1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/ADT/VariadicFunction.h"
29 #include "llvm/CodeGen/IntrinsicLowering.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 static cl::opt<int> ReciprocalEstimateRefinementSteps(
71 "x86-recip-refinement-steps", cl::init(1),
72 cl::desc("Specify the number of Newton-Raphson iterations applied to the "
73 "result of the hardware reciprocal estimate instruction."),
76 // Forward declarations.
77 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
80 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
81 SelectionDAG &DAG, SDLoc dl,
82 unsigned vectorWidth) {
83 assert((vectorWidth == 128 || vectorWidth == 256) &&
84 "Unsupported vector width");
85 EVT VT = Vec.getValueType();
86 EVT ElVT = VT.getVectorElementType();
87 unsigned Factor = VT.getSizeInBits()/vectorWidth;
88 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
89 VT.getVectorNumElements()/Factor);
91 // Extract from UNDEF is UNDEF.
92 if (Vec.getOpcode() == ISD::UNDEF)
93 return DAG.getUNDEF(ResultVT);
95 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
96 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
98 // This is the index of the first element of the vectorWidth-bit chunk
100 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
103 // If the input is a buildvector just emit a smaller one.
104 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
105 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
106 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
109 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
110 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
113 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
114 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
115 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
116 /// instructions or a simple subregister reference. Idx is an index in the
117 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
118 /// lowering EXTRACT_VECTOR_ELT operations easier.
119 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
120 SelectionDAG &DAG, SDLoc dl) {
121 assert((Vec.getValueType().is256BitVector() ||
122 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
123 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
126 /// Generate a DAG to grab 256-bits from a 512-bit vector.
127 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
128 SelectionDAG &DAG, SDLoc dl) {
129 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
130 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
133 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
134 unsigned IdxVal, SelectionDAG &DAG,
135 SDLoc dl, unsigned vectorWidth) {
136 assert((vectorWidth == 128 || vectorWidth == 256) &&
137 "Unsupported vector width");
138 // Inserting UNDEF is Result
139 if (Vec.getOpcode() == ISD::UNDEF)
141 EVT VT = Vec.getValueType();
142 EVT ElVT = VT.getVectorElementType();
143 EVT ResultVT = Result.getValueType();
145 // Insert the relevant vectorWidth bits.
146 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
148 // This is the index of the first element of the vectorWidth-bit chunk
150 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
153 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal);
154 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
157 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
158 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
159 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
160 /// simple superregister reference. Idx is an index in the 128 bits
161 /// we want. It need not be aligned to a 128-bit boundary. That makes
162 /// lowering INSERT_VECTOR_ELT operations easier.
163 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
164 SelectionDAG &DAG,SDLoc dl) {
165 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
166 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
169 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
170 SelectionDAG &DAG, SDLoc dl) {
171 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
172 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
175 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
176 /// instructions. This is used because creating CONCAT_VECTOR nodes of
177 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
178 /// large BUILD_VECTORS.
179 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
180 unsigned NumElems, SelectionDAG &DAG,
182 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
183 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
186 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
187 unsigned NumElems, SelectionDAG &DAG,
189 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
190 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
193 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
194 const X86Subtarget &STI)
195 : TargetLowering(TM), Subtarget(&STI) {
196 X86ScalarSSEf64 = Subtarget->hasSSE2();
197 X86ScalarSSEf32 = Subtarget->hasSSE1();
198 TD = getDataLayout();
200 // Set up the TargetLowering object.
201 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
203 // X86 is weird. It always uses i8 for shift amounts and setcc results.
204 setBooleanContents(ZeroOrOneBooleanContent);
205 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
206 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
208 // For 64-bit, since we have so many registers, use the ILP scheduler.
209 // For 32-bit, use the register pressure specific scheduling.
210 // For Atom, always use ILP scheduling.
211 if (Subtarget->isAtom())
212 setSchedulingPreference(Sched::ILP);
213 else if (Subtarget->is64Bit())
214 setSchedulingPreference(Sched::ILP);
216 setSchedulingPreference(Sched::RegPressure);
217 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
218 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
220 // Bypass expensive divides on Atom when compiling with O2.
221 if (TM.getOptLevel() >= CodeGenOpt::Default) {
222 if (Subtarget->hasSlowDivide32())
223 addBypassSlowDiv(32, 8);
224 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
225 addBypassSlowDiv(64, 16);
228 if (Subtarget->isTargetKnownWindowsMSVC()) {
229 // Setup Windows compiler runtime calls.
230 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
231 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
232 setLibcallName(RTLIB::SREM_I64, "_allrem");
233 setLibcallName(RTLIB::UREM_I64, "_aullrem");
234 setLibcallName(RTLIB::MUL_I64, "_allmul");
235 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
236 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
237 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
238 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
239 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
241 // The _ftol2 runtime function has an unusual calling conv, which
242 // is modeled by a special pseudo-instruction.
243 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
244 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
245 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
246 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
249 if (Subtarget->isTargetDarwin()) {
250 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
251 setUseUnderscoreSetJmp(false);
252 setUseUnderscoreLongJmp(false);
253 } else if (Subtarget->isTargetWindowsGNU()) {
254 // MS runtime is weird: it exports _setjmp, but longjmp!
255 setUseUnderscoreSetJmp(true);
256 setUseUnderscoreLongJmp(false);
258 setUseUnderscoreSetJmp(true);
259 setUseUnderscoreLongJmp(true);
262 // Set up the register classes.
263 addRegisterClass(MVT::i8, &X86::GR8RegClass);
264 addRegisterClass(MVT::i16, &X86::GR16RegClass);
265 addRegisterClass(MVT::i32, &X86::GR32RegClass);
266 if (Subtarget->is64Bit())
267 addRegisterClass(MVT::i64, &X86::GR64RegClass);
269 for (MVT VT : MVT::integer_valuetypes())
270 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
272 // We don't accept any truncstore of integer registers.
273 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
274 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
275 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
276 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
277 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
278 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
280 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
282 // SETOEQ and SETUNE require checking two conditions.
283 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
284 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
285 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
286 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
287 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
288 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
290 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
292 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
293 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
294 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
298 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
299 } else if (!TM.Options.UseSoftFloat) {
300 // We have an algorithm for SSE2->double, and we turn this into a
301 // 64-bit FILD followed by conditional FADD for other targets.
302 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
303 // We have an algorithm for SSE2, and we turn this into a 64-bit
304 // FILD for other targets.
305 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
308 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
310 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
311 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
313 if (!TM.Options.UseSoftFloat) {
314 // SSE has no i16 to fp conversion, only i32
315 if (X86ScalarSSEf32) {
316 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
317 // f32 and f64 cases are Legal, f80 case is not
318 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
320 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
321 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
324 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
325 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
328 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
329 // are Legal, f80 is custom lowered.
330 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
331 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
333 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
335 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
336 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
338 if (X86ScalarSSEf32) {
339 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
340 // f32 and f64 cases are Legal, f80 case is not
341 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
343 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
344 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
347 // Handle FP_TO_UINT by promoting the destination to a larger signed
349 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
350 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
351 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
353 if (Subtarget->is64Bit()) {
354 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
355 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
356 } else if (!TM.Options.UseSoftFloat) {
357 // Since AVX is a superset of SSE3, only check for SSE here.
358 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
359 // Expand FP_TO_UINT into a select.
360 // FIXME: We would like to use a Custom expander here eventually to do
361 // the optimal thing for SSE vs. the default expansion in the legalizer.
362 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
364 // With SSE3 we can use fisttpll to convert to a signed i64; without
365 // SSE, we're stuck with a fistpll.
366 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
369 if (isTargetFTOL()) {
370 // Use the _ftol2 runtime function, which has a pseudo-instruction
371 // to handle its weird calling convention.
372 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
375 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
376 if (!X86ScalarSSEf64) {
377 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
378 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
379 if (Subtarget->is64Bit()) {
380 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
381 // Without SSE, i64->f64 goes through memory.
382 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
386 // Scalar integer divide and remainder are lowered to use operations that
387 // produce two results, to match the available instructions. This exposes
388 // the two-result form to trivial CSE, which is able to combine x/y and x%y
389 // into a single instruction.
391 // Scalar integer multiply-high is also lowered to use two-result
392 // operations, to match the available instructions. However, plain multiply
393 // (low) operations are left as Legal, as there are single-result
394 // instructions for this in x86. Using the two-result multiply instructions
395 // when both high and low results are needed must be arranged by dagcombine.
396 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
398 setOperationAction(ISD::MULHS, VT, Expand);
399 setOperationAction(ISD::MULHU, VT, Expand);
400 setOperationAction(ISD::SDIV, VT, Expand);
401 setOperationAction(ISD::UDIV, VT, Expand);
402 setOperationAction(ISD::SREM, VT, Expand);
403 setOperationAction(ISD::UREM, VT, Expand);
405 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
406 setOperationAction(ISD::ADDC, VT, Custom);
407 setOperationAction(ISD::ADDE, VT, Custom);
408 setOperationAction(ISD::SUBC, VT, Custom);
409 setOperationAction(ISD::SUBE, VT, Custom);
412 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
413 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
414 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
415 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
416 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
417 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
418 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
419 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
420 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
421 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
422 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
423 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
424 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
425 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
426 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
427 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
430 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
431 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
432 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
433 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
434 setOperationAction(ISD::FREM , MVT::f32 , Expand);
435 setOperationAction(ISD::FREM , MVT::f64 , Expand);
436 setOperationAction(ISD::FREM , MVT::f80 , Expand);
437 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
439 // Promote the i8 variants and force them on up to i32 which has a shorter
441 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
442 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
443 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
444 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
445 if (Subtarget->hasBMI()) {
446 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
447 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
448 if (Subtarget->is64Bit())
449 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
451 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
452 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
453 if (Subtarget->is64Bit())
454 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
457 if (Subtarget->hasLZCNT()) {
458 // When promoting the i8 variants, force them to i32 for a shorter
460 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
461 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
462 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
463 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
464 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
465 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
466 if (Subtarget->is64Bit())
467 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
469 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
470 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
471 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
472 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
473 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
474 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
475 if (Subtarget->is64Bit()) {
476 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
477 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
481 // Special handling for half-precision floating point conversions.
482 // If we don't have F16C support, then lower half float conversions
483 // into library calls.
484 if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
485 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
486 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
489 // There's never any support for operations beyond MVT::f32.
490 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
491 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
492 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
493 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
495 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
496 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
497 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
498 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
499 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
500 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
502 if (Subtarget->hasPOPCNT()) {
503 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
505 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
506 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
507 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
508 if (Subtarget->is64Bit())
509 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
512 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
514 if (!Subtarget->hasMOVBE())
515 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
517 // These should be promoted to a larger select which is supported.
518 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
519 // X86 wants to expand cmov itself.
520 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
521 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
522 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
523 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
524 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
525 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
526 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
527 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
528 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
529 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
530 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
531 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
532 if (Subtarget->is64Bit()) {
533 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
534 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
536 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
537 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
538 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
539 // support continuation, user-level threading, and etc.. As a result, no
540 // other SjLj exception interfaces are implemented and please don't build
541 // your own exception handling based on them.
542 // LLVM/Clang supports zero-cost DWARF exception handling.
543 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
544 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
547 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
548 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
549 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
550 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
551 if (Subtarget->is64Bit())
552 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
553 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
554 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
555 if (Subtarget->is64Bit()) {
556 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
557 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
558 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
559 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
560 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
562 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
563 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
564 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
565 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
566 if (Subtarget->is64Bit()) {
567 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
568 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
569 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
572 if (Subtarget->hasSSE1())
573 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
575 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
577 // Expand certain atomics
578 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
580 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
581 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
582 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
585 if (Subtarget->hasCmpxchg16b()) {
586 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
589 // FIXME - use subtarget debug flags
590 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
591 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
592 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
595 if (Subtarget->is64Bit()) {
596 setExceptionPointerRegister(X86::RAX);
597 setExceptionSelectorRegister(X86::RDX);
599 setExceptionPointerRegister(X86::EAX);
600 setExceptionSelectorRegister(X86::EDX);
602 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
603 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
605 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
606 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
608 setOperationAction(ISD::TRAP, MVT::Other, Legal);
609 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
611 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
612 setOperationAction(ISD::VASTART , MVT::Other, Custom);
613 setOperationAction(ISD::VAEND , MVT::Other, Expand);
614 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
615 // TargetInfo::X86_64ABIBuiltinVaList
616 setOperationAction(ISD::VAARG , MVT::Other, Custom);
617 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
619 // TargetInfo::CharPtrBuiltinVaList
620 setOperationAction(ISD::VAARG , MVT::Other, Expand);
621 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
624 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
625 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
627 setOperationAction(ISD::DYNAMIC_STACKALLOC, getPointerTy(), Custom);
629 if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
630 // f32 and f64 use SSE.
631 // Set up the FP register classes.
632 addRegisterClass(MVT::f32, &X86::FR32RegClass);
633 addRegisterClass(MVT::f64, &X86::FR64RegClass);
635 // Use ANDPD to simulate FABS.
636 setOperationAction(ISD::FABS , MVT::f64, Custom);
637 setOperationAction(ISD::FABS , MVT::f32, Custom);
639 // Use XORP to simulate FNEG.
640 setOperationAction(ISD::FNEG , MVT::f64, Custom);
641 setOperationAction(ISD::FNEG , MVT::f32, Custom);
643 // Use ANDPD and ORPD to simulate FCOPYSIGN.
644 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
645 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
647 // Lower this to FGETSIGNx86 plus an AND.
648 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
649 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
651 // We don't support sin/cos/fmod
652 setOperationAction(ISD::FSIN , MVT::f64, Expand);
653 setOperationAction(ISD::FCOS , MVT::f64, Expand);
654 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
655 setOperationAction(ISD::FSIN , MVT::f32, Expand);
656 setOperationAction(ISD::FCOS , MVT::f32, Expand);
657 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
659 // Expand FP immediates into loads from the stack, except for the special
661 addLegalFPImmediate(APFloat(+0.0)); // xorpd
662 addLegalFPImmediate(APFloat(+0.0f)); // xorps
663 } else if (!TM.Options.UseSoftFloat && X86ScalarSSEf32) {
664 // Use SSE for f32, x87 for f64.
665 // Set up the FP register classes.
666 addRegisterClass(MVT::f32, &X86::FR32RegClass);
667 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
669 // Use ANDPS to simulate FABS.
670 setOperationAction(ISD::FABS , MVT::f32, Custom);
672 // Use XORP to simulate FNEG.
673 setOperationAction(ISD::FNEG , MVT::f32, Custom);
675 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
677 // Use ANDPS and ORPS to simulate FCOPYSIGN.
678 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
679 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
681 // We don't support sin/cos/fmod
682 setOperationAction(ISD::FSIN , MVT::f32, Expand);
683 setOperationAction(ISD::FCOS , MVT::f32, Expand);
684 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
686 // Special cases we handle for FP constants.
687 addLegalFPImmediate(APFloat(+0.0f)); // xorps
688 addLegalFPImmediate(APFloat(+0.0)); // FLD0
689 addLegalFPImmediate(APFloat(+1.0)); // FLD1
690 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
691 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
693 if (!TM.Options.UnsafeFPMath) {
694 setOperationAction(ISD::FSIN , MVT::f64, Expand);
695 setOperationAction(ISD::FCOS , MVT::f64, Expand);
696 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
698 } else if (!TM.Options.UseSoftFloat) {
699 // f32 and f64 in x87.
700 // Set up the FP register classes.
701 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
702 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
704 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
705 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
706 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
707 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
709 if (!TM.Options.UnsafeFPMath) {
710 setOperationAction(ISD::FSIN , MVT::f64, Expand);
711 setOperationAction(ISD::FSIN , MVT::f32, Expand);
712 setOperationAction(ISD::FCOS , MVT::f64, Expand);
713 setOperationAction(ISD::FCOS , MVT::f32, Expand);
714 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
715 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
717 addLegalFPImmediate(APFloat(+0.0)); // FLD0
718 addLegalFPImmediate(APFloat(+1.0)); // FLD1
719 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
720 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
721 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
722 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
723 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
724 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
727 // We don't support FMA.
728 setOperationAction(ISD::FMA, MVT::f64, Expand);
729 setOperationAction(ISD::FMA, MVT::f32, Expand);
731 // Long double always uses X87.
732 if (!TM.Options.UseSoftFloat) {
733 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
734 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
735 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
737 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
738 addLegalFPImmediate(TmpFlt); // FLD0
740 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
743 APFloat TmpFlt2(+1.0);
744 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
746 addLegalFPImmediate(TmpFlt2); // FLD1
747 TmpFlt2.changeSign();
748 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
751 if (!TM.Options.UnsafeFPMath) {
752 setOperationAction(ISD::FSIN , MVT::f80, Expand);
753 setOperationAction(ISD::FCOS , MVT::f80, Expand);
754 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
757 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
758 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
759 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
760 setOperationAction(ISD::FRINT, MVT::f80, Expand);
761 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
762 setOperationAction(ISD::FMA, MVT::f80, Expand);
765 // Always use a library call for pow.
766 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
767 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
768 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
770 setOperationAction(ISD::FLOG, MVT::f80, Expand);
771 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
772 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
773 setOperationAction(ISD::FEXP, MVT::f80, Expand);
774 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
775 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
776 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
778 // First set operation action for all vector types to either promote
779 // (for widening) or expand (for scalarization). Then we will selectively
780 // turn on ones that can be effectively codegen'd.
781 for (MVT VT : MVT::vector_valuetypes()) {
782 setOperationAction(ISD::ADD , VT, Expand);
783 setOperationAction(ISD::SUB , VT, Expand);
784 setOperationAction(ISD::FADD, VT, Expand);
785 setOperationAction(ISD::FNEG, VT, Expand);
786 setOperationAction(ISD::FSUB, VT, Expand);
787 setOperationAction(ISD::MUL , VT, Expand);
788 setOperationAction(ISD::FMUL, VT, Expand);
789 setOperationAction(ISD::SDIV, VT, Expand);
790 setOperationAction(ISD::UDIV, VT, Expand);
791 setOperationAction(ISD::FDIV, VT, Expand);
792 setOperationAction(ISD::SREM, VT, Expand);
793 setOperationAction(ISD::UREM, VT, Expand);
794 setOperationAction(ISD::LOAD, VT, Expand);
795 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
797 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
798 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
799 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
800 setOperationAction(ISD::FABS, VT, Expand);
801 setOperationAction(ISD::FSIN, VT, Expand);
802 setOperationAction(ISD::FSINCOS, VT, Expand);
803 setOperationAction(ISD::FCOS, VT, Expand);
804 setOperationAction(ISD::FSINCOS, VT, Expand);
805 setOperationAction(ISD::FREM, VT, Expand);
806 setOperationAction(ISD::FMA, VT, Expand);
807 setOperationAction(ISD::FPOWI, VT, Expand);
808 setOperationAction(ISD::FSQRT, VT, Expand);
809 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
810 setOperationAction(ISD::FFLOOR, VT, Expand);
811 setOperationAction(ISD::FCEIL, VT, Expand);
812 setOperationAction(ISD::FTRUNC, VT, Expand);
813 setOperationAction(ISD::FRINT, VT, Expand);
814 setOperationAction(ISD::FNEARBYINT, VT, Expand);
815 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
816 setOperationAction(ISD::MULHS, VT, Expand);
817 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
818 setOperationAction(ISD::MULHU, VT, Expand);
819 setOperationAction(ISD::SDIVREM, VT, Expand);
820 setOperationAction(ISD::UDIVREM, VT, Expand);
821 setOperationAction(ISD::FPOW, VT, Expand);
822 setOperationAction(ISD::CTPOP, VT, Expand);
823 setOperationAction(ISD::CTTZ, VT, Expand);
824 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
825 setOperationAction(ISD::CTLZ, VT, Expand);
826 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
827 setOperationAction(ISD::SHL, VT, Expand);
828 setOperationAction(ISD::SRA, VT, Expand);
829 setOperationAction(ISD::SRL, VT, Expand);
830 setOperationAction(ISD::ROTL, VT, Expand);
831 setOperationAction(ISD::ROTR, VT, Expand);
832 setOperationAction(ISD::BSWAP, VT, Expand);
833 setOperationAction(ISD::SETCC, VT, Expand);
834 setOperationAction(ISD::FLOG, VT, Expand);
835 setOperationAction(ISD::FLOG2, VT, Expand);
836 setOperationAction(ISD::FLOG10, VT, Expand);
837 setOperationAction(ISD::FEXP, VT, Expand);
838 setOperationAction(ISD::FEXP2, VT, Expand);
839 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
840 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
841 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
842 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
843 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
844 setOperationAction(ISD::TRUNCATE, VT, Expand);
845 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
846 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
847 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
848 setOperationAction(ISD::VSELECT, VT, Expand);
849 setOperationAction(ISD::SELECT_CC, VT, Expand);
850 for (MVT InnerVT : MVT::vector_valuetypes()) {
851 setTruncStoreAction(InnerVT, VT, Expand);
853 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
854 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
856 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
857 // types, we have to deal with them whether we ask for Expansion or not.
858 // Setting Expand causes its own optimisation problems though, so leave
860 if (VT.getVectorElementType() == MVT::i1)
861 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
865 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
866 // with -msoft-float, disable use of MMX as well.
867 if (!TM.Options.UseSoftFloat && Subtarget->hasMMX()) {
868 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
869 // No operations on x86mmx supported, everything uses intrinsics.
872 // MMX-sized vectors (other than x86mmx) are expected to be expanded
873 // into smaller operations.
874 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
875 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
876 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
877 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
878 setOperationAction(ISD::AND, MVT::v8i8, Expand);
879 setOperationAction(ISD::AND, MVT::v4i16, Expand);
880 setOperationAction(ISD::AND, MVT::v2i32, Expand);
881 setOperationAction(ISD::AND, MVT::v1i64, Expand);
882 setOperationAction(ISD::OR, MVT::v8i8, Expand);
883 setOperationAction(ISD::OR, MVT::v4i16, Expand);
884 setOperationAction(ISD::OR, MVT::v2i32, Expand);
885 setOperationAction(ISD::OR, MVT::v1i64, Expand);
886 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
887 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
888 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
889 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
890 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
891 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
892 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
893 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
895 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
896 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
897 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
898 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
899 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
900 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
901 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
902 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
904 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE1()) {
905 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
907 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
908 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
909 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
910 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
911 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
912 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
913 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
914 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
915 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
916 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
917 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
918 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
919 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
920 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
923 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE2()) {
924 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
926 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
927 // registers cannot be used even for integer operations.
928 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
929 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
930 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
931 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
933 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
934 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
935 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
936 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
937 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
938 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
939 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
940 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
941 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
942 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
943 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
944 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
945 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
946 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
947 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
948 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
949 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
950 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
951 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
952 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
953 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
954 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
956 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
957 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
958 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
959 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
961 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
962 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
963 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
964 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
965 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
967 // Only provide customized ctpop vector bit twiddling for vector types we
968 // know to perform better than using the popcnt instructions on each vector
969 // element. If popcnt isn't supported, always provide the custom version.
970 if (!Subtarget->hasPOPCNT()) {
971 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
972 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
975 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
976 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
977 MVT VT = (MVT::SimpleValueType)i;
978 // Do not attempt to custom lower non-power-of-2 vectors
979 if (!isPowerOf2_32(VT.getVectorNumElements()))
981 // Do not attempt to custom lower non-128-bit vectors
982 if (!VT.is128BitVector())
984 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
985 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
986 setOperationAction(ISD::VSELECT, VT, Custom);
987 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
990 // We support custom legalizing of sext and anyext loads for specific
991 // memory vector types which we can load as a scalar (or sequence of
992 // scalars) and extend in-register to a legal 128-bit vector type. For sext
993 // loads these must work with a single scalar load.
994 for (MVT VT : MVT::integer_vector_valuetypes()) {
995 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
996 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
997 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
998 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
999 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
1000 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
1001 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
1002 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
1003 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
1006 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
1007 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
1008 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
1009 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
1010 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
1011 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
1013 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
1015 if (Subtarget->is64Bit()) {
1016 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1020 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
1021 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
1022 MVT VT = (MVT::SimpleValueType)i;
1024 // Do not attempt to promote non-128-bit vectors
1025 if (!VT.is128BitVector())
1028 setOperationAction(ISD::AND, VT, Promote);
1029 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
1030 setOperationAction(ISD::OR, VT, Promote);
1031 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
1032 setOperationAction(ISD::XOR, VT, Promote);
1033 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
1034 setOperationAction(ISD::LOAD, VT, Promote);
1035 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
1036 setOperationAction(ISD::SELECT, VT, Promote);
1037 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
1040 // Custom lower v2i64 and v2f64 selects.
1041 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
1042 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
1043 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
1044 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
1046 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1047 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1049 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
1050 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
1051 // As there is no 64-bit GPR available, we need build a special custom
1052 // sequence to convert from v2i32 to v2f32.
1053 if (!Subtarget->is64Bit())
1054 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
1056 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
1057 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
1059 for (MVT VT : MVT::fp_vector_valuetypes())
1060 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
1062 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
1063 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
1064 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
1067 if (!TM.Options.UseSoftFloat && Subtarget->hasSSE41()) {
1068 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
1069 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
1070 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
1071 setOperationAction(ISD::FRINT, MVT::f32, Legal);
1072 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
1073 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
1074 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
1075 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
1076 setOperationAction(ISD::FRINT, MVT::f64, Legal);
1077 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
1079 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
1080 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
1081 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
1082 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
1083 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
1084 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
1085 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
1086 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
1087 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
1088 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
1090 // FIXME: Do we need to handle scalar-to-vector here?
1091 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
1093 // We directly match byte blends in the backend as they match the VSELECT
1095 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
1097 // SSE41 brings specific instructions for doing vector sign extend even in
1098 // cases where we don't have SRA.
1099 for (MVT VT : MVT::integer_vector_valuetypes()) {
1100 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
1101 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
1102 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
1105 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
1106 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1107 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1108 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1109 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1110 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1111 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1113 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
1114 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1115 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1116 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1117 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1118 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1120 // i8 and i16 vectors are custom because the source register and source
1121 // source memory operand types are not the same width. f32 vectors are
1122 // custom since the immediate controlling the insert encodes additional
1124 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1125 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1126 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1127 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1129 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1130 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1131 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1132 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1134 // FIXME: these should be Legal, but that's only for the case where
1135 // the index is constant. For now custom expand to deal with that.
1136 if (Subtarget->is64Bit()) {
1137 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1138 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1142 if (Subtarget->hasSSE2()) {
1143 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1144 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1146 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1147 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1149 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1150 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1152 // In the customized shift lowering, the legal cases in AVX2 will be
1154 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1155 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1157 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1158 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1160 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1163 if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
1164 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1165 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1166 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1167 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1168 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1169 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1171 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1172 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1173 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1175 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1176 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1177 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1178 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1179 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1180 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1181 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1182 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1183 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1184 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1185 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1186 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1188 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1189 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1190 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1191 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1192 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1193 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1194 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1195 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1196 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1197 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1198 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1199 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1201 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1202 // even though v8i16 is a legal type.
1203 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1204 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1205 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1207 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1208 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1209 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1211 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1212 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1214 for (MVT VT : MVT::fp_vector_valuetypes())
1215 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1217 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1218 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1220 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1221 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1223 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1224 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1226 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1227 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1228 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1229 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1232 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1233 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1235 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1236 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1237 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1238 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1239 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1240 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1241 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1242 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1243 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1244 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1245 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1246 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1248 if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
1249 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1250 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1251 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1252 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1253 setOperationAction(ISD::FMA, MVT::f32, Legal);
1254 setOperationAction(ISD::FMA, MVT::f64, Legal);
1257 if (Subtarget->hasInt256()) {
1258 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1259 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1260 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1261 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1263 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1264 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1265 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1266 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1268 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1269 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1270 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1271 // Don't lower v32i8 because there is no 128-bit byte mul
1273 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1274 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1275 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1276 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1278 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1279 // when we have a 256bit-wide blend with immediate.
1280 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1282 // Only provide customized ctpop vector bit twiddling for vector types we
1283 // know to perform better than using the popcnt instructions on each
1284 // vector element. If popcnt isn't supported, always provide the custom
1286 if (!Subtarget->hasPOPCNT())
1287 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1289 // Custom CTPOP always performs better on natively supported v8i32
1290 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1292 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1293 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1294 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1295 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1296 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1297 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1298 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1300 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1301 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1302 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1303 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1304 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1305 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1307 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1308 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1309 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1310 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1312 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1313 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1314 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1315 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1317 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1318 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1319 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1320 // Don't lower v32i8 because there is no 128-bit byte mul
1323 // In the customized shift lowering, the legal cases in AVX2 will be
1325 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1326 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1328 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1329 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1331 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1333 // Custom lower several nodes for 256-bit types.
1334 for (MVT VT : MVT::vector_valuetypes()) {
1335 if (VT.getScalarSizeInBits() >= 32) {
1336 setOperationAction(ISD::MLOAD, VT, Legal);
1337 setOperationAction(ISD::MSTORE, VT, Legal);
1339 // Extract subvector is special because the value type
1340 // (result) is 128-bit but the source is 256-bit wide.
1341 if (VT.is128BitVector()) {
1342 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1344 // Do not attempt to custom lower other non-256-bit vectors
1345 if (!VT.is256BitVector())
1348 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1349 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1350 setOperationAction(ISD::VSELECT, VT, Custom);
1351 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1352 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1353 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1354 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1355 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1358 if (Subtarget->hasInt256())
1359 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1362 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1363 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1364 MVT VT = (MVT::SimpleValueType)i;
1366 // Do not attempt to promote non-256-bit vectors
1367 if (!VT.is256BitVector())
1370 setOperationAction(ISD::AND, VT, Promote);
1371 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1372 setOperationAction(ISD::OR, VT, Promote);
1373 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1374 setOperationAction(ISD::XOR, VT, Promote);
1375 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1376 setOperationAction(ISD::LOAD, VT, Promote);
1377 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1378 setOperationAction(ISD::SELECT, VT, Promote);
1379 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1383 if (!TM.Options.UseSoftFloat && Subtarget->hasAVX512()) {
1384 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1385 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1386 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1387 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1389 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1390 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1391 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1393 for (MVT VT : MVT::fp_vector_valuetypes())
1394 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1396 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1397 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1398 setOperationAction(ISD::XOR, MVT::i1, Legal);
1399 setOperationAction(ISD::OR, MVT::i1, Legal);
1400 setOperationAction(ISD::AND, MVT::i1, Legal);
1401 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1402 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1403 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1404 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1405 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1407 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1408 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1409 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1410 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1411 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1412 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1414 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1415 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1416 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1417 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1418 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1419 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1420 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1421 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1423 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1424 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1425 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1426 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1427 if (Subtarget->is64Bit()) {
1428 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1429 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1430 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1431 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1433 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1434 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1435 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1436 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1437 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1438 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1439 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1440 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1441 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1442 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1443 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1444 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1445 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1446 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1448 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1449 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1450 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1451 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1452 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1453 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1454 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1455 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1456 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1457 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1458 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1459 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1460 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1462 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1463 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1464 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1465 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1466 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1467 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1468 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1469 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1470 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1471 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1473 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1474 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1475 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1476 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1477 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1478 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1480 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1481 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1483 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1485 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1486 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1487 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1488 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1489 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1490 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1491 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1492 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1493 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1495 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1496 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1498 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1499 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1501 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1503 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1504 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1506 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1507 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1509 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1510 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1512 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1513 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1514 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1515 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1516 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1517 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1519 if (Subtarget->hasCDI()) {
1520 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1521 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1524 // Custom lower several nodes.
1525 for (MVT VT : MVT::vector_valuetypes()) {
1526 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1527 // Extract subvector is special because the value type
1528 // (result) is 256/128-bit but the source is 512-bit wide.
1529 if (VT.is128BitVector() || VT.is256BitVector()) {
1530 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1532 if (VT.getVectorElementType() == MVT::i1)
1533 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1535 // Do not attempt to custom lower other non-512-bit vectors
1536 if (!VT.is512BitVector())
1539 if ( EltSize >= 32) {
1540 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1541 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1542 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1543 setOperationAction(ISD::VSELECT, VT, Legal);
1544 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1545 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1546 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1547 setOperationAction(ISD::MLOAD, VT, Legal);
1548 setOperationAction(ISD::MSTORE, VT, Legal);
1551 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1552 MVT VT = (MVT::SimpleValueType)i;
1554 // Do not attempt to promote non-512-bit vectors.
1555 if (!VT.is512BitVector())
1558 setOperationAction(ISD::SELECT, VT, Promote);
1559 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1563 if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
1564 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1565 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1567 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1568 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1570 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1571 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1572 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1573 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1574 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1575 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1576 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1577 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1578 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1580 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1581 const MVT VT = (MVT::SimpleValueType)i;
1583 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1585 // Do not attempt to promote non-512-bit vectors.
1586 if (!VT.is512BitVector())
1590 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1591 setOperationAction(ISD::VSELECT, VT, Legal);
1596 if (!TM.Options.UseSoftFloat && Subtarget->hasVLX()) {
1597 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1598 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1600 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1601 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1602 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Legal);
1604 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1605 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1606 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1607 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1608 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1609 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1612 // We want to custom lower some of our intrinsics.
1613 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1614 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1615 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1616 if (!Subtarget->is64Bit())
1617 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1619 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1620 // handle type legalization for these operations here.
1622 // FIXME: We really should do custom legalization for addition and
1623 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1624 // than generic legalization for 64-bit multiplication-with-overflow, though.
1625 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1626 // Add/Sub/Mul with overflow operations are custom lowered.
1628 setOperationAction(ISD::SADDO, VT, Custom);
1629 setOperationAction(ISD::UADDO, VT, Custom);
1630 setOperationAction(ISD::SSUBO, VT, Custom);
1631 setOperationAction(ISD::USUBO, VT, Custom);
1632 setOperationAction(ISD::SMULO, VT, Custom);
1633 setOperationAction(ISD::UMULO, VT, Custom);
1637 if (!Subtarget->is64Bit()) {
1638 // These libcalls are not available in 32-bit.
1639 setLibcallName(RTLIB::SHL_I128, nullptr);
1640 setLibcallName(RTLIB::SRL_I128, nullptr);
1641 setLibcallName(RTLIB::SRA_I128, nullptr);
1644 // Combine sin / cos into one node or libcall if possible.
1645 if (Subtarget->hasSinCos()) {
1646 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1647 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1648 if (Subtarget->isTargetDarwin()) {
1649 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1650 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1651 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1652 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1656 if (Subtarget->isTargetWin64()) {
1657 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1658 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1659 setOperationAction(ISD::SREM, MVT::i128, Custom);
1660 setOperationAction(ISD::UREM, MVT::i128, Custom);
1661 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1662 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1665 // We have target-specific dag combine patterns for the following nodes:
1666 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1667 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1668 setTargetDAGCombine(ISD::BITCAST);
1669 setTargetDAGCombine(ISD::VSELECT);
1670 setTargetDAGCombine(ISD::SELECT);
1671 setTargetDAGCombine(ISD::SHL);
1672 setTargetDAGCombine(ISD::SRA);
1673 setTargetDAGCombine(ISD::SRL);
1674 setTargetDAGCombine(ISD::OR);
1675 setTargetDAGCombine(ISD::AND);
1676 setTargetDAGCombine(ISD::ADD);
1677 setTargetDAGCombine(ISD::FADD);
1678 setTargetDAGCombine(ISD::FSUB);
1679 setTargetDAGCombine(ISD::FMA);
1680 setTargetDAGCombine(ISD::SUB);
1681 setTargetDAGCombine(ISD::LOAD);
1682 setTargetDAGCombine(ISD::MLOAD);
1683 setTargetDAGCombine(ISD::STORE);
1684 setTargetDAGCombine(ISD::MSTORE);
1685 setTargetDAGCombine(ISD::ZERO_EXTEND);
1686 setTargetDAGCombine(ISD::ANY_EXTEND);
1687 setTargetDAGCombine(ISD::SIGN_EXTEND);
1688 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1689 setTargetDAGCombine(ISD::TRUNCATE);
1690 setTargetDAGCombine(ISD::SINT_TO_FP);
1691 setTargetDAGCombine(ISD::SETCC);
1692 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1693 setTargetDAGCombine(ISD::BUILD_VECTOR);
1694 setTargetDAGCombine(ISD::MUL);
1695 setTargetDAGCombine(ISD::XOR);
1697 computeRegisterProperties(Subtarget->getRegisterInfo());
1699 // On Darwin, -Os means optimize for size without hurting performance,
1700 // do not reduce the limit.
1701 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1702 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1703 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1704 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1705 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1706 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1707 setPrefLoopAlignment(4); // 2^4 bytes.
1709 // Predictable cmov don't hurt on atom because it's in-order.
1710 PredictableSelectIsExpensive = !Subtarget->isAtom();
1711 EnableExtLdPromotion = true;
1712 setPrefFunctionAlignment(4); // 2^4 bytes.
1714 verifyIntrinsicTables();
1717 // This has so far only been implemented for 64-bit MachO.
1718 bool X86TargetLowering::useLoadStackGuardNode() const {
1719 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1722 TargetLoweringBase::LegalizeTypeAction
1723 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1724 if (ExperimentalVectorWideningLegalization &&
1725 VT.getVectorNumElements() != 1 &&
1726 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1727 return TypeWidenVector;
1729 return TargetLoweringBase::getPreferredVectorAction(VT);
1732 EVT X86TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
1734 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1736 const unsigned NumElts = VT.getVectorNumElements();
1737 const EVT EltVT = VT.getVectorElementType();
1738 if (VT.is512BitVector()) {
1739 if (Subtarget->hasAVX512())
1740 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1741 EltVT == MVT::f32 || EltVT == MVT::f64)
1743 case 8: return MVT::v8i1;
1744 case 16: return MVT::v16i1;
1746 if (Subtarget->hasBWI())
1747 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1749 case 32: return MVT::v32i1;
1750 case 64: return MVT::v64i1;
1754 if (VT.is256BitVector() || VT.is128BitVector()) {
1755 if (Subtarget->hasVLX())
1756 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1757 EltVT == MVT::f32 || EltVT == MVT::f64)
1759 case 2: return MVT::v2i1;
1760 case 4: return MVT::v4i1;
1761 case 8: return MVT::v8i1;
1763 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1764 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1766 case 8: return MVT::v8i1;
1767 case 16: return MVT::v16i1;
1768 case 32: return MVT::v32i1;
1772 return VT.changeVectorElementTypeToInteger();
1775 /// Helper for getByValTypeAlignment to determine
1776 /// the desired ByVal argument alignment.
1777 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1780 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1781 if (VTy->getBitWidth() == 128)
1783 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1784 unsigned EltAlign = 0;
1785 getMaxByValAlign(ATy->getElementType(), EltAlign);
1786 if (EltAlign > MaxAlign)
1787 MaxAlign = EltAlign;
1788 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1789 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1790 unsigned EltAlign = 0;
1791 getMaxByValAlign(STy->getElementType(i), EltAlign);
1792 if (EltAlign > MaxAlign)
1793 MaxAlign = EltAlign;
1800 /// Return the desired alignment for ByVal aggregate
1801 /// function arguments in the caller parameter area. For X86, aggregates
1802 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1803 /// are at 4-byte boundaries.
1804 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1805 if (Subtarget->is64Bit()) {
1806 // Max of 8 and alignment of type.
1807 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1814 if (Subtarget->hasSSE1())
1815 getMaxByValAlign(Ty, Align);
1819 /// Returns the target specific optimal type for load
1820 /// and store operations as a result of memset, memcpy, and memmove
1821 /// lowering. If DstAlign is zero that means it's safe to destination
1822 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1823 /// means there isn't a need to check it against alignment requirement,
1824 /// probably because the source does not need to be loaded. If 'IsMemset' is
1825 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1826 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1827 /// source is constant so it does not need to be loaded.
1828 /// It returns EVT::Other if the type should be determined using generic
1829 /// target-independent logic.
1831 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1832 unsigned DstAlign, unsigned SrcAlign,
1833 bool IsMemset, bool ZeroMemset,
1835 MachineFunction &MF) const {
1836 const Function *F = MF.getFunction();
1837 if ((!IsMemset || ZeroMemset) &&
1838 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1840 (Subtarget->isUnalignedMemAccessFast() ||
1841 ((DstAlign == 0 || DstAlign >= 16) &&
1842 (SrcAlign == 0 || SrcAlign >= 16)))) {
1844 if (Subtarget->hasInt256())
1846 if (Subtarget->hasFp256())
1849 if (Subtarget->hasSSE2())
1851 if (Subtarget->hasSSE1())
1853 } else if (!MemcpyStrSrc && Size >= 8 &&
1854 !Subtarget->is64Bit() &&
1855 Subtarget->hasSSE2()) {
1856 // Do not use f64 to lower memcpy if source is string constant. It's
1857 // better to use i32 to avoid the loads.
1861 if (Subtarget->is64Bit() && Size >= 8)
1866 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1868 return X86ScalarSSEf32;
1869 else if (VT == MVT::f64)
1870 return X86ScalarSSEf64;
1875 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1880 *Fast = Subtarget->isUnalignedMemAccessFast();
1884 /// Return the entry encoding for a jump table in the
1885 /// current function. The returned value is a member of the
1886 /// MachineJumpTableInfo::JTEntryKind enum.
1887 unsigned X86TargetLowering::getJumpTableEncoding() const {
1888 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1890 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1891 Subtarget->isPICStyleGOT())
1892 return MachineJumpTableInfo::EK_Custom32;
1894 // Otherwise, use the normal jump table encoding heuristics.
1895 return TargetLowering::getJumpTableEncoding();
1899 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1900 const MachineBasicBlock *MBB,
1901 unsigned uid,MCContext &Ctx) const{
1902 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1903 Subtarget->isPICStyleGOT());
1904 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1906 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1907 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1910 /// Returns relocation base for the given PIC jumptable.
1911 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1912 SelectionDAG &DAG) const {
1913 if (!Subtarget->is64Bit())
1914 // This doesn't have SDLoc associated with it, but is not really the
1915 // same as a Register.
1916 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy());
1920 /// This returns the relocation base for the given PIC jumptable,
1921 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1922 const MCExpr *X86TargetLowering::
1923 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1924 MCContext &Ctx) const {
1925 // X86-64 uses RIP relative addressing based on the jump table label.
1926 if (Subtarget->isPICStyleRIPRel())
1927 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1929 // Otherwise, the reference is relative to the PIC base.
1930 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1933 std::pair<const TargetRegisterClass *, uint8_t>
1934 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1936 const TargetRegisterClass *RRC = nullptr;
1938 switch (VT.SimpleTy) {
1940 return TargetLowering::findRepresentativeClass(TRI, VT);
1941 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1942 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1945 RRC = &X86::VR64RegClass;
1947 case MVT::f32: case MVT::f64:
1948 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1949 case MVT::v4f32: case MVT::v2f64:
1950 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1952 RRC = &X86::VR128RegClass;
1955 return std::make_pair(RRC, Cost);
1958 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1959 unsigned &Offset) const {
1960 if (!Subtarget->isTargetLinux())
1963 if (Subtarget->is64Bit()) {
1964 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1966 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1978 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
1979 unsigned DestAS) const {
1980 assert(SrcAS != DestAS && "Expected different address spaces!");
1982 return SrcAS < 256 && DestAS < 256;
1985 //===----------------------------------------------------------------------===//
1986 // Return Value Calling Convention Implementation
1987 //===----------------------------------------------------------------------===//
1989 #include "X86GenCallingConv.inc"
1992 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1993 MachineFunction &MF, bool isVarArg,
1994 const SmallVectorImpl<ISD::OutputArg> &Outs,
1995 LLVMContext &Context) const {
1996 SmallVector<CCValAssign, 16> RVLocs;
1997 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
1998 return CCInfo.CheckReturn(Outs, RetCC_X86);
2001 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2002 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2007 X86TargetLowering::LowerReturn(SDValue Chain,
2008 CallingConv::ID CallConv, bool isVarArg,
2009 const SmallVectorImpl<ISD::OutputArg> &Outs,
2010 const SmallVectorImpl<SDValue> &OutVals,
2011 SDLoc dl, SelectionDAG &DAG) const {
2012 MachineFunction &MF = DAG.getMachineFunction();
2013 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2015 SmallVector<CCValAssign, 16> RVLocs;
2016 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2017 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2020 SmallVector<SDValue, 6> RetOps;
2021 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2022 // Operand #1 = Bytes To Pop
2023 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
2026 // Copy the result values into the output registers.
2027 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2028 CCValAssign &VA = RVLocs[i];
2029 assert(VA.isRegLoc() && "Can only return in registers!");
2030 SDValue ValToCopy = OutVals[i];
2031 EVT ValVT = ValToCopy.getValueType();
2033 // Promote values to the appropriate types.
2034 if (VA.getLocInfo() == CCValAssign::SExt)
2035 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2036 else if (VA.getLocInfo() == CCValAssign::ZExt)
2037 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2038 else if (VA.getLocInfo() == CCValAssign::AExt)
2039 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2040 else if (VA.getLocInfo() == CCValAssign::BCvt)
2041 ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
2043 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2044 "Unexpected FP-extend for return value.");
2046 // If this is x86-64, and we disabled SSE, we can't return FP values,
2047 // or SSE or MMX vectors.
2048 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2049 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2050 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2051 report_fatal_error("SSE register return with SSE disabled");
2053 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2054 // llvm-gcc has never done it right and no one has noticed, so this
2055 // should be OK for now.
2056 if (ValVT == MVT::f64 &&
2057 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2058 report_fatal_error("SSE2 register return with SSE2 disabled");
2060 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2061 // the RET instruction and handled by the FP Stackifier.
2062 if (VA.getLocReg() == X86::FP0 ||
2063 VA.getLocReg() == X86::FP1) {
2064 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2065 // change the value to the FP stack register class.
2066 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2067 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2068 RetOps.push_back(ValToCopy);
2069 // Don't emit a copytoreg.
2073 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2074 // which is returned in RAX / RDX.
2075 if (Subtarget->is64Bit()) {
2076 if (ValVT == MVT::x86mmx) {
2077 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2078 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
2079 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2081 // If we don't have SSE2 available, convert to v4f32 so the generated
2082 // register is legal.
2083 if (!Subtarget->hasSSE2())
2084 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
2089 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2090 Flag = Chain.getValue(1);
2091 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2094 // The x86-64 ABIs require that for returning structs by value we copy
2095 // the sret argument into %rax/%eax (depending on ABI) for the return.
2096 // Win32 requires us to put the sret argument to %eax as well.
2097 // We saved the argument into a virtual register in the entry block,
2098 // so now we copy the value out and into %rax/%eax.
2100 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2101 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2102 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2103 // either case FuncInfo->setSRetReturnReg() will have been called.
2104 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2105 assert((Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) &&
2106 "No need for an sret register");
2107 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg, getPointerTy());
2110 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2111 X86::RAX : X86::EAX;
2112 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2113 Flag = Chain.getValue(1);
2115 // RAX/EAX now acts like a return value.
2116 RetOps.push_back(DAG.getRegister(RetValReg, getPointerTy()));
2119 RetOps[0] = Chain; // Update chain.
2121 // Add the flag if we have it.
2123 RetOps.push_back(Flag);
2125 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2128 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2129 if (N->getNumValues() != 1)
2131 if (!N->hasNUsesOfValue(1, 0))
2134 SDValue TCChain = Chain;
2135 SDNode *Copy = *N->use_begin();
2136 if (Copy->getOpcode() == ISD::CopyToReg) {
2137 // If the copy has a glue operand, we conservatively assume it isn't safe to
2138 // perform a tail call.
2139 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2141 TCChain = Copy->getOperand(0);
2142 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2145 bool HasRet = false;
2146 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2148 if (UI->getOpcode() != X86ISD::RET_FLAG)
2150 // If we are returning more than one value, we can definitely
2151 // not make a tail call see PR19530
2152 if (UI->getNumOperands() > 4)
2154 if (UI->getNumOperands() == 4 &&
2155 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2168 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2169 ISD::NodeType ExtendKind) const {
2171 // TODO: Is this also valid on 32-bit?
2172 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2173 ReturnMVT = MVT::i8;
2175 ReturnMVT = MVT::i32;
2177 EVT MinVT = getRegisterType(Context, ReturnMVT);
2178 return VT.bitsLT(MinVT) ? MinVT : VT;
2181 /// Lower the result values of a call into the
2182 /// appropriate copies out of appropriate physical registers.
2185 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2186 CallingConv::ID CallConv, bool isVarArg,
2187 const SmallVectorImpl<ISD::InputArg> &Ins,
2188 SDLoc dl, SelectionDAG &DAG,
2189 SmallVectorImpl<SDValue> &InVals) const {
2191 // Assign locations to each value returned by this call.
2192 SmallVector<CCValAssign, 16> RVLocs;
2193 bool Is64Bit = Subtarget->is64Bit();
2194 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2196 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2198 // Copy all of the result registers out of their specified physreg.
2199 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2200 CCValAssign &VA = RVLocs[i];
2201 EVT CopyVT = VA.getValVT();
2203 // If this is x86-64, and we disabled SSE, we can't return FP values
2204 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2205 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2206 report_fatal_error("SSE register return with SSE disabled");
2209 // If we prefer to use the value in xmm registers, copy it out as f80 and
2210 // use a truncate to move it from fp stack reg to xmm reg.
2211 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2212 isScalarFPTypeInSSEReg(VA.getValVT()))
2215 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2216 CopyVT, InFlag).getValue(1);
2217 SDValue Val = Chain.getValue(0);
2219 if (CopyVT != VA.getValVT())
2220 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2221 // This truncation won't change the value.
2222 DAG.getIntPtrConstant(1));
2224 InFlag = Chain.getValue(2);
2225 InVals.push_back(Val);
2231 //===----------------------------------------------------------------------===//
2232 // C & StdCall & Fast Calling Convention implementation
2233 //===----------------------------------------------------------------------===//
2234 // StdCall calling convention seems to be standard for many Windows' API
2235 // routines and around. It differs from C calling convention just a little:
2236 // callee should clean up the stack, not caller. Symbols should be also
2237 // decorated in some fancy way :) It doesn't support any vector arguments.
2238 // For info on fast calling convention see Fast Calling Convention (tail call)
2239 // implementation LowerX86_32FastCCCallTo.
2241 /// CallIsStructReturn - Determines whether a call uses struct return
2243 enum StructReturnType {
2248 static StructReturnType
2249 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2251 return NotStructReturn;
2253 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2254 if (!Flags.isSRet())
2255 return NotStructReturn;
2256 if (Flags.isInReg())
2257 return RegStructReturn;
2258 return StackStructReturn;
2261 /// Determines whether a function uses struct return semantics.
2262 static StructReturnType
2263 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2265 return NotStructReturn;
2267 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2268 if (!Flags.isSRet())
2269 return NotStructReturn;
2270 if (Flags.isInReg())
2271 return RegStructReturn;
2272 return StackStructReturn;
2275 /// Make a copy of an aggregate at address specified by "Src" to address
2276 /// "Dst" with size and alignment information specified by the specific
2277 /// parameter attribute. The copy will be passed as a byval function parameter.
2279 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2280 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2282 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
2284 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2285 /*isVolatile*/false, /*AlwaysInline=*/true,
2286 MachinePointerInfo(), MachinePointerInfo());
2289 /// Return true if the calling convention is one that
2290 /// supports tail call optimization.
2291 static bool IsTailCallConvention(CallingConv::ID CC) {
2292 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2293 CC == CallingConv::HiPE);
2296 /// \brief Return true if the calling convention is a C calling convention.
2297 static bool IsCCallConvention(CallingConv::ID CC) {
2298 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2299 CC == CallingConv::X86_64_SysV);
2302 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2303 if (!CI->isTailCall() || getTargetMachine().Options.DisableTailCalls)
2307 CallingConv::ID CalleeCC = CS.getCallingConv();
2308 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2314 /// Return true if the function is being made into
2315 /// a tailcall target by changing its ABI.
2316 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2317 bool GuaranteedTailCallOpt) {
2318 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2322 X86TargetLowering::LowerMemArgument(SDValue Chain,
2323 CallingConv::ID CallConv,
2324 const SmallVectorImpl<ISD::InputArg> &Ins,
2325 SDLoc dl, SelectionDAG &DAG,
2326 const CCValAssign &VA,
2327 MachineFrameInfo *MFI,
2329 // Create the nodes corresponding to a load from this parameter slot.
2330 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2331 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2332 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2333 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2336 // If value is passed by pointer we have address passed instead of the value
2338 if (VA.getLocInfo() == CCValAssign::Indirect)
2339 ValVT = VA.getLocVT();
2341 ValVT = VA.getValVT();
2343 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2344 // changed with more analysis.
2345 // In case of tail call optimization mark all arguments mutable. Since they
2346 // could be overwritten by lowering of arguments in case of a tail call.
2347 if (Flags.isByVal()) {
2348 unsigned Bytes = Flags.getByValSize();
2349 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2350 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2351 return DAG.getFrameIndex(FI, getPointerTy());
2353 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2354 VA.getLocMemOffset(), isImmutable);
2355 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
2356 return DAG.getLoad(ValVT, dl, Chain, FIN,
2357 MachinePointerInfo::getFixedStack(FI),
2358 false, false, false, 0);
2362 // FIXME: Get this from tablegen.
2363 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2364 const X86Subtarget *Subtarget) {
2365 assert(Subtarget->is64Bit());
2367 if (Subtarget->isCallingConvWin64(CallConv)) {
2368 static const MCPhysReg GPR64ArgRegsWin64[] = {
2369 X86::RCX, X86::RDX, X86::R8, X86::R9
2371 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2374 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2375 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2377 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2380 // FIXME: Get this from tablegen.
2381 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2382 CallingConv::ID CallConv,
2383 const X86Subtarget *Subtarget) {
2384 assert(Subtarget->is64Bit());
2385 if (Subtarget->isCallingConvWin64(CallConv)) {
2386 // The XMM registers which might contain var arg parameters are shadowed
2387 // in their paired GPR. So we only need to save the GPR to their home
2389 // TODO: __vectorcall will change this.
2393 const Function *Fn = MF.getFunction();
2394 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2395 assert(!(MF.getTarget().Options.UseSoftFloat && NoImplicitFloatOps) &&
2396 "SSE register cannot be used when SSE is disabled!");
2397 if (MF.getTarget().Options.UseSoftFloat || NoImplicitFloatOps ||
2398 !Subtarget->hasSSE1())
2399 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2403 static const MCPhysReg XMMArgRegs64Bit[] = {
2404 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2405 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2407 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2411 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2412 CallingConv::ID CallConv,
2414 const SmallVectorImpl<ISD::InputArg> &Ins,
2417 SmallVectorImpl<SDValue> &InVals)
2419 MachineFunction &MF = DAG.getMachineFunction();
2420 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2422 const Function* Fn = MF.getFunction();
2423 if (Fn->hasExternalLinkage() &&
2424 Subtarget->isTargetCygMing() &&
2425 Fn->getName() == "main")
2426 FuncInfo->setForceFramePointer(true);
2428 MachineFrameInfo *MFI = MF.getFrameInfo();
2429 bool Is64Bit = Subtarget->is64Bit();
2430 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2432 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2433 "Var args not supported with calling convention fastcc, ghc or hipe");
2435 // Assign locations to all of the incoming arguments.
2436 SmallVector<CCValAssign, 16> ArgLocs;
2437 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2439 // Allocate shadow area for Win64
2441 CCInfo.AllocateStack(32, 8);
2443 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2445 unsigned LastVal = ~0U;
2447 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2448 CCValAssign &VA = ArgLocs[i];
2449 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2451 assert(VA.getValNo() != LastVal &&
2452 "Don't support value assigned to multiple locs yet");
2454 LastVal = VA.getValNo();
2456 if (VA.isRegLoc()) {
2457 EVT RegVT = VA.getLocVT();
2458 const TargetRegisterClass *RC;
2459 if (RegVT == MVT::i32)
2460 RC = &X86::GR32RegClass;
2461 else if (Is64Bit && RegVT == MVT::i64)
2462 RC = &X86::GR64RegClass;
2463 else if (RegVT == MVT::f32)
2464 RC = &X86::FR32RegClass;
2465 else if (RegVT == MVT::f64)
2466 RC = &X86::FR64RegClass;
2467 else if (RegVT.is512BitVector())
2468 RC = &X86::VR512RegClass;
2469 else if (RegVT.is256BitVector())
2470 RC = &X86::VR256RegClass;
2471 else if (RegVT.is128BitVector())
2472 RC = &X86::VR128RegClass;
2473 else if (RegVT == MVT::x86mmx)
2474 RC = &X86::VR64RegClass;
2475 else if (RegVT == MVT::i1)
2476 RC = &X86::VK1RegClass;
2477 else if (RegVT == MVT::v8i1)
2478 RC = &X86::VK8RegClass;
2479 else if (RegVT == MVT::v16i1)
2480 RC = &X86::VK16RegClass;
2481 else if (RegVT == MVT::v32i1)
2482 RC = &X86::VK32RegClass;
2483 else if (RegVT == MVT::v64i1)
2484 RC = &X86::VK64RegClass;
2486 llvm_unreachable("Unknown argument type!");
2488 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2489 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2491 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2492 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2494 if (VA.getLocInfo() == CCValAssign::SExt)
2495 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2496 DAG.getValueType(VA.getValVT()));
2497 else if (VA.getLocInfo() == CCValAssign::ZExt)
2498 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2499 DAG.getValueType(VA.getValVT()));
2500 else if (VA.getLocInfo() == CCValAssign::BCvt)
2501 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
2503 if (VA.isExtInLoc()) {
2504 // Handle MMX values passed in XMM regs.
2505 if (RegVT.isVector())
2506 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2508 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2511 assert(VA.isMemLoc());
2512 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2515 // If value is passed via pointer - do a load.
2516 if (VA.getLocInfo() == CCValAssign::Indirect)
2517 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2518 MachinePointerInfo(), false, false, false, 0);
2520 InVals.push_back(ArgValue);
2523 if (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC()) {
2524 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2525 // The x86-64 ABIs require that for returning structs by value we copy
2526 // the sret argument into %rax/%eax (depending on ABI) for the return.
2527 // Win32 requires us to put the sret argument to %eax as well.
2528 // Save the argument into a virtual register so that we can access it
2529 // from the return points.
2530 if (Ins[i].Flags.isSRet()) {
2531 unsigned Reg = FuncInfo->getSRetReturnReg();
2533 MVT PtrTy = getPointerTy();
2534 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2535 FuncInfo->setSRetReturnReg(Reg);
2537 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2538 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2544 unsigned StackSize = CCInfo.getNextStackOffset();
2545 // Align stack specially for tail calls.
2546 if (FuncIsMadeTailCallSafe(CallConv,
2547 MF.getTarget().Options.GuaranteedTailCallOpt))
2548 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2550 // If the function takes variable number of arguments, make a frame index for
2551 // the start of the first vararg value... for expansion of llvm.va_start. We
2552 // can skip this if there are no va_start calls.
2553 if (MFI->hasVAStart() &&
2554 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2555 CallConv != CallingConv::X86_ThisCall))) {
2556 FuncInfo->setVarArgsFrameIndex(
2557 MFI->CreateFixedObject(1, StackSize, true));
2560 // Figure out if XMM registers are in use.
2561 assert(!(MF.getTarget().Options.UseSoftFloat &&
2562 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2563 "SSE register cannot be used when SSE is disabled!");
2565 // 64-bit calling conventions support varargs and register parameters, so we
2566 // have to do extra work to spill them in the prologue.
2567 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2568 // Find the first unallocated argument registers.
2569 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2570 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2571 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2572 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2573 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2574 "SSE register cannot be used when SSE is disabled!");
2576 // Gather all the live in physical registers.
2577 SmallVector<SDValue, 6> LiveGPRs;
2578 SmallVector<SDValue, 8> LiveXMMRegs;
2580 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2581 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2583 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2585 if (!ArgXMMs.empty()) {
2586 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2587 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2588 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2589 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2590 LiveXMMRegs.push_back(
2591 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2596 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2597 // Get to the caller-allocated home save location. Add 8 to account
2598 // for the return address.
2599 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2600 FuncInfo->setRegSaveFrameIndex(
2601 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2602 // Fixup to set vararg frame on shadow area (4 x i64).
2604 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2606 // For X86-64, if there are vararg parameters that are passed via
2607 // registers, then we must store them to their spots on the stack so
2608 // they may be loaded by deferencing the result of va_next.
2609 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2610 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2611 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2612 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2615 // Store the integer parameter registers.
2616 SmallVector<SDValue, 8> MemOps;
2617 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2619 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2620 for (SDValue Val : LiveGPRs) {
2621 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
2622 DAG.getIntPtrConstant(Offset));
2624 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2625 MachinePointerInfo::getFixedStack(
2626 FuncInfo->getRegSaveFrameIndex(), Offset),
2628 MemOps.push_back(Store);
2632 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2633 // Now store the XMM (fp + vector) parameter registers.
2634 SmallVector<SDValue, 12> SaveXMMOps;
2635 SaveXMMOps.push_back(Chain);
2636 SaveXMMOps.push_back(ALVal);
2637 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2638 FuncInfo->getRegSaveFrameIndex()));
2639 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2640 FuncInfo->getVarArgsFPOffset()));
2641 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2643 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2644 MVT::Other, SaveXMMOps));
2647 if (!MemOps.empty())
2648 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2651 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2652 // Find the largest legal vector type.
2653 MVT VecVT = MVT::Other;
2654 // FIXME: Only some x86_32 calling conventions support AVX512.
2655 if (Subtarget->hasAVX512() &&
2656 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2657 CallConv == CallingConv::Intel_OCL_BI)))
2658 VecVT = MVT::v16f32;
2659 else if (Subtarget->hasAVX())
2661 else if (Subtarget->hasSSE2())
2664 // We forward some GPRs and some vector types.
2665 SmallVector<MVT, 2> RegParmTypes;
2666 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2667 RegParmTypes.push_back(IntVT);
2668 if (VecVT != MVT::Other)
2669 RegParmTypes.push_back(VecVT);
2671 // Compute the set of forwarded registers. The rest are scratch.
2672 SmallVectorImpl<ForwardedRegister> &Forwards =
2673 FuncInfo->getForwardedMustTailRegParms();
2674 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2676 // Conservatively forward AL on x86_64, since it might be used for varargs.
2677 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2678 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2679 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2682 // Copy all forwards from physical to virtual registers.
2683 for (ForwardedRegister &F : Forwards) {
2684 // FIXME: Can we use a less constrained schedule?
2685 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2686 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2687 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2691 // Some CCs need callee pop.
2692 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2693 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2694 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2696 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2697 // If this is an sret function, the return should pop the hidden pointer.
2698 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2699 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2700 argsAreStructReturn(Ins) == StackStructReturn)
2701 FuncInfo->setBytesToPopOnReturn(4);
2705 // RegSaveFrameIndex is X86-64 only.
2706 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2707 if (CallConv == CallingConv::X86_FastCall ||
2708 CallConv == CallingConv::X86_ThisCall)
2709 // fastcc functions can't have varargs.
2710 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2713 FuncInfo->setArgumentStackSize(StackSize);
2719 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2720 SDValue StackPtr, SDValue Arg,
2721 SDLoc dl, SelectionDAG &DAG,
2722 const CCValAssign &VA,
2723 ISD::ArgFlagsTy Flags) const {
2724 unsigned LocMemOffset = VA.getLocMemOffset();
2725 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
2726 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
2727 if (Flags.isByVal())
2728 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2730 return DAG.getStore(Chain, dl, Arg, PtrOff,
2731 MachinePointerInfo::getStack(LocMemOffset),
2735 /// Emit a load of return address if tail call
2736 /// optimization is performed and it is required.
2738 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2739 SDValue &OutRetAddr, SDValue Chain,
2740 bool IsTailCall, bool Is64Bit,
2741 int FPDiff, SDLoc dl) const {
2742 // Adjust the Return address stack slot.
2743 EVT VT = getPointerTy();
2744 OutRetAddr = getReturnAddressFrameIndex(DAG);
2746 // Load the "old" Return address.
2747 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2748 false, false, false, 0);
2749 return SDValue(OutRetAddr.getNode(), 1);
2752 /// Emit a store of the return address if tail call
2753 /// optimization is performed and it is required (FPDiff!=0).
2754 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2755 SDValue Chain, SDValue RetAddrFrIdx,
2756 EVT PtrVT, unsigned SlotSize,
2757 int FPDiff, SDLoc dl) {
2758 // Store the return address to the appropriate stack slot.
2759 if (!FPDiff) return Chain;
2760 // Calculate the new stack slot for the return address.
2761 int NewReturnAddrFI =
2762 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2764 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2765 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2766 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2772 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2773 SmallVectorImpl<SDValue> &InVals) const {
2774 SelectionDAG &DAG = CLI.DAG;
2776 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2777 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2778 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2779 SDValue Chain = CLI.Chain;
2780 SDValue Callee = CLI.Callee;
2781 CallingConv::ID CallConv = CLI.CallConv;
2782 bool &isTailCall = CLI.IsTailCall;
2783 bool isVarArg = CLI.IsVarArg;
2785 MachineFunction &MF = DAG.getMachineFunction();
2786 bool Is64Bit = Subtarget->is64Bit();
2787 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2788 StructReturnType SR = callIsStructReturn(Outs);
2789 bool IsSibcall = false;
2790 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2792 if (MF.getTarget().Options.DisableTailCalls)
2795 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2797 // Force this to be a tail call. The verifier rules are enough to ensure
2798 // that we can lower this successfully without moving the return address
2801 } else if (isTailCall) {
2802 // Check if it's really possible to do a tail call.
2803 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2804 isVarArg, SR != NotStructReturn,
2805 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2806 Outs, OutVals, Ins, DAG);
2808 // Sibcalls are automatically detected tailcalls which do not require
2810 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2817 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2818 "Var args not supported with calling convention fastcc, ghc or hipe");
2820 // Analyze operands of the call, assigning locations to each operand.
2821 SmallVector<CCValAssign, 16> ArgLocs;
2822 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2824 // Allocate shadow area for Win64
2826 CCInfo.AllocateStack(32, 8);
2828 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2830 // Get a count of how many bytes are to be pushed on the stack.
2831 unsigned NumBytes = CCInfo.getNextStackOffset();
2833 // This is a sibcall. The memory operands are available in caller's
2834 // own caller's stack.
2836 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2837 IsTailCallConvention(CallConv))
2838 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2841 if (isTailCall && !IsSibcall && !IsMustTail) {
2842 // Lower arguments at fp - stackoffset + fpdiff.
2843 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2845 FPDiff = NumBytesCallerPushed - NumBytes;
2847 // Set the delta of movement of the returnaddr stackslot.
2848 // But only set if delta is greater than previous delta.
2849 if (FPDiff < X86Info->getTCReturnAddrDelta())
2850 X86Info->setTCReturnAddrDelta(FPDiff);
2853 unsigned NumBytesToPush = NumBytes;
2854 unsigned NumBytesToPop = NumBytes;
2856 // If we have an inalloca argument, all stack space has already been allocated
2857 // for us and be right at the top of the stack. We don't support multiple
2858 // arguments passed in memory when using inalloca.
2859 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2861 if (!ArgLocs.back().isMemLoc())
2862 report_fatal_error("cannot use inalloca attribute on a register "
2864 if (ArgLocs.back().getLocMemOffset() != 0)
2865 report_fatal_error("any parameter with the inalloca attribute must be "
2866 "the only memory argument");
2870 Chain = DAG.getCALLSEQ_START(
2871 Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
2873 SDValue RetAddrFrIdx;
2874 // Load return address for tail calls.
2875 if (isTailCall && FPDiff)
2876 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2877 Is64Bit, FPDiff, dl);
2879 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2880 SmallVector<SDValue, 8> MemOpChains;
2883 // Walk the register/memloc assignments, inserting copies/loads. In the case
2884 // of tail call optimization arguments are handle later.
2885 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
2886 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2887 // Skip inalloca arguments, they have already been written.
2888 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2889 if (Flags.isInAlloca())
2892 CCValAssign &VA = ArgLocs[i];
2893 EVT RegVT = VA.getLocVT();
2894 SDValue Arg = OutVals[i];
2895 bool isByVal = Flags.isByVal();
2897 // Promote the value if needed.
2898 switch (VA.getLocInfo()) {
2899 default: llvm_unreachable("Unknown loc info!");
2900 case CCValAssign::Full: break;
2901 case CCValAssign::SExt:
2902 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2904 case CCValAssign::ZExt:
2905 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2907 case CCValAssign::AExt:
2908 if (RegVT.is128BitVector()) {
2909 // Special case: passing MMX values in XMM registers.
2910 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2911 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2912 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2914 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2916 case CCValAssign::BCvt:
2917 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2919 case CCValAssign::Indirect: {
2920 // Store the argument.
2921 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2922 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2923 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2924 MachinePointerInfo::getFixedStack(FI),
2931 if (VA.isRegLoc()) {
2932 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2933 if (isVarArg && IsWin64) {
2934 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2935 // shadow reg if callee is a varargs function.
2936 unsigned ShadowReg = 0;
2937 switch (VA.getLocReg()) {
2938 case X86::XMM0: ShadowReg = X86::RCX; break;
2939 case X86::XMM1: ShadowReg = X86::RDX; break;
2940 case X86::XMM2: ShadowReg = X86::R8; break;
2941 case X86::XMM3: ShadowReg = X86::R9; break;
2944 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2946 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2947 assert(VA.isMemLoc());
2948 if (!StackPtr.getNode())
2949 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
2951 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2952 dl, DAG, VA, Flags));
2956 if (!MemOpChains.empty())
2957 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
2959 if (Subtarget->isPICStyleGOT()) {
2960 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2963 RegsToPass.push_back(std::make_pair(unsigned(X86::EBX),
2964 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), getPointerTy())));
2966 // If we are tail calling and generating PIC/GOT style code load the
2967 // address of the callee into ECX. The value in ecx is used as target of
2968 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2969 // for tail calls on PIC/GOT architectures. Normally we would just put the
2970 // address of GOT into ebx and then call target@PLT. But for tail calls
2971 // ebx would be restored (since ebx is callee saved) before jumping to the
2974 // Note: The actual moving to ECX is done further down.
2975 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2976 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2977 !G->getGlobal()->hasProtectedVisibility())
2978 Callee = LowerGlobalAddress(Callee, DAG);
2979 else if (isa<ExternalSymbolSDNode>(Callee))
2980 Callee = LowerExternalSymbol(Callee, DAG);
2984 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
2985 // From AMD64 ABI document:
2986 // For calls that may call functions that use varargs or stdargs
2987 // (prototype-less calls or calls to functions containing ellipsis (...) in
2988 // the declaration) %al is used as hidden argument to specify the number
2989 // of SSE registers used. The contents of %al do not need to match exactly
2990 // the number of registers, but must be an ubound on the number of SSE
2991 // registers used and is in the range 0 - 8 inclusive.
2993 // Count the number of XMM registers allocated.
2994 static const MCPhysReg XMMArgRegs[] = {
2995 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2996 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2998 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
2999 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3000 && "SSE registers cannot be used when SSE is disabled");
3002 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3003 DAG.getConstant(NumXMMRegs, MVT::i8)));
3006 if (isVarArg && IsMustTail) {
3007 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3008 for (const auto &F : Forwards) {
3009 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3010 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3014 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3015 // don't need this because the eligibility check rejects calls that require
3016 // shuffling arguments passed in memory.
3017 if (!IsSibcall && isTailCall) {
3018 // Force all the incoming stack arguments to be loaded from the stack
3019 // before any new outgoing arguments are stored to the stack, because the
3020 // outgoing stack slots may alias the incoming argument stack slots, and
3021 // the alias isn't otherwise explicit. This is slightly more conservative
3022 // than necessary, because it means that each store effectively depends
3023 // on every argument instead of just those arguments it would clobber.
3024 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3026 SmallVector<SDValue, 8> MemOpChains2;
3029 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3030 CCValAssign &VA = ArgLocs[i];
3033 assert(VA.isMemLoc());
3034 SDValue Arg = OutVals[i];
3035 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3036 // Skip inalloca arguments. They don't require any work.
3037 if (Flags.isInAlloca())
3039 // Create frame index.
3040 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3041 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3042 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3043 FIN = DAG.getFrameIndex(FI, getPointerTy());
3045 if (Flags.isByVal()) {
3046 // Copy relative to framepointer.
3047 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
3048 if (!StackPtr.getNode())
3049 StackPtr = DAG.getCopyFromReg(Chain, dl,
3050 RegInfo->getStackRegister(),
3052 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
3054 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3058 // Store relative to framepointer.
3059 MemOpChains2.push_back(
3060 DAG.getStore(ArgChain, dl, Arg, FIN,
3061 MachinePointerInfo::getFixedStack(FI),
3066 if (!MemOpChains2.empty())
3067 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3069 // Store the return address to the appropriate stack slot.
3070 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3071 getPointerTy(), RegInfo->getSlotSize(),
3075 // Build a sequence of copy-to-reg nodes chained together with token chain
3076 // and flag operands which copy the outgoing args into registers.
3078 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3079 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3080 RegsToPass[i].second, InFlag);
3081 InFlag = Chain.getValue(1);
3084 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3085 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3086 // In the 64-bit large code model, we have to make all calls
3087 // through a register, since the call instruction's 32-bit
3088 // pc-relative offset may not be large enough to hold the whole
3090 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3091 // If the callee is a GlobalAddress node (quite common, every direct call
3092 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3094 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3096 // We should use extra load for direct calls to dllimported functions in
3098 const GlobalValue *GV = G->getGlobal();
3099 if (!GV->hasDLLImportStorageClass()) {
3100 unsigned char OpFlags = 0;
3101 bool ExtraLoad = false;
3102 unsigned WrapperKind = ISD::DELETED_NODE;
3104 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3105 // external symbols most go through the PLT in PIC mode. If the symbol
3106 // has hidden or protected visibility, or if it is static or local, then
3107 // we don't need to use the PLT - we can directly call it.
3108 if (Subtarget->isTargetELF() &&
3109 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3110 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3111 OpFlags = X86II::MO_PLT;
3112 } else if (Subtarget->isPICStyleStubAny() &&
3113 (GV->isDeclaration() || GV->isWeakForLinker()) &&
3114 (!Subtarget->getTargetTriple().isMacOSX() ||
3115 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3116 // PC-relative references to external symbols should go through $stub,
3117 // unless we're building with the leopard linker or later, which
3118 // automatically synthesizes these stubs.
3119 OpFlags = X86II::MO_DARWIN_STUB;
3120 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3121 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3122 // If the function is marked as non-lazy, generate an indirect call
3123 // which loads from the GOT directly. This avoids runtime overhead
3124 // at the cost of eager binding (and one extra byte of encoding).
3125 OpFlags = X86II::MO_GOTPCREL;
3126 WrapperKind = X86ISD::WrapperRIP;
3130 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
3131 G->getOffset(), OpFlags);
3133 // Add a wrapper if needed.
3134 if (WrapperKind != ISD::DELETED_NODE)
3135 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
3136 // Add extra indirection if needed.
3138 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
3139 MachinePointerInfo::getGOT(),
3140 false, false, false, 0);
3142 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3143 unsigned char OpFlags = 0;
3145 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3146 // external symbols should go through the PLT.
3147 if (Subtarget->isTargetELF() &&
3148 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3149 OpFlags = X86II::MO_PLT;
3150 } else if (Subtarget->isPICStyleStubAny() &&
3151 (!Subtarget->getTargetTriple().isMacOSX() ||
3152 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3153 // PC-relative references to external symbols should go through $stub,
3154 // unless we're building with the leopard linker or later, which
3155 // automatically synthesizes these stubs.
3156 OpFlags = X86II::MO_DARWIN_STUB;
3159 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
3161 } else if (Subtarget->isTarget64BitILP32() &&
3162 Callee->getValueType(0) == MVT::i32) {
3163 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3164 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3167 // Returns a chain & a flag for retval copy to use.
3168 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3169 SmallVector<SDValue, 8> Ops;
3171 if (!IsSibcall && isTailCall) {
3172 Chain = DAG.getCALLSEQ_END(Chain,
3173 DAG.getIntPtrConstant(NumBytesToPop, true),
3174 DAG.getIntPtrConstant(0, true), InFlag, dl);
3175 InFlag = Chain.getValue(1);
3178 Ops.push_back(Chain);
3179 Ops.push_back(Callee);
3182 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
3184 // Add argument registers to the end of the list so that they are known live
3186 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3187 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3188 RegsToPass[i].second.getValueType()));
3190 // Add a register mask operand representing the call-preserved registers.
3191 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
3192 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
3193 assert(Mask && "Missing call preserved mask for calling convention");
3194 Ops.push_back(DAG.getRegisterMask(Mask));
3196 if (InFlag.getNode())
3197 Ops.push_back(InFlag);
3201 //// If this is the first return lowered for this function, add the regs
3202 //// to the liveout set for the function.
3203 // This isn't right, although it's probably harmless on x86; liveouts
3204 // should be computed from returns not tail calls. Consider a void
3205 // function making a tail call to a function returning int.
3206 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3209 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3210 InFlag = Chain.getValue(1);
3212 // Create the CALLSEQ_END node.
3213 unsigned NumBytesForCalleeToPop;
3214 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3215 DAG.getTarget().Options.GuaranteedTailCallOpt))
3216 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3217 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3218 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3219 SR == StackStructReturn)
3220 // If this is a call to a struct-return function, the callee
3221 // pops the hidden struct pointer, so we have to push it back.
3222 // This is common for Darwin/X86, Linux & Mingw32 targets.
3223 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3224 NumBytesForCalleeToPop = 4;
3226 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3228 // Returns a flag for retval copy to use.
3230 Chain = DAG.getCALLSEQ_END(Chain,
3231 DAG.getIntPtrConstant(NumBytesToPop, true),
3232 DAG.getIntPtrConstant(NumBytesForCalleeToPop,
3235 InFlag = Chain.getValue(1);
3238 // Handle result values, copying them out of physregs into vregs that we
3240 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3241 Ins, dl, DAG, InVals);
3244 //===----------------------------------------------------------------------===//
3245 // Fast Calling Convention (tail call) implementation
3246 //===----------------------------------------------------------------------===//
3248 // Like std call, callee cleans arguments, convention except that ECX is
3249 // reserved for storing the tail called function address. Only 2 registers are
3250 // free for argument passing (inreg). Tail call optimization is performed
3252 // * tailcallopt is enabled
3253 // * caller/callee are fastcc
3254 // On X86_64 architecture with GOT-style position independent code only local
3255 // (within module) calls are supported at the moment.
