT SVT = SrcVT.getVectorElementType();
19279 // Widen the vector in input in the case of MVT::v2i32.
19280 // Example: from MVT::v2i32 to MVT::v4i32.
19281 SmallVector<SDValue, 16> Elts;
19282 for (unsigned i = 0, e = NumElts; i != e; ++i)
19283 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19284 DAG.getIntPtrConstant(i)));
19286 // Explicitly mark the extra elements as Undef.
19287 SDValue Undef = DAG.getUNDEF(SVT);
19288 for (unsigned i = NumElts, e = NumElts * 2; i != e; ++i)
19289 Elts.push_back(Undef);
19291 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19292 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19293 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
19294 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19295 DAG.getIntPtrConstant(0));
19298 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19299 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19300 assert((DstVT == MVT::i64 ||
19301 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19302 "Unexpected custom BITCAST");
19303 // i64 <=> MMX conversions are Legal.
19304 if (SrcVT==MVT::i64 && DstVT.isVector())
19306 if (DstVT==MVT::i64 && SrcVT.isVector())
19308 // MMX <=> MMX conversions are Legal.
19309 if (SrcVT.isVector() && DstVT.isVector())
19311 // All other conversions need to be expanded.
19315 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19316 SelectionDAG &DAG) {
19317 SDNode *Node = Op.getNode();
19320 Op = Op.getOperand(0);
19321 EVT VT = Op.getValueType();
19322 assert((VT.is128BitVector() || VT.is256BitVector()) &&
19323 "CTPOP lowering only implemented for 128/256-bit wide vector types");
19325 unsigned NumElts = VT.getVectorNumElements();
19326 EVT EltVT = VT.getVectorElementType();
19327 unsigned Len = EltVT.getSizeInBits();
19329 // This is the vectorized version of the "best" algorithm from
19330 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19331 // with a minor tweak to use a series of adds + shifts instead of vector
19332 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
19334 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
19335 // v8i32 => Always profitable
19337 // FIXME: There a couple of possible improvements:
19339 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
19340 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
19342 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
19343 "CTPOP not implemented for this vector element type.");
19345 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
19346 // extra legalization.
19347 bool NeedsBitcast = EltVT == MVT::i32;
19348 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
19350 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
19351 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
19352 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
19354 // v = v - ((v >> 1) & 0x55555555...)
19355 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
19356 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
19357 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
19359 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19361 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
19362 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
19364 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
19366 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
19367 if (VT != And.getValueType())
19368 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19369 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
19371 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19372 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
19373 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
19374 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
19375 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
19377 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
19378 if (NeedsBitcast) {
19379 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
19380 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
19381 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
19384 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
19385 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
19386 if (VT != AndRHS.getValueType()) {
19387 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
19388 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
19390 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
19392 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19393 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
19394 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
19395 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
19396 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19398 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
19399 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
19400 if (NeedsBitcast) {
19401 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19402 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
19404 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
19405 if (VT != And.getValueType())
19406 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19408 // The algorithm mentioned above uses:
19409 // v = (v * 0x01010101...) >> (Len - 8)
19411 // Change it to use vector adds + vector shifts which yield faster results on
19412 // Haswell than using vector integer multiplication.
19414 // For i32 elements:
19415 // v = v + (v >> 8)
19416 // v = v + (v >> 16)
19418 // For i64 elements:
19419 // v = v + (v >> 8)
19420 // v = v + (v >> 16)
19421 // v = v + (v >> 32)
19424 SmallVector<SDValue, 8> Csts;
19425 for (unsigned i = 8; i <= Len/2; i *= 2) {
19426 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
19427 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
19428 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
19429 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
19433 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
19434 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
19435 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
19436 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
19437 if (NeedsBitcast) {
19438 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
19439 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
19441 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
19442 if (VT != And.getValueType())
19443 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
19448 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19449 SDNode *Node = Op.getNode();
19451 EVT T = Node->getValueType(0);
19452 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19453 DAG.getConstant(0, T), Node->getOperand(2));
19454 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19455 cast<AtomicSDNode>(Node)->getMemoryVT(),
19456 Node->getOperand(0),
19457 Node->getOperand(1), negOp,
19458 cast<AtomicSDNode>(Node)->getMemOperand(),
19459 cast<AtomicSDNode>(Node)->getOrdering(),
19460 cast<AtomicSDNode>(Node)->getSynchScope());
19463 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19464 SDNode *Node = Op.getNode();
19466 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19468 // Convert seq_cst store -> xchg
19469 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19470 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19471 // (The only way to get a 16-byte store is cmpxchg16b)
19472 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19473 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19474 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19475 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19476 cast<AtomicSDNode>(Node)->getMemoryVT(),
19477 Node->getOperand(0),
19478 Node->getOperand(1), Node->getOperand(2),
19479 cast<AtomicSDNode>(Node)->getMemOperand(),
19480 cast<AtomicSDNode>(Node)->getOrdering(),
19481 cast<AtomicSDNode>(Node)->getSynchScope());
19482 return Swap.getValue(1);
19484 // Other atomic stores have a simple pattern.
19488 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19489 EVT VT = Op.getNode()->getSimpleValueType(0);
19491 // Let legalize expand this if it isn't a legal type yet.
19492 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19495 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19498 bool ExtraOp = false;
19499 switch (Op.getOpcode()) {
19500 default: llvm_unreachable("Invalid code");
19501 case ISD::ADDC: Opc = X86ISD::ADD; break;
19502 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19503 case ISD::SUBC: Opc = X86ISD::SUB; break;
19504 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19508 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19510 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19511 Op.getOperand(1), Op.getOperand(2));
19514 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19515 SelectionDAG &DAG) {
19516 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19518 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19519 // which returns the values as { float, float } (in XMM0) or
19520 // { double, double } (which is returned in XMM0, XMM1).
19522 SDValue Arg = Op.getOperand(0);
19523 EVT ArgVT = Arg.getValueType();
19524 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19526 TargetLowering::ArgListTy Args;
19527 TargetLowering::ArgListEntry Entry;
19531 Entry.isSExt = false;
19532 Entry.isZExt = false;
19533 Args.push_back(Entry);
19535 bool isF64 = ArgVT == MVT::f64;
19536 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19537 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19538 // the results are returned via SRet in memory.
19539 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19540 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19541 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
19543 Type *RetTy = isF64
19544 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19545 : (Type*)VectorType::get(ArgTy, 4);
19547 TargetLowering::CallLoweringInfo CLI(DAG);
19548 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19549 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19551 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19554 // Returned in xmm0 and xmm1.
19555 return CallResult.first;
19557 // Returned in bits 0:31 and 32:64 xmm0.
19558 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19559 CallResult.first, DAG.getIntPtrConstant(0));
19560 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19561 CallResult.first, DAG.getIntPtrConstant(1));
19562 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19563 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19566 /// LowerOperation - Provide custom lowering hooks for some operations.
19568 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19569 switch (Op.getOpcode()) {
19570 default: llvm_unreachable("Should not custom lower this!");
19571 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
19572 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19573 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19574 return LowerCMP_SWAP(Op, Subtarget, DAG);
19575 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19576 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19577 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19578 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19579 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
19580 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
19581 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19582 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19583 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19584 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19585 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19586 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19587 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19588 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19589 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19590 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19591 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19592 case ISD::SHL_PARTS:
19593 case ISD::SRA_PARTS:
19594 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19595 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19596 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19597 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19598 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19599 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19600 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19601 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19602 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19603 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19604 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19606 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19607 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19608 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19609 case ISD::SETCC: return LowerSETCC(Op, DAG);
19610 case ISD::SELECT: return LowerSELECT(Op, DAG);
19611 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19612 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19613 case ISD::VASTART: return LowerVASTART(Op, DAG);
19614 case ISD::VAARG: return LowerVAARG(Op, DAG);
19615 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19616 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19617 case ISD::INTRINSIC_VOID:
19618 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19619 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19620 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19621 case ISD::FRAME_TO_ARGS_OFFSET:
19622 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19623 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19624 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19625 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19626 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19627 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19628 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19629 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19630 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
19631 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
19632 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
19633 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19634 case ISD::UMUL_LOHI:
19635 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19638 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19644 case ISD::UMULO: return LowerXALUO(Op, DAG);
19645 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19646 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19650 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19651 case ISD::ADD: return LowerADD(Op, DAG);
19652 case ISD::SUB: return LowerSUB(Op, DAG);
19653 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19657 /// ReplaceNodeResults - Replace a node with an illegal result type
19658 /// with a new node built out of custom code.
19659 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19660 SmallVectorImpl<SDValue>&Results,
19661 SelectionDAG &DAG) const {
19663 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19664 switch (N->getOpcode()) {
19666 llvm_unreachable("Do not know how to custom type legalize this operation!");
19667 case ISD::SIGN_EXTEND_INREG:
19672 // We don't want to expand or promote these.
19679 case ISD::UDIVREM: {
19680 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19681 Results.push_back(V);
19684 case ISD::FP_TO_SINT:
19685 case ISD::FP_TO_UINT: {
19686 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19688 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
19691 std::pair<SDValue,SDValue> Vals =
19692 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19693 SDValue FIST = Vals.first, StackSlot = Vals.second;
19694 if (FIST.getNode()) {
19695 EVT VT = N->getValueType(0);
19696 // Return a load from the stack slot.
19697 if (StackSlot.getNode())
19698 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19699 MachinePointerInfo(),
19700 false, false, false, 0));
19702 Results.push_back(FIST);
19706 case ISD::UINT_TO_FP: {
19707 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19708 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19709 N->getValueType(0) != MVT::v2f32)
19711 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19713 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
19715 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19716 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19717 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
19718 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
19719 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19720 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19723 case ISD::FP_ROUND: {
19724 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19726 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19727 Results.push_back(V);
19730 case ISD::INTRINSIC_W_CHAIN: {
19731 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19733 default : llvm_unreachable("Do not know how to custom type "
19734 "legalize this intrinsic operation!");
19735 case Intrinsic::x86_rdtsc:
19736 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19738 case Intrinsic::x86_rdtscp:
19739 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19741 case Intrinsic::x86_rdpmc:
19742 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19745 case ISD::READCYCLECOUNTER: {
19746 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19749 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19750 EVT T = N->getValueType(0);
19751 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19752 bool Regs64bit = T == MVT::i128;
19753 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19754 SDValue cpInL, cpInH;
19755 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19756 DAG.getConstant(0, HalfT));
19757 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19758 DAG.getConstant(1, HalfT));
19759 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19760 Regs64bit ? X86::RAX : X86::EAX,
19762 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19763 Regs64bit ? X86::RDX : X86::EDX,
19764 cpInH, cpInL.getValue(1));
19765 SDValue swapInL, swapInH;
19766 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19767 DAG.getConstant(0, HalfT));
19768 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19769 DAG.getConstant(1, HalfT));
19770 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19771 Regs64bit ? X86::RBX : X86::EBX,
19772 swapInL, cpInH.getValue(1));
19773 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19774 Regs64bit ? X86::RCX : X86::ECX,
19775 swapInH, swapInL.getValue(1));
19776 SDValue Ops[] = { swapInH.getValue(0),
19778 swapInH.getValue(1) };
19779 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19780 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19781 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19782 X86ISD::LCMPXCHG8_DAG;
19783 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19784 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19785 Regs64bit ? X86::RAX : X86::EAX,
19786 HalfT, Result.getValue(1));
19787 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19788 Regs64bit ? X86::RDX : X86::EDX,
19789 HalfT, cpOutL.getValue(2));
19790 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19792 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19793 MVT::i32, cpOutH.getValue(2));
19795 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19796 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
19797 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19799 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19800 Results.push_back(Success);
19801 Results.push_back(EFLAGS.getValue(1));
19804 case ISD::ATOMIC_SWAP:
19805 case ISD::ATOMIC_LOAD_ADD:
19806 case ISD::ATOMIC_LOAD_SUB:
19807 case ISD::ATOMIC_LOAD_AND:
19808 case ISD::ATOMIC_LOAD_OR:
19809 case ISD::ATOMIC_LOAD_XOR:
19810 case ISD::ATOMIC_LOAD_NAND:
19811 case ISD::ATOMIC_LOAD_MIN:
19812 case ISD::ATOMIC_LOAD_MAX:
19813 case ISD::ATOMIC_LOAD_UMIN:
19814 case ISD::ATOMIC_LOAD_UMAX:
19815 case ISD::ATOMIC_LOAD: {
19816 // Delegate to generic TypeLegalization. Situations we can really handle
19817 // should have already been dealt with by AtomicExpandPass.cpp.
19820 case ISD::BITCAST: {
19821 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19822 EVT DstVT = N->getValueType(0);
19823 EVT SrcVT = N->getOperand(0)->getValueType(0);
19825 if (SrcVT != MVT::f64 ||
19826 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19829 unsigned NumElts = DstVT.getVectorNumElements();
19830 EVT SVT = DstVT.getVectorElementType();
19831 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19832 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19833 MVT::v2f64, N->getOperand(0));
19834 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
19836 if (ExperimentalVectorWideningLegalization) {
19837 // If we are legalizing vectors by widening, we already have the desired
19838 // legal vector type, just return it.
19839 Results.push_back(ToVecInt);
19843 SmallVector<SDValue, 8> Elts;
19844 for (unsigned i = 0, e = NumElts; i != e; ++i)
19845 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19846 ToVecInt, DAG.getIntPtrConstant(i)));
19848 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19853 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19855 default: return nullptr;
19856 case X86ISD::BSF: return "X86ISD::BSF";
19857 case X86ISD::BSR: return "X86ISD::BSR";
19858 case X86ISD::SHLD: return "X86ISD::SHLD";
19859 case X86ISD::SHRD: return "X86ISD::SHRD";
19860 case X86ISD::FAND: return "X86ISD::FAND";
19861 case X86ISD::FANDN: return "X86ISD::FANDN";
19862 case X86ISD::FOR: return "X86ISD::FOR";
19863 case X86ISD::FXOR: return "X86ISD::FXOR";
19864 case X86ISD::FSRL: return "X86ISD::FSRL";
19865 case X86ISD::FILD: return "X86ISD::FILD";
19866 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19867 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19868 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19869 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19870 case X86ISD::FLD: return "X86ISD::FLD";
19871 case X86ISD::FST: return "X86ISD::FST";
19872 case X86ISD::CALL: return "X86ISD::CALL";
19873 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19874 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19875 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19876 case X86ISD::BT: return "X86ISD::BT";
19877 case X86ISD::CMP: return "X86ISD::CMP";
19878 case X86ISD::COMI: return "X86ISD::COMI";
19879 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19880 case X86ISD::CMPM: return "X86ISD::CMPM";
19881 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19882 case X86ISD::SETCC: return "X86ISD::SETCC";
19883 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19884 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19885 case X86ISD::CMOV: return "X86ISD::CMOV";
19886 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19887 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19888 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19889 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19890 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19891 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19892 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19893 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19894 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19895 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19896 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19897 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19898 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19899 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19900 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19901 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19902 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19903 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19904 case X86ISD::HADD: return "X86ISD::HADD";
19905 case X86ISD::HSUB: return "X86ISD::HSUB";
19906 case X86ISD::FHADD: return "X86ISD::FHADD";
19907 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19908 case X86ISD::UMAX: return "X86ISD::UMAX";
19909 case X86ISD::UMIN: return "X86ISD::UMIN";
19910 case X86ISD::SMAX: return "X86ISD::SMAX";
19911 case X86ISD::SMIN: return "X86ISD::SMIN";
19912 case X86ISD::FMAX: return "X86ISD::FMAX";
19913 case X86ISD::FMIN: return "X86ISD::FMIN";
19914 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19915 case X86ISD::FMINC: return "X86ISD::FMINC";
19916 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19917 case X86ISD::FRCP: return "X86ISD::FRCP";
19918 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19919 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19920 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19921 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19922 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19923 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19924 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19925 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19926 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19927 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19928 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19929 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19930 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19931 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19932 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19933 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19934 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19935 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
19936 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19937 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19938 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19939 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19940 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19941 case X86ISD::VSHL: return "X86ISD::VSHL";
19942 case X86ISD::VSRL: return "X86ISD::VSRL";
19943 case X86ISD::VSRA: return "X86ISD::VSRA";
19944 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19945 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19946 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19947 case X86ISD::CMPP: return "X86ISD::CMPP";
19948 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19949 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19950 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19951 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19952 case X86ISD::ADD: return "X86ISD::ADD";
19953 case X86ISD::SUB: return "X86ISD::SUB";
19954 case X86ISD::ADC: return "X86ISD::ADC";
19955 case X86ISD::SBB: return "X86ISD::SBB";
19956 case X86ISD::SMUL: return "X86ISD::SMUL";
19957 case X86ISD::UMUL: return "X86ISD::UMUL";
19958 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19959 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19960 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19961 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19962 case X86ISD::INC: return "X86ISD::INC";
19963 case X86ISD::DEC: return "X86ISD::DEC";
19964 case X86ISD::OR: return "X86ISD::OR";
19965 case X86ISD::XOR: return "X86ISD::XOR";
19966 case X86ISD::AND: return "X86ISD::AND";
19967 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19968 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19969 case X86ISD::PTEST: return "X86ISD::PTEST";
19970 case X86ISD::TESTP: return "X86ISD::TESTP";
19971 case X86ISD::TESTM: return "X86ISD::TESTM";
19972 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19973 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19974 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19975 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19976 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19977 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19978 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19979 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19980 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19981 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19982 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19983 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19984 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19985 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19986 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19987 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19988 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19989 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19990 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19991 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19992 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19993 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19994 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19995 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
19996 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19997 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19998 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19999 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20000 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20001 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20002 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20003 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20004 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20005 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20006 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20007 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20008 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20009 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20010 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
20011 case X86ISD::SAHF: return "X86ISD::SAHF";
20012 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20013 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20014 case X86ISD::FMADD: return "X86ISD::FMADD";
20015 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20016 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20017 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20018 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20019 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20020 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20021 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20022 case X86ISD::XTEST: return "X86ISD::XTEST";
20023 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20024 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20025 case X86ISD::SELECT: return "X86ISD::SELECT";
20029 // isLegalAddressingMode - Return true if the addressing mode represented
20030 // by AM is legal for this target, for a load/store of the specified type.
20031 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
20033 // X86 supports extremely general addressing modes.
20034 CodeModel::Model M = getTargetMachine().getCodeModel();
20035 Reloc::Model R = getTargetMachine().getRelocationModel();
20037 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20038 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20043 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20045 // If a reference to this global requires an extra load, we can't fold it.
20046 if (isGlobalStubReference(GVFlags))
20049 // If BaseGV requires a register for the PIC base, we cannot also have a
20050 // BaseReg specified.
20051 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20054 // If lower 4G is not available, then we must use rip-relative addressing.
20055 if ((M != CodeModel::Small || R != Reloc::Static) &&
20056 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20060 switch (AM.Scale) {
20066 // These scales always work.
20071 // These scales are formed with basereg+scalereg. Only accept if there is
20076 default: // Other stuff never works.
20083 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20084 unsigned Bits = Ty->getScalarSizeInBits();
20086 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20087 // particularly cheaper than those without.
20091 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20092 // variable shifts just as cheap as scalar ones.
20093 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20096 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20097 // fully general vector.
20101 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20102 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20104 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20105 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20106 return NumBits1 > NumBits2;
20109 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20110 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20113 if (!isTypeLegal(EVT::getEVT(Ty1)))
20116 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20118 // Assuming the caller doesn't have a zeroext or signext return parameter,
20119 // truncation all the way down to i1 is valid.
20123 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20124 return isInt<32>(Imm);
20127 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20128 // Can also use sub to handle negated immediates.
20129 return isInt<32>(Imm);
20132 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20133 if (!VT1.isInteger() || !VT2.isInteger())
20135 unsigned NumBits1 = VT1.getSizeInBits();
20136 unsigned NumBits2 = VT2.getSizeInBits();
20137 return NumBits1 > NumBits2;
20140 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20141 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20142 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20145 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20146 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20147 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20150 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20151 EVT VT1 = Val.getValueType();
20152 if (isZExtFree(VT1, VT2))
20155 if (Val.getOpcode() != ISD::LOAD)
20158 if (!VT1.isSimple() || !VT1.isInteger() ||
20159 !VT2.isSimple() || !VT2.isInteger())
20162 switch (VT1.getSimpleVT().SimpleTy) {
20167 // X86 has 8, 16, and 32-bit zero-extending loads.
20175 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20176 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
20179 VT = VT.getScalarType();
20181 if (!VT.isSimple())
20184 switch (VT.getSimpleVT().SimpleTy) {
20195 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20196 // i16 instructions are longer (0x66 prefix) and potentially slower.
20197 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20200 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20201 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20202 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20203 /// are assumed to be legal.
20205 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20207 if (!VT.isSimple())
20210 MVT SVT = VT.getSimpleVT();
20212 // Very little shuffling can be done for 64-bit vectors right now.
20213 if (VT.getSizeInBits() == 64)
20216 // This is an experimental legality test that is tailored to match the
20217 // legality test of the experimental lowering more closely. They are gated
20218 // separately to ease testing of performance differences.
20219 if (ExperimentalVectorShuffleLegality)
20220 // We only care that the types being shuffled are legal. The lowering can
20221 // handle any possible shuffle mask that results.
20222 return isTypeLegal(SVT);
20224 // If this is a single-input shuffle with no 128 bit lane crossings we can
20225 // lower it into pshufb.
20226 if ((SVT.is128BitVector() && Subtarget->hasSSSE3()) ||
20227 (SVT.is256BitVector() && Subtarget->hasInt256())) {
20228 bool isLegal = true;
20229 for (unsigned I = 0, E = M.size(); I != E; ++I) {
20230 if (M[I] >= (int)SVT.getVectorNumElements() ||
20231 ShuffleCrosses128bitLane(SVT, I, M[I])) {
20240 // FIXME: blends, shifts.
20241 return (SVT.getVectorNumElements() == 2 ||
20242 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
20243 isMOVLMask(M, SVT) ||
20244 isCommutedMOVLMask(M, SVT) ||
20245 isMOVHLPSMask(M, SVT) ||
20246 isSHUFPMask(M, SVT) ||
20247 isSHUFPMask(M, SVT, /* Commuted */ true) ||
20248 isPSHUFDMask(M, SVT) ||
20249 isPSHUFDMask(M, SVT, /* SecondOperand */ true) ||
20250 isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
20251 isPSHUFLWMask(M, SVT, Subtarget->hasInt256()) ||
20252 isPALIGNRMask(M, SVT, Subtarget) ||
20253 isUNPCKLMask(M, SVT, Subtarget->hasInt256()) ||
20254 isUNPCKHMask(M, SVT, Subtarget->hasInt256()) ||
20255 isUNPCKL_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20256 isUNPCKH_v_undef_Mask(M, SVT, Subtarget->hasInt256()) ||
20257 isBlendMask(M, SVT, Subtarget->hasSSE41(), Subtarget->hasInt256()) ||
20258 (Subtarget->hasSSE41() && isINSERTPSMask(M, SVT)));
20262 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20264 if (!VT.isSimple())
20267 MVT SVT = VT.getSimpleVT();
20269 // This is an experimental legality test that is tailored to match the
20270 // legality test of the experimental lowering more closely. They are gated
20271 // separately to ease testing of performance differences.