3256 // To keep the stack aligned according to platform abi the function
3257 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3258 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3259 // If a tail called function callee has more arguments than the caller the
3260 // caller needs to make sure that there is room to move the RETADDR to. This is
3261 // achieved by reserving an area the size of the argument delta right after the
3262 // original RETADDR, but before the saved framepointer or the spilled registers
3263 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3275 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
3276 /// for a 16 byte align requirement.
3278 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3279 SelectionDAG& DAG) const {
3280 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3281 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3282 unsigned StackAlignment = TFI.getStackAlignment();
3283 uint64_t AlignMask = StackAlignment - 1;
3284 int64_t Offset = StackSize;
3285 unsigned SlotSize = RegInfo->getSlotSize();
3286 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3287 // Number smaller than 12 so just add the difference.
3288 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3290 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3291 Offset = ((~AlignMask) & Offset) + StackAlignment +
3292 (StackAlignment-SlotSize);
3297 /// MatchingStackOffset - Return true if the given stack call argument is
3298 /// already available in the same position (relatively) of the caller's
3299 /// incoming argument stack.
3301 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3302 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3303 const X86InstrInfo *TII) {
3304 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3306 if (Arg.getOpcode() == ISD::CopyFromReg) {
3307 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3308 if (!TargetRegisterInfo::isVirtualRegister(VR))
3310 MachineInstr *Def = MRI->getVRegDef(VR);
3313 if (!Flags.isByVal()) {
3314 if (!TII->isLoadFromStackSlot(Def, FI))
3317 unsigned Opcode = Def->getOpcode();
3318 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3319 Opcode == X86::LEA64_32r) &&
3320 Def->getOperand(1).isFI()) {
3321 FI = Def->getOperand(1).getIndex();
3322 Bytes = Flags.getByValSize();
3326 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3327 if (Flags.isByVal())
3328 // ByVal argument is passed in as a pointer but it's now being
3329 // dereferenced. e.g.
3330 // define @foo(%struct.X* %A) {
3331 // tail call @bar(%struct.X* byval %A)
3334 SDValue Ptr = Ld->getBasePtr();
3335 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3338 FI = FINode->getIndex();
3339 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3340 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3341 FI = FINode->getIndex();
3342 Bytes = Flags.getByValSize();
3346 assert(FI != INT_MAX);
3347 if (!MFI->isFixedObjectIndex(FI))
3349 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3352 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
3353 /// for tail call optimization. Targets which want to do tail call
3354 /// optimization should implement this function.
3356 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3357 CallingConv::ID CalleeCC,
3359 bool isCalleeStructRet,
3360 bool isCallerStructRet,
3362 const SmallVectorImpl<ISD::OutputArg> &Outs,
3363 const SmallVectorImpl<SDValue> &OutVals,
3364 const SmallVectorImpl<ISD::InputArg> &Ins,
3365 SelectionDAG &DAG) const {
3366 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3369 // If -tailcallopt is specified, make fastcc functions tail-callable.
3370 const MachineFunction &MF = DAG.getMachineFunction();
3371 const Function *CallerF = MF.getFunction();
3373 // If the function return type is x86_fp80 and the callee return type is not,
3374 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3375 // perform a tailcall optimization here.
3376 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3379 CallingConv::ID CallerCC = CallerF->getCallingConv();
3380 bool CCMatch = CallerCC == CalleeCC;
3381 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3382 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3384 // Win64 functions have extra shadow space for argument homing. Don't do the
3385 // sibcall if the caller and callee have mismatched expectations for this
3387 if (IsCalleeWin64 != IsCallerWin64)
3390 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3391 if (IsTailCallConvention(CalleeCC) && CCMatch)
3396 // Look for obvious safe cases to perform tail call optimization that do not
3397 // require ABI changes. This is what gcc calls sibcall.
3399 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3400 // emit a special epilogue.
3401 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3402 if (RegInfo->needsStackRealignment(MF))
3405 // Also avoid sibcall optimization if either caller or callee uses struct
3406 // return semantics.
3407 if (isCalleeStructRet || isCallerStructRet)
3410 // An stdcall/thiscall caller is expected to clean up its arguments; the
3411 // callee isn't going to do that.
3412 // FIXME: this is more restrictive than needed. We could produce a tailcall
3413 // when the stack adjustment matches. For example, with a thiscall that takes
3414 // only one argument.
3415 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3416 CallerCC == CallingConv::X86_ThisCall))
3419 // Do not sibcall optimize vararg calls unless all arguments are passed via
3421 if (isVarArg && !Outs.empty()) {
3423 // Optimizing for varargs on Win64 is unlikely to be safe without
3424 // additional testing.
3425 if (IsCalleeWin64 || IsCallerWin64)
3428 SmallVector<CCValAssign, 16> ArgLocs;
3429 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3432 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3433 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3434 if (!ArgLocs[i].isRegLoc())
3438 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3439 // stack. Therefore, if it's not used by the call it is not safe to optimize
3440 // this into a sibcall.
3441 bool Unused = false;
3442 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3449 SmallVector<CCValAssign, 16> RVLocs;
3450 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3452 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3453 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3454 CCValAssign &VA = RVLocs[i];
3455 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3460 // If the calling conventions do not match, then we'd better make sure the
3461 // results are returned in the same way as what the caller expects.
3463 SmallVector<CCValAssign, 16> RVLocs1;
3464 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3466 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3468 SmallVector<CCValAssign, 16> RVLocs2;
3469 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3471 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3473 if (RVLocs1.size() != RVLocs2.size())
3475 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3476 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3478 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3480 if (RVLocs1[i].isRegLoc()) {
3481 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3484 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3490 // If the callee takes no arguments then go on to check the results of the
3492 if (!Outs.empty()) {
3493 // Check if stack adjustment is needed. For now, do not do this if any
3494 // argument is passed on the stack.
3495 SmallVector<CCValAssign, 16> ArgLocs;
3496 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3499 // Allocate shadow area for Win64
3501 CCInfo.AllocateStack(32, 8);
3503 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3504 if (CCInfo.getNextStackOffset()) {
3505 MachineFunction &MF = DAG.getMachineFunction();
3506 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3509 // Check if the arguments are already laid out in the right way as
3510 // the caller's fixed stack objects.
3511 MachineFrameInfo *MFI = MF.getFrameInfo();
3512 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3513 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3514 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3515 CCValAssign &VA = ArgLocs[i];
3516 SDValue Arg = OutVals[i];
3517 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3518 if (VA.getLocInfo() == CCValAssign::Indirect)
3520 if (!VA.isRegLoc()) {
3521 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3528 // If the tailcall address may be in a register, then make sure it's
3529 // possible to register allocate for it. In 32-bit, the call address can
3530 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3531 // callee-saved registers are restored. These happen to be the same
3532 // registers used to pass 'inreg' arguments so watch out for those.
3533 if (!Subtarget->is64Bit() &&
3534 ((!isa<GlobalAddressSDNode>(Callee) &&
3535 !isa<ExternalSymbolSDNode>(Callee)) ||
3536 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3537 unsigned NumInRegs = 0;
3538 // In PIC we need an extra register to formulate the address computation
3540 unsigned MaxInRegs =
3541 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3543 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3544 CCValAssign &VA = ArgLocs[i];
3547 unsigned Reg = VA.getLocReg();
3550 case X86::EAX: case X86::EDX: case X86::ECX:
3551 if (++NumInRegs == MaxInRegs)
3563 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3564 const TargetLibraryInfo *libInfo) const {
3565 return X86::createFastISel(funcInfo, libInfo);
3568 //===----------------------------------------------------------------------===//
3569 // Other Lowering Hooks
3570 //===----------------------------------------------------------------------===//
3572 static bool MayFoldLoad(SDValue Op) {
3573 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3576 static bool MayFoldIntoStore(SDValue Op) {
3577 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3580 static bool isTargetShuffle(unsigned Opcode) {
3582 default: return false;
3583 case X86ISD::BLENDI:
3584 case X86ISD::PSHUFB:
3585 case X86ISD::PSHUFD:
3586 case X86ISD::PSHUFHW:
3587 case X86ISD::PSHUFLW:
3589 case X86ISD::PALIGNR:
3590 case X86ISD::MOVLHPS:
3591 case X86ISD::MOVLHPD:
3592 case X86ISD::MOVHLPS:
3593 case X86ISD::MOVLPS:
3594 case X86ISD::MOVLPD:
3595 case X86ISD::MOVSHDUP:
3596 case X86ISD::MOVSLDUP:
3597 case X86ISD::MOVDDUP:
3600 case X86ISD::UNPCKL:
3601 case X86ISD::UNPCKH:
3602 case X86ISD::VPERMILPI:
3603 case X86ISD::VPERM2X128:
3604 case X86ISD::VPERMI:
3609 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3610 SDValue V1, unsigned TargetMask,
3611 SelectionDAG &DAG) {
3613 default: llvm_unreachable("Unknown x86 shuffle node");
3614 case X86ISD::PSHUFD:
3615 case X86ISD::PSHUFHW:
3616 case X86ISD::PSHUFLW:
3617 case X86ISD::VPERMILPI:
3618 case X86ISD::VPERMI:
3619 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
3623 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3624 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3626 default: llvm_unreachable("Unknown x86 shuffle node");
3627 case X86ISD::MOVLHPS:
3628 case X86ISD::MOVLHPD:
3629 case X86ISD::MOVHLPS:
3630 case X86ISD::MOVLPS:
3631 case X86ISD::MOVLPD:
3634 case X86ISD::UNPCKL:
3635 case X86ISD::UNPCKH:
3636 return DAG.getNode(Opc, dl, VT, V1, V2);
3640 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3641 MachineFunction &MF = DAG.getMachineFunction();
3642 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3643 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3644 int ReturnAddrIndex = FuncInfo->getRAIndex();
3646 if (ReturnAddrIndex == 0) {
3647 // Set up a frame object for the return address.
3648 unsigned SlotSize = RegInfo->getSlotSize();
3649 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3652 FuncInfo->setRAIndex(ReturnAddrIndex);
3655 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
3658 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3659 bool hasSymbolicDisplacement) {
3660 // Offset should fit into 32 bit immediate field.
3661 if (!isInt<32>(Offset))
3664 // If we don't have a symbolic displacement - we don't have any extra
3666 if (!hasSymbolicDisplacement)
3669 // FIXME: Some tweaks might be needed for medium code model.
3670 if (M != CodeModel::Small && M != CodeModel::Kernel)
3673 // For small code model we assume that latest object is 16MB before end of 31
3674 // bits boundary. We may also accept pretty large negative constants knowing
3675 // that all objects are in the positive half of address space.
3676 if (M == CodeModel::Small && Offset < 16*1024*1024)
3679 // For kernel code model we know that all object resist in the negative half
3680 // of 32bits address space. We may not accept negative offsets, since they may
3681 // be just off and we may accept pretty large positive ones.
3682 if (M == CodeModel::Kernel && Offset >= 0)
3688 /// isCalleePop - Determines whether the callee is required to pop its
3689 /// own arguments. Callee pop is necessary to support tail calls.
3690 bool X86::isCalleePop(CallingConv::ID CallingConv,
3691 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3692 switch (CallingConv) {
3695 case CallingConv::X86_StdCall:
3696 case CallingConv::X86_FastCall:
3697 case CallingConv::X86_ThisCall:
3699 case CallingConv::Fast:
3700 case CallingConv::GHC:
3701 case CallingConv::HiPE:
3708 /// \brief Return true if the condition is an unsigned comparison operation.
3709 static bool isX86CCUnsigned(unsigned X86CC) {
3711 default: llvm_unreachable("Invalid integer condition!");
3712 case X86::COND_E: return true;
3713 case X86::COND_G: return false;
3714 case X86::COND_GE: return false;
3715 case X86::COND_L: return false;
3716 case X86::COND_LE: return false;
3717 case X86::COND_NE: return true;
3718 case X86::COND_B: return true;
3719 case X86::COND_A: return true;
3720 case X86::COND_BE: return true;
3721 case X86::COND_AE: return true;
3723 llvm_unreachable("covered switch fell through?!");
3726 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
3727 /// specific condition code, returning the condition code and the LHS/RHS of the
3728 /// comparison to make.
3729 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
3730 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3732 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3733 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3734 // X > -1 -> X == 0, jump !sign.
3735 RHS = DAG.getConstant(0, RHS.getValueType());
3736 return X86::COND_NS;
3738 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3739 // X < 0 -> X == 0, jump on sign.
3742 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3744 RHS = DAG.getConstant(0, RHS.getValueType());
3745 return X86::COND_LE;
3749 switch (SetCCOpcode) {
3750 default: llvm_unreachable("Invalid integer condition!");
3751 case ISD::SETEQ: return X86::COND_E;
3752 case ISD::SETGT: return X86::COND_G;
3753 case ISD::SETGE: return X86::COND_GE;
3754 case ISD::SETLT: return X86::COND_L;
3755 case ISD::SETLE: return X86::COND_LE;
3756 case ISD::SETNE: return X86::COND_NE;
3757 case ISD::SETULT: return X86::COND_B;
3758 case ISD::SETUGT: return X86::COND_A;
3759 case ISD::SETULE: return X86::COND_BE;
3760 case ISD::SETUGE: return X86::COND_AE;
3764 // First determine if it is required or is profitable to flip the operands.
3766 // If LHS is a foldable load, but RHS is not, flip the condition.
3767 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3768 !ISD::isNON_EXTLoad(RHS.getNode())) {
3769 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3770 std::swap(LHS, RHS);
3773 switch (SetCCOpcode) {
3779 std::swap(LHS, RHS);
3783 // On a floating point condition, the flags are set as follows:
3785 // 0 | 0 | 0 | X > Y
3786 // 0 | 0 | 1 | X < Y
3787 // 1 | 0 | 0 | X == Y
3788 // 1 | 1 | 1 | unordered
3789 switch (SetCCOpcode) {
3790 default: llvm_unreachable("Condcode should be pre-legalized away");
3792 case ISD::SETEQ: return X86::COND_E;
3793 case ISD::SETOLT: // flipped
3795 case ISD::SETGT: return X86::COND_A;
3796 case ISD::SETOLE: // flipped
3798 case ISD::SETGE: return X86::COND_AE;
3799 case ISD::SETUGT: // flipped
3801 case ISD::SETLT: return X86::COND_B;
3802 case ISD::SETUGE: // flipped
3804 case ISD::SETLE: return X86::COND_BE;
3806 case ISD::SETNE: return X86::COND_NE;
3807 case ISD::SETUO: return X86::COND_P;
3808 case ISD::SETO: return X86::COND_NP;
3810 case ISD::SETUNE: return X86::COND_INVALID;
3814 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3815 /// code. Current x86 isa includes the following FP cmov instructions:
3816 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3817 static bool hasFPCMov(unsigned X86CC) {
3833 /// isFPImmLegal - Returns true if the target can instruction select the
3834 /// specified FP immediate natively. If false, the legalizer will
3835 /// materialize the FP immediate as a load from a constant pool.
3836 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3837 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3838 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3844 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3845 ISD::LoadExtType ExtTy,
3847 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3848 // relocation target a movq or addq instruction: don't let the load shrink.
3849 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3850 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3851 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3852 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3856 /// \brief Returns true if it is beneficial to convert a load of a constant
3857 /// to just the constant itself.
3858 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
3860 assert(Ty->isIntegerTy());
3862 unsigned BitSize = Ty->getPrimitiveSizeInBits();
3863 if (BitSize == 0 || BitSize > 64)
3868 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
3869 unsigned Index) const {
3870 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
3873 return (Index == 0 || Index == ResVT.getVectorNumElements());
3876 bool X86TargetLowering::isCheapToSpeculateCttz() const {
3877 // Speculate cttz only if we can directly use TZCNT.
3878 return Subtarget->hasBMI();
3881 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
3882 // Speculate ctlz only if we can directly use LZCNT.
3883 return Subtarget->hasLZCNT();
3886 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3887 /// the specified range (L, H].
3888 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3889 return (Val < 0) || (Val >= Low && Val < Hi);
3892 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3893 /// specified value.
3894 static bool isUndefOrEqual(int Val, int CmpVal) {
3895 return (Val < 0 || Val == CmpVal);
3898 /// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
3899 /// from position Pos and ending in Pos+Size, falls within the specified
3900 /// sequential range (Low, Low+Size]. or is undef.
3901 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
3902 unsigned Pos, unsigned Size, int Low) {
3903 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3904 if (!isUndefOrEqual(Mask[i], Low))
3909 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3910 /// the two vector operands have swapped position.
3911 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
3912 unsigned NumElems) {
3913 for (unsigned i = 0; i != NumElems; ++i) {
3917 else if (idx < (int)NumElems)
3918 Mask[i] = idx + NumElems;
3920 Mask[i] = idx - NumElems;
3924 /// isVEXTRACTIndex - Return true if the specified
3925 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3926 /// suitable for instruction that extract 128 or 256 bit vectors
3927 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
3928 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3929 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3932 // The index should be aligned on a vecWidth-bit boundary.
3934 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3936 MVT VT = N->getSimpleValueType(0);
3937 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3938 bool Result = (Index * ElSize) % vecWidth == 0;
3943 /// isVINSERTIndex - Return true if the specified INSERT_SUBVECTOR
3944 /// operand specifies a subvector insert that is suitable for input to
3945 /// insertion of 128 or 256-bit subvectors
3946 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
3947 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
3948 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3950 // The index should be aligned on a vecWidth-bit boundary.
3952 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3954 MVT VT = N->getSimpleValueType(0);
3955 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
3956 bool Result = (Index * ElSize) % vecWidth == 0;
3961 bool X86::isVINSERT128Index(SDNode *N) {
3962 return isVINSERTIndex(N, 128);
3965 bool X86::isVINSERT256Index(SDNode *N) {
3966 return isVINSERTIndex(N, 256);
3969 bool X86::isVEXTRACT128Index(SDNode *N) {
3970 return isVEXTRACTIndex(N, 128);
3973 bool X86::isVEXTRACT256Index(SDNode *N) {
3974 return isVEXTRACTIndex(N, 256);
3977 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
3978 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3979 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3980 llvm_unreachable("Illegal extract subvector for VEXTRACT");
3983 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3985 MVT VecVT = N->getOperand(0).getSimpleValueType();
3986 MVT ElVT = VecVT.getVectorElementType();
3988 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
3989 return Index / NumElemsPerChunk;
3992 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
3993 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
3994 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3995 llvm_unreachable("Illegal insert subvector for VINSERT");
3998 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4000 MVT VecVT = N->getSimpleValueType(0);
4001 MVT ElVT = VecVT.getVectorElementType();
4003 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4004 return Index / NumElemsPerChunk;
4007 /// getExtractVEXTRACT128Immediate - Return the appropriate immediate
4008 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4009 /// and VINSERTI128 instructions.
4010 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4011 return getExtractVEXTRACTImmediate(N, 128);
4014 /// getExtractVEXTRACT256Immediate - Return the appropriate immediate
4015 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF64x4
4016 /// and VINSERTI64x4 instructions.
4017 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4018 return getExtractVEXTRACTImmediate(N, 256);
4021 /// getInsertVINSERT128Immediate - Return the appropriate immediate
4022 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4023 /// and VINSERTI128 instructions.
4024 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4025 return getInsertVINSERTImmediate(N, 128);
4028 /// getInsertVINSERT256Immediate - Return the appropriate immediate
4029 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF46x4
4030 /// and VINSERTI64x4 instructions.
4031 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4032 return getInsertVINSERTImmediate(N, 256);
4035 /// isZero - Returns true if Elt is a constant integer zero
4036 static bool isZero(SDValue V) {
4037 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4038 return C && C->isNullValue();
4041 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4043 bool X86::isZeroNode(SDValue Elt) {
4046 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4047 return CFP->getValueAPF().isPosZero();
4051 /// getZeroVector - Returns a vector of specified type with all zero elements.
4053 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4054 SelectionDAG &DAG, SDLoc dl) {
4055 assert(VT.isVector() && "Expected a vector type");
4057 // Always build SSE zero vectors as <4 x i32> bitcasted
4058 // to their dest type. This ensures they get CSE'd.
4060 if (VT.is128BitVector()) { // SSE
4061 if (Subtarget->hasSSE2()) { // SSE2
4062 SDValue Cst = DAG.getConstant(0, MVT::i32);
4063 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4065 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4066 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4068 } else if (VT.is256BitVector()) { // AVX
4069 if (Subtarget->hasInt256()) { // AVX2
4070 SDValue Cst = DAG.getConstant(0, MVT::i32);
4071 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4072 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4074 // 256-bit logic and arithmetic instructions in AVX are all
4075 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4076 SDValue Cst = DAG.getConstantFP(+0.0, MVT::f32);
4077 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4078 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4080 } else if (VT.is512BitVector()) { // AVX-512
4081 SDValue Cst = DAG.getConstant(0, MVT::i32);
4082 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4083 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4084 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4085 } else if (VT.getScalarType() == MVT::i1) {
4086 assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
4087 SDValue Cst = DAG.getConstant(0, MVT::i1);
4088 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
4089 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4091 llvm_unreachable("Unexpected vector type");
4093 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4096 /// getOnesVector - Returns a vector of specified type with all bits set.
4097 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4098 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4099 /// Then bitcast to their original type, ensuring they get CSE'd.
4100 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4102 assert(VT.isVector() && "Expected a vector type");
4104 SDValue Cst = DAG.getConstant(~0U, MVT::i32);
4106 if (VT.is256BitVector()) {
4107 if (HasInt256) { // AVX2
4108 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4109 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4111 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4112 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4114 } else if (VT.is128BitVector()) {
4115 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4117 llvm_unreachable("Unexpected vector type");
4119 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4122 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4123 /// operation of specified width.
4124 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
4126 unsigned NumElems = VT.getVectorNumElements();
4127 SmallVector<int, 8> Mask;
4128 Mask.push_back(NumElems);
4129 for (unsigned i = 1; i != NumElems; ++i)
4131 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4134 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4135 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4137 unsigned NumElems = VT.getVectorNumElements();
4138 SmallVector<int, 8> Mask;
4139 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4141 Mask.push_back(i + NumElems);
4143 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4146 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4147 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4149 unsigned NumElems = VT.getVectorNumElements();
4150 SmallVector<int, 8> Mask;
4151 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4152 Mask.push_back(i + Half);
4153 Mask.push_back(i + NumElems + Half);
4155 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4158 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4159 /// vector of zero or undef vector. This produces a shuffle where the low
4160 /// element of V2 is swizzled into the zero/undef vector, landing at element
4161 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4162 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4164 const X86Subtarget *Subtarget,
4165 SelectionDAG &DAG) {
4166 MVT VT = V2.getSimpleValueType();
4168 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4169 unsigned NumElems = VT.getVectorNumElements();
4170 SmallVector<int, 16> MaskVec;
4171 for (unsigned i = 0; i != NumElems; ++i)
4172 // If this is the insertion idx, put the low elt of V2 here.
4173 MaskVec.push_back(i == Idx ? NumElems : i);
4174 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4177 /// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
4178 /// target specific opcode. Returns true if the Mask could be calculated. Sets
4179 /// IsUnary to true if only uses one source. Note that this will set IsUnary for
4180 /// shuffles which use a single input multiple times, and in those cases it will
4181 /// adjust the mask to only have indices within that single input.
4182 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4183 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4184 unsigned NumElems = VT.getVectorNumElements();
4188 bool IsFakeUnary = false;
4189 switch(N->getOpcode()) {
4190 case X86ISD::BLENDI:
4191 ImmN = N->getOperand(N->getNumOperands()-1);
4192 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4195 ImmN = N->getOperand(N->getNumOperands()-1);
4196 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4197 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4199 case X86ISD::UNPCKH:
4200 DecodeUNPCKHMask(VT, Mask);
4201 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4203 case X86ISD::UNPCKL:
4204 DecodeUNPCKLMask(VT, Mask);
4205 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4207 case X86ISD::MOVHLPS:
4208 DecodeMOVHLPSMask(NumElems, Mask);
4209 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4211 case X86ISD::MOVLHPS:
4212 DecodeMOVLHPSMask(NumElems, Mask);
4213 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4215 case X86ISD::PALIGNR:
4216 ImmN = N->getOperand(N->getNumOperands()-1);
4217 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4219 case X86ISD::PSHUFD:
4220 case X86ISD::VPERMILPI:
4221 ImmN = N->getOperand(N->getNumOperands()-1);
4222 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4225 case X86ISD::PSHUFHW:
4226 ImmN = N->getOperand(N->getNumOperands()-1);
4227 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4230 case X86ISD::PSHUFLW:
4231 ImmN = N->getOperand(N->getNumOperands()-1);
4232 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4235 case X86ISD::PSHUFB: {
4237 SDValue MaskNode = N->getOperand(1);
4238 while (MaskNode->getOpcode() == ISD::BITCAST)
4239 MaskNode = MaskNode->getOperand(0);
4241 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4242 // If we have a build-vector, then things are easy.
4243 EVT VT = MaskNode.getValueType();
4244 assert(VT.isVector() &&
4245 "Can't produce a non-vector with a build_vector!");
4246 if (!VT.isInteger())
4249 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4251 SmallVector<uint64_t, 32> RawMask;
4252 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4253 SDValue Op = MaskNode->getOperand(i);
4254 if (Op->getOpcode() == ISD::UNDEF) {
4255 RawMask.push_back((uint64_t)SM_SentinelUndef);
4258 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4261 APInt MaskElement = CN->getAPIntValue();
4263 // We now have to decode the element which could be any integer size and
4264 // extract each byte of it.
4265 for (int j = 0; j < NumBytesPerElement; ++j) {
4266 // Note that this is x86 and so always little endian: the low byte is
4267 // the first byte of the mask.
4268 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4269 MaskElement = MaskElement.lshr(8);
4272 DecodePSHUFBMask(RawMask, Mask);
4276 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4280 SDValue Ptr = MaskLoad->getBasePtr();
4281 if (Ptr->getOpcode() == X86ISD::Wrapper)
4282 Ptr = Ptr->getOperand(0);
4284 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4285 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4288 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4289 DecodePSHUFBMask(C, Mask);
4297 case X86ISD::VPERMI:
4298 ImmN = N->getOperand(N->getNumOperands()-1);
4299 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4304 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4306 case X86ISD::VPERM2X128:
4307 ImmN = N->getOperand(N->getNumOperands()-1);
4308 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4309 if (Mask.empty()) return false;
4311 case X86ISD::MOVSLDUP:
4312 DecodeMOVSLDUPMask(VT, Mask);
4315 case X86ISD::MOVSHDUP:
4316 DecodeMOVSHDUPMask(VT, Mask);
4319 case X86ISD::MOVDDUP:
4320 DecodeMOVDDUPMask(VT, Mask);
4323 case X86ISD::MOVLHPD:
4324 case X86ISD::MOVLPD:
4325 case X86ISD::MOVLPS:
4326 // Not yet implemented
4328 default: llvm_unreachable("unknown target shuffle node");
4331 // If we have a fake unary shuffle, the shuffle mask is spread across two
4332 // inputs that are actually the same node. Re-map the mask to always point
4333 // into the first input.
4336 if (M >= (int)Mask.size())
4342 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4343 /// element of the result of the vector shuffle.
4344 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4347 return SDValue(); // Limit search depth.
4349 SDValue V = SDValue(N, 0);
4350 EVT VT = V.getValueType();
4351 unsigned Opcode = V.getOpcode();
4353 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4354 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4355 int Elt = SV->getMaskElt(Index);
4358 return DAG.getUNDEF(VT.getVectorElementType());
4360 unsigned NumElems = VT.getVectorNumElements();
4361 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4362 : SV->getOperand(1);
4363 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4366 // Recurse into target specific vector shuffles to find scalars.
4367 if (isTargetShuffle(Opcode)) {
4368 MVT ShufVT = V.getSimpleValueType();
4369 unsigned NumElems = ShufVT.getVectorNumElements();
4370 SmallVector<int, 16> ShuffleMask;
4373 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4376 int Elt = ShuffleMask[Index];
4378 return DAG.getUNDEF(ShufVT.getVectorElementType());
4380 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4382 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4386 // Actual nodes that may contain scalar elements
4387 if (Opcode == ISD::BITCAST) {
4388 V = V.getOperand(0);
4389 EVT SrcVT = V.getValueType();
4390 unsigned NumElems = VT.getVectorNumElements();
4392 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4396 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4397 return (Index == 0) ? V.getOperand(0)
4398 : DAG.getUNDEF(VT.getVectorElementType());
4400 if (V.getOpcode() == ISD::BUILD_VECTOR)
4401 return V.getOperand(Index);
4406 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4408 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4409 unsigned NumNonZero, unsigned NumZero,
4411 const X86Subtarget* Subtarget,
4412 const TargetLowering &TLI) {
4419 for (unsigned i = 0; i < 16; ++i) {
4420 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4421 if (ThisIsNonZero && First) {
4423 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4425 V = DAG.getUNDEF(MVT::v8i16);
4430 SDValue ThisElt, LastElt;
4431 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4432 if (LastIsNonZero) {
4433 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4434 MVT::i16, Op.getOperand(i-1));
4436 if (ThisIsNonZero) {
4437 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4438 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4439 ThisElt, DAG.getConstant(8, MVT::i8));
4441 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4445 if (ThisElt.getNode())
4446 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4447 DAG.getIntPtrConstant(i/2));
4451 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4454 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4456 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4457 unsigned NumNonZero, unsigned NumZero,
4459 const X86Subtarget* Subtarget,
4460 const TargetLowering &TLI) {
4467 for (unsigned i = 0; i < 8; ++i) {
4468 bool isNonZero = (NonZeros & (1 << i)) != 0;
4472 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4474 V = DAG.getUNDEF(MVT::v8i16);
4477 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4478 MVT::v8i16, V, Op.getOperand(i),
4479 DAG.getIntPtrConstant(i));
4486 /// LowerBuildVectorv4x32 - Custom lower build_vector of v4i32 or v4f32.
4487 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4488 const X86Subtarget *Subtarget,
4489 const TargetLowering &TLI) {
4490 // Find all zeroable elements.
4491 std::bitset<4> Zeroable;
4492 for (int i=0; i < 4; ++i) {
4493 SDValue Elt = Op->getOperand(i);
4494 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4496 assert(Zeroable.size() - Zeroable.count() > 1 &&
4497 "We expect at least two non-zero elements!");
4499 // We only know how to deal with build_vector nodes where elements are either
4500 // zeroable or extract_vector_elt with constant index.
4501 SDValue FirstNonZero;
4502 unsigned FirstNonZeroIdx;
4503 for (unsigned i=0; i < 4; ++i) {
4506 SDValue Elt = Op->getOperand(i);
4507 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4508 !isa<ConstantSDNode>(Elt.getOperand(1)))
4510 // Make sure that this node is extracting from a 128-bit vector.
4511 MVT VT = Elt.getOperand(0).getSimpleValueType();
4512 if (!VT.is128BitVector())
4514 if (!FirstNonZero.getNode()) {
4516 FirstNonZeroIdx = i;
4520 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4521 SDValue V1 = FirstNonZero.getOperand(0);
4522 MVT VT = V1.getSimpleValueType();
4524 // See if this build_vector can be lowered as a blend with zero.
4526 unsigned EltMaskIdx, EltIdx;
4528 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4529 if (Zeroable[EltIdx]) {
4530 // The zero vector will be on the right hand side.
4531 Mask[EltIdx] = EltIdx+4;
4535 Elt = Op->getOperand(EltIdx);
4536 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4537 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4538 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4540 Mask[EltIdx] = EltIdx;
4544 // Let the shuffle legalizer deal with blend operations.
4545 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4546 if (V1.getSimpleValueType() != VT)
4547 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4548 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4551 // See if we can lower this build_vector to a INSERTPS.
4552 if (!Subtarget->hasSSE41())
4555 SDValue V2 = Elt.getOperand(0);
4556 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4559 bool CanFold = true;
4560 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4564 SDValue Current = Op->getOperand(i);
4565 SDValue SrcVector = Current->getOperand(0);
4568 CanFold = SrcVector == V1 &&
4569 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4575 assert(V1.getNode() && "Expected at least two non-zero elements!");
4576 if (V1.getSimpleValueType() != MVT::v4f32)
4577 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4578 if (V2.getSimpleValueType() != MVT::v4f32)
4579 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4581 // Ok, we can emit an INSERTPS instruction.
4582 unsigned ZMask = Zeroable.to_ulong();
4584 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4585 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4586 SDValue Result = DAG.getNode(X86ISD::INSERTPS, SDLoc(Op), MVT::v4f32, V1, V2,
4587 DAG.getIntPtrConstant(InsertPSMask));
4588 return DAG.getNode(ISD::BITCAST, SDLoc(Op), VT, Result);
4591 /// Return a vector logical shift node.
4592 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4593 unsigned NumBits, SelectionDAG &DAG,
4594 const TargetLowering &TLI, SDLoc dl) {
4595 assert(VT.is128BitVector() && "Unknown type for VShift");
4596 MVT ShVT = MVT::v2i64;
4597 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4598 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4599 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(SrcOp.getValueType());
4600 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4601 SDValue ShiftVal = DAG.getConstant(NumBits/8, ScalarShiftTy);
4602 return DAG.getNode(ISD::BITCAST, dl, VT,
4603 DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4607 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4609 // Check if the scalar load can be widened into a vector load. And if
4610 // the address is "base + cst" see if the cst can be "absorbed" into
4611 // the shuffle mask.
4612 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4613 SDValue Ptr = LD->getBasePtr();
4614 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4616 EVT PVT = LD->getValueType(0);
4617 if (PVT != MVT::i32 && PVT != MVT::f32)
4622 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4623 FI = FINode->getIndex();
4625 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4626 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4627 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4628 Offset = Ptr.getConstantOperandVal(1);
4629 Ptr = Ptr.getOperand(0);
4634 // FIXME: 256-bit vector instructions don't require a strict alignment,
4635 // improve this code to support it better.
4636 unsigned RequiredAlign = VT.getSizeInBits()/8;
4637 SDValue Chain = LD->getChain();
4638 // Make sure the stack object alignment is at least 16 or 32.
4639 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4640 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4641 if (MFI->isFixedObjectIndex(FI)) {
4642 // Can't change the alignment. FIXME: It's possible to compute
4643 // the exact stack offset and reference FI + adjust offset instead.
4644 // If someone *really* cares about this. That's the way to implement it.
4647 MFI->setObjectAlignment(FI, RequiredAlign);
4651 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4652 // Ptr + (Offset & ~15).
4655 if ((Offset % RequiredAlign) & 3)
4657 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4659 Ptr = DAG.getNode(ISD::ADD, SDLoc(Ptr), Ptr.getValueType(),
4660 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4662 int EltNo = (Offset - StartOffset) >> 2;
4663 unsigned NumElems = VT.getVectorNumElements();
4665 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4666 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4667 LD->getPointerInfo().getWithOffset(StartOffset),
4668 false, false, false, 0);
4670 SmallVector<int, 8> Mask(NumElems, EltNo);
4672 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4678 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4679 /// elements can be replaced by a single large load which has the same value as
4680 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4682 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4684 /// FIXME: we'd also like to handle the case where the last elements are zero
4685 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4686 /// There's even a handy isZeroNode for that purpose.
4687 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4688 SDLoc &DL, SelectionDAG &DAG,
4689 bool isAfterLegalize) {
4690 unsigned NumElems = Elts.size();
4692 LoadSDNode *LDBase = nullptr;
4693 unsigned LastLoadedElt = -1U;
4695 // For each element in the initializer, see if we've found a load or an undef.
4696 // If we don't find an initial load element, or later load elements are
4697 // non-consecutive, bail out.
4698 for (unsigned i = 0; i < NumElems; ++i) {
4699 SDValue Elt = Elts[i];
4700 // Look through a bitcast.
4701 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
4702 Elt = Elt.getOperand(0);
4703 if (!Elt.getNode() ||
4704 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4707 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4709 LDBase = cast<LoadSDNode>(Elt.getNode());
4713 if (Elt.getOpcode() == ISD::UNDEF)
4716 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4717 EVT LdVT = Elt.getValueType();
4718 // Each loaded element must be the correct fractional portion of the
4719 // requested vector load.
4720 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
4722 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
4727 // If we have found an entire vector of loads and undefs, then return a large
4728 // load of the entire vector width starting at the base pointer. If we found
4729 // consecutive loads for the low half, generate a vzext_load node.
4730 if (LastLoadedElt == NumElems - 1) {
4731 assert(LDBase && "Did not find base load for merging consecutive loads");
4732 EVT EltVT = LDBase->getValueType(0);
4733 // Ensure that the input vector size for the merged loads matches the
4734 // cumulative size of the input elements.
4735 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
4738 if (isAfterLegalize &&
4739 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
4742 SDValue NewLd = SDValue();
4744 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4745 LDBase->getPointerInfo(), LDBase->isVolatile(),
4746 LDBase->isNonTemporal(), LDBase->isInvariant(),
4747 LDBase->getAlignment());
4749 if (LDBase->hasAnyUseOfValue(1)) {
4750 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4752 SDValue(NewLd.getNode(), 1));
4753 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4754 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4755 SDValue(NewLd.getNode(), 1));
4761 //TODO: The code below fires only for for loading the low v2i32 / v2f32
4762 //of a v4i32 / v4f32. It's probably worth generalizing.
4763 EVT EltVT = VT.getVectorElementType();
4764 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
4765 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4766 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4767 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4769 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
4770 LDBase->getPointerInfo(),
4771 LDBase->getAlignment(),
4772 false/*isVolatile*/, true/*ReadMem*/,
4775 // Make sure the newly-created LOAD is in the same position as LDBase in
4776 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
4777 // update uses of LDBase's output chain to use the TokenFactor.
4778 if (LDBase->hasAnyUseOfValue(1)) {
4779 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
4780 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
4781 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
4782 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
4783 SDValue(ResNode.getNode(), 1));
4786 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4791 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
4792 /// to generate a splat value for the following cases:
4793 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
4794 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
4795 /// a scalar load, or a constant.
4796 /// The VBROADCAST node is returned when a pattern is found,
4797 /// or SDValue() otherwise.
4798 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
4799 SelectionDAG &DAG) {
4800 // VBROADCAST requires AVX.
4801 // TODO: Splats could be generated for non-AVX CPUs using SSE
4802 // instructions, but there's less potential gain for only 128-bit vectors.
4803 if (!Subtarget->hasAVX())
4806 MVT VT = Op.getSimpleValueType();
4809 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
4810 "Unsupported vector type for broadcast.");
4815 switch (Op.getOpcode()) {
4817 // Unknown pattern found.
4820 case ISD::BUILD_VECTOR: {
4821 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
4822 BitVector UndefElements;
4823 SDValue Splat = BVOp->getSplatValue(&UndefElements);
4825 // We need a splat of a single value to use broadcast, and it doesn't
4826 // make any sense if the value is only in one element of the vector.
4827 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
4831 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4832 Ld.getOpcode() == ISD::ConstantFP);
4834 // Make sure that all of the users of a non-constant load are from the
4835 // BUILD_VECTOR node.
4836 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
4841 case ISD::VECTOR_SHUFFLE: {
4842 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4844 // Shuffles must have a splat mask where the first element is
4846 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
4849 SDValue Sc = Op.getOperand(0);
4850 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
4851 Sc.getOpcode() != ISD::BUILD_VECTOR) {
4853 if (!Subtarget->hasInt256())
4856 // Use the register form of the broadcast instruction available on AVX2.
4857 if (VT.getSizeInBits() >= 256)
4858 Sc = Extract128BitVector(Sc, 0, DAG, dl);
4859 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
4862 Ld = Sc.getOperand(0);
4863 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
4864 Ld.getOpcode() == ISD::ConstantFP);
4866 // The scalar_to_vector node and the suspected
4867 // load node must have exactly one user.
4868 // Constants may have multiple users.
4870 // AVX-512 has register version of the broadcast
4871 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
4872 Ld.getValueType().getSizeInBits() >= 32;
4873 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
4880 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
4881 bool IsGE256 = (VT.getSizeInBits() >= 256);
4883 // When optimizing for size, generate up to 5 extra bytes for a broadcast
4884 // instruction to save 8 or more bytes of constant pool data.
4885 // TODO: If multiple splats are generated to load the same constant,
4886 // it may be detrimental to overall size. There needs to be a way to detect
4887 // that condition to know if this is truly a size win.
4888 const Function *F = DAG.getMachineFunction().getFunction();
4889 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
4891 // Handle broadcasting a single constant scalar from the constant pool
4893 // On Sandybridge (no AVX2), it is still better to load a constant vector
4894 // from the constant pool and not to broadcast it from a scalar.
4895 // But override that restriction when optimizing for size.
4896 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
4897 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
4898 EVT CVT = Ld.getValueType();
4899 assert(!CVT.isVector() && "Must not broadcast a vector type");
4901 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
4902 // For size optimization, also splat v2f64 and v2i64, and for size opt
4903 // with AVX2, also splat i8 and i16.
4904 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
4905 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4906 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
4907 const Constant *C = nullptr;
4908 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
4909 C = CI->getConstantIntValue();
4910 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
4911 C = CF->getConstantFPValue();
4913 assert(C && "Invalid constant type");
4915 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4916 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
4917 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
4918 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
4919 MachinePointerInfo::getConstantPool(),
4920 false, false, false, Alignment);
4922 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4926 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
4928 // Handle AVX2 in-register broadcasts.
4929 if (!IsLoad && Subtarget->hasInt256() &&
4930 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
4931 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4933 // The scalar source must be a normal load.
4937 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
4938 (Subtarget->hasVLX() && ScalarSize == 64))
4939 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4941 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
4942 // double since there is no vbroadcastsd xmm
4943 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
4944 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
4945 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
4948 // Unsupported broadcast.
4952 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
4953 /// underlying vector and index.
4955 /// Modifies \p ExtractedFromVec to the real vector and returns the real
4957 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
4959 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
4960 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
4963 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
4965 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
4967 // (extract_vector_elt (vector_shuffle<2,u,u,u>
4968 // (extract_subvector (v8f32 %vreg0), Constant<4>),
4971 // In this case the vector is the extract_subvector expression and the index
4972 // is 2, as specified by the shuffle.
4973 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
4974 SDValue ShuffleVec = SVOp->getOperand(0);
4975 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
4976 assert(ShuffleVecVT.getVectorElementType() ==
4977 ExtractedFromVec.getSimpleValueType().getVectorElementType());
4979 int ShuffleIdx = SVOp->getMaskElt(Idx);
4980 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
4981 ExtractedFromVec = ShuffleVec;
4987 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
4988 MVT VT = Op.getSimpleValueType();
4990 // Skip if insert_vec_elt is not supported.
4991 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4992 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
4996 unsigned NumElems = Op.getNumOperands();
5000 SmallVector<unsigned, 4> InsertIndices;
5001 SmallVector<int, 8> Mask(NumElems, -1);
5003 for (unsigned i = 0; i != NumElems; ++i) {
5004 unsigned Opc = Op.getOperand(i).getOpcode();
5006 if (Opc == ISD::UNDEF)
5009 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5010 // Quit if more than 1 elements need inserting.
5011 if (InsertIndices.size() > 1)
5014 InsertIndices.push_back(i);
5018 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5019 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5020 // Quit if non-constant index.
5021 if (!isa<ConstantSDNode>(ExtIdx))
5023 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5025 // Quit if extracted from vector of different type.
5026 if (ExtractedFromVec.getValueType() != VT)
5029 if (!VecIn1.getNode())
5030 VecIn1 = ExtractedFromVec;
5031 else if (VecIn1 != ExtractedFromVec) {
5032 if (!VecIn2.getNode())
5033 VecIn2 = ExtractedFromVec;
5034 else if (VecIn2 != ExtractedFromVec)
5035 // Quit if more than 2 vectors to shuffle
5039 if (ExtractedFromVec == VecIn1)
5041 else if (ExtractedFromVec == VecIn2)
5042 Mask[i] = Idx + NumElems;
5045 if (!VecIn1.getNode())
5048 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5049 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5050 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5051 unsigned Idx = InsertIndices[i];
5052 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5053 DAG.getIntPtrConstant(Idx));
5059 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5061 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5063 MVT VT = Op.getSimpleValueType();
5064 assert((VT.getVectorElementType() == MVT::i1) && (VT.getSizeInBits() <= 16) &&
5065 "Unexpected type in LowerBUILD_VECTORvXi1!");
5068 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5069 SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
5070 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5071 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5074 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5075 SDValue Cst = DAG.getTargetConstant(1, MVT::i1);
5076 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5077 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5080 bool AllContants = true;
5081 uint64_t Immediate = 0;
5082 int NonConstIdx = -1;
5083 bool IsSplat = true;
5084 unsigned NumNonConsts = 0;
5085 unsigned NumConsts = 0;
5086 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5087 SDValue In = Op.getOperand(idx);
5088 if (In.getOpcode() == ISD::UNDEF)
5090 if (!isa<ConstantSDNode>(In)) {
5091 AllContants = false;
5096 if (cast<ConstantSDNode>(In)->getZExtValue())
5097 Immediate |= (1ULL << idx);
5099 if (In != Op.getOperand(0))
5104 SDValue FullMask = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1,
5105 DAG.getConstant(Immediate, MVT::i16));
5106 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, FullMask,
5107 DAG.getIntPtrConstant(0));
5110 if (NumNonConsts == 1 && NonConstIdx != 0) {
5113 SDValue VecAsImm = DAG.getConstant(Immediate,
5114 MVT::getIntegerVT(VT.getSizeInBits()));
5115 DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
5118 DstVec = DAG.getUNDEF(VT);
5119 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5120 Op.getOperand(NonConstIdx),
5121 DAG.getIntPtrConstant(NonConstIdx));
5123 if (!IsSplat && (NonConstIdx != 0))
5124 llvm_unreachable("Unsupported BUILD_VECTOR operation");
5125 MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
5128 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5129 DAG.getConstant(-1, SelectVT),
5130 DAG.getConstant(0, SelectVT));
5132 Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
5133 DAG.getConstant((Immediate | 1), SelectVT),
5134 DAG.getConstant(Immediate, SelectVT));
5135 return DAG.getNode(ISD::BITCAST, dl, VT, Select);
5138 /// \brief Return true if \p N implements a horizontal binop and return the
5139 /// operands for the horizontal binop into V0 and V1.
5141 /// This is a helper function of PerformBUILD_VECTORCombine.
5142 /// This function checks that the build_vector \p N in input implements a
5143 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5144 /// operation to match.
5145 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5146 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5147 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5150 /// This function only analyzes elements of \p N whose indices are
5151 /// in range [BaseIdx, LastIdx).
5152 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5154 unsigned BaseIdx, unsigned LastIdx,
5155 SDValue &V0, SDValue &V1) {
5156 EVT VT = N->getValueType(0);
5158 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5159 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5160 "Invalid Vector in input!");
5162 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5163 bool CanFold = true;
5164 unsigned ExpectedVExtractIdx = BaseIdx;
5165 unsigned NumElts = LastIdx - BaseIdx;
5166 V0 = DAG.getUNDEF(VT);
5167 V1 = DAG.getUNDEF(VT);
5169 // Check if N implements a horizontal binop.
5170 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5171 SDValue Op = N->getOperand(i + BaseIdx);
5174 if (Op->getOpcode() == ISD::UNDEF) {
5175 // Update the expected vector extract index.
5176 if (i * 2 == NumElts)
5177 ExpectedVExtractIdx = BaseIdx;
5178 ExpectedVExtractIdx += 2;
5182 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5187 SDValue Op0 = Op.getOperand(0);
5188 SDValue Op1 = Op.getOperand(1);
5190 // Try to match the following pattern:
5191 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5192 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5193 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5194 Op0.getOperand(0) == Op1.getOperand(0) &&
5195 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5196 isa<ConstantSDNode>(Op1.getOperand(1)));
5200 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5201 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5203 if (i * 2 < NumElts) {
5204 if (V0.getOpcode() == ISD::UNDEF)
5205 V0 = Op0.getOperand(0);
5207 if (V1.getOpcode() == ISD::UNDEF)
5208 V1 = Op0.getOperand(0);
5209 if (i * 2 == NumElts)
5210 ExpectedVExtractIdx = BaseIdx;
5213 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5214 if (I0 == ExpectedVExtractIdx)
5215 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5216 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5217 // Try to match the following dag sequence:
5218 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5219 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5223 ExpectedVExtractIdx += 2;
5229 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5230 /// a concat_vector.
5232 /// This is a helper function of PerformBUILD_VECTORCombine.
5233 /// This function expects two 256-bit vectors called V0 and V1.
5234 /// At first, each vector is split into two separate 128-bit vectors.
5235 /// Then, the resulting 128-bit vectors are used to implement two
5236 /// horizontal binary operations.
5238 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5240 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5241 /// the two new horizontal binop.
5242 /// When Mode is set, the first horizontal binop dag node would take as input
5243 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5244 /// horizontal binop dag node would take as input the lower 128-bit of V1
5245 /// and the upper 128-bit of V1.
5247 /// HADD V0_LO, V0_HI
5248 /// HADD V1_LO, V1_HI
5250 /// Otherwise, the first horizontal binop dag node takes as input the lower
5251 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5252 /// dag node takes the the upper 128-bit of V0 and the upper 128-bit of V1.
5254 /// HADD V0_LO, V1_LO
5255 /// HADD V0_HI, V1_HI
5257 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5258 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5259 /// the upper 128-bits of the result.
5260 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5261 SDLoc DL, SelectionDAG &DAG,
5262 unsigned X86Opcode, bool Mode,
5263 bool isUndefLO, bool isUndefHI) {
5264 EVT VT = V0.getValueType();
5265 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5266 "Invalid nodes in input!");
5268 unsigned NumElts = VT.getVectorNumElements();
5269 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5270 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5271 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5272 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5273 EVT NewVT = V0_LO.getValueType();
5275 SDValue LO = DAG.getUNDEF(NewVT);
5276 SDValue HI = DAG.getUNDEF(NewVT);
5279 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5280 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5281 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5282 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5283 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5285 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5286 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5287 V1_LO->getOpcode() != ISD::UNDEF))
5288 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5290 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5291 V1_HI->getOpcode() != ISD::UNDEF))
5292 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5295 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5298 /// \brief Try to fold a build_vector that performs an 'addsub' into the
5299 /// sequence of 'vadd + vsub + blendi'.
5300 static SDValue matchAddSub(const BuildVectorSDNode *BV, SelectionDAG &DAG,
5301 const X86Subtarget *Subtarget) {
5303 EVT VT = BV->getValueType(0);
5304 unsigned NumElts = VT.getVectorNumElements();
5305 SDValue InVec0 = DAG.getUNDEF(VT);
5306 SDValue InVec1 = DAG.getUNDEF(VT);
5308 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5309 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5311 // Odd-numbered elements in the input build vector are obtained from
5312 // adding two integer/float elements.
5313 // Even-numbered elements in the input build vector are obtained from
5314 // subtracting two integer/float elements.
5315 unsigned ExpectedOpcode = ISD::FSUB;
5316 unsigned NextExpectedOpcode = ISD::FADD;
5317 bool AddFound = false;
5318 bool SubFound = false;
5320 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5321 SDValue Op = BV->getOperand(i);
5323 // Skip 'undef' values.
5324 unsigned Opcode = Op.getOpcode();
5325 if (Opcode == ISD::UNDEF) {
5326 std::swap(ExpectedOpcode, NextExpectedOpcode);
5330 // Early exit if we found an unexpected opcode.
5331 if (Opcode != ExpectedOpcode)
5334 SDValue Op0 = Op.getOperand(0);
5335 SDValue Op1 = Op.getOperand(1);
5337 // Try to match the following pattern:
5338 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5339 // Early exit if we cannot match that sequence.
5340 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5341 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5342 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5343 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5344 Op0.getOperand(1) != Op1.getOperand(1))
5347 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5351 // We found a valid add/sub node. Update the information accordingly.
5357 // Update InVec0 and InVec1.
5358 if (InVec0.getOpcode() == ISD::UNDEF)
5359 InVec0 = Op0.getOperand(0);
5360 if (InVec1.getOpcode() == ISD::UNDEF)
5361 InVec1 = Op1.getOperand(0);
5363 // Make sure that operands in input to each add/sub node always
5364 // come from a same pair of vectors.
5365 if (InVec0 != Op0.getOperand(0)) {
5366 if (ExpectedOpcode == ISD::FSUB)
5369 // FADD is commutable. Try to commute the operands
5370 // and then test again.
5371 std::swap(Op0, Op1);
5372 if (InVec0 != Op0.getOperand(0))
5376 if (InVec1 != Op1.getOperand(0))
5379 // Update the pair of expected opcodes.
5380 std::swap(ExpectedOpcode, NextExpectedOpcode);
5383 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5384 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5385 InVec1.getOpcode() != ISD::UNDEF)
5386 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5391 static SDValue PerformBUILD_VECTORCombine(SDNode *N, SelectionDAG &DAG,
5392 const X86Subtarget *Subtarget) {
5394 EVT VT = N->getValueType(0);
5395 unsigned NumElts = VT.getVectorNumElements();
5396 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(N);
5397 SDValue InVec0, InVec1;
5399 // Try to match an ADDSUB.
5400 if ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
5401 (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) {
5402 SDValue Value = matchAddSub(BV, DAG, Subtarget);
5403 if (Value.getNode())
5407 // Try to match horizontal ADD/SUB.
5408 unsigned NumUndefsLO = 0;
5409 unsigned NumUndefsHI = 0;
5410 unsigned Half = NumElts/2;
5412 // Count the number of UNDEF operands in the build_vector in input.
5413 for (unsigned i = 0, e = Half; i != e; ++i)
5414 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5417 for (unsigned i = Half, e = NumElts; i != e; ++i)
5418 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5421 // Early exit if this is either a build_vector of all UNDEFs or all the
5422 // operands but one are UNDEF.
5423 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5426 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5427 // Try to match an SSE3 float HADD/HSUB.
5428 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5429 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5431 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5432 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5433 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5434 // Try to match an SSSE3 integer HADD/HSUB.
5435 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5436 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5438 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5439 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5442 if (!Subtarget->hasAVX())
5445 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5446 // Try to match an AVX horizontal add/sub of packed single/double
5447 // precision floating point values from 256-bit vectors.
5448 SDValue InVec2, InVec3;
5449 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5450 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5451 ((InVec0.getOpcode() == ISD::UNDEF ||
5452 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5453 ((InVec1.getOpcode() == ISD::UNDEF ||
5454 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5455 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5457 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5458 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5459 ((InVec0.getOpcode() == ISD::UNDEF ||
5460 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5461 ((InVec1.getOpcode() == ISD::UNDEF ||
5462 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5463 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5464 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5465 // Try to match an AVX2 horizontal add/sub of signed integers.
5466 SDValue InVec2, InVec3;
5468 bool CanFold = true;
5470 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5471 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5472 ((InVec0.getOpcode() == ISD::UNDEF ||
5473 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5474 ((InVec1.getOpcode() == ISD::UNDEF ||
5475 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5476 X86Opcode = X86ISD::HADD;
5477 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5478 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5479 ((InVec0.getOpcode() == ISD::UNDEF ||
5480 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5481 ((InVec1.getOpcode() == ISD::UNDEF ||
5482 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5483 X86Opcode = X86ISD::HSUB;
5488 // Fold this build_vector into a single horizontal add/sub.
5489 // Do this only if the target has AVX2.
5490 if (Subtarget->hasAVX2())
5491 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5493 // Do not try to expand this build_vector into a pair of horizontal
5494 // add/sub if we can emit a pair of scalar add/sub.
5495 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5498 // Convert this build_vector into a pair of horizontal binop followed by
5500 bool isUndefLO = NumUndefsLO == Half;
5501 bool isUndefHI = NumUndefsHI == Half;
5502 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5503 isUndefLO, isUndefHI);
5507 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5508 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5510 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5511 X86Opcode = X86ISD::HADD;
5512 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5513 X86Opcode = X86ISD::HSUB;
5514 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5515 X86Opcode = X86ISD::FHADD;
5516 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5517 X86Opcode = X86ISD::FHSUB;
5521 // Don't try to expand this build_vector into a pair of horizontal add/sub
5522 // if we can simply emit a pair of scalar add/sub.
5523 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5526 // Convert this build_vector into two horizontal add/sub followed by
5528 bool isUndefLO = NumUndefsLO == Half;
5529 bool isUndefHI = NumUndefsHI == Half;
5530 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5531 isUndefLO, isUndefHI);
5538 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5541 MVT VT = Op.getSimpleValueType();
5542 MVT ExtVT = VT.getVectorElementType();
5543 unsigned NumElems = Op.getNumOperands();
5545 // Generate vectors for predicate vectors.
5546 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5547 return LowerBUILD_VECTORvXi1(Op, DAG);
5549 // Vectors containing all zeros can be matched by pxor and xorps later
5550 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5551 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5552 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5553 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5556 return getZeroVector(VT, Subtarget, DAG, dl);
5559 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5560 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5561 // vpcmpeqd on 256-bit vectors.
5562 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5563 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5566 if (!VT.is512BitVector())
5567 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5570 SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG);
5571 if (Broadcast.getNode())
5574 unsigned EVTBits = ExtVT.getSizeInBits();
5576 unsigned NumZero = 0;
5577 unsigned NumNonZero = 0;
5578 unsigned NonZeros = 0;
5579 bool IsAllConstants = true;
5580 SmallSet<SDValue, 8> Values;
5581 for (unsigned i = 0; i < NumElems; ++i) {
5582 SDValue Elt = Op.getOperand(i);
5583 if (Elt.getOpcode() == ISD::UNDEF)
5586 if (Elt.getOpcode() != ISD::Constant &&
5587 Elt.getOpcode() != ISD::ConstantFP)
5588 IsAllConstants = false;
5589 if (X86::isZeroNode(Elt))
5592 NonZeros |= (1 << i);
5597 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5598 if (NumNonZero == 0)
5599 return DAG.getUNDEF(VT);
5601 // Special case for single non-zero, non-undef, element.
5602 if (NumNonZero == 1) {
5603 unsigned Idx = countTrailingZeros(NonZeros);
5604 SDValue Item = Op.getOperand(Idx);
5606 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5607 // the value are obviously zero, truncate the value to i32 and do the
5608 // insertion that way. Only do this if the value is non-constant or if the
5609 // value is a constant being inserted into element 0. It is cheaper to do
5610 // a constant pool load than it is to do a movd + shuffle.
5611 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5612 (!IsAllConstants || Idx == 0)) {
5613 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5615 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5616 EVT VecVT = MVT::v4i32;
5618 // Truncate the value (which may itself be a constant) to i32, and
5619 // convert it to a vector with movd (S2V+shuffle to zero extend).
5620 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5621 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5623 ISD::BITCAST, dl, VT,
5624 getShuffleVectorZeroOrUndef(Item, Idx * 2, true, Subtarget, DAG));
5628 // If we have a constant or non-constant insertion into the low element of
5629 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5630 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5631 // depending on what the source datatype is.
5634 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5636 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5637 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5638 if (VT.is256BitVector() || VT.is512BitVector()) {
5639 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5640 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5641 Item, DAG.getIntPtrConstant(0));
5643 assert(VT.is128BitVector() && "Expected an SSE value type!");
5644 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5645 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5646 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5649 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5650 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5651 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
5652 if (VT.is256BitVector()) {
5653 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
5654 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
5656 assert(VT.is128BitVector() && "Expected an SSE value type!");
5657 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5659 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5663 // Is it a vector logical left shift?
5664 if (NumElems == 2 && Idx == 1 &&
5665 X86::isZeroNode(Op.getOperand(0)) &&
5666 !X86::isZeroNode(Op.getOperand(1))) {
5667 unsigned NumBits = VT.getSizeInBits();
5668 return getVShift(true, VT,
5669 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5670 VT, Op.getOperand(1)),
5671 NumBits/2, DAG, *this, dl);
5674 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5677 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5678 // is a non-constant being inserted into an element other than the low one,
5679 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5680 // movd/movss) to move this into the low element, then shuffle it into
5682 if (EVTBits == 32) {
5683 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5684 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
5688 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5689 if (Values.size() == 1) {
5690 if (EVTBits == 32) {
5691 // Instead of a shuffle like this:
5692 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5693 // Check if it's possible to issue this instead.
5694 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5695 unsigned Idx = countTrailingZeros(NonZeros);
5696 SDValue Item = Op.getOperand(Idx);
5697 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5698 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5703 // A vector full of immediates; various special cases are already
5704 // handled, so this is best done with a single constant-pool load.
5708 // For AVX-length vectors, see if we can use a vector load to get all of the
5709 // elements, otherwise build the individual 128-bit pieces and use
5710 // shuffles to put them in place.
5711 if (VT.is256BitVector() || VT.is512BitVector()) {
5712 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
5714 // Check for a build vector of consecutive loads.
5715 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
5718 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5720 // Build both the lower and upper subvector.
5721 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5722 makeArrayRef(&V[0], NumElems/2));
5723 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
5724 makeArrayRef(&V[NumElems / 2], NumElems/2));
5726 // Recreate the wider vector with the lower and upper part.
5727 if (VT.is256BitVector())
5728 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5729 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
5732 // Let legalizer expand 2-wide build_vectors.
5733 if (EVTBits == 64) {
5734 if (NumNonZero == 1) {
5735 // One half is zero or undef.
5736 unsigned Idx = countTrailingZeros(NonZeros);
5737 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5738 Op.getOperand(Idx));
5739 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
5744 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5745 if (EVTBits == 8 && NumElems == 16) {
5746 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5748 if (V.getNode()) return V;
5751 if (EVTBits == 16 && NumElems == 8) {
5752 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5754 if (V.getNode()) return V;
5757 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
5758 if (EVTBits == 32 && NumElems == 4) {
5759 SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this);
5764 // If element VT is == 32 bits, turn it into a number of shuffles.
5765 SmallVector<SDValue, 8> V(NumElems);
5766 if (NumElems == 4 && NumZero > 0) {
5767 for (unsigned i = 0; i < 4; ++i) {
5768 bool isZero = !(NonZeros & (1 << i));
5770 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
5772 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5775 for (unsigned i = 0; i < 2; ++i) {
5776 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5779 V[i] = V[i*2]; // Must be a zero vector.
5782 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5785 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5788 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5793 bool Reverse1 = (NonZeros & 0x3) == 2;
5794 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5798 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
5799 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
5801 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5804 if (Values.size() > 1 && VT.is128BitVector()) {
5805 // Check for a build vector of consecutive loads.
5806 for (unsigned i = 0; i < NumElems; ++i)
5807 V[i] = Op.getOperand(i);
5809 // Check for elements which are consecutive loads.
5810 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false);
5814 // Check for a build vector from mostly shuffle plus few inserting.
5815 SDValue Sh = buildFromShuffleMostly(Op, DAG);
5819 // For SSE 4.1, use insertps to put the high elements into the low element.
5820 if (Subtarget->hasSSE41()) {
5822 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5823 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5825 Result = DAG.getUNDEF(VT);
5827 for (unsigned i = 1; i < NumElems; ++i) {
5828 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5829 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5830 Op.getOperand(i), DAG.getIntPtrConstant(i));
5835 // Otherwise, expand into a number of unpckl*, start by extending each of
5836 // our (non-undef) elements to the full vector width with the element in the
5837 // bottom slot of the vector (which generates no code for SSE).
5838 for (unsigned i = 0; i < NumElems; ++i) {
5839 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5840 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5842 V[i] = DAG.getUNDEF(VT);
5845 // Next, we iteratively mix elements, e.g. for v4f32:
5846 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5847 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5848 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5849 unsigned EltStride = NumElems >> 1;
5850 while (EltStride != 0) {
5851 for (unsigned i = 0; i < EltStride; ++i) {
5852 // If V[i+EltStride] is undef and this is the first round of mixing,
5853 // then it is safe to just drop this shuffle: V[i] is already in the
5854 // right place, the one element (since it's the first round) being
5855 // inserted as undef can be dropped. This isn't safe for successive
5856 // rounds because they will permute elements within both vectors.
5857 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5858 EltStride == NumElems/2)
5861 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5870 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5871 // to create 256-bit vectors from two other 128-bit ones.
5872 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5874 MVT ResVT = Op.getSimpleValueType();
5876 assert((ResVT.is256BitVector() ||
5877 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
5879 SDValue V1 = Op.getOperand(0);
5880 SDValue V2 = Op.getOperand(1);
5881 unsigned NumElems = ResVT.getVectorNumElements();
5882 if(ResVT.is256BitVector())
5883 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5885 if (Op.getNumOperands() == 4) {
5886 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
5887 ResVT.getVectorNumElements()/2);
5888 SDValue V3 = Op.getOperand(2);
5889 SDValue V4 = Op.getOperand(3);
5890 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
5891 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
5893 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
5896 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5897 MVT LLVM_ATTRIBUTE_UNUSED VT = Op.getSimpleValueType();
5898 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
5899 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
5900 Op.getNumOperands() == 4)));
5902 // AVX can use the vinsertf128 instruction to create 256-bit vectors
5903 // from two other 128-bit ones.
5905 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
5906 return LowerAVXCONCAT_VECTORS(Op, DAG);
5910 //===----------------------------------------------------------------------===//
5911 // Vector shuffle lowering
5913 // This is an experimental code path for lowering vector shuffles on x86. It is
5914 // designed to handle arbitrary vector shuffles and blends, gracefully
5915 // degrading performance as necessary. It works hard to recognize idiomatic
5916 // shuffles and lower them to optimal instruction patterns without leaving
5917 // a framework that allows reasonably efficient handling of all vector shuffle
5919 //===----------------------------------------------------------------------===//
5921 /// \brief Tiny helper function to identify a no-op mask.
5923 /// This is a somewhat boring predicate function. It checks whether the mask
5924 /// array input, which is assumed to be a single-input shuffle mask of the kind
5925 /// used by the X86 shuffle instructions (not a fully general
5926 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
5927 /// in-place shuffle are 'no-op's.
5928 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
5929 for (int i = 0, Size = Mask.size(); i < Size; ++i)
5930 if (Mask[i] != -1 && Mask[i] != i)
5935 /// \brief Helper function to classify a mask as a single-input mask.
5937 /// This isn't a generic single-input test because in the vector shuffle
5938 /// lowering we canonicalize single inputs to be the first input operand. This
5939 /// means we can more quickly test for a single input by only checking whether
5940 /// an input from the second operand exists. We also assume that the size of
5941 /// mask corresponds to the size of the input vectors which isn't true in the
5942 /// fully general case.
5943 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
5945 if (M >= (int)Mask.size())
5950 /// \brief Test whether there are elements crossing 128-bit lanes in this
5953 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
5954 /// and we routinely test for these.
5955 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
5956 int LaneSize = 128 / VT.getScalarSizeInBits();
5957 int Size = Mask.size();
5958 for (int i = 0; i < Size; ++i)
5959 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
5964 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
5966 /// This checks a shuffle mask to see if it is performing the same
5967 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
5968 /// that it is also not lane-crossing. It may however involve a blend from the
5969 /// same lane of a second vector.
5971 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
5972 /// non-trivial to compute in the face of undef lanes. The representation is
5973 /// *not* suitable for use with existing 128-bit shuffles as it will contain
5974 /// entries from both V1 and V2 inputs to the wider mask.
5976 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
5977 SmallVectorImpl<int> &RepeatedMask) {
5978 int LaneSize = 128 / VT.getScalarSizeInBits();
5979 RepeatedMask.resize(LaneSize, -1);
5980 int Size = Mask.size();
5981 for (int i = 0; i < Size; ++i) {
5984 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
5985 // This entry crosses lanes, so there is no way to model this shuffle.
5988 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
5989 if (RepeatedMask[i % LaneSize] == -1)
5990 // This is the first non-undef entry in this slot of a 128-bit lane.
5991 RepeatedMask[i % LaneSize] =
5992 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
5993 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
5994 // Found a mismatch with the repeated mask.
6000 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6003 /// This is a fast way to test a shuffle mask against a fixed pattern:
6005 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6007 /// It returns true if the mask is exactly as wide as the argument list, and
6008 /// each element of the mask is either -1 (signifying undef) or the value given
6009 /// in the argument.
6010 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6011 ArrayRef<int> ExpectedMask) {
6012 if (Mask.size() != ExpectedMask.size())
6015 int Size = Mask.size();
6017 // If the values are build vectors, we can look through them to find
6018 // equivalent inputs that make the shuffles equivalent.
6019 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6020 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6022 for (int i = 0; i < Size; ++i)
6023 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6024 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6025 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6026 if (!MaskBV || !ExpectedBV ||
6027 MaskBV->getOperand(Mask[i] % Size) !=
6028 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6035 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6037 /// This helper function produces an 8-bit shuffle immediate corresponding to
6038 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6039 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6042 /// NB: We rely heavily on "undef" masks preserving the input lane.
6043 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask,
6044 SelectionDAG &DAG) {
6045 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6046 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6047 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6048 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6049 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6052 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6053 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6054 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6055 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6056 return DAG.getConstant(Imm, MVT::i8);
6059 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6061 /// This is used as a fallback approach when first class blend instructions are
6062 /// unavailable. Currently it is only suitable for integer vectors, but could
6063 /// be generalized for floating point vectors if desirable.
6064 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6065 SDValue V2, ArrayRef<int> Mask,
6066 SelectionDAG &DAG) {
6067 assert(VT.isInteger() && "Only supports integer vector types!");
6068 MVT EltVT = VT.getScalarType();
6069 int NumEltBits = EltVT.getSizeInBits();
6070 SDValue Zero = DAG.getConstant(0, EltVT);
6071 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), EltVT);
6072 SmallVector<SDValue, 16> MaskOps;
6073 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6074 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6075 return SDValue(); // Shuffled input!
6076 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6079 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6080 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6081 // We have to cast V2 around.
6082 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6083 V2 = DAG.getNode(ISD::BITCAST, DL, VT,
6084 DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6085 DAG.getNode(ISD::BITCAST, DL, MaskVT, V1Mask),
6086 DAG.getNode(ISD::BITCAST, DL, MaskVT, V2)));
6087 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6090 /// \brief Try to emit a blend instruction for a shuffle.
6092 /// This doesn't do any checks for the availability of instructions for blending
6093 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6094 /// be matched in the backend with the type given. What it does check for is
6095 /// that the shuffle mask is in fact a blend.
6096 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6097 SDValue V2, ArrayRef<int> Mask,
6098 const X86Subtarget *Subtarget,
6099 SelectionDAG &DAG) {
6100 unsigned BlendMask = 0;
6101 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6102 if (Mask[i] >= Size) {
6103 if (Mask[i] != i + Size)
6104 return SDValue(); // Shuffled V2 input!
6105 BlendMask |= 1u << i;
6108 if (Mask[i] >= 0 && Mask[i] != i)
6109 return SDValue(); // Shuffled V1 input!
6111 switch (VT.SimpleTy) {
6116 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6117 DAG.getConstant(BlendMask, MVT::i8));
6121 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6125 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6126 // that instruction.
6127 if (Subtarget->hasAVX2()) {
6128 // Scale the blend by the number of 32-bit dwords per element.
6129 int Scale = VT.getScalarSizeInBits() / 32;
6131 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6132 if (Mask[i] >= Size)
6133 for (int j = 0; j < Scale; ++j)
6134 BlendMask |= 1u << (i * Scale + j);
6136 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6137 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6138 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6139 return DAG.getNode(ISD::BITCAST, DL, VT,
6140 DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6141 DAG.getConstant(BlendMask, MVT::i8)));
6145 // For integer shuffles we need to expand the mask and cast the inputs to
6146 // v8i16s prior to blending.
6147 int Scale = 8 / VT.getVectorNumElements();
6149 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6150 if (Mask[i] >= Size)
6151 for (int j = 0; j < Scale; ++j)
6152 BlendMask |= 1u << (i * Scale + j);
6154 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
6155 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
6156 return DAG.getNode(ISD::BITCAST, DL, VT,
6157 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6158 DAG.getConstant(BlendMask, MVT::i8)));
6162 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6163 SmallVector<int, 8> RepeatedMask;
6164 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6165 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6166 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6168 for (int i = 0; i < 8; ++i)
6169 if (RepeatedMask[i] >= 16)
6170 BlendMask |= 1u << i;
6171 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6172 DAG.getConstant(BlendMask, MVT::i8));
6178 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6179 "256-bit byte-blends require AVX2 support!");
6181 // Scale the blend by the number of bytes per element.
6182 int Scale = VT.getScalarSizeInBits() / 8;
6184 // This form of blend is always done on bytes. Compute the byte vector
6186 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6188 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6189 // mix of LLVM's code generator and the x86 backend. We tell the code
6190 // generator that boolean values in the elements of an x86 vector register
6191 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6192 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6193 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6194 // of the element (the remaining are ignored) and 0 in that high bit would
6195 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6196 // the LLVM model for boolean values in vector elements gets the relevant
6197 // bit set, it is set backwards and over constrained relative to x86's
6199 SmallVector<SDValue, 32> VSELECTMask;
6200 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6201 for (int j = 0; j < Scale; ++j)
6202 VSELECTMask.push_back(
6203 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6204 : DAG.getConstant(Mask[i] < Size ? -1 : 0, MVT::i8));
6206 V1 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V1);
6207 V2 = DAG.getNode(ISD::BITCAST, DL, BlendVT, V2);
6209 ISD::BITCAST, DL, VT,
6210 DAG.getNode(ISD::VSELECT, DL, BlendVT,
6211 DAG.getNode(ISD::BUILD_VECTOR, DL, BlendVT, VSELECTMask),
6216 llvm_unreachable("Not a supported integer vector type!");
6220 /// \brief Try to lower as a blend of elements from two inputs followed by
6221 /// a single-input permutation.
6223 /// This matches the pattern where we can blend elements from two inputs and
6224 /// then reduce the shuffle to a single-input permutation.
6225 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6228 SelectionDAG &DAG) {
6229 // We build up the blend mask while checking whether a blend is a viable way
6230 // to reduce the shuffle.
6231 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6232 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6234 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6238 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6240 if (BlendMask[Mask[i] % Size] == -1)
6241 BlendMask[Mask[i] % Size] = Mask[i];
6242 else if (BlendMask[Mask[i] % Size] != Mask[i])
6243 return SDValue(); // Can't blend in the needed input!
6245 PermuteMask[i] = Mask[i] % Size;
6248 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6249 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6252 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6253 /// blends and permutes.
6255 /// This matches the extremely common pattern for handling combined
6256 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6257 /// operations. It will try to pick the best arrangement of shuffles and
6259 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6263 SelectionDAG &DAG) {
6264 // Shuffle the input elements into the desired positions in V1 and V2 and
6265 // blend them together.
6266 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6267 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6268 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6269 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6270 if (Mask[i] >= 0 && Mask[i] < Size) {
6271 V1Mask[i] = Mask[i];
6273 } else if (Mask[i] >= Size) {
6274 V2Mask[i] = Mask[i] - Size;
6275 BlendMask[i] = i + Size;
6278 // Try to lower with the simpler initial blend strategy unless one of the
6279 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6280 // shuffle may be able to fold with a load or other benefit. However, when
6281 // we'll have to do 2x as many shuffles in order to achieve this, blending
6282 // first is a better strategy.
6283 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6284 if (SDValue BlendPerm =
6285 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6288 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6289 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6290 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6293 /// \brief Try to lower a vector shuffle as a byte rotation.
6295 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6296 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6297 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6298 /// try to generically lower a vector shuffle through such an pattern. It
6299 /// does not check for the profitability of lowering either as PALIGNR or
6300 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6301 /// This matches shuffle vectors that look like:
6303 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6305 /// Essentially it concatenates V1 and V2, shifts right by some number of
6306 /// elements, and takes the low elements as the result. Note that while this is
6307 /// specified as a *right shift* because x86 is little-endian, it is a *left
6308 /// rotate* of the vector lanes.
6309 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6312 const X86Subtarget *Subtarget,
6313 SelectionDAG &DAG) {
6314 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6316 int NumElts = Mask.size();
6317 int NumLanes = VT.getSizeInBits() / 128;
6318 int NumLaneElts = NumElts / NumLanes;
6320 // We need to detect various ways of spelling a rotation:
6321 // [11, 12, 13, 14, 15, 0, 1, 2]
6322 // [-1, 12, 13, 14, -1, -1, 1, -1]
6323 // [-1, -1, -1, -1, -1, -1, 1, 2]
6324 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6325 // [-1, 4, 5, 6, -1, -1, 9, -1]
6326 // [-1, 4, 5, 6, -1, -1, -1, -1]
6329 for (int l = 0; l < NumElts; l += NumLaneElts) {
6330 for (int i = 0; i < NumLaneElts; ++i) {
6331 if (Mask[l + i] == -1)
6333 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6335 // Get the mod-Size index and lane correct it.
6336 int LaneIdx = (Mask[l + i] % NumElts) - l;
6337 // Make sure it was in this lane.
6338 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6341 // Determine where a rotated vector would have started.
6342 int StartIdx = i - LaneIdx;
6344 // The identity rotation isn't interesting, stop.
6347 // If we found the tail of a vector the rotation must be the missing
6348 // front. If we found the head of a vector, it must be how much of the
6350 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6353 Rotation = CandidateRotation;
6354 else if (Rotation != CandidateRotation)
6355 // The rotations don't match, so we can't match this mask.
6358 // Compute which value this mask is pointing at.
6359 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6361 // Compute which of the two target values this index should be assigned
6362 // to. This reflects whether the high elements are remaining or the low
6363 // elements are remaining.
6364 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6366 // Either set up this value if we've not encountered it before, or check
6367 // that it remains consistent.
6370 else if (TargetV != MaskV)
6371 // This may be a rotation, but it pulls from the inputs in some
6372 // unsupported interleaving.
6377 // Check that we successfully analyzed the mask, and normalize the results.
6378 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6379 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6385 // The actual rotate instruction rotates bytes, so we need to scale the
6386 // rotation based on how many bytes are in the vector lane.
6387 int Scale = 16 / NumLaneElts;
6389 // SSSE3 targets can use the palignr instruction.
6390 if (Subtarget->hasSSSE3()) {
6391 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6392 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6393 Lo = DAG.getNode(ISD::BITCAST, DL, AlignVT, Lo);
6394 Hi = DAG.getNode(ISD::BITCAST, DL, AlignVT, Hi);
6396 return DAG.getNode(ISD::BITCAST, DL, VT,
6397 DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6398 DAG.getConstant(Rotation * Scale, MVT::i8)));
6401 assert(VT.getSizeInBits() == 128 &&
6402 "Rotate-based lowering only supports 128-bit lowering!");
6403 assert(Mask.size() <= 16 &&
6404 "Can shuffle at most 16 bytes in a 128-bit vector!");
6406 // Default SSE2 implementation
6407 int LoByteShift = 16 - Rotation * Scale;
6408 int HiByteShift = Rotation * Scale;
6410 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6411 Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Lo);
6412 Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Hi);
6414 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6415 DAG.getConstant(LoByteShift, MVT::i8));
6416 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6417 DAG.getConstant(HiByteShift, MVT::i8));
6418 return DAG.getNode(ISD::BITCAST, DL, VT,
6419 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6422 /// \brief Compute whether each element of a shuffle is zeroable.
6424 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6425 /// Either it is an undef element in the shuffle mask, the element of the input
6426 /// referenced is undef, or the element of the input referenced is known to be
6427 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6428 /// as many lanes with this technique as possible to simplify the remaining
6430 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6431 SDValue V1, SDValue V2) {
6432 SmallBitVector Zeroable(Mask.size(), false);
6434 while (V1.getOpcode() == ISD::BITCAST)
6435 V1 = V1->getOperand(0);
6436 while (V2.getOpcode() == ISD::BITCAST)
6437 V2 = V2->getOperand(0);
6439 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6440 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6442 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6444 // Handle the easy cases.
6445 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6450 // If this is an index into a build_vector node (which has the same number
6451 // of elements), dig out the input value and use it.
6452 SDValue V = M < Size ? V1 : V2;
6453 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6456 SDValue Input = V.getOperand(M % Size);
6457 // The UNDEF opcode check really should be dead code here, but not quite
6458 // worth asserting on (it isn't invalid, just unexpected).
6459 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6466 /// \brief Try to emit a bitmask instruction for a shuffle.
6468 /// This handles cases where we can model a blend exactly as a bitmask due to
6469 /// one of the inputs being zeroable.
6470 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6471 SDValue V2, ArrayRef<int> Mask,
6472 SelectionDAG &DAG) {
6473 MVT EltVT = VT.getScalarType();
6474 int NumEltBits = EltVT.getSizeInBits();
6475 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6476 SDValue Zero = DAG.getConstant(0, IntEltVT);
6477 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), IntEltVT);
6478 if (EltVT.isFloatingPoint()) {
6479 Zero = DAG.getNode(ISD::BITCAST, DL, EltVT, Zero);
6480 AllOnes = DAG.getNode(ISD::BITCAST, DL, EltVT, AllOnes);
6482 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6483 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6485 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6488 if (Mask[i] % Size != i)
6489 return SDValue(); // Not a blend.
6491 V = Mask[i] < Size ? V1 : V2;
6492 else if (V != (Mask[i] < Size ? V1 : V2))
6493 return SDValue(); // Can only let one input through the mask.
6495 VMaskOps[i] = AllOnes;
6498 return SDValue(); // No non-zeroable elements!
6500 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6501 V = DAG.getNode(VT.isFloatingPoint()
6502 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6507 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6509 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6510 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6511 /// matches elements from one of the input vectors shuffled to the left or
6512 /// right with zeroable elements 'shifted in'. It handles both the strictly
6513 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6516 /// PSHL : (little-endian) left bit shift.
6517 /// [ zz, 0, zz, 2 ]
6518 /// [ -1, 4, zz, -1 ]
6519 /// PSRL : (little-endian) right bit shift.
6521 /// [ -1, -1, 7, zz]
6522 /// PSLLDQ : (little-endian) left byte shift
6523 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6524 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6525 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6526 /// PSRLDQ : (little-endian) right byte shift
6527 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6528 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6529 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6530 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6531 SDValue V2, ArrayRef<int> Mask,
6532 SelectionDAG &DAG) {
6533 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6535 int Size = Mask.size();
6536 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6538 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6539 for (int i = 0; i < Size; i += Scale)
6540 for (int j = 0; j < Shift; ++j)
6541 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6547 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6548 for (int i = 0; i != Size; i += Scale) {
6549 unsigned Pos = Left ? i + Shift : i;
6550 unsigned Low = Left ? i : i + Shift;
6551 unsigned Len = Scale - Shift;
6552 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6553 Low + (V == V1 ? 0 : Size)))
6557 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6558 bool ByteShift = ShiftEltBits > 64;
6559 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6560 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6561 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6563 // Normalize the scale for byte shifts to still produce an i64 element
6565 Scale = ByteShift ? Scale / 2 : Scale;
6567 // We need to round trip through the appropriate type for the shift.
6568 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6569 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6570 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6571 "Illegal integer vector type");
6572 V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
6574 V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
6575 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6578 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6579 // keep doubling the size of the integer elements up to that. We can
6580 // then shift the elements of the integer vector by whole multiples of
6581 // their width within the elements of the larger integer vector. Test each
6582 // multiple to see if we can find a match with the moved element indices
6583 // and that the shifted in elements are all zeroable.
6584 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6585 for (int Shift = 1; Shift != Scale; ++Shift)
6586 for (bool Left : {true, false})
6587 if (CheckZeros(Shift, Scale, Left))
6588 for (SDValue V : {V1, V2})
6589 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6596 /// \brief Lower a vector shuffle as a zero or any extension.
6598 /// Given a specific number of elements, element bit width, and extension
6599 /// stride, produce either a zero or any extension based on the available
6600 /// features of the subtarget.
6601 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6602 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
6603 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6604 assert(Scale > 1 && "Need a scale to extend.");
6605 int NumElements = VT.getVectorNumElements();
6606 int EltBits = VT.getScalarSizeInBits();
6607 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
6608 "Only 8, 16, and 32 bit elements can be extended.");
6609 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
6611 // Found a valid zext mask! Try various lowering strategies based on the
6612 // input type and available ISA extensions.
6613 if (Subtarget->hasSSE41()) {
6614 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
6615 NumElements / Scale);
6616 return DAG.getNode(ISD::BITCAST, DL, VT,
6617 DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
6620 // For any extends we can cheat for larger element sizes and use shuffle
6621 // instructions that can fold with a load and/or copy.
6622 if (AnyExt && EltBits == 32) {
6623 int PSHUFDMask[4] = {0, -1, 1, -1};
6625 ISD::BITCAST, DL, VT,
6626 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6627 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6628 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
6630 if (AnyExt && EltBits == 16 && Scale > 2) {
6631 int PSHUFDMask[4] = {0, -1, 0, -1};
6632 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
6633 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, InputV),
6634 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG));
6635 int PSHUFHWMask[4] = {1, -1, -1, -1};
6637 ISD::BITCAST, DL, VT,
6638 DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
6639 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, InputV),
6640 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DAG)));
6643 // If this would require more than 2 unpack instructions to expand, use
6644 // pshufb when available. We can only use more than 2 unpack instructions
6645 // when zero extending i8 elements which also makes it easier to use pshufb.
6646 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
6647 assert(NumElements == 16 && "Unexpected byte vector width!");
6648 SDValue PSHUFBMask[16];
6649 for (int i = 0; i < 16; ++i)
6651 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, MVT::i8);
6652 InputV = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, InputV);
6653 return DAG.getNode(ISD::BITCAST, DL, VT,
6654 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
6655 DAG.getNode(ISD::BUILD_VECTOR, DL,
6656 MVT::v16i8, PSHUFBMask)));
6659 // Otherwise emit a sequence of unpacks.
6661 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
6662 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
6663 : getZeroVector(InputVT, Subtarget, DAG, DL);
6664 InputV = DAG.getNode(ISD::BITCAST, DL, InputVT, InputV);
6665 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
6669 } while (Scale > 1);
6670 return DAG.getNode(ISD::BITCAST, DL, VT, InputV);
6673 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
6675 /// This routine will try to do everything in its power to cleverly lower
6676 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
6677 /// check for the profitability of this lowering, it tries to aggressively
6678 /// match this pattern. It will use all of the micro-architectural details it
6679 /// can to emit an efficient lowering. It handles both blends with all-zero
6680 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
6681 /// masking out later).
6683 /// The reason we have dedicated lowering for zext-style shuffles is that they
6684 /// are both incredibly common and often quite performance sensitive.
6685 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
6686 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6687 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6688 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6690 int Bits = VT.getSizeInBits();
6691 int NumElements = VT.getVectorNumElements();
6692 assert(VT.getScalarSizeInBits() <= 32 &&
6693 "Exceeds 32-bit integer zero extension limit");
6694 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
6696 // Define a helper function to check a particular ext-scale and lower to it if
6698 auto Lower = [&](int Scale) -> SDValue {
6701 for (int i = 0; i < NumElements; ++i) {
6703 continue; // Valid anywhere but doesn't tell us anything.
6704 if (i % Scale != 0) {
6705 // Each of the extended elements need to be zeroable.
6709 // We no longer are in the anyext case.
6714 // Each of the base elements needs to be consecutive indices into the
6715 // same input vector.
6716 SDValue V = Mask[i] < NumElements ? V1 : V2;
6719 else if (InputV != V)
6720 return SDValue(); // Flip-flopping inputs.
6722 if (Mask[i] % NumElements != i / Scale)
6723 return SDValue(); // Non-consecutive strided elements.
6726 // If we fail to find an input, we have a zero-shuffle which should always
6727 // have already been handled.
6728 // FIXME: Maybe handle this here in case during blending we end up with one?
6732 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
6733 DL, VT, Scale, AnyExt, InputV, Subtarget, DAG);
6736 // The widest scale possible for extending is to a 64-bit integer.
6737 assert(Bits % 64 == 0 &&
6738 "The number of bits in a vector must be divisible by 64 on x86!");
6739 int NumExtElements = Bits / 64;
6741 // Each iteration, try extending the elements half as much, but into twice as
6743 for (; NumExtElements < NumElements; NumExtElements *= 2) {
6744 assert(NumElements % NumExtElements == 0 &&
6745 "The input vector size must be divisible by the extended size.");
6746 if (SDValue V = Lower(NumElements / NumExtElements))
6750 // General extends failed, but 128-bit vectors may be able to use MOVQ.
6754 // Returns one of the source operands if the shuffle can be reduced to a
6755 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
6756 auto CanZExtLowHalf = [&]() {
6757 for (int i = NumElements / 2; i != NumElements; ++i)
6760 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
6762 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
6767 if (SDValue V = CanZExtLowHalf()) {
6768 V = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V);
6769 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
6770 return DAG.getNode(ISD::BITCAST, DL, VT, V);
6773 // No viable ext lowering found.
6777 /// \brief Try to get a scalar value for a specific element of a vector.
6779 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
6780 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
6781 SelectionDAG &DAG) {
6782 MVT VT = V.getSimpleValueType();
6783 MVT EltVT = VT.getVectorElementType();
6784 while (V.getOpcode() == ISD::BITCAST)
6785 V = V.getOperand(0);
6786 // If the bitcasts shift the element size, we can't extract an equivalent
6788 MVT NewVT = V.getSimpleValueType();
6789 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
6792 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6793 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR))
6794 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, V.getOperand(Idx));
6799 /// \brief Helper to test for a load that can be folded with x86 shuffles.
6801 /// This is particularly important because the set of instructions varies
6802 /// significantly based on whether the operand is a load or not.
6803 static bool isShuffleFoldableLoad(SDValue V) {
6804 while (V.getOpcode() == ISD::BITCAST)
6805 V = V.getOperand(0);
6807 return ISD::isNON_EXTLoad(V.getNode());
6810 /// \brief Try to lower insertion of a single element into a zero vector.
6812 /// This is a common pattern that we have especially efficient patterns to lower
6813 /// across all subtarget feature sets.
6814 static SDValue lowerVectorShuffleAsElementInsertion(
6815 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
6816 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
6817 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6819 MVT EltVT = VT.getVectorElementType();
6821 int V2Index = std::find_if(Mask.begin(), Mask.end(),
6822 [&Mask](int M) { return M >= (int)Mask.size(); }) -
6824 bool IsV1Zeroable = true;
6825 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6826 if (i != V2Index && !Zeroable[i]) {
6827 IsV1Zeroable = false;
6831 // Check for a single input from a SCALAR_TO_VECTOR node.
6832 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
6833 // all the smarts here sunk into that routine. However, the current
6834 // lowering of BUILD_VECTOR makes that nearly impossible until the old
6835 // vector shuffle lowering is dead.
6836 if (SDValue V2S = getScalarValueForVectorElement(
6837 V2, Mask[V2Index] - Mask.size(), DAG)) {
6838 // We need to zext the scalar if it is smaller than an i32.
6839 V2S = DAG.getNode(ISD::BITCAST, DL, EltVT, V2S);
6840 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
6841 // Using zext to expand a narrow element won't work for non-zero
6846 // Zero-extend directly to i32.
6848 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
6850 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
6851 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
6852 EltVT == MVT::i16) {
6853 // Either not inserting from the low element of the input or the input
6854 // element size is too small to use VZEXT_MOVL to clear the high bits.
6858 if (!IsV1Zeroable) {
6859 // If V1 can't be treated as a zero vector we have fewer options to lower
6860 // this. We can't support integer vectors or non-zero targets cheaply, and
6861 // the V1 elements can't be permuted in any way.
6862 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
6863 if (!VT.isFloatingPoint() || V2Index != 0)
6865 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
6866 V1Mask[V2Index] = -1;
6867 if (!isNoopShuffleMask(V1Mask))
6869 // This is essentially a special case blend operation, but if we have
6870 // general purpose blend operations, they are always faster. Bail and let
6871 // the rest of the lowering handle these as blends.
6872 if (Subtarget->hasSSE41())
6875 // Otherwise, use MOVSD or MOVSS.
6876 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
6877 "Only two types of floating point element types to handle!");
6878 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
6882 // This lowering only works for the low element with floating point vectors.
6883 if (VT.isFloatingPoint() && V2Index != 0)
6886 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
6888 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6891 // If we have 4 or fewer lanes we can cheaply shuffle the element into
6892 // the desired position. Otherwise it is more efficient to do a vector
6893 // shift left. We know that we can do a vector shift left because all
6894 // the inputs are zero.
6895 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
6896 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
6897 V2Shuffle[V2Index] = 0;
6898 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
6900 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, V2);
6902 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
6904 V2Index * EltVT.getSizeInBits()/8,
6905 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(MVT::v2i64)));
6906 V2 = DAG.getNode(ISD::BITCAST, DL, VT, V2);
6912 /// \brief Try to lower broadcast of a single element.
6914 /// For convenience, this code also bundles all of the subtarget feature set
6915 /// filtering. While a little annoying to re-dispatch on type here, there isn't
6916 /// a convenient way to factor it out.
6917 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
6919 const X86Subtarget *Subtarget,
6920 SelectionDAG &DAG) {
6921 if (!Subtarget->hasAVX())
6923 if (VT.isInteger() && !Subtarget->hasAVX2())
6926 // Check that the mask is a broadcast.
6927 int BroadcastIdx = -1;
6929 if (M >= 0 && BroadcastIdx == -1)
6931 else if (M >= 0 && M != BroadcastIdx)
6934 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
6935 "a sorted mask where the broadcast "
6938 // Go up the chain of (vector) values to try and find a scalar load that
6939 // we can combine with the broadcast.
6941 switch (V.getOpcode()) {
6942 case ISD::CONCAT_VECTORS: {
6943 int OperandSize = Mask.size() / V.getNumOperands();
6944 V = V.getOperand(BroadcastIdx / OperandSize);
6945 BroadcastIdx %= OperandSize;
6949 case ISD::INSERT_SUBVECTOR: {
6950 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
6951 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
6955 int BeginIdx = (int)ConstantIdx->getZExtValue();
6957 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
6958 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
6959 BroadcastIdx -= BeginIdx;
6970 // Check if this is a broadcast of a scalar. We special case lowering
6971 // for scalars so that we can more effectively fold with loads.
6972 if (V.getOpcode() == ISD::BUILD_VECTOR ||
6973 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
6974 V = V.getOperand(BroadcastIdx);
6976 // If the scalar isn't a load we can't broadcast from it in AVX1, only with
6978 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
6980 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
6981 // We can't broadcast from a vector register w/o AVX2, and we can only
6982 // broadcast from the zero-element of a vector register.
6986 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
6989 // Check for whether we can use INSERTPS to perform the shuffle. We only use
6990 // INSERTPS when the V1 elements are already in the correct locations
6991 // because otherwise we can just always use two SHUFPS instructions which
6992 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
6993 // perform INSERTPS if a single V1 element is out of place and all V2
6994 // elements are zeroable.
6995 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
6997 SelectionDAG &DAG) {
6998 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
6999 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7000 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7001 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7003 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7006 int V1DstIndex = -1;
7007 int V2DstIndex = -1;
7008 bool V1UsedInPlace = false;
7010 for (int i = 0; i < 4; ++i) {
7011 // Synthesize a zero mask from the zeroable elements (includes undefs).
7017 // Flag if we use any V1 inputs in place.
7019 V1UsedInPlace = true;
7023 // We can only insert a single non-zeroable element.
7024 if (V1DstIndex != -1 || V2DstIndex != -1)
7028 // V1 input out of place for insertion.
7031 // V2 input for insertion.
7036 // Don't bother if we have no (non-zeroable) element for insertion.
7037 if (V1DstIndex == -1 && V2DstIndex == -1)
7040 // Determine element insertion src/dst indices. The src index is from the
7041 // start of the inserted vector, not the start of the concatenated vector.
7042 unsigned V2SrcIndex = 0;
7043 if (V1DstIndex != -1) {
7044 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7045 // and don't use the original V2 at all.
7046 V2SrcIndex = Mask[V1DstIndex];
7047 V2DstIndex = V1DstIndex;
7050 V2SrcIndex = Mask[V2DstIndex] - 4;
7053 // If no V1 inputs are used in place, then the result is created only from
7054 // the zero mask and the V2 insertion - so remove V1 dependency.
7056 V1 = DAG.getUNDEF(MVT::v4f32);
7058 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7059 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7061 // Insert the V2 element into the desired position.
7063 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7064 DAG.getConstant(InsertPSMask, MVT::i8));
7067 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7068 /// UNPCK instruction.
7070 /// This specifically targets cases where we end up with alternating between
7071 /// the two inputs, and so can permute them into something that feeds a single
7072 /// UNPCK instruction. Note that this routine only targets integer vectors
7073 /// because for floating point vectors we have a generalized SHUFPS lowering
7074 /// strategy that handles everything that doesn't *exactly* match an unpack,
7075 /// making this clever lowering unnecessary.
7076 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7077 SDValue V2, ArrayRef<int> Mask,
7078 SelectionDAG &DAG) {
7079 assert(!VT.isFloatingPoint() &&
7080 "This routine only supports integer vectors.");
7081 assert(!isSingleInputShuffleMask(Mask) &&
7082 "This routine should only be used when blending two inputs.");
7083 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7085 int Size = Mask.size();
7087 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7088 return M >= 0 && M % Size < Size / 2;
7090 int NumHiInputs = std::count_if(
7091 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7093 bool UnpackLo = NumLoInputs >= NumHiInputs;
7095 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7096 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7097 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7099 for (int i = 0; i < Size; ++i) {
7103 // Each element of the unpack contains Scale elements from this mask.
7104 int UnpackIdx = i / Scale;
7106 // We only handle the case where V1 feeds the first slots of the unpack.
7107 // We rely on canonicalization to ensure this is the case.
7108 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7111 // Setup the mask for this input. The indexing is tricky as we have to
7112 // handle the unpack stride.
7113 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7114 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7118 // If we will have to shuffle both inputs to use the unpack, check whether
7119 // we can just unpack first and shuffle the result. If so, skip this unpack.
7120 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7121 !isNoopShuffleMask(V2Mask))
7124 // Shuffle the inputs into place.
7125 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7126 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7128 // Cast the inputs to the type we will use to unpack them.
7129 V1 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V1);
7130 V2 = DAG.getNode(ISD::BITCAST, DL, UnpackVT, V2);
7132 // Unpack the inputs and cast the result back to the desired type.
7133 return DAG.getNode(ISD::BITCAST, DL, VT,
7134 DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH,
7135 DL, UnpackVT, V1, V2));
7138 // We try each unpack from the largest to the smallest to try and find one
7139 // that fits this mask.
7140 int OrigNumElements = VT.getVectorNumElements();
7141 int OrigScalarSize = VT.getScalarSizeInBits();
7142 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7143 int Scale = ScalarSize / OrigScalarSize;
7144 int NumElements = OrigNumElements / Scale;
7145 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7146 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7150 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7152 if (NumLoInputs == 0 || NumHiInputs == 0) {
7153 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7154 "We have to have *some* inputs!");
7155 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7157 // FIXME: We could consider the total complexity of the permute of each
7158 // possible unpacking. Or at the least we should consider how many
7159 // half-crossings are created.
7160 // FIXME: We could consider commuting the unpacks.
7162 SmallVector<int, 32> PermMask;
7163 PermMask.assign(Size, -1);
7164 for (int i = 0; i < Size; ++i) {
7168 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7171 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7173 return DAG.getVectorShuffle(
7174 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7176 DAG.getUNDEF(VT), PermMask);
7182 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7184 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7185 /// support for floating point shuffles but not integer shuffles. These
7186 /// instructions will incur a domain crossing penalty on some chips though so
7187 /// it is better to avoid lowering through this for integer vectors where
7189 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7190 const X86Subtarget *Subtarget,
7191 SelectionDAG &DAG) {
7193 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7194 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7195 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7196 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7197 ArrayRef<int> Mask = SVOp->getMask();
7198 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7200 if (isSingleInputShuffleMask(Mask)) {
7201 // Use low duplicate instructions for masks that match their pattern.
7202 if (Subtarget->hasSSE3())
7203 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7204 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7206 // Straight shuffle of a single input vector. Simulate this by using the
7207 // single input as both of the "inputs" to this instruction..
7208 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7210 if (Subtarget->hasAVX()) {
7211 // If we have AVX, we can use VPERMILPS which will allow folding a load
7212 // into the shuffle.
7213 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7214 DAG.getConstant(SHUFPDMask, MVT::i8));
7217 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V1,
7218 DAG.getConstant(SHUFPDMask, MVT::i8));
7220 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7221 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7223 // If we have a single input, insert that into V1 if we can do so cheaply.
7224 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7225 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7226 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7228 // Try inverting the insertion since for v2 masks it is easy to do and we
7229 // can't reliably sort the mask one way or the other.
7230 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7231 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7232 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7233 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7237 // Try to use one of the special instruction patterns to handle two common
7238 // blend patterns if a zero-blend above didn't work.
7239 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7240 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7241 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7242 // We can either use a special instruction to load over the low double or
7243 // to move just the low double.
7245 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7247 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7249 if (Subtarget->hasSSE41())
7250 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7254 // Use dedicated unpack instructions for masks that match their pattern.
7255 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7256 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7257 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7258 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7260 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7261 return DAG.getNode(X86ISD::SHUFP, SDLoc(Op), MVT::v2f64, V1, V2,
7262 DAG.getConstant(SHUFPDMask, MVT::i8));
7265 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7267 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7268 /// the integer unit to minimize domain crossing penalties. However, for blends
7269 /// it falls back to the floating point shuffle operation with appropriate bit
7271 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7272 const X86Subtarget *Subtarget,
7273 SelectionDAG &DAG) {
7275 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7276 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7277 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7278 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7279 ArrayRef<int> Mask = SVOp->getMask();
7280 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7282 if (isSingleInputShuffleMask(Mask)) {
7283 // Check for being able to broadcast a single element.
7284 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7285 Mask, Subtarget, DAG))
7288 // Straight shuffle of a single input vector. For everything from SSE2
7289 // onward this has a single fast instruction with no scary immediates.
7290 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7291 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V1);
7292 int WidenedMask[4] = {
7293 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7294 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7296 ISD::BITCAST, DL, MVT::v2i64,
7297 DAG.getNode(X86ISD::PSHUFD, SDLoc(Op), MVT::v4i32, V1,
7298 getV4X86ShuffleImm8ForMask(WidenedMask, DAG)));
7300 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7301 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7302 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7303 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7305 // If we have a blend of two PACKUS operations an the blend aligns with the
7306 // low and half halves, we can just merge the PACKUS operations. This is
7307 // particularly important as it lets us merge shuffles that this routine itself
7309 auto GetPackNode = [](SDValue V) {
7310 while (V.getOpcode() == ISD::BITCAST)
7311 V = V.getOperand(0);
7313 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7315 if (SDValue V1Pack = GetPackNode(V1))
7316 if (SDValue V2Pack = GetPackNode(V2))
7317 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7318 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7319 Mask[0] == 0 ? V1Pack.getOperand(0)
7320 : V1Pack.getOperand(1),
7321 Mask[1] == 2 ? V2Pack.getOperand(0)
7322 : V2Pack.getOperand(1)));
7324 // Try to use shift instructions.
7326 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7329 // When loading a scalar and then shuffling it into a vector we can often do
7330 // the insertion cheaply.
7331 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7332 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7334 // Try inverting the insertion since for v2 masks it is easy to do and we
7335 // can't reliably sort the mask one way or the other.
7336 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7337 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7338 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7341 // We have different paths for blend lowering, but they all must use the
7342 // *exact* same predicate.
7343 bool IsBlendSupported = Subtarget->hasSSE41();
7344 if (IsBlendSupported)
7345 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7349 // Use dedicated unpack instructions for masks that match their pattern.
7350 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7351 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7352 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7353 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7355 // Try to use byte rotation instructions.
7356 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7357 if (Subtarget->hasSSSE3())
7358 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7359 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7362 // If we have direct support for blends, we should lower by decomposing into
7363 // a permute. That will be faster than the domain cross.
7364 if (IsBlendSupported)
7365 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7368 // We implement this with SHUFPD which is pretty lame because it will likely
7369 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7370 // However, all the alternatives are still more cycles and newer chips don't
7371 // have this problem. It would be really nice if x86 had better shuffles here.
7372 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V1);
7373 V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, V2);
7374 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7375 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7378 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7380 /// This is used to disable more specialized lowerings when the shufps lowering
7381 /// will happen to be efficient.
7382 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7383 // This routine only handles 128-bit shufps.
7384 assert(Mask.size() == 4 && "Unsupported mask size!");
7386 // To lower with a single SHUFPS we need to have the low half and high half
7387 // each requiring a single input.
7388 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7390 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7396 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7398 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7399 /// It makes no assumptions about whether this is the *best* lowering, it simply
7401 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7402 ArrayRef<int> Mask, SDValue V1,
7403 SDValue V2, SelectionDAG &DAG) {
7404 SDValue LowV = V1, HighV = V2;
7405 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7408 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7410 if (NumV2Elements == 1) {
7412 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7415 // Compute the index adjacent to V2Index and in the same half by toggling
7417 int V2AdjIndex = V2Index ^ 1;
7419 if (Mask[V2AdjIndex] == -1) {
7420 // Handles all the cases where we have a single V2 element and an undef.
7421 // This will only ever happen in the high lanes because we commute the
7422 // vector otherwise.
7424 std::swap(LowV, HighV);
7425 NewMask[V2Index] -= 4;
7427 // Handle the case where the V2 element ends up adjacent to a V1 element.
7428 // To make this work, blend them together as the first step.
7429 int V1Index = V2AdjIndex;
7430 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7431 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7432 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7434 // Now proceed to reconstruct the final blend as we have the necessary
7435 // high or low half formed.
7442 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7443 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
7445 } else if (NumV2Elements == 2) {
7446 if (Mask[0] < 4 && Mask[1] < 4) {
7447 // Handle the easy case where we have V1 in the low lanes and V2 in the
7451 } else if (Mask[2] < 4 && Mask[3] < 4) {
7452 // We also handle the reversed case because this utility may get called
7453 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
7454 // arrange things in the right direction.
7460 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
7461 // trying to place elements directly, just blend them and set up the final
7462 // shuffle to place them.
7464 // The first two blend mask elements are for V1, the second two are for
7466 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
7467 Mask[2] < 4 ? Mask[2] : Mask[3],
7468 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
7469 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
7470 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
7471 getV4X86ShuffleImm8ForMask(BlendMask, DAG));
7473 // Now we do a normal shuffle of V1 by giving V1 as both operands to
7476 NewMask[0] = Mask[0] < 4 ? 0 : 2;
7477 NewMask[1] = Mask[0] < 4 ? 2 : 0;
7478 NewMask[2] = Mask[2] < 4 ? 1 : 3;
7479 NewMask[3] = Mask[2] < 4 ? 3 : 1;
7482 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
7483 getV4X86ShuffleImm8ForMask(NewMask, DAG));
7486 /// \brief Lower 4-lane 32-bit floating point shuffles.
7488 /// Uses instructions exclusively from the floating point unit to minimize
7489 /// domain crossing penalties, as these are sufficient to implement all v4f32
7491 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7492 const X86Subtarget *Subtarget,
7493 SelectionDAG &DAG) {
7495 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7496 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7497 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7498 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7499 ArrayRef<int> Mask = SVOp->getMask();
7500 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7503 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7505 if (NumV2Elements == 0) {
7506 // Check for being able to broadcast a single element.
7507 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
7508 Mask, Subtarget, DAG))
7511 // Use even/odd duplicate instructions for masks that match their pattern.
7512 if (Subtarget->hasSSE3()) {
7513 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
7514 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
7515 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
7516 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
7519 if (Subtarget->hasAVX()) {
7520 // If we have AVX, we can use VPERMILPS which will allow folding a load
7521 // into the shuffle.
7522 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
7523 getV4X86ShuffleImm8ForMask(Mask, DAG));
7526 // Otherwise, use a straight shuffle of a single input vector. We pass the
7527 // input vector to both operands to simulate this with a SHUFPS.
7528 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
7529 getV4X86ShuffleImm8ForMask(Mask, DAG));
7532 // There are special ways we can lower some single-element blends. However, we
7533 // have custom ways we can lower more complex single-element blends below that
7534 // we defer to if both this and BLENDPS fail to match, so restrict this to
7535 // when the V2 input is targeting element 0 of the mask -- that is the fast
7537 if (NumV2Elements == 1 && Mask[0] >= 4)
7538 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
7539 Mask, Subtarget, DAG))
7542 if (Subtarget->hasSSE41()) {
7543 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
7547 // Use INSERTPS if we can complete the shuffle efficiently.
7548 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
7551 if (!isSingleSHUFPSMask(Mask))
7552 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
7553 DL, MVT::v4f32, V1, V2, Mask, DAG))
7557 // Use dedicated unpack instructions for masks that match their pattern.
7558 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7559 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
7560 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7561 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
7562 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7563 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
7564 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7565 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
7567 // Otherwise fall back to a SHUFPS lowering strategy.
7568 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
7571 /// \brief Lower 4-lane i32 vector shuffles.
7573 /// We try to handle these with integer-domain shuffles where we can, but for
7574 /// blends we use the floating point domain blend instructions.
7575 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7576 const X86Subtarget *Subtarget,
7577 SelectionDAG &DAG) {
7579 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
7580 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7581 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
7582 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7583 ArrayRef<int> Mask = SVOp->getMask();
7584 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7586 // Whenever we can lower this as a zext, that instruction is strictly faster
7587 // than any alternative. It also allows us to fold memory operands into the
7588 // shuffle in many cases.
7589 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
7590 Mask, Subtarget, DAG))
7594 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7596 if (NumV2Elements == 0) {
7597 // Check for being able to broadcast a single element.
7598 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
7599 Mask, Subtarget, DAG))
7602 // Straight shuffle of a single input vector. For everything from SSE2
7603 // onward this has a single fast instruction with no scary immediates.
7604 // We coerce the shuffle pattern to be compatible with UNPCK instructions
7605 // but we aren't actually going to use the UNPCK instruction because doing
7606 // so prevents folding a load into this instruction or making a copy.
7607 const int UnpackLoMask[] = {0, 0, 1, 1};
7608 const int UnpackHiMask[] = {2, 2, 3, 3};
7609 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
7610 Mask = UnpackLoMask;
7611 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
7612 Mask = UnpackHiMask;
7614 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7615 getV4X86ShuffleImm8ForMask(Mask, DAG));
7618 // Try to use shift instructions.
7620 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
7623 // There are special ways we can lower some single-element blends.
7624 if (NumV2Elements == 1)
7625 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
7626 Mask, Subtarget, DAG))
7629 // We have different paths for blend lowering, but they all must use the
7630 // *exact* same predicate.
7631 bool IsBlendSupported = Subtarget->hasSSE41();
7632 if (IsBlendSupported)
7633 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
7637 if (SDValue Masked =
7638 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
7641 // Use dedicated unpack instructions for masks that match their pattern.
7642 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
7643 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
7644 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
7645 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
7646 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
7647 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
7648 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
7649 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
7651 // Try to use byte rotation instructions.
7652 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7653 if (Subtarget->hasSSSE3())
7654 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7655 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
7658 // If we have direct support for blends, we should lower by decomposing into
7659 // a permute. That will be faster than the domain cross.
7660 if (IsBlendSupported)
7661 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
7664 // Try to lower by permuting the inputs into an unpack instruction.
7665 if (SDValue Unpack =
7666 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
7669 // We implement this with SHUFPS because it can blend from two vectors.
7670 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
7671 // up the inputs, bypassing domain shift penalties that we would encur if we
7672 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
7674 return DAG.getNode(ISD::BITCAST, DL, MVT::v4i32,
7675 DAG.getVectorShuffle(
7677 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V1),
7678 DAG.getNode(ISD::BITCAST, DL, MVT::v4f32, V2), Mask));
7681 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
7682 /// shuffle lowering, and the most complex part.
7684 /// The lowering strategy is to try to form pairs of input lanes which are
7685 /// targeted at the same half of the final vector, and then use a dword shuffle
7686 /// to place them onto the right half, and finally unpack the paired lanes into
7687 /// their final position.
7689 /// The exact breakdown of how to form these dword pairs and align them on the
7690 /// correct sides is really tricky. See the comments within the function for
7691 /// more of the details.
7692 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
7693 SDLoc DL, SDValue V, MutableArrayRef<int> Mask,
7694 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7695 assert(V.getSimpleValueType() == MVT::v8i16 && "Bad input type!");
7696 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
7697 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
7699 SmallVector<int, 4> LoInputs;
7700 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
7701 [](int M) { return M >= 0; });
7702 std::sort(LoInputs.begin(), LoInputs.end());
7703 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
7704 SmallVector<int, 4> HiInputs;
7705 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
7706 [](int M) { return M >= 0; });
7707 std::sort(HiInputs.begin(), HiInputs.end());
7708 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
7710 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
7711 int NumHToL = LoInputs.size() - NumLToL;
7713 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
7714 int NumHToH = HiInputs.size() - NumLToH;
7715 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
7716 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
7717 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
7718 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
7720 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
7721 // such inputs we can swap two of the dwords across the half mark and end up
7722 // with <=2 inputs to each half in each half. Once there, we can fall through
7723 // to the generic code below. For example:
7725 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7726 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
7728 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
7729 // and an existing 2-into-2 on the other half. In this case we may have to
7730 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
7731 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
7732 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
7733 // because any other situation (including a 3-into-1 or 1-into-3 in the other
7734 // half than the one we target for fixing) will be fixed when we re-enter this
7735 // path. We will also combine away any sequence of PSHUFD instructions that
7736 // result into a single instruction. Here is an example of the tricky case:
7738 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
7739 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
7741 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
7743 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
7744 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
7746 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
7747 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
7749 // The result is fine to be handled by the generic logic.
7750 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
7751 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
7752 int AOffset, int BOffset) {
7753 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
7754 "Must call this with A having 3 or 1 inputs from the A half.");
7755 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
7756 "Must call this with B having 1 or 3 inputs from the B half.");
7757 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
7758 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
7760 // Compute the index of dword with only one word among the three inputs in
7761 // a half by taking the sum of the half with three inputs and subtracting
7762 // the sum of the actual three inputs. The difference is the remaining
7765 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
7766 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
7767 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
7768 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
7769 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
7770 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
7771 int TripleNonInputIdx =
7772 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
7773 TripleDWord = TripleNonInputIdx / 2;
7775 // We use xor with one to compute the adjacent DWord to whichever one the
7777 OneInputDWord = (OneInput / 2) ^ 1;
7779 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
7780 // and BToA inputs. If there is also such a problem with the BToB and AToB
7781 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
7782 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
7783 // is essential that we don't *create* a 3<-1 as then we might oscillate.
7784 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
7785 // Compute how many inputs will be flipped by swapping these DWords. We
7787 // to balance this to ensure we don't form a 3-1 shuffle in the other
7789 int NumFlippedAToBInputs =
7790 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
7791 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
7792 int NumFlippedBToBInputs =
7793 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
7794 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
7795 if ((NumFlippedAToBInputs == 1 &&
7796 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
7797 (NumFlippedBToBInputs == 1 &&
7798 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
7799 // We choose whether to fix the A half or B half based on whether that
7800 // half has zero flipped inputs. At zero, we may not be able to fix it
7801 // with that half. We also bias towards fixing the B half because that
7802 // will more commonly be the high half, and we have to bias one way.
7803 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
7804 ArrayRef<int> Inputs) {
7805 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
7806 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
7807 PinnedIdx ^ 1) != Inputs.end();
7808 // Determine whether the free index is in the flipped dword or the
7809 // unflipped dword based on where the pinned index is. We use this bit
7810 // in an xor to conditionally select the adjacent dword.
7811 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
7812 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7813 FixFreeIdx) != Inputs.end();
7814 if (IsFixIdxInput == IsFixFreeIdxInput)
7816 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
7817 FixFreeIdx) != Inputs.end();
7818 assert(IsFixIdxInput != IsFixFreeIdxInput &&
7819 "We need to be changing the number of flipped inputs!");
7820 int PSHUFHalfMask[] = {0, 1, 2, 3};
7821 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
7822 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
7824 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DAG));
7827 if (M != -1 && M == FixIdx)
7829 else if (M != -1 && M == FixFreeIdx)
7832 if (NumFlippedBToBInputs != 0) {
7834 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7835 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
7837 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
7839 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
7840 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
7845 int PSHUFDMask[] = {0, 1, 2, 3};
7846 PSHUFDMask[ADWord] = BDWord;
7847 PSHUFDMask[BDWord] = ADWord;
7848 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
7849 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7850 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
7851 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
7853 // Adjust the mask to match the new locations of A and B.
7855 if (M != -1 && M/2 == ADWord)
7856 M = 2 * BDWord + M % 2;
7857 else if (M != -1 && M/2 == BDWord)
7858 M = 2 * ADWord + M % 2;
7860 // Recurse back into this routine to re-compute state now that this isn't
7861 // a 3 and 1 problem.
7862 return DAG.getVectorShuffle(MVT::v8i16, DL, V, DAG.getUNDEF(MVT::v8i16),
7865 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
7866 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
7867 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
7868 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
7870 // At this point there are at most two inputs to the low and high halves from
7871 // each half. That means the inputs can always be grouped into dwords and
7872 // those dwords can then be moved to the correct half with a dword shuffle.
7873 // We use at most one low and one high word shuffle to collect these paired
7874 // inputs into dwords, and finally a dword shuffle to place them.
7875 int PSHUFLMask[4] = {-1, -1, -1, -1};
7876 int PSHUFHMask[4] = {-1, -1, -1, -1};
7877 int PSHUFDMask[4] = {-1, -1, -1, -1};
7879 // First fix the masks for all the inputs that are staying in their
7880 // original halves. This will then dictate the targets of the cross-half
7882 auto fixInPlaceInputs =
7883 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
7884 MutableArrayRef<int> SourceHalfMask,
7885 MutableArrayRef<int> HalfMask, int HalfOffset) {
7886 if (InPlaceInputs.empty())
7888 if (InPlaceInputs.size() == 1) {
7889 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7890 InPlaceInputs[0] - HalfOffset;
7891 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
7894 if (IncomingInputs.empty()) {
7895 // Just fix all of the in place inputs.
7896 for (int Input : InPlaceInputs) {
7897 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
7898 PSHUFDMask[Input / 2] = Input / 2;
7903 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
7904 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
7905 InPlaceInputs[0] - HalfOffset;
7906 // Put the second input next to the first so that they are packed into
7907 // a dword. We find the adjacent index by toggling the low bit.
7908 int AdjIndex = InPlaceInputs[0] ^ 1;
7909 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
7910 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
7911 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
7913 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
7914 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
7916 // Now gather the cross-half inputs and place them into a free dword of
7917 // their target half.
7918 // FIXME: This operation could almost certainly be simplified dramatically to
7919 // look more like the 3-1 fixing operation.
7920 auto moveInputsToRightHalf = [&PSHUFDMask](
7921 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
7922 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
7923 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
7925 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
7926 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
7928 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
7930 int LowWord = Word & ~1;
7931 int HighWord = Word | 1;
7932 return isWordClobbered(SourceHalfMask, LowWord) ||
7933 isWordClobbered(SourceHalfMask, HighWord);
7936 if (IncomingInputs.empty())
7939 if (ExistingInputs.empty()) {
7940 // Map any dwords with inputs from them into the right half.
7941 for (int Input : IncomingInputs) {
7942 // If the source half mask maps over the inputs, turn those into
7943 // swaps and use the swapped lane.
7944 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
7945 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
7946 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
7947 Input - SourceOffset;
7948 // We have to swap the uses in our half mask in one sweep.
7949 for (int &M : HalfMask)
7950 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
7952 else if (M == Input)
7953 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7955 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
7956 Input - SourceOffset &&
7957 "Previous placement doesn't match!");
7959 // Note that this correctly re-maps both when we do a swap and when
7960 // we observe the other side of the swap above. We rely on that to
7961 // avoid swapping the members of the input list directly.
7962 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
7965 // Map the input's dword into the correct half.
7966 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
7967 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
7969 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
7971 "Previous placement doesn't match!");
7974 // And just directly shift any other-half mask elements to be same-half
7975 // as we will have mirrored the dword containing the element into the
7976 // same position within that half.
7977 for (int &M : HalfMask)
7978 if (M >= SourceOffset && M < SourceOffset + 4) {
7979 M = M - SourceOffset + DestOffset;
7980 assert(M >= 0 && "This should never wrap below zero!");
7985 // Ensure we have the input in a viable dword of its current half. This
7986 // is particularly tricky because the original position may be clobbered
7987 // by inputs being moved and *staying* in that half.
7988 if (IncomingInputs.size() == 1) {
7989 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
7990 int InputFixed = std::find(std::begin(SourceHalfMask),
7991 std::end(SourceHalfMask), -1) -
7992 std::begin(SourceHalfMask) + SourceOffset;
7993 SourceHalfMask[InputFixed - SourceOffset] =
7994 IncomingInputs[0] - SourceOffset;
7995 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
7997 IncomingInputs[0] = InputFixed;
7999 } else if (IncomingInputs.size() == 2) {
8000 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8001 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8002 // We have two non-adjacent or clobbered inputs we need to extract from
8003 // the source half. To do this, we need to map them into some adjacent
8004 // dword slot in the source mask.
8005 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8006 IncomingInputs[1] - SourceOffset};
8008 // If there is a free slot in the source half mask adjacent to one of
8009 // the inputs, place the other input in it. We use (Index XOR 1) to
8010 // compute an adjacent index.
8011 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8012 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8013 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8014 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8015 InputsFixed[1] = InputsFixed[0] ^ 1;
8016 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8017 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8018 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8019 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8020 InputsFixed[0] = InputsFixed[1] ^ 1;
8021 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8022 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8023 // The two inputs are in the same DWord but it is clobbered and the
8024 // adjacent DWord isn't used at all. Move both inputs to the free
8026 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8027 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8028 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8029 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8031 // The only way we hit this point is if there is no clobbering
8032 // (because there are no off-half inputs to this half) and there is no
8033 // free slot adjacent to one of the inputs. In this case, we have to
8034 // swap an input with a non-input.
8035 for (int i = 0; i < 4; ++i)
8036 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8037 "We can't handle any clobbers here!");
8038 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8039 "Cannot have adjacent inputs here!");
8041 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8042 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8044 // We also have to update the final source mask in this case because
8045 // it may need to undo the above swap.
8046 for (int &M : FinalSourceHalfMask)
8047 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8048 M = InputsFixed[1] + SourceOffset;
8049 else if (M == InputsFixed[1] + SourceOffset)
8050 M = (InputsFixed[0] ^ 1) + SourceOffset;
8052 InputsFixed[1] = InputsFixed[0] ^ 1;
8055 // Point everything at the fixed inputs.
8056 for (int &M : HalfMask)
8057 if (M == IncomingInputs[0])
8058 M = InputsFixed[0] + SourceOffset;
8059 else if (M == IncomingInputs[1])
8060 M = InputsFixed[1] + SourceOffset;
8062 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8063 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8066 llvm_unreachable("Unhandled input size!");
8069 // Now hoist the DWord down to the right half.
8070 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8071 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8072 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8073 for (int &M : HalfMask)
8074 for (int Input : IncomingInputs)
8076 M = FreeDWord * 2 + Input % 2;
8078 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8079 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8080 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8081 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8083 // Now enact all the shuffles we've computed to move the inputs into their
8085 if (!isNoopShuffleMask(PSHUFLMask))
8086 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8087 getV4X86ShuffleImm8ForMask(PSHUFLMask, DAG));
8088 if (!isNoopShuffleMask(PSHUFHMask))
8089 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8090 getV4X86ShuffleImm8ForMask(PSHUFHMask, DAG));
8091 if (!isNoopShuffleMask(PSHUFDMask))
8092 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8093 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
8094 DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V),
8095 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
8097 // At this point, each half should contain all its inputs, and we can then
8098 // just shuffle them into their final position.
8099 assert(std::count_if(LoMask.begin(), LoMask.end(),
8100 [](int M) { return M >= 4; }) == 0 &&
8101 "Failed to lift all the high half inputs to the low mask!");
8102 assert(std::count_if(HiMask.begin(), HiMask.end(),
8103 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8104 "Failed to lift all the low half inputs to the high mask!");
8106 // Do a half shuffle for the low mask.
8107 if (!isNoopShuffleMask(LoMask))
8108 V = DAG.getNode(X86ISD::PSHUFLW, DL, MVT::v8i16, V,
8109 getV4X86ShuffleImm8ForMask(LoMask, DAG));
8111 // Do a half shuffle with the high mask after shifting its values down.
8112 for (int &M : HiMask)
8115 if (!isNoopShuffleMask(HiMask))
8116 V = DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16, V,
8117 getV4X86ShuffleImm8ForMask(HiMask, DAG));
8122 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8123 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8124 SDValue V2, ArrayRef<int> Mask,
8125 SelectionDAG &DAG, bool &V1InUse,
8127 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8133 int Size = Mask.size();
8134 int Scale = 16 / Size;
8135 for (int i = 0; i < 16; ++i) {
8136 if (Mask[i / Scale] == -1) {
8137 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8139 const int ZeroMask = 0x80;
8140 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8142 int V2Idx = Mask[i / Scale] < Size
8144 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8145 if (Zeroable[i / Scale])
8146 V1Idx = V2Idx = ZeroMask;
8147 V1Mask[i] = DAG.getConstant(V1Idx, MVT::i8);
8148 V2Mask[i] = DAG.getConstant(V2Idx, MVT::i8);
8149 V1InUse |= (ZeroMask != V1Idx);
8150 V2InUse |= (ZeroMask != V2Idx);
8155 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8156 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V1),
8157 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8159 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8160 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, V2),
8161 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8163 // If we need shuffled inputs from both, blend the two.
8165 if (V1InUse && V2InUse)
8166 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8168 V = V1InUse ? V1 : V2;
8170 // Cast the result back to the correct type.
8171 return DAG.getNode(ISD::BITCAST, DL, VT, V);
8174 /// \brief Generic lowering of 8-lane i16 shuffles.
8176 /// This handles both single-input shuffles and combined shuffle/blends with
8177 /// two inputs. The single input shuffles are immediately delegated to
8178 /// a dedicated lowering routine.
8180 /// The blends are lowered in one of three fundamental ways. If there are few
8181 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8182 /// of the input is significantly cheaper when lowered as an interleaving of
8183 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8184 /// halves of the inputs separately (making them have relatively few inputs)
8185 /// and then concatenate them.
8186 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8187 const X86Subtarget *Subtarget,
8188 SelectionDAG &DAG) {
8190 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8191 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8192 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8193 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8194 ArrayRef<int> OrigMask = SVOp->getMask();
8195 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8196 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8197 MutableArrayRef<int> Mask(MaskStorage);
8199 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8201 // Whenever we can lower this as a zext, that instruction is strictly faster
8202 // than any alternative.
8203 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8204 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8207 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8209 auto isV2 = [](int M) { return M >= 8; };
8211 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8213 if (NumV2Inputs == 0) {
8214 // Check for being able to broadcast a single element.
8215 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8216 Mask, Subtarget, DAG))
8219 // Try to use shift instructions.
8221 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8224 // Use dedicated unpack instructions for masks that match their pattern.
8225 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8226 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8227 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8228 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8230 // Try to use byte rotation instructions.
8231 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8232 Mask, Subtarget, DAG))
8235 return lowerV8I16GeneralSingleInputVectorShuffle(DL, V1, Mask, Subtarget,
8239 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8240 "All single-input shuffles should be canonicalized to be V1-input "
8243 // Try to use shift instructions.
8245 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8248 // There are special ways we can lower some single-element blends.
8249 if (NumV2Inputs == 1)
8250 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8251 Mask, Subtarget, DAG))
8254 // We have different paths for blend lowering, but they all must use the
8255 // *exact* same predicate.
8256 bool IsBlendSupported = Subtarget->hasSSE41();
8257 if (IsBlendSupported)
8258 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8262 if (SDValue Masked =
8263 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8266 // Use dedicated unpack instructions for masks that match their pattern.
8267 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8268 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8269 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8270 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8272 // Try to use byte rotation instructions.
8273 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8274 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8277 if (SDValue BitBlend =
8278 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8281 if (SDValue Unpack =
8282 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8285 // If we can't directly blend but can use PSHUFB, that will be better as it
8286 // can both shuffle and set up the inefficient blend.
8287 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8288 bool V1InUse, V2InUse;
8289 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8293 // We can always bit-blend if we have to so the fallback strategy is to
8294 // decompose into single-input permutes and blends.
8295 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8299 /// \brief Check whether a compaction lowering can be done by dropping even
8300 /// elements and compute how many times even elements must be dropped.
8302 /// This handles shuffles which take every Nth element where N is a power of
8303 /// two. Example shuffle masks:
8305 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8306 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8307 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8308 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8309 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8310 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8312 /// Any of these lanes can of course be undef.
8314 /// This routine only supports N <= 3.
8315 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8318 /// \returns N above, or the number of times even elements must be dropped if
8319 /// there is such a number. Otherwise returns zero.
8320 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8321 // Figure out whether we're looping over two inputs or just one.
8322 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8324 // The modulus for the shuffle vector entries is based on whether this is
8325 // a single input or not.
8326 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8327 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8328 "We should only be called with masks with a power-of-2 size!");
8330 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8332 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8333 // and 2^3 simultaneously. This is because we may have ambiguity with
8334 // partially undef inputs.
8335 bool ViableForN[3] = {true, true, true};
8337 for (int i = 0, e = Mask.size(); i < e; ++i) {
8338 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8343 bool IsAnyViable = false;
8344 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8345 if (ViableForN[j]) {
8348 // The shuffle mask must be equal to (i * 2^N) % M.
8349 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8352 ViableForN[j] = false;
8354 // Early exit if we exhaust the possible powers of two.
8359 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8363 // Return 0 as there is no viable power of two.
8367 /// \brief Generic lowering of v16i8 shuffles.
8369 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8370 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8371 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8372 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8374 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8375 const X86Subtarget *Subtarget,
8376 SelectionDAG &DAG) {
8378 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8379 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8380 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8381 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8382 ArrayRef<int> Mask = SVOp->getMask();
8383 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8385 // Try to use shift instructions.
8387 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8390 // Try to use byte rotation instructions.
8391 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8392 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8395 // Try to use a zext lowering.
8396 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8397 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8401 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8403 // For single-input shuffles, there are some nicer lowering tricks we can use.
8404 if (NumV2Elements == 0) {
8405 // Check for being able to broadcast a single element.
8406 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8407 Mask, Subtarget, DAG))
8410 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8411 // Notably, this handles splat and partial-splat shuffles more efficiently.
8412 // However, it only makes sense if the pre-duplication shuffle simplifies
8413 // things significantly. Currently, this means we need to be able to
8414 // express the pre-duplication shuffle as an i16 shuffle.
8416 // FIXME: We should check for other patterns which can be widened into an
8417 // i16 shuffle as well.
8418 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8419 for (int i = 0; i < 16; i += 2)
8420 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8425 auto tryToWidenViaDuplication = [&]() -> SDValue {
8426 if (!canWidenViaDuplication(Mask))
8428 SmallVector<int, 4> LoInputs;
8429 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
8430 [](int M) { return M >= 0 && M < 8; });
8431 std::sort(LoInputs.begin(), LoInputs.end());
8432 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
8434 SmallVector<int, 4> HiInputs;
8435 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
8436 [](int M) { return M >= 8; });
8437 std::sort(HiInputs.begin(), HiInputs.end());
8438 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
8441 bool TargetLo = LoInputs.size() >= HiInputs.size();
8442 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
8443 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
8445 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
8446 SmallDenseMap<int, int, 8> LaneMap;
8447 for (int I : InPlaceInputs) {
8448 PreDupI16Shuffle[I/2] = I/2;
8451 int j = TargetLo ? 0 : 4, je = j + 4;
8452 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
8453 // Check if j is already a shuffle of this input. This happens when
8454 // there are two adjacent bytes after we move the low one.
8455 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
8456 // If we haven't yet mapped the input, search for a slot into which
8458 while (j < je && PreDupI16Shuffle[j] != -1)
8462 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
8465 // Map this input with the i16 shuffle.
8466 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
8469 // Update the lane map based on the mapping we ended up with.
8470 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
8473 ISD::BITCAST, DL, MVT::v16i8,
8474 DAG.getVectorShuffle(MVT::v8i16, DL,
8475 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8476 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
8478 // Unpack the bytes to form the i16s that will be shuffled into place.
8479 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8480 MVT::v16i8, V1, V1);
8482 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8483 for (int i = 0; i < 16; ++i)
8484 if (Mask[i] != -1) {
8485 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
8486 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
8487 if (PostDupI16Shuffle[i / 2] == -1)
8488 PostDupI16Shuffle[i / 2] = MappedMask;
8490 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
8491 "Conflicting entrties in the original shuffle!");
8494 ISD::BITCAST, DL, MVT::v16i8,
8495 DAG.getVectorShuffle(MVT::v8i16, DL,
8496 DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
8497 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
8499 if (SDValue V = tryToWidenViaDuplication())
8503 // Use dedicated unpack instructions for masks that match their pattern.
8504 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8505 0, 16, 1, 17, 2, 18, 3, 19,
8507 4, 20, 5, 21, 6, 22, 7, 23}))
8508 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
8509 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
8510 8, 24, 9, 25, 10, 26, 11, 27,
8512 12, 28, 13, 29, 14, 30, 15, 31}))
8513 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
8515 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
8516 // with PSHUFB. It is important to do this before we attempt to generate any
8517 // blends but after all of the single-input lowerings. If the single input
8518 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
8519 // want to preserve that and we can DAG combine any longer sequences into
8520 // a PSHUFB in the end. But once we start blending from multiple inputs,
8521 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
8522 // and there are *very* few patterns that would actually be faster than the
8523 // PSHUFB approach because of its ability to zero lanes.
8525 // FIXME: The only exceptions to the above are blends which are exact
8526 // interleavings with direct instructions supporting them. We currently don't
8527 // handle those well here.
8528 if (Subtarget->hasSSSE3()) {
8529 bool V1InUse = false;
8530 bool V2InUse = false;
8532 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
8533 DAG, V1InUse, V2InUse);
8535 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
8536 // do so. This avoids using them to handle blends-with-zero which is
8537 // important as a single pshufb is significantly faster for that.
8538 if (V1InUse && V2InUse) {
8539 if (Subtarget->hasSSE41())
8540 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
8541 Mask, Subtarget, DAG))
8544 // We can use an unpack to do the blending rather than an or in some
8545 // cases. Even though the or may be (very minorly) more efficient, we
8546 // preference this lowering because there are common cases where part of
8547 // the complexity of the shuffles goes away when we do the final blend as
8549 // FIXME: It might be worth trying to detect if the unpack-feeding
8550 // shuffles will both be pshufb, in which case we shouldn't bother with
8552 if (SDValue Unpack =
8553 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
8560 // There are special ways we can lower some single-element blends.
8561 if (NumV2Elements == 1)
8562 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
8563 Mask, Subtarget, DAG))
8566 if (SDValue BitBlend =
8567 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
8570 // Check whether a compaction lowering can be done. This handles shuffles
8571 // which take every Nth element for some even N. See the helper function for
8574 // We special case these as they can be particularly efficiently handled with
8575 // the PACKUSB instruction on x86 and they show up in common patterns of
8576 // rearranging bytes to truncate wide elements.
8577 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
8578 // NumEvenDrops is the power of two stride of the elements. Another way of
8579 // thinking about it is that we need to drop the even elements this many
8580 // times to get the original input.
8581 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8583 // First we need to zero all the dropped bytes.
8584 assert(NumEvenDrops <= 3 &&
8585 "No support for dropping even elements more than 3 times.");
8586 // We use the mask type to pick which bytes are preserved based on how many
8587 // elements are dropped.
8588 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
8589 SDValue ByteClearMask =
8590 DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
8591 DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
8592 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
8594 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
8596 // Now pack things back together.
8597 V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
8598 V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
8599 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
8600 for (int i = 1; i < NumEvenDrops; ++i) {
8601 Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
8602 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
8608 // Handle multi-input cases by blending single-input shuffles.
8609 if (NumV2Elements > 0)
8610 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
8613 // The fallback path for single-input shuffles widens this into two v8i16
8614 // vectors with unpacks, shuffles those, and then pulls them back together
8618 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8619 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
8620 for (int i = 0; i < 16; ++i)
8622 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
8624 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
8626 SDValue VLoHalf, VHiHalf;
8627 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
8628 // them out and avoid using UNPCK{L,H} to extract the elements of V as
8630 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
8631 [](int M) { return M >= 0 && M % 2 == 1; }) &&
8632 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
8633 [](int M) { return M >= 0 && M % 2 == 1; })) {
8634 // Use a mask to drop the high bytes.
8635 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
8636 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
8637 DAG.getConstant(0x00FF, MVT::v8i16));
8639 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
8640 VHiHalf = DAG.getUNDEF(MVT::v8i16);
8642 // Squash the masks to point directly into VLoHalf.
8643 for (int &M : LoBlendMask)
8646 for (int &M : HiBlendMask)
8650 // Otherwise just unpack the low half of V into VLoHalf and the high half into
8651 // VHiHalf so that we can blend them as i16s.
8652 VLoHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8653 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
8654 VHiHalf = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
8655 DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
8658 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
8659 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
8661 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
8664 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
8666 /// This routine breaks down the specific type of 128-bit shuffle and
8667 /// dispatches to the lowering routines accordingly.
8668 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8669 MVT VT, const X86Subtarget *Subtarget,
8670 SelectionDAG &DAG) {
8671 switch (VT.SimpleTy) {
8673 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8675 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
8677 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8679 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
8681 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
8683 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
8686 llvm_unreachable("Unimplemented!");
8690 /// \brief Helper function to test whether a shuffle mask could be
8691 /// simplified by widening the elements being shuffled.
8693 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
8694 /// leaves it in an unspecified state.
8696 /// NOTE: This must handle normal vector shuffle masks and *target* vector
8697 /// shuffle masks. The latter have the special property of a '-2' representing
8698 /// a zero-ed lane of a vector.
8699 static bool canWidenShuffleElements(ArrayRef<int> Mask,
8700 SmallVectorImpl<int> &WidenedMask) {
8701 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
8702 // If both elements are undef, its trivial.
8703 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
8704 WidenedMask.push_back(SM_SentinelUndef);
8708 // Check for an undef mask and a mask value properly aligned to fit with
8709 // a pair of values. If we find such a case, use the non-undef mask's value.
8710 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
8711 WidenedMask.push_back(Mask[i + 1] / 2);
8714 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
8715 WidenedMask.push_back(Mask[i] / 2);
8719 // When zeroing, we need to spread the zeroing across both lanes to widen.
8720 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
8721 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
8722 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
8723 WidenedMask.push_back(SM_SentinelZero);
8729 // Finally check if the two mask values are adjacent and aligned with
8731 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
8732 WidenedMask.push_back(Mask[i] / 2);
8736 // Otherwise we can't safely widen the elements used in this shuffle.
8739 assert(WidenedMask.size() == Mask.size() / 2 &&
8740 "Incorrect size of mask after widening the elements!");
8745 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
8747 /// This routine just extracts two subvectors, shuffles them independently, and
8748 /// then concatenates them back together. This should work effectively with all
8749 /// AVX vector shuffle types.
8750 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8751 SDValue V2, ArrayRef<int> Mask,
8752 SelectionDAG &DAG) {
8753 assert(VT.getSizeInBits() >= 256 &&
8754 "Only for 256-bit or wider vector shuffles!");
8755 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
8756 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
8758 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
8759 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
8761 int NumElements = VT.getVectorNumElements();
8762 int SplitNumElements = NumElements / 2;
8763 MVT ScalarVT = VT.getScalarType();
8764 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
8766 // Rather than splitting build-vectors, just build two narrower build
8767 // vectors. This helps shuffling with splats and zeros.
8768 auto SplitVector = [&](SDValue V) {
8769 while (V.getOpcode() == ISD::BITCAST)
8770 V = V->getOperand(0);
8772 MVT OrigVT = V.getSimpleValueType();
8773 int OrigNumElements = OrigVT.getVectorNumElements();
8774 int OrigSplitNumElements = OrigNumElements / 2;
8775 MVT OrigScalarVT = OrigVT.getScalarType();
8776 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
8780 auto *BV = dyn_cast<BuildVectorSDNode>(V);
8782 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8783 DAG.getIntPtrConstant(0));
8784 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
8785 DAG.getIntPtrConstant(OrigSplitNumElements));
8788 SmallVector<SDValue, 16> LoOps, HiOps;
8789 for (int i = 0; i < OrigSplitNumElements; ++i) {
8790 LoOps.push_back(BV->getOperand(i));
8791 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
8793 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
8794 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
8796 return std::make_pair(DAG.getNode(ISD::BITCAST, DL, SplitVT, LoV),
8797 DAG.getNode(ISD::BITCAST, DL, SplitVT, HiV));
8800 SDValue LoV1, HiV1, LoV2, HiV2;
8801 std::tie(LoV1, HiV1) = SplitVector(V1);
8802 std::tie(LoV2, HiV2) = SplitVector(V2);
8804 // Now create two 4-way blends of these half-width vectors.
8805 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
8806 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
8807 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
8808 for (int i = 0; i < SplitNumElements; ++i) {
8809 int M = HalfMask[i];
8810 if (M >= NumElements) {
8811 if (M >= NumElements + SplitNumElements)
8815 V2BlendMask.push_back(M - NumElements);
8816 V1BlendMask.push_back(-1);
8817 BlendMask.push_back(SplitNumElements + i);
8818 } else if (M >= 0) {
8819 if (M >= SplitNumElements)
8823 V2BlendMask.push_back(-1);
8824 V1BlendMask.push_back(M);
8825 BlendMask.push_back(i);
8827 V2BlendMask.push_back(-1);
8828 V1BlendMask.push_back(-1);
8829 BlendMask.push_back(-1);
8833 // Because the lowering happens after all combining takes place, we need to
8834 // manually combine these blend masks as much as possible so that we create
8835 // a minimal number of high-level vector shuffle nodes.
8837 // First try just blending the halves of V1 or V2.
8838 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
8839 return DAG.getUNDEF(SplitVT);
8840 if (!UseLoV2 && !UseHiV2)
8841 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8842 if (!UseLoV1 && !UseHiV1)
8843 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8845 SDValue V1Blend, V2Blend;
8846 if (UseLoV1 && UseHiV1) {
8848 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
8850 // We only use half of V1 so map the usage down into the final blend mask.
8851 V1Blend = UseLoV1 ? LoV1 : HiV1;
8852 for (int i = 0; i < SplitNumElements; ++i)
8853 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
8854 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
8856 if (UseLoV2 && UseHiV2) {
8858 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
8860 // We only use half of V2 so map the usage down into the final blend mask.
8861 V2Blend = UseLoV2 ? LoV2 : HiV2;
8862 for (int i = 0; i < SplitNumElements; ++i)
8863 if (BlendMask[i] >= SplitNumElements)
8864 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
8866 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
8868 SDValue Lo = HalfBlend(LoMask);
8869 SDValue Hi = HalfBlend(HiMask);
8870 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
8873 /// \brief Either split a vector in halves or decompose the shuffles and the
8876 /// This is provided as a good fallback for many lowerings of non-single-input
8877 /// shuffles with more than one 128-bit lane. In those cases, we want to select
8878 /// between splitting the shuffle into 128-bit components and stitching those
8879 /// back together vs. extracting the single-input shuffles and blending those
8881 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
8882 SDValue V2, ArrayRef<int> Mask,
8883 SelectionDAG &DAG) {
8884 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
8885 "lower single-input shuffles as it "
8886 "could then recurse on itself.");
8887 int Size = Mask.size();
8889 // If this can be modeled as a broadcast of two elements followed by a blend,
8890 // prefer that lowering. This is especially important because broadcasts can
8891 // often fold with memory operands.
8892 auto DoBothBroadcast = [&] {
8893 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
8896 if (V2BroadcastIdx == -1)
8897 V2BroadcastIdx = M - Size;
8898 else if (M - Size != V2BroadcastIdx)
8900 } else if (M >= 0) {
8901 if (V1BroadcastIdx == -1)
8903 else if (M != V1BroadcastIdx)
8908 if (DoBothBroadcast())
8909 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
8912 // If the inputs all stem from a single 128-bit lane of each input, then we
8913 // split them rather than blending because the split will decompose to
8914 // unusually few instructions.
8915 int LaneCount = VT.getSizeInBits() / 128;
8916 int LaneSize = Size / LaneCount;
8917 SmallBitVector LaneInputs[2];
8918 LaneInputs[0].resize(LaneCount, false);
8919 LaneInputs[1].resize(LaneCount, false);
8920 for (int i = 0; i < Size; ++i)
8922 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
8923 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
8924 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8926 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
8927 // that the decomposed single-input shuffles don't end up here.
8928 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8931 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
8932 /// a permutation and blend of those lanes.
8934 /// This essentially blends the out-of-lane inputs to each lane into the lane
8935 /// from a permuted copy of the vector. This lowering strategy results in four
8936 /// instructions in the worst case for a single-input cross lane shuffle which
8937 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
8938 /// of. Special cases for each particular shuffle pattern should be handled
8939 /// prior to trying this lowering.
8940 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
8941 SDValue V1, SDValue V2,
8943 SelectionDAG &DAG) {
8944 // FIXME: This should probably be generalized for 512-bit vectors as well.
8945 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
8946 int LaneSize = Mask.size() / 2;
8948 // If there are only inputs from one 128-bit lane, splitting will in fact be
8949 // less expensive. The flags track wether the given lane contains an element
8950 // that crosses to another lane.
8951 bool LaneCrossing[2] = {false, false};
8952 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8953 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
8954 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
8955 if (!LaneCrossing[0] || !LaneCrossing[1])
8956 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
8958 if (isSingleInputShuffleMask(Mask)) {
8959 SmallVector<int, 32> FlippedBlendMask;
8960 for (int i = 0, Size = Mask.size(); i < Size; ++i)
8961 FlippedBlendMask.push_back(
8962 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
8964 : Mask[i] % LaneSize +
8965 (i / LaneSize) * LaneSize + Size));
8967 // Flip the vector, and blend the results which should now be in-lane. The
8968 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
8969 // 5 for the high source. The value 3 selects the high half of source 2 and
8970 // the value 2 selects the low half of source 2. We only use source 2 to
8971 // allow folding it into a memory operand.
8972 unsigned PERMMask = 3 | 2 << 4;
8973 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
8974 V1, DAG.getConstant(PERMMask, MVT::i8));
8975 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
8978 // This now reduces to two single-input shuffles of V1 and V2 which at worst
8979 // will be handled by the above logic and a blend of the results, much like
8980 // other patterns in AVX.
8981 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
8984 /// \brief Handle lowering 2-lane 128-bit shuffles.
8985 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
8986 SDValue V2, ArrayRef<int> Mask,
8987 const X86Subtarget *Subtarget,
8988 SelectionDAG &DAG) {
8989 // Blends are faster and handle all the non-lane-crossing cases.
8990 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
8994 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
8995 VT.getVectorNumElements() / 2);
8996 // Check for patterns which can be matched with a single insert of a 128-bit
8998 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1}) ||
8999 isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9000 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9001 DAG.getIntPtrConstant(0));
9002 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9003 Mask[2] < 4 ? V1 : V2, DAG.getIntPtrConstant(0));
9004 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9006 if (isShuffleEquivalent(V1, V2, Mask, {0, 1, 6, 7})) {
9007 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9008 DAG.getIntPtrConstant(0));
9009 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
9010 DAG.getIntPtrConstant(2));
9011 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9014 // Otherwise form a 128-bit permutation.
9015 // FIXME: Detect zero-vector inputs and use the VPERM2X128 to zero that half.
9016 unsigned PermMask = Mask[0] / 2 | (Mask[2] / 2) << 4;
9017 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9018 DAG.getConstant(PermMask, MVT::i8));
9021 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9022 /// shuffling each lane.
9024 /// This will only succeed when the result of fixing the 128-bit lanes results
9025 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9026 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9027 /// the lane crosses early and then use simpler shuffles within each lane.
9029 /// FIXME: It might be worthwhile at some point to support this without
9030 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9031 /// in x86 only floating point has interesting non-repeating shuffles, and even
9032 /// those are still *marginally* more expensive.
9033 static SDValue lowerVectorShuffleByMerging128BitLanes(
9034 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9035 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9036 assert(!isSingleInputShuffleMask(Mask) &&
9037 "This is only useful with multiple inputs.");
9039 int Size = Mask.size();
9040 int LaneSize = 128 / VT.getScalarSizeInBits();
9041 int NumLanes = Size / LaneSize;
9042 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9044 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9045 // check whether the in-128-bit lane shuffles share a repeating pattern.
9046 SmallVector<int, 4> Lanes;
9047 Lanes.resize(NumLanes, -1);
9048 SmallVector<int, 4> InLaneMask;
9049 InLaneMask.resize(LaneSize, -1);
9050 for (int i = 0; i < Size; ++i) {
9054 int j = i / LaneSize;
9057 // First entry we've seen for this lane.
9058 Lanes[j] = Mask[i] / LaneSize;
9059 } else if (Lanes[j] != Mask[i] / LaneSize) {
9060 // This doesn't match the lane selected previously!
9064 // Check that within each lane we have a consistent shuffle mask.
9065 int k = i % LaneSize;
9066 if (InLaneMask[k] < 0) {
9067 InLaneMask[k] = Mask[i] % LaneSize;
9068 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9069 // This doesn't fit a repeating in-lane mask.
9074 // First shuffle the lanes into place.
9075 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9076 VT.getSizeInBits() / 64);
9077 SmallVector<int, 8> LaneMask;
9078 LaneMask.resize(NumLanes * 2, -1);
9079 for (int i = 0; i < NumLanes; ++i)
9080 if (Lanes[i] >= 0) {
9081 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9082 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9085 V1 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V1);
9086 V2 = DAG.getNode(ISD::BITCAST, DL, LaneVT, V2);
9087 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9089 // Cast it back to the type we actually want.
9090 LaneShuffle = DAG.getNode(ISD::BITCAST, DL, VT, LaneShuffle);
9092 // Now do a simple shuffle that isn't lane crossing.
9093 SmallVector<int, 8> NewMask;
9094 NewMask.resize(Size, -1);
9095 for (int i = 0; i < Size; ++i)
9097 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9098 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9099 "Must not introduce lane crosses at this point!");
9101 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9104 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9107 /// This returns true if the elements from a particular input are already in the
9108 /// slot required by the given mask and require no permutation.
9109 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9110 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9111 int Size = Mask.size();
9112 for (int i = 0; i < Size; ++i)
9113 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9119 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9121 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9122 /// isn't available.
9123 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9124 const X86Subtarget *Subtarget,
9125 SelectionDAG &DAG) {
9127 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9128 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9129 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9130 ArrayRef<int> Mask = SVOp->getMask();
9131 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9133 SmallVector<int, 4> WidenedMask;
9134 if (canWidenShuffleElements(Mask, WidenedMask))
9135 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9138 if (isSingleInputShuffleMask(Mask)) {
9139 // Check for being able to broadcast a single element.
9140 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9141 Mask, Subtarget, DAG))
9144 // Use low duplicate instructions for masks that match their pattern.
9145 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9146 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9148 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9149 // Non-half-crossing single input shuffles can be lowerid with an
9150 // interleaved permutation.
9151 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9152 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9153 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9154 DAG.getConstant(VPERMILPMask, MVT::i8));
9157 // With AVX2 we have direct support for this permutation.
9158 if (Subtarget->hasAVX2())
9159 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9160 getV4X86ShuffleImm8ForMask(Mask, DAG));
9162 // Otherwise, fall back.
9163 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9167 // X86 has dedicated unpack instructions that can handle specific blend
9168 // operations: UNPCKH and UNPCKL.
9169 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9170 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9171 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9172 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9173 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9174 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9175 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9176 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9178 // If we have a single input to the zero element, insert that into V1 if we
9179 // can do so cheaply.
9181 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
9182 if (NumV2Elements == 1 && Mask[0] >= 4)
9183 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
9184 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9187 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9191 // Check if the blend happens to exactly fit that of SHUFPD.
9192 if ((Mask[0] == -1 || Mask[0] < 2) &&
9193 (Mask[1] == -1 || (Mask[1] >= 4 && Mask[1] < 6)) &&
9194 (Mask[2] == -1 || (Mask[2] >= 2 && Mask[2] < 4)) &&
9195 (Mask[3] == -1 || Mask[3] >= 6)) {
9196 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 5) << 1) |
9197 ((Mask[2] == 3) << 2) | ((Mask[3] == 7) << 3);
9198 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V1, V2,
9199 DAG.getConstant(SHUFPDMask, MVT::i8));
9201 if ((Mask[0] == -1 || (Mask[0] >= 4 && Mask[0] < 6)) &&
9202 (Mask[1] == -1 || Mask[1] < 2) &&
9203 (Mask[2] == -1 || Mask[2] >= 6) &&
9204 (Mask[3] == -1 || (Mask[3] >= 2 && Mask[3] < 4))) {
9205 unsigned SHUFPDMask = (Mask[0] == 5) | ((Mask[1] == 1) << 1) |
9206 ((Mask[2] == 7) << 2) | ((Mask[3] == 3) << 3);
9207 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f64, V2, V1,
9208 DAG.getConstant(SHUFPDMask, MVT::i8));
9211 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9212 // shuffle. However, if we have AVX2 and either inputs are already in place,
9213 // we will be able to shuffle even across lanes the other input in a single
9214 // instruction so skip this pattern.
9215 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9216 isShuffleMaskInputInPlace(1, Mask))))
9217 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9218 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9221 // If we have AVX2 then we always want to lower with a blend because an v4 we
9222 // can fully permute the elements.
9223 if (Subtarget->hasAVX2())
9224 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9227 // Otherwise fall back on generic lowering.
9228 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9231 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9233 /// This routine is only called when we have AVX2 and thus a reasonable
9234 /// instruction set for v4i64 shuffling..
9235 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9236 const X86Subtarget *Subtarget,
9237 SelectionDAG &DAG) {
9239 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9240 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9241 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9242 ArrayRef<int> Mask = SVOp->getMask();
9243 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9244 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9246 SmallVector<int, 4> WidenedMask;
9247 if (canWidenShuffleElements(Mask, WidenedMask))
9248 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9251 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9255 // Check for being able to broadcast a single element.
9256 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9257 Mask, Subtarget, DAG))
9260 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9261 // use lower latency instructions that will operate on both 128-bit lanes.
9262 SmallVector<int, 2> RepeatedMask;
9263 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9264 if (isSingleInputShuffleMask(Mask)) {
9265 int PSHUFDMask[] = {-1, -1, -1, -1};
9266 for (int i = 0; i < 2; ++i)
9267 if (RepeatedMask[i] >= 0) {
9268 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9269 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9272 ISD::BITCAST, DL, MVT::v4i64,
9273 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9274 DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, V1),
9275 getV4X86ShuffleImm8ForMask(PSHUFDMask, DAG)));
9279 // AVX2 provides a direct instruction for permuting a single input across
9281 if (isSingleInputShuffleMask(Mask))
9282 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9283 getV4X86ShuffleImm8ForMask(Mask, DAG));
9285 // Try to use shift instructions.
9287 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9290 // Use dedicated unpack instructions for masks that match their pattern.
9291 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9292 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9293 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9294 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9295 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9296 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9297 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9298 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9300 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9301 // shuffle. However, if we have AVX2 and either inputs are already in place,
9302 // we will be able to shuffle even across lanes the other input in a single
9303 // instruction so skip this pattern.
9304 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9305 isShuffleMaskInputInPlace(1, Mask))))
9306 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9307 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9310 // Otherwise fall back on generic blend lowering.
9311 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9315 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9317 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9318 /// isn't available.
9319 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9320 const X86Subtarget *Subtarget,
9321 SelectionDAG &DAG) {
9323 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9324 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9325 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9326 ArrayRef<int> Mask = SVOp->getMask();
9327 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9329 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9333 // Check for being able to broadcast a single element.
9334 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9335 Mask, Subtarget, DAG))
9338 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9339 // options to efficiently lower the shuffle.
9340 SmallVector<int, 4> RepeatedMask;
9341 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9342 assert(RepeatedMask.size() == 4 &&
9343 "Repeated masks must be half the mask width!");
9345 // Use even/odd duplicate instructions for masks that match their pattern.
9346 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9347 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9348 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9349 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9351 if (isSingleInputShuffleMask(Mask))
9352 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9353 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9355 // Use dedicated unpack instructions for masks that match their pattern.
9356 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9357 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9358 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9359 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9360 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9361 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9362 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9363 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9365 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9366 // have already handled any direct blends. We also need to squash the
9367 // repeated mask into a simulated v4f32 mask.
9368 for (int i = 0; i < 4; ++i)
9369 if (RepeatedMask[i] >= 8)
9370 RepeatedMask[i] -= 4;
9371 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9374 // If we have a single input shuffle with different shuffle patterns in the
9375 // two 128-bit lanes use the variable mask to VPERMILPS.
9376 if (isSingleInputShuffleMask(Mask)) {
9377 SDValue VPermMask[8];
9378 for (int i = 0; i < 8; ++i)
9379 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9380 : DAG.getConstant(Mask[i], MVT::i32);
9381 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
9383 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
9384 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
9386 if (Subtarget->hasAVX2())
9387 return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32,
9388 DAG.getNode(ISD::BITCAST, DL, MVT::v8f32,
9389 DAG.getNode(ISD::BUILD_VECTOR, DL,
9390 MVT::v8i32, VPermMask)),
9393 // Otherwise, fall back.