20272 if (ExperimentalVectorShuffleLegality)
20273 // The new vector shuffle lowering is very good at managing zero-inputs.
20274 return isShuffleMaskLegal(Mask, VT);
20276 unsigned NumElts = SVT.getVectorNumElements();
20277 // FIXME: This collection of masks seems suspect.
20280 if (NumElts == 4 && SVT.is128BitVector()) {
20281 return (isMOVLMask(Mask, SVT) ||
20282 isCommutedMOVLMask(Mask, SVT, true) ||
20283 isSHUFPMask(Mask, SVT) ||
20284 isSHUFPMask(Mask, SVT, /* Commuted */ true) ||
20285 isBlendMask(Mask, SVT, Subtarget->hasSSE41(),
20286 Subtarget->hasInt256()));
20291 //===----------------------------------------------------------------------===//
20292 // X86 Scheduler Hooks
20293 //===----------------------------------------------------------------------===//
20295 /// Utility function to emit xbegin specifying the start of an RTM region.
20296 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20297 const TargetInstrInfo *TII) {
20298 DebugLoc DL = MI->getDebugLoc();
20300 const BasicBlock *BB = MBB->getBasicBlock();
20301 MachineFunction::iterator I = MBB;
20304 // For the v = xbegin(), we generate
20315 MachineBasicBlock *thisMBB = MBB;
20316 MachineFunction *MF = MBB->getParent();
20317 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20318 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20319 MF->insert(I, mainMBB);
20320 MF->insert(I, sinkMBB);
20322 // Transfer the remainder of BB and its successor edges to sinkMBB.
20323 sinkMBB->splice(sinkMBB->begin(), MBB,
20324 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20325 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20329 // # fallthrough to mainMBB
20330 // # abortion to sinkMBB
20331 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20332 thisMBB->addSuccessor(mainMBB);
20333 thisMBB->addSuccessor(sinkMBB);
20337 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20338 mainMBB->addSuccessor(sinkMBB);
20341 // EAX is live into the sinkMBB
20342 sinkMBB->addLiveIn(X86::EAX);
20343 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20344 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20347 MI->eraseFromParent();
20351 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20352 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20353 // in the .td file.
20354 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20355 const TargetInstrInfo *TII) {
20357 switch (MI->getOpcode()) {
20358 default: llvm_unreachable("illegal opcode!");
20359 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20360 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20361 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20362 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20363 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20364 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20365 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20366 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20369 DebugLoc dl = MI->getDebugLoc();
20370 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20372 unsigned NumArgs = MI->getNumOperands();
20373 for (unsigned i = 1; i < NumArgs; ++i) {
20374 MachineOperand &Op = MI->getOperand(i);
20375 if (!(Op.isReg() && Op.isImplicit()))
20376 MIB.addOperand(Op);
20378 if (MI->hasOneMemOperand())
20379 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20381 BuildMI(*BB, MI, dl,
20382 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20383 .addReg(X86::XMM0);
20385 MI->eraseFromParent();
20389 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20390 // defs in an instruction pattern
20391 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20392 const TargetInstrInfo *TII) {
20394 switch (MI->getOpcode()) {
20395 default: llvm_unreachable("illegal opcode!");
20396 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20397 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20398 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20399 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20400 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20401 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20402 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20403 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20406 DebugLoc dl = MI->getDebugLoc();
20407 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20409 unsigned NumArgs = MI->getNumOperands(); // remove the results
20410 for (unsigned i = 1; i < NumArgs; ++i) {
20411 MachineOperand &Op = MI->getOperand(i);
20412 if (!(Op.isReg() && Op.isImplicit()))
20413 MIB.addOperand(Op);
20415 if (MI->hasOneMemOperand())
20416 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20418 BuildMI(*BB, MI, dl,
20419 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20422 MI->eraseFromParent();
20426 static MachineBasicBlock * EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20427 const TargetInstrInfo *TII,
20428 const X86Subtarget* Subtarget) {
20429 DebugLoc dl = MI->getDebugLoc();
20431 // Address into RAX/EAX, other two args into ECX, EDX.
20432 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20433 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20434 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20435 for (int i = 0; i < X86::AddrNumOperands; ++i)
20436 MIB.addOperand(MI->getOperand(i));
20438 unsigned ValOps = X86::AddrNumOperands;
20439 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20440 .addReg(MI->getOperand(ValOps).getReg());
20441 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20442 .addReg(MI->getOperand(ValOps+1).getReg());
20444 // The instruction doesn't actually take any operands though.
20445 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20447 MI->eraseFromParent(); // The pseudo is gone now.
20451 MachineBasicBlock *
20452 X86TargetLowering::EmitVAARG64WithCustomInserter(
20454 MachineBasicBlock *MBB) const {
20455 // Emit va_arg instruction on X86-64.
20457 // Operands to this pseudo-instruction:
20458 // 0 ) Output : destination address (reg)
20459 // 1-5) Input : va_list address (addr, i64mem)
20460 // 6 ) ArgSize : Size (in bytes) of vararg type
20461 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20462 // 8 ) Align : Alignment of type
20463 // 9 ) EFLAGS (implicit-def)
20465 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20466 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
20468 unsigned DestReg = MI->getOperand(0).getReg();
20469 MachineOperand &Base = MI->getOperand(1);
20470 MachineOperand &Scale = MI->getOperand(2);
20471 MachineOperand &Index = MI->getOperand(3);
20472 MachineOperand &Disp = MI->getOperand(4);
20473 MachineOperand &Segment = MI->getOperand(5);
20474 unsigned ArgSize = MI->getOperand(6).getImm();
20475 unsigned ArgMode = MI->getOperand(7).getImm();
20476 unsigned Align = MI->getOperand(8).getImm();
20478 // Memory Reference
20479 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20480 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20481 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20483 // Machine Information
20484 const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
20485 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20486 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20487 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20488 DebugLoc DL = MI->getDebugLoc();
20490 // struct va_list {
20493 // i64 overflow_area (address)
20494 // i64 reg_save_area (address)
20496 // sizeof(va_list) = 24
20497 // alignment(va_list) = 8
20499 unsigned TotalNumIntRegs = 6;
20500 unsigned TotalNumXMMRegs = 8;
20501 bool UseGPOffset = (ArgMode == 1);
20502 bool UseFPOffset = (ArgMode == 2);
20503 unsigned MaxOffset = TotalNumIntRegs * 8 +
20504 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20506 /* Align ArgSize to a multiple of 8 */
20507 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20508 bool NeedsAlign = (Align > 8);
20510 MachineBasicBlock *thisMBB = MBB;
20511 MachineBasicBlock *overflowMBB;
20512 MachineBasicBlock *offsetMBB;
20513 MachineBasicBlock *endMBB;
20515 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20516 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20517 unsigned OffsetReg = 0;
20519 if (!UseGPOffset && !UseFPOffset) {
20520 // If we only pull from the overflow region, we don't create a branch.
20521 // We don't need to alter control flow.
20522 OffsetDestReg = 0; // unused
20523 OverflowDestReg = DestReg;
20525 offsetMBB = nullptr;
20526 overflowMBB = thisMBB;
20529 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20530 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20531 // If not, pull from overflow_area. (branch to overflowMBB)
20536 // offsetMBB overflowMBB
20541 // Registers for the PHI in endMBB
20542 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20543 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20545 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20546 MachineFunction *MF = MBB->getParent();
20547 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20548 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20549 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20551 MachineFunction::iterator MBBIter = MBB;
20554 // Insert the new basic blocks
20555 MF->insert(MBBIter, offsetMBB);
20556 MF->insert(MBBIter, overflowMBB);
20557 MF->insert(MBBIter, endMBB);
20559 // Transfer the remainder of MBB and its successor edges to endMBB.
20560 endMBB->splice(endMBB->begin(), thisMBB,
20561 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20562 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20564 // Make offsetMBB and overflowMBB successors of thisMBB
20565 thisMBB->addSuccessor(offsetMBB);
20566 thisMBB->addSuccessor(overflowMBB);
20568 // endMBB is a successor of both offsetMBB and overflowMBB
20569 offsetMBB->addSuccessor(endMBB);
20570 overflowMBB->addSuccessor(endMBB);
20572 // Load the offset value into a register
20573 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20574 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20578 .addDisp(Disp, UseFPOffset ? 4 : 0)
20579 .addOperand(Segment)
20580 .setMemRefs(MMOBegin, MMOEnd);
20582 // Check if there is enough room left to pull this argument.
20583 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20585 .addImm(MaxOffset + 8 - ArgSizeA8);
20587 // Branch to "overflowMBB" if offset >= max
20588 // Fall through to "offsetMBB" otherwise
20589 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20590 .addMBB(overflowMBB);
20593 // In offsetMBB, emit code to use the reg_save_area.
20595 assert(OffsetReg != 0);
20597 // Read the reg_save_area address.
20598 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20599 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20604 .addOperand(Segment)
20605 .setMemRefs(MMOBegin, MMOEnd);
20607 // Zero-extend the offset
20608 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20609 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20612 .addImm(X86::sub_32bit);
20614 // Add the offset to the reg_save_area to get the final address.
20615 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20616 .addReg(OffsetReg64)
20617 .addReg(RegSaveReg);
20619 // Compute the offset for the next argument
20620 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20621 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20623 .addImm(UseFPOffset ? 16 : 8);
20625 // Store it back into the va_list.
20626 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20630 .addDisp(Disp, UseFPOffset ? 4 : 0)
20631 .addOperand(Segment)
20632 .addReg(NextOffsetReg)
20633 .setMemRefs(MMOBegin, MMOEnd);
20636 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20641 // Emit code to use overflow area
20644 // Load the overflow_area address into a register.
20645 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20646 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20651 .addOperand(Segment)
20652 .setMemRefs(MMOBegin, MMOEnd);
20654 // If we need to align it, do so. Otherwise, just copy the address
20655 // to OverflowDestReg.
20657 // Align the overflow address
20658 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20659 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20661 // aligned_addr = (addr + (align-1)) & ~(align-1)
20662 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20663 .addReg(OverflowAddrReg)
20666 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20668 .addImm(~(uint64_t)(Align-1));
20670 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20671 .addReg(OverflowAddrReg);
20674 // Compute the next overflow address after this argument.
20675 // (the overflow address should be kept 8-byte aligned)
20676 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20677 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20678 .addReg(OverflowDestReg)
20679 .addImm(ArgSizeA8);
20681 // Store the new overflow address.
20682 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20687 .addOperand(Segment)
20688 .addReg(NextAddrReg)
20689 .setMemRefs(MMOBegin, MMOEnd);
20691 // If we branched, emit the PHI to the front of endMBB.
20693 BuildMI(*endMBB, endMBB->begin(), DL,
20694 TII->get(X86::PHI), DestReg)
20695 .addReg(OffsetDestReg).addMBB(offsetMBB)
20696 .addReg(OverflowDestReg).addMBB(overflowMBB);
20699 // Erase the pseudo instruction
20700 MI->eraseFromParent();
20705 MachineBasicBlock *
20706 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20708 MachineBasicBlock *MBB) const {
20709 // Emit code to save XMM registers to the stack. The ABI says that the
20710 // number of registers to save is given in %al, so it's theoretically
20711 // possible to do an indirect jump trick to avoid saving all of them,
20712 // however this code takes a simpler approach and just executes all
20713 // of the stores if %al is non-zero. It's less code, and it's probably
20714 // easier on the hardware branch predictor, and stores aren't all that
20715 // expensive anyway.
20717 // Create the new basic blocks. One block contains all the XMM stores,
20718 // and one block is the final destination regardless of whether any
20719 // stores were performed.
20720 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20721 MachineFunction *F = MBB->getParent();
20722 MachineFunction::iterator MBBIter = MBB;
20724 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20725 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20726 F->insert(MBBIter, XMMSaveMBB);
20727 F->insert(MBBIter, EndMBB);
20729 // Transfer the remainder of MBB and its successor edges to EndMBB.
20730 EndMBB->splice(EndMBB->begin(), MBB,
20731 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20732 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20734 // The original block will now fall through to the XMM save block.
20735 MBB->addSuccessor(XMMSaveMBB);
20736 // The XMMSaveMBB will fall through to the end block.
20737 XMMSaveMBB->addSuccessor(EndMBB);
20739 // Now add the instructions.
20740 const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
20741 DebugLoc DL = MI->getDebugLoc();
20743 unsigned CountReg = MI->getOperand(0).getReg();
20744 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20745 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20747 if (!Subtarget->isTargetWin64()) {
20748 // If %al is 0, branch around the XMM save block.
20749 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20750 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20751 MBB->addSuccessor(EndMBB);
20754 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20755 // that was just emitted, but clearly shouldn't be "saved".
20756 assert((MI->getNumOperands() <= 3 ||
20757 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20758 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20759 && "Expected last argument to be EFLAGS");
20760 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20761 // In the XMM save block, save all the XMM argument registers.
20762 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20763 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20764 MachineMemOperand *MMO =
20765 F->getMachineMemOperand(
20766 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
20767 MachineMemOperand::MOStore,
20768 /*Size=*/16, /*Align=*/16);
20769 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20770 .addFrameIndex(RegSaveFrameIndex)
20771 .addImm(/*Scale=*/1)
20772 .addReg(/*IndexReg=*/0)
20773 .addImm(/*Disp=*/Offset)
20774 .addReg(/*Segment=*/0)
20775 .addReg(MI->getOperand(i).getReg())
20776 .addMemOperand(MMO);
20779 MI->eraseFromParent(); // The pseudo instruction is gone now.
20784 // The EFLAGS operand of SelectItr might be missing a kill marker
20785 // because there were multiple uses of EFLAGS, and ISel didn't know
20786 // which to mark. Figure out whether SelectItr should have had a
20787 // kill marker, and set it if it should. Returns the correct kill
20789 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20790 MachineBasicBlock* BB,
20791 const TargetRegisterInfo* TRI) {
20792 // Scan forward through BB for a use/def of EFLAGS.
20793 MachineBasicBlock::iterator miI(std::next(SelectItr));
20794 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20795 const MachineInstr& mi = *miI;
20796 if (mi.readsRegister(X86::EFLAGS))
20798 if (mi.definesRegister(X86::EFLAGS))
20799 break; // Should have kill-flag - update below.
20802 // If we hit the end of the block, check whether EFLAGS is live into a
20804 if (miI == BB->end()) {
20805 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20806 sEnd = BB->succ_end();
20807 sItr != sEnd; ++sItr) {
20808 MachineBasicBlock* succ = *sItr;
20809 if (succ->isLiveIn(X86::EFLAGS))
20814 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20815 // out. SelectMI should have a kill flag on EFLAGS.
20816 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20820 MachineBasicBlock *
20821 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20822 MachineBasicBlock *BB) const {
20823 const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
20824 DebugLoc DL = MI->getDebugLoc();
20826 // To "insert" a SELECT_CC instruction, we actually have to insert the
20827 // diamond control-flow pattern. The incoming instruction knows the
20828 // destination vreg to set, the condition code register to branch on, the
20829 // true/false values to select between, and a branch opcode to use.
20830 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20831 MachineFunction::iterator It = BB;
20837 // cmpTY ccX, r1, r2
20839 // fallthrough --> copy0MBB
20840 MachineBasicBlock *thisMBB = BB;
20841 MachineFunction *F = BB->getParent();
20842 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20843 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20844 F->insert(It, copy0MBB);
20845 F->insert(It, sinkMBB);
20847 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20848 // live into the sink and copy blocks.
20849 const TargetRegisterInfo *TRI =
20850 BB->getParent()->getSubtarget().getRegisterInfo();
20851 if (!MI->killsRegister(X86::EFLAGS) &&
20852 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
20853 copy0MBB->addLiveIn(X86::EFLAGS);
20854 sinkMBB->addLiveIn(X86::EFLAGS);
20857 // Transfer the remainder of BB and its successor edges to sinkMBB.
20858 sinkMBB->splice(sinkMBB->begin(), BB,
20859 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20860 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20862 // Add the true and fallthrough blocks as its successors.
20863 BB->addSuccessor(copy0MBB);
20864 BB->addSuccessor(sinkMBB);
20866 // Create the conditional branch instruction.
20868 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
20869 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20872 // %FalseValue = ...
20873 // # fallthrough to sinkMBB
20874 copy0MBB->addSuccessor(sinkMBB);
20877 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20879 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20880 TII->get(X86::PHI), MI->getOperand(0).getReg())
20881 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
20882 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
20884 MI->eraseFromParent(); // The pseudo instruction is gone now.
20888 MachineBasicBlock *
20889 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20890 MachineBasicBlock *BB) const {
20891 MachineFunction *MF = BB->getParent();
20892 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
20893 DebugLoc DL = MI->getDebugLoc();
20894 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20896 assert(MF->shouldSplitStack());
20898 const bool Is64Bit = Subtarget->is64Bit();
20899 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20901 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20902 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20905 // ... [Till the alloca]
20906 // If stacklet is not large enough, jump to mallocMBB
20909 // Allocate by subtracting from RSP
20910 // Jump to continueMBB
20913 // Allocate by call to runtime
20917 // [rest of original BB]
20920 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20921 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20922 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20924 MachineRegisterInfo &MRI = MF->getRegInfo();
20925 const TargetRegisterClass *AddrRegClass =
20926 getRegClassFor(getPointerTy());
20928 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20929 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20930 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20931 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20932 sizeVReg = MI->getOperand(1).getReg(),
20933 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20935 MachineFunction::iterator MBBIter = BB;
20938 MF->insert(MBBIter, bumpMBB);
20939 MF->insert(MBBIter, mallocMBB);
20940 MF->insert(MBBIter, continueMBB);
20942 continueMBB->splice(continueMBB->begin(), BB,
20943 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20944 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20946 // Add code to the main basic block to check if the stack limit has been hit,
20947 // and if so, jump to mallocMBB otherwise to bumpMBB.
20948 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20949 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20950 .addReg(tmpSPVReg).addReg(sizeVReg);
20951 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20952 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20953 .addReg(SPLimitVReg);
20954 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20956 // bumpMBB simply decreases the stack pointer, since we know the current
20957 // stacklet has enough space.
20958 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20959 .addReg(SPLimitVReg);
20960 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20961 .addReg(SPLimitVReg);
20962 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20964 // Calls into a routine in libgcc to allocate more space from the heap.
20965 const uint32_t *RegMask = MF->getTarget()
20966 .getSubtargetImpl()
20967 ->getRegisterInfo()
20968 ->getCallPreservedMask(CallingConv::C);
20970 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20972 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20973 .addExternalSymbol("__morestack_allocate_stack_space")
20974 .addRegMask(RegMask)
20975 .addReg(X86::RDI, RegState::Implicit)
20976 .addReg(X86::RAX, RegState::ImplicitDefine);
20977 } else if (Is64Bit) {
20978 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20980 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20981 .addExternalSymbol("__morestack_allocate_stack_space")
20982 .addRegMask(RegMask)
20983 .addReg(X86::EDI, RegState::Implicit)
20984 .addReg(X86::EAX, RegState::ImplicitDefine);
20986 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20988 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20989 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20990 .addExternalSymbol("__morestack_allocate_stack_space")
20991 .addRegMask(RegMask)
20992 .addReg(X86::EAX, RegState::ImplicitDefine);
20996 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20999 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21000 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21001 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21003 // Set up the CFG correctly.
21004 BB->addSuccessor(bumpMBB);
21005 BB->addSuccessor(mallocMBB);
21006 mallocMBB->addSuccessor(continueMBB);
21007 bumpMBB->addSuccessor(continueMBB);
21009 // Take care of the PHI nodes.
21010 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21011 MI->getOperand(0).getReg())
21012 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21013 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21015 // Delete the original pseudo instruction.
21016 MI->eraseFromParent();
21019 return continueMBB;
21022 MachineBasicBlock *
21023 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21024 MachineBasicBlock *BB) const {
21025 const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
21026 DebugLoc DL = MI->getDebugLoc();
21028 assert(!Subtarget->isTargetMachO());
21030 // The lowering is pretty easy: we're just emitting the call to _alloca. The
21031 // non-trivial part is impdef of ESP.
21033 if (Subtarget->isTargetWin64()) {
21034 if (Subtarget->isTargetCygMing()) {
21035 // ___chkstk(Mingw64):
21036 // Clobbers R10, R11, RAX and EFLAGS.
21038 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
21039 .addExternalSymbol("___chkstk")
21040 .addReg(X86::RAX, RegState::Implicit)
21041 .addReg(X86::RSP, RegState::Implicit)
21042 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
21043 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
21044 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
21046 // __chkstk(MSVCRT): does not update stack pointer.
21047 // Clobbers R10, R11 and EFLAGS.
21048 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
21049 .addExternalSymbol("__chkstk")
21050 .addReg(X86::RAX, RegState::Implicit)
21051 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
21052 // RAX has the offset to be subtracted from RSP.
21053 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
21058 const char *StackProbeSymbol = (Subtarget->isTargetKnownWindowsMSVC() ||
21059 Subtarget->isTargetWindowsItanium())
21063 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
21064 .addExternalSymbol(StackProbeSymbol)
21065 .addReg(X86::EAX, RegState::Implicit)
21066 .addReg(X86::ESP, RegState::Implicit)
21067 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
21068 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
21069 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
21072 MI->eraseFromParent(); // The pseudo instruction is gone now.
21076 MachineBasicBlock *
21077 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21078 MachineBasicBlock *BB) const {
21079 // This is pretty easy. We're taking the value that we received from
21080 // our load from the relocation, sticking it in either RDI (x86-64)
21081 // or EAX and doing an indirect call. The return value will then
21082 // be in the normal return register.
21083 MachineFunction *F = BB->getParent();
21084 const X86InstrInfo *TII =
21085 static_cast<const X86InstrInfo *>(F->getSubtarget().getInstrInfo());
21086 DebugLoc DL = MI->getDebugLoc();
21088 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21089 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21091 // Get a register mask for the lowered call.
21092 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21093 // proper register mask.
21094 const uint32_t *RegMask = F->getTarget()
21095 .getSubtargetImpl()
21096 ->getRegisterInfo()
21097 ->getCallPreservedMask(CallingConv::C);
21098 if (Subtarget->is64Bit()) {
21099 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21100 TII->get(X86::MOV64rm), X86::RDI)
21102 .addImm(0).addReg(0)
21103 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21104 MI->getOperand(3).getTargetFlags())
21106 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21107 addDirectMem(MIB, X86::RDI);
21108 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21109 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21110 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21111 TII->get(X86::MOV32rm), X86::EAX)
21113 .addImm(0).addReg(0)
21114 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21115 MI->getOperand(3).getTargetFlags())
21117 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21118 addDirectMem(MIB, X86::EAX);
21119 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21121 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21122 TII->get(X86::MOV32rm), X86::EAX)
21123 .addReg(TII->getGlobalBaseReg(F))
21124 .addImm(0).addReg(0)
21125 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21126 MI->getOperand(3).getTargetFlags())
21128 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21129 addDirectMem(MIB, X86::EAX);
21130 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21133 MI->eraseFromParent(); // The pseudo instruction is gone now.