9394 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
9398 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9400 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9401 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
9404 // If we have AVX2 then we always want to lower with a blend because at v8 we
9405 // can fully permute the elements.
9406 if (Subtarget->hasAVX2())
9407 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
9410 // Otherwise fall back on generic lowering.
9411 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
9414 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
9416 /// This routine is only called when we have AVX2 and thus a reasonable
9417 /// instruction set for v8i32 shuffling..
9418 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9419 const X86Subtarget *Subtarget,
9420 SelectionDAG &DAG) {
9422 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9423 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
9424 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9425 ArrayRef<int> Mask = SVOp->getMask();
9426 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9427 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
9429 // Whenever we can lower this as a zext, that instruction is strictly faster
9430 // than any alternative. It also allows us to fold memory operands into the
9431 // shuffle in many cases.
9432 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
9433 Mask, Subtarget, DAG))
9436 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
9440 // Check for being able to broadcast a single element.
9441 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
9442 Mask, Subtarget, DAG))
9445 // If the shuffle mask is repeated in each 128-bit lane we can use more
9446 // efficient instructions that mirror the shuffles across the two 128-bit
9448 SmallVector<int, 4> RepeatedMask;
9449 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
9450 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
9451 if (isSingleInputShuffleMask(Mask))
9452 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
9453 getV4X86ShuffleImm8ForMask(RepeatedMask, DAG));
9455 // Use dedicated unpack instructions for masks that match their pattern.
9456 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9457 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
9458 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9459 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
9460 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9461 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
9462 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9463 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
9466 // Try to use shift instructions.
9468 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
9471 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9472 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9475 // If the shuffle patterns aren't repeated but it is a single input, directly
9476 // generate a cross-lane VPERMD instruction.
9477 if (isSingleInputShuffleMask(Mask)) {
9478 SDValue VPermMask[8];
9479 for (int i = 0; i < 8; ++i)
9480 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
9481 : DAG.getConstant(Mask[i], MVT::i32);
9483 X86ISD::VPERMV, DL, MVT::v8i32,
9484 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
9487 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9489 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9490 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
9493 // Otherwise fall back on generic blend lowering.
9494 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
9498 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
9500 /// This routine is only called when we have AVX2 and thus a reasonable
9501 /// instruction set for v16i16 shuffling..
9502 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9503 const X86Subtarget *Subtarget,
9504 SelectionDAG &DAG) {
9506 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9507 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
9508 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9509 ArrayRef<int> Mask = SVOp->getMask();
9510 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9511 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
9513 // Whenever we can lower this as a zext, that instruction is strictly faster
9514 // than any alternative. It also allows us to fold memory operands into the
9515 // shuffle in many cases.
9516 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
9517 Mask, Subtarget, DAG))
9520 // Check for being able to broadcast a single element.
9521 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
9522 Mask, Subtarget, DAG))
9525 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
9529 // Use dedicated unpack instructions for masks that match their pattern.
9530 if (isShuffleEquivalent(V1, V2, Mask,
9531 {// First 128-bit lane:
9532 0, 16, 1, 17, 2, 18, 3, 19,
9533 // Second 128-bit lane:
9534 8, 24, 9, 25, 10, 26, 11, 27}))
9535 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
9536 if (isShuffleEquivalent(V1, V2, Mask,
9537 {// First 128-bit lane:
9538 4, 20, 5, 21, 6, 22, 7, 23,
9539 // Second 128-bit lane:
9540 12, 28, 13, 29, 14, 30, 15, 31}))
9541 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
9543 // Try to use shift instructions.
9545 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
9548 // Try to use byte rotation instructions.
9549 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9550 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9553 if (isSingleInputShuffleMask(Mask)) {
9554 // There are no generalized cross-lane shuffle operations available on i16
9556 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
9557 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
9560 SDValue PSHUFBMask[32];
9561 for (int i = 0; i < 16; ++i) {
9562 if (Mask[i] == -1) {
9563 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
9567 int M = i < 8 ? Mask[i] : Mask[i] - 8;
9568 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
9569 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, MVT::i8);
9570 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, MVT::i8);
9573 ISD::BITCAST, DL, MVT::v16i16,
9575 X86ISD::PSHUFB, DL, MVT::v32i8,
9576 DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, V1),
9577 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask)));
9580 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9582 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9583 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
9586 // Otherwise fall back on generic lowering.
9587 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
9590 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
9592 /// This routine is only called when we have AVX2 and thus a reasonable
9593 /// instruction set for v32i8 shuffling..
9594 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9595 const X86Subtarget *Subtarget,
9596 SelectionDAG &DAG) {
9598 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9599 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
9600 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9601 ArrayRef<int> Mask = SVOp->getMask();
9602 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9603 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
9605 // Whenever we can lower this as a zext, that instruction is strictly faster
9606 // than any alternative. It also allows us to fold memory operands into the
9607 // shuffle in many cases.
9608 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
9609 Mask, Subtarget, DAG))
9612 // Check for being able to broadcast a single element.
9613 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
9614 Mask, Subtarget, DAG))
9617 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
9621 // Use dedicated unpack instructions for masks that match their pattern.
9622 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
9624 if (isShuffleEquivalent(
9626 {// First 128-bit lane:
9627 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
9628 // Second 128-bit lane:
9629 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
9630 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
9631 if (isShuffleEquivalent(
9633 {// First 128-bit lane:
9634 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
9635 // Second 128-bit lane:
9636 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
9637 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
9639 // Try to use shift instructions.
9641 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
9644 // Try to use byte rotation instructions.
9645 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9646 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9649 if (isSingleInputShuffleMask(Mask)) {
9650 // There are no generalized cross-lane shuffle operations available on i8
9652 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
9653 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
9656 SDValue PSHUFBMask[32];
9657 for (int i = 0; i < 32; ++i)
9660 ? DAG.getUNDEF(MVT::i8)
9661 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, MVT::i8);
9664 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
9665 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
9668 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9670 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9671 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
9674 // Otherwise fall back on generic lowering.
9675 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
9678 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
9680 /// This routine either breaks down the specific type of a 256-bit x86 vector
9681 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
9682 /// together based on the available instructions.
9683 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9684 MVT VT, const X86Subtarget *Subtarget,
9685 SelectionDAG &DAG) {
9687 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9688 ArrayRef<int> Mask = SVOp->getMask();
9690 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
9691 // check for those subtargets here and avoid much of the subtarget querying in
9692 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
9693 // ability to manipulate a 256-bit vector with integer types. Since we'll use
9694 // floating point types there eventually, just immediately cast everything to
9695 // a float and operate entirely in that domain.
9696 if (VT.isInteger() && !Subtarget->hasAVX2()) {
9697 int ElementBits = VT.getScalarSizeInBits();
9698 if (ElementBits < 32)
9699 // No floating point type available, decompose into 128-bit vectors.
9700 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9702 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
9703 VT.getVectorNumElements());
9704 V1 = DAG.getNode(ISD::BITCAST, DL, FpVT, V1);
9705 V2 = DAG.getNode(ISD::BITCAST, DL, FpVT, V2);
9706 return DAG.getNode(ISD::BITCAST, DL, VT,
9707 DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
9710 switch (VT.SimpleTy) {
9712 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9714 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9716 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9718 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9720 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9722 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9725 llvm_unreachable("Not a valid 256-bit x86 vector type!");
9729 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
9730 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9731 const X86Subtarget *Subtarget,
9732 SelectionDAG &DAG) {
9734 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9735 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
9736 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9737 ArrayRef<int> Mask = SVOp->getMask();
9738 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9740 // X86 has dedicated unpack instructions that can handle specific blend
9741 // operations: UNPCKH and UNPCKL.
9742 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9743 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
9744 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9745 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
9747 // FIXME: Implement direct support for this type!
9748 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
9751 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
9752 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9753 const X86Subtarget *Subtarget,
9754 SelectionDAG &DAG) {
9756 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9757 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
9758 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9759 ArrayRef<int> Mask = SVOp->getMask();
9760 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9762 // Use dedicated unpack instructions for masks that match their pattern.
9763 if (isShuffleEquivalent(V1, V2, Mask,
9764 {// First 128-bit lane.
9765 0, 16, 1, 17, 4, 20, 5, 21,
9766 // Second 128-bit lane.
9767 8, 24, 9, 25, 12, 28, 13, 29}))
9768 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
9769 if (isShuffleEquivalent(V1, V2, Mask,
9770 {// First 128-bit lane.
9771 2, 18, 3, 19, 6, 22, 7, 23,
9772 // Second 128-bit lane.
9773 10, 26, 11, 27, 14, 30, 15, 31}))
9774 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
9776 // FIXME: Implement direct support for this type!
9777 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
9780 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
9781 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9782 const X86Subtarget *Subtarget,
9783 SelectionDAG &DAG) {
9785 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9786 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
9787 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9788 ArrayRef<int> Mask = SVOp->getMask();
9789 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9791 // X86 has dedicated unpack instructions that can handle specific blend
9792 // operations: UNPCKH and UNPCKL.
9793 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
9794 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
9795 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
9796 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
9798 // FIXME: Implement direct support for this type!
9799 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
9802 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
9803 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9804 const X86Subtarget *Subtarget,
9805 SelectionDAG &DAG) {
9807 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9808 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
9809 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9810 ArrayRef<int> Mask = SVOp->getMask();
9811 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9813 // Use dedicated unpack instructions for masks that match their pattern.
9814 if (isShuffleEquivalent(V1, V2, Mask,
9815 {// First 128-bit lane.
9816 0, 16, 1, 17, 4, 20, 5, 21,
9817 // Second 128-bit lane.
9818 8, 24, 9, 25, 12, 28, 13, 29}))
9819 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
9820 if (isShuffleEquivalent(V1, V2, Mask,
9821 {// First 128-bit lane.
9822 2, 18, 3, 19, 6, 22, 7, 23,
9823 // Second 128-bit lane.
9824 10, 26, 11, 27, 14, 30, 15, 31}))
9825 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
9827 // FIXME: Implement direct support for this type!
9828 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
9831 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
9832 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9833 const X86Subtarget *Subtarget,
9834 SelectionDAG &DAG) {
9836 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9837 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
9838 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9839 ArrayRef<int> Mask = SVOp->getMask();
9840 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
9841 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
9843 // FIXME: Implement direct support for this type!
9844 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
9847 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
9848 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9849 const X86Subtarget *Subtarget,
9850 SelectionDAG &DAG) {
9852 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9853 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
9854 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9855 ArrayRef<int> Mask = SVOp->getMask();
9856 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
9857 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
9859 // FIXME: Implement direct support for this type!
9860 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
9863 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
9865 /// This routine either breaks down the specific type of a 512-bit x86 vector
9866 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
9867 /// together based on the available instructions.
9868 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9869 MVT VT, const X86Subtarget *Subtarget,
9870 SelectionDAG &DAG) {
9872 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9873 ArrayRef<int> Mask = SVOp->getMask();
9874 assert(Subtarget->hasAVX512() &&
9875 "Cannot lower 512-bit vectors w/ basic ISA!");
9877 // Check for being able to broadcast a single element.
9878 if (SDValue Broadcast =
9879 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
9882 // Dispatch to each element type for lowering. If we don't have supprot for
9883 // specific element type shuffles at 512 bits, immediately split them and
9884 // lower them. Each lowering routine of a given type is allowed to assume that
9885 // the requisite ISA extensions for that element type are available.
9886 switch (VT.SimpleTy) {
9888 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9890 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9892 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9894 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9896 if (Subtarget->hasBWI())
9897 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9900 if (Subtarget->hasBWI())
9901 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9905 llvm_unreachable("Not a valid 512-bit x86 vector type!");
9908 // Otherwise fall back on splitting.
9909 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9912 /// \brief Top-level lowering for x86 vector shuffles.
9914 /// This handles decomposition, canonicalization, and lowering of all x86
9915 /// vector shuffles. Most of the specific lowering strategies are encapsulated
9916 /// above in helper routines. The canonicalization attempts to widen shuffles
9917 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
9918 /// s.t. only one of the two inputs needs to be tested, etc.
9919 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
9920 SelectionDAG &DAG) {
9921 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9922 ArrayRef<int> Mask = SVOp->getMask();
9923 SDValue V1 = Op.getOperand(0);
9924 SDValue V2 = Op.getOperand(1);
9925 MVT VT = Op.getSimpleValueType();
9926 int NumElements = VT.getVectorNumElements();
9929 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
9931 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
9932 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
9933 if (V1IsUndef && V2IsUndef)
9934 return DAG.getUNDEF(VT);
9936 // When we create a shuffle node we put the UNDEF node to second operand,
9937 // but in some cases the first operand may be transformed to UNDEF.
9938 // In this case we should just commute the node.
9940 return DAG.getCommutedVectorShuffle(*SVOp);
9942 // Check for non-undef masks pointing at an undef vector and make the masks
9943 // undef as well. This makes it easier to match the shuffle based solely on
9947 if (M >= NumElements) {
9948 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
9949 for (int &M : NewMask)
9950 if (M >= NumElements)
9952 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
9955 // We actually see shuffles that are entirely re-arrangements of a set of
9956 // zero inputs. This mostly happens while decomposing complex shuffles into
9957 // simple ones. Directly lower these as a buildvector of zeros.
9958 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9960 return getZeroVector(VT, Subtarget, DAG, dl);
9962 // Try to collapse shuffles into using a vector type with fewer elements but
9963 // wider element types. We cap this to not form integers or floating point
9964 // elements wider than 64 bits, but it might be interesting to form i128
9965 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
9966 SmallVector<int, 16> WidenedMask;
9967 if (VT.getScalarSizeInBits() < 64 &&
9968 canWidenShuffleElements(Mask, WidenedMask)) {
9969 MVT NewEltVT = VT.isFloatingPoint()
9970 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
9971 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
9972 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
9973 // Make sure that the new vector type is legal. For example, v2f64 isn't
9975 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
9976 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
9977 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
9978 return DAG.getNode(ISD::BITCAST, dl, VT,
9979 DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
9983 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
9984 for (int M : SVOp->getMask())
9987 else if (M < NumElements)
9992 // Commute the shuffle as needed such that more elements come from V1 than
9993 // V2. This allows us to match the shuffle pattern strictly on how many
9994 // elements come from V1 without handling the symmetric cases.
9995 if (NumV2Elements > NumV1Elements)
9996 return DAG.getCommutedVectorShuffle(*SVOp);
9998 // When the number of V1 and V2 elements are the same, try to minimize the
9999 // number of uses of V2 in the low half of the vector. When that is tied,
10000 // ensure that the sum of indices for V1 is equal to or lower than the sum
10001 // indices for V2. When those are equal, try to ensure that the number of odd
10002 // indices for V1 is lower than the number of odd indices for V2.
10003 if (NumV1Elements == NumV2Elements) {
10004 int LowV1Elements = 0, LowV2Elements = 0;
10005 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10006 if (M >= NumElements)
10010 if (LowV2Elements > LowV1Elements) {
10011 return DAG.getCommutedVectorShuffle(*SVOp);
10012 } else if (LowV2Elements == LowV1Elements) {
10013 int SumV1Indices = 0, SumV2Indices = 0;
10014 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10015 if (SVOp->getMask()[i] >= NumElements)
10017 else if (SVOp->getMask()[i] >= 0)
10019 if (SumV2Indices < SumV1Indices) {
10020 return DAG.getCommutedVectorShuffle(*SVOp);
10021 } else if (SumV2Indices == SumV1Indices) {
10022 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10023 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10024 if (SVOp->getMask()[i] >= NumElements)
10025 NumV2OddIndices += i % 2;
10026 else if (SVOp->getMask()[i] >= 0)
10027 NumV1OddIndices += i % 2;
10028 if (NumV2OddIndices < NumV1OddIndices)
10029 return DAG.getCommutedVectorShuffle(*SVOp);
10034 // For each vector width, delegate to a specialized lowering routine.
10035 if (VT.getSizeInBits() == 128)
10036 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10038 if (VT.getSizeInBits() == 256)
10039 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10041 // Force AVX-512 vectors to be scalarized for now.
10042 // FIXME: Implement AVX-512 support!
10043 if (VT.getSizeInBits() == 512)
10044 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10046 llvm_unreachable("Unimplemented!");
10049 // This function assumes its argument is a BUILD_VECTOR of constants or
10050 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10052 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10053 unsigned &MaskValue) {
10055 unsigned NumElems = BuildVector->getNumOperands();
10056 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10057 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10058 unsigned NumElemsInLane = NumElems / NumLanes;
10060 // Blend for v16i16 should be symetric for the both lanes.
10061 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10062 SDValue EltCond = BuildVector->getOperand(i);
10063 SDValue SndLaneEltCond =
10064 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10066 int Lane1Cond = -1, Lane2Cond = -1;
10067 if (isa<ConstantSDNode>(EltCond))
10068 Lane1Cond = !isZero(EltCond);
10069 if (isa<ConstantSDNode>(SndLaneEltCond))
10070 Lane2Cond = !isZero(SndLaneEltCond);
10072 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10073 // Lane1Cond != 0, means we want the first argument.
10074 // Lane1Cond == 0, means we want the second argument.
10075 // The encoding of this argument is 0 for the first argument, 1
10076 // for the second. Therefore, invert the condition.
10077 MaskValue |= !Lane1Cond << i;
10078 else if (Lane1Cond < 0)
10079 MaskValue |= !Lane2Cond << i;
10086 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10087 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10088 const X86Subtarget *Subtarget,
10089 SelectionDAG &DAG) {
10090 SDValue Cond = Op.getOperand(0);
10091 SDValue LHS = Op.getOperand(1);
10092 SDValue RHS = Op.getOperand(2);
10094 MVT VT = Op.getSimpleValueType();
10096 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10098 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10100 // Only non-legal VSELECTs reach this lowering, convert those into generic
10101 // shuffles and re-use the shuffle lowering path for blends.
10102 SmallVector<int, 32> Mask;
10103 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10104 SDValue CondElt = CondBV->getOperand(i);
10106 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10108 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10111 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10112 // A vselect where all conditions and data are constants can be optimized into
10113 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10114 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10115 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10116 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10119 // Try to lower this to a blend-style vector shuffle. This can handle all
10120 // constant condition cases.
10121 SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG);
10122 if (BlendOp.getNode())
10125 // Variable blends are only legal from SSE4.1 onward.
10126 if (!Subtarget->hasSSE41())
10129 // Some types for vselect were previously set to Expand, not Legal or
10130 // Custom. Return an empty SDValue so we fall-through to Expand, after
10131 // the Custom lowering phase.
10132 MVT VT = Op.getSimpleValueType();
10133 switch (VT.SimpleTy) {
10138 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10143 // We couldn't create a "Blend with immediate" node.
10144 // This node should still be legal, but we'll have to emit a blendv*
10149 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10150 MVT VT = Op.getSimpleValueType();
10153 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10156 if (VT.getSizeInBits() == 8) {
10157 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10158 Op.getOperand(0), Op.getOperand(1));
10159 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10160 DAG.getValueType(VT));
10161 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10164 if (VT.getSizeInBits() == 16) {
10165 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10166 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10168 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10169 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10170 DAG.getNode(ISD::BITCAST, dl,
10173 Op.getOperand(1)));
10174 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10175 Op.getOperand(0), Op.getOperand(1));
10176 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10177 DAG.getValueType(VT));
10178 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10181 if (VT == MVT::f32) {
10182 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10183 // the result back to FR32 register. It's only worth matching if the
10184 // result has a single use which is a store or a bitcast to i32. And in
10185 // the case of a store, it's not worth it if the index is a constant 0,
10186 // because a MOVSSmr can be used instead, which is smaller and faster.
10187 if (!Op.hasOneUse())
10189 SDNode *User = *Op.getNode()->use_begin();
10190 if ((User->getOpcode() != ISD::STORE ||
10191 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10192 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10193 (User->getOpcode() != ISD::BITCAST ||
10194 User->getValueType(0) != MVT::i32))
10196 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10197 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
10200 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
10203 if (VT == MVT::i32 || VT == MVT::i64) {
10204 // ExtractPS/pextrq works with constant index.
10205 if (isa<ConstantSDNode>(Op.getOperand(1)))
10211 /// Extract one bit from mask vector, like v16i1 or v8i1.
10212 /// AVX-512 feature.
10214 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10215 SDValue Vec = Op.getOperand(0);
10217 MVT VecVT = Vec.getSimpleValueType();
10218 SDValue Idx = Op.getOperand(1);
10219 MVT EltVT = Op.getSimpleValueType();
10221 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10222 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10223 "Unexpected vector type in ExtractBitFromMaskVector");
10225 // variable index can't be handled in mask registers,
10226 // extend vector to VR512
10227 if (!isa<ConstantSDNode>(Idx)) {
10228 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10229 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10230 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10231 ExtVT.getVectorElementType(), Ext, Idx);
10232 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10235 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10236 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10237 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10238 rc = getRegClassFor(MVT::v16i1);
10239 unsigned MaxSift = rc->getSize()*8 - 1;
10240 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10241 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10242 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10243 DAG.getConstant(MaxSift, MVT::i8));
10244 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10245 DAG.getIntPtrConstant(0));
10249 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10250 SelectionDAG &DAG) const {
10252 SDValue Vec = Op.getOperand(0);
10253 MVT VecVT = Vec.getSimpleValueType();
10254 SDValue Idx = Op.getOperand(1);
10256 if (Op.getSimpleValueType() == MVT::i1)
10257 return ExtractBitFromMaskVector(Op, DAG);
10259 if (!isa<ConstantSDNode>(Idx)) {
10260 if (VecVT.is512BitVector() ||
10261 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10262 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10265 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10266 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10267 MaskEltVT.getSizeInBits());
10269 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10270 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10271 getZeroVector(MaskVT, Subtarget, DAG, dl),
10272 Idx, DAG.getConstant(0, getPointerTy()));
10273 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10274 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(),
10275 Perm, DAG.getConstant(0, getPointerTy()));
10280 // If this is a 256-bit vector result, first extract the 128-bit vector and
10281 // then extract the element from the 128-bit vector.
10282 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10284 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10285 // Get the 128-bit vector.
10286 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10287 MVT EltVT = VecVT.getVectorElementType();
10289 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10291 //if (IdxVal >= NumElems/2)
10292 // IdxVal -= NumElems/2;
10293 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10294 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10295 DAG.getConstant(IdxVal, MVT::i32));
10298 assert(VecVT.is128BitVector() && "Unexpected vector length");
10300 if (Subtarget->hasSSE41()) {
10301 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
10306 MVT VT = Op.getSimpleValueType();
10307 // TODO: handle v16i8.
10308 if (VT.getSizeInBits() == 16) {
10309 SDValue Vec = Op.getOperand(0);
10310 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10312 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10313 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10314 DAG.getNode(ISD::BITCAST, dl,
10316 Op.getOperand(1)));
10317 // Transform it so it match pextrw which produces a 32-bit result.
10318 MVT EltVT = MVT::i32;
10319 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10320 Op.getOperand(0), Op.getOperand(1));
10321 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10322 DAG.getValueType(VT));
10323 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10326 if (VT.getSizeInBits() == 32) {
10327 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10331 // SHUFPS the element to the lowest double word, then movss.
10332 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10333 MVT VVT = Op.getOperand(0).getSimpleValueType();
10334 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10335 DAG.getUNDEF(VVT), Mask);
10336 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10337 DAG.getIntPtrConstant(0));
10340 if (VT.getSizeInBits() == 64) {
10341 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10342 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10343 // to match extract_elt for f64.
10344 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10348 // UNPCKHPD the element to the lowest double word, then movsd.
10349 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10350 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10351 int Mask[2] = { 1, -1 };
10352 MVT VVT = Op.getOperand(0).getSimpleValueType();
10353 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10354 DAG.getUNDEF(VVT), Mask);
10355 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10356 DAG.getIntPtrConstant(0));
10362 /// Insert one bit to mask vector, like v16i1 or v8i1.
10363 /// AVX-512 feature.
10365 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
10367 SDValue Vec = Op.getOperand(0);
10368 SDValue Elt = Op.getOperand(1);
10369 SDValue Idx = Op.getOperand(2);
10370 MVT VecVT = Vec.getSimpleValueType();
10372 if (!isa<ConstantSDNode>(Idx)) {
10373 // Non constant index. Extend source and destination,
10374 // insert element and then truncate the result.
10375 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10376 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
10377 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
10378 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
10379 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
10380 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
10383 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10384 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
10385 if (Vec.getOpcode() == ISD::UNDEF)
10386 return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10387 DAG.getConstant(IdxVal, MVT::i8));
10388 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10389 unsigned MaxSift = rc->getSize()*8 - 1;
10390 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
10391 DAG.getConstant(MaxSift, MVT::i8));
10392 EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
10393 DAG.getConstant(MaxSift - IdxVal, MVT::i8));
10394 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
10397 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
10398 SelectionDAG &DAG) const {
10399 MVT VT = Op.getSimpleValueType();
10400 MVT EltVT = VT.getVectorElementType();
10402 if (EltVT == MVT::i1)
10403 return InsertBitToMaskVector(Op, DAG);
10406 SDValue N0 = Op.getOperand(0);
10407 SDValue N1 = Op.getOperand(1);
10408 SDValue N2 = Op.getOperand(2);
10409 if (!isa<ConstantSDNode>(N2))
10411 auto *N2C = cast<ConstantSDNode>(N2);
10412 unsigned IdxVal = N2C->getZExtValue();
10414 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
10415 // into that, and then insert the subvector back into the result.
10416 if (VT.is256BitVector() || VT.is512BitVector()) {
10417 // Get the desired 128-bit vector half.
10418 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
10420 // Insert the element into the desired half.
10421 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
10422 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
10424 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
10425 DAG.getConstant(IdxIn128, MVT::i32));
10427 // Insert the changed part back to the 256-bit vector
10428 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
10430 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
10432 if (Subtarget->hasSSE41()) {
10433 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
10435 if (VT == MVT::v8i16) {
10436 Opc = X86ISD::PINSRW;
10438 assert(VT == MVT::v16i8);
10439 Opc = X86ISD::PINSRB;
10442 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
10444 if (N1.getValueType() != MVT::i32)
10445 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10446 if (N2.getValueType() != MVT::i32)
10447 N2 = DAG.getIntPtrConstant(IdxVal);
10448 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
10451 if (EltVT == MVT::f32) {
10452 // Bits [7:6] of the constant are the source select. This will always be
10453 // zero here. The DAG Combiner may combine an extract_elt index into
10455 // bits. For example (insert (extract, 3), 2) could be matched by
10457 // the '3' into bits [7:6] of X86ISD::INSERTPS.
10458 // Bits [5:4] of the constant are the destination select. This is the
10459 // value of the incoming immediate.
10460 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
10461 // combine either bitwise AND or insert of float 0.0 to set these bits.
10462 N2 = DAG.getIntPtrConstant(IdxVal << 4);
10463 // Create this as a scalar to vector..
10464 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
10465 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
10468 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
10469 // PINSR* works with constant index.
10474 if (EltVT == MVT::i8)
10477 if (EltVT.getSizeInBits() == 16) {
10478 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
10479 // as its second argument.
10480 if (N1.getValueType() != MVT::i32)
10481 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
10482 if (N2.getValueType() != MVT::i32)
10483 N2 = DAG.getIntPtrConstant(IdxVal);
10484 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
10489 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
10491 MVT OpVT = Op.getSimpleValueType();
10493 // If this is a 256-bit vector result, first insert into a 128-bit
10494 // vector and then insert into the 256-bit vector.
10495 if (!OpVT.is128BitVector()) {
10496 // Insert into a 128-bit vector.
10497 unsigned SizeFactor = OpVT.getSizeInBits()/128;
10498 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
10499 OpVT.getVectorNumElements() / SizeFactor);
10501 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
10503 // Insert the 128-bit vector.
10504 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
10507 if (OpVT == MVT::v1i64 &&
10508 Op.getOperand(0).getValueType() == MVT::i64)
10509 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
10511 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
10512 assert(OpVT.is128BitVector() && "Expected an SSE type!");
10513 return DAG.getNode(ISD::BITCAST, dl, OpVT,
10514 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
10517 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
10518 // a simple subregister reference or explicit instructions to grab
10519 // upper bits of a vector.
10520 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10521 SelectionDAG &DAG) {
10523 SDValue In = Op.getOperand(0);
10524 SDValue Idx = Op.getOperand(1);
10525 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10526 MVT ResVT = Op.getSimpleValueType();
10527 MVT InVT = In.getSimpleValueType();
10529 if (Subtarget->hasFp256()) {
10530 if (ResVT.is128BitVector() &&
10531 (InVT.is256BitVector() || InVT.is512BitVector()) &&
10532 isa<ConstantSDNode>(Idx)) {
10533 return Extract128BitVector(In, IdxVal, DAG, dl);
10535 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
10536 isa<ConstantSDNode>(Idx)) {
10537 return Extract256BitVector(In, IdxVal, DAG, dl);
10543 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
10544 // simple superregister reference or explicit instructions to insert
10545 // the upper bits of a vector.
10546 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
10547 SelectionDAG &DAG) {
10548 if (!Subtarget->hasAVX())
10552 SDValue Vec = Op.getOperand(0);
10553 SDValue SubVec = Op.getOperand(1);
10554 SDValue Idx = Op.getOperand(2);
10556 if (!isa<ConstantSDNode>(Idx))
10559 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10560 MVT OpVT = Op.getSimpleValueType();
10561 MVT SubVecVT = SubVec.getSimpleValueType();
10563 // Fold two 16-byte subvector loads into one 32-byte load:
10564 // (insert_subvector (insert_subvector undef, (load addr), 0),
10565 // (load addr + 16), Elts/2)
10567 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
10568 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
10569 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
10570 !Subtarget->isUnalignedMem32Slow()) {
10571 SDValue SubVec2 = Vec.getOperand(1);
10572 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
10573 if (Idx2->getZExtValue() == 0) {
10574 SDValue Ops[] = { SubVec2, SubVec };
10575 SDValue LD = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false);
10582 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
10583 SubVecVT.is128BitVector())
10584 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
10586 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
10587 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
10592 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
10593 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
10594 // one of the above mentioned nodes. It has to be wrapped because otherwise
10595 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
10596 // be used to form addressing mode. These wrapped nodes will be selected
10599 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
10600 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
10602 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10603 // global base reg.
10604 unsigned char OpFlag = 0;
10605 unsigned WrapperKind = X86ISD::Wrapper;
10606 CodeModel::Model M = DAG.getTarget().getCodeModel();
10608 if (Subtarget->isPICStyleRIPRel() &&
10609 (M == CodeModel::Small || M == CodeModel::Kernel))
10610 WrapperKind = X86ISD::WrapperRIP;
10611 else if (Subtarget->isPICStyleGOT())
10612 OpFlag = X86II::MO_GOTOFF;
10613 else if (Subtarget->isPICStyleStubPIC())
10614 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10616 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
10617 CP->getAlignment(),
10618 CP->getOffset(), OpFlag);
10620 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10621 // With PIC, the address is actually $g + Offset.
10623 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10624 DAG.getNode(X86ISD::GlobalBaseReg,
10625 SDLoc(), getPointerTy()),
10632 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
10633 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
10635 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10636 // global base reg.
10637 unsigned char OpFlag = 0;
10638 unsigned WrapperKind = X86ISD::Wrapper;
10639 CodeModel::Model M = DAG.getTarget().getCodeModel();
10641 if (Subtarget->isPICStyleRIPRel() &&
10642 (M == CodeModel::Small || M == CodeModel::Kernel))
10643 WrapperKind = X86ISD::WrapperRIP;
10644 else if (Subtarget->isPICStyleGOT())
10645 OpFlag = X86II::MO_GOTOFF;
10646 else if (Subtarget->isPICStyleStubPIC())
10647 OpFlag = X86II::MO_PIC_BASE_OFFSET;
10649 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
10652 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10654 // With PIC, the address is actually $g + Offset.
10656 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10657 DAG.getNode(X86ISD::GlobalBaseReg,
10658 SDLoc(), getPointerTy()),
10665 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
10666 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
10668 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10669 // global base reg.
10670 unsigned char OpFlag = 0;
10671 unsigned WrapperKind = X86ISD::Wrapper;
10672 CodeModel::Model M = DAG.getTarget().getCodeModel();
10674 if (Subtarget->isPICStyleRIPRel() &&
10675 (M == CodeModel::Small || M == CodeModel::Kernel)) {
10676 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
10677 OpFlag = X86II::MO_GOTPCREL;
10678 WrapperKind = X86ISD::WrapperRIP;
10679 } else if (Subtarget->isPICStyleGOT()) {
10680 OpFlag = X86II::MO_GOT;
10681 } else if (Subtarget->isPICStyleStubPIC()) {
10682 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
10683 } else if (Subtarget->isPICStyleStubNoDynamic()) {
10684 OpFlag = X86II::MO_DARWIN_NONLAZY;
10687 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
10690 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10692 // With PIC, the address is actually $g + Offset.
10693 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
10694 !Subtarget->is64Bit()) {
10695 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10696 DAG.getNode(X86ISD::GlobalBaseReg,
10697 SDLoc(), getPointerTy()),
10701 // For symbols that require a load from a stub to get the address, emit the
10703 if (isGlobalStubReference(OpFlag))
10704 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
10705 MachinePointerInfo::getGOT(), false, false, false, 0);
10711 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
10712 // Create the TargetBlockAddressAddress node.
10713 unsigned char OpFlags =
10714 Subtarget->ClassifyBlockAddressReference();
10715 CodeModel::Model M = DAG.getTarget().getCodeModel();
10716 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
10717 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
10719 SDValue Result = DAG.getTargetBlockAddress(BA, getPointerTy(), Offset,
10722 if (Subtarget->isPICStyleRIPRel() &&
10723 (M == CodeModel::Small || M == CodeModel::Kernel))
10724 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10726 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10728 // With PIC, the address is actually $g + Offset.
10729 if (isGlobalRelativeToPICBase(OpFlags)) {
10730 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10731 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10739 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
10740 int64_t Offset, SelectionDAG &DAG) const {
10741 // Create the TargetGlobalAddress node, folding in the constant
10742 // offset if it is legal.
10743 unsigned char OpFlags =
10744 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
10745 CodeModel::Model M = DAG.getTarget().getCodeModel();
10747 if (OpFlags == X86II::MO_NO_FLAG &&
10748 X86::isOffsetSuitableForCodeModel(Offset, M)) {
10749 // A direct static reference to a global.
10750 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
10753 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
10756 if (Subtarget->isPICStyleRIPRel() &&
10757 (M == CodeModel::Small || M == CodeModel::Kernel))
10758 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
10760 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
10762 // With PIC, the address is actually $g + Offset.
10763 if (isGlobalRelativeToPICBase(OpFlags)) {
10764 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
10765 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
10769 // For globals that require a load from a stub to get the address, emit the
10771 if (isGlobalStubReference(OpFlags))
10772 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
10773 MachinePointerInfo::getGOT(), false, false, false, 0);
10775 // If there was a non-zero offset that we didn't fold, create an explicit
10776 // addition for it.
10778 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
10779 DAG.getConstant(Offset, getPointerTy()));
10785 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
10786 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
10787 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
10788 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
10792 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
10793 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
10794 unsigned char OperandFlags, bool LocalDynamic = false) {
10795 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
10796 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10798 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10799 GA->getValueType(0),
10803 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
10807 SDValue Ops[] = { Chain, TGA, *InFlag };
10808 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10810 SDValue Ops[] = { Chain, TGA };
10811 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
10814 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
10815 MFI->setAdjustsStack(true);
10816 MFI->setHasCalls(true);
10818 SDValue Flag = Chain.getValue(1);
10819 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
10822 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
10824 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10827 SDLoc dl(GA); // ? function entry point might be better
10828 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10829 DAG.getNode(X86ISD::GlobalBaseReg,
10830 SDLoc(), PtrVT), InFlag);
10831 InFlag = Chain.getValue(1);
10833 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
10836 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
10838 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10840 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
10841 X86::RAX, X86II::MO_TLSGD);
10844 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
10850 // Get the start address of the TLS block for this module.
10851 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
10852 .getInfo<X86MachineFunctionInfo>();
10853 MFI->incNumLocalDynamicTLSAccesses();
10857 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
10858 X86II::MO_TLSLD, /*LocalDynamic=*/true);
10861 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
10862 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
10863 InFlag = Chain.getValue(1);
10864 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
10865 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
10868 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
10872 unsigned char OperandFlags = X86II::MO_DTPOFF;
10873 unsigned WrapperKind = X86ISD::Wrapper;
10874 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
10875 GA->getValueType(0),
10876 GA->getOffset(), OperandFlags);
10877 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10879 // Add x@dtpoff with the base.
10880 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
10883 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
10884 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
10885 const EVT PtrVT, TLSModel::Model model,
10886 bool is64Bit, bool isPIC) {
10889 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
10890 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
10891 is64Bit ? 257 : 256));
10893 SDValue ThreadPointer =
10894 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0),
10895 MachinePointerInfo(Ptr), false, false, false, 0);
10897 unsigned char OperandFlags = 0;
10898 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
10900 unsigned WrapperKind = X86ISD::Wrapper;
10901 if (model == TLSModel::LocalExec) {
10902 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
10903 } else if (model == TLSModel::InitialExec) {
10905 OperandFlags = X86II::MO_GOTTPOFF;
10906 WrapperKind = X86ISD::WrapperRIP;
10908 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
10911 llvm_unreachable("Unexpected model");
10914 // emit "addl x@ntpoff,%eax" (local exec)
10915 // or "addl x@indntpoff,%eax" (initial exec)
10916 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
10918 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
10919 GA->getOffset(), OperandFlags);
10920 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
10922 if (model == TLSModel::InitialExec) {
10923 if (isPIC && !is64Bit) {
10924 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
10925 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
10929 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
10930 MachinePointerInfo::getGOT(), false, false, false, 0);
10933 // The address of the thread local variable is the add of the thread
10934 // pointer with the offset of the variable.
10935 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
10939 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
10941 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
10942 const GlobalValue *GV = GA->getGlobal();
10944 if (Subtarget->isTargetELF()) {
10945 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
10948 case TLSModel::GeneralDynamic:
10949 if (Subtarget->is64Bit())
10950 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
10951 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
10952 case TLSModel::LocalDynamic:
10953 return LowerToTLSLocalDynamicModel(GA, DAG, getPointerTy(),
10954 Subtarget->is64Bit());
10955 case TLSModel::InitialExec:
10956 case TLSModel::LocalExec:
10957 return LowerToTLSExecModel(
10958 GA, DAG, getPointerTy(), model, Subtarget->is64Bit(),
10959 DAG.getTarget().getRelocationModel() == Reloc::PIC_);
10961 llvm_unreachable("Unknown TLS model.");
10964 if (Subtarget->isTargetDarwin()) {
10965 // Darwin only has one model of TLS. Lower to that.
10966 unsigned char OpFlag = 0;
10967 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
10968 X86ISD::WrapperRIP : X86ISD::Wrapper;
10970 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
10971 // global base reg.
10972 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
10973 !Subtarget->is64Bit();
10975 OpFlag = X86II::MO_TLVP_PIC_BASE;
10977 OpFlag = X86II::MO_TLVP;
10979 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
10980 GA->getValueType(0),
10981 GA->getOffset(), OpFlag);
10982 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
10984 // With PIC32, the address is actually $g + Offset.
10986 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
10987 DAG.getNode(X86ISD::GlobalBaseReg,
10988 SDLoc(), getPointerTy()),
10991 // Lowering the machine isd will make sure everything is in the right
10993 SDValue Chain = DAG.getEntryNode();
10994 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
10995 SDValue Args[] = { Chain, Offset };
10996 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
10998 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
10999 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11000 MFI->setAdjustsStack(true);
11002 // And our return value (tls address) is in the standard call return value
11004 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11005 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy(),
11006 Chain.getValue(1));
11009 if (Subtarget->isTargetKnownWindowsMSVC() ||
11010 Subtarget->isTargetWindowsGNU()) {
11011 // Just use the implicit TLS architecture
11012 // Need to generate someting similar to:
11013 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11015 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11016 // mov rcx, qword [rdx+rcx*8]
11017 // mov eax, .tls$:tlsvar
11018 // [rax+rcx] contains the address
11019 // Windows 64bit: gs:0x58
11020 // Windows 32bit: fs:__tls_array
11023 SDValue Chain = DAG.getEntryNode();
11025 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11026 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11027 // use its literal value of 0x2C.
11028 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11029 ? Type::getInt8PtrTy(*DAG.getContext(),
11031 : Type::getInt32PtrTy(*DAG.getContext(),
11035 Subtarget->is64Bit()
11036 ? DAG.getIntPtrConstant(0x58)
11037 : (Subtarget->isTargetWindowsGNU()
11038 ? DAG.getIntPtrConstant(0x2C)
11039 : DAG.getExternalSymbol("_tls_array", getPointerTy()));
11041 SDValue ThreadPointer =
11042 DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
11043 MachinePointerInfo(Ptr), false, false, false, 0);
11045 // Load the _tls_index variable
11046 SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
11047 if (Subtarget->is64Bit())
11048 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
11049 IDX, MachinePointerInfo(), MVT::i32,
11050 false, false, false, 0);
11052 IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
11053 false, false, false, 0);
11055 SDValue Scale = DAG.getConstant(Log2_64_Ceil(TD->getPointerSize()),
11057 IDX = DAG.getNode(ISD::SHL, dl, getPointerTy(), IDX, Scale);
11059 SDValue res = DAG.getNode(ISD::ADD, dl, getPointerTy(), ThreadPointer, IDX);
11060 res = DAG.getLoad(getPointerTy(), dl, Chain, res, MachinePointerInfo(),
11061 false, false, false, 0);
11063 // Get the offset of start of .tls section
11064 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11065 GA->getValueType(0),
11066 GA->getOffset(), X86II::MO_SECREL);
11067 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), TGA);
11069 // The address of the thread local variable is the add of the thread
11070 // pointer with the offset of the variable.
11071 return DAG.getNode(ISD::ADD, dl, getPointerTy(), res, Offset);
11074 llvm_unreachable("TLS not implemented for this target.");
11077 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11078 /// and take a 2 x i32 value to shift plus a shift amount.
11079 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11080 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11081 MVT VT = Op.getSimpleValueType();
11082 unsigned VTBits = VT.getSizeInBits();
11084 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11085 SDValue ShOpLo = Op.getOperand(0);
11086 SDValue ShOpHi = Op.getOperand(1);
11087 SDValue ShAmt = Op.getOperand(2);
11088 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11089 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11091 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11092 DAG.getConstant(VTBits - 1, MVT::i8));
11093 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11094 DAG.getConstant(VTBits - 1, MVT::i8))
11095 : DAG.getConstant(0, VT);
11097 SDValue Tmp2, Tmp3;
11098 if (Op.getOpcode() == ISD::SHL_PARTS) {
11099 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11100 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11102 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11103 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11106 // If the shift amount is larger or equal than the width of a part we can't
11107 // rely on the results of shld/shrd. Insert a test and select the appropriate
11108 // values for large shift amounts.
11109 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11110 DAG.getConstant(VTBits, MVT::i8));
11111 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11112 AndNode, DAG.getConstant(0, MVT::i8));
11115 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
11116 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11117 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11119 if (Op.getOpcode() == ISD::SHL_PARTS) {
11120 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11121 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11123 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11124 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11127 SDValue Ops[2] = { Lo, Hi };
11128 return DAG.getMergeValues(Ops, dl);
11131 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11132 SelectionDAG &DAG) const {
11133 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
11136 if (SrcVT.isVector()) {
11137 if (SrcVT.getVectorElementType() == MVT::i1) {
11138 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11139 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11140 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT,
11141 Op.getOperand(0)));
11146 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11147 "Unknown SINT_TO_FP to lower!");
11149 // These are really Legal; return the operand so the caller accepts it as
11151 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11153 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11154 Subtarget->is64Bit()) {
11158 unsigned Size = SrcVT.getSizeInBits()/8;
11159 MachineFunction &MF = DAG.getMachineFunction();
11160 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11161 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11162 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11164 MachinePointerInfo::getFixedStack(SSFI),
11166 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11169 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11171 SelectionDAG &DAG) const {
11175 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11177 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11179 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11181 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11183 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11184 MachineMemOperand *MMO;
11186 int SSFI = FI->getIndex();
11188 DAG.getMachineFunction()
11189 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11190 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11192 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11193 StackSlot = StackSlot.getOperand(1);
11195 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11196 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11198 Tys, Ops, SrcVT, MMO);
11201 Chain = Result.getValue(1);
11202 SDValue InFlag = Result.getValue(2);
11204 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11205 // shouldn't be necessary except that RFP cannot be live across
11206 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11207 MachineFunction &MF = DAG.getMachineFunction();
11208 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11209 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11210 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11211 Tys = DAG.getVTList(MVT::Other);
11213 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11215 MachineMemOperand *MMO =
11216 DAG.getMachineFunction()
11217 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11218 MachineMemOperand::MOStore, SSFISize, SSFISize);
11220 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11221 Ops, Op.getValueType(), MMO);
11222 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11223 MachinePointerInfo::getFixedStack(SSFI),
11224 false, false, false, 0);
11230 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11231 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11232 SelectionDAG &DAG) const {
11233 // This algorithm is not obvious. Here it is what we're trying to output:
11236 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11237 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11239 haddpd %xmm0, %xmm0
11241 pshufd $0x4e, %xmm0, %xmm1
11247 LLVMContext *Context = DAG.getContext();
11249 // Build some magic constants.
11250 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11251 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11252 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
11254 SmallVector<Constant*,2> CV1;
11256 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11257 APInt(64, 0x4330000000000000ULL))));
11259 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11260 APInt(64, 0x4530000000000000ULL))));
11261 Constant *C1 = ConstantVector::get(CV1);
11262 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
11264 // Load the 64-bit value into an XMM register.
11265 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11267 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11268 MachinePointerInfo::getConstantPool(),
11269 false, false, false, 16);
11270 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32,
11271 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, XR1),
11274 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11275 MachinePointerInfo::getConstantPool(),
11276 false, false, false, 16);
11277 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck1);
11278 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11281 if (Subtarget->hasSSE3()) {
11282 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11283 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11285 SDValue S2F = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Sub);
11286 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11288 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11289 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Shuffle),
11293 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11294 DAG.getIntPtrConstant(0));
11297 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
11298 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
11299 SelectionDAG &DAG) const {
11301 // FP constant to bias correct the final result.
11302 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
11305 // Load the 32-bit value into an XMM register.
11306 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
11309 // Zero out the upper parts of the register.
11310 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
11312 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11313 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
11314 DAG.getIntPtrConstant(0));
11316 // Or the load with the bias.
11317 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
11318 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11319 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11320 MVT::v2f64, Load)),
11321 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
11322 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
11323 MVT::v2f64, Bias)));
11324 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
11325 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
11326 DAG.getIntPtrConstant(0));
11328 // Subtract the bias.
11329 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
11331 // Handle final rounding.
11332 EVT DestVT = Op.getValueType();
11334 if (DestVT.bitsLT(MVT::f64))
11335 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
11336 DAG.getIntPtrConstant(0));
11337 if (DestVT.bitsGT(MVT::f64))
11338 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
11340 // Handle final rounding.
11344 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
11345 const X86Subtarget &Subtarget) {
11346 // The algorithm is the following:
11347 // #ifdef __SSE4_1__
11348 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11349 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11350 // (uint4) 0x53000000, 0xaa);
11352 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11353 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11355 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11356 // return (float4) lo + fhi;
11359 SDValue V = Op->getOperand(0);
11360 EVT VecIntVT = V.getValueType();
11361 bool Is128 = VecIntVT == MVT::v4i32;
11362 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
11363 // If we convert to something else than the supported type, e.g., to v4f64,
11365 if (VecFloatVT != Op->getValueType(0))
11368 unsigned NumElts = VecIntVT.getVectorNumElements();
11369 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
11370 "Unsupported custom type");
11371 assert(NumElts <= 8 && "The size of the constant array must be fixed");
11373 // In the #idef/#else code, we have in common:
11374 // - The vector of constants:
11380 // Create the splat vector for 0x4b000000.
11381 SDValue CstLow = DAG.getConstant(0x4b000000, MVT::i32);
11382 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
11383 CstLow, CstLow, CstLow, CstLow};
11384 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11385 makeArrayRef(&CstLowArray[0], NumElts));
11386 // Create the splat vector for 0x53000000.
11387 SDValue CstHigh = DAG.getConstant(0x53000000, MVT::i32);
11388 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
11389 CstHigh, CstHigh, CstHigh, CstHigh};
11390 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11391 makeArrayRef(&CstHighArray[0], NumElts));
11393 // Create the right shift.
11394 SDValue CstShift = DAG.getConstant(16, MVT::i32);
11395 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
11396 CstShift, CstShift, CstShift, CstShift};
11397 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
11398 makeArrayRef(&CstShiftArray[0], NumElts));
11399 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
11402 if (Subtarget.hasSSE41()) {
11403 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
11404 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
11405 SDValue VecCstLowBitcast =
11406 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstLow);
11407 SDValue VecBitcast = DAG.getNode(ISD::BITCAST, DL, VecI16VT, V);
11408 // Low will be bitcasted right away, so do not bother bitcasting back to its
11410 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
11411 VecCstLowBitcast, DAG.getConstant(0xaa, MVT::i32));
11412 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
11413 // (uint4) 0x53000000, 0xaa);
11414 SDValue VecCstHighBitcast =
11415 DAG.getNode(ISD::BITCAST, DL, VecI16VT, VecCstHigh);
11416 SDValue VecShiftBitcast =
11417 DAG.getNode(ISD::BITCAST, DL, VecI16VT, HighShift);
11418 // High will be bitcasted right away, so do not bother bitcasting back to
11419 // its original type.
11420 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
11421 VecCstHighBitcast, DAG.getConstant(0xaa, MVT::i32));
11423 SDValue CstMask = DAG.getConstant(0xffff, MVT::i32);
11424 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
11425 CstMask, CstMask, CstMask);
11426 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
11427 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
11428 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
11430 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
11431 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
11434 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
11435 SDValue CstFAdd = DAG.getConstantFP(
11436 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), MVT::f32);
11437 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
11438 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
11439 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
11440 makeArrayRef(&CstFAddArray[0], NumElts));
11442 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
11443 SDValue HighBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, High);
11445 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
11446 // return (float4) lo + fhi;
11447 SDValue LowBitcast = DAG.getNode(ISD::BITCAST, DL, VecFloatVT, Low);
11448 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
11451 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
11452 SelectionDAG &DAG) const {
11453 SDValue N0 = Op.getOperand(0);
11454 MVT SVT = N0.getSimpleValueType();
11457 switch (SVT.SimpleTy) {
11459 llvm_unreachable("Custom UINT_TO_FP is not supported!");
11464 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
11465 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11466 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
11470 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
11472 llvm_unreachable(nullptr);
11475 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
11476 SelectionDAG &DAG) const {
11477 SDValue N0 = Op.getOperand(0);
11480 if (Op.getValueType().isVector())
11481 return lowerUINT_TO_FP_vec(Op, DAG);
11483 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
11484 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
11485 // the optimization here.
11486 if (DAG.SignBitIsZero(N0))
11487 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
11489 MVT SrcVT = N0.getSimpleValueType();
11490 MVT DstVT = Op.getSimpleValueType();
11491 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
11492 return LowerUINT_TO_FP_i64(Op, DAG);
11493 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
11494 return LowerUINT_TO_FP_i32(Op, DAG);
11495 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
11498 // Make a 64-bit buffer, and use it to build an FILD.
11499 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
11500 if (SrcVT == MVT::i32) {
11501 SDValue WordOff = DAG.getConstant(4, getPointerTy());
11502 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
11503 getPointerTy(), StackSlot, WordOff);
11504 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11505 StackSlot, MachinePointerInfo(),
11507 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
11508 OffsetSlot, MachinePointerInfo(),
11510 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
11514 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
11515 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11516 StackSlot, MachinePointerInfo(),
11518 // For i64 source, we need to add the appropriate power of 2 if the input
11519 // was negative. This is the same as the optimization in
11520 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
11521 // we must be careful to do the computation in x87 extended precision, not
11522 // in SSE. (The generic code can't know it's OK to do this, or how to.)
11523 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
11524 MachineMemOperand *MMO =
11525 DAG.getMachineFunction()
11526 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11527 MachineMemOperand::MOLoad, 8, 8);
11529 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
11530 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
11531 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
11534 APInt FF(32, 0x5F800000ULL);
11536 // Check whether the sign bit is set.
11537 SDValue SignSet = DAG.getSetCC(dl,
11538 getSetCCResultType(*DAG.getContext(), MVT::i64),
11539 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
11542 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
11543 SDValue FudgePtr = DAG.getConstantPool(
11544 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
11547 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
11548 SDValue Zero = DAG.getIntPtrConstant(0);
11549 SDValue Four = DAG.getIntPtrConstant(4);
11550 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
11552 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
11554 // Load the value out, extending it from f32 to f80.
11555 // FIXME: Avoid the extend by constructing the right constant pool?
11556 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
11557 FudgePtr, MachinePointerInfo::getConstantPool(),
11558 MVT::f32, false, false, false, 4);
11559 // Extend everything to 80 bits to force it to be done on x87.
11560 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
11561 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
11564 std::pair<SDValue,SDValue>
11565 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
11566 bool IsSigned, bool IsReplace) const {
11569 EVT DstTy = Op.getValueType();
11571 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
11572 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
11576 assert(DstTy.getSimpleVT() <= MVT::i64 &&
11577 DstTy.getSimpleVT() >= MVT::i16 &&
11578 "Unknown FP_TO_INT to lower!");
11580 // These are really Legal.
11581 if (DstTy == MVT::i32 &&
11582 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11583 return std::make_pair(SDValue(), SDValue());
11584 if (Subtarget->is64Bit() &&
11585 DstTy == MVT::i64 &&
11586 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
11587 return std::make_pair(SDValue(), SDValue());
11589 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
11590 // stack slot, or into the FTOL runtime function.
11591 MachineFunction &MF = DAG.getMachineFunction();
11592 unsigned MemSize = DstTy.getSizeInBits()/8;
11593 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11594 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11597 if (!IsSigned && isIntegerTypeFTOL(DstTy))
11598 Opc = X86ISD::WIN_FTOL;
11600 switch (DstTy.getSimpleVT().SimpleTy) {
11601 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
11602 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
11603 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
11604 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
11607 SDValue Chain = DAG.getEntryNode();
11608 SDValue Value = Op.getOperand(0);
11609 EVT TheVT = Op.getOperand(0).getValueType();
11610 // FIXME This causes a redundant load/store if the SSE-class value is already
11611 // in memory, such as if it is on the callstack.
11612 if (isScalarFPTypeInSSEReg(TheVT)) {
11613 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
11614 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
11615 MachinePointerInfo::getFixedStack(SSFI),
11617 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
11619 Chain, StackSlot, DAG.getValueType(TheVT)
11622 MachineMemOperand *MMO =
11623 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11624 MachineMemOperand::MOLoad, MemSize, MemSize);
11625 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
11626 Chain = Value.getValue(1);
11627 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
11628 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
11631 MachineMemOperand *MMO =
11632 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11633 MachineMemOperand::MOStore, MemSize, MemSize);
11635 if (Opc != X86ISD::WIN_FTOL) {
11636 // Build the FP_TO_INT*_IN_MEM
11637 SDValue Ops[] = { Chain, Value, StackSlot };
11638 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
11640 return std::make_pair(FIST, StackSlot);
11642 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
11643 DAG.getVTList(MVT::Other, MVT::Glue),
11645 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
11646 MVT::i32, ftol.getValue(1));
11647 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
11648 MVT::i32, eax.getValue(2));
11649 SDValue Ops[] = { eax, edx };
11650 SDValue pair = IsReplace
11651 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
11652 : DAG.getMergeValues(Ops, DL);
11653 return std::make_pair(pair, SDValue());
11657 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
11658 const X86Subtarget *Subtarget) {
11659 MVT VT = Op->getSimpleValueType(0);
11660 SDValue In = Op->getOperand(0);
11661 MVT InVT = In.getSimpleValueType();
11664 // Optimize vectors in AVX mode:
11667 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
11668 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
11669 // Concat upper and lower parts.
11672 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
11673 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
11674 // Concat upper and lower parts.
11677 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
11678 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
11679 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
11682 if (Subtarget->hasInt256())
11683 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
11685 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
11686 SDValue Undef = DAG.getUNDEF(InVT);
11687 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
11688 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11689 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
11691 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
11692 VT.getVectorNumElements()/2);
11694 OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
11695 OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
11697 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
11700 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
11701 SelectionDAG &DAG) {
11702 MVT VT = Op->getSimpleValueType(0);
11703 SDValue In = Op->getOperand(0);
11704 MVT InVT = In.getSimpleValueType();
11706 unsigned int NumElts = VT.getVectorNumElements();
11707 if (NumElts != 8 && NumElts != 16)
11710 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
11711 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
11713 EVT ExtVT = (NumElts == 8)? MVT::v8i64 : MVT::v16i32;
11714 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11715 // Now we have only mask extension
11716 assert(InVT.getVectorElementType() == MVT::i1);
11717 SDValue Cst = DAG.getTargetConstant(1, ExtVT.getScalarType());
11718 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11719 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
11720 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11721 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11722 MachinePointerInfo::getConstantPool(),
11723 false, false, false, Alignment);
11725 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, DL, ExtVT, In, Ld);
11726 if (VT.is512BitVector())
11728 return DAG.getNode(X86ISD::VTRUNC, DL, VT, Brcst);
11731 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11732 SelectionDAG &DAG) {
11733 if (Subtarget->hasFp256()) {
11734 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11742 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
11743 SelectionDAG &DAG) {
11745 MVT VT = Op.getSimpleValueType();
11746 SDValue In = Op.getOperand(0);
11747 MVT SVT = In.getSimpleValueType();
11749 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
11750 return LowerZERO_EXTEND_AVX512(Op, DAG);
11752 if (Subtarget->hasFp256()) {
11753 SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
11758 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
11759 VT.getVectorNumElements() != SVT.getVectorNumElements());
11763 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
11765 MVT VT = Op.getSimpleValueType();
11766 SDValue In = Op.getOperand(0);
11767 MVT InVT = In.getSimpleValueType();
11769 if (VT == MVT::i1) {
11770 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
11771 "Invalid scalar TRUNCATE operation");
11772 if (InVT.getSizeInBits() >= 32)
11774 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
11775 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
11777 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
11778 "Invalid TRUNCATE operation");
11780 if (InVT.is512BitVector() || VT.getVectorElementType() == MVT::i1) {
11781 if (VT.getVectorElementType().getSizeInBits() >=8)
11782 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
11784 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
11785 unsigned NumElts = InVT.getVectorNumElements();
11786 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
11787 if (InVT.getSizeInBits() < 512) {
11788 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
11789 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
11793 SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
11794 const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
11795 SDValue CP = DAG.getConstantPool(C, getPointerTy());
11796 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
11797 SDValue Ld = DAG.getLoad(Cst.getValueType(), DL, DAG.getEntryNode(), CP,
11798 MachinePointerInfo::getConstantPool(),
11799 false, false, false, Alignment);
11800 SDValue OneV = DAG.getNode(X86ISD::VBROADCAST, DL, InVT, Ld);
11801 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
11802 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
11805 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
11806 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
11807 if (Subtarget->hasInt256()) {
11808 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
11809 In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
11810 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
11812 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
11813 DAG.getIntPtrConstant(0));
11816 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11817 DAG.getIntPtrConstant(0));
11818 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11819 DAG.getIntPtrConstant(2));
11820 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11821 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11822 static const int ShufMask[] = {0, 2, 4, 6};
11823 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
11826 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
11827 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
11828 if (Subtarget->hasInt256()) {
11829 In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
11831 SmallVector<SDValue,32> pshufbMask;
11832 for (unsigned i = 0; i < 2; ++i) {
11833 pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
11834 pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
11835 pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
11836 pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
11837 pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
11838 pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
11839 pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
11840 pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
11841 for (unsigned j = 0; j < 8; ++j)
11842 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
11844 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
11845 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
11846 In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
11848 static const int ShufMask[] = {0, 2, -1, -1};
11849 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
11851 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
11852 DAG.getIntPtrConstant(0));
11853 return DAG.getNode(ISD::BITCAST, DL, VT, In);
11856 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11857 DAG.getIntPtrConstant(0));
11859 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
11860 DAG.getIntPtrConstant(4));
11862 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
11863 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
11865 // The PSHUFB mask:
11866 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
11867 -1, -1, -1, -1, -1, -1, -1, -1};
11869 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
11870 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
11871 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
11873 OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
11874 OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
11876 // The MOVLHPS Mask:
11877 static const int ShufMask2[] = {0, 1, 4, 5};
11878 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
11879 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
11882 // Handle truncation of V256 to V128 using shuffles.
11883 if (!VT.is128BitVector() || !InVT.is256BitVector())
11886 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
11888 unsigned NumElems = VT.getVectorNumElements();
11889 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
11891 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
11892 // Prepare truncation shuffle mask
11893 for (unsigned i = 0; i != NumElems; ++i)
11894 MaskVec[i] = i * 2;
11895 SDValue V = DAG.getVectorShuffle(NVT, DL,
11896 DAG.getNode(ISD::BITCAST, DL, NVT, In),
11897 DAG.getUNDEF(NVT), &MaskVec[0]);
11898 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
11899 DAG.getIntPtrConstant(0));
11902 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
11903 SelectionDAG &DAG) const {
11904 assert(!Op.getSimpleValueType().isVector());
11906 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11907 /*IsSigned=*/ true, /*IsReplace=*/ false);
11908 SDValue FIST = Vals.first, StackSlot = Vals.second;
11909 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
11910 if (!FIST.getNode()) return Op;
11912 if (StackSlot.getNode())
11913 // Load the result.
11914 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11915 FIST, StackSlot, MachinePointerInfo(),
11916 false, false, false, 0);
11918 // The node is the result.
11922 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
11923 SelectionDAG &DAG) const {
11924 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
11925 /*IsSigned=*/ false, /*IsReplace=*/ false);
11926 SDValue FIST = Vals.first, StackSlot = Vals.second;
11927 assert(FIST.getNode() && "Unexpected failure");
11929 if (StackSlot.getNode())
11930 // Load the result.
11931 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
11932 FIST, StackSlot, MachinePointerInfo(),
11933 false, false, false, 0);
11935 // The node is the result.
11939 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
11941 MVT VT = Op.getSimpleValueType();
11942 SDValue In = Op.getOperand(0);
11943 MVT SVT = In.getSimpleValueType();
11945 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
11947 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
11948 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
11949 In, DAG.getUNDEF(SVT)));
11952 /// The only differences between FABS and FNEG are the mask and the logic op.
11953 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
11954 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
11955 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
11956 "Wrong opcode for lowering FABS or FNEG.");
11958 bool IsFABS = (Op.getOpcode() == ISD::FABS);
11960 // If this is a FABS and it has an FNEG user, bail out to fold the combination
11961 // into an FNABS. We'll lower the FABS after that if it is still in use.
11963 for (SDNode *User : Op->uses())
11964 if (User->getOpcode() == ISD::FNEG)
11967 SDValue Op0 = Op.getOperand(0);
11968 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
11971 MVT VT = Op.getSimpleValueType();
11972 // Assume scalar op for initialization; update for vector if needed.
11973 // Note that there are no scalar bitwise logical SSE/AVX instructions, so we
11974 // generate a 16-byte vector constant and logic op even for the scalar case.
11975 // Using a 16-byte mask allows folding the load of the mask with
11976 // the logic op, so it can save (~4 bytes) on code size.
11978 unsigned NumElts = VT == MVT::f64 ? 2 : 4;
11979 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
11980 // decide if we should generate a 16-byte constant mask when we only need 4 or
11981 // 8 bytes for the scalar case.
11982 if (VT.isVector()) {
11983 EltVT = VT.getVectorElementType();
11984 NumElts = VT.getVectorNumElements();
11987 unsigned EltBits = EltVT.getSizeInBits();
11988 LLVMContext *Context = DAG.getContext();
11989 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
11991 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
11992 Constant *C = ConstantInt::get(*Context, MaskElt);
11993 C = ConstantVector::getSplat(NumElts, C);
11994 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11995 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy());
11996 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11997 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
11998 MachinePointerInfo::getConstantPool(),
11999 false, false, false, Alignment);
12001 if (VT.isVector()) {
12002 // For a vector, cast operands to a vector type, perform the logic op,
12003 // and cast the result back to the original value type.
12004 MVT VecVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
12005 SDValue MaskCasted = DAG.getNode(ISD::BITCAST, dl, VecVT, Mask);
12006 SDValue Operand = IsFNABS ?
12007 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0.getOperand(0)) :
12008 DAG.getNode(ISD::BITCAST, dl, VecVT, Op0);
12009 unsigned BitOp = IsFABS ? ISD::AND : IsFNABS ? ISD::OR : ISD::XOR;
12010 return DAG.getNode(ISD::BITCAST, dl, VT,
12011 DAG.getNode(BitOp, dl, VecVT, Operand, MaskCasted));
12014 // If not vector, then scalar.
12015 unsigned BitOp = IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12016 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12017 return DAG.getNode(BitOp, dl, VT, Operand, Mask);
12020 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12021 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12022 LLVMContext *Context = DAG.getContext();
12023 SDValue Op0 = Op.getOperand(0);
12024 SDValue Op1 = Op.getOperand(1);
12026 MVT VT = Op.getSimpleValueType();
12027 MVT SrcVT = Op1.getSimpleValueType();
12029 // If second operand is smaller, extend it first.
12030 if (SrcVT.bitsLT(VT)) {
12031 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12034 // And if it is bigger, shrink it first.
12035 if (SrcVT.bitsGT(VT)) {
12036 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
12040 // At this point the operands and the result should have the same
12041 // type, and that won't be f80 since that is not custom lowered.
12043 const fltSemantics &Sem =
12044 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12045 const unsigned SizeInBits = VT.getSizeInBits();
12047 SmallVector<Constant *, 4> CV(
12048 VT == MVT::f64 ? 2 : 4,
12049 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12051 // First, clear all bits but the sign bit from the second operand (sign).
12052 CV[0] = ConstantFP::get(*Context,
12053 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12054 Constant *C = ConstantVector::get(CV);
12055 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12056 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
12057 MachinePointerInfo::getConstantPool(),
12058 false, false, false, 16);
12059 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
12061 // Next, clear the sign bit from the first operand (magnitude).
12062 // If it's a constant, we can clear it here.
12063 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12064 APFloat APF = Op0CN->getValueAPF();
12065 // If the magnitude is a positive zero, the sign bit alone is enough.
12066 if (APF.isPosZero())
12069 CV[0] = ConstantFP::get(*Context, APF);
12071 CV[0] = ConstantFP::get(
12073 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12075 C = ConstantVector::get(CV);
12076 CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(), 16);
12077 SDValue Val = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
12078 MachinePointerInfo::getConstantPool(),
12079 false, false, false, 16);
12080 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12081 if (!isa<ConstantFPSDNode>(Op0))
12082 Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Val);
12084 // OR the magnitude value with the sign bit.
12085 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
12088 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12089 SDValue N0 = Op.getOperand(0);
12091 MVT VT = Op.getSimpleValueType();
12093 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12094 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12095 DAG.getConstant(1, VT));
12096 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
12099 // Check whether an OR'd tree is PTEST-able.
12100 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12101 SelectionDAG &DAG) {
12102 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12104 if (!Subtarget->hasSSE41())
12107 if (!Op->hasOneUse())
12110 SDNode *N = Op.getNode();
12113 SmallVector<SDValue, 8> Opnds;
12114 DenseMap<SDValue, unsigned> VecInMap;
12115 SmallVector<SDValue, 8> VecIns;
12116 EVT VT = MVT::Other;
12118 // Recognize a special case where a vector is casted into wide integer to
12120 Opnds.push_back(N->getOperand(0));
12121 Opnds.push_back(N->getOperand(1));
12123 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12124 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12125 // BFS traverse all OR'd operands.
12126 if (I->getOpcode() == ISD::OR) {
12127 Opnds.push_back(I->getOperand(0));
12128 Opnds.push_back(I->getOperand(1));
12129 // Re-evaluate the number of nodes to be traversed.
12130 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12134 // Quit if a non-EXTRACT_VECTOR_ELT
12135 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12138 // Quit if without a constant index.
12139 SDValue Idx = I->getOperand(1);
12140 if (!isa<ConstantSDNode>(Idx))
12143 SDValue ExtractedFromVec = I->getOperand(0);
12144 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12145 if (M == VecInMap.end()) {
12146 VT = ExtractedFromVec.getValueType();
12147 // Quit if not 128/256-bit vector.
12148 if (!VT.is128BitVector() && !VT.is256BitVector())
12150 // Quit if not the same type.
12151 if (VecInMap.begin() != VecInMap.end() &&
12152 VT != VecInMap.begin()->first.getValueType())
12154 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12155 VecIns.push_back(ExtractedFromVec);
12157 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12160 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12161 "Not extracted from 128-/256-bit vector.");
12163 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12165 for (DenseMap<SDValue, unsigned>::const_iterator
12166 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12167 // Quit if not all elements are used.
12168 if (I->second != FullMask)
12172 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12174 // Cast all vectors into TestVT for PTEST.
12175 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12176 VecIns[i] = DAG.getNode(ISD::BITCAST, DL, TestVT, VecIns[i]);
12178 // If more than one full vectors are evaluated, OR them first before PTEST.
12179 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12180 // Each iteration will OR 2 nodes and append the result until there is only
12181 // 1 node left, i.e. the final OR'd value of all vectors.
12182 SDValue LHS = VecIns[Slot];
12183 SDValue RHS = VecIns[Slot + 1];
12184 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12187 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12188 VecIns.back(), VecIns.back());
12191 /// \brief return true if \c Op has a use that doesn't just read flags.
12192 static bool hasNonFlagsUse(SDValue Op) {
12193 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12195 SDNode *User = *UI;
12196 unsigned UOpNo = UI.getOperandNo();
12197 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12198 // Look pass truncate.
12199 UOpNo = User->use_begin().getOperandNo();
12200 User = *User->use_begin();
12203 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12204 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12210 /// Emit nodes that will be selected as "test Op0,Op0", or something
12212 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12213 SelectionDAG &DAG) const {
12214 if (Op.getValueType() == MVT::i1) {
12215 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12216 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12217 DAG.getConstant(0, MVT::i8));
12219 // CF and OF aren't always set the way we want. Determine which
12220 // of these we need.
12221 bool NeedCF = false;
12222 bool NeedOF = false;
12225 case X86::COND_A: case X86::COND_AE:
12226 case X86::COND_B: case X86::COND_BE:
12229 case X86::COND_G: case X86::COND_GE:
12230 case X86::COND_L: case X86::COND_LE:
12231 case X86::COND_O: case X86::COND_NO: {
12232 // Check if we really need to set the
12233 // Overflow flag. If NoSignedWrap is present
12234 // that is not actually needed.
12235 switch (Op->getOpcode()) {
12240 const BinaryWithFlagsSDNode *BinNode =
12241 cast<BinaryWithFlagsSDNode>(Op.getNode());
12242 if (BinNode->hasNoSignedWrap())
12252 // See if we can use the EFLAGS value from the operand instead of
12253 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12254 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12255 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12256 // Emit a CMP with 0, which is the TEST pattern.
12257 //if (Op.getValueType() == MVT::i1)
12258 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12259 // DAG.getConstant(0, MVT::i1));
12260 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12261 DAG.getConstant(0, Op.getValueType()));
12263 unsigned Opcode = 0;
12264 unsigned NumOperands = 0;
12266 // Truncate operations may prevent the merge of the SETCC instruction
12267 // and the arithmetic instruction before it. Attempt to truncate the operands
12268 // of the arithmetic instruction and use a reduced bit-width instruction.
12269 bool NeedTruncation = false;
12270 SDValue ArithOp = Op;
12271 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12272 SDValue Arith = Op->getOperand(0);
12273 // Both the trunc and the arithmetic op need to have one user each.
12274 if (Arith->hasOneUse())
12275 switch (Arith.getOpcode()) {
12282 NeedTruncation = true;
12288 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
12289 // which may be the result of a CAST. We use the variable 'Op', which is the
12290 // non-casted variable when we check for possible users.
12291 switch (ArithOp.getOpcode()) {
12293 // Due to an isel shortcoming, be conservative if this add is likely to be
12294 // selected as part of a load-modify-store instruction. When the root node
12295 // in a match is a store, isel doesn't know how to remap non-chain non-flag
12296 // uses of other nodes in the match, such as the ADD in this case. This
12297 // leads to the ADD being left around and reselected, with the result being
12298 // two adds in the output. Alas, even if none our users are stores, that
12299 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
12300 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
12301 // climbing the DAG back to the root, and it doesn't seem to be worth the
12303 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12304 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12305 if (UI->getOpcode() != ISD::CopyToReg &&
12306 UI->getOpcode() != ISD::SETCC &&
12307 UI->getOpcode() != ISD::STORE)
12310 if (ConstantSDNode *C =
12311 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
12312 // An add of one will be selected as an INC.
12313 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
12314 Opcode = X86ISD::INC;
12319 // An add of negative one (subtract of one) will be selected as a DEC.
12320 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
12321 Opcode = X86ISD::DEC;
12327 // Otherwise use a regular EFLAGS-setting add.
12328 Opcode = X86ISD::ADD;
12333 // If we have a constant logical shift that's only used in a comparison
12334 // against zero turn it into an equivalent AND. This allows turning it into
12335 // a TEST instruction later.
12336 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
12337 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
12338 EVT VT = Op.getValueType();
12339 unsigned BitWidth = VT.getSizeInBits();
12340 unsigned ShAmt = Op->getConstantOperandVal(1);
12341 if (ShAmt >= BitWidth) // Avoid undefined shifts.
12343 APInt Mask = ArithOp.getOpcode() == ISD::SRL
12344 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
12345 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
12346 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
12348 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
12349 DAG.getConstant(Mask, VT));
12350 DAG.ReplaceAllUsesWith(Op, New);
12356 // If the primary and result isn't used, don't bother using X86ISD::AND,
12357 // because a TEST instruction will be better.
12358 if (!hasNonFlagsUse(Op))
12364 // Due to the ISEL shortcoming noted above, be conservative if this op is
12365 // likely to be selected as part of a load-modify-store instruction.
12366 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12367 UE = Op.getNode()->use_end(); UI != UE; ++UI)
12368 if (UI->getOpcode() == ISD::STORE)
12371 // Otherwise use a regular EFLAGS-setting instruction.
12372 switch (ArithOp.getOpcode()) {
12373 default: llvm_unreachable("unexpected operator!");
12374 case ISD::SUB: Opcode = X86ISD::SUB; break;
12375 case ISD::XOR: Opcode = X86ISD::XOR; break;
12376 case ISD::AND: Opcode = X86ISD::AND; break;
12378 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
12379 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
12380 if (EFLAGS.getNode())
12383 Opcode = X86ISD::OR;
12397 return SDValue(Op.getNode(), 1);
12403 // If we found that truncation is beneficial, perform the truncation and
12405 if (NeedTruncation) {
12406 EVT VT = Op.getValueType();
12407 SDValue WideVal = Op->getOperand(0);
12408 EVT WideVT = WideVal.getValueType();
12409 unsigned ConvertedOp = 0;
12410 // Use a target machine opcode to prevent further DAGCombine
12411 // optimizations that may separate the arithmetic operations
12412 // from the setcc node.
12413 switch (WideVal.getOpcode()) {
12415 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
12416 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
12417 case ISD::AND: ConvertedOp = X86ISD::AND; break;
12418 case ISD::OR: ConvertedOp = X86ISD::OR; break;
12419 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
12423 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12424 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
12425 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
12426 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
12427 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
12433 // Emit a CMP with 0, which is the TEST pattern.
12434 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12435 DAG.getConstant(0, Op.getValueType()));
12437 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
12438 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
12440 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
12441 DAG.ReplaceAllUsesWith(Op, New);
12442 return SDValue(New.getNode(), 1);
12445 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
12447 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
12448 SDLoc dl, SelectionDAG &DAG) const {
12449 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
12450 if (C->getAPIntValue() == 0)
12451 return EmitTest(Op0, X86CC, dl, DAG);
12453 if (Op0.getValueType() == MVT::i1)
12454 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
12457 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
12458 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
12459 // Do the comparison at i32 if it's smaller, besides the Atom case.
12460 // This avoids subregister aliasing issues. Keep the smaller reference
12461 // if we're optimizing for size, however, as that'll allow better folding
12462 // of memory operations.
12463 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
12464 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
12465 Attribute::MinSize) &&
12466 !Subtarget->isAtom()) {
12467 unsigned ExtendOp =
12468 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
12469 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
12470 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
12472 // Use SUB instead of CMP to enable CSE between SUB and CMP.
12473 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
12474 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
12476 return SDValue(Sub.getNode(), 1);
12478 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
12481 /// Convert a comparison if required by the subtarget.
12482 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
12483 SelectionDAG &DAG) const {
12484 // If the subtarget does not support the FUCOMI instruction, floating-point
12485 // comparisons have to be converted.
12486 if (Subtarget->hasCMov() ||
12487 Cmp.getOpcode() != X86ISD::CMP ||
12488 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
12489 !Cmp.getOperand(1).getValueType().isFloatingPoint())
12492 // The instruction selector will select an FUCOM instruction instead of
12493 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
12494 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
12495 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
12497 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
12498 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
12499 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
12500 DAG.getConstant(8, MVT::i8));
12501 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
12502 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
12505 /// The minimum architected relative accuracy is 2^-12. We need one
12506 /// Newton-Raphson step to have a good float result (24 bits of precision).
12507 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
12508 DAGCombinerInfo &DCI,
12509 unsigned &RefinementSteps,
12510 bool &UseOneConstNR) const {
12511 // FIXME: We should use instruction latency models to calculate the cost of
12512 // each potential sequence, but this is very hard to do reliably because
12513 // at least Intel's Core* chips have variable timing based on the number of
12514 // significant digits in the divisor and/or sqrt operand.
12515 if (!Subtarget->useSqrtEst())
12518 EVT VT = Op.getValueType();
12520 // SSE1 has rsqrtss and rsqrtps.
12521 // TODO: Add support for AVX512 (v16f32).
12522 // It is likely not profitable to do this for f64 because a double-precision
12523 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
12524 // instructions: convert to single, rsqrtss, convert back to double, refine
12525 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
12526 // along with FMA, this could be a throughput win.
12527 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12528 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12529 RefinementSteps = 1;
12530 UseOneConstNR = false;
12531 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
12536 /// The minimum architected relative accuracy is 2^-12. We need one
12537 /// Newton-Raphson step to have a good float result (24 bits of precision).
12538 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
12539 DAGCombinerInfo &DCI,
12540 unsigned &RefinementSteps) const {
12541 // FIXME: We should use instruction latency models to calculate the cost of
12542 // each potential sequence, but this is very hard to do reliably because
12543 // at least Intel's Core* chips have variable timing based on the number of
12544 // significant digits in the divisor.
12545 if (!Subtarget->useReciprocalEst())
12548 EVT VT = Op.getValueType();
12550 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
12551 // TODO: Add support for AVX512 (v16f32).
12552 // It is likely not profitable to do this for f64 because a double-precision
12553 // reciprocal estimate with refinement on x86 prior to FMA requires
12554 // 15 instructions: convert to single, rcpss, convert back to double, refine
12555 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
12556 // along with FMA, this could be a throughput win.
12557 if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
12558 (Subtarget->hasAVX() && VT == MVT::v8f32)) {
12559 RefinementSteps = ReciprocalEstimateRefinementSteps;
12560 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
12565 static bool isAllOnes(SDValue V) {
12566 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
12567 return C && C->isAllOnesValue();
12570 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
12571 /// if it's possible.
12572 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
12573 SDLoc dl, SelectionDAG &DAG) const {
12574 SDValue Op0 = And.getOperand(0);
12575 SDValue Op1 = And.getOperand(1);
12576 if (Op0.getOpcode() == ISD::TRUNCATE)
12577 Op0 = Op0.getOperand(0);
12578 if (Op1.getOpcode() == ISD::TRUNCATE)
12579 Op1 = Op1.getOperand(0);
12582 if (Op1.getOpcode() == ISD::SHL)
12583 std::swap(Op0, Op1);
12584 if (Op0.getOpcode() == ISD::SHL) {
12585 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
12586 if (And00C->getZExtValue() == 1) {
12587 // If we looked past a truncate, check that it's only truncating away
12589 unsigned BitWidth = Op0.getValueSizeInBits();
12590 unsigned AndBitWidth = And.getValueSizeInBits();
12591 if (BitWidth > AndBitWidth) {
12593 DAG.computeKnownBits(Op0, Zeros, Ones);
12594 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
12598 RHS = Op0.getOperand(1);
12600 } else if (Op1.getOpcode() == ISD::Constant) {
12601 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
12602 uint64_t AndRHSVal = AndRHS->getZExtValue();
12603 SDValue AndLHS = Op0;
12605 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
12606 LHS = AndLHS.getOperand(0);
12607 RHS = AndLHS.getOperand(1);
12610 // Use BT if the immediate can't be encoded in a TEST instruction.
12611 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
12613 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), LHS.getValueType());
12617 if (LHS.getNode()) {
12618 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
12619 // instruction. Since the shift amount is in-range-or-undefined, we know
12620 // that doing a bittest on the i32 value is ok. We extend to i32 because
12621 // the encoding for the i16 version is larger than the i32 version.
12622 // Also promote i16 to i32 for performance / code size reason.
12623 if (LHS.getValueType() == MVT::i8 ||
12624 LHS.getValueType() == MVT::i16)
12625 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
12627 // If the operand types disagree, extend the shift amount to match. Since
12628 // BT ignores high bits (like shifts) we can use anyextend.
12629 if (LHS.getValueType() != RHS.getValueType())
12630 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
12632 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
12633 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
12634 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
12635 DAG.getConstant(Cond, MVT::i8), BT);
12641 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
12643 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
12648 // SSE Condition code mapping:
12657 switch (SetCCOpcode) {
12658 default: llvm_unreachable("Unexpected SETCC condition");
12660 case ISD::SETEQ: SSECC = 0; break;
12662 case ISD::SETGT: Swap = true; // Fallthrough
12664 case ISD::SETOLT: SSECC = 1; break;
12666 case ISD::SETGE: Swap = true; // Fallthrough
12668 case ISD::SETOLE: SSECC = 2; break;
12669 case ISD::SETUO: SSECC = 3; break;
12671 case ISD::SETNE: SSECC = 4; break;
12672 case ISD::SETULE: Swap = true; // Fallthrough
12673 case ISD::SETUGE: SSECC = 5; break;
12674 case ISD::SETULT: Swap = true; // Fallthrough
12675 case ISD::SETUGT: SSECC = 6; break;
12676 case ISD::SETO: SSECC = 7; break;
12678 case ISD::SETONE: SSECC = 8; break;
12681 std::swap(Op0, Op1);
12686 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
12687 // ones, and then concatenate the result back.
12688 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
12689 MVT VT = Op.getSimpleValueType();
12691 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
12692 "Unsupported value type for operation");
12694 unsigned NumElems = VT.getVectorNumElements();
12696 SDValue CC = Op.getOperand(2);
12698 // Extract the LHS vectors
12699 SDValue LHS = Op.getOperand(0);
12700 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
12701 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
12703 // Extract the RHS vectors
12704 SDValue RHS = Op.getOperand(1);
12705 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
12706 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
12708 // Issue the operation on the smaller types and concatenate the result back
12709 MVT EltVT = VT.getVectorElementType();
12710 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
12711 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
12712 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
12713 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
12716 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
12717 const X86Subtarget *Subtarget) {
12718 SDValue Op0 = Op.getOperand(0);
12719 SDValue Op1 = Op.getOperand(1);
12720 SDValue CC = Op.getOperand(2);
12721 MVT VT = Op.getSimpleValueType();
12724 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
12725 Op.getValueType().getScalarType() == MVT::i1 &&
12726 "Cannot set masked compare for this operation");
12728 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12730 bool Unsigned = false;
12733 switch (SetCCOpcode) {
12734 default: llvm_unreachable("Unexpected SETCC condition");
12735 case ISD::SETNE: SSECC = 4; break;
12736 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
12737 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
12738 case ISD::SETLT: Swap = true; //fall-through
12739 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
12740 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
12741 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
12742 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
12743 case ISD::SETULE: Unsigned = true; //fall-through
12744 case ISD::SETLE: SSECC = 2; break;
12748 std::swap(Op0, Op1);
12750 return DAG.getNode(Opc, dl, VT, Op0, Op1);
12751 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
12752 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12753 DAG.getConstant(SSECC, MVT::i8));
12756 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
12757 /// operand \p Op1. If non-trivial (for example because it's not constant)
12758 /// return an empty value.
12759 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
12761 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
12765 MVT VT = Op1.getSimpleValueType();
12766 MVT EVT = VT.getVectorElementType();
12767 unsigned n = VT.getVectorNumElements();
12768 SmallVector<SDValue, 8> ULTOp1;
12770 for (unsigned i = 0; i < n; ++i) {
12771 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
12772 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
12775 // Avoid underflow.
12776 APInt Val = Elt->getAPIntValue();
12780 ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
12783 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
12786 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
12787 SelectionDAG &DAG) {
12788 SDValue Op0 = Op.getOperand(0);
12789 SDValue Op1 = Op.getOperand(1);
12790 SDValue CC = Op.getOperand(2);
12791 MVT VT = Op.getSimpleValueType();
12792 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
12793 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
12798 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
12799 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
12802 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
12803 unsigned Opc = X86ISD::CMPP;
12804 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
12805 assert(VT.getVectorNumElements() <= 16);
12806 Opc = X86ISD::CMPM;
12808 // In the two special cases we can't handle, emit two comparisons.
12811 unsigned CombineOpc;
12812 if (SetCCOpcode == ISD::SETUEQ) {
12813 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
12815 assert(SetCCOpcode == ISD::SETONE);
12816 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
12819 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12820 DAG.getConstant(CC0, MVT::i8));
12821 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
12822 DAG.getConstant(CC1, MVT::i8));
12823 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
12825 // Handle all other FP comparisons here.
12826 return DAG.getNode(Opc, dl, VT, Op0, Op1,
12827 DAG.getConstant(SSECC, MVT::i8));
12830 // Break 256-bit integer vector compare into smaller ones.
12831 if (VT.is256BitVector() && !Subtarget->hasInt256())
12832 return Lower256IntVSETCC(Op, DAG);
12834 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
12835 EVT OpVT = Op1.getValueType();
12836 if (Subtarget->hasAVX512()) {
12837 if (Op1.getValueType().is512BitVector() ||
12838 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
12839 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
12840 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
12842 // In AVX-512 architecture setcc returns mask with i1 elements,
12843 // But there is no compare instruction for i8 and i16 elements in KNL.
12844 // We are not talking about 512-bit operands in this case, these
12845 // types are illegal.
12847 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
12848 OpVT.getVectorElementType().getSizeInBits() >= 8))
12849 return DAG.getNode(ISD::TRUNCATE, dl, VT,
12850 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
12853 // We are handling one of the integer comparisons here. Since SSE only has
12854 // GT and EQ comparisons for integer, swapping operands and multiple
12855 // operations may be required for some comparisons.
12857 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
12858 bool Subus = false;
12860 switch (SetCCOpcode) {
12861 default: llvm_unreachable("Unexpected SETCC condition");
12862 case ISD::SETNE: Invert = true;
12863 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
12864 case ISD::SETLT: Swap = true;
12865 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
12866 case ISD::SETGE: Swap = true;
12867 case ISD::SETLE: Opc = X86ISD::PCMPGT;
12868 Invert = true; break;
12869 case ISD::SETULT: Swap = true;
12870 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
12871 FlipSigns = true; break;
12872 case ISD::SETUGE: Swap = true;
12873 case ISD::SETULE: Opc = X86ISD::PCMPGT;
12874 FlipSigns = true; Invert = true; break;
12877 // Special case: Use min/max operations for SETULE/SETUGE
12878 MVT VET = VT.getVectorElementType();
12880 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
12881 || (Subtarget->hasSSE2() && (VET == MVT::i8));
12884 switch (SetCCOpcode) {
12886 case ISD::SETULE: Opc = X86ISD::UMIN; MinMax = true; break;
12887 case ISD::SETUGE: Opc = X86ISD::UMAX; MinMax = true; break;
12890 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
12893 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
12894 if (!MinMax && hasSubus) {
12895 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
12897 // t = psubus Op0, Op1
12898 // pcmpeq t, <0..0>
12899 switch (SetCCOpcode) {
12901 case ISD::SETULT: {
12902 // If the comparison is against a constant we can turn this into a
12903 // setule. With psubus, setule does not require a swap. This is
12904 // beneficial because the constant in the register is no longer
12905 // destructed as the destination so it can be hoisted out of a loop.
12906 // Only do this pre-AVX since vpcmp* is no longer destructive.
12907 if (Subtarget->hasAVX())
12909 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
12910 if (ULEOp1.getNode()) {
12912 Subus = true; Invert = false; Swap = false;
12916 // Psubus is better than flip-sign because it requires no inversion.
12917 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
12918 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
12922 Opc = X86ISD::SUBUS;
12928 std::swap(Op0, Op1);
12930 // Check that the operation in question is available (most are plain SSE2,
12931 // but PCMPGTQ and PCMPEQQ have different requirements).
12932 if (VT == MVT::v2i64) {
12933 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
12934 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
12936 // First cast everything to the right type.
12937 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12938 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12940 // Since SSE has no unsigned integer comparisons, we need to flip the sign
12941 // bits of the inputs before performing those operations. The lower
12942 // compare is always unsigned.
12945 SB = DAG.getConstant(0x80000000U, MVT::v4i32);
12947 SDValue Sign = DAG.getConstant(0x80000000U, MVT::i32);
12948 SDValue Zero = DAG.getConstant(0x00000000U, MVT::i32);
12949 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
12950 Sign, Zero, Sign, Zero);
12952 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
12953 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
12955 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
12956 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
12957 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
12959 // Create masks for only the low parts/high parts of the 64 bit integers.
12960 static const int MaskHi[] = { 1, 1, 3, 3 };
12961 static const int MaskLo[] = { 0, 0, 2, 2 };
12962 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
12963 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
12964 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
12966 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
12967 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
12970 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12972 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12975 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
12976 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
12977 // pcmpeqd + pshufd + pand.
12978 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
12980 // First cast everything to the right type.
12981 Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
12982 Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
12985 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
12987 // Make sure the lower and upper halves are both all-ones.
12988 static const int Mask[] = { 1, 0, 3, 2 };
12989 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
12990 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
12993 Result = DAG.getNOT(dl, Result, MVT::v4i32);
12995 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
12999 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13000 // bits of the inputs before performing those operations.
13002 EVT EltVT = VT.getVectorElementType();
13003 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), VT);
13004 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13005 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13008 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13010 // If the logical-not of the result is required, perform that now.
13012 Result = DAG.getNOT(dl, Result, VT);
13015 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13018 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13019 getZeroVector(VT, Subtarget, DAG, dl));
13024 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13026 MVT VT = Op.getSimpleValueType();
13028 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13030 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13031 && "SetCC type must be 8-bit or 1-bit integer");
13032 SDValue Op0 = Op.getOperand(0);
13033 SDValue Op1 = Op.getOperand(1);
13035 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13037 // Optimize to BT if possible.
13038 // Lower (X & (1 << N)) == 0 to BT(X, N).
13039 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13040 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13041 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13042 Op1.getOpcode() == ISD::Constant &&
13043 cast<ConstantSDNode>(Op1)->isNullValue() &&
13044 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13045 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13046 if (NewSetCC.getNode()) {
13048 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13053 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13055 if (Op1.getOpcode() == ISD::Constant &&
13056 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13057 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13058 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13060 // If the input is a setcc, then reuse the input setcc or use a new one with
13061 // the inverted condition.
13062 if (Op0.getOpcode() == X86ISD::SETCC) {
13063 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13064 bool Invert = (CC == ISD::SETNE) ^
13065 cast<ConstantSDNode>(Op1)->isNullValue();
13069 CCode = X86::GetOppositeBranchCondition(CCode);
13070 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13071 DAG.getConstant(CCode, MVT::i8),
13072 Op0.getOperand(1));
13074 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13078 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13079 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13080 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13082 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13083 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
13086 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13087 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
13088 if (X86CC == X86::COND_INVALID)
13091 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13092 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13093 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13094 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
13096 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13100 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13101 static bool isX86LogicalCmp(SDValue Op) {
13102 unsigned Opc = Op.getNode()->getOpcode();
13103 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13104 Opc == X86ISD::SAHF)
13106 if (Op.getResNo() == 1 &&
13107 (Opc == X86ISD::ADD ||
13108 Opc == X86ISD::SUB ||
13109 Opc == X86ISD::ADC ||
13110 Opc == X86ISD::SBB ||
13111 Opc == X86ISD::SMUL ||
13112 Opc == X86ISD::UMUL ||
13113 Opc == X86ISD::INC ||
13114 Opc == X86ISD::DEC ||
13115 Opc == X86ISD::OR ||
13116 Opc == X86ISD::XOR ||
13117 Opc == X86ISD::AND))
13120 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13126 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13127 if (V.getOpcode() != ISD::TRUNCATE)
13130 SDValue VOp0 = V.getOperand(0);
13131 unsigned InBits = VOp0.getValueSizeInBits();
13132 unsigned Bits = V.getValueSizeInBits();
13133 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13136 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13137 bool addTest = true;
13138 SDValue Cond = Op.getOperand(0);
13139 SDValue Op1 = Op.getOperand(1);
13140 SDValue Op2 = Op.getOperand(2);
13142 EVT VT = Op1.getValueType();
13145 // Lower fp selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13146 // are available. Otherwise fp cmovs get lowered into a less efficient branch
13147 // sequence later on.
13148 if (Cond.getOpcode() == ISD::SETCC &&
13149 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13150 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13151 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13152 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13153 int SSECC = translateX86FSETCC(
13154 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13157 if (Subtarget->hasAVX512()) {
13158 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13159 DAG.getConstant(SSECC, MVT::i8));
13160 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13162 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13163 DAG.getConstant(SSECC, MVT::i8));
13164 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13165 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13166 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13170 if (Cond.getOpcode() == ISD::SETCC) {
13171 SDValue NewCond = LowerSETCC(Cond, DAG);
13172 if (NewCond.getNode())
13176 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
13177 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
13178 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
13179 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
13180 if (Cond.getOpcode() == X86ISD::SETCC &&
13181 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
13182 isZero(Cond.getOperand(1).getOperand(1))) {
13183 SDValue Cmp = Cond.getOperand(1);
13185 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
13187 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
13188 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
13189 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
13191 SDValue CmpOp0 = Cmp.getOperand(0);
13192 // Apply further optimizations for special cases
13193 // (select (x != 0), -1, 0) -> neg & sbb
13194 // (select (x == 0), 0, -1) -> neg & sbb
13195 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
13196 if (YC->isNullValue() &&
13197 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
13198 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
13199 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
13200 DAG.getConstant(0, CmpOp0.getValueType()),
13202 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13203 DAG.getConstant(X86::COND_B, MVT::i8),
13204 SDValue(Neg.getNode(), 1));
13208 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
13209 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
13210 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13212 SDValue Res = // Res = 0 or -1.
13213 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13214 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
13216 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
13217 Res = DAG.getNOT(DL, Res, Res.getValueType());
13219 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
13220 if (!N2C || !N2C->isNullValue())
13221 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
13226 // Look past (and (setcc_carry (cmp ...)), 1).
13227 if (Cond.getOpcode() == ISD::AND &&
13228 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13229 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13230 if (C && C->getAPIntValue() == 1)
13231 Cond = Cond.getOperand(0);
13234 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13235 // setting operand in place of the X86ISD::SETCC.
13236 unsigned CondOpcode = Cond.getOpcode();
13237 if (CondOpcode == X86ISD::SETCC ||
13238 CondOpcode == X86ISD::SETCC_CARRY) {
13239 CC = Cond.getOperand(0);
13241 SDValue Cmp = Cond.getOperand(1);
13242 unsigned Opc = Cmp.getOpcode();
13243 MVT VT = Op.getSimpleValueType();
13245 bool IllegalFPCMov = false;
13246 if (VT.isFloatingPoint() && !VT.isVector() &&
13247 !isScalarFPTypeInSSEReg(VT)) // FPStack?
13248 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
13250 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
13251 Opc == X86ISD::BT) { // FIXME
13255 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13256 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13257 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13258 Cond.getOperand(0).getValueType() != MVT::i8)) {
13259 SDValue LHS = Cond.getOperand(0);
13260 SDValue RHS = Cond.getOperand(1);
13261 unsigned X86Opcode;
13264 switch (CondOpcode) {
13265 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13266 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13267 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13268 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13269 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13270 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13271 default: llvm_unreachable("unexpected overflowing operator");
13273 if (CondOpcode == ISD::UMULO)
13274 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13277 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13279 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
13281 if (CondOpcode == ISD::UMULO)
13282 Cond = X86Op.getValue(2);
13284 Cond = X86Op.getValue(1);
13286 CC = DAG.getConstant(X86Cond, MVT::i8);
13291 // Look pass the truncate if the high bits are known zero.
13292 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13293 Cond = Cond.getOperand(0);
13295 // We know the result of AND is compared against zero. Try to match
13297 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13298 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
13299 if (NewSetCC.getNode()) {
13300 CC = NewSetCC.getOperand(0);
13301 Cond = NewSetCC.getOperand(1);
13308 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13309 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
13312 // a < b ? -1 : 0 -> RES = ~setcc_carry
13313 // a < b ? 0 : -1 -> RES = setcc_carry
13314 // a >= b ? -1 : 0 -> RES = setcc_carry
13315 // a >= b ? 0 : -1 -> RES = ~setcc_carry
13316 if (Cond.getOpcode() == X86ISD::SUB) {
13317 Cond = ConvertCmpIfNecessary(Cond, DAG);
13318 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
13320 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
13321 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
13322 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
13323 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
13324 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
13325 return DAG.getNOT(DL, Res, Res.getValueType());
13330 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
13331 // widen the cmov and push the truncate through. This avoids introducing a new
13332 // branch during isel and doesn't add any extensions.
13333 if (Op.getValueType() == MVT::i8 &&
13334 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
13335 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
13336 if (T1.getValueType() == T2.getValueType() &&
13337 // Blacklist CopyFromReg to avoid partial register stalls.
13338 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
13339 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
13340 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
13341 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
13345 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
13346 // condition is true.
13347 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
13348 SDValue Ops[] = { Op2, Op1, CC, Cond };
13349 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
13352 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op, const X86Subtarget *Subtarget,
13353 SelectionDAG &DAG) {
13354 MVT VT = Op->getSimpleValueType(0);
13355 SDValue In = Op->getOperand(0);
13356 MVT InVT = In.getSimpleValueType();
13357 MVT VTElt = VT.getVectorElementType();
13358 MVT InVTElt = InVT.getVectorElementType();
13362 if ((InVTElt == MVT::i1) &&
13363 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
13364 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
13366 ((Subtarget->hasBWI() && VT.is512BitVector() &&
13367 VTElt.getSizeInBits() <= 16)) ||
13369 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
13370 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
13372 ((Subtarget->hasDQI() && VT.is512BitVector() &&
13373 VTElt.getSizeInBits() >= 32))))
13374 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13376 unsigned int NumElts = VT.getVectorNumElements();
13378 if (NumElts != 8 && NumElts != 16)
13381 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
13382 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
13383 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
13384 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13387 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13388 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13390 MVT ExtVT = (NumElts == 8) ? MVT::v8i64 : MVT::v16i32;
13391 Constant *C = ConstantInt::get(*DAG.getContext(),
13392 APInt::getAllOnesValue(ExtVT.getScalarType().getSizeInBits()));
13394 SDValue CP = DAG.getConstantPool(C, TLI.getPointerTy());
13395 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
13396 SDValue Ld = DAG.getLoad(ExtVT.getScalarType(), dl, DAG.getEntryNode(), CP,
13397 MachinePointerInfo::getConstantPool(),
13398 false, false, false, Alignment);
13399 SDValue Brcst = DAG.getNode(X86ISD::VBROADCASTM, dl, ExtVT, In, Ld);
13400 if (VT.is512BitVector())
13402 return DAG.getNode(X86ISD::VTRUNC, dl, VT, Brcst);
13405 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13406 SelectionDAG &DAG) {
13407 MVT VT = Op->getSimpleValueType(0);
13408 SDValue In = Op->getOperand(0);
13409 MVT InVT = In.getSimpleValueType();
13412 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
13413 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
13415 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
13416 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
13417 (VT != MVT::v16i16 || InVT != MVT::v16i8))
13420 if (Subtarget->hasInt256())
13421 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
13423 // Optimize vectors in AVX mode
13424 // Sign extend v8i16 to v8i32 and
13427 // Divide input vector into two parts
13428 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
13429 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
13430 // concat the vectors to original VT
13432 unsigned NumElems = InVT.getVectorNumElements();
13433 SDValue Undef = DAG.getUNDEF(InVT);
13435 SmallVector<int,8> ShufMask1(NumElems, -1);
13436 for (unsigned i = 0; i != NumElems/2; ++i)
13439 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
13441 SmallVector<int,8> ShufMask2(NumElems, -1);
13442 for (unsigned i = 0; i != NumElems/2; ++i)
13443 ShufMask2[i] = i + NumElems/2;
13445 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
13447 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
13448 VT.getVectorNumElements()/2);
13450 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
13451 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
13453 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13456 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
13457 // may emit an illegal shuffle but the expansion is still better than scalar
13458 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
13459 // we'll emit a shuffle and a arithmetic shift.
13460 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
13461 // TODO: It is possible to support ZExt by zeroing the undef values during
13462 // the shuffle phase or after the shuffle.
13463 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
13464 SelectionDAG &DAG) {
13465 MVT RegVT = Op.getSimpleValueType();
13466 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
13467 assert(RegVT.isInteger() &&
13468 "We only custom lower integer vector sext loads.");
13470 // Nothing useful we can do without SSE2 shuffles.
13471 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
13473 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
13475 EVT MemVT = Ld->getMemoryVT();
13476 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13477 unsigned RegSz = RegVT.getSizeInBits();
13479 ISD::LoadExtType Ext = Ld->getExtensionType();
13481 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
13482 && "Only anyext and sext are currently implemented.");
13483 assert(MemVT != RegVT && "Cannot extend to the same type");
13484 assert(MemVT.isVector() && "Must load a vector from memory");
13486 unsigned NumElems = RegVT.getVectorNumElements();
13487 unsigned MemSz = MemVT.getSizeInBits();
13488 assert(RegSz > MemSz && "Register size must be greater than the mem size");
13490 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
13491 // The only way in which we have a legal 256-bit vector result but not the
13492 // integer 256-bit operations needed to directly lower a sextload is if we
13493 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
13494 // a 128-bit vector and a normal sign_extend to 256-bits that should get
13495 // correctly legalized. We do this late to allow the canonical form of
13496 // sextload to persist throughout the rest of the DAG combiner -- it wants
13497 // to fold together any extensions it can, and so will fuse a sign_extend
13498 // of an sextload into a sextload targeting a wider value.
13500 if (MemSz == 128) {
13501 // Just switch this to a normal load.
13502 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
13503 "it must be a legal 128-bit vector "
13505 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
13506 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
13507 Ld->isInvariant(), Ld->getAlignment());
13509 assert(MemSz < 128 &&
13510 "Can't extend a type wider than 128 bits to a 256 bit vector!");
13511 // Do an sext load to a 128-bit vector type. We want to use the same
13512 // number of elements, but elements half as wide. This will end up being
13513 // recursively lowered by this routine, but will succeed as we definitely
13514 // have all the necessary features if we're using AVX1.
13516 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
13517 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
13519 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
13520 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
13521 Ld->isNonTemporal(), Ld->isInvariant(),
13522 Ld->getAlignment());
13525 // Replace chain users with the new chain.
13526 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
13527 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
13529 // Finally, do a normal sign-extend to the desired register.
13530 return DAG.getSExtOrTrunc(Load, dl, RegVT);
13533 // All sizes must be a power of two.
13534 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
13535 "Non-power-of-two elements are not custom lowered!");
13537 // Attempt to load the original value using scalar loads.
13538 // Find the largest scalar type that divides the total loaded size.
13539 MVT SclrLoadTy = MVT::i8;
13540 for (MVT Tp : MVT::integer_valuetypes()) {
13541 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
13546 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
13547 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
13549 SclrLoadTy = MVT::f64;
13551 // Calculate the number of scalar loads that we need to perform
13552 // in order to load our vector from memory.
13553 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
13555 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
13556 "Can only lower sext loads with a single scalar load!");
13558 unsigned loadRegZize = RegSz;
13559 if (Ext == ISD::SEXTLOAD && RegSz == 256)
13562 // Represent our vector as a sequence of elements which are the
13563 // largest scalar that we can load.
13564 EVT LoadUnitVecVT = EVT::getVectorVT(
13565 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
13567 // Represent the data using the same element type that is stored in
13568 // memory. In practice, we ''widen'' MemVT.
13570 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
13571 loadRegZize / MemVT.getScalarType().getSizeInBits());
13573 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
13574 "Invalid vector type");
13576 // We can't shuffle using an illegal type.
13577 assert(TLI.isTypeLegal(WideVecVT) &&
13578 "We only lower types that form legal widened vector types");
13580 SmallVector<SDValue, 8> Chains;
13581 SDValue Ptr = Ld->getBasePtr();
13582 SDValue Increment =
13583 DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
13584 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
13586 for (unsigned i = 0; i < NumLoads; ++i) {
13587 // Perform a single load.
13588 SDValue ScalarLoad =
13589 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
13590 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
13591 Ld->getAlignment());
13592 Chains.push_back(ScalarLoad.getValue(1));
13593 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
13594 // another round of DAGCombining.
13596 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
13598 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
13599 ScalarLoad, DAG.getIntPtrConstant(i));
13601 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
13604 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
13606 // Bitcast the loaded value to a vector of the original element type, in
13607 // the size of the target vector type.
13608 SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
13609 unsigned SizeRatio = RegSz / MemSz;
13611 if (Ext == ISD::SEXTLOAD) {
13612 // If we have SSE4.1, we can directly emit a VSEXT node.
13613 if (Subtarget->hasSSE41()) {
13614 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
13615 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13619 // Otherwise we'll shuffle the small elements in the high bits of the
13620 // larger type and perform an arithmetic shift. If the shift is not legal
13621 // it's better to scalarize.
13622 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
13623 "We can't implement a sext load without an arithmetic right shift!");
13625 // Redistribute the loaded elements into the different locations.
13626 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13627 for (unsigned i = 0; i != NumElems; ++i)
13628 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
13630 SDValue Shuff = DAG.getVectorShuffle(
13631 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13633 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13635 // Build the arithmetic shift.
13636 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
13637 MemVT.getVectorElementType().getSizeInBits();
13639 DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
13641 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13645 // Redistribute the loaded elements into the different locations.
13646 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
13647 for (unsigned i = 0; i != NumElems; ++i)
13648 ShuffleVec[i * SizeRatio] = i;
13650 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
13651 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
13653 // Bitcast to the requested type.
13654 Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
13655 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
13659 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
13660 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
13661 // from the AND / OR.
13662 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
13663 Opc = Op.getOpcode();
13664 if (Opc != ISD::OR && Opc != ISD::AND)
13666 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13667 Op.getOperand(0).hasOneUse() &&
13668 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
13669 Op.getOperand(1).hasOneUse());
13672 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
13673 // 1 and that the SETCC node has a single use.
13674 static bool isXor1OfSetCC(SDValue Op) {
13675 if (Op.getOpcode() != ISD::XOR)
13677 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
13678 if (N1C && N1C->getAPIntValue() == 1) {
13679 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
13680 Op.getOperand(0).hasOneUse();
13685 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
13686 bool addTest = true;
13687 SDValue Chain = Op.getOperand(0);
13688 SDValue Cond = Op.getOperand(1);
13689 SDValue Dest = Op.getOperand(2);
13692 bool Inverted = false;
13694 if (Cond.getOpcode() == ISD::SETCC) {
13695 // Check for setcc([su]{add,sub,mul}o == 0).
13696 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
13697 isa<ConstantSDNode>(Cond.getOperand(1)) &&
13698 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
13699 Cond.getOperand(0).getResNo() == 1 &&
13700 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
13701 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
13702 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
13703 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
13704 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
13705 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
13707 Cond = Cond.getOperand(0);
13709 SDValue NewCond = LowerSETCC(Cond, DAG);
13710 if (NewCond.getNode())
13715 // FIXME: LowerXALUO doesn't handle these!!
13716 else if (Cond.getOpcode() == X86ISD::ADD ||
13717 Cond.getOpcode() == X86ISD::SUB ||
13718 Cond.getOpcode() == X86ISD::SMUL ||
13719 Cond.getOpcode() == X86ISD::UMUL)
13720 Cond = LowerXALUO(Cond, DAG);
13723 // Look pass (and (setcc_carry (cmp ...)), 1).
13724 if (Cond.getOpcode() == ISD::AND &&
13725 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
13726 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
13727 if (C && C->getAPIntValue() == 1)
13728 Cond = Cond.getOperand(0);
13731 // If condition flag is set by a X86ISD::CMP, then use it as the condition
13732 // setting operand in place of the X86ISD::SETCC.
13733 unsigned CondOpcode = Cond.getOpcode();
13734 if (CondOpcode == X86ISD::SETCC ||
13735 CondOpcode == X86ISD::SETCC_CARRY) {
13736 CC = Cond.getOperand(0);
13738 SDValue Cmp = Cond.getOperand(1);
13739 unsigned Opc = Cmp.getOpcode();
13740 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
13741 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
13745 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
13749 // These can only come from an arithmetic instruction with overflow,
13750 // e.g. SADDO, UADDO.
13751 Cond = Cond.getNode()->getOperand(1);
13757 CondOpcode = Cond.getOpcode();
13758 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
13759 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
13760 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
13761 Cond.getOperand(0).getValueType() != MVT::i8)) {
13762 SDValue LHS = Cond.getOperand(0);
13763 SDValue RHS = Cond.getOperand(1);
13764 unsigned X86Opcode;
13767 // Keep this in sync with LowerXALUO, otherwise we might create redundant
13768 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
13770 switch (CondOpcode) {
13771 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
13773 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13775 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
13778 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
13779 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
13781 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
13783 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
13786 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
13787 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
13788 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
13789 default: llvm_unreachable("unexpected overflowing operator");
13792 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
13793 if (CondOpcode == ISD::UMULO)
13794 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
13797 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
13799 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
13801 if (CondOpcode == ISD::UMULO)
13802 Cond = X86Op.getValue(2);
13804 Cond = X86Op.getValue(1);
13806 CC = DAG.getConstant(X86Cond, MVT::i8);
13810 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
13811 SDValue Cmp = Cond.getOperand(0).getOperand(1);
13812 if (CondOpc == ISD::OR) {
13813 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
13814 // two branches instead of an explicit OR instruction with a
13816 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13817 isX86LogicalCmp(Cmp)) {
13818 CC = Cond.getOperand(0).getOperand(0);
13819 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13820 Chain, Dest, CC, Cmp);
13821 CC = Cond.getOperand(1).getOperand(0);
13825 } else { // ISD::AND
13826 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
13827 // two branches instead of an explicit AND instruction with a
13828 // separate test. However, we only do this if this block doesn't
13829 // have a fall-through edge, because this requires an explicit
13830 // jmp when the condition is false.
13831 if (Cmp == Cond.getOperand(1).getOperand(1) &&
13832 isX86LogicalCmp(Cmp) &&
13833 Op.getNode()->hasOneUse()) {
13834 X86::CondCode CCode =
13835 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13836 CCode = X86::GetOppositeBranchCondition(CCode);
13837 CC = DAG.getConstant(CCode, MVT::i8);
13838 SDNode *User = *Op.getNode()->use_begin();
13839 // Look for an unconditional branch following this conditional branch.
13840 // We need this because we need to reverse the successors in order
13841 // to implement FCMP_OEQ.
13842 if (User->getOpcode() == ISD::BR) {
13843 SDValue FalseBB = User->getOperand(1);
13845 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13846 assert(NewBR == User);
13850 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13851 Chain, Dest, CC, Cmp);
13852 X86::CondCode CCode =
13853 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
13854 CCode = X86::GetOppositeBranchCondition(CCode);
13855 CC = DAG.getConstant(CCode, MVT::i8);
13861 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
13862 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
13863 // It should be transformed during dag combiner except when the condition
13864 // is set by a arithmetics with overflow node.
13865 X86::CondCode CCode =
13866 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
13867 CCode = X86::GetOppositeBranchCondition(CCode);
13868 CC = DAG.getConstant(CCode, MVT::i8);
13869 Cond = Cond.getOperand(0).getOperand(1);
13871 } else if (Cond.getOpcode() == ISD::SETCC &&
13872 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
13873 // For FCMP_OEQ, we can emit
13874 // two branches instead of an explicit AND instruction with a
13875 // separate test. However, we only do this if this block doesn't
13876 // have a fall-through edge, because this requires an explicit
13877 // jmp when the condition is false.
13878 if (Op.getNode()->hasOneUse()) {
13879 SDNode *User = *Op.getNode()->use_begin();
13880 // Look for an unconditional branch following this conditional branch.
13881 // We need this because we need to reverse the successors in order
13882 // to implement FCMP_OEQ.
13883 if (User->getOpcode() == ISD::BR) {
13884 SDValue FalseBB = User->getOperand(1);
13886 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13887 assert(NewBR == User);
13891 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13892 Cond.getOperand(0), Cond.getOperand(1));
13893 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13894 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13895 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13896 Chain, Dest, CC, Cmp);
13897 CC = DAG.getConstant(X86::COND_P, MVT::i8);
13902 } else if (Cond.getOpcode() == ISD::SETCC &&
13903 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
13904 // For FCMP_UNE, we can emit
13905 // two branches instead of an explicit AND instruction with a
13906 // separate test. However, we only do this if this block doesn't
13907 // have a fall-through edge, because this requires an explicit
13908 // jmp when the condition is false.
13909 if (Op.getNode()->hasOneUse()) {
13910 SDNode *User = *Op.getNode()->use_begin();
13911 // Look for an unconditional branch following this conditional branch.
13912 // We need this because we need to reverse the successors in order
13913 // to implement FCMP_UNE.
13914 if (User->getOpcode() == ISD::BR) {
13915 SDValue FalseBB = User->getOperand(1);
13917 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
13918 assert(NewBR == User);
13921 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
13922 Cond.getOperand(0), Cond.getOperand(1));
13923 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
13924 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
13925 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13926 Chain, Dest, CC, Cmp);
13927 CC = DAG.getConstant(X86::COND_NP, MVT::i8);
13937 // Look pass the truncate if the high bits are known zero.
13938 if (isTruncWithZeroHighBitsInput(Cond, DAG))
13939 Cond = Cond.getOperand(0);
13941 // We know the result of AND is compared against zero. Try to match
13943 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
13944 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
13945 if (NewSetCC.getNode()) {
13946 CC = NewSetCC.getOperand(0);
13947 Cond = NewSetCC.getOperand(1);
13954 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
13955 CC = DAG.getConstant(X86Cond, MVT::i8);
13956 Cond = EmitTest(Cond, X86Cond, dl, DAG);
13958 Cond = ConvertCmpIfNecessary(Cond, DAG);
13959 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
13960 Chain, Dest, CC, Cond);
13963 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
13964 // Calls to _alloca are needed to probe the stack when allocating more than 4k
13965 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
13966 // that the guard pages used by the OS virtual memory manager are allocated in
13967 // correct sequence.
13969 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
13970 SelectionDAG &DAG) const {
13971 MachineFunction &MF = DAG.getMachineFunction();
13972 bool SplitStack = MF.shouldSplitStack();
13973 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
13978 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13979 SDNode* Node = Op.getNode();
13981 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
13982 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
13983 " not tell us which reg is the stack pointer!");
13984 EVT VT = Node->getValueType(0);
13985 SDValue Tmp1 = SDValue(Node, 0);
13986 SDValue Tmp2 = SDValue(Node, 1);
13987 SDValue Tmp3 = Node->getOperand(2);
13988 SDValue Chain = Tmp1.getOperand(0);
13990 // Chain the dynamic stack allocation so that it doesn't modify the stack
13991 // pointer when other instructions are using the stack.
13992 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
13995 SDValue Size = Tmp2.getOperand(1);
13996 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
13997 Chain = SP.getValue(1);
13998 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
13999 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14000 unsigned StackAlign = TFI.getStackAlignment();
14001 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14002 if (Align > StackAlign)
14003 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14004 DAG.getConstant(-(uint64_t)Align, VT));
14005 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14007 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
14008 DAG.getIntPtrConstant(0, true), SDValue(),
14011 SDValue Ops[2] = { Tmp1, Tmp2 };
14012 return DAG.getMergeValues(Ops, dl);
14016 SDValue Chain = Op.getOperand(0);
14017 SDValue Size = Op.getOperand(1);
14018 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14019 EVT VT = Op.getNode()->getValueType(0);
14021 bool Is64Bit = Subtarget->is64Bit();
14022 EVT SPTy = getPointerTy();
14025 MachineRegisterInfo &MRI = MF.getRegInfo();
14028 // The 64 bit implementation of segmented stacks needs to clobber both r10
14029 // r11. This makes it impossible to use it along with nested parameters.
14030 const Function *F = MF.getFunction();
14032 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14034 if (I->hasNestAttr())
14035 report_fatal_error("Cannot use segmented stacks with functions that "
14036 "have nested arguments.");
14039 const TargetRegisterClass *AddrRegClass =
14040 getRegClassFor(getPointerTy());
14041 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14042 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14043 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14044 DAG.getRegister(Vreg, SPTy));
14045 SDValue Ops1[2] = { Value, Chain };
14046 return DAG.getMergeValues(Ops1, dl);
14049 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14051 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14052 Flag = Chain.getValue(1);
14053 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14055 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14057 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14058 unsigned SPReg = RegInfo->getStackRegister();
14059 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14060 Chain = SP.getValue(1);
14063 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14064 DAG.getConstant(-(uint64_t)Align, VT));
14065 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14068 SDValue Ops1[2] = { SP, Chain };
14069 return DAG.getMergeValues(Ops1, dl);
14073 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14074 MachineFunction &MF = DAG.getMachineFunction();
14075 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14077 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14080 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14081 // vastart just stores the address of the VarArgsFrameIndex slot into the
14082 // memory location argument.
14083 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14085 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14086 MachinePointerInfo(SV), false, false, 0);
14090 // gp_offset (0 - 6 * 8)
14091 // fp_offset (48 - 48 + 8 * 16)
14092 // overflow_arg_area (point to parameters coming in memory).
14094 SmallVector<SDValue, 8> MemOps;
14095 SDValue FIN = Op.getOperand(1);
14097 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
14098 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
14100 FIN, MachinePointerInfo(SV), false, false, 0);
14101 MemOps.push_back(Store);
14104 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14105 FIN, DAG.getIntPtrConstant(4));
14106 Store = DAG.getStore(Op.getOperand(0), DL,
14107 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
14109 FIN, MachinePointerInfo(SV, 4), false, false, 0);
14110 MemOps.push_back(Store);
14112 // Store ptr to overflow_arg_area
14113 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14114 FIN, DAG.getIntPtrConstant(4));
14115 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
14117 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
14118 MachinePointerInfo(SV, 8),
14120 MemOps.push_back(Store);
14122 // Store ptr to reg_save_area.
14123 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
14124 FIN, DAG.getIntPtrConstant(8));
14125 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
14127 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
14128 MachinePointerInfo(SV, 16), false, false, 0);
14129 MemOps.push_back(Store);
14130 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
14133 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
14134 assert(Subtarget->is64Bit() &&
14135 "LowerVAARG only handles 64-bit va_arg!");
14136 assert((Subtarget->isTargetLinux() ||
14137 Subtarget->isTargetDarwin()) &&
14138 "Unhandled target in LowerVAARG");
14139 assert(Op.getNode()->getNumOperands() == 4);
14140 SDValue Chain = Op.getOperand(0);
14141 SDValue SrcPtr = Op.getOperand(1);
14142 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14143 unsigned Align = Op.getConstantOperandVal(3);
14146 EVT ArgVT = Op.getNode()->getValueType(0);
14147 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
14148 uint32_t ArgSize = getDataLayout()->getTypeAllocSize(ArgTy);
14151 // Decide which area this value should be read from.
14152 // TODO: Implement the AMD64 ABI in its entirety. This simple
14153 // selection mechanism works only for the basic types.
14154 if (ArgVT == MVT::f80) {
14155 llvm_unreachable("va_arg for f80 not yet implemented");
14156 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
14157 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
14158 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
14159 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
14161 llvm_unreachable("Unhandled argument type in LowerVAARG");
14164 if (ArgMode == 2) {
14165 // Sanity Check: Make sure using fp_offset makes sense.
14166 assert(!DAG.getTarget().Options.UseSoftFloat &&
14167 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
14168 Attribute::NoImplicitFloat)) &&
14169 Subtarget->hasSSE1());
14172 // Insert VAARG_64 node into the DAG
14173 // VAARG_64 returns two values: Variable Argument Address, Chain
14174 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, MVT::i32),
14175 DAG.getConstant(ArgMode, MVT::i8),
14176 DAG.getConstant(Align, MVT::i32)};
14177 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
14178 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
14179 VTs, InstOps, MVT::i64,
14180 MachinePointerInfo(SV),
14182 /*Volatile=*/false,
14184 /*WriteMem=*/true);
14185 Chain = VAARG.getValue(1);
14187 // Load the next argument and return it
14188 return DAG.getLoad(ArgVT, dl,
14191 MachinePointerInfo(),
14192 false, false, false, 0);
14195 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
14196 SelectionDAG &DAG) {
14197 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
14198 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
14199 SDValue Chain = Op.getOperand(0);
14200 SDValue DstPtr = Op.getOperand(1);
14201 SDValue SrcPtr = Op.getOperand(2);
14202 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
14203 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
14206 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
14207 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
14209 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
14212 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
14213 // amount is a constant. Takes immediate version of shift as input.
14214 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
14215 SDValue SrcOp, uint64_t ShiftAmt,
14216 SelectionDAG &DAG) {
14217 MVT ElementType = VT.getVectorElementType();
14219 // Fold this packed shift into its first operand if ShiftAmt is 0.
14223 // Check for ShiftAmt >= element width
14224 if (ShiftAmt >= ElementType.getSizeInBits()) {
14225 if (Opc == X86ISD::VSRAI)
14226 ShiftAmt = ElementType.getSizeInBits() - 1;
14228 return DAG.getConstant(0, VT);
14231 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
14232 && "Unknown target vector shift-by-constant node");
14234 // Fold this packed vector shift into a build vector if SrcOp is a
14235 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
14236 if (VT == SrcOp.getSimpleValueType() &&
14237 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
14238 SmallVector<SDValue, 8> Elts;
14239 unsigned NumElts = SrcOp->getNumOperands();
14240 ConstantSDNode *ND;
14243 default: llvm_unreachable(nullptr);
14244 case X86ISD::VSHLI:
14245 for (unsigned i=0; i!=NumElts; ++i) {
14246 SDValue CurrentOp = SrcOp->getOperand(i);
14247 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14248 Elts.push_back(CurrentOp);
14251 ND = cast<ConstantSDNode>(CurrentOp);
14252 const APInt &C = ND->getAPIntValue();
14253 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), ElementType));
14256 case X86ISD::VSRLI:
14257 for (unsigned i=0; i!=NumElts; ++i) {
14258 SDValue CurrentOp = SrcOp->getOperand(i);
14259 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14260 Elts.push_back(CurrentOp);
14263 ND = cast<ConstantSDNode>(CurrentOp);
14264 const APInt &C = ND->getAPIntValue();
14265 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), ElementType));
14268 case X86ISD::VSRAI:
14269 for (unsigned i=0; i!=NumElts; ++i) {
14270 SDValue CurrentOp = SrcOp->getOperand(i);
14271 if (CurrentOp->getOpcode() == ISD::UNDEF) {
14272 Elts.push_back(CurrentOp);
14275 ND = cast<ConstantSDNode>(CurrentOp);
14276 const APInt &C = ND->getAPIntValue();
14277 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), ElementType));
14282 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
14285 return DAG.getNode(Opc, dl, VT, SrcOp, DAG.getConstant(ShiftAmt, MVT::i8));
14288 // getTargetVShiftNode - Handle vector element shifts where the shift amount
14289 // may or may not be a constant. Takes immediate version of shift as input.
14290 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
14291 SDValue SrcOp, SDValue ShAmt,
14292 SelectionDAG &DAG) {
14293 MVT SVT = ShAmt.getSimpleValueType();
14294 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
14296 // Catch shift-by-constant.
14297 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
14298 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
14299 CShAmt->getZExtValue(), DAG);
14301 // Change opcode to non-immediate version
14303 default: llvm_unreachable("Unknown target vector shift node");
14304 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
14305 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
14306 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
14309 const X86Subtarget &Subtarget =
14310 static_cast<const X86Subtarget &>(DAG.getSubtarget());
14311 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
14312 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
14313 // Let the shuffle legalizer expand this shift amount node.
14314 SDValue Op0 = ShAmt.getOperand(0);
14315 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
14316 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
14318 // Need to build a vector containing shift amount.
14319 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
14320 SmallVector<SDValue, 4> ShOps;
14321 ShOps.push_back(ShAmt);
14322 if (SVT == MVT::i32) {
14323 ShOps.push_back(DAG.getConstant(0, SVT));
14324 ShOps.push_back(DAG.getUNDEF(SVT));
14326 ShOps.push_back(DAG.getUNDEF(SVT));
14328 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
14329 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
14332 // The return type has to be a 128-bit type with the same element
14333 // type as the input type.
14334 MVT EltVT = VT.getVectorElementType();
14335 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
14337 ShAmt = DAG.getNode(ISD::BITCAST, dl, ShVT, ShAmt);
14338 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
14341 /// \brief Return (and \p Op, \p Mask) for compare instructions or
14342 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
14343 /// necessary casting for \p Mask when lowering masking intrinsics.
14344 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
14345 SDValue PreservedSrc,
14346 const X86Subtarget *Subtarget,
14347 SelectionDAG &DAG) {
14348 EVT VT = Op.getValueType();
14349 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
14350 MVT::i1, VT.getVectorNumElements());
14351 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14352 Mask.getValueType().getSizeInBits());
14355 assert(MaskVT.isSimple() && "invalid mask type");
14357 if (isAllOnes(Mask))
14360 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
14361 // are extracted by EXTRACT_SUBVECTOR.
14362 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14363 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14364 DAG.getIntPtrConstant(0));
14366 switch (Op.getOpcode()) {
14368 case X86ISD::PCMPEQM:
14369 case X86ISD::PCMPGTM:
14371 case X86ISD::CMPMU:
14372 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
14374 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14375 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14376 return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
14379 /// \brief Creates an SDNode for a predicated scalar operation.
14380 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
14381 /// The mask is comming as MVT::i8 and it should be truncated
14382 /// to MVT::i1 while lowering masking intrinsics.
14383 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
14384 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
14385 /// a scalar instruction.
14386 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
14387 SDValue PreservedSrc,
14388 const X86Subtarget *Subtarget,
14389 SelectionDAG &DAG) {
14390 if (isAllOnes(Mask))
14393 EVT VT = Op.getValueType();
14395 // The mask should be of type MVT::i1
14396 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
14398 if (PreservedSrc.getOpcode() == ISD::UNDEF)
14399 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
14400 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
14403 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14404 SelectionDAG &DAG) {
14406 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
14407 EVT VT = Op.getValueType();
14408 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
14410 switch(IntrData->Type) {
14411 case INTR_TYPE_1OP:
14412 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
14413 case INTR_TYPE_2OP:
14414 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14416 case INTR_TYPE_3OP:
14417 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
14418 Op.getOperand(2), Op.getOperand(3));
14419 case INTR_TYPE_1OP_MASK_RM: {
14420 SDValue Src = Op.getOperand(1);
14421 SDValue Src0 = Op.getOperand(2);
14422 SDValue Mask = Op.getOperand(3);
14423 SDValue RoundingMode = Op.getOperand(4);
14424 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
14426 Mask, Src0, Subtarget, DAG);
14428 case INTR_TYPE_SCALAR_MASK_RM: {
14429 SDValue Src1 = Op.getOperand(1);
14430 SDValue Src2 = Op.getOperand(2);
14431 SDValue Src0 = Op.getOperand(3);
14432 SDValue Mask = Op.getOperand(4);
14433 // There are 2 kinds of intrinsics in this group:
14434 // (1) With supress-all-exceptions (sae) - 6 operands
14435 // (2) With rounding mode and sae - 7 operands.
14436 if (Op.getNumOperands() == 6) {
14437 SDValue Sae = Op.getOperand(5);
14438 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14440 Mask, Src0, Subtarget, DAG);
14442 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
14443 SDValue RoundingMode = Op.getOperand(5);
14444 SDValue Sae = Op.getOperand(6);
14445 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
14446 RoundingMode, Sae),
14447 Mask, Src0, Subtarget, DAG);
14449 case INTR_TYPE_2OP_MASK: {
14450 SDValue Src1 = Op.getOperand(1);
14451 SDValue Src2 = Op.getOperand(2);
14452 SDValue PassThru = Op.getOperand(3);
14453 SDValue Mask = Op.getOperand(4);
14454 // We specify 2 possible opcodes for intrinsics with rounding modes.
14455 // First, we check if the intrinsic may have non-default rounding mode,
14456 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14457 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14458 if (IntrWithRoundingModeOpcode != 0) {
14459 SDValue Rnd = Op.getOperand(5);
14460 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
14461 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
14462 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14463 dl, Op.getValueType(),
14465 Mask, PassThru, Subtarget, DAG);
14468 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
14470 Mask, PassThru, Subtarget, DAG);
14472 case FMA_OP_MASK: {
14473 SDValue Src1 = Op.getOperand(1);
14474 SDValue Src2 = Op.getOperand(2);
14475 SDValue Src3 = Op.getOperand(3);
14476 SDValue Mask = Op.getOperand(4);
14477 // We specify 2 possible opcodes for intrinsics with rounding modes.
14478 // First, we check if the intrinsic may have non-default rounding mode,
14479 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
14480 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
14481 if (IntrWithRoundingModeOpcode != 0) {
14482 SDValue Rnd = Op.getOperand(5);
14483 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
14484 X86::STATIC_ROUNDING::CUR_DIRECTION)
14485 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
14486 dl, Op.getValueType(),
14487 Src1, Src2, Src3, Rnd),
14488 Mask, Src1, Subtarget, DAG);
14490 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
14491 dl, Op.getValueType(),
14493 Mask, Src1, Subtarget, DAG);
14496 case CMP_MASK_CC: {
14497 // Comparison intrinsics with masks.
14498 // Example of transformation:
14499 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
14500 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
14502 // (v8i1 (insert_subvector undef,
14503 // (v2i1 (and (PCMPEQM %a, %b),
14504 // (extract_subvector
14505 // (v8i1 (bitcast %mask)), 0))), 0))))
14506 EVT VT = Op.getOperand(1).getValueType();
14507 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14508 VT.getVectorNumElements());
14509 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
14510 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14511 Mask.getValueType().getSizeInBits());
14513 if (IntrData->Type == CMP_MASK_CC) {
14514 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14515 Op.getOperand(2), Op.getOperand(3));
14517 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
14518 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
14521 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
14522 DAG.getTargetConstant(0, MaskVT),
14524 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
14525 DAG.getUNDEF(BitcastVT), CmpMask,
14526 DAG.getIntPtrConstant(0));
14527 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Res);
14529 case COMI: { // Comparison intrinsics
14530 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
14531 SDValue LHS = Op.getOperand(1);
14532 SDValue RHS = Op.getOperand(2);
14533 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
14534 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
14535 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
14536 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14537 DAG.getConstant(X86CC, MVT::i8), Cond);
14538 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14541 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
14542 Op.getOperand(1), Op.getOperand(2), DAG);
14544 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
14545 Op.getSimpleValueType(),
14547 Op.getOperand(2), DAG),
14548 Op.getOperand(4), Op.getOperand(3), Subtarget,
14550 case COMPRESS_EXPAND_IN_REG: {
14551 SDValue Mask = Op.getOperand(3);
14552 SDValue DataToCompress = Op.getOperand(1);
14553 SDValue PassThru = Op.getOperand(2);
14554 if (isAllOnes(Mask)) // return data as is
14555 return Op.getOperand(1);
14556 EVT VT = Op.getValueType();
14557 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14558 VT.getVectorNumElements());
14559 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14560 Mask.getValueType().getSizeInBits());
14562 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14563 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14564 DAG.getIntPtrConstant(0));
14566 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToCompress,
14570 SDValue Mask = Op.getOperand(3);
14571 EVT VT = Op.getValueType();
14572 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14573 VT.getVectorNumElements());
14574 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
14575 Mask.getValueType().getSizeInBits());
14577 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
14578 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
14579 DAG.getIntPtrConstant(0));
14580 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
14589 default: return SDValue(); // Don't custom lower most intrinsics.
14591 case Intrinsic::x86_avx512_mask_valign_q_512:
14592 case Intrinsic::x86_avx512_mask_valign_d_512:
14593 // Vector source operands are swapped.
14594 return getVectorMaskingNode(DAG.getNode(X86ISD::VALIGN, dl,
14595 Op.getValueType(), Op.getOperand(2),
14598 Op.getOperand(5), Op.getOperand(4),
14601 // ptest and testp intrinsics. The intrinsic these come from are designed to
14602 // return an integer value, not just an instruction so lower it to the ptest
14603 // or testp pattern and a setcc for the result.
14604 case Intrinsic::x86_sse41_ptestz:
14605 case Intrinsic::x86_sse41_ptestc:
14606 case Intrinsic::x86_sse41_ptestnzc:
14607 case Intrinsic::x86_avx_ptestz_256:
14608 case Intrinsic::x86_avx_ptestc_256:
14609 case Intrinsic::x86_avx_ptestnzc_256:
14610 case Intrinsic::x86_avx_vtestz_ps:
14611 case Intrinsic::x86_avx_vtestc_ps:
14612 case Intrinsic::x86_avx_vtestnzc_ps:
14613 case Intrinsic::x86_avx_vtestz_pd:
14614 case Intrinsic::x86_avx_vtestc_pd:
14615 case Intrinsic::x86_avx_vtestnzc_pd:
14616 case Intrinsic::x86_avx_vtestz_ps_256:
14617 case Intrinsic::x86_avx_vtestc_ps_256:
14618 case Intrinsic::x86_avx_vtestnzc_ps_256:
14619 case Intrinsic::x86_avx_vtestz_pd_256:
14620 case Intrinsic::x86_avx_vtestc_pd_256:
14621 case Intrinsic::x86_avx_vtestnzc_pd_256: {
14622 bool IsTestPacked = false;
14625 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
14626 case Intrinsic::x86_avx_vtestz_ps:
14627 case Intrinsic::x86_avx_vtestz_pd:
14628 case Intrinsic::x86_avx_vtestz_ps_256:
14629 case Intrinsic::x86_avx_vtestz_pd_256:
14630 IsTestPacked = true; // Fallthrough
14631 case Intrinsic::x86_sse41_ptestz:
14632 case Intrinsic::x86_avx_ptestz_256:
14634 X86CC = X86::COND_E;
14636 case Intrinsic::x86_avx_vtestc_ps:
14637 case Intrinsic::x86_avx_vtestc_pd:
14638 case Intrinsic::x86_avx_vtestc_ps_256:
14639 case Intrinsic::x86_avx_vtestc_pd_256:
14640 IsTestPacked = true; // Fallthrough
14641 case Intrinsic::x86_sse41_ptestc:
14642 case Intrinsic::x86_avx_ptestc_256:
14644 X86CC = X86::COND_B;
14646 case Intrinsic::x86_avx_vtestnzc_ps:
14647 case Intrinsic::x86_avx_vtestnzc_pd:
14648 case Intrinsic::x86_avx_vtestnzc_ps_256:
14649 case Intrinsic::x86_avx_vtestnzc_pd_256:
14650 IsTestPacked = true; // Fallthrough
14651 case Intrinsic::x86_sse41_ptestnzc:
14652 case Intrinsic::x86_avx_ptestnzc_256:
14654 X86CC = X86::COND_A;
14658 SDValue LHS = Op.getOperand(1);
14659 SDValue RHS = Op.getOperand(2);
14660 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
14661 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
14662 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14663 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
14664 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14666 case Intrinsic::x86_avx512_kortestz_w:
14667 case Intrinsic::x86_avx512_kortestc_w: {
14668 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
14669 SDValue LHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(1));
14670 SDValue RHS = DAG.getNode(ISD::BITCAST, dl, MVT::v16i1, Op.getOperand(2));
14671 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
14672 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
14673 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
14674 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14677 case Intrinsic::x86_sse42_pcmpistria128:
14678 case Intrinsic::x86_sse42_pcmpestria128:
14679 case Intrinsic::x86_sse42_pcmpistric128:
14680 case Intrinsic::x86_sse42_pcmpestric128:
14681 case Intrinsic::x86_sse42_pcmpistrio128:
14682 case Intrinsic::x86_sse42_pcmpestrio128:
14683 case Intrinsic::x86_sse42_pcmpistris128:
14684 case Intrinsic::x86_sse42_pcmpestris128:
14685 case Intrinsic::x86_sse42_pcmpistriz128:
14686 case Intrinsic::x86_sse42_pcmpestriz128: {
14690 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
14691 case Intrinsic::x86_sse42_pcmpistria128:
14692 Opcode = X86ISD::PCMPISTRI;
14693 X86CC = X86::COND_A;
14695 case Intrinsic::x86_sse42_pcmpestria128:
14696 Opcode = X86ISD::PCMPESTRI;
14697 X86CC = X86::COND_A;
14699 case Intrinsic::x86_sse42_pcmpistric128:
14700 Opcode = X86ISD::PCMPISTRI;
14701 X86CC = X86::COND_B;
14703 case Intrinsic::x86_sse42_pcmpestric128:
14704 Opcode = X86ISD::PCMPESTRI;
14705 X86CC = X86::COND_B;
14707 case Intrinsic::x86_sse42_pcmpistrio128:
14708 Opcode = X86ISD::PCMPISTRI;
14709 X86CC = X86::COND_O;
14711 case Intrinsic::x86_sse42_pcmpestrio128:
14712 Opcode = X86ISD::PCMPESTRI;
14713 X86CC = X86::COND_O;
14715 case Intrinsic::x86_sse42_pcmpistris128:
14716 Opcode = X86ISD::PCMPISTRI;
14717 X86CC = X86::COND_S;
14719 case Intrinsic::x86_sse42_pcmpestris128:
14720 Opcode = X86ISD::PCMPESTRI;
14721 X86CC = X86::COND_S;
14723 case Intrinsic::x86_sse42_pcmpistriz128:
14724 Opcode = X86ISD::PCMPISTRI;
14725 X86CC = X86::COND_E;
14727 case Intrinsic::x86_sse42_pcmpestriz128:
14728 Opcode = X86ISD::PCMPESTRI;
14729 X86CC = X86::COND_E;
14732 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14733 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14734 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
14735 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14736 DAG.getConstant(X86CC, MVT::i8),
14737 SDValue(PCMP.getNode(), 1));
14738 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
14741 case Intrinsic::x86_sse42_pcmpistri128:
14742 case Intrinsic::x86_sse42_pcmpestri128: {
14744 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
14745 Opcode = X86ISD::PCMPISTRI;
14747 Opcode = X86ISD::PCMPESTRI;
14749 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
14750 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
14751 return DAG.getNode(Opcode, dl, VTs, NewOps);
14756 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14757 SDValue Src, SDValue Mask, SDValue Base,
14758 SDValue Index, SDValue ScaleOp, SDValue Chain,
14759 const X86Subtarget * Subtarget) {
14761 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14762 assert(C && "Invalid scale type");
14763 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14764 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14765 Index.getSimpleValueType().getVectorNumElements());
14767 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14769 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14771 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14772 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
14773 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14774 SDValue Segment = DAG.getRegister(0, MVT::i32);
14775 if (Src.getOpcode() == ISD::UNDEF)
14776 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
14777 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14778 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14779 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
14780 return DAG.getMergeValues(RetOps, dl);
14783 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14784 SDValue Src, SDValue Mask, SDValue Base,
14785 SDValue Index, SDValue ScaleOp, SDValue Chain) {
14787 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14788 assert(C && "Invalid scale type");
14789 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14790 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14791 SDValue Segment = DAG.getRegister(0, MVT::i32);
14792 EVT MaskVT = MVT::getVectorVT(MVT::i1,
14793 Index.getSimpleValueType().getVectorNumElements());
14795 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14797 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14799 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14800 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
14801 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
14802 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
14803 return SDValue(Res, 1);
14806 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
14807 SDValue Mask, SDValue Base, SDValue Index,
14808 SDValue ScaleOp, SDValue Chain) {
14810 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
14811 assert(C && "Invalid scale type");
14812 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), MVT::i8);
14813 SDValue Disp = DAG.getTargetConstant(0, MVT::i32);
14814 SDValue Segment = DAG.getRegister(0, MVT::i32);
14816 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
14818 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
14820 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), MaskVT);
14822 MaskInReg = DAG.getNode(ISD::BITCAST, dl, MaskVT, Mask);
14823 //SDVTList VTs = DAG.getVTList(MVT::Other);
14824 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
14825 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
14826 return SDValue(Res, 0);
14829 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
14830 // read performance monitor counters (x86_rdpmc).
14831 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
14832 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14833 SmallVectorImpl<SDValue> &Results) {
14834 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14835 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14838 // The ECX register is used to select the index of the performance counter
14840 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
14842 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
14844 // Reads the content of a 64-bit performance counter and returns it in the
14845 // registers EDX:EAX.
14846 if (Subtarget->is64Bit()) {
14847 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14848 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14851 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14852 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14855 Chain = HI.getValue(1);
14857 if (Subtarget->is64Bit()) {
14858 // The EAX register is loaded with the low-order 32 bits. The EDX register
14859 // is loaded with the supported high-order bits of the counter.
14860 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14861 DAG.getConstant(32, MVT::i8));
14862 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14863 Results.push_back(Chain);
14867 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14868 SDValue Ops[] = { LO, HI };
14869 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14870 Results.push_back(Pair);
14871 Results.push_back(Chain);
14874 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
14875 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
14876 // also used to custom lower READCYCLECOUNTER nodes.
14877 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
14878 SelectionDAG &DAG, const X86Subtarget *Subtarget,
14879 SmallVectorImpl<SDValue> &Results) {
14880 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
14881 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
14884 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
14885 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
14886 // and the EAX register is loaded with the low-order 32 bits.
14887 if (Subtarget->is64Bit()) {
14888 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
14889 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
14892 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
14893 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
14896 SDValue Chain = HI.getValue(1);
14898 if (Opcode == X86ISD::RDTSCP_DAG) {
14899 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
14901 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
14902 // the ECX register. Add 'ecx' explicitly to the chain.
14903 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
14905 // Explicitly store the content of ECX at the location passed in input
14906 // to the 'rdtscp' intrinsic.
14907 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
14908 MachinePointerInfo(), false, false, 0);
14911 if (Subtarget->is64Bit()) {
14912 // The EDX register is loaded with the high-order 32 bits of the MSR, and
14913 // the EAX register is loaded with the low-order 32 bits.
14914 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
14915 DAG.getConstant(32, MVT::i8));
14916 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
14917 Results.push_back(Chain);
14921 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
14922 SDValue Ops[] = { LO, HI };
14923 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
14924 Results.push_back(Pair);
14925 Results.push_back(Chain);
14928 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
14929 SelectionDAG &DAG) {
14930 SmallVector<SDValue, 2> Results;
14932 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
14934 return DAG.getMergeValues(Results, DL);
14938 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
14939 SelectionDAG &DAG) {
14940 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
14942 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
14947 switch(IntrData->Type) {
14949 llvm_unreachable("Unknown Intrinsic Type");
14953 // Emit the node with the right value type.
14954 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
14955 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
14957 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
14958 // Otherwise return the value from Rand, which is always 0, casted to i32.
14959 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
14960 DAG.getConstant(1, Op->getValueType(1)),
14961 DAG.getConstant(X86::COND_B, MVT::i32),
14962 SDValue(Result.getNode(), 1) };
14963 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
14964 DAG.getVTList(Op->getValueType(1), MVT::Glue),
14967 // Return { result, isValid, chain }.
14968 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
14969 SDValue(Result.getNode(), 2));
14972 //gather(v1, mask, index, base, scale);
14973 SDValue Chain = Op.getOperand(0);
14974 SDValue Src = Op.getOperand(2);
14975 SDValue Base = Op.getOperand(3);
14976 SDValue Index = Op.getOperand(4);
14977 SDValue Mask = Op.getOperand(5);
14978 SDValue Scale = Op.getOperand(6);
14979 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain,
14983 //scatter(base, mask, index, v1, scale);
14984 SDValue Chain = Op.getOperand(0);
14985 SDValue Base = Op.getOperand(2);
14986 SDValue Mask = Op.getOperand(3);
14987 SDValue Index = Op.getOperand(4);
14988 SDValue Src = Op.getOperand(5);
14989 SDValue Scale = Op.getOperand(6);
14990 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
14993 SDValue Hint = Op.getOperand(6);
14995 if (dyn_cast<ConstantSDNode> (Hint) == nullptr ||
14996 (HintVal = dyn_cast<ConstantSDNode> (Hint)->getZExtValue()) > 1)
14997 llvm_unreachable("Wrong prefetch hint in intrinsic: should be 0 or 1");
14998 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
14999 SDValue Chain = Op.getOperand(0);
15000 SDValue Mask = Op.getOperand(2);
15001 SDValue Index = Op.getOperand(3);
15002 SDValue Base = Op.getOperand(4);
15003 SDValue Scale = Op.getOperand(5);
15004 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
15006 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
15008 SmallVector<SDValue, 2> Results;
15009 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget, Results);
15010 return DAG.getMergeValues(Results, dl);
15012 // Read Performance Monitoring Counters.
15014 SmallVector<SDValue, 2> Results;
15015 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
15016 return DAG.getMergeValues(Results, dl);
15018 // XTEST intrinsics.
15020 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15021 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
15022 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15023 DAG.getConstant(X86::COND_NE, MVT::i8),
15025 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
15026 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
15027 Ret, SDValue(InTrans.getNode(), 1));
15031 SmallVector<SDValue, 2> Results;
15032 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
15033 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
15034 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
15035 DAG.getConstant(-1, MVT::i8));
15036 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
15037 Op.getOperand(4), GenCF.getValue(1));
15038 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
15039 Op.getOperand(5), MachinePointerInfo(),
15041 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15042 DAG.getConstant(X86::COND_B, MVT::i8),
15044 Results.push_back(SetCC);
15045 Results.push_back(Store);
15046 return DAG.getMergeValues(Results, dl);
15048 case COMPRESS_TO_MEM: {
15050 SDValue Mask = Op.getOperand(4);
15051 SDValue DataToCompress = Op.getOperand(3);
15052 SDValue Addr = Op.getOperand(2);
15053 SDValue Chain = Op.getOperand(0);
15055 if (isAllOnes(Mask)) // return just a store
15056 return DAG.getStore(Chain, dl, DataToCompress, Addr,
15057 MachinePointerInfo(), false, false, 0);
15059 EVT VT = DataToCompress.getValueType();
15060 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15061 VT.getVectorNumElements());
15062 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15063 Mask.getValueType().getSizeInBits());
15064 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15065 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15066 DAG.getIntPtrConstant(0));
15068 SDValue Compressed = DAG.getNode(IntrData->Opc0, dl, VT, VMask,
15069 DataToCompress, DAG.getUNDEF(VT));
15070 return DAG.getStore(Chain, dl, Compressed, Addr,
15071 MachinePointerInfo(), false, false, 0);
15073 case EXPAND_FROM_MEM: {
15075 SDValue Mask = Op.getOperand(4);
15076 SDValue PathThru = Op.getOperand(3);
15077 SDValue Addr = Op.getOperand(2);
15078 SDValue Chain = Op.getOperand(0);
15079 EVT VT = Op.getValueType();
15081 if (isAllOnes(Mask)) // return just a load
15082 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
15084 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15085 VT.getVectorNumElements());
15086 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15087 Mask.getValueType().getSizeInBits());
15088 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15089 DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
15090 DAG.getIntPtrConstant(0));
15092 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
15093 false, false, false, 0);
15095 SDValue Results[] = {
15096 DAG.getNode(IntrData->Opc0, dl, VT, VMask, DataToExpand, PathThru),
15098 return DAG.getMergeValues(Results, dl);
15103 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
15104 SelectionDAG &DAG) const {
15105 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
15106 MFI->setReturnAddressIsTaken(true);
15108 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
15111 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15113 EVT PtrVT = getPointerTy();
15116 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
15117 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15118 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
15119 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15120 DAG.getNode(ISD::ADD, dl, PtrVT,
15121 FrameAddr, Offset),
15122 MachinePointerInfo(), false, false, false, 0);
15125 // Just load the return address.
15126 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
15127 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
15128 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
15131 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
15132 MachineFunction &MF = DAG.getMachineFunction();
15133 MachineFrameInfo *MFI = MF.getFrameInfo();
15134 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15135 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15136 EVT VT = Op.getValueType();
15138 MFI->setFrameAddressIsTaken(true);
15140 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
15141 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
15142 // is not possible to crawl up the stack without looking at the unwind codes
15144 int FrameAddrIndex = FuncInfo->getFAIndex();
15145 if (!FrameAddrIndex) {
15146 // Set up a frame object for the return address.
15147 unsigned SlotSize = RegInfo->getSlotSize();
15148 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
15149 SlotSize, /*Offset=*/INT64_MIN, /*IsImmutable=*/false);
15150 FuncInfo->setFAIndex(FrameAddrIndex);
15152 return DAG.getFrameIndex(FrameAddrIndex, VT);
15155 unsigned FrameReg =
15156 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
15157 SDLoc dl(Op); // FIXME probably not meaningful
15158 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15159 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
15160 (FrameReg == X86::EBP && VT == MVT::i32)) &&
15161 "Invalid Frame Register!");
15162 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
15164 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
15165 MachinePointerInfo(),
15166 false, false, false, 0);
15170 // FIXME? Maybe this could be a TableGen attribute on some registers and
15171 // this table could be generated automatically from RegInfo.
15172 unsigned X86TargetLowering::getRegisterByName(const char* RegName,
15174 unsigned Reg = StringSwitch<unsigned>(RegName)
15175 .Case("esp", X86::ESP)
15176 .Case("rsp", X86::RSP)
15180 report_fatal_error("Invalid register name global variable");
15183 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
15184 SelectionDAG &DAG) const {
15185 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15186 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
15189 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
15190 SDValue Chain = Op.getOperand(0);
15191 SDValue Offset = Op.getOperand(1);
15192 SDValue Handler = Op.getOperand(2);
15195 EVT PtrVT = getPointerTy();
15196 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15197 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
15198 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
15199 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
15200 "Invalid Frame Register!");
15201 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
15202 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
15204 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
15205 DAG.getIntPtrConstant(RegInfo->getSlotSize()));
15206 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
15207 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
15209 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
15211 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
15212 DAG.getRegister(StoreAddrReg, PtrVT));
15215 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
15216 SelectionDAG &DAG) const {
15218 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
15219 DAG.getVTList(MVT::i32, MVT::Other),
15220 Op.getOperand(0), Op.getOperand(1));
15223 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
15224 SelectionDAG &DAG) const {
15226 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
15227 Op.getOperand(0), Op.getOperand(1));
15230 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
15231 return Op.getOperand(0);
15234 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
15235 SelectionDAG &DAG) const {
15236 SDValue Root = Op.getOperand(0);
15237 SDValue Trmp = Op.getOperand(1); // trampoline
15238 SDValue FPtr = Op.getOperand(2); // nested function
15239 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
15242 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15243 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
15245 if (Subtarget->is64Bit()) {
15246 SDValue OutChains[6];
15248 // Large code-model.
15249 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
15250 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
15252 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
15253 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
15255 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
15257 // Load the pointer to the nested function into R11.
15258 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
15259 SDValue Addr = Trmp;
15260 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15261 Addr, MachinePointerInfo(TrmpAddr),
15264 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15265 DAG.getConstant(2, MVT::i64));
15266 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
15267 MachinePointerInfo(TrmpAddr, 2),
15270 // Load the 'nest' parameter value into R10.
15271 // R10 is specified in X86CallingConv.td
15272 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
15273 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15274 DAG.getConstant(10, MVT::i64));
15275 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15276 Addr, MachinePointerInfo(TrmpAddr, 10),
15279 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15280 DAG.getConstant(12, MVT::i64));
15281 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
15282 MachinePointerInfo(TrmpAddr, 12),
15285 // Jump to the nested function.
15286 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
15287 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15288 DAG.getConstant(20, MVT::i64));
15289 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
15290 Addr, MachinePointerInfo(TrmpAddr, 20),
15293 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
15294 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
15295 DAG.getConstant(22, MVT::i64));
15296 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
15297 MachinePointerInfo(TrmpAddr, 22),
15300 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15302 const Function *Func =
15303 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
15304 CallingConv::ID CC = Func->getCallingConv();
15309 llvm_unreachable("Unsupported calling convention");
15310 case CallingConv::C:
15311 case CallingConv::X86_StdCall: {
15312 // Pass 'nest' parameter in ECX.
15313 // Must be kept in sync with X86CallingConv.td
15314 NestReg = X86::ECX;
15316 // Check that ECX wasn't needed by an 'inreg' parameter.
15317 FunctionType *FTy = Func->getFunctionType();
15318 const AttributeSet &Attrs = Func->getAttributes();
15320 if (!Attrs.isEmpty() && !Func->isVarArg()) {
15321 unsigned InRegCount = 0;
15324 for (FunctionType::param_iterator I = FTy->param_begin(),
15325 E = FTy->param_end(); I != E; ++I, ++Idx)
15326 if (Attrs.hasAttribute(Idx, Attribute::InReg))
15327 // FIXME: should only count parameters that are lowered to integers.
15328 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
15330 if (InRegCount > 2) {
15331 report_fatal_error("Nest register in use - reduce number of inreg"
15337 case CallingConv::X86_FastCall:
15338 case CallingConv::X86_ThisCall:
15339 case CallingConv::Fast:
15340 // Pass 'nest' parameter in EAX.
15341 // Must be kept in sync with X86CallingConv.td
15342 NestReg = X86::EAX;
15346 SDValue OutChains[4];
15347 SDValue Addr, Disp;
15349 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15350 DAG.getConstant(10, MVT::i32));
15351 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
15353 // This is storing the opcode for MOV32ri.
15354 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
15355 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
15356 OutChains[0] = DAG.getStore(Root, dl,
15357 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
15358 Trmp, MachinePointerInfo(TrmpAddr),
15361 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15362 DAG.getConstant(1, MVT::i32));
15363 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
15364 MachinePointerInfo(TrmpAddr, 1),
15367 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
15368 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15369 DAG.getConstant(5, MVT::i32));
15370 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
15371 MachinePointerInfo(TrmpAddr, 5),
15374 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
15375 DAG.getConstant(6, MVT::i32));
15376 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
15377 MachinePointerInfo(TrmpAddr, 6),
15380 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
15384 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
15385 SelectionDAG &DAG) const {
15387 The rounding mode is in bits 11:10 of FPSR, and has the following
15389 00 Round to nearest
15394 FLT_ROUNDS, on the other hand, expects the following:
15401 To perform the conversion, we do:
15402 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
15405 MachineFunction &MF = DAG.getMachineFunction();
15406 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15407 unsigned StackAlignment = TFI.getStackAlignment();
15408 MVT VT = Op.getSimpleValueType();
15411 // Save FP Control Word to stack slot
15412 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
15413 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
15415 MachineMemOperand *MMO =
15416 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
15417 MachineMemOperand::MOStore, 2, 2);
15419 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
15420 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
15421 DAG.getVTList(MVT::Other),
15422 Ops, MVT::i16, MMO);
15424 // Load FP Control Word from stack slot
15425 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
15426 MachinePointerInfo(), false, false, false, 0);
15428 // Transform as necessary
15430 DAG.getNode(ISD::SRL, DL, MVT::i16,
15431 DAG.getNode(ISD::AND, DL, MVT::i16,
15432 CWD, DAG.getConstant(0x800, MVT::i16)),
15433 DAG.getConstant(11, MVT::i8));
15435 DAG.getNode(ISD::SRL, DL, MVT::i16,
15436 DAG.getNode(ISD::AND, DL, MVT::i16,
15437 CWD, DAG.getConstant(0x400, MVT::i16)),
15438 DAG.getConstant(9, MVT::i8));
15441 DAG.getNode(ISD::AND, DL, MVT::i16,
15442 DAG.getNode(ISD::ADD, DL, MVT::i16,
15443 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
15444 DAG.getConstant(1, MVT::i16)),
15445 DAG.getConstant(3, MVT::i16));
15447 return DAG.getNode((VT.getSizeInBits() < 16 ?