21137 MachineBasicBlock *
21138 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21139 MachineBasicBlock *MBB) const {
21140 DebugLoc DL = MI->getDebugLoc();
21141 MachineFunction *MF = MBB->getParent();
21142 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
21143 MachineRegisterInfo &MRI = MF->getRegInfo();
21145 const BasicBlock *BB = MBB->getBasicBlock();
21146 MachineFunction::iterator I = MBB;
21149 // Memory Reference
21150 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21151 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21154 unsigned MemOpndSlot = 0;
21156 unsigned CurOp = 0;
21158 DstReg = MI->getOperand(CurOp++).getReg();
21159 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21160 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21161 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21162 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21164 MemOpndSlot = CurOp;
21166 MVT PVT = getPointerTy();
21167 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21168 "Invalid Pointer Size!");
21170 // For v = setjmp(buf), we generate
21173 // buf[LabelOffset] = restoreMBB
21174 // SjLjSetup restoreMBB
21180 // v = phi(main, restore)
21183 // if base pointer being used, load it from frame
21186 MachineBasicBlock *thisMBB = MBB;
21187 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21188 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21189 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21190 MF->insert(I, mainMBB);
21191 MF->insert(I, sinkMBB);
21192 MF->push_back(restoreMBB);
21194 MachineInstrBuilder MIB;
21196 // Transfer the remainder of BB and its successor edges to sinkMBB.
21197 sinkMBB->splice(sinkMBB->begin(), MBB,
21198 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21199 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21202 unsigned PtrStoreOpc = 0;
21203 unsigned LabelReg = 0;
21204 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21205 Reloc::Model RM = MF->getTarget().getRelocationModel();
21206 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21207 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21209 // Prepare IP either in reg or imm.
21210 if (!UseImmLabel) {
21211 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21212 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21213 LabelReg = MRI.createVirtualRegister(PtrRC);
21214 if (Subtarget->is64Bit()) {
21215 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21219 .addMBB(restoreMBB)
21222 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21223 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21224 .addReg(XII->getGlobalBaseReg(MF))
21227 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21231 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21233 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21234 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21235 if (i == X86::AddrDisp)
21236 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21238 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21241 MIB.addReg(LabelReg);
21243 MIB.addMBB(restoreMBB);
21244 MIB.setMemRefs(MMOBegin, MMOEnd);
21246 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21247 .addMBB(restoreMBB);
21249 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
21250 MF->getSubtarget().getRegisterInfo());
21251 MIB.addRegMask(RegInfo->getNoPreservedMask());
21252 thisMBB->addSuccessor(mainMBB);
21253 thisMBB->addSuccessor(restoreMBB);
21257 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21258 mainMBB->addSuccessor(sinkMBB);
21261 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21262 TII->get(X86::PHI), DstReg)
21263 .addReg(mainDstReg).addMBB(mainMBB)
21264 .addReg(restoreDstReg).addMBB(restoreMBB);
21267 if (RegInfo->hasBasePointer(*MF)) {
21268 const X86Subtarget &STI = MF->getTarget().getSubtarget<X86Subtarget>();
21269 const bool Uses64BitFramePtr = STI.isTarget64BitLP64() || STI.isTargetNaCl64();
21270 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21271 X86FI->setRestoreBasePointer(MF);
21272 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21273 unsigned BasePtr = RegInfo->getBaseRegister();
21274 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21275 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21276 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21277 .setMIFlag(MachineInstr::FrameSetup);
21279 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21280 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21281 restoreMBB->addSuccessor(sinkMBB);
21283 MI->eraseFromParent();
21287 MachineBasicBlock *
21288 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21289 MachineBasicBlock *MBB) const {
21290 DebugLoc DL = MI->getDebugLoc();
21291 MachineFunction *MF = MBB->getParent();
21292 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
21293 MachineRegisterInfo &MRI = MF->getRegInfo();
21295 // Memory Reference
21296 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21297 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21299 MVT PVT = getPointerTy();
21300 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21301 "Invalid Pointer Size!");
21303 const TargetRegisterClass *RC =
21304 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21305 unsigned Tmp = MRI.createVirtualRegister(RC);
21306 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21307 const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
21308 MF->getSubtarget().getRegisterInfo());
21309 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21310 unsigned SP = RegInfo->getStackRegister();
21312 MachineInstrBuilder MIB;
21314 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21315 const int64_t SPOffset = 2 * PVT.getStoreSize();
21317 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21318 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21321 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21322 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21323 MIB.addOperand(MI->getOperand(i));
21324 MIB.setMemRefs(MMOBegin, MMOEnd);
21326 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21327 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21328 if (i == X86::AddrDisp)
21329 MIB.addDisp(MI->getOperand(i), LabelOffset);
21331 MIB.addOperand(MI->getOperand(i));
21333 MIB.setMemRefs(MMOBegin, MMOEnd);
21335 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21336 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21337 if (i == X86::AddrDisp)
21338 MIB.addDisp(MI->getOperand(i), SPOffset);
21340 MIB.addOperand(MI->getOperand(i));
21342 MIB.setMemRefs(MMOBegin, MMOEnd);
21344 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21346 MI->eraseFromParent();
21350 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21351 // accumulator loops. Writing back to the accumulator allows the coalescer
21352 // to remove extra copies in the loop.
21353 MachineBasicBlock *
21354 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21355 MachineBasicBlock *MBB) const {
21356 MachineOperand &AddendOp = MI->getOperand(3);
21358 // Bail out early if the addend isn't a register - we can't switch these.
21359 if (!AddendOp.isReg())
21362 MachineFunction &MF = *MBB->getParent();
21363 MachineRegisterInfo &MRI = MF.getRegInfo();
21365 // Check whether the addend is defined by a PHI:
21366 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21367 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21368 if (!AddendDef.isPHI())
21371 // Look for the following pattern:
21373 // %addend = phi [%entry, 0], [%loop, %result]
21375 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21379 // %addend = phi [%entry, 0], [%loop, %result]
21381 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21383 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21384 assert(AddendDef.getOperand(i).isReg());
21385 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21386 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21387 if (&PHISrcInst == MI) {
21388 // Found a matching instruction.
21389 unsigned NewFMAOpc = 0;
21390 switch (MI->getOpcode()) {
21391 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21392 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21393 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21394 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21395 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21396 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21397 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21398 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21399 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21400 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21401 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21402 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21403 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21404 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21405 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21406 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21407 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21408 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21409 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21410 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21412 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21413 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21414 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21415 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21416 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21417 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21418 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21419 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21420 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21421 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21422 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21423 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21424 default: llvm_unreachable("Unrecognized FMA variant.");
21427 const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
21428 MachineInstrBuilder MIB =
21429 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21430 .addOperand(MI->getOperand(0))
21431 .addOperand(MI->getOperand(3))
21432 .addOperand(MI->getOperand(2))
21433 .addOperand(MI->getOperand(1));
21434 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21435 MI->eraseFromParent();
21442 MachineBasicBlock *
21443 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21444 MachineBasicBlock *BB) const {
21445 switch (MI->getOpcode()) {
21446 default: llvm_unreachable("Unexpected instr type to insert");
21447 case X86::TAILJMPd64:
21448 case X86::TAILJMPr64:
21449 case X86::TAILJMPm64:
21450 llvm_unreachable("TAILJMP64 would not be touched here.");
21451 case X86::TCRETURNdi64:
21452 case X86::TCRETURNri64:
21453 case X86::TCRETURNmi64:
21455 case X86::WIN_ALLOCA:
21456 return EmitLoweredWinAlloca(MI, BB);
21457 case X86::SEG_ALLOCA_32:
21458 case X86::SEG_ALLOCA_64:
21459 return EmitLoweredSegAlloca(MI, BB);
21460 case X86::TLSCall_32:
21461 case X86::TLSCall_64:
21462 return EmitLoweredTLSCall(MI, BB);
21463 case X86::CMOV_GR8:
21464 case X86::CMOV_FR32:
21465 case X86::CMOV_FR64:
21466 case X86::CMOV_V4F32:
21467 case X86::CMOV_V2F64:
21468 case X86::CMOV_V2I64:
21469 case X86::CMOV_V8F32:
21470 case X86::CMOV_V4F64:
21471 case X86::CMOV_V4I64:
21472 case X86::CMOV_V16F32:
21473 case X86::CMOV_V8F64:
21474 case X86::CMOV_V8I64:
21475 case X86::CMOV_GR16:
21476 case X86::CMOV_GR32:
21477 case X86::CMOV_RFP32:
21478 case X86::CMOV_RFP64:
21479 case X86::CMOV_RFP80:
21480 return EmitLoweredSelect(MI, BB);
21482 case X86::FP32_TO_INT16_IN_MEM:
21483 case X86::FP32_TO_INT32_IN_MEM:
21484 case X86::FP32_TO_INT64_IN_MEM:
21485 case X86::FP64_TO_INT16_IN_MEM:
21486 case X86::FP64_TO_INT32_IN_MEM:
21487 case X86::FP64_TO_INT64_IN_MEM:
21488 case X86::FP80_TO_INT16_IN_MEM:
21489 case X86::FP80_TO_INT32_IN_MEM:
21490 case X86::FP80_TO_INT64_IN_MEM: {
21491 MachineFunction *F = BB->getParent();
21492 const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
21493 DebugLoc DL = MI->getDebugLoc();
21495 // Change the floating point control register to use "round towards zero"
21496 // mode when truncating to an integer value.
21497 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21498 addFrameReference(BuildMI(*BB, MI, DL,
21499 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21501 // Load the old value of the high byte of the control word...
21503 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21504 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21507 // Set the high part to be round to zero...
21508 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21511 // Reload the modified control word now...
21512 addFrameReference(BuildMI(*BB, MI, DL,
21513 TII->get(X86::FLDCW16m)), CWFrameIdx);
21515 // Restore the memory image of control word to original value
21516 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21519 // Get the X86 opcode to use.
21521 switch (MI->getOpcode()) {
21522 default: llvm_unreachable("illegal opcode!");
21523 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21524 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21525 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21526 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21527 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21528 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21529 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21530 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21531 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21535 MachineOperand &Op = MI->getOperand(0);
21537 AM.BaseType = X86AddressMode::RegBase;
21538 AM.Base.Reg = Op.getReg();
21540 AM.BaseType = X86AddressMode::FrameIndexBase;
21541 AM.Base.FrameIndex = Op.getIndex();
21543 Op = MI->getOperand(1);
21545 AM.Scale = Op.getImm();
21546 Op = MI->getOperand(2);
21548 AM.IndexReg = Op.getImm();
21549 Op = MI->getOperand(3);
21550 if (Op.isGlobal()) {
21551 AM.GV = Op.getGlobal();
21553 AM.Disp = Op.getImm();
21555 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21556 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21558 // Reload the original control word now.
21559 addFrameReference(BuildMI(*BB, MI, DL,
21560 TII->get(X86::FLDCW16m)), CWFrameIdx);
21562 MI->eraseFromParent(); // The pseudo instruction is gone now.
21565 // String/text processing lowering.
21566 case X86::PCMPISTRM128REG:
21567 case X86::VPCMPISTRM128REG:
21568 case X86::PCMPISTRM128MEM:
21569 case X86::VPCMPISTRM128MEM:
21570 case X86::PCMPESTRM128REG:
21571 case X86::VPCMPESTRM128REG:
21572 case X86::PCMPESTRM128MEM:
21573 case X86::VPCMPESTRM128MEM:
21574 assert(Subtarget->hasSSE42() &&
21575 "Target must have SSE4.2 or AVX features enabled");
21576 return EmitPCMPSTRM(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
21578 // String/text processing lowering.
21579 case X86::PCMPISTRIREG:
21580 case X86::VPCMPISTRIREG:
21581 case X86::PCMPISTRIMEM:
21582 case X86::VPCMPISTRIMEM:
21583 case X86::PCMPESTRIREG:
21584 case X86::VPCMPESTRIREG:
21585 case X86::PCMPESTRIMEM:
21586 case X86::VPCMPESTRIMEM:
21587 assert(Subtarget->hasSSE42() &&
21588 "Target must have SSE4.2 or AVX features enabled");
21589 return EmitPCMPSTRI(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
21591 // Thread synchronization.
21593 return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(),
21598 return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
21600 case X86::VASTART_SAVE_XMM_REGS:
21601 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21603 case X86::VAARG_64:
21604 return EmitVAARG64WithCustomInserter(MI, BB);
21606 case X86::EH_SjLj_SetJmp32:
21607 case X86::EH_SjLj_SetJmp64:
21608 return emitEHSjLjSetJmp(MI, BB);
21610 case X86::EH_SjLj_LongJmp32:
21611 case X86::EH_SjLj_LongJmp64:
21612 return emitEHSjLjLongJmp(MI, BB);
21614 case TargetOpcode::STATEPOINT:
21615 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21616 // this point in the process. We diverge later.
21617 return emitPatchPoint(MI, BB);
21619 case TargetOpcode::STACKMAP:
21620 case TargetOpcode::PATCHPOINT:
21621 return emitPatchPoint(MI, BB);
21623 case X86::VFMADDPDr213r:
21624 case X86::VFMADDPSr213r:
21625 case X86::VFMADDSDr213r:
21626 case X86::VFMADDSSr213r:
21627 case X86::VFMSUBPDr213r:
21628 case X86::VFMSUBPSr213r:
21629 case X86::VFMSUBSDr213r:
21630 case X86::VFMSUBSSr213r:
21631 case X86::VFNMADDPDr213r:
21632 case X86::VFNMADDPSr213r:
21633 case X86::VFNMADDSDr213r:
21634 case X86::VFNMADDSSr213r:
21635 case X86::VFNMSUBPDr213r:
21636 case X86::VFNMSUBPSr213r:
21637 case X86::VFNMSUBSDr213r:
21638 case X86::VFNMSUBSSr213r:
21639 case X86::VFMADDSUBPDr213r:
21640 case X86::VFMADDSUBPSr213r:
21641 case X86::VFMSUBADDPDr213r:
21642 case X86::VFMSUBADDPSr213r:
21643 case X86::VFMADDPDr213rY:
21644 case X86::VFMADDPSr213rY:
21645 case X86::VFMSUBPDr213rY:
21646 case X86::VFMSUBPSr213rY:
21647 case X86::VFNMADDPDr213rY:
21648 case X86::VFNMADDPSr213rY:
21649 case X86::VFNMSUBPDr213rY:
21650 case X86::VFNMSUBPSr213rY:
21651 case X86::VFMADDSUBPDr213rY:
21652 case X86::VFMADDSUBPSr213rY:
21653 case X86::VFMSUBADDPDr213rY:
21654 case X86::VFMSUBADDPSr213rY:
21655 return emitFMA3Instr(MI, BB);
21659 //===----------------------------------------------------------------------===//
21660 // X86 Optimization Hooks
21661 //===----------------------------------------------------------------------===//
21663 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21666 const SelectionDAG &DAG,
21667 unsigned Depth) const {
21668 unsigned BitWidth = KnownZero.getBitWidth();
21669 unsigned Opc = Op.getOpcode();
21670 assert((Opc >= ISD::BUILTIN_OP_END ||
21671 Opc == ISD::INTRINSIC_WO_CHAIN ||
21672 Opc == ISD::INTRINSIC_W_CHAIN ||
21673 Opc == ISD::INTRINSIC_VOID) &&
21674 "Should use MaskedValueIsZero if you don't know whether Op"
21675 " is a target node!");
21677 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21691 // These nodes' second result is a boolean.
21692 if (Op.getResNo() == 0)
21695 case X86ISD::SETCC:
21696 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21698 case ISD::INTRINSIC_WO_CHAIN: {
21699 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21700 unsigned NumLoBits = 0;
21703 case Intrinsic::x86_sse_movmsk_ps:
21704 case Intrinsic::x86_avx_movmsk_ps_256:
21705 case Intrinsic::x86_sse2_movmsk_pd:
21706 case Intrinsic::x86_avx_movmsk_pd_256:
21707 case Intrinsic::x86_mmx_pmovmskb:
21708 case Intrinsic::x86_sse2_pmovmskb_128:
21709 case Intrinsic::x86_avx2_pmovmskb: {
21710 // High bits of movmskp{s|d}, pmovmskb are known zero.
21712 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21713 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21714 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21715 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21716 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21717 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21718 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21719 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21721 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21730 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21732 const SelectionDAG &,
21733 unsigned Depth) const {
21734 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21735 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21736 return Op.getValueType().getScalarType().getSizeInBits();
21742 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21743 /// node is a GlobalAddress + offset.
21744 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21745 const GlobalValue* &GA,
21746 int64_t &Offset) const {
21747 if (N->getOpcode() == X86ISD::Wrapper) {
21748 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21749 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21750 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21754 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21757 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21758 /// same as extracting the high 128-bit part of 256-bit vector and then
21759 /// inserting the result into the low part of a new 256-bit vector
21760 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21761 EVT VT = SVOp->getValueType(0);
21762 unsigned NumElems = VT.getVectorNumElements();
21764 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21765 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21766 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21767 SVOp->getMaskElt(j) >= 0)
21773 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21774 /// same as extracting the low 128-bit part of 256-bit vector and then
21775 /// inserting the result into the high part of a new 256-bit vector
21776 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21777 EVT VT = SVOp->getValueType(0);
21778 unsigned NumElems = VT.getVectorNumElements();
21780 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21781 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21782 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21783 SVOp->getMaskElt(j) >= 0)
21789 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21790 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21791 TargetLowering::DAGCombinerInfo &DCI,
21792 const X86Subtarget* Subtarget) {
21794 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21795 SDValue V1 = SVOp->getOperand(0);
21796 SDValue V2 = SVOp->getOperand(1);
21797 EVT VT = SVOp->getValueType(0);
21798 unsigned NumElems = VT.getVectorNumElements();
21800 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21801 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21805 // V UNDEF BUILD_VECTOR UNDEF
21807 // CONCAT_VECTOR CONCAT_VECTOR
21810 // RESULT: V + zero extended
21812 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21813 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21814 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21817 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21820 // To match the shuffle mask, the first half of the mask should
21821 // be exactly the first vector, and all the rest a splat with the
21822 // first element of the second one.
21823 for (unsigned i = 0; i != NumElems/2; ++i)
21824 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21825 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21828 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21829 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21830 if (Ld->hasNUsesOfValue(1, 0)) {
21831 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21832 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21834 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21836 Ld->getPointerInfo(),
21837 Ld->getAlignment(),
21838 false/*isVolatile*/, true/*ReadMem*/,
21839 false/*WriteMem*/);
21841 // Make sure the newly-created LOAD is in the same position as Ld in
21842 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21843 // and update uses of Ld's output chain to use the TokenFactor.
21844 if (Ld->hasAnyUseOfValue(1)) {
21845 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21846 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21847 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21848 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21849 SDValue(ResNode.getNode(), 1));
21852 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
21856 // Emit a zeroed vector and insert the desired subvector on its
21858 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21859 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21860 return DCI.CombineTo(N, InsV);
21863 //===--------------------------------------------------------------------===//
21864 // Combine some shuffles into subvector extracts and inserts:
21867 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21868 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21869 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21870 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21871 return DCI.CombineTo(N, InsV);
21874 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21875 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21876 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21877 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21878 return DCI.CombineTo(N, InsV);
21884 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21887 /// This is the leaf of the recursive combinine below. When we have found some
21888 /// chain of single-use x86 shuffle instructions and accumulated the combined
21889 /// shuffle mask represented by them, this will try to pattern match that mask
21890 /// into either a single instruction if there is a special purpose instruction
21891 /// for this operation, or into a PSHUFB instruction which is a fully general
21892 /// instruction but should only be used to replace chains over a certain depth.
21893 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21894 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21895 TargetLowering::DAGCombinerInfo &DCI,
21896 const X86Subtarget *Subtarget) {
21897 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21899 // Find the operand that enters the chain. Note that multiple uses are OK
21900 // here, we're not going to remove the operand we find.
21901 SDValue Input = Op.getOperand(0);
21902 while (Input.getOpcode() == ISD::BITCAST)
21903 Input = Input.getOperand(0);
21905 MVT VT = Input.getSimpleValueType();
21906 MVT RootVT = Root.getSimpleValueType();
21909 // Just remove no-op shuffle masks.
21910 if (Mask.size() == 1) {
21911 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
21916 // Use the float domain if the operand type is a floating point type.
21917 bool FloatDomain = VT.isFloatingPoint();
21919 // For floating point shuffles, we don't have free copies in the shuffle
21920 // instructions or the ability to load as part of the instruction, so
21921 // canonicalize their shuffles to UNPCK or MOV variants.
21923 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21924 // vectors because it can have a load folded into it that UNPCK cannot. This
21925 // doesn't preclude something switching to the shorter encoding post-RA.
21927 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
21928 bool Lo = Mask.equals(0, 0);
21931 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21932 // is no slower than UNPCKLPD but has the option to fold the input operand
21933 // into even an unaligned memory load.
21934 if (Lo && Subtarget->hasSSE3()) {
21935 Shuffle = X86ISD::MOVDDUP;
21936 ShuffleVT = MVT::v2f64;
21938 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21939 // than the UNPCK variants.
21940 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21941 ShuffleVT = MVT::v4f32;
21943 if (Depth == 1 && Root->getOpcode() == Shuffle)
21944 return false; // Nothing to do!
21945 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
21946 DCI.AddToWorklist(Op.getNode());
21947 if (Shuffle == X86ISD::MOVDDUP)
21948 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21950 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21951 DCI.AddToWorklist(Op.getNode());
21952 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
21956 if (Subtarget->hasSSE3() &&
21957 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
21958 bool Lo = Mask.equals(0, 0, 2, 2);
21959 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21960 MVT ShuffleVT = MVT::v4f32;
21961 if (Depth == 1 && Root->getOpcode() == Shuffle)
21962 return false; // Nothing to do!
21963 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
21964 DCI.AddToWorklist(Op.getNode());
21965 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21966 DCI.AddToWorklist(Op.getNode());
21967 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
21971 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
21972 bool Lo = Mask.equals(0, 0, 1, 1);
21973 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21974 MVT ShuffleVT = MVT::v4f32;
21975 if (Depth == 1 && Root->getOpcode() == Shuffle)
21976 return false; // Nothing to do!
21977 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
21978 DCI.AddToWorklist(Op.getNode());
21979 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21980 DCI.AddToWorklist(Op.getNode());
21981 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
21987 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21988 // variants as none of these have single-instruction variants that are
21989 // superior to the UNPCK formulation.
21990 if (!FloatDomain &&
21991 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
21992 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
21993 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
21994 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
21996 bool Lo = Mask[0] == 0;
21997 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21998 if (Depth == 1 && Root->getOpcode() == Shuffle)
21999 return false; // Nothing to do!
22001 switch (Mask.size()) {
22003 ShuffleVT = MVT::v8i16;
22006 ShuffleVT = MVT::v16i8;
22009 llvm_unreachable("Impossible mask size!");
22011 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
22012 DCI.AddToWorklist(Op.getNode());
22013 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22014 DCI.AddToWorklist(Op.getNode());
22015 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22020 // Don't try to re-form single instruction chains under any circumstances now
22021 // that we've done encoding canonicalization for them.
22025 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22026 // can replace them with a single PSHUFB instruction profitably. Intel's
22027 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22028 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22029 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22030 SmallVector<SDValue, 16> PSHUFBMask;
22031 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
22032 int Ratio = 16 / Mask.size();
22033 for (unsigned i = 0; i < 16; ++i) {
22034 if (Mask[i / Ratio] == SM_SentinelUndef) {
22035 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22038 int M = Mask[i / Ratio] != SM_SentinelZero
22039 ? Ratio * Mask[i / Ratio] + i % Ratio
22041 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
22043 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
22044 DCI.AddToWorklist(Op.getNode());
22045 SDValue PSHUFBMaskOp =
22046 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
22047 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22048 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
22049 DCI.AddToWorklist(Op.getNode());
22050 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
22055 // Failed to find any combines.
22059 /// \brief Fully generic combining of x86 shuffle instructions.
22061 /// This should be the last combine run over the x86 shuffle instructions. Once
22062 /// they have been fully optimized, this will recursively consider all chains
22063 /// of single-use shuffle instructions, build a generic model of the cumulative
22064 /// shuffle operation, and check for simpler instructions which implement this
22065 /// operation. We use this primarily for two purposes:
22067 /// 1) Collapse generic shuffles to specialized single instructions when
22068 /// equivalent. In most cases, this is just an encoding size win, but
22069 /// sometimes we will collapse multiple generic shuffles into a single
22070 /// special-purpose shuffle.
22071 /// 2) Look for sequences of shuffle instructions with 3 or more total
22072 /// instructions, and replace them with the slightly more expensive SSSE3
22073 /// PSHUFB instruction if available. We do this as the last combining step
22074 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22075 /// a suitable short sequence of other instructions. The PHUFB will either
22076 /// use a register or have to read from memory and so is slightly (but only
22077 /// slightly) more expensive than the other shuffle instructions.
22079 /// Because this is inherently a quadratic operation (for each shuffle in
22080 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22081 /// This should never be an issue in practice as the shuffle lowering doesn't
22082 /// produce sequences of more than 8 instructions.
22084 /// FIXME: We will currently miss some cases where the redundant shuffling
22085 /// would simplify under the threshold for PSHUFB formation because of
22086 /// combine-ordering. To fix this, we should do the redundant instruction
22087 /// combining in this recursive walk.
22088 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22089 ArrayRef<int> RootMask,
22090 int Depth, bool HasPSHUFB,
22092 TargetLowering::DAGCombinerInfo &DCI,
22093 const X86Subtarget *Subtarget) {
22094 // Bound the depth of our recursive combine because this is ultimately
22095 // quadratic in nature.
22099 // Directly rip through bitcasts to find the underlying operand.
22100 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22101 Op = Op.getOperand(0);
22103 MVT VT = Op.getSimpleValueType();
22104 if (!VT.isVector())
22105 return false; // Bail if we hit a non-vector.
22106 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
22107 // version should be added.
22108 if (VT.getSizeInBits() != 128)
22111 assert(Root.getSimpleValueType().isVector() &&
22112 "Shuffles operate on vector types!");
22113 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22114 "Can only combine shuffles of the same vector register size.");
22116 if (!isTargetShuffle(Op.getOpcode()))
22118 SmallVector<int, 16> OpMask;
22120 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22121 // We only can combine unary shuffles which we can decode the mask for.
22122 if (!HaveMask || !IsUnary)
22125 assert(VT.getVectorNumElements() == OpMask.size() &&
22126 "Different mask size from vector size!");
22127 assert(((RootMask.size() > OpMask.size() &&
22128 RootMask.size() % OpMask.size() == 0) ||
22129 (OpMask.size() > RootMask.size() &&
22130 OpMask.size() % RootMask.size() == 0) ||
22131 OpMask.size() == RootMask.size()) &&
22132 "The smaller number of elements must divide the larger.");
22133 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22134 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22135 assert(((RootRatio == 1 && OpRatio == 1) ||
22136 (RootRatio == 1) != (OpRatio == 1)) &&
22137 "Must not have a ratio for both incoming and op masks!");
22139 SmallVector<int, 16> Mask;
22140 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22142 // Merge this shuffle operation's mask into our accumulated mask. Note that
22143 // this shuffle's mask will be the first applied to the input, followed by the
22144 // root mask to get us all the way to the root value arrangement. The reason
22145 // for this order is that we are recursing up the operation chain.
22146 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22147 int RootIdx = i / RootRatio;
22148 if (RootMask[RootIdx] < 0) {
22149 // This is a zero or undef lane, we're done.
22150 Mask.push_back(RootMask[RootIdx]);
22154 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22155 int OpIdx = RootMaskedIdx / OpRatio;
22156 if (OpMask[OpIdx] < 0) {
22157 // The incoming lanes are zero or undef, it doesn't matter which ones we
22159 Mask.push_back(OpMask[OpIdx]);
22163 // Ok, we have non-zero lanes, map them through.
22164 Mask.push_back(OpMask[OpIdx] * OpRatio +
22165 RootMaskedIdx % OpRatio);
22168 // See if we can recurse into the operand to combine more things.
22169 switch (Op.getOpcode()) {
22170 case X86ISD::PSHUFB:
22172 case X86ISD::PSHUFD:
22173 case X86ISD::PSHUFHW:
22174 case X86ISD::PSHUFLW:
22175 if (Op.getOperand(0).hasOneUse() &&
22176 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22177 HasPSHUFB, DAG, DCI, Subtarget))
22181 case X86ISD::UNPCKL:
22182 case X86ISD::UNPCKH:
22183 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
22184 // We can't check for single use, we have to check that this shuffle is the only user.
22185 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22186 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22187 HasPSHUFB, DAG, DCI, Subtarget))
22192 // Minor canonicalization of the accumulated shuffle mask to make it easier
22193 // to match below. All this does is detect masks with squential pairs of
22194 // elements, and shrink them to the half-width mask. It does this in a loop
22195 // so it will reduce the size of the mask to the minimal width mask which
22196 // performs an equivalent shuffle.
22197 SmallVector<int, 16> WidenedMask;
22198 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22199 Mask = std::move(WidenedMask);
22200 WidenedMask.clear();
22203 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22207 /// \brief Get the PSHUF-style mask from PSHUF node.
22209 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22210 /// PSHUF-style masks that can be reused with such instructions.
22211 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22212 SmallVector<int, 4> Mask;
22214 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
22218 switch (N.getOpcode()) {
22219 case X86ISD::PSHUFD:
22221 case X86ISD::PSHUFLW:
22224 case X86ISD::PSHUFHW:
22225 Mask.erase(Mask.begin(), Mask.begin() + 4);
22226 for (int &M : Mask)
22230 llvm_unreachable("No valid shuffle instruction found!");
22234 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22236 /// We walk up the chain and look for a combinable shuffle, skipping over
22237 /// shuffles that we could hoist this shuffle's transformation past without
22238 /// altering anything.
22240 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22242 TargetLowering::DAGCombinerInfo &DCI) {
22243 assert(N.getOpcode() == X86ISD::PSHUFD &&
22244 "Called with something other than an x86 128-bit half shuffle!");
22247 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22248 // of the shuffles in the chain so that we can form a fresh chain to replace
22250 SmallVector<SDValue, 8> Chain;
22251 SDValue V = N.getOperand(0);
22252 for (; V.hasOneUse(); V = V.getOperand(0)) {
22253 switch (V.getOpcode()) {
22255 return SDValue(); // Nothing combined!
22258 // Skip bitcasts as we always know the type for the target specific
22262 case X86ISD::PSHUFD:
22263 // Found another dword shuffle.
22266 case X86ISD::PSHUFLW:
22267 // Check that the low words (being shuffled) are the identity in the
22268 // dword shuffle, and the high words are self-contained.
22269 if (Mask[0] != 0 || Mask[1] != 1 ||
22270 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22273 Chain.push_back(V);
22276 case X86ISD::PSHUFHW:
22277 // Check that the high words (being shuffled) are the identity in the
22278 // dword shuffle, and the low words are self-contained.
22279 if (Mask[2] != 2 || Mask[3] != 3 ||
22280 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22283 Chain.push_back(V);
22286 case X86ISD::UNPCKL:
22287 case X86ISD::UNPCKH:
22288 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22289 // shuffle into a preceding word shuffle.
22290 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
22293 // Search for a half-shuffle which we can combine with.
22294 unsigned CombineOp =
22295 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22296 if (V.getOperand(0) != V.getOperand(1) ||
22297 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22299 Chain.push_back(V);
22300 V = V.getOperand(0);
22302 switch (V.getOpcode()) {
22304 return SDValue(); // Nothing to combine.
22306 case X86ISD::PSHUFLW:
22307 case X86ISD::PSHUFHW:
22308 if (V.getOpcode() == CombineOp)
22311 Chain.push_back(V);
22315 V = V.getOperand(0);
22319 } while (V.hasOneUse());
22322 // Break out of the loop if we break out of the switch.
22326 if (!V.hasOneUse())
22327 // We fell out of the loop without finding a viable combining instruction.
22330 // Merge this node's mask and our incoming mask.
22331 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22332 for (int &M : Mask)
22334 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22335 getV4X86ShuffleImm8ForMask(Mask, DAG));
22337 // Rebuild the chain around this new shuffle.
22338 while (!Chain.empty()) {
22339 SDValue W = Chain.pop_back_val();
22341 if (V.getValueType() != W.getOperand(0).getValueType())
22342 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
22344 switch (W.getOpcode()) {
22346 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22348 case X86ISD::UNPCKL:
22349 case X86ISD::UNPCKH:
22350 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22353 case X86ISD::PSHUFD:
22354 case X86ISD::PSHUFLW:
22355 case X86ISD::PSHUFHW:
22356 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22360 if (V.getValueType() != N.getValueType())
22361 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
22363 // Return the new chain to replace N.
22367 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
22369 /// We walk up the chain, skipping shuffles of the other half and looking
22370 /// through shuffles which switch halves trying to find a shuffle of the same
22371 /// pair of dwords.
22372 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22374 TargetLowering::DAGCombinerInfo &DCI) {
22376 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22377 "Called with something other than an x86 128-bit half shuffle!");
22379 unsigned CombineOpcode = N.getOpcode();
22381 // Walk up a single-use chain looking for a combinable shuffle.
22382 SDValue V = N.getOperand(0);
22383 for (; V.hasOneUse(); V = V.getOperand(0)) {
22384 switch (V.getOpcode()) {
22386 return false; // Nothing combined!
22389 // Skip bitcasts as we always know the type for the target specific
22393 case X86ISD::PSHUFLW:
22394 case X86ISD::PSHUFHW:
22395 if (V.getOpcode() == CombineOpcode)
22398 // Other-half shuffles are no-ops.
22401 // Break out of the loop if we break out of the switch.
22405 if (!V.hasOneUse())
22406 // We fell out of the loop without finding a viable combining instruction.
22409 // Combine away the bottom node as its shuffle will be accumulated into
22410 // a preceding shuffle.
22411 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22413 // Record the old value.
22416 // Merge this node's mask and our incoming mask (adjusted to account for all
22417 // the pshufd instructions encountered).
22418 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22419 for (int &M : Mask)
22421 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22422 getV4X86ShuffleImm8ForMask(Mask, DAG));
22424 // Check that the shuffles didn't cancel each other out. If not, we need to
22425 // combine to the new one.
22427 // Replace the combinable shuffle with the combined one, updating all users
22428 // so that we re-evaluate the chain here.
22429 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22434 /// \brief Try to combine x86 target specific shuffles.
22435 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22436 TargetLowering::DAGCombinerInfo &DCI,
22437 const X86Subtarget *Subtarget) {
22439 MVT VT = N.getSimpleValueType();
22440 SmallVector<int, 4> Mask;
22442 switch (N.getOpcode()) {
22443 case X86ISD::PSHUFD:
22444 case X86ISD::PSHUFLW:
22445 case X86ISD::PSHUFHW:
22446 Mask = getPSHUFShuffleMask(N);
22447 assert(Mask.size() == 4);
22453 // Nuke no-op shuffles that show up after combining.
22454 if (isNoopShuffleMask(Mask))
22455 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22457 // Look for simplifications involving one or two shuffle instructions.
22458 SDValue V = N.getOperand(0);
22459 switch (N.getOpcode()) {
22462 case X86ISD::PSHUFLW:
22463 case X86ISD::PSHUFHW:
22464 assert(VT == MVT::v8i16);
22467 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22468 return SDValue(); // We combined away this shuffle, so we're done.
22470 // See if this reduces to a PSHUFD which is no more expensive and can
22471 // combine with more operations. Note that it has to at least flip the
22472 // dwords as otherwise it would have been removed as a no-op.
22473 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
22474 int DMask[] = {0, 1, 2, 3};
22475 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22476 DMask[DOffset + 0] = DOffset + 1;
22477 DMask[DOffset + 1] = DOffset + 0;
22478 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
22479 DCI.AddToWorklist(V.getNode());
22480 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
22481 getV4X86ShuffleImm8ForMask(DMask, DAG));
22482 DCI.AddToWorklist(V.getNode());
22483 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
22486 // Look for shuffle patterns which can be implemented as a single unpack.
22487 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22488 // only works when we have a PSHUFD followed by two half-shuffles.
22489 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22490 (V.getOpcode() == X86ISD::PSHUFLW ||
22491 V.getOpcode() == X86ISD::PSHUFHW) &&
22492 V.getOpcode() != N.getOpcode() &&
22494 SDValue D = V.getOperand(0);
22495 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22496 D = D.getOperand(0);
22497 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22498 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22499 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22500 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22501 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22503 for (int i = 0; i < 4; ++i) {
22504 WordMask[i + NOffset] = Mask[i] + NOffset;
22505 WordMask[i + VOffset] = VMask[i] + VOffset;
22507 // Map the word mask through the DWord mask.
22509 for (int i = 0; i < 8; ++i)
22510 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22511 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
22512 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
22513 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
22514 std::begin(UnpackLoMask)) ||
22515 std::equal(std::begin(MappedMask), std::end(MappedMask),
22516 std::begin(UnpackHiMask))) {
22517 // We can replace all three shuffles with an unpack.
22518 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
22519 DCI.AddToWorklist(V.getNode());
22520 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22522 DL, MVT::v8i16, V, V);
22529 case X86ISD::PSHUFD:
22530 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22539 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22541 /// We combine this directly on the abstract vector shuffle nodes so it is
22542 /// easier to generically match. We also insert dummy vector shuffle nodes for
22543 /// the operands which explicitly discard the lanes which are unused by this
22544 /// operation to try to flow through the rest of the combiner the fact that
22545 /// they're unused.
22546 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22548 EVT VT = N->getValueType(0);
22550 // We only handle target-independent shuffles.
22551 // FIXME: It would be easy and harmless to use the target shuffle mask
22552 // extraction tool to support more.
22553 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22556 auto *SVN = cast<ShuffleVectorSDNode>(N);
22557 ArrayRef<int> Mask = SVN->getMask();
22558 SDValue V1 = N->getOperand(0);
22559 SDValue V2 = N->getOperand(1);
22561 // We require the first shuffle operand to be the SUB node, and the second to
22562 // be the ADD node.
22563 // FIXME: We should support the commuted patterns.
22564 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22567 // If there are other uses of these operations we can't fold them.
22568 if (!V1->hasOneUse() || !V2->hasOneUse())
22571 // Ensure that both operations have the same operands. Note that we can
22572 // commute the FADD operands.
22573 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22574 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22575 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22578 // We're looking for blends between FADD and FSUB nodes. We insist on these
22579 // nodes being lined up in a specific expected pattern.
22580 if (!(isShuffleEquivalent(Mask, 0, 3) ||
22581 isShuffleEquivalent(Mask, 0, 5, 2, 7) ||
22582 isShuffleEquivalent(Mask, 0, 9, 2, 11, 4, 13, 6, 15)))
22585 // Only specific types are legal at this point, assert so we notice if and
22586 // when these change.
22587 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22588 VT == MVT::v4f64) &&
22589 "Unknown vector type encountered!");
22591 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22594 /// PerformShuffleCombine - Performs several different shuffle combines.
22595 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22596 TargetLowering::DAGCombinerInfo &DCI,
22597 const X86Subtarget *Subtarget) {
22599 SDValue N0 = N->getOperand(0);
22600 SDValue N1 = N->getOperand(1);
22601 EVT VT = N->getValueType(0);
22603 // Don't create instructions with illegal types after legalize types has run.
22604 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22605 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22608 // If we have legalized the vector types, look for blends of FADD and FSUB
22609 // nodes that we can fuse into an ADDSUB node.
22610 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22611 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22614 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22615 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22616 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22617 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22619 // During Type Legalization, when promoting illegal vector types,
22620 // the backend might introduce new shuffle dag nodes and bitcasts.
22622 // This code performs the following transformation:
22623 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22624 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22626 // We do this only if both the bitcast and the BINOP dag nodes have
22627 // one use. Also, perform this transformation only if the new binary
22628 // operation is legal. This is to avoid introducing dag nodes that
22629 // potentially need to be further expanded (or custom lowered) into a
22630 // less optimal sequence of dag nodes.
22631 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22632 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22633 N0.getOpcode() == ISD::BITCAST) {
22634 SDValue BC0 = N0.getOperand(0);
22635 EVT SVT = BC0.getValueType();
22636 unsigned Opcode = BC0.getOpcode();
22637 unsigned NumElts = VT.getVectorNumElements();
22639 if (BC0.hasOneUse() && SVT.isVector() &&
22640 SVT.getVectorNumElements() * 2 == NumElts &&
22641 TLI.isOperationLegal(Opcode, VT)) {
22642 bool CanFold = false;
22654 unsigned SVTNumElts = SVT.getVectorNumElements();
22655 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22656 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22657 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22658 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22659 CanFold = SVOp->getMaskElt(i) < 0;
22662 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
22663 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
22664 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22665 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22670 // Only handle 128 wide vector from here on.
22671 if (!VT.is128BitVector())
22674 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22675 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22676 // consecutive, non-overlapping, and in the right order.
22677 SmallVector<SDValue, 16> Elts;
22678 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22679 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22681 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
22685 if (isTargetShuffle(N->getOpcode())) {
22687 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22688 if (Shuffle.getNode())
22691 // Try recursively combining arbitrary sequences of x86 shuffle
22692 // instructions into higher-order shuffles. We do this after combining
22693 // specific PSHUF instruction sequences into their minimal form so that we
22694 // can evaluate how many specialized shuffle instructions are involved in
22695 // a particular chain.
22696 SmallVector<int, 1> NonceMask; // Just a placeholder.
22697 NonceMask.push_back(0);
22698 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22699 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22701 return SDValue(); // This routine will use CombineTo to replace N.
22707 /// PerformTruncateCombine - Converts truncate operation to
22708 /// a sequence of vector shuffle operations.
22709 /// It is possible when we truncate 256-bit vector to 128-bit vector
22710 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
22711 TargetLowering::DAGCombinerInfo &DCI,
22712 const X86Subtarget *Subtarget) {
22716 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22717 /// specific shuffle of a load can be folded into a single element load.
22718 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22719 /// shuffles have been custom lowered so we need to handle those here.
22720 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22721 TargetLowering::DAGCombinerInfo &DCI) {
22722 if (DCI.isBeforeLegalizeOps())
22725 SDValue InVec = N->getOperand(0);
22726 SDValue EltNo = N->getOperand(1);
22728 if (!isa<ConstantSDNode>(EltNo))
22731 EVT OriginalVT = InVec.getValueType();
22733 if (InVec.getOpcode() == ISD::BITCAST) {
22734 // Don't duplicate a load with other uses.
22735 if (!InVec.hasOneUse())
22737 EVT BCVT = InVec.getOperand(0).getValueType();
22738 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22740 InVec = InVec.getOperand(0);
22743 EVT CurrentVT = InVec.getValueType();
22745 if (!isTargetShuffle(InVec.getOpcode()))
22748 // Don't duplicate a load with other uses.
22749 if (!InVec.hasOneUse())
22752 SmallVector<int, 16> ShuffleMask;
22754 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22755 ShuffleMask, UnaryShuffle))
22758 // Select the input vector, guarding against out of range extract vector.
22759 unsigned NumElems = CurrentVT.getVectorNumElements();
22760 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22761 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22762 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22763 : InVec.getOperand(1);
22765 // If inputs to shuffle are the same for both ops, then allow 2 uses
22766 unsigned AllowedUses = InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22768 if (LdNode.getOpcode() == ISD::BITCAST) {
22769 // Don't duplicate a load with other uses.
22770 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22773 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22774 LdNode = LdNode.getOperand(0);
22777 if (!ISD::isNormalLoad(LdNode.getNode()))
22780 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22782 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22785 EVT EltVT = N->getValueType(0);
22786 // If there's a bitcast before the shuffle, check if the load type and
22787 // alignment is valid.
22788 unsigned Align = LN0->getAlignment();
22789 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22790 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
22791 EltVT.getTypeForEVT(*DAG.getContext()));
22793 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22796 // All checks match so transform back to vector_shuffle so that DAG combiner
22797 // can finish the job
22800 // Create shuffle node taking into account the case that its a unary shuffle
22801 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22802 : InVec.getOperand(1);
22803 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22804 InVec.getOperand(0), Shuffle,
22806 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
22807 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22811 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22812 /// generation and convert it from being a bunch of shuffles and extracts
22813 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22814 /// storing the value and loading scalars back, while for x64 we should
22815 /// use 64-bit extracts and shifts.
22816 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22817 TargetLowering::DAGCombinerInfo &DCI) {
22818 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
22819 if (NewOp.getNode())
22822 SDValue InputVector = N->getOperand(0);
22824 // Detect whether we are trying to convert from mmx to i32 and the bitcast
22825 // from mmx to v2i32 has a single usage.
22826 if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
22827 InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
22828 InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
22829 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22830 N->getValueType(0),
22831 InputVector.getNode()->getOperand(0));
22833 // Only operate on vectors of 4 elements, where the alternative shuffling
22834 // gets to be more expensive.
22835 if (InputVector.getValueType() != MVT::v4i32)
22838 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22839 // single use which is a sign-extend or zero-extend, and all elements are
22841 SmallVector<SDNode *, 4> Uses;
22842 unsigned ExtractedElements = 0;
22843 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22844 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22845 if (UI.getUse().getResNo() != InputVector.getResNo())
22848 SDNode *Extract = *UI;
22849 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22852 if (Extract->getValueType(0) != MVT::i32)
22854 if (!Extract->hasOneUse())
22856 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22857 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22859 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22862 // Record which element was extracted.
22863 ExtractedElements |=
22864 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22866 Uses.push_back(Extract);
22869 // If not all the elements were used, this may not be worthwhile.
22870 if (ExtractedElements != 15)
22873 // Ok, we've now decided to do the transformation.
22874 // If 64-bit shifts are legal, use the extract-shift sequence,
22875 // otherwise bounce the vector off the cache.
22876 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22878 SDLoc dl(InputVector);
22880 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22881 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
22882 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
22883 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22884 DAG.getConstant(0, VecIdxTy));
22885 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22886 DAG.getConstant(1, VecIdxTy));
22888 SDValue ShAmt = DAG.getConstant(32,
22889 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
22890 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22891 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22892 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22893 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22894 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22895 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22897 // Store the value to a temporary stack slot.
22898 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22899 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22900 MachinePointerInfo(), false, false, 0);
22902 EVT ElementType = InputVector.getValueType().getVectorElementType();
22903 unsigned EltSize = ElementType.getSizeInBits() / 8;
22905 // Replace each use (extract) with a load of the appropriate element.
22906 for (unsigned i = 0; i < 4; ++i) {
22907 uint64_t Offset = EltSize * i;
22908 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
22910 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
22911 StackPtr, OffsetVal);
22913 // Load the scalar.