15448 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
15451 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
15452 MVT VT = Op.getSimpleValueType();
15454 unsigned NumBits = VT.getSizeInBits();
15457 Op = Op.getOperand(0);
15458 if (VT == MVT::i8) {
15459 // Zero extend to i32 since there is not an i8 bsr.
15461 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15464 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
15465 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15466 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15468 // If src is zero (i.e. bsr sets ZF), returns NumBits.
15471 DAG.getConstant(NumBits+NumBits-1, OpVT),
15472 DAG.getConstant(X86::COND_E, MVT::i8),
15475 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
15477 // Finally xor with NumBits-1.
15478 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15481 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15485 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
15486 MVT VT = Op.getSimpleValueType();
15488 unsigned NumBits = VT.getSizeInBits();
15491 Op = Op.getOperand(0);
15492 if (VT == MVT::i8) {
15493 // Zero extend to i32 since there is not an i8 bsr.
15495 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
15498 // Issue a bsr (scan bits in reverse).
15499 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
15500 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
15502 // And xor with NumBits-1.
15503 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
15506 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
15510 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
15511 MVT VT = Op.getSimpleValueType();
15512 unsigned NumBits = VT.getSizeInBits();
15514 Op = Op.getOperand(0);
15516 // Issue a bsf (scan bits forward) which also sets EFLAGS.
15517 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
15518 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
15520 // If src is zero (i.e. bsf sets ZF), returns NumBits.
15523 DAG.getConstant(NumBits, VT),
15524 DAG.getConstant(X86::COND_E, MVT::i8),
15527 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
15530 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
15531 // ones, and then concatenate the result back.
15532 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
15533 MVT VT = Op.getSimpleValueType();
15535 assert(VT.is256BitVector() && VT.isInteger() &&
15536 "Unsupported value type for operation");
15538 unsigned NumElems = VT.getVectorNumElements();
15541 // Extract the LHS vectors
15542 SDValue LHS = Op.getOperand(0);
15543 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
15544 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
15546 // Extract the RHS vectors
15547 SDValue RHS = Op.getOperand(1);
15548 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
15549 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
15551 MVT EltVT = VT.getVectorElementType();
15552 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
15554 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
15555 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
15556 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
15559 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
15560 assert(Op.getSimpleValueType().is256BitVector() &&
15561 Op.getSimpleValueType().isInteger() &&
15562 "Only handle AVX 256-bit vector integer operation");
15563 return Lower256IntArith(Op, DAG);
15566 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
15567 assert(Op.getSimpleValueType().is256BitVector() &&
15568 Op.getSimpleValueType().isInteger() &&
15569 "Only handle AVX 256-bit vector integer operation");
15570 return Lower256IntArith(Op, DAG);
15573 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
15574 SelectionDAG &DAG) {
15576 MVT VT = Op.getSimpleValueType();
15578 // Decompose 256-bit ops into smaller 128-bit ops.
15579 if (VT.is256BitVector() && !Subtarget->hasInt256())
15580 return Lower256IntArith(Op, DAG);
15582 SDValue A = Op.getOperand(0);
15583 SDValue B = Op.getOperand(1);
15585 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
15586 if (VT == MVT::v4i32) {
15587 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
15588 "Should not custom lower when pmuldq is available!");
15590 // Extract the odd parts.
15591 static const int UnpackMask[] = { 1, -1, 3, -1 };
15592 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
15593 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
15595 // Multiply the even parts.
15596 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
15597 // Now multiply odd parts.
15598 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
15600 Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
15601 Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
15603 // Merge the two vectors back together with a shuffle. This expands into 2
15605 static const int ShufMask[] = { 0, 4, 2, 6 };
15606 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
15609 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
15610 "Only know how to lower V2I64/V4I64/V8I64 multiply");
15612 // Ahi = psrlqi(a, 32);
15613 // Bhi = psrlqi(b, 32);
15615 // AloBlo = pmuludq(a, b);
15616 // AloBhi = pmuludq(a, Bhi);
15617 // AhiBlo = pmuludq(Ahi, b);
15619 // AloBhi = psllqi(AloBhi, 32);
15620 // AhiBlo = psllqi(AhiBlo, 32);
15621 // return AloBlo + AloBhi + AhiBlo;
15623 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
15624 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
15626 // Bit cast to 32-bit vectors for MULUDQ
15627 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
15628 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
15629 A = DAG.getNode(ISD::BITCAST, dl, MulVT, A);
15630 B = DAG.getNode(ISD::BITCAST, dl, MulVT, B);
15631 Ahi = DAG.getNode(ISD::BITCAST, dl, MulVT, Ahi);
15632 Bhi = DAG.getNode(ISD::BITCAST, dl, MulVT, Bhi);
15634 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
15635 SDValue AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
15636 SDValue AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
15638 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
15639 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
15641 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
15642 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
15645 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
15646 assert(Subtarget->isTargetWin64() && "Unexpected target");
15647 EVT VT = Op.getValueType();
15648 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
15649 "Unexpected return type for lowering");
15653 switch (Op->getOpcode()) {
15654 default: llvm_unreachable("Unexpected request for libcall!");
15655 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
15656 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
15657 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
15658 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
15659 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
15660 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
15664 SDValue InChain = DAG.getEntryNode();
15666 TargetLowering::ArgListTy Args;
15667 TargetLowering::ArgListEntry Entry;
15668 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
15669 EVT ArgVT = Op->getOperand(i).getValueType();
15670 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
15671 "Unexpected argument type for lowering");
15672 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
15673 Entry.Node = StackPtr;
15674 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
15676 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15677 Entry.Ty = PointerType::get(ArgTy,0);
15678 Entry.isSExt = false;
15679 Entry.isZExt = false;
15680 Args.push_back(Entry);
15683 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
15686 TargetLowering::CallLoweringInfo CLI(DAG);
15687 CLI.setDebugLoc(dl).setChain(InChain)
15688 .setCallee(getLibcallCallingConv(LC),
15689 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
15690 Callee, std::move(Args), 0)
15691 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
15693 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
15694 return DAG.getNode(ISD::BITCAST, dl, VT, CallInfo.first);
15697 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
15698 SelectionDAG &DAG) {
15699 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
15700 EVT VT = Op0.getValueType();
15703 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
15704 (VT == MVT::v8i32 && Subtarget->hasInt256()));
15706 // PMULxD operations multiply each even value (starting at 0) of LHS with
15707 // the related value of RHS and produce a widen result.
15708 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15709 // => <2 x i64> <ae|cg>
15711 // In other word, to have all the results, we need to perform two PMULxD:
15712 // 1. one with the even values.
15713 // 2. one with the odd values.
15714 // To achieve #2, with need to place the odd values at an even position.
15716 // Place the odd value at an even position (basically, shift all values 1
15717 // step to the left):
15718 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
15719 // <a|b|c|d> => <b|undef|d|undef>
15720 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
15721 // <e|f|g|h> => <f|undef|h|undef>
15722 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
15724 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
15726 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
15727 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
15729 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
15730 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
15731 // => <2 x i64> <ae|cg>
15732 SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
15733 DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
15734 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
15735 // => <2 x i64> <bf|dh>
15736 SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
15737 DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
15739 // Shuffle it back into the right order.
15740 SDValue Highs, Lows;
15741 if (VT == MVT::v8i32) {
15742 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
15743 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15744 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
15745 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15747 const int HighMask[] = {1, 5, 3, 7};
15748 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
15749 const int LowMask[] = {0, 4, 2, 6};
15750 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
15753 // If we have a signed multiply but no PMULDQ fix up the high parts of a
15754 // unsigned multiply.
15755 if (IsSigned && !Subtarget->hasSSE41()) {
15757 DAG.getConstant(31, DAG.getTargetLoweringInfo().getShiftAmountTy(VT));
15758 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
15759 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
15760 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
15761 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
15763 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
15764 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
15767 // The first result of MUL_LOHI is actually the low value, followed by the
15769 SDValue Ops[] = {Lows, Highs};
15770 return DAG.getMergeValues(Ops, dl);
15773 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
15774 const X86Subtarget *Subtarget) {
15775 MVT VT = Op.getSimpleValueType();
15777 SDValue R = Op.getOperand(0);
15778 SDValue Amt = Op.getOperand(1);
15780 // Optimize shl/srl/sra with constant shift amount.
15781 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
15782 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
15783 uint64_t ShiftAmt = ShiftConst->getZExtValue();
15785 if (VT == MVT::v2i64 || VT == MVT::v4i32 || VT == MVT::v8i16 ||
15786 (Subtarget->hasInt256() &&
15787 (VT == MVT::v4i64 || VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15788 (Subtarget->hasAVX512() &&
15789 (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15790 if (Op.getOpcode() == ISD::SHL)
15791 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15793 if (Op.getOpcode() == ISD::SRL)
15794 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15796 if (Op.getOpcode() == ISD::SRA && VT != MVT::v2i64 && VT != MVT::v4i64)
15797 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15801 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
15802 unsigned NumElts = VT.getVectorNumElements();
15803 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
15805 if (Op.getOpcode() == ISD::SHL) {
15806 // Make a large shift.
15807 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
15809 SHL = DAG.getNode(ISD::BITCAST, dl, VT, SHL);
15810 // Zero out the rightmost bits.
15811 SmallVector<SDValue, 32> V(
15812 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), MVT::i8));
15813 return DAG.getNode(ISD::AND, dl, VT, SHL,
15814 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15816 if (Op.getOpcode() == ISD::SRL) {
15817 // Make a large shift.
15818 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
15820 SRL = DAG.getNode(ISD::BITCAST, dl, VT, SRL);
15821 // Zero out the leftmost bits.
15822 SmallVector<SDValue, 32> V(
15823 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, MVT::i8));
15824 return DAG.getNode(ISD::AND, dl, VT, SRL,
15825 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
15827 if (Op.getOpcode() == ISD::SRA) {
15828 if (ShiftAmt == 7) {
15829 // R s>> 7 === R s< 0
15830 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
15831 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
15834 // R s>> a === ((R u>> a) ^ m) - m
15835 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
15836 SmallVector<SDValue, 32> V(NumElts,
15837 DAG.getConstant(128 >> ShiftAmt, MVT::i8));
15838 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
15839 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
15840 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
15843 llvm_unreachable("Unknown shift opcode.");
15848 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15849 if (!Subtarget->is64Bit() &&
15850 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
15851 Amt.getOpcode() == ISD::BITCAST &&
15852 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
15853 Amt = Amt.getOperand(0);
15854 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
15855 VT.getVectorNumElements();
15856 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
15857 uint64_t ShiftAmt = 0;
15858 for (unsigned i = 0; i != Ratio; ++i) {
15859 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
15863 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
15865 // Check remaining shift amounts.
15866 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
15867 uint64_t ShAmt = 0;
15868 for (unsigned j = 0; j != Ratio; ++j) {
15869 ConstantSDNode *C =
15870 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
15874 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
15876 if (ShAmt != ShiftAmt)
15879 switch (Op.getOpcode()) {
15881 llvm_unreachable("Unknown shift opcode!");
15883 return getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, R, ShiftAmt,
15886 return getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt,
15889 return getTargetVShiftByConstNode(X86ISD::VSRAI, dl, VT, R, ShiftAmt,
15897 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
15898 const X86Subtarget* Subtarget) {
15899 MVT VT = Op.getSimpleValueType();
15901 SDValue R = Op.getOperand(0);
15902 SDValue Amt = Op.getOperand(1);
15904 if ((VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) ||
15905 VT == MVT::v4i32 || VT == MVT::v8i16 ||
15906 (Subtarget->hasInt256() &&
15907 ((VT == MVT::v4i64 && Op.getOpcode() != ISD::SRA) ||
15908 VT == MVT::v8i32 || VT == MVT::v16i16)) ||
15909 (Subtarget->hasAVX512() && (VT == MVT::v8i64 || VT == MVT::v16i32))) {
15911 EVT EltVT = VT.getVectorElementType();
15913 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
15914 // Check if this build_vector node is doing a splat.
15915 // If so, then set BaseShAmt equal to the splat value.
15916 BaseShAmt = BV->getSplatValue();
15917 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
15918 BaseShAmt = SDValue();
15920 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
15921 Amt = Amt.getOperand(0);
15923 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
15924 if (SVN && SVN->isSplat()) {
15925 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
15926 SDValue InVec = Amt.getOperand(0);
15927 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
15928 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
15929 "Unexpected shuffle index found!");
15930 BaseShAmt = InVec.getOperand(SplatIdx);
15931 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
15932 if (ConstantSDNode *C =
15933 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
15934 if (C->getZExtValue() == SplatIdx)
15935 BaseShAmt = InVec.getOperand(1);
15940 // Avoid introducing an extract element from a shuffle.
15941 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
15942 DAG.getIntPtrConstant(SplatIdx));
15946 if (BaseShAmt.getNode()) {
15947 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
15948 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
15949 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
15950 else if (EltVT.bitsLT(MVT::i32))
15951 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
15953 switch (Op.getOpcode()) {
15955 llvm_unreachable("Unknown shift opcode!");
15957 switch (VT.SimpleTy) {
15958 default: return SDValue();
15967 return getTargetVShiftNode(X86ISD::VSHLI, dl, VT, R, BaseShAmt, DAG);
15970 switch (VT.SimpleTy) {
15971 default: return SDValue();
15978 return getTargetVShiftNode(X86ISD::VSRAI, dl, VT, R, BaseShAmt, DAG);
15981 switch (VT.SimpleTy) {
15982 default: return SDValue();
15991 return getTargetVShiftNode(X86ISD::VSRLI, dl, VT, R, BaseShAmt, DAG);
15997 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
15998 if (!Subtarget->is64Bit() &&
15999 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
16000 (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
16001 Amt.getOpcode() == ISD::BITCAST &&
16002 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
16003 Amt = Amt.getOperand(0);
16004 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
16005 VT.getVectorNumElements();
16006 std::vector<SDValue> Vals(Ratio);
16007 for (unsigned i = 0; i != Ratio; ++i)
16008 Vals[i] = Amt.getOperand(i);
16009 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
16010 for (unsigned j = 0; j != Ratio; ++j)
16011 if (Vals[j] != Amt.getOperand(i + j))
16014 switch (Op.getOpcode()) {
16016 llvm_unreachable("Unknown shift opcode!");
16018 return DAG.getNode(X86ISD::VSHL, dl, VT, R, Op.getOperand(1));
16020 return DAG.getNode(X86ISD::VSRL, dl, VT, R, Op.getOperand(1));
16022 return DAG.getNode(X86ISD::VSRA, dl, VT, R, Op.getOperand(1));
16029 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
16030 SelectionDAG &DAG) {
16031 MVT VT = Op.getSimpleValueType();
16033 SDValue R = Op.getOperand(0);
16034 SDValue Amt = Op.getOperand(1);
16037 assert(VT.isVector() && "Custom lowering only for vector shifts!");
16038 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
16040 V = LowerScalarImmediateShift(Op, DAG, Subtarget);
16044 V = LowerScalarVariableShift(Op, DAG, Subtarget);
16048 if (Subtarget->hasAVX512() && (VT == MVT::v16i32 || VT == MVT::v8i64))
16050 // AVX2 has VPSLLV/VPSRAV/VPSRLV.
16051 if (Subtarget->hasInt256()) {
16052 if (Op.getOpcode() == ISD::SRL &&
16053 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16054 VT == MVT::v4i64 || VT == MVT::v8i32))
16056 if (Op.getOpcode() == ISD::SHL &&
16057 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
16058 VT == MVT::v4i64 || VT == MVT::v8i32))
16060 if (Op.getOpcode() == ISD::SRA && (VT == MVT::v4i32 || VT == MVT::v8i32))
16064 // If possible, lower this packed shift into a vector multiply instead of
16065 // expanding it into a sequence of scalar shifts.
16066 // Do this only if the vector shift count is a constant build_vector.
16067 if (Op.getOpcode() == ISD::SHL &&
16068 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
16069 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
16070 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16071 SmallVector<SDValue, 8> Elts;
16072 EVT SVT = VT.getScalarType();
16073 unsigned SVTBits = SVT.getSizeInBits();
16074 const APInt &One = APInt(SVTBits, 1);
16075 unsigned NumElems = VT.getVectorNumElements();
16077 for (unsigned i=0; i !=NumElems; ++i) {
16078 SDValue Op = Amt->getOperand(i);
16079 if (Op->getOpcode() == ISD::UNDEF) {
16080 Elts.push_back(Op);
16084 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
16085 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
16086 uint64_t ShAmt = C.getZExtValue();
16087 if (ShAmt >= SVTBits) {
16088 Elts.push_back(DAG.getUNDEF(SVT));
16091 Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
16093 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
16094 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
16097 // Lower SHL with variable shift amount.
16098 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
16099 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
16101 Op = DAG.getNode(ISD::ADD, dl, VT, Op, DAG.getConstant(0x3f800000U, VT));
16102 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
16103 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
16104 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
16107 // If possible, lower this shift as a sequence of two shifts by
16108 // constant plus a MOVSS/MOVSD instead of scalarizing it.
16110 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
16112 // Could be rewritten as:
16113 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
16115 // The advantage is that the two shifts from the example would be
16116 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
16117 // the vector shift into four scalar shifts plus four pairs of vector
16119 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
16120 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
16121 unsigned TargetOpcode = X86ISD::MOVSS;
16122 bool CanBeSimplified;
16123 // The splat value for the first packed shift (the 'X' from the example).
16124 SDValue Amt1 = Amt->getOperand(0);
16125 // The splat value for the second packed shift (the 'Y' from the example).
16126 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
16127 Amt->getOperand(2);
16129 // See if it is possible to replace this node with a sequence of
16130 // two shifts followed by a MOVSS/MOVSD
16131 if (VT == MVT::v4i32) {
16132 // Check if it is legal to use a MOVSS.
16133 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
16134 Amt2 == Amt->getOperand(3);
16135 if (!CanBeSimplified) {
16136 // Otherwise, check if we can still simplify this node using a MOVSD.
16137 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
16138 Amt->getOperand(2) == Amt->getOperand(3);
16139 TargetOpcode = X86ISD::MOVSD;
16140 Amt2 = Amt->getOperand(2);
16143 // Do similar checks for the case where the machine value type
16145 CanBeSimplified = Amt1 == Amt->getOperand(1);
16146 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
16147 CanBeSimplified = Amt2 == Amt->getOperand(i);
16149 if (!CanBeSimplified) {
16150 TargetOpcode = X86ISD::MOVSD;
16151 CanBeSimplified = true;
16152 Amt2 = Amt->getOperand(4);
16153 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
16154 CanBeSimplified = Amt1 == Amt->getOperand(i);
16155 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
16156 CanBeSimplified = Amt2 == Amt->getOperand(j);
16160 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
16161 isa<ConstantSDNode>(Amt2)) {
16162 // Replace this node with two shifts followed by a MOVSS/MOVSD.
16163 EVT CastVT = MVT::v4i32;
16165 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
16166 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
16168 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
16169 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
16170 if (TargetOpcode == X86ISD::MOVSD)
16171 CastVT = MVT::v2i64;
16172 SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
16173 SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
16174 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
16176 return DAG.getNode(ISD::BITCAST, dl, VT, Result);
16180 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
16181 assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
16184 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(5, VT));
16185 Op = DAG.getNode(ISD::BITCAST, dl, VT, Op);
16187 // Turn 'a' into a mask suitable for VSELECT
16188 SDValue VSelM = DAG.getConstant(0x80, VT);
16189 SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16190 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16192 SDValue CM1 = DAG.getConstant(0x0f, VT);
16193 SDValue CM2 = DAG.getConstant(0x3f, VT);
16195 // r = VSELECT(r, psllw(r & (char16)15, 4), a);
16196 SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
16197 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 4, DAG);
16198 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16199 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16202 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16203 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16204 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16206 // r = VSELECT(r, psllw(r & (char16)63, 2), a);
16207 M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
16208 M = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, MVT::v8i16, M, 2, DAG);
16209 M = DAG.getNode(ISD::BITCAST, dl, VT, M);
16210 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
16213 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
16214 OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
16215 OpVSel = DAG.getNode(X86ISD::PCMPEQ, dl, VT, OpVSel, VSelM);
16217 // return VSELECT(r, r+r, a);
16218 R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
16219 DAG.getNode(ISD::ADD, dl, VT, R, R), R);
16223 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
16224 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
16225 // solution better.
16226 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
16227 MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
16229 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
16230 R = DAG.getNode(ExtOpc, dl, NewVT, R);
16231 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
16232 return DAG.getNode(ISD::TRUNCATE, dl, VT,
16233 DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
16236 // Decompose 256-bit shifts into smaller 128-bit shifts.
16237 if (VT.is256BitVector()) {
16238 unsigned NumElems = VT.getVectorNumElements();
16239 MVT EltVT = VT.getVectorElementType();
16240 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16242 // Extract the two vectors
16243 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
16244 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
16246 // Recreate the shift amount vectors
16247 SDValue Amt1, Amt2;
16248 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
16249 // Constant shift amount
16250 SmallVector<SDValue, 4> Amt1Csts;
16251 SmallVector<SDValue, 4> Amt2Csts;
16252 for (unsigned i = 0; i != NumElems/2; ++i)
16253 Amt1Csts.push_back(Amt->getOperand(i));
16254 for (unsigned i = NumElems/2; i != NumElems; ++i)
16255 Amt2Csts.push_back(Amt->getOperand(i));
16257 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
16258 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
16260 // Variable shift amount
16261 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
16262 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
16265 // Issue new vector shifts for the smaller types
16266 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
16267 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
16269 // Concatenate the result back
16270 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
16276 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
16277 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
16278 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
16279 // looks for this combo and may remove the "setcc" instruction if the "setcc"
16280 // has only one use.
16281 SDNode *N = Op.getNode();
16282 SDValue LHS = N->getOperand(0);
16283 SDValue RHS = N->getOperand(1);
16284 unsigned BaseOp = 0;
16287 switch (Op.getOpcode()) {
16288 default: llvm_unreachable("Unknown ovf instruction!");
16290 // A subtract of one will be selected as a INC. Note that INC doesn't
16291 // set CF, so we can't do this for UADDO.
16292 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16294 BaseOp = X86ISD::INC;
16295 Cond = X86::COND_O;
16298 BaseOp = X86ISD::ADD;
16299 Cond = X86::COND_O;
16302 BaseOp = X86ISD::ADD;
16303 Cond = X86::COND_B;
16306 // A subtract of one will be selected as a DEC. Note that DEC doesn't
16307 // set CF, so we can't do this for USUBO.
16308 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
16310 BaseOp = X86ISD::DEC;
16311 Cond = X86::COND_O;
16314 BaseOp = X86ISD::SUB;
16315 Cond = X86::COND_O;
16318 BaseOp = X86ISD::SUB;
16319 Cond = X86::COND_B;
16322 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
16323 Cond = X86::COND_O;
16325 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
16326 if (N->getValueType(0) == MVT::i8) {
16327 BaseOp = X86ISD::UMUL8;
16328 Cond = X86::COND_O;
16331 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
16333 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
16336 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
16337 DAG.getConstant(X86::COND_O, MVT::i32),
16338 SDValue(Sum.getNode(), 2));
16340 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16344 // Also sets EFLAGS.
16345 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
16346 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
16349 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
16350 DAG.getConstant(Cond, MVT::i32),
16351 SDValue(Sum.getNode(), 1));
16353 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
16356 /// Returns true if the operand type is exactly twice the native width, and
16357 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
16358 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
16359 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
16360 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
16361 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
16364 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
16365 else if (OpWidth == 128)
16366 return Subtarget->hasCmpxchg16b();
16371 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
16372 return needsCmpXchgNb(SI->getValueOperand()->getType());
16375 // Note: this turns large loads into lock cmpxchg8b/16b.
16376 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
16377 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
16378 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
16379 return needsCmpXchgNb(PTy->getElementType());
16382 bool X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
16383 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16384 const Type *MemType = AI->getType();
16386 // If the operand is too big, we must see if cmpxchg8/16b is available
16387 // and default to library calls otherwise.
16388 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16389 return needsCmpXchgNb(MemType);
16391 AtomicRMWInst::BinOp Op = AI->getOperation();
16394 llvm_unreachable("Unknown atomic operation");
16395 case AtomicRMWInst::Xchg:
16396 case AtomicRMWInst::Add:
16397 case AtomicRMWInst::Sub:
16398 // It's better to use xadd, xsub or xchg for these in all cases.
16400 case AtomicRMWInst::Or:
16401 case AtomicRMWInst::And:
16402 case AtomicRMWInst::Xor:
16403 // If the atomicrmw's result isn't actually used, we can just add a "lock"
16404 // prefix to a normal instruction for these operations.
16405 return !AI->use_empty();
16406 case AtomicRMWInst::Nand:
16407 case AtomicRMWInst::Max:
16408 case AtomicRMWInst::Min:
16409 case AtomicRMWInst::UMax:
16410 case AtomicRMWInst::UMin:
16411 // These always require a non-trivial set of data operations on x86. We must
16412 // use a cmpxchg loop.
16417 static bool hasMFENCE(const X86Subtarget& Subtarget) {
16418 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
16419 // no-sse2). There isn't any reason to disable it if the target processor
16421 return Subtarget.hasSSE2() || Subtarget.is64Bit();
16425 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
16426 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
16427 const Type *MemType = AI->getType();
16428 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
16429 // there is no benefit in turning such RMWs into loads, and it is actually
16430 // harmful as it introduces a mfence.
16431 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
16434 auto Builder = IRBuilder<>(AI);
16435 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
16436 auto SynchScope = AI->getSynchScope();
16437 // We must restrict the ordering to avoid generating loads with Release or
16438 // ReleaseAcquire orderings.
16439 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
16440 auto Ptr = AI->getPointerOperand();
16442 // Before the load we need a fence. Here is an example lifted from
16443 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
16446 // x.store(1, relaxed);
16447 // r1 = y.fetch_add(0, release);
16449 // y.fetch_add(42, acquire);
16450 // r2 = x.load(relaxed);
16451 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
16452 // lowered to just a load without a fence. A mfence flushes the store buffer,
16453 // making the optimization clearly correct.
16454 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
16455 // otherwise, we might be able to be more agressive on relaxed idempotent
16456 // rmw. In practice, they do not look useful, so we don't try to be
16457 // especially clever.
16458 if (SynchScope == SingleThread) {
16459 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
16460 // the IR level, so we must wrap it in an intrinsic.
16462 } else if (hasMFENCE(*Subtarget)) {
16463 Function *MFence = llvm::Intrinsic::getDeclaration(M,
16464 Intrinsic::x86_sse2_mfence);
16465 Builder.CreateCall(MFence);
16467 // FIXME: it might make sense to use a locked operation here but on a
16468 // different cache-line to prevent cache-line bouncing. In practice it
16469 // is probably a small win, and x86 processors without mfence are rare
16470 // enough that we do not bother.
16474 // Finally we can emit the atomic load.
16475 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
16476 AI->getType()->getPrimitiveSizeInBits());
16477 Loaded->setAtomic(Order, SynchScope);
16478 AI->replaceAllUsesWith(Loaded);
16479 AI->eraseFromParent();
16483 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
16484 SelectionDAG &DAG) {
16486 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
16487 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
16488 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
16489 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
16491 // The only fence that needs an instruction is a sequentially-consistent
16492 // cross-thread fence.
16493 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
16494 if (hasMFENCE(*Subtarget))
16495 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
16497 SDValue Chain = Op.getOperand(0);
16498 SDValue Zero = DAG.getConstant(0, MVT::i32);
16500 DAG.getRegister(X86::ESP, MVT::i32), // Base
16501 DAG.getTargetConstant(1, MVT::i8), // Scale
16502 DAG.getRegister(0, MVT::i32), // Index
16503 DAG.getTargetConstant(0, MVT::i32), // Disp
16504 DAG.getRegister(0, MVT::i32), // Segment.
16508 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
16509 return SDValue(Res, 0);
16512 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
16513 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
16516 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
16517 SelectionDAG &DAG) {
16518 MVT T = Op.getSimpleValueType();
16522 switch(T.SimpleTy) {
16523 default: llvm_unreachable("Invalid value type!");
16524 case MVT::i8: Reg = X86::AL; size = 1; break;
16525 case MVT::i16: Reg = X86::AX; size = 2; break;
16526 case MVT::i32: Reg = X86::EAX; size = 4; break;
16528 assert(Subtarget->is64Bit() && "Node not type legal!");
16529 Reg = X86::RAX; size = 8;
16532 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
16533 Op.getOperand(2), SDValue());
16534 SDValue Ops[] = { cpIn.getValue(0),
16537 DAG.getTargetConstant(size, MVT::i8),
16538 cpIn.getValue(1) };
16539 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16540 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
16541 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
16545 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
16546 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
16547 MVT::i32, cpOut.getValue(2));
16548 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
16549 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
16551 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
16552 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
16553 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
16557 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
16558 SelectionDAG &DAG) {
16559 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
16560 MVT DstVT = Op.getSimpleValueType();
16562 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
16563 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
16564 if (DstVT != MVT::f64)
16565 // This conversion needs to be expanded.
16568 SDValue InVec = Op->getOperand(0);
16570 unsigned NumElts = SrcVT.getVectorNumElements();
16571 EVT SVT = SrcVT.getVectorElementType();
16573 // Widen the vector in input in the case of MVT::v2i32.
16574 // Example: from MVT::v2i32 to MVT::v4i32.
16575 SmallVector<SDValue, 16> Elts;
16576 for (unsigned i = 0, e = NumElts; i != e; ++i)
16577 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
16578 DAG.getIntPtrConstant(i)));
16580 // Explicitly mark the extra elements as Undef.
16581 Elts.append(NumElts, DAG.getUNDEF(SVT));
16583 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
16584 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
16585 SDValue ToV2F64 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, BV);
16586 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
16587 DAG.getIntPtrConstant(0));
16590 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
16591 Subtarget->hasMMX() && "Unexpected custom BITCAST");
16592 assert((DstVT == MVT::i64 ||
16593 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
16594 "Unexpected custom BITCAST");
16595 // i64 <=> MMX conversions are Legal.
16596 if (SrcVT==MVT::i64 && DstVT.isVector())
16598 if (DstVT==MVT::i64 && SrcVT.isVector())
16600 // MMX <=> MMX conversions are Legal.
16601 if (SrcVT.isVector() && DstVT.isVector())
16603 // All other conversions need to be expanded.
16607 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
16608 SelectionDAG &DAG) {
16609 SDNode *Node = Op.getNode();
16612 Op = Op.getOperand(0);
16613 EVT VT = Op.getValueType();
16614 assert((VT.is128BitVector() || VT.is256BitVector()) &&
16615 "CTPOP lowering only implemented for 128/256-bit wide vector types");
16617 unsigned NumElts = VT.getVectorNumElements();
16618 EVT EltVT = VT.getVectorElementType();
16619 unsigned Len = EltVT.getSizeInBits();
16621 // This is the vectorized version of the "best" algorithm from
16622 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
16623 // with a minor tweak to use a series of adds + shifts instead of vector
16624 // multiplications. Implemented for the v2i64, v4i64, v4i32, v8i32 types:
16626 // v2i64, v4i64, v4i32 => Only profitable w/ popcnt disabled
16627 // v8i32 => Always profitable
16629 // FIXME: There a couple of possible improvements:
16631 // 1) Support for i8 and i16 vectors (needs measurements if popcnt enabled).
16632 // 2) Use strategies from http://wm.ite.pl/articles/sse-popcount.html
16634 assert(EltVT.isInteger() && (Len == 32 || Len == 64) && Len % 8 == 0 &&
16635 "CTPOP not implemented for this vector element type.");
16637 // X86 canonicalize ANDs to vXi64, generate the appropriate bitcasts to avoid
16638 // extra legalization.
16639 bool NeedsBitcast = EltVT == MVT::i32;
16640 MVT BitcastVT = VT.is256BitVector() ? MVT::v4i64 : MVT::v2i64;
16642 SDValue Cst55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), EltVT);
16643 SDValue Cst33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), EltVT);
16644 SDValue Cst0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), EltVT);
16646 // v = v - ((v >> 1) & 0x55555555...)
16647 SmallVector<SDValue, 8> Ones(NumElts, DAG.getConstant(1, EltVT));
16648 SDValue OnesV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ones);
16649 SDValue Srl = DAG.getNode(ISD::SRL, dl, VT, Op, OnesV);
16651 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16653 SmallVector<SDValue, 8> Mask55(NumElts, Cst55);
16654 SDValue M55 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask55);
16656 M55 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M55);
16658 SDValue And = DAG.getNode(ISD::AND, dl, Srl.getValueType(), Srl, M55);
16659 if (VT != And.getValueType())
16660 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16661 SDValue Sub = DAG.getNode(ISD::SUB, dl, VT, Op, And);
16663 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
16664 SmallVector<SDValue, 8> Mask33(NumElts, Cst33);
16665 SDValue M33 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask33);
16666 SmallVector<SDValue, 8> Twos(NumElts, DAG.getConstant(2, EltVT));
16667 SDValue TwosV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Twos);
16669 Srl = DAG.getNode(ISD::SRL, dl, VT, Sub, TwosV);
16670 if (NeedsBitcast) {
16671 Srl = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Srl);
16672 M33 = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M33);
16673 Sub = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Sub);
16676 SDValue AndRHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Srl, M33);
16677 SDValue AndLHS = DAG.getNode(ISD::AND, dl, M33.getValueType(), Sub, M33);
16678 if (VT != AndRHS.getValueType()) {
16679 AndRHS = DAG.getNode(ISD::BITCAST, dl, VT, AndRHS);
16680 AndLHS = DAG.getNode(ISD::BITCAST, dl, VT, AndLHS);
16682 SDValue Add = DAG.getNode(ISD::ADD, dl, VT, AndLHS, AndRHS);
16684 // v = (v + (v >> 4)) & 0x0F0F0F0F...
16685 SmallVector<SDValue, 8> Fours(NumElts, DAG.getConstant(4, EltVT));
16686 SDValue FoursV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Fours);
16687 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, FoursV);
16688 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16690 SmallVector<SDValue, 8> Mask0F(NumElts, Cst0F);
16691 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Mask0F);
16692 if (NeedsBitcast) {
16693 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16694 M0F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M0F);
16696 And = DAG.getNode(ISD::AND, dl, M0F.getValueType(), Add, M0F);
16697 if (VT != And.getValueType())
16698 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16700 // The algorithm mentioned above uses:
16701 // v = (v * 0x01010101...) >> (Len - 8)
16703 // Change it to use vector adds + vector shifts which yield faster results on
16704 // Haswell than using vector integer multiplication.
16706 // For i32 elements:
16707 // v = v + (v >> 8)
16708 // v = v + (v >> 16)
16710 // For i64 elements:
16711 // v = v + (v >> 8)
16712 // v = v + (v >> 16)
16713 // v = v + (v >> 32)
16716 SmallVector<SDValue, 8> Csts;
16717 for (unsigned i = 8; i <= Len/2; i *= 2) {
16718 Csts.assign(NumElts, DAG.getConstant(i, EltVT));
16719 SDValue CstsV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Csts);
16720 Srl = DAG.getNode(ISD::SRL, dl, VT, Add, CstsV);
16721 Add = DAG.getNode(ISD::ADD, dl, VT, Add, Srl);
16725 // The result is on the least significant 6-bits on i32 and 7-bits on i64.
16726 SDValue Cst3F = DAG.getConstant(APInt(Len, Len == 32 ? 0x3F : 0x7F), EltVT);
16727 SmallVector<SDValue, 8> Cst3FV(NumElts, Cst3F);
16728 SDValue M3F = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Cst3FV);
16729 if (NeedsBitcast) {
16730 Add = DAG.getNode(ISD::BITCAST, dl, BitcastVT, Add);
16731 M3F = DAG.getNode(ISD::BITCAST, dl, BitcastVT, M3F);
16733 And = DAG.getNode(ISD::AND, dl, M3F.getValueType(), Add, M3F);
16734 if (VT != And.getValueType())
16735 And = DAG.getNode(ISD::BITCAST, dl, VT, And);
16740 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
16741 SDNode *Node = Op.getNode();
16743 EVT T = Node->getValueType(0);
16744 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
16745 DAG.getConstant(0, T), Node->getOperand(2));
16746 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
16747 cast<AtomicSDNode>(Node)->getMemoryVT(),
16748 Node->getOperand(0),
16749 Node->getOperand(1), negOp,
16750 cast<AtomicSDNode>(Node)->getMemOperand(),
16751 cast<AtomicSDNode>(Node)->getOrdering(),
16752 cast<AtomicSDNode>(Node)->getSynchScope());
16755 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
16756 SDNode *Node = Op.getNode();
16758 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
16760 // Convert seq_cst store -> xchg
16761 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
16762 // FIXME: On 32-bit, store -> fist or movq would be more efficient
16763 // (The only way to get a 16-byte store is cmpxchg16b)
16764 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
16765 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
16766 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
16767 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
16768 cast<AtomicSDNode>(Node)->getMemoryVT(),
16769 Node->getOperand(0),
16770 Node->getOperand(1), Node->getOperand(2),
16771 cast<AtomicSDNode>(Node)->getMemOperand(),
16772 cast<AtomicSDNode>(Node)->getOrdering(),
16773 cast<AtomicSDNode>(Node)->getSynchScope());
16774 return Swap.getValue(1);
16776 // Other atomic stores have a simple pattern.
16780 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
16781 EVT VT = Op.getNode()->getSimpleValueType(0);
16783 // Let legalize expand this if it isn't a legal type yet.
16784 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
16787 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16790 bool ExtraOp = false;
16791 switch (Op.getOpcode()) {
16792 default: llvm_unreachable("Invalid code");
16793 case ISD::ADDC: Opc = X86ISD::ADD; break;
16794 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
16795 case ISD::SUBC: Opc = X86ISD::SUB; break;
16796 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
16800 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16802 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
16803 Op.getOperand(1), Op.getOperand(2));
16806 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
16807 SelectionDAG &DAG) {
16808 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
16810 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
16811 // which returns the values as { float, float } (in XMM0) or
16812 // { double, double } (which is returned in XMM0, XMM1).
16814 SDValue Arg = Op.getOperand(0);
16815 EVT ArgVT = Arg.getValueType();
16816 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
16818 TargetLowering::ArgListTy Args;
16819 TargetLowering::ArgListEntry Entry;
16823 Entry.isSExt = false;
16824 Entry.isZExt = false;
16825 Args.push_back(Entry);
16827 bool isF64 = ArgVT == MVT::f64;
16828 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
16829 // the small struct {f32, f32} is returned in (eax, edx). For f64,
16830 // the results are returned via SRet in memory.
16831 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
16832 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16833 SDValue Callee = DAG.getExternalSymbol(LibcallName, TLI.getPointerTy());
16835 Type *RetTy = isF64
16836 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
16837 : (Type*)VectorType::get(ArgTy, 4);
16839 TargetLowering::CallLoweringInfo CLI(DAG);
16840 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
16841 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
16843 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
16846 // Returned in xmm0 and xmm1.
16847 return CallResult.first;
16849 // Returned in bits 0:31 and 32:64 xmm0.
16850 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16851 CallResult.first, DAG.getIntPtrConstant(0));
16852 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
16853 CallResult.first, DAG.getIntPtrConstant(1));
16854 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
16855 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
16858 /// LowerOperation - Provide custom lowering hooks for some operations.
16860 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
16861 switch (Op.getOpcode()) {
16862 default: llvm_unreachable("Should not custom lower this!");
16863 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
16864 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
16865 return LowerCMP_SWAP(Op, Subtarget, DAG);
16866 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
16867 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
16868 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
16869 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
16870 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
16871 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
16872 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
16873 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
16874 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
16875 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
16876 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
16877 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
16878 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
16879 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
16880 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
16881 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
16882 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
16883 case ISD::SHL_PARTS:
16884 case ISD::SRA_PARTS:
16885 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
16886 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
16887 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
16888 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
16889 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
16890 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
16891 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
16892 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
16893 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
16894 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
16895 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
16897 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
16898 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
16899 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
16900 case ISD::SETCC: return LowerSETCC(Op, DAG);
16901 case ISD::SELECT: return LowerSELECT(Op, DAG);
16902 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
16903 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
16904 case ISD::VASTART: return LowerVASTART(Op, DAG);
16905 case ISD::VAARG: return LowerVAARG(Op, DAG);
16906 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
16907 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
16908 case ISD::INTRINSIC_VOID:
16909 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
16910 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
16911 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
16912 case ISD::FRAME_TO_ARGS_OFFSET:
16913 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
16914 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
16915 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
16916 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
16917 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
16918 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
16919 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
16920 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
16921 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
16922 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
16923 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
16924 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
16925 case ISD::UMUL_LOHI:
16926 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
16929 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
16935 case ISD::UMULO: return LowerXALUO(Op, DAG);
16936 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
16937 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
16941 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
16942 case ISD::ADD: return LowerADD(Op, DAG);
16943 case ISD::SUB: return LowerSUB(Op, DAG);
16944 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
16948 /// ReplaceNodeResults - Replace a node with an illegal result type
16949 /// with a new node built out of custom code.
16950 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
16951 SmallVectorImpl<SDValue>&Results,
16952 SelectionDAG &DAG) const {
16954 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16955 switch (N->getOpcode()) {
16957 llvm_unreachable("Do not know how to custom type legalize this operation!");
16958 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
16959 case X86ISD::FMINC:
16961 case X86ISD::FMAXC:
16962 case X86ISD::FMAX: {
16963 EVT VT = N->getValueType(0);
16964 if (VT != MVT::v2f32)
16965 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
16966 SDValue UNDEF = DAG.getUNDEF(VT);
16967 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
16968 N->getOperand(0), UNDEF);
16969 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
16970 N->getOperand(1), UNDEF);
16971 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
16974 case ISD::SIGN_EXTEND_INREG:
16979 // We don't want to expand or promote these.
16986 case ISD::UDIVREM: {
16987 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
16988 Results.push_back(V);
16991 case ISD::FP_TO_SINT:
16992 case ISD::FP_TO_UINT: {
16993 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
16995 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
16998 std::pair<SDValue,SDValue> Vals =
16999 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
17000 SDValue FIST = Vals.first, StackSlot = Vals.second;
17001 if (FIST.getNode()) {
17002 EVT VT = N->getValueType(0);
17003 // Return a load from the stack slot.
17004 if (StackSlot.getNode())
17005 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
17006 MachinePointerInfo(),
17007 false, false, false, 0));
17009 Results.push_back(FIST);
17013 case ISD::UINT_TO_FP: {
17014 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17015 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
17016 N->getValueType(0) != MVT::v2f32)
17018 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
17020 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
17022 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
17023 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
17024 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
17025 Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
17026 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
17027 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
17030 case ISD::FP_ROUND: {
17031 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
17033 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
17034 Results.push_back(V);
17037 case ISD::INTRINSIC_W_CHAIN: {
17038 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
17040 default : llvm_unreachable("Do not know how to custom type "
17041 "legalize this intrinsic operation!");
17042 case Intrinsic::x86_rdtsc:
17043 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17045 case Intrinsic::x86_rdtscp:
17046 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
17048 case Intrinsic::x86_rdpmc:
17049 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
17052 case ISD::READCYCLECOUNTER: {
17053 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
17056 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
17057 EVT T = N->getValueType(0);
17058 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
17059 bool Regs64bit = T == MVT::i128;
17060 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
17061 SDValue cpInL, cpInH;
17062 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17063 DAG.getConstant(0, HalfT));
17064 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
17065 DAG.getConstant(1, HalfT));
17066 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
17067 Regs64bit ? X86::RAX : X86::EAX,
17069 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
17070 Regs64bit ? X86::RDX : X86::EDX,
17071 cpInH, cpInL.getValue(1));
17072 SDValue swapInL, swapInH;
17073 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17074 DAG.getConstant(0, HalfT));
17075 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
17076 DAG.getConstant(1, HalfT));
17077 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
17078 Regs64bit ? X86::RBX : X86::EBX,
17079 swapInL, cpInH.getValue(1));
17080 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
17081 Regs64bit ? X86::RCX : X86::ECX,
17082 swapInH, swapInL.getValue(1));
17083 SDValue Ops[] = { swapInH.getValue(0),
17085 swapInH.getValue(1) };
17086 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
17087 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
17088 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
17089 X86ISD::LCMPXCHG8_DAG;
17090 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
17091 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
17092 Regs64bit ? X86::RAX : X86::EAX,
17093 HalfT, Result.getValue(1));
17094 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
17095 Regs64bit ? X86::RDX : X86::EDX,
17096 HalfT, cpOutL.getValue(2));
17097 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
17099 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
17100 MVT::i32, cpOutH.getValue(2));
17102 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17103 DAG.getConstant(X86::COND_E, MVT::i8), EFLAGS);
17104 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
17106 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
17107 Results.push_back(Success);
17108 Results.push_back(EFLAGS.getValue(1));
17111 case ISD::ATOMIC_SWAP:
17112 case ISD::ATOMIC_LOAD_ADD:
17113 case ISD::ATOMIC_LOAD_SUB:
17114 case ISD::ATOMIC_LOAD_AND:
17115 case ISD::ATOMIC_LOAD_OR:
17116 case ISD::ATOMIC_LOAD_XOR:
17117 case ISD::ATOMIC_LOAD_NAND:
17118 case ISD::ATOMIC_LOAD_MIN:
17119 case ISD::ATOMIC_LOAD_MAX:
17120 case ISD::ATOMIC_LOAD_UMIN:
17121 case ISD::ATOMIC_LOAD_UMAX:
17122 case ISD::ATOMIC_LOAD: {
17123 // Delegate to generic TypeLegalization. Situations we can really handle
17124 // should have already been dealt with by AtomicExpandPass.cpp.
17127 case ISD::BITCAST: {
17128 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
17129 EVT DstVT = N->getValueType(0);
17130 EVT SrcVT = N->getOperand(0)->getValueType(0);
17132 if (SrcVT != MVT::f64 ||
17133 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
17136 unsigned NumElts = DstVT.getVectorNumElements();
17137 EVT SVT = DstVT.getVectorElementType();
17138 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
17139 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
17140 MVT::v2f64, N->getOperand(0));
17141 SDValue ToVecInt = DAG.getNode(ISD::BITCAST, dl, WiderVT, Expanded);
17143 if (ExperimentalVectorWideningLegalization) {
17144 // If we are legalizing vectors by widening, we already have the desired
17145 // legal vector type, just return it.
17146 Results.push_back(ToVecInt);
17150 SmallVector<SDValue, 8> Elts;
17151 for (unsigned i = 0, e = NumElts; i != e; ++i)
17152 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
17153 ToVecInt, DAG.getIntPtrConstant(i)));
17155 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
17160 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
17162 default: return nullptr;
17163 case X86ISD::BSF: return "X86ISD::BSF";
17164 case X86ISD::BSR: return "X86ISD::BSR";
17165 case X86ISD::SHLD: return "X86ISD::SHLD";
17166 case X86ISD::SHRD: return "X86ISD::SHRD";
17167 case X86ISD::FAND: return "X86ISD::FAND";
17168 case X86ISD::FANDN: return "X86ISD::FANDN";
17169 case X86ISD::FOR: return "X86ISD::FOR";
17170 case X86ISD::FXOR: return "X86ISD::FXOR";
17171 case X86ISD::FSRL: return "X86ISD::FSRL";
17172 case X86ISD::FILD: return "X86ISD::FILD";
17173 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
17174 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
17175 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
17176 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
17177 case X86ISD::FLD: return "X86ISD::FLD";
17178 case X86ISD::FST: return "X86ISD::FST";
17179 case X86ISD::CALL: return "X86ISD::CALL";
17180 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
17181 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
17182 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
17183 case X86ISD::BT: return "X86ISD::BT";
17184 case X86ISD::CMP: return "X86ISD::CMP";
17185 case X86ISD::COMI: return "X86ISD::COMI";
17186 case X86ISD::UCOMI: return "X86ISD::UCOMI";
17187 case X86ISD::CMPM: return "X86ISD::CMPM";
17188 case X86ISD::CMPMU: return "X86ISD::CMPMU";
17189 case X86ISD::SETCC: return "X86ISD::SETCC";
17190 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
17191 case X86ISD::FSETCC: return "X86ISD::FSETCC";
17192 case X86ISD::CMOV: return "X86ISD::CMOV";
17193 case X86ISD::BRCOND: return "X86ISD::BRCOND";
17194 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
17195 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
17196 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
17197 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
17198 case X86ISD::Wrapper: return "X86ISD::Wrapper";
17199 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
17200 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
17201 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
17202 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
17203 case X86ISD::PINSRB: return "X86ISD::PINSRB";
17204 case X86ISD::PINSRW: return "X86ISD::PINSRW";
17205 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
17206 case X86ISD::ANDNP: return "X86ISD::ANDNP";
17207 case X86ISD::PSIGN: return "X86ISD::PSIGN";
17208 case X86ISD::BLENDI: return "X86ISD::BLENDI";
17209 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
17210 case X86ISD::SUBUS: return "X86ISD::SUBUS";
17211 case X86ISD::HADD: return "X86ISD::HADD";
17212 case X86ISD::HSUB: return "X86ISD::HSUB";
17213 case X86ISD::FHADD: return "X86ISD::FHADD";
17214 case X86ISD::FHSUB: return "X86ISD::FHSUB";
17215 case X86ISD::UMAX: return "X86ISD::UMAX";
17216 case X86ISD::UMIN: return "X86ISD::UMIN";
17217 case X86ISD::SMAX: return "X86ISD::SMAX";
17218 case X86ISD::SMIN: return "X86ISD::SMIN";
17219 case X86ISD::FMAX: return "X86ISD::FMAX";
17220 case X86ISD::FMIN: return "X86ISD::FMIN";
17221 case X86ISD::FMAXC: return "X86ISD::FMAXC";
17222 case X86ISD::FMINC: return "X86ISD::FMINC";
17223 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
17224 case X86ISD::FRCP: return "X86ISD::FRCP";
17225 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
17226 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
17227 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
17228 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
17229 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
17230 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
17231 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
17232 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
17233 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
17234 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
17235 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
17236 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
17237 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
17238 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
17239 case X86ISD::VZEXT: return "X86ISD::VZEXT";
17240 case X86ISD::VSEXT: return "X86ISD::VSEXT";
17241 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
17242 case X86ISD::VTRUNCM: return "X86ISD::VTRUNCM";
17243 case X86ISD::VINSERT: return "X86ISD::VINSERT";
17244 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
17245 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
17246 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
17247 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
17248 case X86ISD::VSHL: return "X86ISD::VSHL";
17249 case X86ISD::VSRL: return "X86ISD::VSRL";
17250 case X86ISD::VSRA: return "X86ISD::VSRA";
17251 case X86ISD::VSHLI: return "X86ISD::VSHLI";
17252 case X86ISD::VSRLI: return "X86ISD::VSRLI";
17253 case X86ISD::VSRAI: return "X86ISD::VSRAI";
17254 case X86ISD::CMPP: return "X86ISD::CMPP";
17255 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
17256 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
17257 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
17258 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
17259 case X86ISD::ADD: return "X86ISD::ADD";
17260 case X86ISD::SUB: return "X86ISD::SUB";
17261 case X86ISD::ADC: return "X86ISD::ADC";
17262 case X86ISD::SBB: return "X86ISD::SBB";
17263 case X86ISD::SMUL: return "X86ISD::SMUL";
17264 case X86ISD::UMUL: return "X86ISD::UMUL";
17265 case X86ISD::SMUL8: return "X86ISD::SMUL8";
17266 case X86ISD::UMUL8: return "X86ISD::UMUL8";
17267 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
17268 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
17269 case X86ISD::INC: return "X86ISD::INC";
17270 case X86ISD::DEC: return "X86ISD::DEC";
17271 case X86ISD::OR: return "X86ISD::OR";
17272 case X86ISD::XOR: return "X86ISD::XOR";
17273 case X86ISD::AND: return "X86ISD::AND";
17274 case X86ISD::BEXTR: return "X86ISD::BEXTR";
17275 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
17276 case X86ISD::PTEST: return "X86ISD::PTEST";
17277 case X86ISD::TESTP: return "X86ISD::TESTP";
17278 case X86ISD::TESTM: return "X86ISD::TESTM";
17279 case X86ISD::TESTNM: return "X86ISD::TESTNM";
17280 case X86ISD::KORTEST: return "X86ISD::KORTEST";
17281 case X86ISD::PACKSS: return "X86ISD::PACKSS";
17282 case X86ISD::PACKUS: return "X86ISD::PACKUS";
17283 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
17284 case X86ISD::VALIGN: return "X86ISD::VALIGN";
17285 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
17286 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
17287 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
17288 case X86ISD::SHUFP: return "X86ISD::SHUFP";
17289 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
17290 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
17291 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
17292 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
17293 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
17294 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
17295 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
17296 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
17297 case X86ISD::MOVSD: return "X86ISD::MOVSD";
17298 case X86ISD::MOVSS: return "X86ISD::MOVSS";
17299 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
17300 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
17301 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
17302 case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
17303 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
17304 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
17305 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
17306 case X86ISD::VPERMV: return "X86ISD::VPERMV";
17307 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
17308 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
17309 case X86ISD::VPERMI: return "X86ISD::VPERMI";
17310 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
17311 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
17312 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
17313 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
17314 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
17315 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
17316 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
17317 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
17318 case X86ISD::SAHF: return "X86ISD::SAHF";
17319 case X86ISD::RDRAND: return "X86ISD::RDRAND";
17320 case X86ISD::RDSEED: return "X86ISD::RDSEED";
17321 case X86ISD::FMADD: return "X86ISD::FMADD";
17322 case X86ISD::FMSUB: return "X86ISD::FMSUB";
17323 case X86ISD::FNMADD: return "X86ISD::FNMADD";
17324 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
17325 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
17326 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
17327 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
17328 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
17329 case X86ISD::XTEST: return "X86ISD::XTEST";
17330 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
17331 case X86ISD::EXPAND: return "X86ISD::EXPAND";
17332 case X86ISD::SELECT: return "X86ISD::SELECT";
17333 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
17334 case X86ISD::RCP28: return "X86ISD::RCP28";
17335 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
17336 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
17337 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
17338 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
17339 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
17343 // isLegalAddressingMode - Return true if the addressing mode represented
17344 // by AM is legal for this target, for a load/store of the specified type.
17345 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
17347 // X86 supports extremely general addressing modes.
17348 CodeModel::Model M = getTargetMachine().getCodeModel();
17349 Reloc::Model R = getTargetMachine().getRelocationModel();
17351 // X86 allows a sign-extended 32-bit immediate field as a displacement.
17352 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
17357 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
17359 // If a reference to this global requires an extra load, we can't fold it.
17360 if (isGlobalStubReference(GVFlags))
17363 // If BaseGV requires a register for the PIC base, we cannot also have a
17364 // BaseReg specified.
17365 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
17368 // If lower 4G is not available, then we must use rip-relative addressing.
17369 if ((M != CodeModel::Small || R != Reloc::Static) &&
17370 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
17374 switch (AM.Scale) {
17380 // These scales always work.
17385 // These scales are formed with basereg+scalereg. Only accept if there is
17390 default: // Other stuff never works.
17397 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
17398 unsigned Bits = Ty->getScalarSizeInBits();
17400 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
17401 // particularly cheaper than those without.
17405 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
17406 // variable shifts just as cheap as scalar ones.
17407 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
17410 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
17411 // fully general vector.
17415 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
17416 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17418 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
17419 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
17420 return NumBits1 > NumBits2;
17423 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
17424 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
17427 if (!isTypeLegal(EVT::getEVT(Ty1)))
17430 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
17432 // Assuming the caller doesn't have a zeroext or signext return parameter,
17433 // truncation all the way down to i1 is valid.
17437 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
17438 return isInt<32>(Imm);
17441 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
17442 // Can also use sub to handle negated immediates.
17443 return isInt<32>(Imm);
17446 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
17447 if (!VT1.isInteger() || !VT2.isInteger())
17449 unsigned NumBits1 = VT1.getSizeInBits();
17450 unsigned NumBits2 = VT2.getSizeInBits();
17451 return NumBits1 > NumBits2;
17454 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
17455 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17456 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
17459 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
17460 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
17461 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
17464 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
17465 EVT VT1 = Val.getValueType();
17466 if (isZExtFree(VT1, VT2))
17469 if (Val.getOpcode() != ISD::LOAD)
17472 if (!VT1.isSimple() || !VT1.isInteger() ||
17473 !VT2.isSimple() || !VT2.isInteger())
17476 switch (VT1.getSimpleVT().SimpleTy) {
17481 // X86 has 8, 16, and 32-bit zero-extending loads.
17488 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
17491 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
17492 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4()))
17495 VT = VT.getScalarType();
17497 if (!VT.isSimple())
17500 switch (VT.getSimpleVT().SimpleTy) {
17511 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
17512 // i16 instructions are longer (0x66 prefix) and potentially slower.
17513 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
17516 /// isShuffleMaskLegal - Targets can use this to indicate that they only
17517 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
17518 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
17519 /// are assumed to be legal.
17521 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
17523 if (!VT.isSimple())
17526 // Very little shuffling can be done for 64-bit vectors right now.
17527 if (VT.getSizeInBits() == 64)
17530 // We only care that the types being shuffled are legal. The lowering can
17531 // handle any possible shuffle mask that results.
17532 return isTypeLegal(VT.getSimpleVT());
17536 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
17538 // Just delegate to the generic legality, clear masks aren't special.
17539 return isShuffleMaskLegal(Mask, VT);
17542 //===----------------------------------------------------------------------===//
17543 // X86 Scheduler Hooks
17544 //===----------------------------------------------------------------------===//
17546 /// Utility function to emit xbegin specifying the start of an RTM region.
17547 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
17548 const TargetInstrInfo *TII) {
17549 DebugLoc DL = MI->getDebugLoc();
17551 const BasicBlock *BB = MBB->getBasicBlock();
17552 MachineFunction::iterator I = MBB;
17555 // For the v = xbegin(), we generate
17566 MachineBasicBlock *thisMBB = MBB;
17567 MachineFunction *MF = MBB->getParent();
17568 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
17569 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
17570 MF->insert(I, mainMBB);
17571 MF->insert(I, sinkMBB);
17573 // Transfer the remainder of BB and its successor edges to sinkMBB.
17574 sinkMBB->splice(sinkMBB->begin(), MBB,
17575 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17576 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
17580 // # fallthrough to mainMBB
17581 // # abortion to sinkMBB
17582 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
17583 thisMBB->addSuccessor(mainMBB);
17584 thisMBB->addSuccessor(sinkMBB);
17588 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
17589 mainMBB->addSuccessor(sinkMBB);
17592 // EAX is live into the sinkMBB
17593 sinkMBB->addLiveIn(X86::EAX);
17594 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
17595 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17598 MI->eraseFromParent();
17602 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
17603 // or XMM0_V32I8 in AVX all of this code can be replaced with that
17604 // in the .td file.
17605 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
17606 const TargetInstrInfo *TII) {
17608 switch (MI->getOpcode()) {
17609 default: llvm_unreachable("illegal opcode!");
17610 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
17611 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
17612 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
17613 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
17614 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
17615 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
17616 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
17617 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
17620 DebugLoc dl = MI->getDebugLoc();
17621 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17623 unsigned NumArgs = MI->getNumOperands();
17624 for (unsigned i = 1; i < NumArgs; ++i) {
17625 MachineOperand &Op = MI->getOperand(i);
17626 if (!(Op.isReg() && Op.isImplicit()))
17627 MIB.addOperand(Op);
17629 if (MI->hasOneMemOperand())
17630 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17632 BuildMI(*BB, MI, dl,
17633 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17634 .addReg(X86::XMM0);
17636 MI->eraseFromParent();
17640 // FIXME: Custom handling because TableGen doesn't support multiple implicit
17641 // defs in an instruction pattern
17642 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
17643 const TargetInstrInfo *TII) {
17645 switch (MI->getOpcode()) {
17646 default: llvm_unreachable("illegal opcode!");
17647 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
17648 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
17649 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
17650 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
17651 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
17652 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
17653 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
17654 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
17657 DebugLoc dl = MI->getDebugLoc();
17658 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
17660 unsigned NumArgs = MI->getNumOperands(); // remove the results
17661 for (unsigned i = 1; i < NumArgs; ++i) {
17662 MachineOperand &Op = MI->getOperand(i);
17663 if (!(Op.isReg() && Op.isImplicit()))
17664 MIB.addOperand(Op);
17666 if (MI->hasOneMemOperand())
17667 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
17669 BuildMI(*BB, MI, dl,
17670 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
17673 MI->eraseFromParent();
17677 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
17678 const X86Subtarget *Subtarget) {
17679 DebugLoc dl = MI->getDebugLoc();
17680 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17681 // Address into RAX/EAX, other two args into ECX, EDX.
17682 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
17683 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
17684 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
17685 for (int i = 0; i < X86::AddrNumOperands; ++i)
17686 MIB.addOperand(MI->getOperand(i));
17688 unsigned ValOps = X86::AddrNumOperands;
17689 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
17690 .addReg(MI->getOperand(ValOps).getReg());
17691 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
17692 .addReg(MI->getOperand(ValOps+1).getReg());
17694 // The instruction doesn't actually take any operands though.
17695 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
17697 MI->eraseFromParent(); // The pseudo is gone now.
17701 MachineBasicBlock *
17702 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
17703 MachineBasicBlock *MBB) const {
17704 // Emit va_arg instruction on X86-64.
17706 // Operands to this pseudo-instruction:
17707 // 0 ) Output : destination address (reg)
17708 // 1-5) Input : va_list address (addr, i64mem)
17709 // 6 ) ArgSize : Size (in bytes) of vararg type
17710 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
17711 // 8 ) Align : Alignment of type
17712 // 9 ) EFLAGS (implicit-def)
17714 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
17715 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
17717 unsigned DestReg = MI->getOperand(0).getReg();
17718 MachineOperand &Base = MI->getOperand(1);
17719 MachineOperand &Scale = MI->getOperand(2);
17720 MachineOperand &Index = MI->getOperand(3);
17721 MachineOperand &Disp = MI->getOperand(4);
17722 MachineOperand &Segment = MI->getOperand(5);
17723 unsigned ArgSize = MI->getOperand(6).getImm();
17724 unsigned ArgMode = MI->getOperand(7).getImm();
17725 unsigned Align = MI->getOperand(8).getImm();
17727 // Memory Reference
17728 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
17729 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
17730 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
17732 // Machine Information
17733 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17734 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
17735 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
17736 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
17737 DebugLoc DL = MI->getDebugLoc();
17739 // struct va_list {
17742 // i64 overflow_area (address)
17743 // i64 reg_save_area (address)
17745 // sizeof(va_list) = 24
17746 // alignment(va_list) = 8
17748 unsigned TotalNumIntRegs = 6;
17749 unsigned TotalNumXMMRegs = 8;
17750 bool UseGPOffset = (ArgMode == 1);
17751 bool UseFPOffset = (ArgMode == 2);
17752 unsigned MaxOffset = TotalNumIntRegs * 8 +
17753 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
17755 /* Align ArgSize to a multiple of 8 */
17756 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
17757 bool NeedsAlign = (Align > 8);
17759 MachineBasicBlock *thisMBB = MBB;
17760 MachineBasicBlock *overflowMBB;
17761 MachineBasicBlock *offsetMBB;
17762 MachineBasicBlock *endMBB;
17764 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
17765 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
17766 unsigned OffsetReg = 0;
17768 if (!UseGPOffset && !UseFPOffset) {
17769 // If we only pull from the overflow region, we don't create a branch.
17770 // We don't need to alter control flow.
17771 OffsetDestReg = 0; // unused
17772 OverflowDestReg = DestReg;
17774 offsetMBB = nullptr;
17775 overflowMBB = thisMBB;
17778 // First emit code to check if gp_offset (or fp_offset) is below the bound.
17779 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
17780 // If not, pull from overflow_area. (branch to overflowMBB)
17785 // offsetMBB overflowMBB
17790 // Registers for the PHI in endMBB
17791 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
17792 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
17794 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17795 MachineFunction *MF = MBB->getParent();
17796 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17797 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17798 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
17800 MachineFunction::iterator MBBIter = MBB;
17803 // Insert the new basic blocks
17804 MF->insert(MBBIter, offsetMBB);
17805 MF->insert(MBBIter, overflowMBB);
17806 MF->insert(MBBIter, endMBB);
17808 // Transfer the remainder of MBB and its successor edges to endMBB.
17809 endMBB->splice(endMBB->begin(), thisMBB,
17810 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
17811 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
17813 // Make offsetMBB and overflowMBB successors of thisMBB
17814 thisMBB->addSuccessor(offsetMBB);
17815 thisMBB->addSuccessor(overflowMBB);
17817 // endMBB is a successor of both offsetMBB and overflowMBB
17818 offsetMBB->addSuccessor(endMBB);
17819 overflowMBB->addSuccessor(endMBB);
17821 // Load the offset value into a register
17822 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17823 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
17827 .addDisp(Disp, UseFPOffset ? 4 : 0)
17828 .addOperand(Segment)
17829 .setMemRefs(MMOBegin, MMOEnd);
17831 // Check if there is enough room left to pull this argument.
17832 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
17834 .addImm(MaxOffset + 8 - ArgSizeA8);
17836 // Branch to "overflowMBB" if offset >= max
17837 // Fall through to "offsetMBB" otherwise
17838 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
17839 .addMBB(overflowMBB);
17842 // In offsetMBB, emit code to use the reg_save_area.
17844 assert(OffsetReg != 0);
17846 // Read the reg_save_area address.
17847 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
17848 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
17853 .addOperand(Segment)
17854 .setMemRefs(MMOBegin, MMOEnd);
17856 // Zero-extend the offset
17857 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
17858 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
17861 .addImm(X86::sub_32bit);
17863 // Add the offset to the reg_save_area to get the final address.
17864 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
17865 .addReg(OffsetReg64)
17866 .addReg(RegSaveReg);
17868 // Compute the offset for the next argument
17869 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
17870 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
17872 .addImm(UseFPOffset ? 16 : 8);
17874 // Store it back into the va_list.
17875 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
17879 .addDisp(Disp, UseFPOffset ? 4 : 0)
17880 .addOperand(Segment)
17881 .addReg(NextOffsetReg)
17882 .setMemRefs(MMOBegin, MMOEnd);
17885 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
17890 // Emit code to use overflow area
17893 // Load the overflow_area address into a register.
17894 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
17895 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
17900 .addOperand(Segment)
17901 .setMemRefs(MMOBegin, MMOEnd);
17903 // If we need to align it, do so. Otherwise, just copy the address
17904 // to OverflowDestReg.
17906 // Align the overflow address
17907 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
17908 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
17910 // aligned_addr = (addr + (align-1)) & ~(align-1)
17911 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
17912 .addReg(OverflowAddrReg)
17915 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
17917 .addImm(~(uint64_t)(Align-1));
17919 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
17920 .addReg(OverflowAddrReg);
17923 // Compute the next overflow address after this argument.
17924 // (the overflow address should be kept 8-byte aligned)
17925 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
17926 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
17927 .addReg(OverflowDestReg)
17928 .addImm(ArgSizeA8);
17930 // Store the new overflow address.
17931 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
17936 .addOperand(Segment)
17937 .addReg(NextAddrReg)
17938 .setMemRefs(MMOBegin, MMOEnd);
17940 // If we branched, emit the PHI to the front of endMBB.
17942 BuildMI(*endMBB, endMBB->begin(), DL,
17943 TII->get(X86::PHI), DestReg)
17944 .addReg(OffsetDestReg).addMBB(offsetMBB)
17945 .addReg(OverflowDestReg).addMBB(overflowMBB);
17948 // Erase the pseudo instruction
17949 MI->eraseFromParent();
17954 MachineBasicBlock *
17955 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
17957 MachineBasicBlock *MBB) const {
17958 // Emit code to save XMM registers to the stack. The ABI says that the
17959 // number of registers to save is given in %al, so it's theoretically
17960 // possible to do an indirect jump trick to avoid saving all of them,
17961 // however this code takes a simpler approach and just executes all
17962 // of the stores if %al is non-zero. It's less code, and it's probably
17963 // easier on the hardware branch predictor, and stores aren't all that
17964 // expensive anyway.
17966 // Create the new basic blocks. One block contains all the XMM stores,
17967 // and one block is the final destination regardless of whether any
17968 // stores were performed.
17969 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
17970 MachineFunction *F = MBB->getParent();
17971 MachineFunction::iterator MBBIter = MBB;
17973 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
17974 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
17975 F->insert(MBBIter, XMMSaveMBB);
17976 F->insert(MBBIter, EndMBB);
17978 // Transfer the remainder of MBB and its successor edges to EndMBB.
17979 EndMBB->splice(EndMBB->begin(), MBB,
17980 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
17981 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
17983 // The original block will now fall through to the XMM save block.
17984 MBB->addSuccessor(XMMSaveMBB);
17985 // The XMMSaveMBB will fall through to the end block.
17986 XMMSaveMBB->addSuccessor(EndMBB);
17988 // Now add the instructions.
17989 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
17990 DebugLoc DL = MI->getDebugLoc();
17992 unsigned CountReg = MI->getOperand(0).getReg();
17993 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
17994 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
17996 if (!Subtarget->isTargetWin64()) {
17997 // If %al is 0, branch around the XMM save block.
17998 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
17999 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
18000 MBB->addSuccessor(EndMBB);
18003 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
18004 // that was just emitted, but clearly shouldn't be "saved".
18005 assert((MI->getNumOperands() <= 3 ||
18006 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
18007 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
18008 && "Expected last argument to be EFLAGS");
18009 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
18010 // In the XMM save block, save all the XMM argument registers.
18011 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
18012 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
18013 MachineMemOperand *MMO =
18014 F->getMachineMemOperand(
18015 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
18016 MachineMemOperand::MOStore,
18017 /*Size=*/16, /*Align=*/16);
18018 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
18019 .addFrameIndex(RegSaveFrameIndex)
18020 .addImm(/*Scale=*/1)
18021 .addReg(/*IndexReg=*/0)
18022 .addImm(/*Disp=*/Offset)
18023 .addReg(/*Segment=*/0)
18024 .addReg(MI->getOperand(i).getReg())
18025 .addMemOperand(MMO);
18028 MI->eraseFromParent(); // The pseudo instruction is gone now.
18033 // The EFLAGS operand of SelectItr might be missing a kill marker
18034 // because there were multiple uses of EFLAGS, and ISel didn't know
18035 // which to mark. Figure out whether SelectItr should have had a
18036 // kill marker, and set it if it should. Returns the correct kill
18038 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
18039 MachineBasicBlock* BB,
18040 const TargetRegisterInfo* TRI) {
18041 // Scan forward through BB for a use/def of EFLAGS.
18042 MachineBasicBlock::iterator miI(std::next(SelectItr));
18043 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
18044 const MachineInstr& mi = *miI;
18045 if (mi.readsRegister(X86::EFLAGS))
18047 if (mi.definesRegister(X86::EFLAGS))
18048 break; // Should have kill-flag - update below.
18051 // If we hit the end of the block, check whether EFLAGS is live into a
18053 if (miI == BB->end()) {
18054 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
18055 sEnd = BB->succ_end();
18056 sItr != sEnd; ++sItr) {
18057 MachineBasicBlock* succ = *sItr;
18058 if (succ->isLiveIn(X86::EFLAGS))
18063 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
18064 // out. SelectMI should have a kill flag on EFLAGS.
18065 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
18069 MachineBasicBlock *
18070 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
18071 MachineBasicBlock *BB) const {
18072 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18073 DebugLoc DL = MI->getDebugLoc();
18075 // To "insert" a SELECT_CC instruction, we actually have to insert the
18076 // diamond control-flow pattern. The incoming instruction knows the
18077 // destination vreg to set, the condition code register to branch on, the
18078 // true/false values to select between, and a branch opcode to use.
18079 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18080 MachineFunction::iterator It = BB;
18086 // cmpTY ccX, r1, r2
18088 // fallthrough --> copy0MBB
18089 MachineBasicBlock *thisMBB = BB;
18090 MachineFunction *F = BB->getParent();
18091 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
18092 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
18093 F->insert(It, copy0MBB);
18094 F->insert(It, sinkMBB);
18096 // If the EFLAGS register isn't dead in the terminator, then claim that it's
18097 // live into the sink and copy blocks.
18098 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
18099 if (!MI->killsRegister(X86::EFLAGS) &&
18100 !checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
18101 copy0MBB->addLiveIn(X86::EFLAGS);
18102 sinkMBB->addLiveIn(X86::EFLAGS);
18105 // Transfer the remainder of BB and its successor edges to sinkMBB.
18106 sinkMBB->splice(sinkMBB->begin(), BB,
18107 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18108 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
18110 // Add the true and fallthrough blocks as its successors.
18111 BB->addSuccessor(copy0MBB);
18112 BB->addSuccessor(sinkMBB);
18114 // Create the conditional branch instruction.
18116 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
18117 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
18120 // %FalseValue = ...
18121 // # fallthrough to sinkMBB
18122 copy0MBB->addSuccessor(sinkMBB);
18125 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
18127 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18128 TII->get(X86::PHI), MI->getOperand(0).getReg())
18129 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
18130 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
18132 MI->eraseFromParent(); // The pseudo instruction is gone now.
18136 MachineBasicBlock *
18137 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
18138 MachineBasicBlock *BB) const {
18139 MachineFunction *MF = BB->getParent();
18140 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18141 DebugLoc DL = MI->getDebugLoc();
18142 const BasicBlock *LLVM_BB = BB->getBasicBlock();
18144 assert(MF->shouldSplitStack());
18146 const bool Is64Bit = Subtarget->is64Bit();
18147 const bool IsLP64 = Subtarget->isTarget64BitLP64();
18149 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
18150 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
18153 // ... [Till the alloca]
18154 // If stacklet is not large enough, jump to mallocMBB
18157 // Allocate by subtracting from RSP
18158 // Jump to continueMBB
18161 // Allocate by call to runtime
18165 // [rest of original BB]
18168 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18169 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18170 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
18172 MachineRegisterInfo &MRI = MF->getRegInfo();
18173 const TargetRegisterClass *AddrRegClass =
18174 getRegClassFor(getPointerTy());
18176 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18177 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
18178 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
18179 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
18180 sizeVReg = MI->getOperand(1).getReg(),
18181 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
18183 MachineFunction::iterator MBBIter = BB;
18186 MF->insert(MBBIter, bumpMBB);
18187 MF->insert(MBBIter, mallocMBB);
18188 MF->insert(MBBIter, continueMBB);
18190 continueMBB->splice(continueMBB->begin(), BB,
18191 std::next(MachineBasicBlock::iterator(MI)), BB->end());
18192 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
18194 // Add code to the main basic block to check if the stack limit has been hit,
18195 // and if so, jump to mallocMBB otherwise to bumpMBB.
18196 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
18197 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
18198 .addReg(tmpSPVReg).addReg(sizeVReg);
18199 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
18200 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
18201 .addReg(SPLimitVReg);
18202 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
18204 // bumpMBB simply decreases the stack pointer, since we know the current
18205 // stacklet has enough space.
18206 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
18207 .addReg(SPLimitVReg);
18208 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
18209 .addReg(SPLimitVReg);
18210 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18212 // Calls into a routine in libgcc to allocate more space from the heap.
18213 const uint32_t *RegMask =
18214 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18216 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
18218 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18219 .addExternalSymbol("__morestack_allocate_stack_space")
18220 .addRegMask(RegMask)
18221 .addReg(X86::RDI, RegState::Implicit)
18222 .addReg(X86::RAX, RegState::ImplicitDefine);
18223 } else if (Is64Bit) {
18224 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
18226 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
18227 .addExternalSymbol("__morestack_allocate_stack_space")
18228 .addRegMask(RegMask)
18229 .addReg(X86::EDI, RegState::Implicit)
18230 .addReg(X86::EAX, RegState::ImplicitDefine);
18232 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
18234 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
18235 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
18236 .addExternalSymbol("__morestack_allocate_stack_space")
18237 .addRegMask(RegMask)
18238 .addReg(X86::EAX, RegState::ImplicitDefine);
18242 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
18245 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
18246 .addReg(IsLP64 ? X86::RAX : X86::EAX);
18247 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
18249 // Set up the CFG correctly.
18250 BB->addSuccessor(bumpMBB);
18251 BB->addSuccessor(mallocMBB);
18252 mallocMBB->addSuccessor(continueMBB);
18253 bumpMBB->addSuccessor(continueMBB);
18255 // Take care of the PHI nodes.
18256 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
18257 MI->getOperand(0).getReg())
18258 .addReg(mallocPtrVReg).addMBB(mallocMBB)
18259 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
18261 // Delete the original pseudo instruction.
18262 MI->eraseFromParent();
18265 return continueMBB;
18268 MachineBasicBlock *
18269 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
18270 MachineBasicBlock *BB) const {
18271 DebugLoc DL = MI->getDebugLoc();
18273 assert(!Subtarget->isTargetMachO());
18275 X86FrameLowering::emitStackProbeCall(*BB->getParent(), *BB, MI, DL);
18277 MI->eraseFromParent(); // The pseudo instruction is gone now.
18281 MachineBasicBlock *
18282 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
18283 MachineBasicBlock *BB) const {
18284 // This is pretty easy. We're taking the value that we received from
18285 // our load from the relocation, sticking it in either RDI (x86-64)
18286 // or EAX and doing an indirect call. The return value will then
18287 // be in the normal return register.
18288 MachineFunction *F = BB->getParent();
18289 const X86InstrInfo *TII = Subtarget->getInstrInfo();
18290 DebugLoc DL = MI->getDebugLoc();
18292 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
18293 assert(MI->getOperand(3).isGlobal() && "This should be a global");
18295 // Get a register mask for the lowered call.
18296 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
18297 // proper register mask.