22914 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22915 ScalarAddr, MachinePointerInfo(),
22916 false, false, false, 0);
22921 // Replace the extracts
22922 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22923 UE = Uses.end(); UI != UE; ++UI) {
22924 SDNode *Extract = *UI;
22926 SDValue Idx = Extract->getOperand(1);
22927 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22928 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22931 // The replacement was made in place; don't return anything.
22935 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
22936 static std::pair<unsigned, bool>
22937 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
22938 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
22939 if (!VT.isVector())
22940 return std::make_pair(0, false);
22942 bool NeedSplit = false;
22943 switch (VT.getSimpleVT().SimpleTy) {
22944 default: return std::make_pair(0, false);
22947 if (!Subtarget->hasVLX())
22948 return std::make_pair(0, false);
22952 if (!Subtarget->hasBWI())
22953 return std::make_pair(0, false);
22957 if (!Subtarget->hasAVX512())
22958 return std::make_pair(0, false);
22963 if (!Subtarget->hasAVX2())
22965 if (!Subtarget->hasAVX())
22966 return std::make_pair(0, false);
22971 if (!Subtarget->hasSSE2())
22972 return std::make_pair(0, false);
22975 // SSE2 has only a small subset of the operations.
22976 bool hasUnsigned = Subtarget->hasSSE41() ||
22977 (Subtarget->hasSSE2() && VT == MVT::v16i8);
22978 bool hasSigned = Subtarget->hasSSE41() ||
22979 (Subtarget->hasSSE2() && VT == MVT::v8i16);
22981 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22984 // Check for x CC y ? x : y.
22985 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22986 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22991 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
22994 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
22997 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23000 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23002 // Check for x CC y ? y : x -- a min/max with reversed arms.
23003 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23004 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23009 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
23012 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
23015 Opc = hasSigned ? X86ISD::SMAX : 0; break;
23018 Opc = hasSigned ? X86ISD::SMIN : 0; break;
23022 return std::make_pair(Opc, NeedSplit);
23026 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23027 const X86Subtarget *Subtarget) {
23029 SDValue Cond = N->getOperand(0);
23030 SDValue LHS = N->getOperand(1);
23031 SDValue RHS = N->getOperand(2);
23033 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23034 SDValue CondSrc = Cond->getOperand(0);
23035 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23036 Cond = CondSrc->getOperand(0);
23039 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23042 // A vselect where all conditions and data are constants can be optimized into
23043 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23044 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23045 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23048 unsigned MaskValue = 0;
23049 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23052 MVT VT = N->getSimpleValueType(0);
23053 unsigned NumElems = VT.getVectorNumElements();
23054 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23055 for (unsigned i = 0; i < NumElems; ++i) {
23056 // Be sure we emit undef where we can.
23057 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23058 ShuffleMask[i] = -1;
23060 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23063 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23064 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23066 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23069 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23071 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23072 TargetLowering::DAGCombinerInfo &DCI,
23073 const X86Subtarget *Subtarget) {
23075 SDValue Cond = N->getOperand(0);
23076 // Get the LHS/RHS of the select.
23077 SDValue LHS = N->getOperand(1);
23078 SDValue RHS = N->getOperand(2);
23079 EVT VT = LHS.getValueType();
23080 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23082 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23083 // instructions match the semantics of the common C idiom x<y?x:y but not
23084 // x<=y?x:y, because of how they handle negative zero (which can be
23085 // ignored in unsafe-math mode).
23086 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23087 VT != MVT::f80 && TLI.isTypeLegal(VT) &&
23088 (Subtarget->hasSSE2() ||
23089 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23090 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23092 unsigned Opcode = 0;
23093 // Check for x CC y ? x : y.
23094 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23095 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23099 // Converting this to a min would handle NaNs incorrectly, and swapping
23100 // the operands would cause it to handle comparisons between positive
23101 // and negative zero incorrectly.
23102 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23103 if (!DAG.getTarget().Options.UnsafeFPMath &&
23104 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23106 std::swap(LHS, RHS);
23108 Opcode = X86ISD::FMIN;
23111 // Converting this to a min would handle comparisons between positive
23112 // and negative zero incorrectly.
23113 if (!DAG.getTarget().Options.UnsafeFPMath &&
23114 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23116 Opcode = X86ISD::FMIN;
23119 // Converting this to a min would handle both negative zeros and NaNs
23120 // incorrectly, but we can swap the operands to fix both.
23121 std::swap(LHS, RHS);
23125 Opcode = X86ISD::FMIN;
23129 // Converting this to a max would handle comparisons between positive
23130 // and negative zero incorrectly.
23131 if (!DAG.getTarget().Options.UnsafeFPMath &&
23132 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23134 Opcode = X86ISD::FMAX;
23137 // Converting this to a max would handle NaNs incorrectly, and swapping
23138 // the operands would cause it to handle comparisons between positive
23139 // and negative zero incorrectly.
23140 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23141 if (!DAG.getTarget().Options.UnsafeFPMath &&
23142 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23144 std::swap(LHS, RHS);
23146 Opcode = X86ISD::FMAX;
23149 // Converting this to a max would handle both negative zeros and NaNs
23150 // incorrectly, but we can swap the operands to fix both.
23151 std::swap(LHS, RHS);
23155 Opcode = X86ISD::FMAX;
23158 // Check for x CC y ? y : x -- a min/max with reversed arms.
23159 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23160 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23164 // Converting this to a min would handle comparisons between positive
23165 // and negative zero incorrectly, and swapping the operands would
23166 // cause it to handle NaNs incorrectly.
23167 if (!DAG.getTarget().Options.UnsafeFPMath &&
23168 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23169 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23171 std::swap(LHS, RHS);
23173 Opcode = X86ISD::FMIN;
23176 // Converting this to a min would handle NaNs incorrectly.
23177 if (!DAG.getTarget().Options.UnsafeFPMath &&
23178 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23180 Opcode = X86ISD::FMIN;
23183 // Converting this to a min would handle both negative zeros and NaNs
23184 // incorrectly, but we can swap the operands to fix both.
23185 std::swap(LHS, RHS);
23189 Opcode = X86ISD::FMIN;
23193 // Converting this to a max would handle NaNs incorrectly.
23194 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23196 Opcode = X86ISD::FMAX;
23199 // Converting this to a max would handle comparisons between positive
23200 // and negative zero incorrectly, and swapping the operands would
23201 // cause it to handle NaNs incorrectly.
23202 if (!DAG.getTarget().Options.UnsafeFPMath &&
23203 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23204 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23206 std::swap(LHS, RHS);
23208 Opcode = X86ISD::FMAX;
23211 // Converting this to a max would handle both negative zeros and NaNs
23212 // incorrectly, but we can swap the operands to fix both.
23213 std::swap(LHS, RHS);
23217 Opcode = X86ISD::FMAX;
23223 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23226 EVT CondVT = Cond.getValueType();
23227 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23228 CondVT.getVectorElementType() == MVT::i1) {
23229 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23230 // lowering on KNL. In this case we convert it to
23231 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23232 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23233 // Since SKX these selects have a proper lowering.
23234 EVT OpVT = LHS.getValueType();
23235 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23236 (OpVT.getVectorElementType() == MVT::i8 ||
23237 OpVT.getVectorElementType() == MVT::i16) &&
23238 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23239 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23240 DCI.AddToWorklist(Cond.getNode());
23241 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23244 // If this is a select between two integer constants, try to do some
23246 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23247 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23248 // Don't do this for crazy integer types.
23249 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23250 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23251 // so that TrueC (the true value) is larger than FalseC.
23252 bool NeedsCondInvert = false;
23254 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23255 // Efficiently invertible.
23256 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23257 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23258 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23259 NeedsCondInvert = true;
23260 std::swap(TrueC, FalseC);
23263 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23264 if (FalseC->getAPIntValue() == 0 &&
23265 TrueC->getAPIntValue().isPowerOf2()) {
23266 if (NeedsCondInvert) // Invert the condition if needed.
23267 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23268 DAG.getConstant(1, Cond.getValueType()));
23270 // Zero extend the condition if needed.
23271 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23273 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23274 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23275 DAG.getConstant(ShAmt, MVT::i8));
23278 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23279 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23280 if (NeedsCondInvert) // Invert the condition if needed.
23281 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23282 DAG.getConstant(1, Cond.getValueType()));
23284 // Zero extend the condition if needed.
23285 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23286 FalseC->getValueType(0), Cond);
23287 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23288 SDValue(FalseC, 0));
23291 // Optimize cases that will turn into an LEA instruction. This requires
23292 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23293 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23294 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23295 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23297 bool isFastMultiplier = false;
23299 switch ((unsigned char)Diff) {
23301 case 1: // result = add base, cond
23302 case 2: // result = lea base( , cond*2)
23303 case 3: // result = lea base(cond, cond*2)
23304 case 4: // result = lea base( , cond*4)
23305 case 5: // result = lea base(cond, cond*4)
23306 case 8: // result = lea base( , cond*8)
23307 case 9: // result = lea base(cond, cond*8)
23308 isFastMultiplier = true;
23313 if (isFastMultiplier) {
23314 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23315 if (NeedsCondInvert) // Invert the condition if needed.
23316 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23317 DAG.getConstant(1, Cond.getValueType()));
23319 // Zero extend the condition if needed.
23320 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23322 // Scale the condition by the difference.
23324 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23325 DAG.getConstant(Diff, Cond.getValueType()));
23327 // Add the base if non-zero.
23328 if (FalseC->getAPIntValue() != 0)
23329 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23330 SDValue(FalseC, 0));
23337 // Canonicalize max and min:
23338 // (x > y) ? x : y -> (x >= y) ? x : y
23339 // (x < y) ? x : y -> (x <= y) ? x : y
23340 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23341 // the need for an extra compare
23342 // against zero. e.g.
23343 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23345 // testl %edi, %edi
23347 // cmovgl %edi, %eax
23351 // cmovsl %eax, %edi
23352 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23353 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23354 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23355 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23360 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23361 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23362 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23363 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23368 // Early exit check
23369 if (!TLI.isTypeLegal(VT))
23372 // Match VSELECTs into subs with unsigned saturation.
23373 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23374 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23375 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23376 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23377 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23379 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23380 // left side invert the predicate to simplify logic below.
23382 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23384 CC = ISD::getSetCCInverse(CC, true);
23385 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23389 if (Other.getNode() && Other->getNumOperands() == 2 &&
23390 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23391 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23392 SDValue CondRHS = Cond->getOperand(1);
23394 // Look for a general sub with unsigned saturation first.
23395 // x >= y ? x-y : 0 --> subus x, y
23396 // x > y ? x-y : 0 --> subus x, y
23397 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23398 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23399 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23401 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23402 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23403 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23404 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23405 // If the RHS is a constant we have to reverse the const
23406 // canonicalization.
23407 // x > C-1 ? x+-C : 0 --> subus x, C
23408 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23409 CondRHSConst->getAPIntValue() ==
23410 (-OpRHSConst->getAPIntValue() - 1))
23411 return DAG.getNode(
23412 X86ISD::SUBUS, DL, VT, OpLHS,
23413 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
23415 // Another special case: If C was a sign bit, the sub has been
23416 // canonicalized into a xor.
23417 // FIXME: Would it be better to use computeKnownBits to determine
23418 // whether it's safe to decanonicalize the xor?
23419 // x s< 0 ? x^C : 0 --> subus x, C
23420 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23421 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23422 OpRHSConst->getAPIntValue().isSignBit())
23423 // Note that we have to rebuild the RHS constant here to ensure we
23424 // don't rely on particular values of undef lanes.
23425 return DAG.getNode(
23426 X86ISD::SUBUS, DL, VT, OpLHS,
23427 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
23432 // Try to match a min/max vector operation.
23433 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
23434 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
23435 unsigned Opc = ret.first;
23436 bool NeedSplit = ret.second;
23438 if (Opc && NeedSplit) {
23439 unsigned NumElems = VT.getVectorNumElements();
23440 // Extract the LHS vectors
23441 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
23442 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
23444 // Extract the RHS vectors
23445 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
23446 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
23448 // Create min/max for each subvector
23449 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
23450 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
23452 // Merge the result
23453 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
23455 return DAG.getNode(Opc, DL, VT, LHS, RHS);
23458 // Simplify vector selection if condition value type matches vselect
23460 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23461 assert(Cond.getValueType().isVector() &&
23462 "vector select expects a vector selector!");
23464 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23465 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23467 // Try invert the condition if true value is not all 1s and false value
23469 if (!TValIsAllOnes && !FValIsAllZeros &&
23470 // Check if the selector will be produced by CMPP*/PCMP*
23471 Cond.getOpcode() == ISD::SETCC &&
23472 // Check if SETCC has already been promoted
23473 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
23474 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23475 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23477 if (TValIsAllZeros || FValIsAllOnes) {
23478 SDValue CC = Cond.getOperand(2);
23479 ISD::CondCode NewCC =
23480 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23481 Cond.getOperand(0).getValueType().isInteger());
23482 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23483 std::swap(LHS, RHS);
23484 TValIsAllOnes = FValIsAllOnes;
23485 FValIsAllZeros = TValIsAllZeros;
23489 if (TValIsAllOnes || FValIsAllZeros) {
23492 if (TValIsAllOnes && FValIsAllZeros)
23494 else if (TValIsAllOnes)
23495 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
23496 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
23497 else if (FValIsAllZeros)
23498 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23499 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
23501 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
23505 // If we know that this node is legal then we know that it is going to be
23506 // matched by one of the SSE/AVX BLEND instructions. These instructions only
23507 // depend on the highest bit in each word. Try to use SimplifyDemandedBits
23508 // to simplify previous instructions.
23509 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23510 !DCI.isBeforeLegalize() &&
23511 // We explicitly check against v8i16 and v16i16 because, although
23512 // they're marked as Custom, they might only be legal when Cond is a
23513 // build_vector of constants. This will be taken care in a later
23515 (TLI.isOperationLegalOrCustom(ISD::VSELECT, VT) && VT != MVT::v16i16 &&
23516 VT != MVT::v8i16) &&
23517 // Don't optimize vector of constants. Those are handled by
23518 // the generic code and all the bits must be properly set for
23519 // the generic optimizer.
23520 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23521 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23523 // Don't optimize vector selects that map to mask-registers.
23527 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23528 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23530 APInt KnownZero, KnownOne;
23531 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23532 DCI.isBeforeLegalizeOps());
23533 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23534 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23536 // If we changed the computation somewhere in the DAG, this change
23537 // will affect all users of Cond.
23538 // Make sure it is fine and update all the nodes so that we do not
23539 // use the generic VSELECT anymore. Otherwise, we may perform
23540 // wrong optimizations as we messed up with the actual expectation
23541 // for the vector boolean values.
23542 if (Cond != TLO.Old) {
23543 // Check all uses of that condition operand to check whether it will be
23544 // consumed by non-BLEND instructions, which may depend on all bits are
23546 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23548 if (I->getOpcode() != ISD::VSELECT)
23549 // TODO: Add other opcodes eventually lowered into BLEND.
23552 // Update all the users of the condition, before committing the change,
23553 // so that the VSELECT optimizations that expect the correct vector
23554 // boolean value will not be triggered.
23555 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23557 DAG.ReplaceAllUsesOfValueWith(
23559 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23560 Cond, I->getOperand(1), I->getOperand(2)));
23561 DCI.CommitTargetLoweringOpt(TLO);
23564 // At this point, only Cond is changed. Change the condition
23565 // just for N to keep the opportunity to optimize all other
23566 // users their own way.
23567 DAG.ReplaceAllUsesOfValueWith(
23569 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23570 TLO.New, N->getOperand(1), N->getOperand(2)));
23575 // We should generate an X86ISD::BLENDI from a vselect if its argument
23576 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23577 // constants. This specific pattern gets generated when we split a
23578 // selector for a 512 bit vector in a machine without AVX512 (but with
23579 // 256-bit vectors), during legalization:
23581 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23583 // Iff we find this pattern and the build_vectors are built from
23584 // constants, we translate the vselect into a shuffle_vector that we
23585 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23586 if ((N->getOpcode() == ISD::VSELECT ||
23587 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23588 !DCI.isBeforeLegalize()) {
23589 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23590 if (Shuffle.getNode())
23597 // Check whether a boolean test is testing a boolean value generated by
23598 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23601 // Simplify the following patterns:
23602 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23603 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23604 // to (Op EFLAGS Cond)
23606 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23607 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23608 // to (Op EFLAGS !Cond)
23610 // where Op could be BRCOND or CMOV.
23612 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23613 // Quit if not CMP and SUB with its value result used.
23614 if (Cmp.getOpcode() != X86ISD::CMP &&
23615 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23618 // Quit if not used as a boolean value.
23619 if (CC != X86::COND_E && CC != X86::COND_NE)
23622 // Check CMP operands. One of them should be 0 or 1 and the other should be
23623 // an SetCC or extended from it.
23624 SDValue Op1 = Cmp.getOperand(0);
23625 SDValue Op2 = Cmp.getOperand(1);
23628 const ConstantSDNode* C = nullptr;
23629 bool needOppositeCond = (CC == X86::COND_E);
23630 bool checkAgainstTrue = false; // Is it a comparison against 1?
23632 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23634 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23636 else // Quit if all operands are not constants.
23639 if (C->getZExtValue() == 1) {
23640 needOppositeCond = !needOppositeCond;
23641 checkAgainstTrue = true;
23642 } else if (C->getZExtValue() != 0)
23643 // Quit if the constant is neither 0 or 1.
23646 bool truncatedToBoolWithAnd = false;
23647 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23648 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23649 SetCC.getOpcode() == ISD::TRUNCATE ||
23650 SetCC.getOpcode() == ISD::AND) {
23651 if (SetCC.getOpcode() == ISD::AND) {
23653 ConstantSDNode *CS;
23654 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23655 CS->getZExtValue() == 1)
23657 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23658 CS->getZExtValue() == 1)
23662 SetCC = SetCC.getOperand(OpIdx);
23663 truncatedToBoolWithAnd = true;
23665 SetCC = SetCC.getOperand(0);
23668 switch (SetCC.getOpcode()) {
23669 case X86ISD::SETCC_CARRY:
23670 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23671 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23672 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23673 // truncated to i1 using 'and'.
23674 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23676 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23677 "Invalid use of SETCC_CARRY!");
23679 case X86ISD::SETCC:
23680 // Set the condition code or opposite one if necessary.
23681 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23682 if (needOppositeCond)
23683 CC = X86::GetOppositeBranchCondition(CC);
23684 return SetCC.getOperand(1);
23685 case X86ISD::CMOV: {
23686 // Check whether false/true value has canonical one, i.e. 0 or 1.
23687 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23688 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23689 // Quit if true value is not a constant.
23692 // Quit if false value is not a constant.
23694 SDValue Op = SetCC.getOperand(0);
23695 // Skip 'zext' or 'trunc' node.
23696 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23697 Op.getOpcode() == ISD::TRUNCATE)
23698 Op = Op.getOperand(0);
23699 // A special case for rdrand/rdseed, where 0 is set if false cond is
23701 if ((Op.getOpcode() != X86ISD::RDRAND &&
23702 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23705 // Quit if false value is not the constant 0 or 1.
23706 bool FValIsFalse = true;
23707 if (FVal && FVal->getZExtValue() != 0) {
23708 if (FVal->getZExtValue() != 1)
23710 // If FVal is 1, opposite cond is needed.
23711 needOppositeCond = !needOppositeCond;
23712 FValIsFalse = false;
23714 // Quit if TVal is not the constant opposite of FVal.
23715 if (FValIsFalse && TVal->getZExtValue() != 1)
23717 if (!FValIsFalse && TVal->getZExtValue() != 0)
23719 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23720 if (needOppositeCond)
23721 CC = X86::GetOppositeBranchCondition(CC);
23722 return SetCC.getOperand(3);
23729 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23730 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23731 TargetLowering::DAGCombinerInfo &DCI,
23732 const X86Subtarget *Subtarget) {
23735 // If the flag operand isn't dead, don't touch this CMOV.
23736 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23739 SDValue FalseOp = N->getOperand(0);
23740 SDValue TrueOp = N->getOperand(1);
23741 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23742 SDValue Cond = N->getOperand(3);
23744 if (CC == X86::COND_E || CC == X86::COND_NE) {
23745 switch (Cond.getOpcode()) {
23749 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23750 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23751 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23757 Flags = checkBoolTestSetCCCombine(Cond, CC);
23758 if (Flags.getNode() &&
23759 // Extra check as FCMOV only supports a subset of X86 cond.
23760 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23761 SDValue Ops[] = { FalseOp, TrueOp,
23762 DAG.getConstant(CC, MVT::i8), Flags };
23763 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23766 // If this is a select between two integer constants, try to do some
23767 // optimizations. Note that the operands are ordered the opposite of SELECT
23769 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23770 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23771 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23772 // larger than FalseC (the false value).
23773 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23774 CC = X86::GetOppositeBranchCondition(CC);
23775 std::swap(TrueC, FalseC);
23776 std::swap(TrueOp, FalseOp);
23779 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23780 // This is efficient for any integer data type (including i8/i16) and
23782 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23783 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23784 DAG.getConstant(CC, MVT::i8), Cond);
23786 // Zero extend the condition if needed.
23787 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23789 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23790 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23791 DAG.getConstant(ShAmt, MVT::i8));
23792 if (N->getNumValues() == 2) // Dead flag value?
23793 return DCI.CombineTo(N, Cond, SDValue());
23797 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23798 // for any integer data type, including i8/i16.
23799 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23800 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23801 DAG.getConstant(CC, MVT::i8), Cond);
23803 // Zero extend the condition if needed.
23804 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23805 FalseC->getValueType(0), Cond);
23806 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23807 SDValue(FalseC, 0));
23809 if (N->getNumValues() == 2) // Dead flag value?
23810 return DCI.CombineTo(N, Cond, SDValue());
23814 // Optimize cases that will turn into an LEA instruction. This requires
23815 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23816 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23817 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23818 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23820 bool isFastMultiplier = false;
23822 switch ((unsigned char)Diff) {
23824 case 1: // result = add base, cond
23825 case 2: // result = lea base( , cond*2)
23826 case 3: // result = lea base(cond, cond*2)
23827 case 4: // result = lea base( , cond*4)
23828 case 5: // result = lea base(cond, cond*4)
23829 case 8: // result = lea base( , cond*8)
23830 case 9: // result = lea base(cond, cond*8)
23831 isFastMultiplier = true;
23836 if (isFastMultiplier) {
23837 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23838 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23839 DAG.getConstant(CC, MVT::i8), Cond);
23840 // Zero extend the condition if needed.
23841 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23843 // Scale the condition by the difference.
23845 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23846 DAG.getConstant(Diff, Cond.getValueType()));
23848 // Add the base if non-zero.
23849 if (FalseC->getAPIntValue() != 0)
23850 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23851 SDValue(FalseC, 0));
23852 if (N->getNumValues() == 2) // Dead flag value?
23853 return DCI.CombineTo(N, Cond, SDValue());
23860 // Handle these cases:
23861 // (select (x != c), e, c) -> select (x != c), e, x),
23862 // (select (x == c), c, e) -> select (x == c), x, e)
23863 // where the c is an integer constant, and the "select" is the combination
23864 // of CMOV and CMP.
23866 // The rationale for this change is that the conditional-move from a constant
23867 // needs two instructions, however, conditional-move from a register needs
23868 // only one instruction.
23870 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23871 // some instruction-combining opportunities. This opt needs to be
23872 // postponed as late as possible.
23874 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23875 // the DCI.xxxx conditions are provided to postpone the optimization as
23876 // late as possible.
23878 ConstantSDNode *CmpAgainst = nullptr;
23879 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23880 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23881 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23883 if (CC == X86::COND_NE &&
23884 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23885 CC = X86::GetOppositeBranchCondition(CC);
23886 std::swap(TrueOp, FalseOp);
23889 if (CC == X86::COND_E &&
23890 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23891 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23892 DAG.getConstant(CC, MVT::i8), Cond };
23893 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23901 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
23902 const X86Subtarget *Subtarget) {
23903 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
23905 default: return SDValue();
23906 // SSE/AVX/AVX2 blend intrinsics.
23907 case Intrinsic::x86_avx2_pblendvb:
23908 case Intrinsic::x86_avx2_pblendw:
23909 case Intrinsic::x86_avx2_pblendd_128:
23910 case Intrinsic::x86_avx2_pblendd_256:
23911 // Don't try to simplify this intrinsic if we don't have AVX2.
23912 if (!Subtarget->hasAVX2())
23915 case Intrinsic::x86_avx_blend_pd_256:
23916 case Intrinsic::x86_avx_blend_ps_256:
23917 case Intrinsic::x86_avx_blendv_pd_256:
23918 case Intrinsic::x86_avx_blendv_ps_256:
23919 // Don't try to simplify this intrinsic if we don't have AVX.
23920 if (!Subtarget->hasAVX())
23923 case Intrinsic::x86_sse41_pblendw:
23924 case Intrinsic::x86_sse41_blendpd:
23925 case Intrinsic::x86_sse41_blendps:
23926 case Intrinsic::x86_sse41_blendvps:
23927 case Intrinsic::x86_sse41_blendvpd:
23928 case Intrinsic::x86_sse41_pblendvb: {
23929 SDValue Op0 = N->getOperand(1);
23930 SDValue Op1 = N->getOperand(2);
23931 SDValue Mask = N->getOperand(3);
23933 // Don't try to simplify this intrinsic if we don't have SSE4.1.
23934 if (!Subtarget->hasSSE41())
23937 // fold (blend A, A, Mask) -> A
23940 // fold (blend A, B, allZeros) -> A
23941 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
23943 // fold (blend A, B, allOnes) -> B
23944 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
23947 // Simplify the case where the mask is a constant i32 value.
23948 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
23949 if (C->isNullValue())
23951 if (C->isAllOnesValue())
23958 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
23959 case Intrinsic::x86_sse2_psrai_w:
23960 case Intrinsic::x86_sse2_psrai_d:
23961 case Intrinsic::x86_avx2_psrai_w:
23962 case Intrinsic::x86_avx2_psrai_d:
23963 case Intrinsic::x86_sse2_psra_w:
23964 case Intrinsic::x86_sse2_psra_d:
23965 case Intrinsic::x86_avx2_psra_w:
23966 case Intrinsic::x86_avx2_psra_d: {
23967 SDValue Op0 = N->getOperand(1);
23968 SDValue Op1 = N->getOperand(2);
23969 EVT VT = Op0.getValueType();
23970 assert(VT.isVector() && "Expected a vector type!");
23972 if (isa<BuildVectorSDNode>(Op1))
23973 Op1 = Op1.getOperand(0);
23975 if (!isa<ConstantSDNode>(Op1))
23978 EVT SVT = VT.getVectorElementType();
23979 unsigned SVTBits = SVT.getSizeInBits();
23981 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
23982 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
23983 uint64_t ShAmt = C.getZExtValue();
23985 // Don't try to convert this shift into a ISD::SRA if the shift
23986 // count is bigger than or equal to the element size.
23987 if (ShAmt >= SVTBits)
23990 // Trivial case: if the shift count is zero, then fold this
23991 // into the first operand.
23995 // Replace this packed shift intrinsic with a target independent
23997 SDValue Splat = DAG.getConstant(C, VT);
23998 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
24003 /// PerformMulCombine - Optimize a single multiply with constant into two
24004 /// in order to implement it with two cheaper instructions, e.g.
24005 /// LEA + SHL, LEA + LEA.
24006 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24007 TargetLowering::DAGCombinerInfo &DCI) {
24008 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24011 EVT VT = N->getValueType(0);
24012 if (VT != MVT::i64)
24015 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24018 uint64_t MulAmt = C->getZExtValue();
24019 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24022 uint64_t MulAmt1 = 0;
24023 uint64_t MulAmt2 = 0;
24024 if ((MulAmt % 9) == 0) {
24026 MulAmt2 = MulAmt / 9;
24027 } else if ((MulAmt % 5) == 0) {
24029 MulAmt2 = MulAmt / 5;
24030 } else if ((MulAmt % 3) == 0) {
24032 MulAmt2 = MulAmt / 3;
24035 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24038 if (isPowerOf2_64(MulAmt2) &&
24039 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24040 // If second multiplifer is pow2, issue it first. We want the multiply by
24041 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24043 std::swap(MulAmt1, MulAmt2);
24046 if (isPowerOf2_64(MulAmt1))
24047 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24048 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
24050 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24051 DAG.getConstant(MulAmt1, VT));
24053 if (isPowerOf2_64(MulAmt2))
24054 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24055 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
24057 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24058 DAG.getConstant(MulAmt2, VT));
24060 // Do not add new nodes to DAG combiner worklist.
24061 DCI.CombineTo(N, NewMul, false);
24066 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24067 SDValue N0 = N->getOperand(0);
24068 SDValue N1 = N->getOperand(1);
24069 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24070 EVT VT = N0.getValueType();
24072 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24073 // since the result of setcc_c is all zero's or all ones.
24074 if (VT.isInteger() && !VT.isVector() &&
24075 N1C && N0.getOpcode() == ISD::AND &&
24076 N0.getOperand(1).getOpcode() == ISD::Constant) {
24077 SDValue N00 = N0.getOperand(0);
24078 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
24079 ((N00.getOpcode() == ISD::ANY_EXTEND ||
24080 N00.getOpcode() == ISD::ZERO_EXTEND) &&
24081 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
24082 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24083 APInt ShAmt = N1C->getAPIntValue();
24084 Mask = Mask.shl(ShAmt);
24086 return DAG.getNode(ISD::AND, SDLoc(N), VT,
24087 N00, DAG.getConstant(Mask, VT));
24091 // Hardware support for vector shifts is sparse which makes us scalarize the
24092 // vector operations in many cases. Also, on sandybridge ADD is faster than
24094 // (shl V, 1) -> add V,V
24095 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24096 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24097 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24098 // We shift all of the values by one. In many cases we do not have
24099 // hardware support for this operation. This is better expressed as an ADD
24101 if (N1SplatC->getZExtValue() == 1)
24102 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24108 /// \brief Returns a vector of 0s if the node in input is a vector logical
24109 /// shift by a constant amount which is known to be bigger than or equal
24110 /// to the vector element size in bits.
24111 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24112 const X86Subtarget *Subtarget) {
24113 EVT VT = N->getValueType(0);
24115 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24116 (!Subtarget->hasInt256() ||
24117 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24120 SDValue Amt = N->getOperand(1);
24122 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24123 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24124 APInt ShiftAmt = AmtSplat->getAPIntValue();
24125 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24127 // SSE2/AVX2 logical shifts always return a vector of 0s
24128 // if the shift amount is bigger than or equal to
24129 // the element size. The constant shift amount will be
24130 // encoded as a 8-bit immediate.
24131 if (ShiftAmt.trunc(8).uge(MaxAmount))
24132 return getZeroVector(VT, Subtarget, DAG, DL);
24138 /// PerformShiftCombine - Combine shifts.
24139 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24140 TargetLowering::DAGCombinerInfo &DCI,
24141 const X86Subtarget *Subtarget) {
24142 if (N->getOpcode() == ISD::SHL) {
24143 SDValue V = PerformSHLCombine(N, DAG);
24144 if (V.getNode()) return V;
24147 if (N->getOpcode() != ISD::SRA) {
24148 // Try to fold this logical shift into a zero vector.
24149 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
24150 if (V.getNode()) return V;
24156 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24157 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24158 // and friends. Likewise for OR -> CMPNEQSS.
24159 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24160 TargetLowering::DAGCombinerInfo &DCI,
24161 const X86Subtarget *Subtarget) {
24164 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24165 // we're requiring SSE2 for both.
24166 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24167 SDValue N0 = N->getOperand(0);
24168 SDValue N1 = N->getOperand(1);
24169 SDValue CMP0 = N0->getOperand(1);
24170 SDValue CMP1 = N1->getOperand(1);
24173 // The SETCCs should both refer to the same CMP.
24174 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24177 SDValue CMP00 = CMP0->getOperand(0);
24178 SDValue CMP01 = CMP0->getOperand(1);
24179 EVT VT = CMP00.getValueType();
24181 if (VT == MVT::f32 || VT == MVT::f64) {
24182 bool ExpectingFlags = false;
24183 // Check for any users that want flags:
24184 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24185 !ExpectingFlags && UI != UE; ++UI)
24186 switch (UI->getOpcode()) {
24191 ExpectingFlags = true;
24193 case ISD::CopyToReg:
24194 case ISD::SIGN_EXTEND:
24195 case ISD::ZERO_EXTEND:
24196 case ISD::ANY_EXTEND:
24200 if (!ExpectingFlags) {
24201 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24202 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24204 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24205 X86::CondCode tmp = cc0;
24210 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24211 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24212 // FIXME: need symbolic constants for these magic numbers.
24213 // See X86ATTInstPrinter.cpp:printSSECC().
24214 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24215 if (Subtarget->hasAVX512()) {
24216 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24217 CMP01, DAG.getConstant(x86cc, MVT::i8));
24218 if (N->getValueType(0) != MVT::i1)
24219 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24223 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24224 CMP00.getValueType(), CMP00, CMP01,
24225 DAG.getConstant(x86cc, MVT::i8));
24227 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24228 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24230 if (is64BitFP && !Subtarget->is64Bit()) {
24231 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24232 // 64-bit integer, since that's not a legal type. Since
24233 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24234 // bits, but can do this little dance to extract the lowest 32 bits
24235 // and work with those going forward.
24236 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24238 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
24240 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24241 Vector32, DAG.getIntPtrConstant(0));
24245 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
24246 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24247 DAG.getConstant(1, IntVT));
24248 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
24249 return OneBitOfTruth;
24257 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24258 /// so it can be folded inside ANDNP.
24259 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24260 EVT VT = N->getValueType(0);
24262 // Match direct AllOnes for 128 and 256-bit vectors
24263 if (ISD::isBuildVectorAllOnes(N))
24266 // Look through a bit convert.
24267 if (N->getOpcode() == ISD::BITCAST)
24268 N = N->getOperand(0).getNode();
24270 // Sometimes the operand may come from a insert_subvector building a 256-bit
24272 if (VT.is256BitVector() &&
24273 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24274 SDValue V1 = N->getOperand(0);
24275 SDValue V2 = N->getOperand(1);
24277 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24278 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24279 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24280 ISD::isBuildVectorAllOnes(V2.getNode()))
24287 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24288 // register. In most cases we actually compare or select YMM-sized registers
24289 // and mixing the two types creates horrible code. This method optimizes
24290 // some of the transition sequences.
24291 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24292 TargetLowering::DAGCombinerInfo &DCI,
24293 const X86Subtarget *Subtarget) {
24294 EVT VT = N->getValueType(0);
24295 if (!VT.is256BitVector())
24298 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24299 N->getOpcode() == ISD::ZERO_EXTEND ||
24300 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24302 SDValue Narrow = N->getOperand(0);
24303 EVT NarrowVT = Narrow->getValueType(0);
24304 if (!NarrowVT.is128BitVector())
24307 if (Narrow->getOpcode() != ISD::XOR &&
24308 Narrow->getOpcode() != ISD::AND &&
24309 Narrow->getOpcode() != ISD::OR)
24312 SDValue N0 = Narrow->getOperand(0);
24313 SDValue N1 = Narrow->getOperand(1);
24316 // The Left side has to be a trunc.
24317 if (N0.getOpcode() != ISD::TRUNCATE)
24320 // The type of the truncated inputs.
24321 EVT WideVT = N0->getOperand(0)->getValueType(0);
24325 // The right side has to be a 'trunc' or a constant vector.
24326 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24327 ConstantSDNode *RHSConstSplat = nullptr;
24328 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24329 RHSConstSplat = RHSBV->getConstantSplatNode();
24330 if (!RHSTrunc && !RHSConstSplat)
24333 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24335 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24338 // Set N0 and N1 to hold the inputs to the new wide operation.
24339 N0 = N0->getOperand(0);
24340 if (RHSConstSplat) {
24341 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24342 SDValue(RHSConstSplat, 0));
24343 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24344 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24345 } else if (RHSTrunc) {
24346 N1 = N1->getOperand(0);
24349 // Generate the wide operation.
24350 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24351 unsigned Opcode = N->getOpcode();
24353 case ISD::ANY_EXTEND:
24355 case ISD::ZERO_EXTEND: {
24356 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24357 APInt Mask = APInt::getAllOnesValue(InBits);
24358 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24359 return DAG.getNode(ISD::AND, DL, VT,
24360 Op, DAG.getConstant(Mask, VT));
24362 case ISD::SIGN_EXTEND:
24363 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24364 Op, DAG.getValueType(NarrowVT));
24366 llvm_unreachable("Unexpected opcode");
24370 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24371 TargetLowering::DAGCombinerInfo &DCI,
24372 const X86Subtarget *Subtarget) {
24373 EVT VT = N->getValueType(0);
24374 if (DCI.isBeforeLegalizeOps())
24377 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
24381 // Create BEXTR instructions
24382 // BEXTR is ((X >> imm) & (2**size-1))
24383 if (VT == MVT::i32 || VT == MVT::i64) {
24384 SDValue N0 = N->getOperand(0);
24385 SDValue N1 = N->getOperand(1);
24388 // Check for BEXTR.
24389 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24390 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24391 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24392 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24393 if (MaskNode && ShiftNode) {
24394 uint64_t Mask = MaskNode->getZExtValue();
24395 uint64_t Shift = ShiftNode->getZExtValue();
24396 if (isMask_64(Mask)) {
24397 uint64_t MaskSize = CountPopulation_64(Mask);
24398 if (Shift + MaskSize <= VT.getSizeInBits())
24399 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24400 DAG.getConstant(Shift | (MaskSize << 8), VT));
24408 // Want to form ANDNP nodes:
24409 // 1) In the hopes of then easily combining them with OR and AND nodes
24410 // to form PBLEND/PSIGN.
24411 // 2) To match ANDN packed intrinsics
24412 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24415 SDValue N0 = N->getOperand(0);
24416 SDValue N1 = N->getOperand(1);
24419 // Check LHS for vnot
24420 if (N0.getOpcode() == ISD::XOR &&
24421 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24422 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24423 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24425 // Check RHS for vnot
24426 if (N1.getOpcode() == ISD::XOR &&
24427 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24428 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24429 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24434 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24435 TargetLowering::DAGCombinerInfo &DCI,
24436 const X86Subtarget *Subtarget) {
24437 if (DCI.isBeforeLegalizeOps())
24440 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
24444 SDValue N0 = N->getOperand(0);
24445 SDValue N1 = N->getOperand(1);
24446 EVT VT = N->getValueType(0);
24448 // look for psign/blend
24449 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24450 if (!Subtarget->hasSSSE3() ||
24451 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24454 // Canonicalize pandn to RHS
24455 if (N0.getOpcode() == X86ISD::ANDNP)
24457 // or (and (m, y), (pandn m, x))
24458 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24459 SDValue Mask = N1.getOperand(0);
24460 SDValue X = N1.getOperand(1);
24462 if (N0.getOperand(0) == Mask)
24463 Y = N0.getOperand(1);
24464 if (N0.getOperand(1) == Mask)
24465 Y = N0.getOperand(0);
24467 // Check to see if the mask appeared in both the AND and ANDNP and
24471 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24472 // Look through mask bitcast.
24473 if (Mask.getOpcode() == ISD::BITCAST)
24474 Mask = Mask.getOperand(0);
24475 if (X.getOpcode() == ISD::BITCAST)
24476 X = X.getOperand(0);
24477 if (Y.getOpcode() == ISD::BITCAST)
24478 Y = Y.getOperand(0);
24480 EVT MaskVT = Mask.getValueType();
24482 // Validate that the Mask operand is a vector sra node.
24483 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24484 // there is no psrai.b
24485 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24486 unsigned SraAmt = ~0;
24487 if (Mask.getOpcode() == ISD::SRA) {
24488 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24489 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24490 SraAmt = AmtConst->getZExtValue();
24491 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24492 SDValue SraC = Mask.getOperand(1);
24493 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24495 if ((SraAmt + 1) != EltBits)
24500 // Now we know we at least have a plendvb with the mask val. See if
24501 // we can form a psignb/w/d.
24502 // psign = x.type == y.type == mask.type && y = sub(0, x);
24503 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24504 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24505 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24506 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24507 "Unsupported VT for PSIGN");
24508 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24509 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
24511 // PBLENDVB only available on SSE 4.1
24512 if (!Subtarget->hasSSE41())
24515 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24517 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
24518 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
24519 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
24520 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24521 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
24525 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24528 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24529 MachineFunction &MF = DAG.getMachineFunction();
24530 bool OptForSize = MF.getFunction()->getAttributes().
24531 hasAttribute(AttributeSet::FunctionIndex, Attribute::OptimizeForSize);
24533 // SHLD/SHRD instructions have lower register pressure, but on some
24534 // platforms they have higher latency than the equivalent
24535 // series of shifts/or that would otherwise be generated.
24536 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24537 // have higher latencies and we are not optimizing for size.
24538 if (!OptForSize && Subtarget->isSHLDSlow())
24541 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24543 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24545 if (!N0.hasOneUse() || !N1.hasOneUse())
24548 SDValue ShAmt0 = N0.getOperand(1);
24549 if (ShAmt0.getValueType() != MVT::i8)
24551 SDValue ShAmt1 = N1.getOperand(1);
24552 if (ShAmt1.getValueType() != MVT::i8)
24554 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24555 ShAmt0 = ShAmt0.getOperand(0);
24556 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24557 ShAmt1 = ShAmt1.getOperand(0);
24560 unsigned Opc = X86ISD::SHLD;
24561 SDValue Op0 = N0.getOperand(0);
24562 SDValue Op1 = N1.getOperand(0);
24563 if (ShAmt0.getOpcode() == ISD::SUB) {
24564 Opc = X86ISD::SHRD;
24565 std::swap(Op0, Op1);
24566 std::swap(ShAmt0, ShAmt1);
24569 unsigned Bits = VT.getSizeInBits();
24570 if (ShAmt1.getOpcode() == ISD::SUB) {
24571 SDValue Sum = ShAmt1.getOperand(0);
24572 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24573 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24574 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24575 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24576 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24577 return DAG.getNode(Opc, DL, VT,
24579 DAG.getNode(ISD::TRUNCATE, DL,
24582 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24583 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24585 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24586 return DAG.getNode(Opc, DL, VT,
24587 N0.getOperand(0), N1.getOperand(0),
24588 DAG.getNode(ISD::TRUNCATE, DL,
24595 // Generate NEG and CMOV for integer abs.
24596 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24597 EVT VT = N->getValueType(0);
24599 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24600 // 8-bit integer abs to NEG and CMOV.
24601 if (VT.isInteger() && VT.getSizeInBits() == 8)
24604 SDValue N0 = N->getOperand(0);
24605 SDValue N1 = N->getOperand(1);
24608 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24609 // and change it to SUB and CMOV.
24610 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24611 N0.getOpcode() == ISD::ADD &&
24612 N0.getOperand(1) == N1 &&
24613 N1.getOpcode() == ISD::SRA &&
24614 N1.getOperand(0) == N0.getOperand(0))
24615 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24616 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24617 // Generate SUB & CMOV.
24618 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24619 DAG.getConstant(0, VT), N0.getOperand(0));
24621 SDValue Ops[] = { N0.getOperand(0), Neg,
24622 DAG.getConstant(X86::COND_GE, MVT::i8),
24623 SDValue(Neg.getNode(), 1) };
24624 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24629 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
24630 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24631 TargetLowering::DAGCombinerInfo &DCI,
24632 const X86Subtarget *Subtarget) {
24633 if (DCI.isBeforeLegalizeOps())
24636 if (Subtarget->hasCMov()) {
24637 SDValue RV = performIntegerAbsCombine(N, DAG);
24645 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24646 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24647 TargetLowering::DAGCombinerInfo &DCI,
24648 const X86Subtarget *Subtarget) {
24649 LoadSDNode *Ld = cast<LoadSDNode>(N);
24650 EVT RegVT = Ld->getValueType(0);
24651 EVT MemVT = Ld->getMemoryVT();
24653 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24655 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24656 // into two 16-byte operations.
24657 ISD::LoadExtType Ext = Ld->getExtensionType();
24658 unsigned Alignment = Ld->getAlignment();
24659 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
24660 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24661 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
24662 unsigned NumElems = RegVT.getVectorNumElements();
24666 SDValue Ptr = Ld->getBasePtr();
24667 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
24669 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24671 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24672 Ld->getPointerInfo(), Ld->isVolatile(),
24673 Ld->isNonTemporal(), Ld->isInvariant(),
24675 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24676 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24677 Ld->getPointerInfo(), Ld->isVolatile(),
24678 Ld->isNonTemporal(), Ld->isInvariant(),
24679 std::min(16U, Alignment));
24680 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24682 Load2.getValue(1));
24684 SDValue NewVec = DAG.getUNDEF(RegVT);
24685 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24686 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24687 return DCI.CombineTo(N, NewVec, TF, true);
24693 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24694 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24695 const X86Subtarget *Subtarget) {
24696 StoreSDNode *St = cast<StoreSDNode>(N);
24697 EVT VT = St->getValue().getValueType();
24698 EVT StVT = St->getMemoryVT();
24700 SDValue StoredVal = St->getOperand(1);
24701 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24703 // If we are saving a concatenation of two XMM registers and 32-byte stores
24704 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24705 unsigned Alignment = St->getAlignment();
24706 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
24707 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24708 StVT == VT && !IsAligned) {
24709 unsigned NumElems = VT.getVectorNumElements();
24713 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24714 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24716 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
24717 SDValue Ptr0 = St->getBasePtr();
24718 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24720 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24721 St->getPointerInfo(), St->isVolatile(),
24722 St->isNonTemporal(), Alignment);
24723 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24724 St->getPointerInfo(), St->isVolatile(),
24725 St->isNonTemporal(),
24726 std::min(16U, Alignment));
24727 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24730 // Optimize trunc store (of multiple scalars) to shuffle and store.
24731 // First, pack all of the elements in one place. Next, store to memory
24732 // in fewer chunks.
24733 if (St->isTruncatingStore() && VT.isVector()) {
24734 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24735 unsigned NumElems = VT.getVectorNumElements();
24736 assert(StVT != VT && "Cannot truncate to the same type");
24737 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24738 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24740 // From, To sizes and ElemCount must be pow of two
24741 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24742 // We are going to use the original vector elt for storing.
24743 // Accumulated smaller vector elements must be a multiple of the store size.