18298 const uint32_t *RegMask =
18299 Subtarget->getRegisterInfo()->getCallPreservedMask(CallingConv::C);
18300 if (Subtarget->is64Bit()) {
18301 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18302 TII->get(X86::MOV64rm), X86::RDI)
18304 .addImm(0).addReg(0)
18305 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18306 MI->getOperand(3).getTargetFlags())
18308 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
18309 addDirectMem(MIB, X86::RDI);
18310 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
18311 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
18312 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18313 TII->get(X86::MOV32rm), X86::EAX)
18315 .addImm(0).addReg(0)
18316 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18317 MI->getOperand(3).getTargetFlags())
18319 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18320 addDirectMem(MIB, X86::EAX);
18321 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18323 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
18324 TII->get(X86::MOV32rm), X86::EAX)
18325 .addReg(TII->getGlobalBaseReg(F))
18326 .addImm(0).addReg(0)
18327 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
18328 MI->getOperand(3).getTargetFlags())
18330 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
18331 addDirectMem(MIB, X86::EAX);
18332 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
18335 MI->eraseFromParent(); // The pseudo instruction is gone now.
18339 MachineBasicBlock *
18340 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
18341 MachineBasicBlock *MBB) const {
18342 DebugLoc DL = MI->getDebugLoc();
18343 MachineFunction *MF = MBB->getParent();
18344 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18345 MachineRegisterInfo &MRI = MF->getRegInfo();
18347 const BasicBlock *BB = MBB->getBasicBlock();
18348 MachineFunction::iterator I = MBB;
18351 // Memory Reference
18352 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18353 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18356 unsigned MemOpndSlot = 0;
18358 unsigned CurOp = 0;
18360 DstReg = MI->getOperand(CurOp++).getReg();
18361 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
18362 assert(RC->hasType(MVT::i32) && "Invalid destination!");
18363 unsigned mainDstReg = MRI.createVirtualRegister(RC);
18364 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
18366 MemOpndSlot = CurOp;
18368 MVT PVT = getPointerTy();
18369 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18370 "Invalid Pointer Size!");
18372 // For v = setjmp(buf), we generate
18375 // buf[LabelOffset] = restoreMBB
18376 // SjLjSetup restoreMBB
18382 // v = phi(main, restore)
18385 // if base pointer being used, load it from frame
18388 MachineBasicBlock *thisMBB = MBB;
18389 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
18390 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
18391 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
18392 MF->insert(I, mainMBB);
18393 MF->insert(I, sinkMBB);
18394 MF->push_back(restoreMBB);
18396 MachineInstrBuilder MIB;
18398 // Transfer the remainder of BB and its successor edges to sinkMBB.
18399 sinkMBB->splice(sinkMBB->begin(), MBB,
18400 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
18401 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
18404 unsigned PtrStoreOpc = 0;
18405 unsigned LabelReg = 0;
18406 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18407 Reloc::Model RM = MF->getTarget().getRelocationModel();
18408 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
18409 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
18411 // Prepare IP either in reg or imm.
18412 if (!UseImmLabel) {
18413 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
18414 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
18415 LabelReg = MRI.createVirtualRegister(PtrRC);
18416 if (Subtarget->is64Bit()) {
18417 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
18421 .addMBB(restoreMBB)
18424 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
18425 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
18426 .addReg(XII->getGlobalBaseReg(MF))
18429 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
18433 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
18435 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
18436 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18437 if (i == X86::AddrDisp)
18438 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
18440 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
18443 MIB.addReg(LabelReg);
18445 MIB.addMBB(restoreMBB);
18446 MIB.setMemRefs(MMOBegin, MMOEnd);
18448 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
18449 .addMBB(restoreMBB);
18451 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18452 MIB.addRegMask(RegInfo->getNoPreservedMask());
18453 thisMBB->addSuccessor(mainMBB);
18454 thisMBB->addSuccessor(restoreMBB);
18458 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
18459 mainMBB->addSuccessor(sinkMBB);
18462 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
18463 TII->get(X86::PHI), DstReg)
18464 .addReg(mainDstReg).addMBB(mainMBB)
18465 .addReg(restoreDstReg).addMBB(restoreMBB);
18468 if (RegInfo->hasBasePointer(*MF)) {
18469 const bool Uses64BitFramePtr =
18470 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
18471 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
18472 X86FI->setRestoreBasePointer(MF);
18473 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
18474 unsigned BasePtr = RegInfo->getBaseRegister();
18475 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
18476 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
18477 FramePtr, true, X86FI->getRestoreBasePointerOffset())
18478 .setMIFlag(MachineInstr::FrameSetup);
18480 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
18481 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
18482 restoreMBB->addSuccessor(sinkMBB);
18484 MI->eraseFromParent();
18488 MachineBasicBlock *
18489 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
18490 MachineBasicBlock *MBB) const {
18491 DebugLoc DL = MI->getDebugLoc();
18492 MachineFunction *MF = MBB->getParent();
18493 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18494 MachineRegisterInfo &MRI = MF->getRegInfo();
18496 // Memory Reference
18497 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
18498 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
18500 MVT PVT = getPointerTy();
18501 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
18502 "Invalid Pointer Size!");
18504 const TargetRegisterClass *RC =
18505 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
18506 unsigned Tmp = MRI.createVirtualRegister(RC);
18507 // Since FP is only updated here but NOT referenced, it's treated as GPR.
18508 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
18509 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
18510 unsigned SP = RegInfo->getStackRegister();
18512 MachineInstrBuilder MIB;
18514 const int64_t LabelOffset = 1 * PVT.getStoreSize();
18515 const int64_t SPOffset = 2 * PVT.getStoreSize();
18517 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
18518 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
18521 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
18522 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
18523 MIB.addOperand(MI->getOperand(i));
18524 MIB.setMemRefs(MMOBegin, MMOEnd);
18526 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
18527 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18528 if (i == X86::AddrDisp)
18529 MIB.addDisp(MI->getOperand(i), LabelOffset);
18531 MIB.addOperand(MI->getOperand(i));
18533 MIB.setMemRefs(MMOBegin, MMOEnd);
18535 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
18536 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
18537 if (i == X86::AddrDisp)
18538 MIB.addDisp(MI->getOperand(i), SPOffset);
18540 MIB.addOperand(MI->getOperand(i));
18542 MIB.setMemRefs(MMOBegin, MMOEnd);
18544 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
18546 MI->eraseFromParent();
18550 // Replace 213-type (isel default) FMA3 instructions with 231-type for
18551 // accumulator loops. Writing back to the accumulator allows the coalescer
18552 // to remove extra copies in the loop.
18553 MachineBasicBlock *
18554 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
18555 MachineBasicBlock *MBB) const {
18556 MachineOperand &AddendOp = MI->getOperand(3);
18558 // Bail out early if the addend isn't a register - we can't switch these.
18559 if (!AddendOp.isReg())
18562 MachineFunction &MF = *MBB->getParent();
18563 MachineRegisterInfo &MRI = MF.getRegInfo();
18565 // Check whether the addend is defined by a PHI:
18566 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
18567 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
18568 if (!AddendDef.isPHI())
18571 // Look for the following pattern:
18573 // %addend = phi [%entry, 0], [%loop, %result]
18575 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
18579 // %addend = phi [%entry, 0], [%loop, %result]
18581 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
18583 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
18584 assert(AddendDef.getOperand(i).isReg());
18585 MachineOperand PHISrcOp = AddendDef.getOperand(i);
18586 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
18587 if (&PHISrcInst == MI) {
18588 // Found a matching instruction.
18589 unsigned NewFMAOpc = 0;
18590 switch (MI->getOpcode()) {
18591 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
18592 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
18593 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
18594 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
18595 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
18596 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
18597 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
18598 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
18599 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
18600 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
18601 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
18602 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
18603 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
18604 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
18605 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
18606 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
18607 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
18608 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
18609 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
18610 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
18612 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
18613 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
18614 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
18615 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
18616 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
18617 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
18618 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
18619 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
18620 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
18621 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
18622 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
18623 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
18624 default: llvm_unreachable("Unrecognized FMA variant.");
18627 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
18628 MachineInstrBuilder MIB =
18629 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
18630 .addOperand(MI->getOperand(0))
18631 .addOperand(MI->getOperand(3))
18632 .addOperand(MI->getOperand(2))
18633 .addOperand(MI->getOperand(1));
18634 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
18635 MI->eraseFromParent();
18642 MachineBasicBlock *
18643 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
18644 MachineBasicBlock *BB) const {
18645 switch (MI->getOpcode()) {
18646 default: llvm_unreachable("Unexpected instr type to insert");
18647 case X86::TAILJMPd64:
18648 case X86::TAILJMPr64:
18649 case X86::TAILJMPm64:
18650 case X86::TAILJMPd64_REX:
18651 case X86::TAILJMPr64_REX:
18652 case X86::TAILJMPm64_REX:
18653 llvm_unreachable("TAILJMP64 would not be touched here.");
18654 case X86::TCRETURNdi64:
18655 case X86::TCRETURNri64:
18656 case X86::TCRETURNmi64:
18658 case X86::WIN_ALLOCA:
18659 return EmitLoweredWinAlloca(MI, BB);
18660 case X86::SEG_ALLOCA_32:
18661 case X86::SEG_ALLOCA_64:
18662 return EmitLoweredSegAlloca(MI, BB);
18663 case X86::TLSCall_32:
18664 case X86::TLSCall_64:
18665 return EmitLoweredTLSCall(MI, BB);
18666 case X86::CMOV_GR8:
18667 case X86::CMOV_FR32:
18668 case X86::CMOV_FR64:
18669 case X86::CMOV_V4F32:
18670 case X86::CMOV_V2F64:
18671 case X86::CMOV_V2I64:
18672 case X86::CMOV_V8F32:
18673 case X86::CMOV_V4F64:
18674 case X86::CMOV_V4I64:
18675 case X86::CMOV_V16F32:
18676 case X86::CMOV_V8F64:
18677 case X86::CMOV_V8I64:
18678 case X86::CMOV_GR16:
18679 case X86::CMOV_GR32:
18680 case X86::CMOV_RFP32:
18681 case X86::CMOV_RFP64:
18682 case X86::CMOV_RFP80:
18683 return EmitLoweredSelect(MI, BB);
18685 case X86::FP32_TO_INT16_IN_MEM:
18686 case X86::FP32_TO_INT32_IN_MEM:
18687 case X86::FP32_TO_INT64_IN_MEM:
18688 case X86::FP64_TO_INT16_IN_MEM:
18689 case X86::FP64_TO_INT32_IN_MEM:
18690 case X86::FP64_TO_INT64_IN_MEM:
18691 case X86::FP80_TO_INT16_IN_MEM:
18692 case X86::FP80_TO_INT32_IN_MEM:
18693 case X86::FP80_TO_INT64_IN_MEM: {
18694 MachineFunction *F = BB->getParent();
18695 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
18696 DebugLoc DL = MI->getDebugLoc();
18698 // Change the floating point control register to use "round towards zero"
18699 // mode when truncating to an integer value.
18700 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
18701 addFrameReference(BuildMI(*BB, MI, DL,
18702 TII->get(X86::FNSTCW16m)), CWFrameIdx);
18704 // Load the old value of the high byte of the control word...
18706 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
18707 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
18710 // Set the high part to be round to zero...
18711 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
18714 // Reload the modified control word now...
18715 addFrameReference(BuildMI(*BB, MI, DL,
18716 TII->get(X86::FLDCW16m)), CWFrameIdx);
18718 // Restore the memory image of control word to original value
18719 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
18722 // Get the X86 opcode to use.
18724 switch (MI->getOpcode()) {
18725 default: llvm_unreachable("illegal opcode!");
18726 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
18727 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
18728 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
18729 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
18730 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
18731 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
18732 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
18733 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
18734 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
18738 MachineOperand &Op = MI->getOperand(0);
18740 AM.BaseType = X86AddressMode::RegBase;
18741 AM.Base.Reg = Op.getReg();
18743 AM.BaseType = X86AddressMode::FrameIndexBase;
18744 AM.Base.FrameIndex = Op.getIndex();
18746 Op = MI->getOperand(1);
18748 AM.Scale = Op.getImm();
18749 Op = MI->getOperand(2);
18751 AM.IndexReg = Op.getImm();
18752 Op = MI->getOperand(3);
18753 if (Op.isGlobal()) {
18754 AM.GV = Op.getGlobal();
18756 AM.Disp = Op.getImm();
18758 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
18759 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
18761 // Reload the original control word now.
18762 addFrameReference(BuildMI(*BB, MI, DL,
18763 TII->get(X86::FLDCW16m)), CWFrameIdx);
18765 MI->eraseFromParent(); // The pseudo instruction is gone now.
18768 // String/text processing lowering.
18769 case X86::PCMPISTRM128REG:
18770 case X86::VPCMPISTRM128REG:
18771 case X86::PCMPISTRM128MEM:
18772 case X86::VPCMPISTRM128MEM:
18773 case X86::PCMPESTRM128REG:
18774 case X86::VPCMPESTRM128REG:
18775 case X86::PCMPESTRM128MEM:
18776 case X86::VPCMPESTRM128MEM:
18777 assert(Subtarget->hasSSE42() &&
18778 "Target must have SSE4.2 or AVX features enabled");
18779 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
18781 // String/text processing lowering.
18782 case X86::PCMPISTRIREG:
18783 case X86::VPCMPISTRIREG:
18784 case X86::PCMPISTRIMEM:
18785 case X86::VPCMPISTRIMEM:
18786 case X86::PCMPESTRIREG:
18787 case X86::VPCMPESTRIREG:
18788 case X86::PCMPESTRIMEM:
18789 case X86::VPCMPESTRIMEM:
18790 assert(Subtarget->hasSSE42() &&
18791 "Target must have SSE4.2 or AVX features enabled");
18792 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
18794 // Thread synchronization.
18796 return EmitMonitor(MI, BB, Subtarget);
18800 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
18802 case X86::VASTART_SAVE_XMM_REGS:
18803 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
18805 case X86::VAARG_64:
18806 return EmitVAARG64WithCustomInserter(MI, BB);
18808 case X86::EH_SjLj_SetJmp32:
18809 case X86::EH_SjLj_SetJmp64:
18810 return emitEHSjLjSetJmp(MI, BB);
18812 case X86::EH_SjLj_LongJmp32:
18813 case X86::EH_SjLj_LongJmp64:
18814 return emitEHSjLjLongJmp(MI, BB);
18816 case TargetOpcode::STATEPOINT:
18817 // As an implementation detail, STATEPOINT shares the STACKMAP format at
18818 // this point in the process. We diverge later.
18819 return emitPatchPoint(MI, BB);
18821 case TargetOpcode::STACKMAP:
18822 case TargetOpcode::PATCHPOINT:
18823 return emitPatchPoint(MI, BB);
18825 case X86::VFMADDPDr213r:
18826 case X86::VFMADDPSr213r:
18827 case X86::VFMADDSDr213r:
18828 case X86::VFMADDSSr213r:
18829 case X86::VFMSUBPDr213r:
18830 case X86::VFMSUBPSr213r:
18831 case X86::VFMSUBSDr213r:
18832 case X86::VFMSUBSSr213r:
18833 case X86::VFNMADDPDr213r:
18834 case X86::VFNMADDPSr213r:
18835 case X86::VFNMADDSDr213r:
18836 case X86::VFNMADDSSr213r:
18837 case X86::VFNMSUBPDr213r:
18838 case X86::VFNMSUBPSr213r:
18839 case X86::VFNMSUBSDr213r:
18840 case X86::VFNMSUBSSr213r:
18841 case X86::VFMADDSUBPDr213r:
18842 case X86::VFMADDSUBPSr213r:
18843 case X86::VFMSUBADDPDr213r:
18844 case X86::VFMSUBADDPSr213r:
18845 case X86::VFMADDPDr213rY:
18846 case X86::VFMADDPSr213rY:
18847 case X86::VFMSUBPDr213rY:
18848 case X86::VFMSUBPSr213rY:
18849 case X86::VFNMADDPDr213rY:
18850 case X86::VFNMADDPSr213rY:
18851 case X86::VFNMSUBPDr213rY:
18852 case X86::VFNMSUBPSr213rY:
18853 case X86::VFMADDSUBPDr213rY:
18854 case X86::VFMADDSUBPSr213rY:
18855 case X86::VFMSUBADDPDr213rY:
18856 case X86::VFMSUBADDPSr213rY:
18857 return emitFMA3Instr(MI, BB);
18861 //===----------------------------------------------------------------------===//
18862 // X86 Optimization Hooks
18863 //===----------------------------------------------------------------------===//
18865 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
18868 const SelectionDAG &DAG,
18869 unsigned Depth) const {
18870 unsigned BitWidth = KnownZero.getBitWidth();
18871 unsigned Opc = Op.getOpcode();
18872 assert((Opc >= ISD::BUILTIN_OP_END ||
18873 Opc == ISD::INTRINSIC_WO_CHAIN ||
18874 Opc == ISD::INTRINSIC_W_CHAIN ||
18875 Opc == ISD::INTRINSIC_VOID) &&
18876 "Should use MaskedValueIsZero if you don't know whether Op"
18877 " is a target node!");
18879 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
18893 // These nodes' second result is a boolean.
18894 if (Op.getResNo() == 0)
18897 case X86ISD::SETCC:
18898 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
18900 case ISD::INTRINSIC_WO_CHAIN: {
18901 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
18902 unsigned NumLoBits = 0;
18905 case Intrinsic::x86_sse_movmsk_ps:
18906 case Intrinsic::x86_avx_movmsk_ps_256:
18907 case Intrinsic::x86_sse2_movmsk_pd:
18908 case Intrinsic::x86_avx_movmsk_pd_256:
18909 case Intrinsic::x86_mmx_pmovmskb:
18910 case Intrinsic::x86_sse2_pmovmskb_128:
18911 case Intrinsic::x86_avx2_pmovmskb: {
18912 // High bits of movmskp{s|d}, pmovmskb are known zero.
18914 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
18915 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
18916 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
18917 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
18918 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
18919 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
18920 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
18921 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
18923 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
18932 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
18934 const SelectionDAG &,
18935 unsigned Depth) const {
18936 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
18937 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
18938 return Op.getValueType().getScalarType().getSizeInBits();
18944 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
18945 /// node is a GlobalAddress + offset.
18946 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
18947 const GlobalValue* &GA,
18948 int64_t &Offset) const {
18949 if (N->getOpcode() == X86ISD::Wrapper) {
18950 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
18951 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
18952 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
18956 return TargetLowering::isGAPlusOffset(N, GA, Offset);
18959 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
18960 /// same as extracting the high 128-bit part of 256-bit vector and then
18961 /// inserting the result into the low part of a new 256-bit vector
18962 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
18963 EVT VT = SVOp->getValueType(0);
18964 unsigned NumElems = VT.getVectorNumElements();
18966 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
18967 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
18968 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18969 SVOp->getMaskElt(j) >= 0)
18975 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
18976 /// same as extracting the low 128-bit part of 256-bit vector and then
18977 /// inserting the result into the high part of a new 256-bit vector
18978 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
18979 EVT VT = SVOp->getValueType(0);
18980 unsigned NumElems = VT.getVectorNumElements();
18982 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
18983 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
18984 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
18985 SVOp->getMaskElt(j) >= 0)
18991 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
18992 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
18993 TargetLowering::DAGCombinerInfo &DCI,
18994 const X86Subtarget* Subtarget) {
18996 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
18997 SDValue V1 = SVOp->getOperand(0);
18998 SDValue V2 = SVOp->getOperand(1);
18999 EVT VT = SVOp->getValueType(0);
19000 unsigned NumElems = VT.getVectorNumElements();
19002 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
19003 V2.getOpcode() == ISD::CONCAT_VECTORS) {
19007 // V UNDEF BUILD_VECTOR UNDEF
19009 // CONCAT_VECTOR CONCAT_VECTOR
19012 // RESULT: V + zero extended
19014 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
19015 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
19016 V1.getOperand(1).getOpcode() != ISD::UNDEF)
19019 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
19022 // To match the shuffle mask, the first half of the mask should
19023 // be exactly the first vector, and all the rest a splat with the
19024 // first element of the second one.
19025 for (unsigned i = 0; i != NumElems/2; ++i)
19026 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
19027 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
19030 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
19031 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
19032 if (Ld->hasNUsesOfValue(1, 0)) {
19033 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
19034 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
19036 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
19038 Ld->getPointerInfo(),
19039 Ld->getAlignment(),
19040 false/*isVolatile*/, true/*ReadMem*/,
19041 false/*WriteMem*/);
19043 // Make sure the newly-created LOAD is in the same position as Ld in
19044 // terms of dependency. We create a TokenFactor for Ld and ResNode,
19045 // and update uses of Ld's output chain to use the TokenFactor.
19046 if (Ld->hasAnyUseOfValue(1)) {
19047 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
19048 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
19049 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
19050 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
19051 SDValue(ResNode.getNode(), 1));
19054 return DAG.getNode(ISD::BITCAST, dl, VT, ResNode);
19058 // Emit a zeroed vector and insert the desired subvector on its
19060 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
19061 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
19062 return DCI.CombineTo(N, InsV);
19065 //===--------------------------------------------------------------------===//
19066 // Combine some shuffles into subvector extracts and inserts:
19069 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
19070 if (isShuffleHigh128VectorInsertLow(SVOp)) {
19071 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
19072 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
19073 return DCI.CombineTo(N, InsV);
19076 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
19077 if (isShuffleLow128VectorInsertHigh(SVOp)) {
19078 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
19079 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
19080 return DCI.CombineTo(N, InsV);
19086 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
19089 /// This is the leaf of the recursive combinine below. When we have found some
19090 /// chain of single-use x86 shuffle instructions and accumulated the combined
19091 /// shuffle mask represented by them, this will try to pattern match that mask
19092 /// into either a single instruction if there is a special purpose instruction
19093 /// for this operation, or into a PSHUFB instruction which is a fully general
19094 /// instruction but should only be used to replace chains over a certain depth.
19095 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
19096 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
19097 TargetLowering::DAGCombinerInfo &DCI,
19098 const X86Subtarget *Subtarget) {
19099 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
19101 // Find the operand that enters the chain. Note that multiple uses are OK
19102 // here, we're not going to remove the operand we find.
19103 SDValue Input = Op.getOperand(0);
19104 while (Input.getOpcode() == ISD::BITCAST)
19105 Input = Input.getOperand(0);
19107 MVT VT = Input.getSimpleValueType();
19108 MVT RootVT = Root.getSimpleValueType();
19111 // Just remove no-op shuffle masks.
19112 if (Mask.size() == 1) {
19113 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
19118 // Use the float domain if the operand type is a floating point type.
19119 bool FloatDomain = VT.isFloatingPoint();
19121 // For floating point shuffles, we don't have free copies in the shuffle
19122 // instructions or the ability to load as part of the instruction, so
19123 // canonicalize their shuffles to UNPCK or MOV variants.
19125 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
19126 // vectors because it can have a load folded into it that UNPCK cannot. This
19127 // doesn't preclude something switching to the shorter encoding post-RA.
19129 if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
19130 bool Lo = Mask.equals(0, 0);
19133 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
19134 // is no slower than UNPCKLPD but has the option to fold the input operand
19135 // into even an unaligned memory load.
19136 if (Lo && Subtarget->hasSSE3()) {
19137 Shuffle = X86ISD::MOVDDUP;
19138 ShuffleVT = MVT::v2f64;
19140 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
19141 // than the UNPCK variants.
19142 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
19143 ShuffleVT = MVT::v4f32;
19145 if (Depth == 1 && Root->getOpcode() == Shuffle)
19146 return false; // Nothing to do!
19147 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19148 DCI.AddToWorklist(Op.getNode());
19149 if (Shuffle == X86ISD::MOVDDUP)
19150 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19152 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19153 DCI.AddToWorklist(Op.getNode());
19154 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19158 if (Subtarget->hasSSE3() &&
19159 (Mask.equals(0, 0, 2, 2) || Mask.equals(1, 1, 3, 3))) {
19160 bool Lo = Mask.equals(0, 0, 2, 2);
19161 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
19162 MVT ShuffleVT = MVT::v4f32;
19163 if (Depth == 1 && Root->getOpcode() == Shuffle)
19164 return false; // Nothing to do!
19165 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19166 DCI.AddToWorklist(Op.getNode());
19167 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
19168 DCI.AddToWorklist(Op.getNode());
19169 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19173 if (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3)) {
19174 bool Lo = Mask.equals(0, 0, 1, 1);
19175 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19176 MVT ShuffleVT = MVT::v4f32;
19177 if (Depth == 1 && Root->getOpcode() == Shuffle)
19178 return false; // Nothing to do!
19179 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19180 DCI.AddToWorklist(Op.getNode());
19181 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19182 DCI.AddToWorklist(Op.getNode());
19183 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19189 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
19190 // variants as none of these have single-instruction variants that are
19191 // superior to the UNPCK formulation.
19192 if (!FloatDomain &&
19193 (Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
19194 Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
19195 Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
19196 Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
19198 bool Lo = Mask[0] == 0;
19199 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
19200 if (Depth == 1 && Root->getOpcode() == Shuffle)
19201 return false; // Nothing to do!
19203 switch (Mask.size()) {
19205 ShuffleVT = MVT::v8i16;
19208 ShuffleVT = MVT::v16i8;
19211 llvm_unreachable("Impossible mask size!");
19213 Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
19214 DCI.AddToWorklist(Op.getNode());
19215 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
19216 DCI.AddToWorklist(Op.getNode());
19217 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19222 // Don't try to re-form single instruction chains under any circumstances now
19223 // that we've done encoding canonicalization for them.
19227 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
19228 // can replace them with a single PSHUFB instruction profitably. Intel's
19229 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
19230 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
19231 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
19232 SmallVector<SDValue, 16> PSHUFBMask;
19233 assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
19234 int Ratio = 16 / Mask.size();
19235 for (unsigned i = 0; i < 16; ++i) {
19236 if (Mask[i / Ratio] == SM_SentinelUndef) {
19237 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
19240 int M = Mask[i / Ratio] != SM_SentinelZero
19241 ? Ratio * Mask[i / Ratio] + i % Ratio
19243 PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
19245 Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
19246 DCI.AddToWorklist(Op.getNode());
19247 SDValue PSHUFBMaskOp =
19248 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
19249 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
19250 Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
19251 DCI.AddToWorklist(Op.getNode());
19252 DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
19257 // Failed to find any combines.
19261 /// \brief Fully generic combining of x86 shuffle instructions.
19263 /// This should be the last combine run over the x86 shuffle instructions. Once
19264 /// they have been fully optimized, this will recursively consider all chains
19265 /// of single-use shuffle instructions, build a generic model of the cumulative
19266 /// shuffle operation, and check for simpler instructions which implement this
19267 /// operation. We use this primarily for two purposes:
19269 /// 1) Collapse generic shuffles to specialized single instructions when
19270 /// equivalent. In most cases, this is just an encoding size win, but
19271 /// sometimes we will collapse multiple generic shuffles into a single
19272 /// special-purpose shuffle.
19273 /// 2) Look for sequences of shuffle instructions with 3 or more total
19274 /// instructions, and replace them with the slightly more expensive SSSE3
19275 /// PSHUFB instruction if available. We do this as the last combining step
19276 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
19277 /// a suitable short sequence of other instructions. The PHUFB will either
19278 /// use a register or have to read from memory and so is slightly (but only
19279 /// slightly) more expensive than the other shuffle instructions.
19281 /// Because this is inherently a quadratic operation (for each shuffle in
19282 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
19283 /// This should never be an issue in practice as the shuffle lowering doesn't
19284 /// produce sequences of more than 8 instructions.
19286 /// FIXME: We will currently miss some cases where the redundant shuffling
19287 /// would simplify under the threshold for PSHUFB formation because of
19288 /// combine-ordering. To fix this, we should do the redundant instruction
19289 /// combining in this recursive walk.
19290 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
19291 ArrayRef<int> RootMask,
19292 int Depth, bool HasPSHUFB,
19294 TargetLowering::DAGCombinerInfo &DCI,
19295 const X86Subtarget *Subtarget) {
19296 // Bound the depth of our recursive combine because this is ultimately
19297 // quadratic in nature.
19301 // Directly rip through bitcasts to find the underlying operand.
19302 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
19303 Op = Op.getOperand(0);
19305 MVT VT = Op.getSimpleValueType();
19306 if (!VT.isVector())
19307 return false; // Bail if we hit a non-vector.
19308 // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
19309 // version should be added.
19310 if (VT.getSizeInBits() != 128)
19313 assert(Root.getSimpleValueType().isVector() &&
19314 "Shuffles operate on vector types!");
19315 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
19316 "Can only combine shuffles of the same vector register size.");
19318 if (!isTargetShuffle(Op.getOpcode()))
19320 SmallVector<int, 16> OpMask;
19322 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
19323 // We only can combine unary shuffles which we can decode the mask for.
19324 if (!HaveMask || !IsUnary)
19327 assert(VT.getVectorNumElements() == OpMask.size() &&
19328 "Different mask size from vector size!");
19329 assert(((RootMask.size() > OpMask.size() &&
19330 RootMask.size() % OpMask.size() == 0) ||
19331 (OpMask.size() > RootMask.size() &&
19332 OpMask.size() % RootMask.size() == 0) ||
19333 OpMask.size() == RootMask.size()) &&
19334 "The smaller number of elements must divide the larger.");
19335 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
19336 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
19337 assert(((RootRatio == 1 && OpRatio == 1) ||
19338 (RootRatio == 1) != (OpRatio == 1)) &&
19339 "Must not have a ratio for both incoming and op masks!");
19341 SmallVector<int, 16> Mask;
19342 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
19344 // Merge this shuffle operation's mask into our accumulated mask. Note that
19345 // this shuffle's mask will be the first applied to the input, followed by the
19346 // root mask to get us all the way to the root value arrangement. The reason
19347 // for this order is that we are recursing up the operation chain.
19348 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
19349 int RootIdx = i / RootRatio;
19350 if (RootMask[RootIdx] < 0) {
19351 // This is a zero or undef lane, we're done.
19352 Mask.push_back(RootMask[RootIdx]);
19356 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
19357 int OpIdx = RootMaskedIdx / OpRatio;
19358 if (OpMask[OpIdx] < 0) {
19359 // The incoming lanes are zero or undef, it doesn't matter which ones we
19361 Mask.push_back(OpMask[OpIdx]);
19365 // Ok, we have non-zero lanes, map them through.
19366 Mask.push_back(OpMask[OpIdx] * OpRatio +
19367 RootMaskedIdx % OpRatio);
19370 // See if we can recurse into the operand to combine more things.
19371 switch (Op.getOpcode()) {
19372 case X86ISD::PSHUFB:
19374 case X86ISD::PSHUFD:
19375 case X86ISD::PSHUFHW:
19376 case X86ISD::PSHUFLW:
19377 if (Op.getOperand(0).hasOneUse() &&
19378 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19379 HasPSHUFB, DAG, DCI, Subtarget))
19383 case X86ISD::UNPCKL:
19384 case X86ISD::UNPCKH:
19385 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
19386 // We can't check for single use, we have to check that this shuffle is the only user.
19387 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
19388 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
19389 HasPSHUFB, DAG, DCI, Subtarget))
19394 // Minor canonicalization of the accumulated shuffle mask to make it easier
19395 // to match below. All this does is detect masks with squential pairs of
19396 // elements, and shrink them to the half-width mask. It does this in a loop
19397 // so it will reduce the size of the mask to the minimal width mask which
19398 // performs an equivalent shuffle.
19399 SmallVector<int, 16> WidenedMask;
19400 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
19401 Mask = std::move(WidenedMask);
19402 WidenedMask.clear();
19405 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
19409 /// \brief Get the PSHUF-style mask from PSHUF node.
19411 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
19412 /// PSHUF-style masks that can be reused with such instructions.
19413 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
19414 SmallVector<int, 4> Mask;
19416 bool HaveMask = getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), Mask, IsUnary);
19420 switch (N.getOpcode()) {
19421 case X86ISD::PSHUFD:
19423 case X86ISD::PSHUFLW:
19426 case X86ISD::PSHUFHW:
19427 Mask.erase(Mask.begin(), Mask.begin() + 4);
19428 for (int &M : Mask)
19432 llvm_unreachable("No valid shuffle instruction found!");
19436 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
19438 /// We walk up the chain and look for a combinable shuffle, skipping over
19439 /// shuffles that we could hoist this shuffle's transformation past without
19440 /// altering anything.
19442 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
19444 TargetLowering::DAGCombinerInfo &DCI) {
19445 assert(N.getOpcode() == X86ISD::PSHUFD &&
19446 "Called with something other than an x86 128-bit half shuffle!");
19449 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
19450 // of the shuffles in the chain so that we can form a fresh chain to replace
19452 SmallVector<SDValue, 8> Chain;
19453 SDValue V = N.getOperand(0);
19454 for (; V.hasOneUse(); V = V.getOperand(0)) {
19455 switch (V.getOpcode()) {
19457 return SDValue(); // Nothing combined!
19460 // Skip bitcasts as we always know the type for the target specific
19464 case X86ISD::PSHUFD:
19465 // Found another dword shuffle.
19468 case X86ISD::PSHUFLW:
19469 // Check that the low words (being shuffled) are the identity in the
19470 // dword shuffle, and the high words are self-contained.
19471 if (Mask[0] != 0 || Mask[1] != 1 ||
19472 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
19475 Chain.push_back(V);
19478 case X86ISD::PSHUFHW:
19479 // Check that the high words (being shuffled) are the identity in the
19480 // dword shuffle, and the low words are self-contained.
19481 if (Mask[2] != 2 || Mask[3] != 3 ||
19482 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
19485 Chain.push_back(V);
19488 case X86ISD::UNPCKL:
19489 case X86ISD::UNPCKH:
19490 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
19491 // shuffle into a preceding word shuffle.
19492 if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
19495 // Search for a half-shuffle which we can combine with.
19496 unsigned CombineOp =
19497 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
19498 if (V.getOperand(0) != V.getOperand(1) ||
19499 !V->isOnlyUserOf(V.getOperand(0).getNode()))
19501 Chain.push_back(V);
19502 V = V.getOperand(0);
19504 switch (V.getOpcode()) {
19506 return SDValue(); // Nothing to combine.
19508 case X86ISD::PSHUFLW:
19509 case X86ISD::PSHUFHW:
19510 if (V.getOpcode() == CombineOp)
19513 Chain.push_back(V);
19517 V = V.getOperand(0);
19521 } while (V.hasOneUse());
19524 // Break out of the loop if we break out of the switch.
19528 if (!V.hasOneUse())
19529 // We fell out of the loop without finding a viable combining instruction.
19532 // Merge this node's mask and our incoming mask.
19533 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19534 for (int &M : Mask)
19536 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
19537 getV4X86ShuffleImm8ForMask(Mask, DAG));
19539 // Rebuild the chain around this new shuffle.
19540 while (!Chain.empty()) {
19541 SDValue W = Chain.pop_back_val();
19543 if (V.getValueType() != W.getOperand(0).getValueType())
19544 V = DAG.getNode(ISD::BITCAST, DL, W.getOperand(0).getValueType(), V);
19546 switch (W.getOpcode()) {
19548 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
19550 case X86ISD::UNPCKL:
19551 case X86ISD::UNPCKH:
19552 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
19555 case X86ISD::PSHUFD:
19556 case X86ISD::PSHUFLW:
19557 case X86ISD::PSHUFHW:
19558 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
19562 if (V.getValueType() != N.getValueType())
19563 V = DAG.getNode(ISD::BITCAST, DL, N.getValueType(), V);
19565 // Return the new chain to replace N.
19569 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
19571 /// We walk up the chain, skipping shuffles of the other half and looking
19572 /// through shuffles which switch halves trying to find a shuffle of the same
19573 /// pair of dwords.
19574 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
19576 TargetLowering::DAGCombinerInfo &DCI) {
19578 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
19579 "Called with something other than an x86 128-bit half shuffle!");
19581 unsigned CombineOpcode = N.getOpcode();
19583 // Walk up a single-use chain looking for a combinable shuffle.
19584 SDValue V = N.getOperand(0);
19585 for (; V.hasOneUse(); V = V.getOperand(0)) {
19586 switch (V.getOpcode()) {
19588 return false; // Nothing combined!
19591 // Skip bitcasts as we always know the type for the target specific
19595 case X86ISD::PSHUFLW:
19596 case X86ISD::PSHUFHW:
19597 if (V.getOpcode() == CombineOpcode)
19600 // Other-half shuffles are no-ops.
19603 // Break out of the loop if we break out of the switch.
19607 if (!V.hasOneUse())
19608 // We fell out of the loop without finding a viable combining instruction.
19611 // Combine away the bottom node as its shuffle will be accumulated into
19612 // a preceding shuffle.
19613 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19615 // Record the old value.
19618 // Merge this node's mask and our incoming mask (adjusted to account for all
19619 // the pshufd instructions encountered).
19620 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19621 for (int &M : Mask)
19623 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
19624 getV4X86ShuffleImm8ForMask(Mask, DAG));
19626 // Check that the shuffles didn't cancel each other out. If not, we need to
19627 // combine to the new one.
19629 // Replace the combinable shuffle with the combined one, updating all users
19630 // so that we re-evaluate the chain here.
19631 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
19636 /// \brief Try to combine x86 target specific shuffles.
19637 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
19638 TargetLowering::DAGCombinerInfo &DCI,
19639 const X86Subtarget *Subtarget) {
19641 MVT VT = N.getSimpleValueType();
19642 SmallVector<int, 4> Mask;
19644 switch (N.getOpcode()) {
19645 case X86ISD::PSHUFD:
19646 case X86ISD::PSHUFLW:
19647 case X86ISD::PSHUFHW:
19648 Mask = getPSHUFShuffleMask(N);
19649 assert(Mask.size() == 4);
19655 // Nuke no-op shuffles that show up after combining.
19656 if (isNoopShuffleMask(Mask))
19657 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
19659 // Look for simplifications involving one or two shuffle instructions.
19660 SDValue V = N.getOperand(0);
19661 switch (N.getOpcode()) {
19664 case X86ISD::PSHUFLW:
19665 case X86ISD::PSHUFHW:
19666 assert(VT == MVT::v8i16);
19669 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
19670 return SDValue(); // We combined away this shuffle, so we're done.
19672 // See if this reduces to a PSHUFD which is no more expensive and can
19673 // combine with more operations. Note that it has to at least flip the
19674 // dwords as otherwise it would have been removed as a no-op.
19675 if (Mask[0] == 2 && Mask[1] == 3 && Mask[2] == 0 && Mask[3] == 1) {
19676 int DMask[] = {0, 1, 2, 3};
19677 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
19678 DMask[DOffset + 0] = DOffset + 1;
19679 DMask[DOffset + 1] = DOffset + 0;
19680 V = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, V);
19681 DCI.AddToWorklist(V.getNode());
19682 V = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V,
19683 getV4X86ShuffleImm8ForMask(DMask, DAG));
19684 DCI.AddToWorklist(V.getNode());
19685 return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
19688 // Look for shuffle patterns which can be implemented as a single unpack.
19689 // FIXME: This doesn't handle the location of the PSHUFD generically, and
19690 // only works when we have a PSHUFD followed by two half-shuffles.
19691 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
19692 (V.getOpcode() == X86ISD::PSHUFLW ||
19693 V.getOpcode() == X86ISD::PSHUFHW) &&
19694 V.getOpcode() != N.getOpcode() &&
19696 SDValue D = V.getOperand(0);
19697 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
19698 D = D.getOperand(0);
19699 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
19700 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
19701 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
19702 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19703 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
19705 for (int i = 0; i < 4; ++i) {
19706 WordMask[i + NOffset] = Mask[i] + NOffset;
19707 WordMask[i + VOffset] = VMask[i] + VOffset;
19709 // Map the word mask through the DWord mask.
19711 for (int i = 0; i < 8; ++i)
19712 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
19713 const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
19714 const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
19715 if (std::equal(std::begin(MappedMask), std::end(MappedMask),
19716 std::begin(UnpackLoMask)) ||
19717 std::equal(std::begin(MappedMask), std::end(MappedMask),
19718 std::begin(UnpackHiMask))) {
19719 // We can replace all three shuffles with an unpack.
19720 V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
19721 DCI.AddToWorklist(V.getNode());
19722 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
19724 DL, MVT::v8i16, V, V);
19731 case X86ISD::PSHUFD:
19732 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
19741 /// \brief Try to combine a shuffle into a target-specific add-sub node.
19743 /// We combine this directly on the abstract vector shuffle nodes so it is
19744 /// easier to generically match. We also insert dummy vector shuffle nodes for
19745 /// the operands which explicitly discard the lanes which are unused by this
19746 /// operation to try to flow through the rest of the combiner the fact that
19747 /// they're unused.
19748 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
19750 EVT VT = N->getValueType(0);
19752 // We only handle target-independent shuffles.
19753 // FIXME: It would be easy and harmless to use the target shuffle mask
19754 // extraction tool to support more.
19755 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
19758 auto *SVN = cast<ShuffleVectorSDNode>(N);
19759 ArrayRef<int> Mask = SVN->getMask();
19760 SDValue V1 = N->getOperand(0);
19761 SDValue V2 = N->getOperand(1);
19763 // We require the first shuffle operand to be the SUB node, and the second to
19764 // be the ADD node.
19765 // FIXME: We should support the commuted patterns.
19766 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
19769 // If there are other uses of these operations we can't fold them.
19770 if (!V1->hasOneUse() || !V2->hasOneUse())
19773 // Ensure that both operations have the same operands. Note that we can
19774 // commute the FADD operands.
19775 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
19776 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
19777 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
19780 // We're looking for blends between FADD and FSUB nodes. We insist on these
19781 // nodes being lined up in a specific expected pattern.
19782 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
19783 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
19784 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
19787 // Only specific types are legal at this point, assert so we notice if and
19788 // when these change.
19789 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
19790 VT == MVT::v4f64) &&
19791 "Unknown vector type encountered!");
19793 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
19796 /// PerformShuffleCombine - Performs several different shuffle combines.
19797 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
19798 TargetLowering::DAGCombinerInfo &DCI,
19799 const X86Subtarget *Subtarget) {
19801 SDValue N0 = N->getOperand(0);
19802 SDValue N1 = N->getOperand(1);
19803 EVT VT = N->getValueType(0);
19805 // Don't create instructions with illegal types after legalize types has run.
19806 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19807 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
19810 // If we have legalized the vector types, look for blends of FADD and FSUB
19811 // nodes that we can fuse into an ADDSUB node.
19812 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
19813 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
19816 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
19817 if (Subtarget->hasFp256() && VT.is256BitVector() &&
19818 N->getOpcode() == ISD::VECTOR_SHUFFLE)
19819 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
19821 // During Type Legalization, when promoting illegal vector types,
19822 // the backend might introduce new shuffle dag nodes and bitcasts.
19824 // This code performs the following transformation:
19825 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
19826 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
19828 // We do this only if both the bitcast and the BINOP dag nodes have
19829 // one use. Also, perform this transformation only if the new binary
19830 // operation is legal. This is to avoid introducing dag nodes that
19831 // potentially need to be further expanded (or custom lowered) into a
19832 // less optimal sequence of dag nodes.
19833 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
19834 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
19835 N0.getOpcode() == ISD::BITCAST) {
19836 SDValue BC0 = N0.getOperand(0);
19837 EVT SVT = BC0.getValueType();
19838 unsigned Opcode = BC0.getOpcode();
19839 unsigned NumElts = VT.getVectorNumElements();
19841 if (BC0.hasOneUse() && SVT.isVector() &&
19842 SVT.getVectorNumElements() * 2 == NumElts &&
19843 TLI.isOperationLegal(Opcode, VT)) {
19844 bool CanFold = false;
19856 unsigned SVTNumElts = SVT.getVectorNumElements();
19857 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
19858 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
19859 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
19860 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
19861 CanFold = SVOp->getMaskElt(i) < 0;
19864 SDValue BC00 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(0));
19865 SDValue BC01 = DAG.getNode(ISD::BITCAST, dl, VT, BC0.getOperand(1));
19866 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
19867 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
19872 // Only handle 128 wide vector from here on.
19873 if (!VT.is128BitVector())
19876 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
19877 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
19878 // consecutive, non-overlapping, and in the right order.
19879 SmallVector<SDValue, 16> Elts;
19880 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
19881 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
19883 SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true);
19887 if (isTargetShuffle(N->getOpcode())) {
19889 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
19890 if (Shuffle.getNode())
19893 // Try recursively combining arbitrary sequences of x86 shuffle
19894 // instructions into higher-order shuffles. We do this after combining
19895 // specific PSHUF instruction sequences into their minimal form so that we
19896 // can evaluate how many specialized shuffle instructions are involved in
19897 // a particular chain.
19898 SmallVector<int, 1> NonceMask; // Just a placeholder.
19899 NonceMask.push_back(0);
19900 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
19901 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
19903 return SDValue(); // This routine will use CombineTo to replace N.
19909 /// PerformTruncateCombine - Converts truncate operation to
19910 /// a sequence of vector shuffle operations.
19911 /// It is possible when we truncate 256-bit vector to 128-bit vector
19912 static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
19913 TargetLowering::DAGCombinerInfo &DCI,
19914 const X86Subtarget *Subtarget) {
19918 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
19919 /// specific shuffle of a load can be folded into a single element load.
19920 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
19921 /// shuffles have been custom lowered so we need to handle those here.
19922 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
19923 TargetLowering::DAGCombinerInfo &DCI) {
19924 if (DCI.isBeforeLegalizeOps())
19927 SDValue InVec = N->getOperand(0);
19928 SDValue EltNo = N->getOperand(1);
19930 if (!isa<ConstantSDNode>(EltNo))
19933 EVT OriginalVT = InVec.getValueType();
19935 if (InVec.getOpcode() == ISD::BITCAST) {
19936 // Don't duplicate a load with other uses.
19937 if (!InVec.hasOneUse())
19939 EVT BCVT = InVec.getOperand(0).getValueType();
19940 if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
19942 InVec = InVec.getOperand(0);
19945 EVT CurrentVT = InVec.getValueType();
19947 if (!isTargetShuffle(InVec.getOpcode()))
19950 // Don't duplicate a load with other uses.
19951 if (!InVec.hasOneUse())
19954 SmallVector<int, 16> ShuffleMask;
19956 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
19957 ShuffleMask, UnaryShuffle))
19960 // Select the input vector, guarding against out of range extract vector.
19961 unsigned NumElems = CurrentVT.getVectorNumElements();
19962 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
19963 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
19964 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
19965 : InVec.getOperand(1);
19967 // If inputs to shuffle are the same for both ops, then allow 2 uses
19968 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
19969 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
19971 if (LdNode.getOpcode() == ISD::BITCAST) {
19972 // Don't duplicate a load with other uses.
19973 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
19976 AllowedUses = 1; // only allow 1 load use if we have a bitcast
19977 LdNode = LdNode.getOperand(0);
19980 if (!ISD::isNormalLoad(LdNode.getNode()))
19983 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
19985 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
19988 EVT EltVT = N->getValueType(0);
19989 // If there's a bitcast before the shuffle, check if the load type and
19990 // alignment is valid.
19991 unsigned Align = LN0->getAlignment();
19992 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19993 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
19994 EltVT.getTypeForEVT(*DAG.getContext()));
19996 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
19999 // All checks match so transform back to vector_shuffle so that DAG combiner
20000 // can finish the job
20003 // Create shuffle node taking into account the case that its a unary shuffle
20004 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
20005 : InVec.getOperand(1);
20006 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
20007 InVec.getOperand(0), Shuffle,
20009 Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
20010 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
20014 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
20015 /// special and don't usually play with other vector types, it's better to
20016 /// handle them early to be sure we emit efficient code by avoiding
20017 /// store-load conversions.
20018 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
20019 if (N->getValueType(0) != MVT::x86mmx ||
20020 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
20021 N->getOperand(0)->getValueType(0) != MVT::v2i32)
20024 SDValue V = N->getOperand(0);
20025 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
20026 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
20027 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
20028 N->getValueType(0), V.getOperand(0));
20033 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
20034 /// generation and convert it from being a bunch of shuffles and extracts
20035 /// into a somewhat faster sequence. For i686, the best sequence is apparently
20036 /// storing the value and loading scalars back, while for x64 we should
20037 /// use 64-bit extracts and shifts.
20038 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
20039 TargetLowering::DAGCombinerInfo &DCI) {
20040 SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI);
20041 if (NewOp.getNode())
20044 SDValue InputVector = N->getOperand(0);
20046 // Detect mmx to i32 conversion through a v2i32 elt extract.
20047 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
20048 N->getValueType(0) == MVT::i32 &&
20049 InputVector.getValueType() == MVT::v2i32) {
20051 // The bitcast source is a direct mmx result.
20052 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
20053 if (MMXSrc.getValueType() == MVT::x86mmx)
20054 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20055 N->getValueType(0),
20056 InputVector.getNode()->getOperand(0));
20058 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
20059 SDValue MMXSrcOp = MMXSrc.getOperand(0);
20060 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
20061 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
20062 MMXSrcOp.getOpcode() == ISD::BITCAST &&
20063 MMXSrcOp.getValueType() == MVT::v1i64 &&
20064 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
20065 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
20066 N->getValueType(0),
20067 MMXSrcOp.getOperand(0));
20070 // Only operate on vectors of 4 elements, where the alternative shuffling
20071 // gets to be more expensive.
20072 if (InputVector.getValueType() != MVT::v4i32)
20075 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
20076 // single use which is a sign-extend or zero-extend, and all elements are
20078 SmallVector<SDNode *, 4> Uses;
20079 unsigned ExtractedElements = 0;
20080 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
20081 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
20082 if (UI.getUse().getResNo() != InputVector.getResNo())
20085 SDNode *Extract = *UI;
20086 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
20089 if (Extract->getValueType(0) != MVT::i32)
20091 if (!Extract->hasOneUse())
20093 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
20094 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
20096 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
20099 // Record which element was extracted.
20100 ExtractedElements |=
20101 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
20103 Uses.push_back(Extract);
20106 // If not all the elements were used, this may not be worthwhile.
20107 if (ExtractedElements != 15)
20110 // Ok, we've now decided to do the transformation.
20111 // If 64-bit shifts are legal, use the extract-shift sequence,
20112 // otherwise bounce the vector off the cache.
20113 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20115 SDLoc dl(InputVector);
20117 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
20118 SDValue Cst = DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, InputVector);
20119 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy();
20120 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20121 DAG.getConstant(0, VecIdxTy));
20122 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
20123 DAG.getConstant(1, VecIdxTy));
20125 SDValue ShAmt = DAG.getConstant(32,
20126 DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64));
20127 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
20128 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20129 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
20130 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
20131 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
20132 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
20134 // Store the value to a temporary stack slot.
20135 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
20136 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
20137 MachinePointerInfo(), false, false, 0);
20139 EVT ElementType = InputVector.getValueType().getVectorElementType();
20140 unsigned EltSize = ElementType.getSizeInBits() / 8;
20142 // Replace each use (extract) with a load of the appropriate element.
20143 for (unsigned i = 0; i < 4; ++i) {
20144 uint64_t Offset = EltSize * i;
20145 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
20147 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
20148 StackPtr, OffsetVal);
20150 // Load the scalar.
20151 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
20152 ScalarAddr, MachinePointerInfo(),
20153 false, false, false, 0);
20158 // Replace the extracts
20159 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
20160 UE = Uses.end(); UI != UE; ++UI) {
20161 SDNode *Extract = *UI;
20163 SDValue Idx = Extract->getOperand(1);
20164 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
20165 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
20168 // The replacement was made in place; don't return anything.
20172 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
20173 static std::pair<unsigned, bool>
20174 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
20175 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
20176 if (!VT.isVector())
20177 return std::make_pair(0, false);
20179 bool NeedSplit = false;
20180 switch (VT.getSimpleVT().SimpleTy) {
20181 default: return std::make_pair(0, false);
20184 if (!Subtarget->hasVLX())
20185 return std::make_pair(0, false);
20189 if (!Subtarget->hasBWI())
20190 return std::make_pair(0, false);
20194 if (!Subtarget->hasAVX512())
20195 return std::make_pair(0, false);
20200 if (!Subtarget->hasAVX2())
20202 if (!Subtarget->hasAVX())
20203 return std::make_pair(0, false);
20208 if (!Subtarget->hasSSE2())
20209 return std::make_pair(0, false);
20212 // SSE2 has only a small subset of the operations.
20213 bool hasUnsigned = Subtarget->hasSSE41() ||
20214 (Subtarget->hasSSE2() && VT == MVT::v16i8);
20215 bool hasSigned = Subtarget->hasSSE41() ||
20216 (Subtarget->hasSSE2() && VT == MVT::v8i16);
20218 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20221 // Check for x CC y ? x : y.
20222 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20223 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20228 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20231 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20234 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20237 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20239 // Check for x CC y ? y : x -- a min/max with reversed arms.
20240 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20241 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20246 Opc = hasUnsigned ? X86ISD::UMAX : 0; break;
20249 Opc = hasUnsigned ? X86ISD::UMIN : 0; break;
20252 Opc = hasSigned ? X86ISD::SMAX : 0; break;
20255 Opc = hasSigned ? X86ISD::SMIN : 0; break;
20259 return std::make_pair(Opc, NeedSplit);
20263 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
20264 const X86Subtarget *Subtarget) {
20266 SDValue Cond = N->getOperand(0);
20267 SDValue LHS = N->getOperand(1);
20268 SDValue RHS = N->getOperand(2);
20270 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
20271 SDValue CondSrc = Cond->getOperand(0);
20272 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
20273 Cond = CondSrc->getOperand(0);
20276 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
20279 // A vselect where all conditions and data are constants can be optimized into
20280 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
20281 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
20282 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
20285 unsigned MaskValue = 0;
20286 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
20289 MVT VT = N->getSimpleValueType(0);
20290 unsigned NumElems = VT.getVectorNumElements();
20291 SmallVector<int, 8> ShuffleMask(NumElems, -1);
20292 for (unsigned i = 0; i < NumElems; ++i) {
20293 // Be sure we emit undef where we can.
20294 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
20295 ShuffleMask[i] = -1;
20297 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
20300 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20301 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
20303 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
20306 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
20308 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
20309 TargetLowering::DAGCombinerInfo &DCI,
20310 const X86Subtarget *Subtarget) {
20312 SDValue Cond = N->getOperand(0);
20313 // Get the LHS/RHS of the select.
20314 SDValue LHS = N->getOperand(1);
20315 SDValue RHS = N->getOperand(2);
20316 EVT VT = LHS.getValueType();
20317 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
20319 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
20320 // instructions match the semantics of the common C idiom x<y?x:y but not
20321 // x<=y?x:y, because of how they handle negative zero (which can be
20322 // ignored in unsafe-math mode).
20323 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
20324 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
20325 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
20326 (Subtarget->hasSSE2() ||
20327 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
20328 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20330 unsigned Opcode = 0;
20331 // Check for x CC y ? x : y.
20332 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20333 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20337 // Converting this to a min would handle NaNs incorrectly, and swapping
20338 // the operands would cause it to handle comparisons between positive
20339 // and negative zero incorrectly.
20340 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20341 if (!DAG.getTarget().Options.UnsafeFPMath &&
20342 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20344 std::swap(LHS, RHS);
20346 Opcode = X86ISD::FMIN;
20349 // Converting this to a min would handle comparisons between positive
20350 // and negative zero incorrectly.
20351 if (!DAG.getTarget().Options.UnsafeFPMath &&
20352 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20354 Opcode = X86ISD::FMIN;
20357 // Converting this to a min would handle both negative zeros and NaNs
20358 // incorrectly, but we can swap the operands to fix both.
20359 std::swap(LHS, RHS);
20363 Opcode = X86ISD::FMIN;
20367 // Converting this to a max would handle comparisons between positive
20368 // and negative zero incorrectly.
20369 if (!DAG.getTarget().Options.UnsafeFPMath &&
20370 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
20372 Opcode = X86ISD::FMAX;
20375 // Converting this to a max would handle NaNs incorrectly, and swapping
20376 // the operands would cause it to handle comparisons between positive
20377 // and negative zero incorrectly.
20378 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
20379 if (!DAG.getTarget().Options.UnsafeFPMath &&
20380 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
20382 std::swap(LHS, RHS);
20384 Opcode = X86ISD::FMAX;
20387 // Converting this to a max would handle both negative zeros and NaNs
20388 // incorrectly, but we can swap the operands to fix both.
20389 std::swap(LHS, RHS);
20393 Opcode = X86ISD::FMAX;
20396 // Check for x CC y ? y : x -- a min/max with reversed arms.
20397 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
20398 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
20402 // Converting this to a min would handle comparisons between positive
20403 // and negative zero incorrectly, and swapping the operands would
20404 // cause it to handle NaNs incorrectly.
20405 if (!DAG.getTarget().Options.UnsafeFPMath &&
20406 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
20407 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20409 std::swap(LHS, RHS);
20411 Opcode = X86ISD::FMIN;
20414 // Converting this to a min would handle NaNs incorrectly.
20415 if (!DAG.getTarget().Options.UnsafeFPMath &&
20416 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
20418 Opcode = X86ISD::FMIN;
20421 // Converting this to a min would handle both negative zeros and NaNs
20422 // incorrectly, but we can swap the operands to fix both.
20423 std::swap(LHS, RHS);
20427 Opcode = X86ISD::FMIN;
20431 // Converting this to a max would handle NaNs incorrectly.
20432 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20434 Opcode = X86ISD::FMAX;
20437 // Converting this to a max would handle comparisons between positive
20438 // and negative zero incorrectly, and swapping the operands would
20439 // cause it to handle NaNs incorrectly.
20440 if (!DAG.getTarget().Options.UnsafeFPMath &&
20441 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
20442 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
20444 std::swap(LHS, RHS);
20446 Opcode = X86ISD::FMAX;
20449 // Converting this to a max would handle both negative zeros and NaNs
20450 // incorrectly, but we can swap the operands to fix both.
20451 std::swap(LHS, RHS);
20455 Opcode = X86ISD::FMAX;
20461 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
20464 EVT CondVT = Cond.getValueType();
20465 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
20466 CondVT.getVectorElementType() == MVT::i1) {
20467 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
20468 // lowering on KNL. In this case we convert it to
20469 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
20470 // The same situation for all 128 and 256-bit vectors of i8 and i16.
20471 // Since SKX these selects have a proper lowering.
20472 EVT OpVT = LHS.getValueType();
20473 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
20474 (OpVT.getVectorElementType() == MVT::i8 ||
20475 OpVT.getVectorElementType() == MVT::i16) &&
20476 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
20477 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
20478 DCI.AddToWorklist(Cond.getNode());
20479 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
20482 // If this is a select between two integer constants, try to do some
20484 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
20485 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
20486 // Don't do this for crazy integer types.
20487 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
20488 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
20489 // so that TrueC (the true value) is larger than FalseC.
20490 bool NeedsCondInvert = false;
20492 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
20493 // Efficiently invertible.
20494 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
20495 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
20496 isa<ConstantSDNode>(Cond.getOperand(1))))) {
20497 NeedsCondInvert = true;
20498 std::swap(TrueC, FalseC);
20501 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
20502 if (FalseC->getAPIntValue() == 0 &&
20503 TrueC->getAPIntValue().isPowerOf2()) {
20504 if (NeedsCondInvert) // Invert the condition if needed.
20505 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20506 DAG.getConstant(1, Cond.getValueType()));
20508 // Zero extend the condition if needed.
20509 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
20511 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
20512 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
20513 DAG.getConstant(ShAmt, MVT::i8));
20516 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
20517 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
20518 if (NeedsCondInvert) // Invert the condition if needed.
20519 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20520 DAG.getConstant(1, Cond.getValueType()));
20522 // Zero extend the condition if needed.
20523 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
20524 FalseC->getValueType(0), Cond);
20525 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20526 SDValue(FalseC, 0));
20529 // Optimize cases that will turn into an LEA instruction. This requires
20530 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
20531 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
20532 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
20533 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
20535 bool isFastMultiplier = false;
20537 switch ((unsigned char)Diff) {
20539 case 1: // result = add base, cond
20540 case 2: // result = lea base( , cond*2)
20541 case 3: // result = lea base(cond, cond*2)
20542 case 4: // result = lea base( , cond*4)
20543 case 5: // result = lea base(cond, cond*4)
20544 case 8: // result = lea base( , cond*8)
20545 case 9: // result = lea base(cond, cond*8)
20546 isFastMultiplier = true;
20551 if (isFastMultiplier) {
20552 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
20553 if (NeedsCondInvert) // Invert the condition if needed.
20554 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
20555 DAG.getConstant(1, Cond.getValueType()));
20557 // Zero extend the condition if needed.
20558 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
20560 // Scale the condition by the difference.
20562 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
20563 DAG.getConstant(Diff, Cond.getValueType()));
20565 // Add the base if non-zero.
20566 if (FalseC->getAPIntValue() != 0)
20567 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
20568 SDValue(FalseC, 0));
20575 // Canonicalize max and min:
20576 // (x > y) ? x : y -> (x >= y) ? x : y
20577 // (x < y) ? x : y -> (x <= y) ? x : y
20578 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
20579 // the need for an extra compare
20580 // against zero. e.g.
20581 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
20583 // testl %edi, %edi
20585 // cmovgl %edi, %eax
20589 // cmovsl %eax, %edi
20590 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
20591 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
20592 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
20593 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20598 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
20599 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
20600 Cond.getOperand(0), Cond.getOperand(1), NewCC);
20601 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
20606 // Early exit check
20607 if (!TLI.isTypeLegal(VT))
20610 // Match VSELECTs into subs with unsigned saturation.
20611 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
20612 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
20613 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
20614 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
20615 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
20617 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
20618 // left side invert the predicate to simplify logic below.
20620 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
20622 CC = ISD::getSetCCInverse(CC, true);
20623 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
20627 if (Other.getNode() && Other->getNumOperands() == 2 &&
20628 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
20629 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
20630 SDValue CondRHS = Cond->getOperand(1);
20632 // Look for a general sub with unsigned saturation first.
20633 // x >= y ? x-y : 0 --> subus x, y
20634 // x > y ? x-y : 0 --> subus x, y
20635 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
20636 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
20637 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
20639 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
20640 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
20641 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
20642 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
20643 // If the RHS is a constant we have to reverse the const
20644 // canonicalization.
20645 // x > C-1 ? x+-C : 0 --> subus x, C
20646 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
20647 CondRHSConst->getAPIntValue() ==
20648 (-OpRHSConst->getAPIntValue() - 1))
20649 return DAG.getNode(
20650 X86ISD::SUBUS, DL, VT, OpLHS,
20651 DAG.getConstant(-OpRHSConst->getAPIntValue(), VT));
20653 // Another special case: If C was a sign bit, the sub has been
20654 // canonicalized into a xor.
20655 // FIXME: Would it be better to use computeKnownBits to determine
20656 // whether it's safe to decanonicalize the xor?
20657 // x s< 0 ? x^C : 0 --> subus x, C
20658 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
20659 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
20660 OpRHSConst->getAPIntValue().isSignBit())
20661 // Note that we have to rebuild the RHS constant here to ensure we
20662 // don't rely on particular values of undef lanes.
20663 return DAG.getNode(
20664 X86ISD::SUBUS, DL, VT, OpLHS,
20665 DAG.getConstant(OpRHSConst->getAPIntValue(), VT));
20670 // Try to match a min/max vector operation.
20671 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
20672 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
20673 unsigned Opc = ret.first;
20674 bool NeedSplit = ret.second;
20676 if (Opc && NeedSplit) {
20677 unsigned NumElems = VT.getVectorNumElements();
20678 // Extract the LHS vectors
20679 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
20680 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
20682 // Extract the RHS vectors
20683 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
20684 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
20686 // Create min/max for each subvector
20687 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
20688 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
20690 // Merge the result
20691 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
20693 return DAG.getNode(Opc, DL, VT, LHS, RHS);
20696 // Simplify vector selection if condition value type matches vselect
20698 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
20699 assert(Cond.getValueType().isVector() &&
20700 "vector select expects a vector selector!");
20702 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
20703 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
20705 // Try invert the condition if true value is not all 1s and false value
20707 if (!TValIsAllOnes && !FValIsAllZeros &&
20708 // Check if the selector will be produced by CMPP*/PCMP*
20709 Cond.getOpcode() == ISD::SETCC &&
20710 // Check if SETCC has already been promoted
20711 TLI.getSetCCResultType(*DAG.getContext(), VT) == CondVT) {
20712 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
20713 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
20715 if (TValIsAllZeros || FValIsAllOnes) {
20716 SDValue CC = Cond.getOperand(2);
20717 ISD::CondCode NewCC =
20718 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
20719 Cond.getOperand(0).getValueType().isInteger());
20720 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
20721 std::swap(LHS, RHS);
20722 TValIsAllOnes = FValIsAllOnes;
20723 FValIsAllZeros = TValIsAllZeros;
20727 if (TValIsAllOnes || FValIsAllZeros) {
20730 if (TValIsAllOnes && FValIsAllZeros)
20732 else if (TValIsAllOnes)
20733 Ret = DAG.getNode(ISD::OR, DL, CondVT, Cond,
20734 DAG.getNode(ISD::BITCAST, DL, CondVT, RHS));
20735 else if (FValIsAllZeros)
20736 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
20737 DAG.getNode(ISD::BITCAST, DL, CondVT, LHS));
20739 return DAG.getNode(ISD::BITCAST, DL, VT, Ret);
20743 // We should generate an X86ISD::BLENDI from a vselect if its argument
20744 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
20745 // constants. This specific pattern gets generated when we split a
20746 // selector for a 512 bit vector in a machine without AVX512 (but with
20747 // 256-bit vectors), during legalization:
20749 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
20751 // Iff we find this pattern and the build_vectors are built from
20752 // constants, we translate the vselect into a shuffle_vector that we
20753 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
20754 if ((N->getOpcode() == ISD::VSELECT ||
20755 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
20756 !DCI.isBeforeLegalize()) {
20757 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
20758 if (Shuffle.getNode())
20762 // If this is a *dynamic* select (non-constant condition) and we can match
20763 // this node with one of the variable blend instructions, restructure the
20764 // condition so that the blends can use the high bit of each element and use
20765 // SimplifyDemandedBits to simplify the condition operand.
20766 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
20767 !DCI.isBeforeLegalize() &&
20768 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
20769 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
20771 // Don't optimize vector selects that map to mask-registers.
20775 // We can only handle the cases where VSELECT is directly legal on the
20776 // subtarget. We custom lower VSELECT nodes with constant conditions and
20777 // this makes it hard to see whether a dynamic VSELECT will correctly
20778 // lower, so we both check the operation's status and explicitly handle the
20779 // cases where a *dynamic* blend will fail even though a constant-condition
20780 // blend could be custom lowered.
20781 // FIXME: We should find a better way to handle this class of problems.
20782 // Potentially, we should combine constant-condition vselect nodes
20783 // pre-legalization into shuffles and not mark as many types as custom
20785 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
20787 if (!Subtarget->hasSSE41() || VT == MVT::v16i16 || VT == MVT::v8i16)
20790 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
20791 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
20793 APInt KnownZero, KnownOne;
20794 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
20795 DCI.isBeforeLegalizeOps());
20796 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
20797 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
20799 // If we changed the computation somewhere in the DAG, this change
20800 // will affect all users of Cond.
20801 // Make sure it is fine and update all the nodes so that we do not
20802 // use the generic VSELECT anymore. Otherwise, we may perform
20803 // wrong optimizations as we messed up with the actual expectation
20804 // for the vector boolean values.
20805 if (Cond != TLO.Old) {
20806 // Check all uses of that condition operand to check whether it will be
20807 // consumed by non-BLEND instructions, which may depend on all bits are
20809 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20811 if (I->getOpcode() != ISD::VSELECT)
20812 // TODO: Add other opcodes eventually lowered into BLEND.
20815 // Update all the users of the condition, before committing the change,
20816 // so that the VSELECT optimizations that expect the correct vector
20817 // boolean value will not be triggered.
20818 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
20820 DAG.ReplaceAllUsesOfValueWith(
20822 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
20823 Cond, I->getOperand(1), I->getOperand(2)));
20824 DCI.CommitTargetLoweringOpt(TLO);
20827 // At this point, only Cond is changed. Change the condition
20828 // just for N to keep the opportunity to optimize all other
20829 // users their own way.
20830 DAG.ReplaceAllUsesOfValueWith(
20832 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
20833 TLO.New, N->getOperand(1), N->getOperand(2)));
20841 // Check whether a boolean test is testing a boolean value generated by
20842 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
20845 // Simplify the following patterns:
20846 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
20847 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
20848 // to (Op EFLAGS Cond)
20850 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
20851 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
20852 // to (Op EFLAGS !Cond)
20854 // where Op could be BRCOND or CMOV.
20856 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
20857 // Quit if not CMP and SUB with its value result used.
20858 if (Cmp.getOpcode() != X86ISD::CMP &&
20859 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
20862 // Quit if not used as a boolean value.
20863 if (CC != X86::COND_E && CC != X86::COND_NE)
20866 // Check CMP operands. One of them should be 0 or 1 and the other should be
20867 // an SetCC or extended from it.
20868 SDValue Op1 = Cmp.getOperand(0);
20869 SDValue Op2 = Cmp.getOperand(1);
20872 const ConstantSDNode* C = nullptr;
20873 bool needOppositeCond = (CC == X86::COND_E);
20874 bool checkAgainstTrue = false; // Is it a comparison against 1?
20876 if ((C = dyn_cast<ConstantSDNode>(Op1)))
20878 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
20880 else // Quit if all operands are not constants.
20883 if (C->getZExtValue() == 1) {
20884 needOppositeCond = !needOppositeCond;
20885 checkAgainstTrue = true;
20886 } else if (C->getZExtValue() != 0)
20887 // Quit if the constant is neither 0 or 1.
20890 bool truncatedToBoolWithAnd = false;
20891 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
20892 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
20893 SetCC.getOpcode() == ISD::TRUNCATE ||
20894 SetCC.getOpcode() == ISD::AND) {
20895 if (SetCC.getOpcode() == ISD::AND) {
20897 ConstantSDNode *CS;
20898 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
20899 CS->getZExtValue() == 1)
20901 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
20902 CS->getZExtValue() == 1)
20906 SetCC = SetCC.getOperand(OpIdx);
20907 truncatedToBoolWithAnd = true;
20909 SetCC = SetCC.getOperand(0);
20912 switch (SetCC.getOpcode()) {
20913 case X86ISD::SETCC_CARRY:
20914 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
20915 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
20916 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
20917 // truncated to i1 using 'and'.
20918 if (checkAgainstTrue && !truncatedToBoolWithAnd)
20920 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
20921 "Invalid use of SETCC_CARRY!");
20923 case X86ISD::SETCC:
20924 // Set the condition code or opposite one if necessary.
20925 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
20926 if (needOppositeCond)
20927 CC = X86::GetOppositeBranchCondition(CC);
20928 return SetCC.getOperand(1);
20929 case X86ISD::CMOV: {
20930 // Check whether false/true value has canonical one, i.e. 0 or 1.
20931 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
20932 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
20933 // Quit if true value is not a constant.
20936 // Quit if false value is not a constant.
20938 SDValue Op = SetCC.getOperand(0);
20939 // Skip 'zext' or 'trunc' node.
20940 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
20941 Op.getOpcode() == ISD::TRUNCATE)
20942 Op = Op.getOperand(0);
20943 // A special case for rdrand/rdseed, where 0 is set if false cond is
20945 if ((Op.getOpcode() != X86ISD::RDRAND &&
20946 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
20949 // Quit if false value is not the constant 0 or 1.
20950 bool FValIsFalse = true;
20951 if (FVal && FVal->getZExtValue() != 0) {
20952 if (FVal->getZExtValue() != 1)
20954 // If FVal is 1, opposite cond is needed.
20955 needOppositeCond = !needOppositeCond;
20956 FValIsFalse = false;
20958 // Quit if TVal is not the constant opposite of FVal.
20959 if (FValIsFalse && TVal->getZExtValue() != 1)
20961 if (!FValIsFalse && TVal->getZExtValue() != 0)
20963 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
20964 if (needOppositeCond)
20965 CC = X86::GetOppositeBranchCondition(CC);
20966 return SetCC.getOperand(3);
20973 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
20974 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
20975 TargetLowering::DAGCombinerInfo &DCI,
20976 const X86Subtarget *Subtarget) {
20979 // If the flag operand isn't dead, don't touch this CMOV.
20980 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
20983 SDValue FalseOp = N->getOperand(0);
20984 SDValue TrueOp = N->getOperand(1);
20985 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
20986 SDValue Cond = N->getOperand(3);
20988 if (CC == X86::COND_E || CC == X86::COND_NE) {
20989 switch (Cond.getOpcode()) {
20993 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
20994 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
20995 return (CC == X86::COND_E) ? FalseOp : TrueOp;
21001 Flags = checkBoolTestSetCCCombine(Cond, CC);
21002 if (Flags.getNode() &&
21003 // Extra check as FCMOV only supports a subset of X86 cond.
21004 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
21005 SDValue Ops[] = { FalseOp, TrueOp,
21006 DAG.getConstant(CC, MVT::i8), Flags };
21007 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
21010 // If this is a select between two integer constants, try to do some
21011 // optimizations. Note that the operands are ordered the opposite of SELECT
21013 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
21014 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
21015 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
21016 // larger than FalseC (the false value).
21017 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
21018 CC = X86::GetOppositeBranchCondition(CC);
21019 std::swap(TrueC, FalseC);
21020 std::swap(TrueOp, FalseOp);
21023 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
21024 // This is efficient for any integer data type (including i8/i16) and
21026 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
21027 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21028 DAG.getConstant(CC, MVT::i8), Cond);
21030 // Zero extend the condition if needed.
21031 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
21033 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
21034 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
21035 DAG.getConstant(ShAmt, MVT::i8));
21036 if (N->getNumValues() == 2) // Dead flag value?
21037 return DCI.CombineTo(N, Cond, SDValue());
21041 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
21042 // for any integer data type, including i8/i16.
21043 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
21044 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21045 DAG.getConstant(CC, MVT::i8), Cond);
21047 // Zero extend the condition if needed.
21048 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
21049 FalseC->getValueType(0), Cond);
21050 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21051 SDValue(FalseC, 0));
21053 if (N->getNumValues() == 2) // Dead flag value?
21054 return DCI.CombineTo(N, Cond, SDValue());
21058 // Optimize cases that will turn into an LEA instruction. This requires
21059 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
21060 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
21061 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
21062 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
21064 bool isFastMultiplier = false;
21066 switch ((unsigned char)Diff) {
21068 case 1: // result = add base, cond
21069 case 2: // result = lea base( , cond*2)
21070 case 3: // result = lea base(cond, cond*2)
21071 case 4: // result = lea base( , cond*4)
21072 case 5: // result = lea base(cond, cond*4)
21073 case 8: // result = lea base( , cond*8)
21074 case 9: // result = lea base(cond, cond*8)
21075 isFastMultiplier = true;
21080 if (isFastMultiplier) {
21081 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
21082 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
21083 DAG.getConstant(CC, MVT::i8), Cond);
21084 // Zero extend the condition if needed.
21085 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
21087 // Scale the condition by the difference.
21089 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
21090 DAG.getConstant(Diff, Cond.getValueType()));
21092 // Add the base if non-zero.
21093 if (FalseC->getAPIntValue() != 0)
21094 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
21095 SDValue(FalseC, 0));
21096 if (N->getNumValues() == 2) // Dead flag value?
21097 return DCI.CombineTo(N, Cond, SDValue());
21104 // Handle these cases:
21105 // (select (x != c), e, c) -> select (x != c), e, x),
21106 // (select (x == c), c, e) -> select (x == c), x, e)
21107 // where the c is an integer constant, and the "select" is the combination
21108 // of CMOV and CMP.
21110 // The rationale for this change is that the conditional-move from a constant
21111 // needs two instructions, however, conditional-move from a register needs
21112 // only one instruction.
21114 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
21115 // some instruction-combining opportunities. This opt needs to be
21116 // postponed as late as possible.
21118 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
21119 // the DCI.xxxx conditions are provided to postpone the optimization as
21120 // late as possible.
21122 ConstantSDNode *CmpAgainst = nullptr;
21123 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
21124 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
21125 !isa<ConstantSDNode>(Cond.getOperand(0))) {
21127 if (CC == X86::COND_NE &&
21128 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
21129 CC = X86::GetOppositeBranchCondition(CC);
21130 std::swap(TrueOp, FalseOp);
21133 if (CC == X86::COND_E &&
21134 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
21135 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
21136 DAG.getConstant(CC, MVT::i8), Cond };
21137 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
21145 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
21146 const X86Subtarget *Subtarget) {
21147 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
21149 default: return SDValue();
21150 // SSE/AVX/AVX2 blend intrinsics.
21151 case Intrinsic::x86_avx2_pblendvb:
21152 case Intrinsic::x86_avx2_pblendw:
21153 case Intrinsic::x86_avx2_pblendd_128:
21154 case Intrinsic::x86_avx2_pblendd_256:
21155 // Don't try to simplify this intrinsic if we don't have AVX2.
21156 if (!Subtarget->hasAVX2())
21159 case Intrinsic::x86_avx_blend_pd_256:
21160 case Intrinsic::x86_avx_blend_ps_256:
21161 case Intrinsic::x86_avx_blendv_pd_256:
21162 case Intrinsic::x86_avx_blendv_ps_256:
21163 // Don't try to simplify this intrinsic if we don't have AVX.
21164 if (!Subtarget->hasAVX())
21167 case Intrinsic::x86_sse41_pblendw:
21168 case Intrinsic::x86_sse41_blendpd:
21169 case Intrinsic::x86_sse41_blendps:
21170 case Intrinsic::x86_sse41_blendvps:
21171 case Intrinsic::x86_sse41_blendvpd:
21172 case Intrinsic::x86_sse41_pblendvb: {
21173 SDValue Op0 = N->getOperand(1);
21174 SDValue Op1 = N->getOperand(2);
21175 SDValue Mask = N->getOperand(3);
21177 // Don't try to simplify this intrinsic if we don't have SSE4.1.
21178 if (!Subtarget->hasSSE41())
21181 // fold (blend A, A, Mask) -> A
21184 // fold (blend A, B, allZeros) -> A
21185 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
21187 // fold (blend A, B, allOnes) -> B
21188 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
21191 // Simplify the case where the mask is a constant i32 value.
21192 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
21193 if (C->isNullValue())
21195 if (C->isAllOnesValue())
21202 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
21203 case Intrinsic::x86_sse2_psrai_w:
21204 case Intrinsic::x86_sse2_psrai_d:
21205 case Intrinsic::x86_avx2_psrai_w:
21206 case Intrinsic::x86_avx2_psrai_d:
21207 case Intrinsic::x86_sse2_psra_w:
21208 case Intrinsic::x86_sse2_psra_d:
21209 case Intrinsic::x86_avx2_psra_w:
21210 case Intrinsic::x86_avx2_psra_d: {
21211 SDValue Op0 = N->getOperand(1);
21212 SDValue Op1 = N->getOperand(2);
21213 EVT VT = Op0.getValueType();
21214 assert(VT.isVector() && "Expected a vector type!");
21216 if (isa<BuildVectorSDNode>(Op1))
21217 Op1 = Op1.getOperand(0);
21219 if (!isa<ConstantSDNode>(Op1))
21222 EVT SVT = VT.getVectorElementType();
21223 unsigned SVTBits = SVT.getSizeInBits();
21225 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
21226 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
21227 uint64_t ShAmt = C.getZExtValue();
21229 // Don't try to convert this shift into a ISD::SRA if the shift
21230 // count is bigger than or equal to the element size.