24744 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24746 unsigned SizeRatio = FromSz / ToSz;
24748 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24750 // Create a type on which we perform the shuffle
24751 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24752 StVT.getScalarType(), NumElems*SizeRatio);
24754 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24756 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
24757 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24758 for (unsigned i = 0; i != NumElems; ++i)
24759 ShuffleVec[i] = i * SizeRatio;
24761 // Can't shuffle using an illegal type.
24762 if (!TLI.isTypeLegal(WideVecVT))
24765 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24766 DAG.getUNDEF(WideVecVT),
24768 // At this point all of the data is stored at the bottom of the
24769 // register. We now need to save it to mem.
24771 // Find the largest store unit
24772 MVT StoreType = MVT::i8;
24773 for (MVT Tp : MVT::integer_valuetypes()) {
24774 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24778 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24779 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24780 (64 <= NumElems * ToSz))
24781 StoreType = MVT::f64;
24783 // Bitcast the original vector into a vector of store-size units
24784 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24785 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24786 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24787 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
24788 SmallVector<SDValue, 8> Chains;
24789 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
24790 TLI.getPointerTy());
24791 SDValue Ptr = St->getBasePtr();
24793 // Perform one or more big stores into memory.
24794 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24795 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24796 StoreType, ShuffWide,
24797 DAG.getIntPtrConstant(i));
24798 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24799 St->getPointerInfo(), St->isVolatile(),
24800 St->isNonTemporal(), St->getAlignment());
24801 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24802 Chains.push_back(Ch);
24805 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24808 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24809 // the FP state in cases where an emms may be missing.
24810 // A preferable solution to the general problem is to figure out the right
24811 // places to insert EMMS. This qualifies as a quick hack.
24813 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24814 if (VT.getSizeInBits() != 64)
24817 const Function *F = DAG.getMachineFunction().getFunction();
24818 bool NoImplicitFloatOps = F->getAttributes().
24819 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
24820 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
24821 && Subtarget->hasSSE2();
24822 if ((VT.isVector() ||
24823 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24824 isa<LoadSDNode>(St->getValue()) &&
24825 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24826 St->getChain().hasOneUse() && !St->isVolatile()) {
24827 SDNode* LdVal = St->getValue().getNode();
24828 LoadSDNode *Ld = nullptr;
24829 int TokenFactorIndex = -1;
24830 SmallVector<SDValue, 8> Ops;
24831 SDNode* ChainVal = St->getChain().getNode();
24832 // Must be a store of a load. We currently handle two cases: the load
24833 // is a direct child, and it's under an intervening TokenFactor. It is
24834 // possible to dig deeper under nested TokenFactors.
24835 if (ChainVal == LdVal)
24836 Ld = cast<LoadSDNode>(St->getChain());
24837 else if (St->getValue().hasOneUse() &&
24838 ChainVal->getOpcode() == ISD::TokenFactor) {
24839 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24840 if (ChainVal->getOperand(i).getNode() == LdVal) {
24841 TokenFactorIndex = i;
24842 Ld = cast<LoadSDNode>(St->getValue());
24844 Ops.push_back(ChainVal->getOperand(i));
24848 if (!Ld || !ISD::isNormalLoad(Ld))
24851 // If this is not the MMX case, i.e. we are just turning i64 load/store
24852 // into f64 load/store, avoid the transformation if there are multiple
24853 // uses of the loaded value.
24854 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24859 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24860 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24862 if (Subtarget->is64Bit() || F64IsLegal) {
24863 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24864 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24865 Ld->getPointerInfo(), Ld->isVolatile(),
24866 Ld->isNonTemporal(), Ld->isInvariant(),
24867 Ld->getAlignment());
24868 SDValue NewChain = NewLd.getValue(1);
24869 if (TokenFactorIndex != -1) {
24870 Ops.push_back(NewChain);
24871 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24873 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24874 St->getPointerInfo(),
24875 St->isVolatile(), St->isNonTemporal(),
24876 St->getAlignment());
24879 // Otherwise, lower to two pairs of 32-bit loads / stores.
24880 SDValue LoAddr = Ld->getBasePtr();
24881 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24882 DAG.getConstant(4, MVT::i32));
24884 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24885 Ld->getPointerInfo(),
24886 Ld->isVolatile(), Ld->isNonTemporal(),
24887 Ld->isInvariant(), Ld->getAlignment());
24888 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24889 Ld->getPointerInfo().getWithOffset(4),
24890 Ld->isVolatile(), Ld->isNonTemporal(),
24892 MinAlign(Ld->getAlignment(), 4));
24894 SDValue NewChain = LoLd.getValue(1);
24895 if (TokenFactorIndex != -1) {
24896 Ops.push_back(LoLd);
24897 Ops.push_back(HiLd);
24898 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24901 LoAddr = St->getBasePtr();
24902 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24903 DAG.getConstant(4, MVT::i32));
24905 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24906 St->getPointerInfo(),
24907 St->isVolatile(), St->isNonTemporal(),
24908 St->getAlignment());
24909 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24910 St->getPointerInfo().getWithOffset(4),
24912 St->isNonTemporal(),
24913 MinAlign(St->getAlignment(), 4));
24914 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24919 /// Return 'true' if this vector operation is "horizontal"
24920 /// and return the operands for the horizontal operation in LHS and RHS. A
24921 /// horizontal operation performs the binary operation on successive elements
24922 /// of its first operand, then on successive elements of its second operand,
24923 /// returning the resulting values in a vector. For example, if
24924 /// A = < float a0, float a1, float a2, float a3 >
24926 /// B = < float b0, float b1, float b2, float b3 >
24927 /// then the result of doing a horizontal operation on A and B is
24928 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24929 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24930 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24931 /// set to A, RHS to B, and the routine returns 'true'.
24932 /// Note that the binary operation should have the property that if one of the
24933 /// operands is UNDEF then the result is UNDEF.
24934 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24935 // Look for the following pattern: if
24936 // A = < float a0, float a1, float a2, float a3 >
24937 // B = < float b0, float b1, float b2, float b3 >
24939 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24940 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24941 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24942 // which is A horizontal-op B.
24944 // At least one of the operands should be a vector shuffle.
24945 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24946 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24949 MVT VT = LHS.getSimpleValueType();
24951 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24952 "Unsupported vector type for horizontal add/sub");
24954 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24955 // operate independently on 128-bit lanes.
24956 unsigned NumElts = VT.getVectorNumElements();
24957 unsigned NumLanes = VT.getSizeInBits()/128;
24958 unsigned NumLaneElts = NumElts / NumLanes;
24959 assert((NumLaneElts % 2 == 0) &&
24960 "Vector type should have an even number of elements in each lane");
24961 unsigned HalfLaneElts = NumLaneElts/2;
24963 // View LHS in the form
24964 // LHS = VECTOR_SHUFFLE A, B, LMask
24965 // If LHS is not a shuffle then pretend it is the shuffle
24966 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24967 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24970 SmallVector<int, 16> LMask(NumElts);
24971 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24972 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24973 A = LHS.getOperand(0);
24974 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24975 B = LHS.getOperand(1);
24976 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24977 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24979 if (LHS.getOpcode() != ISD::UNDEF)
24981 for (unsigned i = 0; i != NumElts; ++i)
24985 // Likewise, view RHS in the form
24986 // RHS = VECTOR_SHUFFLE C, D, RMask
24988 SmallVector<int, 16> RMask(NumElts);
24989 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24990 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24991 C = RHS.getOperand(0);
24992 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24993 D = RHS.getOperand(1);
24994 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24995 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24997 if (RHS.getOpcode() != ISD::UNDEF)
24999 for (unsigned i = 0; i != NumElts; ++i)
25003 // Check that the shuffles are both shuffling the same vectors.
25004 if (!(A == C && B == D) && !(A == D && B == C))
25007 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25008 if (!A.getNode() && !B.getNode())
25011 // If A and B occur in reverse order in RHS, then "swap" them (which means
25012 // rewriting the mask).
25014 CommuteVectorShuffleMask(RMask, NumElts);
25016 // At this point LHS and RHS are equivalent to
25017 // LHS = VECTOR_SHUFFLE A, B, LMask
25018 // RHS = VECTOR_SHUFFLE A, B, RMask
25019 // Check that the masks correspond to performing a horizontal operation.
25020 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25021 for (unsigned i = 0; i != NumLaneElts; ++i) {
25022 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25024 // Ignore any UNDEF components.
25025 if (LIdx < 0 || RIdx < 0 ||
25026 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25027 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25030 // Check that successive elements are being operated on. If not, this is
25031 // not a horizontal operation.
25032 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25033 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25034 if (!(LIdx == Index && RIdx == Index + 1) &&
25035 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25040 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25041 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25045 /// Do target-specific dag combines on floating point adds.
25046 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25047 const X86Subtarget *Subtarget) {
25048 EVT VT = N->getValueType(0);
25049 SDValue LHS = N->getOperand(0);
25050 SDValue RHS = N->getOperand(1);
25052 // Try to synthesize horizontal adds from adds of shuffles.
25053 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25054 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25055 isHorizontalBinOp(LHS, RHS, true))
25056 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25060 /// Do target-specific dag combines on floating point subs.
25061 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25062 const X86Subtarget *Subtarget) {
25063 EVT VT = N->getValueType(0);
25064 SDValue LHS = N->getOperand(0);
25065 SDValue RHS = N->getOperand(1);
25067 // Try to synthesize horizontal subs from subs of shuffles.
25068 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25069 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25070 isHorizontalBinOp(LHS, RHS, false))
25071 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25075 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25076 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
25077 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25078 // F[X]OR(0.0, x) -> x
25079 // F[X]OR(x, 0.0) -> x
25080 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25081 if (C->getValueAPF().isPosZero())
25082 return N->getOperand(1);
25083 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25084 if (C->getValueAPF().isPosZero())
25085 return N->getOperand(0);
25089 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25090 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25091 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25093 // Only perform optimizations if UnsafeMath is used.
25094 if (!DAG.getTarget().Options.UnsafeFPMath)
25097 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25098 // into FMINC and FMAXC, which are Commutative operations.
25099 unsigned NewOp = 0;
25100 switch (N->getOpcode()) {
25101 default: llvm_unreachable("unknown opcode");
25102 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25103 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25106 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25107 N->getOperand(0), N->getOperand(1));
25110 /// Do target-specific dag combines on X86ISD::FAND nodes.
25111 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25112 // FAND(0.0, x) -> 0.0
25113 // FAND(x, 0.0) -> 0.0
25114 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25115 if (C->getValueAPF().isPosZero())
25116 return N->getOperand(0);
25117 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25118 if (C->getValueAPF().isPosZero())
25119 return N->getOperand(1);
25123 /// Do target-specific dag combines on X86ISD::FANDN nodes
25124 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25125 // FANDN(x, 0.0) -> 0.0
25126 // FANDN(0.0, x) -> x
25127 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25128 if (C->getValueAPF().isPosZero())
25129 return N->getOperand(1);
25130 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25131 if (C->getValueAPF().isPosZero())
25132 return N->getOperand(1);
25136 static SDValue PerformBTCombine(SDNode *N,
25138 TargetLowering::DAGCombinerInfo &DCI) {
25139 // BT ignores high bits in the bit index operand.
25140 SDValue Op1 = N->getOperand(1);
25141 if (Op1.hasOneUse()) {
25142 unsigned BitWidth = Op1.getValueSizeInBits();
25143 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25144 APInt KnownZero, KnownOne;
25145 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25146 !DCI.isBeforeLegalizeOps());
25147 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25148 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25149 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25150 DCI.CommitTargetLoweringOpt(TLO);
25155 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25156 SDValue Op = N->getOperand(0);
25157 if (Op.getOpcode() == ISD::BITCAST)
25158 Op = Op.getOperand(0);
25159 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25160 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25161 VT.getVectorElementType().getSizeInBits() ==
25162 OpVT.getVectorElementType().getSizeInBits()) {
25163 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25168 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25169 const X86Subtarget *Subtarget) {
25170 EVT VT = N->getValueType(0);
25171 if (!VT.isVector())
25174 SDValue N0 = N->getOperand(0);
25175 SDValue N1 = N->getOperand(1);
25176 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25179 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25180 // both SSE and AVX2 since there is no sign-extended shift right
25181 // operation on a vector with 64-bit elements.
25182 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25183 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25184 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25185 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25186 SDValue N00 = N0.getOperand(0);
25188 // EXTLOAD has a better solution on AVX2,
25189 // it may be replaced with X86ISD::VSEXT node.
25190 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25191 if (!ISD::isNormalLoad(N00.getNode()))
25194 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25195 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25197 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25203 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25204 TargetLowering::DAGCombinerInfo &DCI,
25205 const X86Subtarget *Subtarget) {
25206 SDValue N0 = N->getOperand(0);
25207 EVT VT = N->getValueType(0);
25209 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25210 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25211 // This exposes the sext to the sdivrem lowering, so that it directly extends
25212 // from AH (which we otherwise need to do contortions to access).
25213 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25214 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
25216 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25217 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
25218 N0.getOperand(0), N0.getOperand(1));
25219 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25220 return R.getValue(1);
25223 if (!DCI.isBeforeLegalizeOps())
25226 if (!Subtarget->hasFp256())
25229 if (VT.isVector() && VT.getSizeInBits() == 256) {
25230 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
25238 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25239 const X86Subtarget* Subtarget) {
25241 EVT VT = N->getValueType(0);
25243 // Let legalize expand this if it isn't a legal type yet.
25244 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25247 EVT ScalarVT = VT.getScalarType();
25248 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25249 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
25252 SDValue A = N->getOperand(0);
25253 SDValue B = N->getOperand(1);
25254 SDValue C = N->getOperand(2);
25256 bool NegA = (A.getOpcode() == ISD::FNEG);
25257 bool NegB = (B.getOpcode() == ISD::FNEG);
25258 bool NegC = (C.getOpcode() == ISD::FNEG);
25260 // Negative multiplication when NegA xor NegB
25261 bool NegMul = (NegA != NegB);
25263 A = A.getOperand(0);
25265 B = B.getOperand(0);
25267 C = C.getOperand(0);
25271 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25273 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25275 return DAG.getNode(Opcode, dl, VT, A, B, C);
25278 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25279 TargetLowering::DAGCombinerInfo &DCI,
25280 const X86Subtarget *Subtarget) {
25281 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25282 // (and (i32 x86isd::setcc_carry), 1)
25283 // This eliminates the zext. This transformation is necessary because
25284 // ISD::SETCC is always legalized to i8.
25286 SDValue N0 = N->getOperand(0);
25287 EVT VT = N->getValueType(0);
25289 if (N0.getOpcode() == ISD::AND &&
25291 N0.getOperand(0).hasOneUse()) {
25292 SDValue N00 = N0.getOperand(0);
25293 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25295 if (!C || C->getZExtValue() != 1)
25297 return DAG.getNode(ISD::AND, dl, VT,
25298 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25299 N00.getOperand(0), N00.getOperand(1)),
25300 DAG.getConstant(1, VT));
25304 if (N0.getOpcode() == ISD::TRUNCATE &&
25306 N0.getOperand(0).hasOneUse()) {
25307 SDValue N00 = N0.getOperand(0);
25308 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25309 return DAG.getNode(ISD::AND, dl, VT,
25310 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25311 N00.getOperand(0), N00.getOperand(1)),
25312 DAG.getConstant(1, VT));
25315 if (VT.is256BitVector()) {
25316 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
25321 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25322 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25323 // This exposes the zext to the udivrem lowering, so that it directly extends
25324 // from AH (which we otherwise need to do contortions to access).
25325 if (N0.getOpcode() == ISD::UDIVREM &&
25326 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25327 (VT == MVT::i32 || VT == MVT::i64)) {
25328 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25329 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25330 N0.getOperand(0), N0.getOperand(1));
25331 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25332 return R.getValue(1);
25338 // Optimize x == -y --> x+y == 0
25339 // x != -y --> x+y != 0
25340 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25341 const X86Subtarget* Subtarget) {
25342 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25343 SDValue LHS = N->getOperand(0);
25344 SDValue RHS = N->getOperand(1);
25345 EVT VT = N->getValueType(0);
25348 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25349 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25350 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25351 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
25352 LHS.getValueType(), RHS, LHS.getOperand(1));
25353 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
25354 addV, DAG.getConstant(0, addV.getValueType()), CC);
25356 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25357 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25358 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25359 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
25360 RHS.getValueType(), LHS, RHS.getOperand(1));
25361 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
25362 addV, DAG.getConstant(0, addV.getValueType()), CC);
25365 if (VT.getScalarType() == MVT::i1) {
25366 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25367 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25368 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
25369 if (!IsSEXT0 && !IsVZero0)
25371 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
25372 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25373 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25375 if (!IsSEXT1 && !IsVZero1)
25378 if (IsSEXT0 && IsVZero1) {
25379 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
25380 if (CC == ISD::SETEQ)
25381 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25382 return LHS.getOperand(0);
25384 if (IsSEXT1 && IsVZero0) {
25385 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
25386 if (CC == ISD::SETEQ)
25387 return DAG.getNOT(DL, RHS.getOperand(0), VT);
25388 return RHS.getOperand(0);
25395 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25396 const X86Subtarget *Subtarget) {
25398 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25399 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25400 "X86insertps is only defined for v4x32");
25402 SDValue Ld = N->getOperand(1);
25403 if (MayFoldLoad(Ld)) {
25404 // Extract the countS bits from the immediate so we can get the proper
25405 // address when narrowing the vector load to a specific element.
25406 // When the second source op is a memory address, interps doesn't use
25407 // countS and just gets an f32 from that address.
25408 unsigned DestIndex =
25409 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25410 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25414 // Create this as a scalar to vector to match the instruction pattern.
25415 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25416 // countS bits are ignored when loading from memory on insertps, which
25417 // means we don't need to explicitly set them to 0.
25418 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25419 LoadScalarToVector, N->getOperand(2));
25422 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25423 // as "sbb reg,reg", since it can be extended without zext and produces
25424 // an all-ones bit which is more useful than 0/1 in some cases.
25425 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25428 return DAG.getNode(ISD::AND, DL, VT,
25429 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25430 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
25431 DAG.getConstant(1, VT));
25432 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25433 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25434 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25435 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
25438 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25439 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25440 TargetLowering::DAGCombinerInfo &DCI,
25441 const X86Subtarget *Subtarget) {
25443 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25444 SDValue EFLAGS = N->getOperand(1);
25446 if (CC == X86::COND_A) {
25447 // Try to convert COND_A into COND_B in an attempt to facilitate
25448 // materializing "setb reg".
25450 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25451 // cannot take an immediate as its first operand.
25453 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25454 EFLAGS.getValueType().isInteger() &&
25455 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25456 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25457 EFLAGS.getNode()->getVTList(),
25458 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25459 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25460 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25464 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25465 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25467 if (CC == X86::COND_B)
25468 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25472 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
25473 if (Flags.getNode()) {
25474 SDValue Cond = DAG.getConstant(CC, MVT::i8);
25475 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25481 // Optimize branch condition evaluation.
25483 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25484 TargetLowering::DAGCombinerInfo &DCI,
25485 const X86Subtarget *Subtarget) {
25487 SDValue Chain = N->getOperand(0);
25488 SDValue Dest = N->getOperand(1);
25489 SDValue EFLAGS = N->getOperand(3);
25490 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25494 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
25495 if (Flags.getNode()) {
25496 SDValue Cond = DAG.getConstant(CC, MVT::i8);
25497 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25504 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25505 SelectionDAG &DAG) {
25506 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25507 // optimize away operation when it's from a constant.
25509 // The general transformation is:
25510 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25511 // AND(VECTOR_CMP(x,y), constant2)
25512 // constant2 = UNARYOP(constant)
25514 // Early exit if this isn't a vector operation, the operand of the
25515 // unary operation isn't a bitwise AND, or if the sizes of the operations
25516 // aren't the same.
25517 EVT VT = N->getValueType(0);
25518 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25519 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25520 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25523 // Now check that the other operand of the AND is a constant. We could
25524 // make the transformation for non-constant splats as well, but it's unclear
25525 // that would be a benefit as it would not eliminate any operations, just
25526 // perform one more step in scalar code before moving to the vector unit.
25527 if (BuildVectorSDNode *BV =
25528 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25529 // Bail out if the vector isn't a constant.
25530 if (!BV->isConstant())
25533 // Everything checks out. Build up the new and improved node.
25535 EVT IntVT = BV->getValueType(0);
25536 // Create a new constant of the appropriate type for the transformed
25538 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25539 // The AND node needs bitcasts to/from an integer vector type around it.
25540 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
25541 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25542 N->getOperand(0)->getOperand(0), MaskConst);
25543 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
25550 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25551 const X86TargetLowering *XTLI) {
25552 // First try to optimize away the conversion entirely when it's
25553 // conditionally from a constant. Vectors only.
25554 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
25555 if (Res != SDValue())
25558 // Now move on to more general possibilities.
25559 SDValue Op0 = N->getOperand(0);
25560 EVT InVT = Op0->getValueType(0);
25562 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
25563 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
25565 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
25566 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25567 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
25570 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25571 // a 32-bit target where SSE doesn't support i64->FP operations.
25572 if (Op0.getOpcode() == ISD::LOAD) {
25573 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25574 EVT VT = Ld->getValueType(0);
25575 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
25576 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25577 !XTLI->getSubtarget()->is64Bit() &&
25579 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
25580 Ld->getChain(), Op0, DAG);
25581 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25588 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25589 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25590 X86TargetLowering::DAGCombinerInfo &DCI) {
25591 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25592 // the result is either zero or one (depending on the input carry bit).
25593 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25594 if (X86::isZeroNode(N->getOperand(0)) &&
25595 X86::isZeroNode(N->getOperand(1)) &&
25596 // We don't have a good way to replace an EFLAGS use, so only do this when
25598 SDValue(N, 1).use_empty()) {
25600 EVT VT = N->getValueType(0);
25601 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
25602 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25603 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25604 DAG.getConstant(X86::COND_B,MVT::i8),
25606 DAG.getConstant(1, VT));
25607 return DCI.CombineTo(N, Res1, CarryOut);
25613 // fold (add Y, (sete X, 0)) -> adc 0, Y
25614 // (add Y, (setne X, 0)) -> sbb -1, Y
25615 // (sub (sete X, 0), Y) -> sbb 0, Y
25616 // (sub (setne X, 0), Y) -> adc -1, Y
25617 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25620 // Look through ZExts.
25621 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25622 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25625 SDValue SetCC = Ext.getOperand(0);
25626 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25629 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25630 if (CC != X86::COND_E && CC != X86::COND_NE)
25633 SDValue Cmp = SetCC.getOperand(1);
25634 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25635 !X86::isZeroNode(Cmp.getOperand(1)) ||
25636 !Cmp.getOperand(0).getValueType().isInteger())
25639 SDValue CmpOp0 = Cmp.getOperand(0);
25640 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25641 DAG.getConstant(1, CmpOp0.getValueType()));
25643 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25644 if (CC == X86::COND_NE)
25645 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25646 DL, OtherVal.getValueType(), OtherVal,
25647 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
25648 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25649 DL, OtherVal.getValueType(), OtherVal,
25650 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
25653 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25654 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25655 const X86Subtarget *Subtarget) {
25656 EVT VT = N->getValueType(0);
25657 SDValue Op0 = N->getOperand(0);
25658 SDValue Op1 = N->getOperand(1);
25660 // Try to synthesize horizontal adds from adds of shuffles.