21231 if (ShAmt >= SVTBits)
21234 // Trivial case: if the shift count is zero, then fold this
21235 // into the first operand.
21239 // Replace this packed shift intrinsic with a target independent
21241 SDValue Splat = DAG.getConstant(C, VT);
21242 return DAG.getNode(ISD::SRA, SDLoc(N), VT, Op0, Splat);
21247 /// PerformMulCombine - Optimize a single multiply with constant into two
21248 /// in order to implement it with two cheaper instructions, e.g.
21249 /// LEA + SHL, LEA + LEA.
21250 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
21251 TargetLowering::DAGCombinerInfo &DCI) {
21252 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
21255 EVT VT = N->getValueType(0);
21256 if (VT != MVT::i64 && VT != MVT::i32)
21259 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
21262 uint64_t MulAmt = C->getZExtValue();
21263 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
21266 uint64_t MulAmt1 = 0;
21267 uint64_t MulAmt2 = 0;
21268 if ((MulAmt % 9) == 0) {
21270 MulAmt2 = MulAmt / 9;
21271 } else if ((MulAmt % 5) == 0) {
21273 MulAmt2 = MulAmt / 5;
21274 } else if ((MulAmt % 3) == 0) {
21276 MulAmt2 = MulAmt / 3;
21279 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
21282 if (isPowerOf2_64(MulAmt2) &&
21283 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
21284 // If second multiplifer is pow2, issue it first. We want the multiply by
21285 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
21287 std::swap(MulAmt1, MulAmt2);
21290 if (isPowerOf2_64(MulAmt1))
21291 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
21292 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
21294 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
21295 DAG.getConstant(MulAmt1, VT));
21297 if (isPowerOf2_64(MulAmt2))
21298 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
21299 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
21301 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
21302 DAG.getConstant(MulAmt2, VT));
21304 // Do not add new nodes to DAG combiner worklist.
21305 DCI.CombineTo(N, NewMul, false);
21310 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
21311 SDValue N0 = N->getOperand(0);
21312 SDValue N1 = N->getOperand(1);
21313 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
21314 EVT VT = N0.getValueType();
21316 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
21317 // since the result of setcc_c is all zero's or all ones.
21318 if (VT.isInteger() && !VT.isVector() &&
21319 N1C && N0.getOpcode() == ISD::AND &&
21320 N0.getOperand(1).getOpcode() == ISD::Constant) {
21321 SDValue N00 = N0.getOperand(0);
21322 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
21323 ((N00.getOpcode() == ISD::ANY_EXTEND ||
21324 N00.getOpcode() == ISD::ZERO_EXTEND) &&
21325 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
21326 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
21327 APInt ShAmt = N1C->getAPIntValue();
21328 Mask = Mask.shl(ShAmt);
21330 return DAG.getNode(ISD::AND, SDLoc(N), VT,
21331 N00, DAG.getConstant(Mask, VT));
21335 // Hardware support for vector shifts is sparse which makes us scalarize the
21336 // vector operations in many cases. Also, on sandybridge ADD is faster than
21338 // (shl V, 1) -> add V,V
21339 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
21340 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
21341 assert(N0.getValueType().isVector() && "Invalid vector shift type");
21342 // We shift all of the values by one. In many cases we do not have
21343 // hardware support for this operation. This is better expressed as an ADD
21345 if (N1SplatC->getZExtValue() == 1)
21346 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
21352 /// \brief Returns a vector of 0s if the node in input is a vector logical
21353 /// shift by a constant amount which is known to be bigger than or equal
21354 /// to the vector element size in bits.
21355 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
21356 const X86Subtarget *Subtarget) {
21357 EVT VT = N->getValueType(0);
21359 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
21360 (!Subtarget->hasInt256() ||
21361 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
21364 SDValue Amt = N->getOperand(1);
21366 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
21367 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
21368 APInt ShiftAmt = AmtSplat->getAPIntValue();
21369 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
21371 // SSE2/AVX2 logical shifts always return a vector of 0s
21372 // if the shift amount is bigger than or equal to
21373 // the element size. The constant shift amount will be
21374 // encoded as a 8-bit immediate.
21375 if (ShiftAmt.trunc(8).uge(MaxAmount))
21376 return getZeroVector(VT, Subtarget, DAG, DL);
21382 /// PerformShiftCombine - Combine shifts.
21383 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
21384 TargetLowering::DAGCombinerInfo &DCI,
21385 const X86Subtarget *Subtarget) {
21386 if (N->getOpcode() == ISD::SHL) {
21387 SDValue V = PerformSHLCombine(N, DAG);
21388 if (V.getNode()) return V;
21391 if (N->getOpcode() != ISD::SRA) {
21392 // Try to fold this logical shift into a zero vector.
21393 SDValue V = performShiftToAllZeros(N, DAG, Subtarget);
21394 if (V.getNode()) return V;
21400 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
21401 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
21402 // and friends. Likewise for OR -> CMPNEQSS.
21403 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
21404 TargetLowering::DAGCombinerInfo &DCI,
21405 const X86Subtarget *Subtarget) {
21408 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
21409 // we're requiring SSE2 for both.
21410 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
21411 SDValue N0 = N->getOperand(0);
21412 SDValue N1 = N->getOperand(1);
21413 SDValue CMP0 = N0->getOperand(1);
21414 SDValue CMP1 = N1->getOperand(1);
21417 // The SETCCs should both refer to the same CMP.
21418 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
21421 SDValue CMP00 = CMP0->getOperand(0);
21422 SDValue CMP01 = CMP0->getOperand(1);
21423 EVT VT = CMP00.getValueType();
21425 if (VT == MVT::f32 || VT == MVT::f64) {
21426 bool ExpectingFlags = false;
21427 // Check for any users that want flags:
21428 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
21429 !ExpectingFlags && UI != UE; ++UI)
21430 switch (UI->getOpcode()) {
21435 ExpectingFlags = true;
21437 case ISD::CopyToReg:
21438 case ISD::SIGN_EXTEND:
21439 case ISD::ZERO_EXTEND:
21440 case ISD::ANY_EXTEND:
21444 if (!ExpectingFlags) {
21445 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
21446 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
21448 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
21449 X86::CondCode tmp = cc0;
21454 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
21455 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
21456 // FIXME: need symbolic constants for these magic numbers.
21457 // See X86ATTInstPrinter.cpp:printSSECC().
21458 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
21459 if (Subtarget->hasAVX512()) {
21460 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
21461 CMP01, DAG.getConstant(x86cc, MVT::i8));
21462 if (N->getValueType(0) != MVT::i1)
21463 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
21467 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
21468 CMP00.getValueType(), CMP00, CMP01,
21469 DAG.getConstant(x86cc, MVT::i8));
21471 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
21472 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
21474 if (is64BitFP && !Subtarget->is64Bit()) {
21475 // On a 32-bit target, we cannot bitcast the 64-bit float to a
21476 // 64-bit integer, since that's not a legal type. Since
21477 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
21478 // bits, but can do this little dance to extract the lowest 32 bits
21479 // and work with those going forward.
21480 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
21482 SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
21484 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
21485 Vector32, DAG.getIntPtrConstant(0));
21489 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
21490 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
21491 DAG.getConstant(1, IntVT));
21492 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
21493 return OneBitOfTruth;
21501 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
21502 /// so it can be folded inside ANDNP.
21503 static bool CanFoldXORWithAllOnes(const SDNode *N) {
21504 EVT VT = N->getValueType(0);
21506 // Match direct AllOnes for 128 and 256-bit vectors
21507 if (ISD::isBuildVectorAllOnes(N))
21510 // Look through a bit convert.
21511 if (N->getOpcode() == ISD::BITCAST)
21512 N = N->getOperand(0).getNode();
21514 // Sometimes the operand may come from a insert_subvector building a 256-bit
21516 if (VT.is256BitVector() &&
21517 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
21518 SDValue V1 = N->getOperand(0);
21519 SDValue V2 = N->getOperand(1);
21521 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
21522 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
21523 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
21524 ISD::isBuildVectorAllOnes(V2.getNode()))
21531 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
21532 // register. In most cases we actually compare or select YMM-sized registers
21533 // and mixing the two types creates horrible code. This method optimizes
21534 // some of the transition sequences.
21535 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
21536 TargetLowering::DAGCombinerInfo &DCI,
21537 const X86Subtarget *Subtarget) {
21538 EVT VT = N->getValueType(0);
21539 if (!VT.is256BitVector())
21542 assert((N->getOpcode() == ISD::ANY_EXTEND ||
21543 N->getOpcode() == ISD::ZERO_EXTEND ||
21544 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
21546 SDValue Narrow = N->getOperand(0);
21547 EVT NarrowVT = Narrow->getValueType(0);
21548 if (!NarrowVT.is128BitVector())
21551 if (Narrow->getOpcode() != ISD::XOR &&
21552 Narrow->getOpcode() != ISD::AND &&
21553 Narrow->getOpcode() != ISD::OR)
21556 SDValue N0 = Narrow->getOperand(0);
21557 SDValue N1 = Narrow->getOperand(1);
21560 // The Left side has to be a trunc.
21561 if (N0.getOpcode() != ISD::TRUNCATE)
21564 // The type of the truncated inputs.
21565 EVT WideVT = N0->getOperand(0)->getValueType(0);
21569 // The right side has to be a 'trunc' or a constant vector.
21570 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
21571 ConstantSDNode *RHSConstSplat = nullptr;
21572 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
21573 RHSConstSplat = RHSBV->getConstantSplatNode();
21574 if (!RHSTrunc && !RHSConstSplat)
21577 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21579 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
21582 // Set N0 and N1 to hold the inputs to the new wide operation.
21583 N0 = N0->getOperand(0);
21584 if (RHSConstSplat) {
21585 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
21586 SDValue(RHSConstSplat, 0));
21587 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
21588 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
21589 } else if (RHSTrunc) {
21590 N1 = N1->getOperand(0);
21593 // Generate the wide operation.
21594 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
21595 unsigned Opcode = N->getOpcode();
21597 case ISD::ANY_EXTEND:
21599 case ISD::ZERO_EXTEND: {
21600 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
21601 APInt Mask = APInt::getAllOnesValue(InBits);
21602 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
21603 return DAG.getNode(ISD::AND, DL, VT,
21604 Op, DAG.getConstant(Mask, VT));
21606 case ISD::SIGN_EXTEND:
21607 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
21608 Op, DAG.getValueType(NarrowVT));
21610 llvm_unreachable("Unexpected opcode");
21614 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
21615 TargetLowering::DAGCombinerInfo &DCI,
21616 const X86Subtarget *Subtarget) {
21617 SDValue N0 = N->getOperand(0);
21618 SDValue N1 = N->getOperand(1);
21621 // A vector zext_in_reg may be represented as a shuffle,
21622 // feeding into a bitcast (this represents anyext) feeding into
21623 // an and with a mask.
21624 // We'd like to try to combine that into a shuffle with zero
21625 // plus a bitcast, removing the and.
21626 if (N0.getOpcode() != ISD::BITCAST ||
21627 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
21630 // The other side of the AND should be a splat of 2^C, where C
21631 // is the number of bits in the source type.
21632 if (N1.getOpcode() == ISD::BITCAST)
21633 N1 = N1.getOperand(0);
21634 if (N1.getOpcode() != ISD::BUILD_VECTOR)
21636 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
21638 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
21639 EVT SrcType = Shuffle->getValueType(0);
21641 // We expect a single-source shuffle
21642 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
21645 unsigned SrcSize = SrcType.getScalarSizeInBits();
21647 APInt SplatValue, SplatUndef;
21648 unsigned SplatBitSize;
21650 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
21651 SplatBitSize, HasAnyUndefs))
21654 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
21655 // Make sure the splat matches the mask we expect
21656 if (SplatBitSize > ResSize ||
21657 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
21660 // Make sure the input and output size make sense
21661 if (SrcSize >= ResSize || ResSize % SrcSize)
21664 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
21665 // The number of u's between each two values depends on the ratio between
21666 // the source and dest type.
21667 unsigned ZextRatio = ResSize / SrcSize;
21668 bool IsZext = true;
21669 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
21670 if (i % ZextRatio) {
21671 if (Shuffle->getMaskElt(i) > 0) {
21677 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
21678 // Expected element number
21688 // Ok, perform the transformation - replace the shuffle with
21689 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
21690 // (instead of undef) where the k elements come from the zero vector.
21691 SmallVector<int, 8> Mask;
21692 unsigned NumElems = SrcType.getVectorNumElements();
21693 for (unsigned i = 0; i < NumElems; ++i)
21695 Mask.push_back(NumElems);
21697 Mask.push_back(i / ZextRatio);
21699 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
21700 Shuffle->getOperand(0), DAG.getConstant(0, SrcType), Mask);
21701 return DAG.getNode(ISD::BITCAST, DL, N0.getValueType(), NewShuffle);
21704 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
21705 TargetLowering::DAGCombinerInfo &DCI,
21706 const X86Subtarget *Subtarget) {
21707 if (DCI.isBeforeLegalizeOps())
21710 SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget);
21711 if (Zext.getNode())
21714 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21718 EVT VT = N->getValueType(0);
21719 SDValue N0 = N->getOperand(0);
21720 SDValue N1 = N->getOperand(1);
21723 // Create BEXTR instructions
21724 // BEXTR is ((X >> imm) & (2**size-1))
21725 if (VT == MVT::i32 || VT == MVT::i64) {
21726 // Check for BEXTR.
21727 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
21728 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
21729 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
21730 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
21731 if (MaskNode && ShiftNode) {
21732 uint64_t Mask = MaskNode->getZExtValue();
21733 uint64_t Shift = ShiftNode->getZExtValue();
21734 if (isMask_64(Mask)) {
21735 uint64_t MaskSize = countPopulation(Mask);
21736 if (Shift + MaskSize <= VT.getSizeInBits())
21737 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
21738 DAG.getConstant(Shift | (MaskSize << 8), VT));
21746 // Want to form ANDNP nodes:
21747 // 1) In the hopes of then easily combining them with OR and AND nodes
21748 // to form PBLEND/PSIGN.
21749 // 2) To match ANDN packed intrinsics
21750 if (VT != MVT::v2i64 && VT != MVT::v4i64)
21753 // Check LHS for vnot
21754 if (N0.getOpcode() == ISD::XOR &&
21755 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
21756 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
21757 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
21759 // Check RHS for vnot
21760 if (N1.getOpcode() == ISD::XOR &&
21761 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
21762 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
21763 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
21768 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
21769 TargetLowering::DAGCombinerInfo &DCI,
21770 const X86Subtarget *Subtarget) {
21771 if (DCI.isBeforeLegalizeOps())
21774 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
21778 SDValue N0 = N->getOperand(0);
21779 SDValue N1 = N->getOperand(1);
21780 EVT VT = N->getValueType(0);
21782 // look for psign/blend
21783 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
21784 if (!Subtarget->hasSSSE3() ||
21785 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
21788 // Canonicalize pandn to RHS
21789 if (N0.getOpcode() == X86ISD::ANDNP)
21791 // or (and (m, y), (pandn m, x))
21792 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
21793 SDValue Mask = N1.getOperand(0);
21794 SDValue X = N1.getOperand(1);
21796 if (N0.getOperand(0) == Mask)
21797 Y = N0.getOperand(1);
21798 if (N0.getOperand(1) == Mask)
21799 Y = N0.getOperand(0);
21801 // Check to see if the mask appeared in both the AND and ANDNP and
21805 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
21806 // Look through mask bitcast.
21807 if (Mask.getOpcode() == ISD::BITCAST)
21808 Mask = Mask.getOperand(0);
21809 if (X.getOpcode() == ISD::BITCAST)
21810 X = X.getOperand(0);
21811 if (Y.getOpcode() == ISD::BITCAST)
21812 Y = Y.getOperand(0);
21814 EVT MaskVT = Mask.getValueType();
21816 // Validate that the Mask operand is a vector sra node.
21817 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
21818 // there is no psrai.b
21819 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
21820 unsigned SraAmt = ~0;
21821 if (Mask.getOpcode() == ISD::SRA) {
21822 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
21823 if (auto *AmtConst = AmtBV->getConstantSplatNode())
21824 SraAmt = AmtConst->getZExtValue();
21825 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
21826 SDValue SraC = Mask.getOperand(1);
21827 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
21829 if ((SraAmt + 1) != EltBits)
21834 // Now we know we at least have a plendvb with the mask val. See if
21835 // we can form a psignb/w/d.
21836 // psign = x.type == y.type == mask.type && y = sub(0, x);
21837 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
21838 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
21839 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
21840 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
21841 "Unsupported VT for PSIGN");
21842 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
21843 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21845 // PBLENDVB only available on SSE 4.1
21846 if (!Subtarget->hasSSE41())
21849 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
21851 X = DAG.getNode(ISD::BITCAST, DL, BlendVT, X);
21852 Y = DAG.getNode(ISD::BITCAST, DL, BlendVT, Y);
21853 Mask = DAG.getNode(ISD::BITCAST, DL, BlendVT, Mask);
21854 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
21855 return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
21859 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
21862 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
21863 MachineFunction &MF = DAG.getMachineFunction();
21865 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
21867 // SHLD/SHRD instructions have lower register pressure, but on some
21868 // platforms they have higher latency than the equivalent
21869 // series of shifts/or that would otherwise be generated.
21870 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
21871 // have higher latencies and we are not optimizing for size.
21872 if (!OptForSize && Subtarget->isSHLDSlow())
21875 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
21877 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
21879 if (!N0.hasOneUse() || !N1.hasOneUse())
21882 SDValue ShAmt0 = N0.getOperand(1);
21883 if (ShAmt0.getValueType() != MVT::i8)
21885 SDValue ShAmt1 = N1.getOperand(1);
21886 if (ShAmt1.getValueType() != MVT::i8)
21888 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
21889 ShAmt0 = ShAmt0.getOperand(0);
21890 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
21891 ShAmt1 = ShAmt1.getOperand(0);
21894 unsigned Opc = X86ISD::SHLD;
21895 SDValue Op0 = N0.getOperand(0);
21896 SDValue Op1 = N1.getOperand(0);
21897 if (ShAmt0.getOpcode() == ISD::SUB) {
21898 Opc = X86ISD::SHRD;
21899 std::swap(Op0, Op1);
21900 std::swap(ShAmt0, ShAmt1);
21903 unsigned Bits = VT.getSizeInBits();
21904 if (ShAmt1.getOpcode() == ISD::SUB) {
21905 SDValue Sum = ShAmt1.getOperand(0);
21906 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
21907 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
21908 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
21909 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
21910 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
21911 return DAG.getNode(Opc, DL, VT,
21913 DAG.getNode(ISD::TRUNCATE, DL,
21916 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
21917 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
21919 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
21920 return DAG.getNode(Opc, DL, VT,
21921 N0.getOperand(0), N1.getOperand(0),
21922 DAG.getNode(ISD::TRUNCATE, DL,
21929 // Generate NEG and CMOV for integer abs.
21930 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
21931 EVT VT = N->getValueType(0);
21933 // Since X86 does not have CMOV for 8-bit integer, we don't convert
21934 // 8-bit integer abs to NEG and CMOV.
21935 if (VT.isInteger() && VT.getSizeInBits() == 8)
21938 SDValue N0 = N->getOperand(0);
21939 SDValue N1 = N->getOperand(1);
21942 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
21943 // and change it to SUB and CMOV.
21944 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
21945 N0.getOpcode() == ISD::ADD &&
21946 N0.getOperand(1) == N1 &&
21947 N1.getOpcode() == ISD::SRA &&
21948 N1.getOperand(0) == N0.getOperand(0))
21949 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
21950 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
21951 // Generate SUB & CMOV.
21952 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
21953 DAG.getConstant(0, VT), N0.getOperand(0));
21955 SDValue Ops[] = { N0.getOperand(0), Neg,
21956 DAG.getConstant(X86::COND_GE, MVT::i8),
21957 SDValue(Neg.getNode(), 1) };
21958 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
21963 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
21964 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
21965 TargetLowering::DAGCombinerInfo &DCI,
21966 const X86Subtarget *Subtarget) {
21967 if (DCI.isBeforeLegalizeOps())
21970 if (Subtarget->hasCMov()) {
21971 SDValue RV = performIntegerAbsCombine(N, DAG);
21979 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
21980 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
21981 TargetLowering::DAGCombinerInfo &DCI,
21982 const X86Subtarget *Subtarget) {
21983 LoadSDNode *Ld = cast<LoadSDNode>(N);
21984 EVT RegVT = Ld->getValueType(0);
21985 EVT MemVT = Ld->getMemoryVT();
21987 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21989 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
21990 // into two 16-byte operations.
21991 ISD::LoadExtType Ext = Ld->getExtensionType();
21992 unsigned Alignment = Ld->getAlignment();
21993 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
21994 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
21995 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
21996 unsigned NumElems = RegVT.getVectorNumElements();
22000 SDValue Ptr = Ld->getBasePtr();
22001 SDValue Increment = DAG.getConstant(16, TLI.getPointerTy());
22003 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
22005 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22006 Ld->getPointerInfo(), Ld->isVolatile(),
22007 Ld->isNonTemporal(), Ld->isInvariant(),
22009 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22010 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
22011 Ld->getPointerInfo(), Ld->isVolatile(),
22012 Ld->isNonTemporal(), Ld->isInvariant(),
22013 std::min(16U, Alignment));
22014 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22016 Load2.getValue(1));
22018 SDValue NewVec = DAG.getUNDEF(RegVT);
22019 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
22020 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
22021 return DCI.CombineTo(N, NewVec, TF, true);
22027 /// PerformMLOADCombine - Resolve extending loads
22028 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
22029 TargetLowering::DAGCombinerInfo &DCI,
22030 const X86Subtarget *Subtarget) {
22031 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
22032 if (Mld->getExtensionType() != ISD::SEXTLOAD)
22035 EVT VT = Mld->getValueType(0);
22036 unsigned NumElems = VT.getVectorNumElements();
22037 EVT LdVT = Mld->getMemoryVT();
22040 assert(LdVT != VT && "Cannot extend to the same type");
22041 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
22042 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
22043 // From, To sizes and ElemCount must be pow of two
22044 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22045 "Unexpected size for extending masked load");
22047 unsigned SizeRatio = ToSz / FromSz;
22048 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
22050 // Create a type on which we perform the shuffle
22051 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22052 LdVT.getScalarType(), NumElems*SizeRatio);
22053 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22055 // Convert Src0 value
22056 SDValue WideSrc0 = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mld->getSrc0());
22057 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
22058 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22059 for (unsigned i = 0; i != NumElems; ++i)
22060 ShuffleVec[i] = i * SizeRatio;
22062 // Can't shuffle using an illegal type.
22063 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22064 && "WideVecVT should be legal");
22065 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
22066 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
22068 // Prepare the new mask
22070 SDValue Mask = Mld->getMask();
22071 if (Mask.getValueType() == VT) {
22072 // Mask and original value have the same type
22073 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22074 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22075 for (unsigned i = 0; i != NumElems; ++i)
22076 ShuffleVec[i] = i * SizeRatio;
22077 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22078 ShuffleVec[i] = NumElems*SizeRatio;
22079 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22080 DAG.getConstant(0, WideVecVT),
22084 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22085 unsigned WidenNumElts = NumElems*SizeRatio;
22086 unsigned MaskNumElts = VT.getVectorNumElements();
22087 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22090 unsigned NumConcat = WidenNumElts / MaskNumElts;
22091 SmallVector<SDValue, 16> Ops(NumConcat);
22092 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22094 for (unsigned i = 1; i != NumConcat; ++i)
22097 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22100 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
22101 Mld->getBasePtr(), NewMask, WideSrc0,
22102 Mld->getMemoryVT(), Mld->getMemOperand(),
22104 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
22105 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
22108 /// PerformMSTORECombine - Resolve truncating stores
22109 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
22110 const X86Subtarget *Subtarget) {
22111 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
22112 if (!Mst->isTruncatingStore())
22115 EVT VT = Mst->getValue().getValueType();
22116 unsigned NumElems = VT.getVectorNumElements();
22117 EVT StVT = Mst->getMemoryVT();
22120 assert(StVT != VT && "Cannot truncate to the same type");
22121 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22122 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22124 // From, To sizes and ElemCount must be pow of two
22125 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
22126 "Unexpected size for truncating masked store");
22127 // We are going to use the original vector elt for storing.
22128 // Accumulated smaller vector elements must be a multiple of the store size.
22129 assert (((NumElems * FromSz) % ToSz) == 0 &&
22130 "Unexpected ratio for truncating masked store");
22132 unsigned SizeRatio = FromSz / ToSz;
22133 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22135 // Create a type on which we perform the shuffle
22136 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22137 StVT.getScalarType(), NumElems*SizeRatio);
22139 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22141 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mst->getValue());
22142 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
22143 for (unsigned i = 0; i != NumElems; ++i)
22144 ShuffleVec[i] = i * SizeRatio;
22146 // Can't shuffle using an illegal type.
22147 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
22148 && "WideVecVT should be legal");
22150 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22151 DAG.getUNDEF(WideVecVT),
22155 SDValue Mask = Mst->getMask();
22156 if (Mask.getValueType() == VT) {
22157 // Mask and original value have the same type
22158 NewMask = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Mask);
22159 for (unsigned i = 0; i != NumElems; ++i)
22160 ShuffleVec[i] = i * SizeRatio;
22161 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
22162 ShuffleVec[i] = NumElems*SizeRatio;
22163 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
22164 DAG.getConstant(0, WideVecVT),
22168 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
22169 unsigned WidenNumElts = NumElems*SizeRatio;
22170 unsigned MaskNumElts = VT.getVectorNumElements();
22171 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
22174 unsigned NumConcat = WidenNumElts / MaskNumElts;
22175 SmallVector<SDValue, 16> Ops(NumConcat);
22176 SDValue ZeroVal = DAG.getConstant(0, Mask.getValueType());
22178 for (unsigned i = 1; i != NumConcat; ++i)
22181 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
22184 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
22185 NewMask, StVT, Mst->getMemOperand(), false);
22187 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
22188 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
22189 const X86Subtarget *Subtarget) {
22190 StoreSDNode *St = cast<StoreSDNode>(N);
22191 EVT VT = St->getValue().getValueType();
22192 EVT StVT = St->getMemoryVT();
22194 SDValue StoredVal = St->getOperand(1);
22195 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22197 // If we are saving a concatenation of two XMM registers and 32-byte stores
22198 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
22199 unsigned Alignment = St->getAlignment();
22200 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
22201 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
22202 StVT == VT && !IsAligned) {
22203 unsigned NumElems = VT.getVectorNumElements();
22207 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
22208 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
22210 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
22211 SDValue Ptr0 = St->getBasePtr();
22212 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
22214 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
22215 St->getPointerInfo(), St->isVolatile(),
22216 St->isNonTemporal(), Alignment);
22217 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
22218 St->getPointerInfo(), St->isVolatile(),
22219 St->isNonTemporal(),
22220 std::min(16U, Alignment));
22221 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
22224 // Optimize trunc store (of multiple scalars) to shuffle and store.
22225 // First, pack all of the elements in one place. Next, store to memory
22226 // in fewer chunks.
22227 if (St->isTruncatingStore() && VT.isVector()) {
22228 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22229 unsigned NumElems = VT.getVectorNumElements();
22230 assert(StVT != VT && "Cannot truncate to the same type");
22231 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
22232 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
22234 // From, To sizes and ElemCount must be pow of two
22235 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
22236 // We are going to use the original vector elt for storing.
22237 // Accumulated smaller vector elements must be a multiple of the store size.
22238 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
22240 unsigned SizeRatio = FromSz / ToSz;
22242 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
22244 // Create a type on which we perform the shuffle
22245 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
22246 StVT.getScalarType(), NumElems*SizeRatio);
22248 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
22250 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
22251 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
22252 for (unsigned i = 0; i != NumElems; ++i)
22253 ShuffleVec[i] = i * SizeRatio;
22255 // Can't shuffle using an illegal type.
22256 if (!TLI.isTypeLegal(WideVecVT))
22259 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
22260 DAG.getUNDEF(WideVecVT),
22262 // At this point all of the data is stored at the bottom of the
22263 // register. We now need to save it to mem.
22265 // Find the largest store unit
22266 MVT StoreType = MVT::i8;
22267 for (MVT Tp : MVT::integer_valuetypes()) {
22268 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
22272 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
22273 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
22274 (64 <= NumElems * ToSz))
22275 StoreType = MVT::f64;
22277 // Bitcast the original vector into a vector of store-size units
22278 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
22279 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
22280 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
22281 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
22282 SmallVector<SDValue, 8> Chains;
22283 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
22284 TLI.getPointerTy());
22285 SDValue Ptr = St->getBasePtr();
22287 // Perform one or more big stores into memory.
22288 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
22289 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
22290 StoreType, ShuffWide,
22291 DAG.getIntPtrConstant(i));
22292 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
22293 St->getPointerInfo(), St->isVolatile(),
22294 St->isNonTemporal(), St->getAlignment());
22295 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
22296 Chains.push_back(Ch);
22299 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
22302 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
22303 // the FP state in cases where an emms may be missing.
22304 // A preferable solution to the general problem is to figure out the right
22305 // places to insert EMMS. This qualifies as a quick hack.
22307 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
22308 if (VT.getSizeInBits() != 64)
22311 const Function *F = DAG.getMachineFunction().getFunction();
22312 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
22313 bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
22314 && Subtarget->hasSSE2();
22315 if ((VT.isVector() ||
22316 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
22317 isa<LoadSDNode>(St->getValue()) &&
22318 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
22319 St->getChain().hasOneUse() && !St->isVolatile()) {
22320 SDNode* LdVal = St->getValue().getNode();
22321 LoadSDNode *Ld = nullptr;
22322 int TokenFactorIndex = -1;
22323 SmallVector<SDValue, 8> Ops;
22324 SDNode* ChainVal = St->getChain().getNode();
22325 // Must be a store of a load. We currently handle two cases: the load
22326 // is a direct child, and it's under an intervening TokenFactor. It is
22327 // possible to dig deeper under nested TokenFactors.
22328 if (ChainVal == LdVal)
22329 Ld = cast<LoadSDNode>(St->getChain());
22330 else if (St->getValue().hasOneUse() &&
22331 ChainVal->getOpcode() == ISD::TokenFactor) {
22332 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
22333 if (ChainVal->getOperand(i).getNode() == LdVal) {
22334 TokenFactorIndex = i;
22335 Ld = cast<LoadSDNode>(St->getValue());
22337 Ops.push_back(ChainVal->getOperand(i));
22341 if (!Ld || !ISD::isNormalLoad(Ld))
22344 // If this is not the MMX case, i.e. we are just turning i64 load/store
22345 // into f64 load/store, avoid the transformation if there are multiple
22346 // uses of the loaded value.
22347 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
22352 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
22353 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
22355 if (Subtarget->is64Bit() || F64IsLegal) {
22356 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
22357 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
22358 Ld->getPointerInfo(), Ld->isVolatile(),
22359 Ld->isNonTemporal(), Ld->isInvariant(),
22360 Ld->getAlignment());
22361 SDValue NewChain = NewLd.getValue(1);
22362 if (TokenFactorIndex != -1) {
22363 Ops.push_back(NewChain);
22364 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22366 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
22367 St->getPointerInfo(),
22368 St->isVolatile(), St->isNonTemporal(),
22369 St->getAlignment());
22372 // Otherwise, lower to two pairs of 32-bit loads / stores.
22373 SDValue LoAddr = Ld->getBasePtr();
22374 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
22375 DAG.getConstant(4, MVT::i32));
22377 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
22378 Ld->getPointerInfo(),
22379 Ld->isVolatile(), Ld->isNonTemporal(),
22380 Ld->isInvariant(), Ld->getAlignment());
22381 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
22382 Ld->getPointerInfo().getWithOffset(4),
22383 Ld->isVolatile(), Ld->isNonTemporal(),
22385 MinAlign(Ld->getAlignment(), 4));
22387 SDValue NewChain = LoLd.getValue(1);
22388 if (TokenFactorIndex != -1) {
22389 Ops.push_back(LoLd);
22390 Ops.push_back(HiLd);
22391 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
22394 LoAddr = St->getBasePtr();
22395 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
22396 DAG.getConstant(4, MVT::i32));
22398 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
22399 St->getPointerInfo(),
22400 St->isVolatile(), St->isNonTemporal(),
22401 St->getAlignment());
22402 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
22403 St->getPointerInfo().getWithOffset(4),
22405 St->isNonTemporal(),
22406 MinAlign(St->getAlignment(), 4));
22407 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
22412 /// Return 'true' if this vector operation is "horizontal"
22413 /// and return the operands for the horizontal operation in LHS and RHS. A
22414 /// horizontal operation performs the binary operation on successive elements
22415 /// of its first operand, then on successive elements of its second operand,
22416 /// returning the resulting values in a vector. For example, if
22417 /// A = < float a0, float a1, float a2, float a3 >
22419 /// B = < float b0, float b1, float b2, float b3 >
22420 /// then the result of doing a horizontal operation on A and B is
22421 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
22422 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
22423 /// A horizontal-op B, for some already available A and B, and if so then LHS is
22424 /// set to A, RHS to B, and the routine returns 'true'.
22425 /// Note that the binary operation should have the property that if one of the
22426 /// operands is UNDEF then the result is UNDEF.
22427 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
22428 // Look for the following pattern: if
22429 // A = < float a0, float a1, float a2, float a3 >
22430 // B = < float b0, float b1, float b2, float b3 >
22432 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
22433 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
22434 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
22435 // which is A horizontal-op B.
22437 // At least one of the operands should be a vector shuffle.
22438 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
22439 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
22442 MVT VT = LHS.getSimpleValueType();
22444 assert((VT.is128BitVector() || VT.is256BitVector()) &&
22445 "Unsupported vector type for horizontal add/sub");
22447 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
22448 // operate independently on 128-bit lanes.
22449 unsigned NumElts = VT.getVectorNumElements();
22450 unsigned NumLanes = VT.getSizeInBits()/128;
22451 unsigned NumLaneElts = NumElts / NumLanes;
22452 assert((NumLaneElts % 2 == 0) &&
22453 "Vector type should have an even number of elements in each lane");
22454 unsigned HalfLaneElts = NumLaneElts/2;
22456 // View LHS in the form
22457 // LHS = VECTOR_SHUFFLE A, B, LMask
22458 // If LHS is not a shuffle then pretend it is the shuffle
22459 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
22460 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
22463 SmallVector<int, 16> LMask(NumElts);
22464 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22465 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
22466 A = LHS.getOperand(0);
22467 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
22468 B = LHS.getOperand(1);
22469 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
22470 std::copy(Mask.begin(), Mask.end(), LMask.begin());
22472 if (LHS.getOpcode() != ISD::UNDEF)
22474 for (unsigned i = 0; i != NumElts; ++i)
22478 // Likewise, view RHS in the form
22479 // RHS = VECTOR_SHUFFLE C, D, RMask
22481 SmallVector<int, 16> RMask(NumElts);
22482 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
22483 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
22484 C = RHS.getOperand(0);
22485 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
22486 D = RHS.getOperand(1);
22487 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
22488 std::copy(Mask.begin(), Mask.end(), RMask.begin());
22490 if (RHS.getOpcode() != ISD::UNDEF)
22492 for (unsigned i = 0; i != NumElts; ++i)
22496 // Check that the shuffles are both shuffling the same vectors.
22497 if (!(A == C && B == D) && !(A == D && B == C))
22500 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
22501 if (!A.getNode() && !B.getNode())
22504 // If A and B occur in reverse order in RHS, then "swap" them (which means
22505 // rewriting the mask).
22507 CommuteVectorShuffleMask(RMask, NumElts);
22509 // At this point LHS and RHS are equivalent to
22510 // LHS = VECTOR_SHUFFLE A, B, LMask
22511 // RHS = VECTOR_SHUFFLE A, B, RMask
22512 // Check that the masks correspond to performing a horizontal operation.
22513 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
22514 for (unsigned i = 0; i != NumLaneElts; ++i) {
22515 int LIdx = LMask[i+l], RIdx = RMask[i+l];
22517 // Ignore any UNDEF components.
22518 if (LIdx < 0 || RIdx < 0 ||
22519 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
22520 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
22523 // Check that successive elements are being operated on. If not, this is
22524 // not a horizontal operation.
22525 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
22526 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
22527 if (!(LIdx == Index && RIdx == Index + 1) &&
22528 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
22533 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
22534 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
22538 /// Do target-specific dag combines on floating point adds.
22539 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
22540 const X86Subtarget *Subtarget) {
22541 EVT VT = N->getValueType(0);
22542 SDValue LHS = N->getOperand(0);
22543 SDValue RHS = N->getOperand(1);
22545 // Try to synthesize horizontal adds from adds of shuffles.
22546 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22547 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22548 isHorizontalBinOp(LHS, RHS, true))
22549 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
22553 /// Do target-specific dag combines on floating point subs.
22554 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
22555 const X86Subtarget *Subtarget) {
22556 EVT VT = N->getValueType(0);
22557 SDValue LHS = N->getOperand(0);
22558 SDValue RHS = N->getOperand(1);
22560 // Try to synthesize horizontal subs from subs of shuffles.
22561 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
22562 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
22563 isHorizontalBinOp(LHS, RHS, false))
22564 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
22568 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
22569 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
22570 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
22572 // F[X]OR(0.0, x) -> x
22573 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22574 if (C->getValueAPF().isPosZero())
22575 return N->getOperand(1);
22577 // F[X]OR(x, 0.0) -> x
22578 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22579 if (C->getValueAPF().isPosZero())
22580 return N->getOperand(0);
22584 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
22585 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
22586 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
22588 // Only perform optimizations if UnsafeMath is used.
22589 if (!DAG.getTarget().Options.UnsafeFPMath)
22592 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
22593 // into FMINC and FMAXC, which are Commutative operations.
22594 unsigned NewOp = 0;
22595 switch (N->getOpcode()) {
22596 default: llvm_unreachable("unknown opcode");
22597 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
22598 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
22601 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
22602 N->getOperand(0), N->getOperand(1));
22605 /// Do target-specific dag combines on X86ISD::FAND nodes.
22606 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
22607 // FAND(0.0, x) -> 0.0
22608 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22609 if (C->getValueAPF().isPosZero())
22610 return N->getOperand(0);
22612 // FAND(x, 0.0) -> 0.0
22613 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22614 if (C->getValueAPF().isPosZero())
22615 return N->getOperand(1);
22620 /// Do target-specific dag combines on X86ISD::FANDN nodes
22621 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
22622 // FANDN(0.0, x) -> x
22623 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
22624 if (C->getValueAPF().isPosZero())
22625 return N->getOperand(1);
22627 // FANDN(x, 0.0) -> 0.0
22628 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
22629 if (C->getValueAPF().isPosZero())
22630 return N->getOperand(1);
22635 static SDValue PerformBTCombine(SDNode *N,
22637 TargetLowering::DAGCombinerInfo &DCI) {
22638 // BT ignores high bits in the bit index operand.
22639 SDValue Op1 = N->getOperand(1);
22640 if (Op1.hasOneUse()) {
22641 unsigned BitWidth = Op1.getValueSizeInBits();
22642 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
22643 APInt KnownZero, KnownOne;
22644 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
22645 !DCI.isBeforeLegalizeOps());
22646 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22647 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
22648 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
22649 DCI.CommitTargetLoweringOpt(TLO);
22654 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
22655 SDValue Op = N->getOperand(0);
22656 if (Op.getOpcode() == ISD::BITCAST)
22657 Op = Op.getOperand(0);
22658 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
22659 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
22660 VT.getVectorElementType().getSizeInBits() ==
22661 OpVT.getVectorElementType().getSizeInBits()) {
22662 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
22667 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
22668 const X86Subtarget *Subtarget) {
22669 EVT VT = N->getValueType(0);
22670 if (!VT.isVector())
22673 SDValue N0 = N->getOperand(0);
22674 SDValue N1 = N->getOperand(1);
22675 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
22678 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
22679 // both SSE and AVX2 since there is no sign-extended shift right
22680 // operation on a vector with 64-bit elements.
22681 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
22682 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
22683 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
22684 N0.getOpcode() == ISD::SIGN_EXTEND)) {
22685 SDValue N00 = N0.getOperand(0);
22687 // EXTLOAD has a better solution on AVX2,
22688 // it may be replaced with X86ISD::VSEXT node.
22689 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
22690 if (!ISD::isNormalLoad(N00.getNode()))
22693 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
22694 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
22696 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
22702 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
22703 TargetLowering::DAGCombinerInfo &DCI,
22704 const X86Subtarget *Subtarget) {
22705 SDValue N0 = N->getOperand(0);
22706 EVT VT = N->getValueType(0);
22708 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
22709 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
22710 // This exposes the sext to the sdivrem lowering, so that it directly extends
22711 // from AH (which we otherwise need to do contortions to access).
22712 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
22713 N0.getValueType() == MVT::i8 && VT == MVT::i32) {
22715 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22716 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, dl, NodeTys,
22717 N0.getOperand(0), N0.getOperand(1));
22718 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22719 return R.getValue(1);
22722 if (!DCI.isBeforeLegalizeOps())
22725 if (!Subtarget->hasFp256())
22728 if (VT.isVector() && VT.getSizeInBits() == 256) {
22729 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22737 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
22738 const X86Subtarget* Subtarget) {
22740 EVT VT = N->getValueType(0);
22742 // Let legalize expand this if it isn't a legal type yet.
22743 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
22746 EVT ScalarVT = VT.getScalarType();
22747 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
22748 (!Subtarget->hasFMA() && !Subtarget->hasFMA4()))
22751 SDValue A = N->getOperand(0);
22752 SDValue B = N->getOperand(1);
22753 SDValue C = N->getOperand(2);
22755 bool NegA = (A.getOpcode() == ISD::FNEG);
22756 bool NegB = (B.getOpcode() == ISD::FNEG);
22757 bool NegC = (C.getOpcode() == ISD::FNEG);
22759 // Negative multiplication when NegA xor NegB
22760 bool NegMul = (NegA != NegB);
22762 A = A.getOperand(0);
22764 B = B.getOperand(0);
22766 C = C.getOperand(0);
22770 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
22772 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
22774 return DAG.getNode(Opcode, dl, VT, A, B, C);
22777 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
22778 TargetLowering::DAGCombinerInfo &DCI,
22779 const X86Subtarget *Subtarget) {
22780 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
22781 // (and (i32 x86isd::setcc_carry), 1)
22782 // This eliminates the zext. This transformation is necessary because
22783 // ISD::SETCC is always legalized to i8.
22785 SDValue N0 = N->getOperand(0);
22786 EVT VT = N->getValueType(0);
22788 if (N0.getOpcode() == ISD::AND &&
22790 N0.getOperand(0).hasOneUse()) {
22791 SDValue N00 = N0.getOperand(0);
22792 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22793 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
22794 if (!C || C->getZExtValue() != 1)
22796 return DAG.getNode(ISD::AND, dl, VT,
22797 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22798 N00.getOperand(0), N00.getOperand(1)),
22799 DAG.getConstant(1, VT));
22803 if (N0.getOpcode() == ISD::TRUNCATE &&
22805 N0.getOperand(0).hasOneUse()) {
22806 SDValue N00 = N0.getOperand(0);
22807 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
22808 return DAG.getNode(ISD::AND, dl, VT,
22809 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
22810 N00.getOperand(0), N00.getOperand(1)),
22811 DAG.getConstant(1, VT));
22814 if (VT.is256BitVector()) {
22815 SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
22820 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
22821 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
22822 // This exposes the zext to the udivrem lowering, so that it directly extends
22823 // from AH (which we otherwise need to do contortions to access).
22824 if (N0.getOpcode() == ISD::UDIVREM &&
22825 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
22826 (VT == MVT::i32 || VT == MVT::i64)) {
22827 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
22828 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
22829 N0.getOperand(0), N0.getOperand(1));
22830 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
22831 return R.getValue(1);
22837 // Optimize x == -y --> x+y == 0
22838 // x != -y --> x+y != 0
22839 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
22840 const X86Subtarget* Subtarget) {
22841 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
22842 SDValue LHS = N->getOperand(0);
22843 SDValue RHS = N->getOperand(1);
22844 EVT VT = N->getValueType(0);
22847 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
22848 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
22849 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
22850 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22851 LHS.getValueType(), RHS, LHS.getOperand(1));
22852 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22853 addV, DAG.getConstant(0, addV.getValueType()), CC);
22855 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
22856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
22857 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
22858 SDValue addV = DAG.getNode(ISD::ADD, SDLoc(N),
22859 RHS.getValueType(), LHS, RHS.getOperand(1));
22860 return DAG.getSetCC(SDLoc(N), N->getValueType(0),
22861 addV, DAG.getConstant(0, addV.getValueType()), CC);
22864 if (VT.getScalarType() == MVT::i1) {
22865 bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
22866 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22867 bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
22868 if (!IsSEXT0 && !IsVZero0)
22870 bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
22871 (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
22872 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
22874 if (!IsSEXT1 && !IsVZero1)
22877 if (IsSEXT0 && IsVZero1) {
22878 assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
22879 if (CC == ISD::SETEQ)
22880 return DAG.getNOT(DL, LHS.getOperand(0), VT);
22881 return LHS.getOperand(0);
22883 if (IsSEXT1 && IsVZero0) {
22884 assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
22885 if (CC == ISD::SETEQ)
22886 return DAG.getNOT(DL, RHS.getOperand(0), VT);
22887 return RHS.getOperand(0);
22894 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
22895 SelectionDAG &DAG) {
22897 MVT VT = Load->getSimpleValueType(0);
22898 MVT EVT = VT.getVectorElementType();
22899 SDValue Addr = Load->getOperand(1);
22900 SDValue NewAddr = DAG.getNode(
22901 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
22902 DAG.getConstant(Index * EVT.getStoreSize(), Addr.getSimpleValueType()));
22905 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
22906 DAG.getMachineFunction().getMachineMemOperand(
22907 Load->getMemOperand(), 0, EVT.getStoreSize()));
22911 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
22912 const X86Subtarget *Subtarget) {
22914 MVT VT = N->getOperand(1)->getSimpleValueType(0);
22915 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
22916 "X86insertps is only defined for v4x32");
22918 SDValue Ld = N->getOperand(1);
22919 if (MayFoldLoad(Ld)) {
22920 // Extract the countS bits from the immediate so we can get the proper
22921 // address when narrowing the vector load to a specific element.
22922 // When the second source op is a memory address, insertps doesn't use
22923 // countS and just gets an f32 from that address.
22924 unsigned DestIndex =
22925 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
22927 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
22929 // Create this as a scalar to vector to match the instruction pattern.
22930 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
22931 // countS bits are ignored when loading from memory on insertps, which
22932 // means we don't need to explicitly set them to 0.
22933 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
22934 LoadScalarToVector, N->getOperand(2));
22939 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
22940 SDValue V0 = N->getOperand(0);
22941 SDValue V1 = N->getOperand(1);
22943 EVT VT = N->getValueType(0);
22945 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
22946 // operands and changing the mask to 1. This saves us a bunch of
22947 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
22948 // x86InstrInfo knows how to commute this back after instruction selection
22949 // if it would help register allocation.
22951 // TODO: If optimizing for size or a processor that doesn't suffer from
22952 // partial register update stalls, this should be transformed into a MOVSD
22953 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
22955 if (VT == MVT::v2f64)
22956 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
22957 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
22958 SDValue NewMask = DAG.getConstant(1, MVT::i8);
22959 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
22965 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
22966 // as "sbb reg,reg", since it can be extended without zext and produces
22967 // an all-ones bit which is more useful than 0/1 in some cases.
22968 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
22971 return DAG.getNode(ISD::AND, DL, VT,
22972 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
22973 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
22974 DAG.getConstant(1, VT));
22975 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
22976 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
22977 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
22978 DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS));
22981 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
22982 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
22983 TargetLowering::DAGCombinerInfo &DCI,
22984 const X86Subtarget *Subtarget) {
22986 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
22987 SDValue EFLAGS = N->getOperand(1);
22989 if (CC == X86::COND_A) {
22990 // Try to convert COND_A into COND_B in an attempt to facilitate
22991 // materializing "setb reg".
22993 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
22994 // cannot take an immediate as its first operand.
22996 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
22997 EFLAGS.getValueType().isInteger() &&
22998 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
22999 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
23000 EFLAGS.getNode()->getVTList(),
23001 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
23002 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
23003 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
23007 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
23008 // a zext and produces an all-ones bit which is more useful than 0/1 in some
23010 if (CC == X86::COND_B)
23011 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
23015 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23016 if (Flags.getNode()) {
23017 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23018 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
23024 // Optimize branch condition evaluation.
23026 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
23027 TargetLowering::DAGCombinerInfo &DCI,
23028 const X86Subtarget *Subtarget) {
23030 SDValue Chain = N->getOperand(0);
23031 SDValue Dest = N->getOperand(1);
23032 SDValue EFLAGS = N->getOperand(3);
23033 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
23037 Flags = checkBoolTestSetCCCombine(EFLAGS, CC);
23038 if (Flags.getNode()) {
23039 SDValue Cond = DAG.getConstant(CC, MVT::i8);
23040 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
23047 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
23048 SelectionDAG &DAG) {
23049 // Take advantage of vector comparisons producing 0 or -1 in each lane to
23050 // optimize away operation when it's from a constant.
23052 // The general transformation is:
23053 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
23054 // AND(VECTOR_CMP(x,y), constant2)
23055 // constant2 = UNARYOP(constant)
23057 // Early exit if this isn't a vector operation, the operand of the
23058 // unary operation isn't a bitwise AND, or if the sizes of the operations
23059 // aren't the same.
23060 EVT VT = N->getValueType(0);
23061 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
23062 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
23063 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
23066 // Now check that the other operand of the AND is a constant. We could
23067 // make the transformation for non-constant splats as well, but it's unclear
23068 // that would be a benefit as it would not eliminate any operations, just
23069 // perform one more step in scalar code before moving to the vector unit.
23070 if (BuildVectorSDNode *BV =
23071 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
23072 // Bail out if the vector isn't a constant.
23073 if (!BV->isConstant())
23076 // Everything checks out. Build up the new and improved node.
23078 EVT IntVT = BV->getValueType(0);
23079 // Create a new constant of the appropriate type for the transformed
23081 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
23082 // The AND node needs bitcasts to/from an integer vector type around it.
23083 SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
23084 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
23085 N->getOperand(0)->getOperand(0), MaskConst);
23086 SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
23093 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
23094 const X86Subtarget *Subtarget) {
23095 // First try to optimize away the conversion entirely when it's
23096 // conditionally from a constant. Vectors only.
23097 SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
23098 if (Res != SDValue())
23101 // Now move on to more general possibilities.
23102 SDValue Op0 = N->getOperand(0);
23103 EVT InVT = Op0->getValueType(0);
23105 // SINT_TO_FP(v4i8) -> SINT_TO_FP(SEXT(v4i8 to v4i32))
23106 if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
23108 MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
23109 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
23110 return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
23113 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
23114 // a 32-bit target where SSE doesn't support i64->FP operations.
23115 if (Op0.getOpcode() == ISD::LOAD) {
23116 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
23117 EVT VT = Ld->getValueType(0);
23118 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
23119 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
23120 !Subtarget->is64Bit() && VT == MVT::i64) {
23121 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
23122 SDValue(N, 0), Ld->getValueType(0), Ld->getChain(), Op0, DAG);
23123 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
23130 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
23131 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
23132 X86TargetLowering::DAGCombinerInfo &DCI) {
23133 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
23134 // the result is either zero or one (depending on the input carry bit).
23135 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
23136 if (X86::isZeroNode(N->getOperand(0)) &&
23137 X86::isZeroNode(N->getOperand(1)) &&
23138 // We don't have a good way to replace an EFLAGS use, so only do this when
23140 SDValue(N, 1).use_empty()) {
23142 EVT VT = N->getValueType(0);
23143 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
23144 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
23145 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
23146 DAG.getConstant(X86::COND_B,MVT::i8),
23148 DAG.getConstant(1, VT));
23149 return DCI.CombineTo(N, Res1, CarryOut);
23155 // fold (add Y, (sete X, 0)) -> adc 0, Y
23156 // (add Y, (setne X, 0)) -> sbb -1, Y
23157 // (sub (sete X, 0), Y) -> sbb 0, Y
23158 // (sub (setne X, 0), Y) -> adc -1, Y
23159 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
23162 // Look through ZExts.
23163 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
23164 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
23167 SDValue SetCC = Ext.getOperand(0);
23168 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
23171 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
23172 if (CC != X86::COND_E && CC != X86::COND_NE)
23175 SDValue Cmp = SetCC.getOperand(1);
23176 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
23177 !X86::isZeroNode(Cmp.getOperand(1)) ||
23178 !Cmp.getOperand(0).getValueType().isInteger())
23181 SDValue CmpOp0 = Cmp.getOperand(0);
23182 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
23183 DAG.getConstant(1, CmpOp0.getValueType()));
23185 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
23186 if (CC == X86::COND_NE)
23187 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
23188 DL, OtherVal.getValueType(), OtherVal,
23189 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
23190 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
23191 DL, OtherVal.getValueType(), OtherVal,
23192 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
23195 /// PerformADDCombine - Do target-specific dag combines on integer adds.
23196 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
23197 const X86Subtarget *Subtarget) {
23198 EVT VT = N->getValueType(0);
23199 SDValue Op0 = N->getOperand(0);
23200 SDValue Op1 = N->getOperand(1);
23202 // Try to synthesize horizontal adds from adds of shuffles.
23203 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23204 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23205 isHorizontalBinOp(Op0, Op1, true))
23206 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
23208 return OptimizeConditionalInDecrement(N, DAG);
23211 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
23212 const X86Subtarget *Subtarget) {
23213 SDValue Op0 = N->getOperand(0);
23214 SDValue Op1 = N->getOperand(1);
23216 // X86 can't encode an immediate LHS of a sub. See if we can push the
23217 // negation into a preceding instruction.
23218 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
23219 // If the RHS of the sub is a XOR with one use and a constant, invert the
23220 // immediate. Then add one to the LHS of the sub so we can turn
23221 // X-Y -> X+~Y+1, saving one register.
23222 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
23223 isa<ConstantSDNode>(Op1.getOperand(1))) {
23224 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
23225 EVT VT = Op0.getValueType();
23226 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
23228 DAG.getConstant(~XorC, VT));
23229 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
23230 DAG.getConstant(C->getAPIntValue()+1, VT));
23234 // Try to synthesize horizontal adds from adds of shuffles.
23235 EVT VT = N->getValueType(0);
23236 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
23237 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
23238 isHorizontalBinOp(Op0, Op1, true))
23239 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
23241 return OptimizeConditionalInDecrement(N, DAG);
23244 /// performVZEXTCombine - Performs build vector combines
23245 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
23246 TargetLowering::DAGCombinerInfo &DCI,
23247 const X86Subtarget *Subtarget) {
23249 MVT VT = N->getSimpleValueType(0);
23250 SDValue Op = N->getOperand(0);
23251 MVT OpVT = Op.getSimpleValueType();
23252 MVT OpEltVT = OpVT.getVectorElementType();
23253 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
23255 // (vzext (bitcast (vzext (x)) -> (vzext x)
23257 while (V.getOpcode() == ISD::BITCAST)
23258 V = V.getOperand(0);
23260 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
23261 MVT InnerVT = V.getSimpleValueType();
23262 MVT InnerEltVT = InnerVT.getVectorElementType();
23264 // If the element sizes match exactly, we can just do one larger vzext. This
23265 // is always an exact type match as vzext operates on integer types.
23266 if (OpEltVT == InnerEltVT) {
23267 assert(OpVT == InnerVT && "Types must match for vzext!");
23268 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
23271 // The only other way we can combine them is if only a single element of the
23272 // inner vzext is used in the input to the outer vzext.
23273 if (InnerEltVT.getSizeInBits() < InputBits)
23276 // In this case, the inner vzext is completely dead because we're going to
23277 // only look at bits inside of the low element. Just do the outer vzext on
23278 // a bitcast of the input to the inner.
23279 return DAG.getNode(X86ISD::VZEXT, DL, VT,
23280 DAG.getNode(ISD::BITCAST, DL, OpVT, V));
23283 // Check if we can bypass extracting and re-inserting an element of an input
23284 // vector. Essentialy:
23285 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
23286 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
23287 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
23288 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
23289 SDValue ExtractedV = V.getOperand(0);
23290 SDValue OrigV = ExtractedV.getOperand(0);
23291 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
23292 if (ExtractIdx->getZExtValue() == 0) {
23293 MVT OrigVT = OrigV.getSimpleValueType();
23294 // Extract a subvector if necessary...
23295 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
23296 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
23297 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
23298 OrigVT.getVectorNumElements() / Ratio);
23299 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
23300 DAG.getIntPtrConstant(0));
23302 Op = DAG.getNode(ISD::BITCAST, DL, OpVT, OrigV);
23303 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
23310 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
23311 DAGCombinerInfo &DCI) const {
23312 SelectionDAG &DAG = DCI.DAG;
23313 switch (N->getOpcode()) {
23315 case ISD::EXTRACT_VECTOR_ELT:
23316 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
23319 case X86ISD::SHRUNKBLEND:
23320 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
23321 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
23322 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
23323 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
23324 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
23325 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
23326 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
23329 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
23330 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
23331 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
23332 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
23333 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
23334 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
23335 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
23336 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
23337 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
23338 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
23339 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
23341 case X86ISD::FOR: return PerformFORCombine(N, DAG);
23343 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
23344 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
23345 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
23346 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
23347 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
23348 case ISD::ANY_EXTEND:
23349 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
23350 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
23351 case ISD::SIGN_EXTEND_INREG:
23352 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
23353 case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
23354 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
23355 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
23356 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
23357 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
23358 case X86ISD::SHUFP: // Handle all target specific shuffles
23359 case X86ISD::PALIGNR:
23360 case X86ISD::UNPCKH:
23361 case X86ISD::UNPCKL:
23362 case X86ISD::MOVHLPS:
23363 case X86ISD::MOVLHPS:
23364 case X86ISD::PSHUFB:
23365 case X86ISD::PSHUFD:
23366 case X86ISD::PSHUFHW:
23367 case X86ISD::PSHUFLW:
23368 case X86ISD::MOVSS:
23369 case X86ISD::MOVSD:
23370 case X86ISD::VPERMILPI:
23371 case X86ISD::VPERM2X128:
23372 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
23373 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
23374 case ISD::INTRINSIC_WO_CHAIN:
23375 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
23376 case X86ISD::INSERTPS: {
23377 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
23378 return PerformINSERTPSCombine(N, DAG, Subtarget);
23381 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
23382 case ISD::BUILD_VECTOR: return PerformBUILD_VECTORCombine(N, DAG, Subtarget);
23388 /// isTypeDesirableForOp - Return true if the target has native support for
23389 /// the specified value type and it is 'desirable' to use the type for the
23390 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
23391 /// instruction encodings are longer and some i16 instructions are slow.
23392 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
23393 if (!isTypeLegal(VT))
23395 if (VT != MVT::i16)
23402 case ISD::SIGN_EXTEND:
23403 case ISD::ZERO_EXTEND:
23404 case ISD::ANY_EXTEND:
23417 /// IsDesirableToPromoteOp - This method query the target whether it is
23418 /// beneficial for dag combiner to promote the specified node. If true, it
23419 /// should return the desired promotion type by reference.
23420 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
23421 EVT VT = Op.getValueType();
23422 if (VT != MVT::i16)
23425 bool Promote = false;
23426 bool Commute = false;
23427 switch (Op.getOpcode()) {
23430 LoadSDNode *LD = cast<LoadSDNode>(Op);
23431 // If the non-extending load has a single use and it's not live out, then it
23432 // might be folded.
23433 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
23434 Op.hasOneUse()*/) {
23435 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
23436 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
23437 // The only case where we'd want to promote LOAD (rather then it being
23438 // promoted as an operand is when it's only use is liveout.
23439 if (UI->getOpcode() != ISD::CopyToReg)
23446 case ISD::SIGN_EXTEND:
23447 case ISD::ZERO_EXTEND:
23448 case ISD::ANY_EXTEND:
23453 SDValue N0 = Op.getOperand(0);
23454 // Look out for (store (shl (load), x)).
23455 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
23468 SDValue N0 = Op.getOperand(0);
23469 SDValue N1 = Op.getOperand(1);
23470 if (!Commute && MayFoldLoad(N1))
23472 // Avoid disabling potential load folding opportunities.
23473 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
23475 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
23485 //===----------------------------------------------------------------------===//
23486 // X86 Inline Assembly Support
23487 //===----------------------------------------------------------------------===//
23490 // Helper to match a string separated by whitespace.
23491 bool matchAsmImpl(StringRef s, ArrayRef<const StringRef *> args) {
23492 s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
23494 for (unsigned i = 0, e = args.size(); i != e; ++i) {
23495 StringRef piece(*args[i]);
23496 if (!s.startswith(piece)) // Check if the piece matches.
23499 s = s.substr(piece.size());
23500 StringRef::size_type pos = s.find_first_not_of(" \t");
23501 if (pos == 0) // We matched a prefix.
23509 const VariadicFunction1<bool, StringRef, StringRef, matchAsmImpl> matchAsm={};
23512 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
23514 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
23515 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
23516 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
23517 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
23519 if (AsmPieces.size() == 3)
23521 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
23528 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
23529 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
23531 std::string AsmStr = IA->getAsmString();
23533 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
23534 if (!Ty || Ty->getBitWidth() % 16 != 0)
23537 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
23538 SmallVector<StringRef, 4> AsmPieces;
23539 SplitString(AsmStr, AsmPieces, ";\n");
23541 switch (AsmPieces.size()) {
23542 default: return false;
23544 // FIXME: this should verify that we are targeting a 486 or better. If not,
23545 // we will turn this bswap into something that will be lowered to logical
23546 // ops instead of emitting the bswap asm. For now, we don't support 486 or
23547 // lower so don't worry about this.
23549 if (matchAsm(AsmPieces[0], "bswap", "$0") ||
23550 matchAsm(AsmPieces[0], "bswapl", "$0") ||
23551 matchAsm(AsmPieces[0], "bswapq", "$0") ||
23552 matchAsm(AsmPieces[0], "bswap", "${0:q}") ||
23553 matchAsm(AsmPieces[0], "bswapl", "${0:q}") ||
23554 matchAsm(AsmPieces[0], "bswapq", "${0:q}")) {
23555 // No need to check constraints, nothing other than the equivalent of
23556 // "=r,0" would be valid here.
23557 return IntrinsicLowering::LowerToByteSwap(CI);
23560 // rorw $$8, ${0:w} --> llvm.bswap.i16
23561 if (CI->getType()->isIntegerTy(16) &&
23562 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23563 (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") ||
23564 matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}"))) {
23566 const std::string &ConstraintsStr = IA->getConstraintString();
23567 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23568 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23569 if (clobbersFlagRegisters(AsmPieces))
23570 return IntrinsicLowering::LowerToByteSwap(CI);
23574 if (CI->getType()->isIntegerTy(32) &&
23575 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
23576 matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}") &&
23577 matchAsm(AsmPieces[1], "rorl", "$$16,", "$0") &&
23578 matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}")) {
23580 const std::string &ConstraintsStr = IA->getConstraintString();
23581 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
23582 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
23583 if (clobbersFlagRegisters(AsmPieces))
23584 return IntrinsicLowering::LowerToByteSwap(CI);
23587 if (CI->getType()->isIntegerTy(64)) {
23588 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
23589 if (Constraints.size() >= 2 &&
23590 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
23591 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
23592 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
23593 if (matchAsm(AsmPieces[0], "bswap", "%eax") &&
23594 matchAsm(AsmPieces[1], "bswap", "%edx") &&
23595 matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx"))
23596 return IntrinsicLowering::LowerToByteSwap(CI);
23604 /// getConstraintType - Given a constraint letter, return the type of
23605 /// constraint it is for this target.
23606 X86TargetLowering::ConstraintType
23607 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
23608 if (Constraint.size() == 1) {
23609 switch (Constraint[0]) {
23620 return C_RegisterClass;
23644 return TargetLowering::getConstraintType(Constraint);
23647 /// Examine constraint type and operand type and determine a weight value.
23648 /// This object must already have been set up with the operand type
23649 /// and the current alternative constraint selected.
23650 TargetLowering::ConstraintWeight
23651 X86TargetLowering::getSingleConstraintMatchWeight(
23652 AsmOperandInfo &info, const char *constraint) const {
23653 ConstraintWeight weight = CW_Invalid;
23654 Value *CallOperandVal = info.CallOperandVal;
23655 // If we don't have a value, we can't do a match,
23656 // but allow it at the lowest weight.
23657 if (!CallOperandVal)
23659 Type *type = CallOperandVal->getType();
23660 // Look at the constraint type.
23661 switch (*constraint) {
23663 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
23674 if (CallOperandVal->getType()->isIntegerTy())
23675 weight = CW_SpecificReg;
23680 if (type->isFloatingPointTy())
23681 weight = CW_SpecificReg;
23684 if (type->isX86_MMXTy() && Subtarget->hasMMX())
23685 weight = CW_SpecificReg;
23689 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
23690 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
23691 weight = CW_Register;
23694 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
23695 if (C->getZExtValue() <= 31)
23696 weight = CW_Constant;
23700 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23701 if (C->getZExtValue() <= 63)
23702 weight = CW_Constant;
23706 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23707 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
23708 weight = CW_Constant;
23712 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23713 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
23714 weight = CW_Constant;
23718 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23719 if (C->getZExtValue() <= 3)
23720 weight = CW_Constant;
23724 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23725 if (C->getZExtValue() <= 0xff)
23726 weight = CW_Constant;
23731 if (dyn_cast<ConstantFP>(CallOperandVal)) {
23732 weight = CW_Constant;
23736 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23737 if ((C->getSExtValue() >= -0x80000000LL) &&
23738 (C->getSExtValue() <= 0x7fffffffLL))
23739 weight = CW_Constant;
23743 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
23744 if (C->getZExtValue() <= 0xffffffff)
23745 weight = CW_Constant;
23752 /// LowerXConstraint - try to replace an X constraint, which matches anything,
23753 /// with another that has more specific requirements based on the type of the
23754 /// corresponding operand.
23755 const char *X86TargetLowering::
23756 LowerXConstraint(EVT ConstraintVT) const {
23757 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
23758 // 'f' like normal targets.
23759 if (ConstraintVT.isFloatingPoint()) {
23760 if (Subtarget->hasSSE2())
23762 if (Subtarget->hasSSE1())
23766 return TargetLowering::LowerXConstraint(ConstraintVT);
23769 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
23770 /// vector. If it is invalid, don't add anything to Ops.
23771 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
23772 std::string &Constraint,
23773 std::vector<SDValue>&Ops,
23774 SelectionDAG &DAG) const {
23777 // Only support length 1 constraints for now.
23778 if (Constraint.length() > 1) return;
23780 char ConstraintLetter = Constraint[0];
23781 switch (ConstraintLetter) {
23784 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23785 if (C->getZExtValue() <= 31) {
23786 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23792 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23793 if (C->getZExtValue() <= 63) {
23794 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23800 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23801 if (isInt<8>(C->getSExtValue())) {
23802 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23809 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
23810 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
23811 Result = DAG.getTargetConstant(C->getSExtValue(), Op.getValueType());
23817 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23818 if (C->getZExtValue() <= 3) {
23819 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23825 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23826 if (C->getZExtValue() <= 255) {
23827 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23833 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23834 if (C->getZExtValue() <= 127) {
23835 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23841 // 32-bit signed value
23842 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23843 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23844 C->getSExtValue())) {
23845 // Widen to 64 bits here to get it sign extended.
23846 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
23849 // FIXME gcc accepts some relocatable values here too, but only in certain
23850 // memory models; it's complicated.
23855 // 32-bit unsigned value
23856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
23857 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
23858 C->getZExtValue())) {
23859 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
23863 // FIXME gcc accepts some relocatable values here too, but only in certain
23864 // memory models; it's complicated.
23868 // Literal immediates are always ok.
23869 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
23870 // Widen to 64 bits here to get it sign extended.
23871 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
23875 // In any sort of PIC mode addresses need to be computed at runtime by
23876 // adding in a register or some sort of table lookup. These can't
23877 // be used as immediates.
23878 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
23881 // If we are in non-pic codegen mode, we allow the address of a global (with
23882 // an optional displacement) to be used with 'i'.
23883 GlobalAddressSDNode *GA = nullptr;
23884 int64_t Offset = 0;
23886 // Match either (GA), (GA+C), (GA+C1+C2), etc.
23888 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
23889 Offset += GA->getOffset();
23891 } else if (Op.getOpcode() == ISD::ADD) {
23892 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23893 Offset += C->getZExtValue();
23894 Op = Op.getOperand(0);
23897 } else if (Op.getOpcode() == ISD::SUB) {
23898 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
23899 Offset += -C->getZExtValue();
23900 Op = Op.getOperand(0);
23905 // Otherwise, this isn't something we can handle, reject it.
23909 const GlobalValue *GV = GA->getGlobal();
23910 // If we require an extra load to get this address, as in PIC mode, we
23911 // can't accept it.
23912 if (isGlobalStubReference(
23913 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
23916 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
23917 GA->getValueType(0), Offset);
23922 if (Result.getNode()) {
23923 Ops.push_back(Result);
23926 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
23929 std::pair<unsigned, const TargetRegisterClass*>
23930 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
23932 // First, see if this is a constraint that directly corresponds to an LLVM
23934 if (Constraint.size() == 1) {
23935 // GCC Constraint Letters
23936 switch (Constraint[0]) {
23938 // TODO: Slight differences here in allocation order and leaving
23939 // RIP in the class. Do they matter any more here than they do
23940 // in the normal allocation?
23941 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
23942 if (Subtarget->is64Bit()) {
23943 if (VT == MVT::i32 || VT == MVT::f32)
23944 return std::make_pair(0U, &X86::GR32RegClass);
23945 if (VT == MVT::i16)
23946 return std::make_pair(0U, &X86::GR16RegClass);
23947 if (VT == MVT::i8 || VT == MVT::i1)
23948 return std::make_pair(0U, &X86::GR8RegClass);
23949 if (VT == MVT::i64 || VT == MVT::f64)
23950 return std::make_pair(0U, &X86::GR64RegClass);
23953 // 32-bit fallthrough
23954 case 'Q': // Q_REGS
23955 if (VT == MVT::i32 || VT == MVT::f32)
23956 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
23957 if (VT == MVT::i16)
23958 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
23959 if (VT == MVT::i8 || VT == MVT::i1)
23960 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
23961 if (VT == MVT::i64)
23962 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
23964 case 'r': // GENERAL_REGS
23965 case 'l': // INDEX_REGS
23966 if (VT == MVT::i8 || VT == MVT::i1)
23967 return std::make_pair(0U, &X86::GR8RegClass);
23968 if (VT == MVT::i16)
23969 return std::make_pair(0U, &X86::GR16RegClass);
23970 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
23971 return std::make_pair(0U, &X86::GR32RegClass);
23972 return std::make_pair(0U, &X86::GR64RegClass);
23973 case 'R': // LEGACY_REGS
23974 if (VT == MVT::i8 || VT == MVT::i1)
23975 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
23976 if (VT == MVT::i16)
23977 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
23978 if (VT == MVT::i32 || !Subtarget->is64Bit())
23979 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
23980 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
23981 case 'f': // FP Stack registers.
23982 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
23983 // value to the correct fpstack register class.
23984 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
23985 return std::make_pair(0U, &X86::RFP32RegClass);
23986 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
23987 return std::make_pair(0U, &X86::RFP64RegClass);
23988 return std::make_pair(0U, &X86::RFP80RegClass);
23989 case 'y': // MMX_REGS if MMX allowed.
23990 if (!Subtarget->hasMMX()) break;
23991 return std::make_pair(0U, &X86::VR64RegClass);
23992 case 'Y': // SSE_REGS if SSE2 allowed
23993 if (!Subtarget->hasSSE2()) break;
23995 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
23996 if (!Subtarget->hasSSE1()) break;
23998 switch (VT.SimpleTy) {
24000 // Scalar SSE types.
24003 return std::make_pair(0U, &X86::FR32RegClass);
24006 return std::make_pair(0U, &X86::FR64RegClass);
24014 return std::make_pair(0U, &X86::VR128RegClass);
24022 return std::make_pair(0U, &X86::VR256RegClass);
24027 return std::make_pair(0U, &X86::VR512RegClass);
24033 // Use the default implementation in TargetLowering to convert the register
24034 // constraint into a member of a register class.
24035 std::pair<unsigned, const TargetRegisterClass*> Res;
24036 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
24038 // Not found as a standard register?
24040 // Map st(0) -> st(7) -> ST0
24041 if (Constraint.size() == 7 && Constraint[0] == '{' &&
24042 tolower(Constraint[1]) == 's' &&
24043 tolower(Constraint[2]) == 't' &&
24044 Constraint[3] == '(' &&
24045 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
24046 Constraint[5] == ')' &&
24047 Constraint[6] == '}') {
24049 Res.first = X86::FP0+Constraint[4]-'0';
24050 Res.second = &X86::RFP80RegClass;
24054 // GCC allows "st(0)" to be called just plain "st".
24055 if (StringRef("{st}").equals_lower(Constraint)) {
24056 Res.first = X86::FP0;
24057 Res.second = &X86::RFP80RegClass;
24062 if (StringRef("{flags}").equals_lower(Constraint)) {
24063 Res.first = X86::EFLAGS;
24064 Res.second = &X86::CCRRegClass;
24068 // 'A' means EAX + EDX.
24069 if (Constraint == "A") {
24070 Res.first = X86::EAX;
24071 Res.second = &X86::GR32_ADRegClass;
24077 // Otherwise, check to see if this is a register class of the wrong value
24078 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
24079 // turn into {ax},{dx}.
24080 if (Res.second->hasType(VT))
24081 return Res; // Correct type already, nothing to do.
24083 // All of the single-register GCC register classes map their values onto
24084 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
24085 // really want an 8-bit or 32-bit register, map to the appropriate register
24086 // class and return the appropriate register.
24087 if (Res.second == &X86::GR16RegClass) {
24088 if (VT == MVT::i8 || VT == MVT::i1) {
24089 unsigned DestReg = 0;
24090 switch (Res.first) {
24092 case X86::AX: DestReg = X86::AL; break;
24093 case X86::DX: DestReg = X86::DL; break;
24094 case X86::CX: DestReg = X86::CL; break;
24095 case X86::BX: DestReg = X86::BL; break;
24098 Res.first = DestReg;
24099 Res.second = &X86::GR8RegClass;
24101 } else if (VT == MVT::i32 || VT == MVT::f32) {
24102 unsigned DestReg = 0;
24103 switch (Res.first) {
24105 case X86::AX: DestReg = X86::EAX; break;
24106 case X86::DX: DestReg = X86::EDX; break;
24107 case X86::CX: DestReg = X86::ECX; break;
24108 case X86::BX: DestReg = X86::EBX; break;
24109 case X86::SI: DestReg = X86::ESI; break;
24110 case X86::DI: DestReg = X86::EDI; break;
24111 case X86::BP: DestReg = X86::EBP; break;
24112 case X86::SP: DestReg = X86::ESP; break;
24115 Res.first = DestReg;
24116 Res.second = &X86::GR32RegClass;
24118 } else if (VT == MVT::i64 || VT == MVT::f64) {
24119 unsigned DestReg = 0;
24120 switch (Res.first) {
24122 case X86::AX: DestReg = X86::RAX; break;
24123 case X86::DX: DestReg = X86::RDX; break;
24124 case X86::CX: DestReg = X86::RCX; break;
24125 case X86::BX: DestReg = X86::RBX; break;
24126 case X86::SI: DestReg = X86::RSI; break;
24127 case X86::DI: DestReg = X86::RDI; break;
24128 case X86::BP: DestReg = X86::RBP; break;
24129 case X86::SP: DestReg = X86::RSP; break;
24132 Res.first = DestReg;
24133 Res.second = &X86::GR64RegClass;
24136 } else if (Res.second == &X86::FR32RegClass ||
24137 Res.second == &X86::FR64RegClass ||
24138 Res.second == &X86::VR128RegClass ||
24139 Res.second == &X86::VR256RegClass ||
24140 Res.second == &X86::FR32XRegClass ||
24141 Res.second == &X86::FR64XRegClass ||
24142 Res.second == &X86::VR128XRegClass ||
24143 Res.second == &X86::VR256XRegClass ||
24144 Res.second == &X86::VR512RegClass) {
24145 // Handle references to XMM physical registers that got mapped into the
24146 // wrong class. This can happen with constraints like {xmm0} where the
24147 // target independent register mapper will just pick the first match it can
24148 // find, ignoring the required type.
24150 if (VT == MVT::f32 || VT == MVT::i32)
24151 Res.second = &X86::FR32RegClass;
24152 else if (VT == MVT::f64 || VT == MVT::i64)
24153 Res.second = &X86::FR64RegClass;
24154 else if (X86::VR128RegClass.hasType(VT))
24155 Res.second = &X86::VR128RegClass;
24156 else if (X86::VR256RegClass.hasType(VT))
24157 Res.second = &X86::VR256RegClass;
24158 else if (X86::VR512RegClass.hasType(VT))
24159 Res.second = &X86::VR512RegClass;
24165 int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
24167 // Scaling factors are not free at all.
24168 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
24169 // will take 2 allocations in the out of order engine instead of 1
24170 // for plain addressing mode, i.e. inst (reg1).
24172 // vaddps (%rsi,%drx), %ymm0, %ymm1
24173 // Requires two allocations (one for the load, one for the computation)
24175 // vaddps (%rsi), %ymm0, %ymm1
24176 // Requires just 1 allocation, i.e., freeing allocations for other operations
24177 // and having less micro operations to execute.
24179 // For some X86 architectures, this is even worse because for instance for
24180 // stores, the complex addressing mode forces the instruction to use the
24181 // "load" ports instead of the dedicated "store" port.
24182 // E.g., on Haswell:
24183 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
24184 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
24185 if (isLegalAddressingMode(AM, Ty))
24186 // Scale represents reg2 * scale, thus account for 1
24187 // as soon as we use a second register.
24188 return AM.Scale != 0;
24192 bool X86TargetLowering::isTargetFTOL() const {
24193 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();