25661 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25662 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25663 isHorizontalBinOp(Op0, Op1, true))
25664 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25666 return OptimizeConditionalInDecrement(N, DAG);
25669 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25670 const X86Subtarget *Subtarget) {
25671 SDValue Op0 = N->getOperand(0);
25672 SDValue Op1 = N->getOperand(1);
25674 // X86 can't encode an immediate LHS of a sub. See if we can push the
25675 // negation into a preceding instruction.
25676 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25677 // If the RHS of the sub is a XOR with one use and a constant, invert the
25678 // immediate. Then add one to the LHS of the sub so we can turn
25679 // X-Y -> X+~Y+1, saving one register.
25680 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25681 isa<ConstantSDNode>(Op1.getOperand(1))) {
25682 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25683 EVT VT = Op0.getValueType();
25684 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25686 DAG.getConstant(~XorC, VT));
25687 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25688 DAG.getConstant(C->getAPIntValue()+1, VT));
25692 // Try to synthesize horizontal adds from adds of shuffles.
25693 EVT VT = N->getValueType(0);
25694 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25695 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25696 isHorizontalBinOp(Op0, Op1, true))
25697 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25699 return OptimizeConditionalInDecrement(N, DAG);
25702 /// performVZEXTCombine - Performs build vector combines
25703 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25704 TargetLowering::DAGCombinerInfo &DCI,
25705 const X86Subtarget *Subtarget) {
25707 MVT VT = N->getSimpleValueType(0);
25708 SDValue Op = N->getOperand(0);
25709 MVT OpVT = Op.getSimpleValueType();
25710 MVT OpEltVT = OpVT.getVectorElementType();
25711 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25713 // (vzext (bitcast (vzext (x)) -> (vzext x)
25715 while (V.getOpcode() == ISD::BITCAST)
25716 V = V.getOperand(0);
25718 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25719 MVT InnerVT = V.getSimpleValueType();
25720 MVT InnerEltVT = InnerVT.getVectorElementType();
25722 // If the element sizes match exactly, we can just do one larger vzext. This
25723 // is always an exact type match as vzext operates on integer types.
25724 if (OpEltVT == InnerEltVT) {
25725 assert(OpVT == InnerVT && "Types must match for vzext!");
25726 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25729 // The only other way we can combine them is if only a single element of the
25730 // inner vzext is used in the input to the outer vzext.
25731 if (InnerEltVT.getSizeInBits() < InputBits)
25734 // In this case, the inner vzext is completely dead because we're going to
25735 // only look at bits inside of the low element. Just do the outer vzext on
25736 // a bitcast of the input to the inner.
25737 return DAG.getNode(X86ISD::VZEXT, DL, VT,
25738 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
25741 // Check if we can bypass extracting and re-inserting an element of an input
25742 // vector. Essentialy:
25743 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25744 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25745 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25746 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25747 SDValue ExtractedV = V.getOperand(0);
25748 SDValue OrigV = ExtractedV.getOperand(0);
25749 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25750 if (ExtractIdx->getZExtValue() == 0) {
25751 MVT OrigVT = OrigV.getSimpleValueType();
25752 // Extract a subvector if necessary...
25753 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25754 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25755 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25756 OrigVT.getVectorNumElements() / Ratio);
25757 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25758 DAG.getIntPtrConstant(0));
25760 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
25761 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25768 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25769 DAGCombinerInfo &DCI) const {
25770 SelectionDAG &DAG = DCI.DAG;
25771 switch (N->getOpcode()) {
25773 case ISD::EXTRACT_VECTOR_ELT:
25774 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25777 case X86ISD::SHRUNKBLEND:
25778 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25779 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25780 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25781 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25782 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25783 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25786 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25787 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25788 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25789 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25790 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25791 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25792 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
25793 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25794 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25796 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25798 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25799 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25800 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25801 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25802 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25803 case ISD::ANY_EXTEND:
25804 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25805 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25806 case ISD::SIGN_EXTEND_INREG:
25807 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25808 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
25809 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25810 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25811 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25812 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25813 case X86ISD::SHUFP: // Handle all target specific shuffles
25814 case X86ISD::PALIGNR:
25815 case X86ISD::UNPCKH:
25816 case X86ISD::UNPCKL:
25817 case X86ISD::MOVHLPS:
25818 case X86ISD::MOVLHPS:
25819 case X86ISD::PSHUFB:
25820 case X86ISD::PSHUFD:
25821 case X86ISD::PSHUFHW:
25822 case X86ISD::PSHUFLW:
25823 case X86ISD::MOVSS:
25824 case X86ISD::MOVSD:
25825 case X86ISD::VPERMILPI:
25826 case X86ISD::VPERM2X128:
25827 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25828 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25829 case ISD::INTRINSIC_WO_CHAIN:
25830 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
25831 case X86ISD::INSERTPS:
25832 return PerformINSERTPSCombine(N, DAG, Subtarget);
25833 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
25839 /// isTypeDesirableForOp - Return true if the target has native support for
25840 /// the specified value type and it is 'desirable' to use the type for the
25841 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25842 /// instruction encodings are longer and some i16 instructions are slow.
25843 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25844 if (!isTypeLegal(VT))
25846 if (VT != MVT::i16)
25853 case ISD::SIGN_EXTEND:
25854 case ISD::ZERO_EXTEND:
25855 case ISD::ANY_EXTEND:
25868 /// IsDesirableToPromoteOp - This method query the target whether it is
25869 /// beneficial for dag combiner to promote the specified node. If true, it
25870 /// should return the desired promotion type by reference.
25871 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25872 EVT VT = Op.getValueType();
25873 if (VT != MVT::i16)
25876 bool Promote = false;
25877 bool Commute = false;
25878 switch (Op.getOpcode()) {
25881 LoadSDNode *LD = cast<LoadSDNode>(Op);
25882 // If the non-extending load has a single use and it's not live out, then it
25883 // might be folded.
25884 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25885 Op.hasOneUse()*/) {
25886 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25887 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25888 // The only case where we'd want to promote LOAD (rather then it being
25889 // promoted as an operand is when it's only use is liveout.
25890 if (UI->getOpcode() != ISD::CopyToReg)
25897 case ISD::SIGN_EXTEND:
25898 case ISD::ZERO_EXTEND:
25899 case ISD::ANY_EXTEND:
25904 SDValue N0 = Op.getOperand(0);
25905 // Look out for (store (shl (load), x)).
25906 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25919 SDValue N0 = Op.getOperand(0);
25920 SDValue N1 = Op.getOperand(1);
25921 if (!Commute && MayFoldLoad(N1))
25923 // Avoid disabling potential load folding opportunities.
25924 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25926 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25936 //===----------------------------------------------------------------------===//
25937 // X86 Inline Assembly Support
25938 //===----------------------------------------------------------------------===//
25941 // Helper to match a string separated by whitespace.
25942 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
25943 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
25945 for (unsigned i = 0, e = args.size(); i != e; ++i) {
25946 StringRef piece(*args[i]);
25947 if (!s.startswith(piece)) // Check if the piece matches.
25950 s = s.substr(piece.size());
25951 StringRef::size_type pos = s.find_first_not_of(" \t");
25952 if (pos == 0) // We matched a prefix.
25960 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
25963 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
25965 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
25966 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
25967 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
25968 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
25970 if (AsmPieces.size() == 3)
25972 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
25979 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
25980 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
25982 std::string AsmStr = IA->getAsmString();
25984 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
25985 if (!Ty || Ty->getBitWidth() % 16 != 0)
25988 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
25989 SmallVector<StringRef, 4> AsmPieces;
25990 SplitString(AsmStr, AsmPieces, ";\n");
25992 switch (AsmPieces.size()) {
25993 default: return false;
25995 // FIXME: this should verify that we are targeting a 486 or better. If not,
25996 // we will turn this bswap into something that will be lowered to logical
25997 // ops instead of emitting the bswap asm. For now, we don't support 486 or
25998 // lower so don't worry about this.
26000 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
26001 matchAsm(AsmPieces[0], "bswapl", "$0") ||
26002 matchAsm(AsmPieces[0], "bswapq", "$0") ||
26003 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
26004 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
26005 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
26006 // No need to check constraints, nothing other than the equivalent of
26007 // "=r,0" would be valid here.
26008 return IntrinsicLowering::LowerToByteSwap(CI);
26011 // rorw $$8, ${0:w} --> llvm.bswap.i16
26012 if (CI->getType()->isIntegerTy(16) &&
26013 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26014 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
26015 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
26017 const std::string &ConstraintsStr = IA->getConstraintString();
26018 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26019 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26020 if (clobbersFlagRegisters(AsmPieces))
26021 return IntrinsicLowering::LowerToByteSwap(CI);
26025 if (CI->getType()->isIntegerTy(32) &&
26026 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26027 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
26028 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
26029 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
26031 const std::string &ConstraintsStr = IA->getConstraintString();
26032 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26033 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26034 if (clobbersFlagRegisters(AsmPieces))
26035 return IntrinsicLowering::LowerToByteSwap(CI);
26038 if (CI->getType()->isIntegerTy(64)) {
26039 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26040 if (Constraints.size() >= 2 &&
26041 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26042 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26043 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26044 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
26045 matchAsm(AsmPieces[1], "bswap", "%edx") &&
26046 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
26047 return IntrinsicLowering::LowerToByteSwap(CI);
26055 /// getConstraintType - Given a constraint letter, return the type of
26056 /// constraint it is for this target.
26057 X86TargetLowering::ConstraintType
26058 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
26059 if (Constraint.size() == 1) {
26060 switch (Constraint[0]) {
26071 return C_RegisterClass;
26095 return TargetLowering::getConstraintType(Constraint);
26098 /// Examine constraint type and operand type and determine a weight value.
26099 /// This object must already have been set up with the operand type
26100 /// and the current alternative constraint selected.
26101 TargetLowering::ConstraintWeight
26102 X86TargetLowering::getSingleConstraintMatchWeight(
26103 AsmOperandInfo &info, const char *constraint) const {
26104 ConstraintWeight weight = CW_Invalid;
26105 Value *CallOperandVal = info.CallOperandVal;
26106 // If we don't have a value, we can't do a match,
26107 // but allow it at the lowest weight.
26108 if (!CallOperandVal)
26110 Type *type = CallOperandVal->getType();
26111 // Look at the constraint type.
26112 switch (*constraint) {
26114 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26125 if (CallOperandVal->getType()->isIntegerTy())
26126 weight = CW_SpecificReg;
26131 if (type->isFloatingPointTy())
26132 weight = CW_SpecificReg;
26135 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26136 weight = CW_SpecificReg;
26140 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26141 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26142 weight = CW_Register;
26145 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26146 if (C->getZExtValue() <= 31)
26147 weight = CW_Constant;
26151 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26152 if (C->getZExtValue() <= 63)
26153 weight = CW_Constant;
26157 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26158 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26159 weight = CW_Constant;
26163 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26164 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26165 weight = CW_Constant;
26169 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26170 if (C->getZExtValue() <= 3)
26171 weight = CW_Constant;
26175 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26176 if (C->getZExtValue() <= 0xff)
26177 weight = CW_Constant;
26182 if (dyn_cast<ConstantFP>(CallOperandVal)) {
26183 weight = CW_Constant;
26187 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26188 if ((C->getSExtValue() >= -0x80000000LL) &&
26189 (C->getSExtValue() <= 0x7fffffffLL))
26190 weight = CW_Constant;
26194 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26195 if (C->getZExtValue() <= 0xffffffff)
26196 weight = CW_Constant;
26203 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26204 /// with another that has more specific requirements based on the type of the
26205 /// corresponding operand.
26206 const char *X86TargetLowering::
26207 LowerXConstraint(EVT ConstraintVT) const {
26208 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26209 // 'f' like normal targets.
26210 if (ConstraintVT.isFloatingPoint()) {
26211 if (Subtarget->hasSSE2())
26213 if (Subtarget->hasSSE1())
26217 return TargetLowering::LowerXConstraint(ConstraintVT);
26220 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26221 /// vector. If it is invalid, don't add anything to Ops.
26222 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26223 std::string &Constraint,
26224 std::vector<SDValue>&Ops,
26225 SelectionDAG &DAG) const {
26228 // Only support length 1 constraints for now.
26229 if (Constraint.length() > 1) return;
26231 char ConstraintLetter = Constraint[0];
26232 switch (ConstraintLetter) {
26235 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26236 if (C->getZExtValue() <= 31) {
26237 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26243 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26244 if (C->getZExtValue() <= 63) {
26245 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26251 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26252 if (isInt<8>(C->getSExtValue())) {
26253 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26259 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26260 if (C->getZExtValue() <= 255) {
26261 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26267 // 32-bit signed value
26268 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26269 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26270 C->getSExtValue())) {
26271 // Widen to 64 bits here to get it sign extended.
26272 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
26275 // FIXME gcc accepts some relocatable values here too, but only in certain
26276 // memory models; it's complicated.
26281 // 32-bit unsigned value
26282 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26283 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26284 C->getZExtValue())) {
26285 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
26289 // FIXME gcc accepts some relocatable values here too, but only in certain
26290 // memory models; it's complicated.
26294 // Literal immediates are always ok.
26295 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26296 // Widen to 64 bits here to get it sign extended.
26297 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
26301 // In any sort of PIC mode addresses need to be computed at runtime by
26302 // adding in a register or some sort of table lookup. These can't
26303 // be used as immediates.
26304 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26307 // If we are in non-pic codegen mode, we allow the address of a global (with
26308 // an optional displacement) to be used with 'i'.
26309 GlobalAddressSDNode *GA = nullptr;
26310 int64_t Offset = 0;
26312 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26314 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26315 Offset += GA->getOffset();
26317 } else if (Op.getOpcode() == ISD::ADD) {
26318 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26319 Offset += C->getZExtValue();
26320 Op = Op.getOperand(0);
26323 } else if (Op.getOpcode() == ISD::SUB) {
26324 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26325 Offset += -C->getZExtValue();
26326 Op = Op.getOperand(0);
26331 // Otherwise, this isn't something we can handle, reject it.
26335 const GlobalValue *GV = GA->getGlobal();
26336 // If we require an extra load to get this address, as in PIC mode, we
26337 // can't accept it.
26338 if (isGlobalStubReference(
26339 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26342 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26343 GA->getValueType(0), Offset);
26348 if (Result.getNode()) {
26349 Ops.push_back(Result);
26352 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26355 std::pair<unsigned, const TargetRegisterClass*>
26356 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
26358 // First, see if this is a constraint that directly corresponds to an LLVM
26360 if (Constraint.size() == 1) {
26361 // GCC Constraint Letters
26362 switch (Constraint[0]) {
26364 // TODO: Slight differences here in allocation order and leaving
26365 // RIP in the class. Do they matter any more here than they do
26366 // in the normal allocation?
26367 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26368 if (Subtarget->is64Bit()) {
26369 if (VT == MVT::i32 || VT == MVT::f32)
26370 return std::make_pair(0U, &X86::GR32RegClass);
26371 if (VT == MVT::i16)
26372 return std::make_pair(0U, &X86::GR16RegClass);
26373 if (VT == MVT::i8 || VT == MVT::i1)
26374 return std::make_pair(0U, &X86::GR8RegClass);
26375 if (VT == MVT::i64 || VT == MVT::f64)
26376 return std::make_pair(0U, &X86::GR64RegClass);
26379 // 32-bit fallthrough
26380 case 'Q': // Q_REGS
26381 if (VT == MVT::i32 || VT == MVT::f32)
26382 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26383 if (VT == MVT::i16)
26384 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26385 if (VT == MVT::i8 || VT == MVT::i1)
26386 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26387 if (VT == MVT::i64)
26388 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26390 case 'r': // GENERAL_REGS
26391 case 'l': // INDEX_REGS
26392 if (VT == MVT::i8 || VT == MVT::i1)
26393 return std::make_pair(0U, &X86::GR8RegClass);
26394 if (VT == MVT::i16)
26395 return std::make_pair(0U, &X86::GR16RegClass);
26396 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26397 return std::make_pair(0U, &X86::GR32RegClass);
26398 return std::make_pair(0U, &X86::GR64RegClass);
26399 case 'R': // LEGACY_REGS
26400 if (VT == MVT::i8 || VT == MVT::i1)
26401 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26402 if (VT == MVT::i16)
26403 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26404 if (VT == MVT::i32 || !Subtarget->is64Bit())
26405 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26406 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26407 case 'f': // FP Stack registers.
26408 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26409 // value to the correct fpstack register class.
26410 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26411 return std::make_pair(0U, &X86::RFP32RegClass);
26412 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26413 return std::make_pair(0U, &X86::RFP64RegClass);
26414 return std::make_pair(0U, &X86::RFP80RegClass);
26415 case 'y': // MMX_REGS if MMX allowed.
26416 if (!Subtarget->hasMMX()) break;
26417 return std::make_pair(0U, &X86::VR64RegClass);
26418 case 'Y': // SSE_REGS if SSE2 allowed
26419 if (!Subtarget->hasSSE2()) break;
26421 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26422 if (!Subtarget->hasSSE1()) break;
26424 switch (VT.SimpleTy) {
26426 // Scalar SSE types.
26429 return std::make_pair(0U, &X86::FR32RegClass);
26432 return std::make_pair(0U, &X86::FR64RegClass);
26440 return std::make_pair(0U, &X86::VR128RegClass);
26448 return std::make_pair(0U, &X86::VR256RegClass);
26453 return std::make_pair(0U, &X86::VR512RegClass);
26459 // Use the default implementation in TargetLowering to convert the register
26460 // constraint into a member of a register class.
26461 std::pair<unsigned, const TargetRegisterClass*> Res;
26462 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
26464 // Not found as a standard register?
26466 // Map st(0) -> st(7) -> ST0
26467 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26468 tolower(Constraint[1]) == 's' &&
26469 tolower(Constraint[2]) == 't' &&
26470 Constraint[3] == '(' &&
26471 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26472 Constraint[5] == ')' &&
26473 Constraint[6] == '}') {
26475 Res.first = X86::FP0+Constraint[4]-'0';
26476 Res.second = &X86::RFP80RegClass;
26480 // GCC allows "st(0)" to be called just plain "st".
26481 if (StringRef("{st}").equals_lower(Constraint)) {
26482 Res.first = X86::FP0;
26483 Res.second = &X86::RFP80RegClass;
26488 if (StringRef("{flags}").equals_lower(Constraint)) {
26489 Res.first = X86::EFLAGS;
26490 Res.second = &X86::CCRRegClass;
26494 // 'A' means EAX + EDX.
26495 if (Constraint == "A") {
26496 Res.first = X86::EAX;
26497 Res.second = &X86::GR32_ADRegClass;
26503 // Otherwise, check to see if this is a register class of the wrong value
26504 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26505 // turn into {ax},{dx}.
26506 if (Res.second->hasType(VT))
26507 return Res; // Correct type already, nothing to do.
26509 // All of the single-register GCC register classes map their values onto
26510 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
26511 // really want an 8-bit or 32-bit register, map to the appropriate register
26512 // class and return the appropriate register.
26513 if (Res.second == &X86::GR16RegClass) {
26514 if (VT == MVT::i8 || VT == MVT::i1) {
26515 unsigned DestReg = 0;
26516 switch (Res.first) {
26518 case X86::AX: DestReg = X86::AL; break;
26519 case X86::DX: DestReg = X86::DL; break;
26520 case X86::CX: DestReg = X86::CL; break;
26521 case X86::BX: DestReg = X86::BL; break;
26524 Res.first = DestReg;
26525 Res.second = &X86::GR8RegClass;
26527 } else if (VT == MVT::i32 || VT == MVT::f32) {
26528 unsigned DestReg = 0;
26529 switch (Res.first) {
26531 case X86::AX: DestReg = X86::EAX; break;
26532 case X86::DX: DestReg = X86::EDX; break;
26533 case X86::CX: DestReg = X86::ECX; break;
26534 case X86::BX: DestReg = X86::EBX; break;
26535 case X86::SI: DestReg = X86::ESI; break;
26536 case X86::DI: DestReg = X86::EDI; break;
26537 case X86::BP: DestReg = X86::EBP; break;
26538 case X86::SP: DestReg = X86::ESP; break;
26541 Res.first = DestReg;
26542 Res.second = &X86::GR32RegClass;
26544 } else if (VT == MVT::i64 || VT == MVT::f64) {
26545 unsigned DestReg = 0;
26546 switch (Res.first) {
26548 case X86::AX: DestReg = X86::RAX; break;
26549 case X86::DX: DestReg = X86::RDX; break;
26550 case X86::CX: DestReg = X86::RCX; break;
26551 case X86::BX: DestReg = X86::RBX; break;
26552 case X86::SI: DestReg = X86::RSI; break;
26553 case X86::DI: DestReg = X86::RDI; break;
26554 case X86::BP: DestReg = X86::RBP; break;
26555 case X86::SP: DestReg = X86::RSP; break;
26558 Res.first = DestReg;
26559 Res.second = &X86::GR64RegClass;
26562 } else if (Res.second == &X86::FR32RegClass ||
26563 Res.second == &X86::FR64RegClass ||
26564 Res.second == &X86::VR128RegClass ||
26565 Res.second == &X86::VR256RegClass ||
26566 Res.second == &X86::FR32XRegClass ||
26567 Res.second == &X86::FR64XRegClass ||
26568 Res.second == &X86::VR128XRegClass ||
26569 Res.second == &X86::VR256XRegClass ||
26570 Res.second == &X86::VR512RegClass) {
26571 // Handle references to XMM physical registers that got mapped into the
26572 // wrong class. This can happen with constraints like {xmm0} where the
26573 // target independent register mapper will just pick the first match it can
26574 // find, ignoring the required type.
26576 if (VT == MVT::f32 || VT == MVT::i32)
26577 Res.second = &X86::FR32RegClass;
26578 else if (VT == MVT::f64 || VT == MVT::i64)
26579 Res.second = &X86::FR64RegClass;
26580 else if (X86::VR128RegClass.hasType(VT))
26581 Res.second = &X86::VR128RegClass;
26582 else if (X86::VR256RegClass.hasType(VT))
26583 Res.second = &X86::VR256RegClass;
26584 else if (X86::VR512RegClass.hasType(VT))
26585 Res.second = &X86::VR512RegClass;
26591 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
26593 // Scaling factors are not free at all.
26594 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26595 // will take 2 allocations in the out of order engine instead of 1
26596 // for plain addressing mode, i.e. inst (reg1).
26598 // vaddps (%rsi,%drx), %ymm0, %ymm1
26599 // Requires two allocations (one for the load, one for the computation)
26601 // vaddps (%rsi), %ymm0, %ymm1
26602 // Requires just 1 allocation, i.e., freeing allocations for other operations
26603 // and having less micro operations to execute.
26605 // For some X86 architectures, this is even worse because for instance for
26606 // stores, the complex addressing mode forces the instruction to use the
26607 // "load" ports instead of the dedicated "store" port.
26608 // E.g., on Haswell:
26609 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26610 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26611 if (isLegalAddressingMode(AM, Ty))
26612 // Scale represents reg2 * scale, thus account for 1
26613 // as soon as we use a second register.
26614 return AM.Scale != 0;
26618 bool X86TargetLowering::isTargetFTOL() const {
26619 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();