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 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/IntrinsicLowering.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/MC/MCAsmInfo.h"
40 #include "llvm/MC/MCContext.h"
41 #include "llvm/MC/MCExpr.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/ADT/BitVector.h"
44 #include "llvm/ADT/SmallSet.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/StringExtras.h"
47 #include "llvm/ADT/VectorExtras.h"
48 #include "llvm/Support/CallSite.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/Dwarf.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
55 using namespace dwarf;
57 STATISTIC(NumTailCalls, "Number of tail calls");
59 // Forward declarations.
60 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
63 static SDValue Insert128BitVector(SDValue Result,
69 static SDValue Extract128BitVector(SDValue Vec,
74 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
77 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
78 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
79 /// simple subregister reference. Idx is an index in the 128 bits we
80 /// want. It need not be aligned to a 128-bit bounday. That makes
81 /// lowering EXTRACT_VECTOR_ELT operations easier.
82 static SDValue Extract128BitVector(SDValue Vec,
86 EVT VT = Vec.getValueType();
87 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
89 EVT ElVT = VT.getVectorElementType();
91 int Factor = VT.getSizeInBits() / 128;
93 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(),
95 VT.getVectorNumElements() / Factor);
97 // Extract from UNDEF is UNDEF.
98 if (Vec.getOpcode() == ISD::UNDEF)
99 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
101 if (isa<ConstantSDNode>(Idx)) {
102 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
104 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
105 // we can match to VEXTRACTF128.
106 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
108 // This is the index of the first element of the 128-bit chunk
110 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
113 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
115 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
124 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
125 /// sets things up to match to an AVX VINSERTF128 instruction or a
126 /// simple superregister reference. Idx is an index in the 128 bits
127 /// we want. It need not be aligned to a 128-bit bounday. That makes
128 /// lowering INSERT_VECTOR_ELT operations easier.
129 static SDValue Insert128BitVector(SDValue Result,
134 if (isa<ConstantSDNode>(Idx)) {
135 EVT VT = Vec.getValueType();
136 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
138 EVT ElVT = VT.getVectorElementType();
140 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
142 EVT ResultVT = Result.getValueType();
144 // Insert the relevant 128 bits.
145 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
147 // This is the index of the first element of the 128-bit chunk
149 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
152 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
154 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
162 /// Given two vectors, concat them.
163 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
164 DebugLoc dl = Lower.getDebugLoc();
166 assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
168 EVT VT = EVT::getVectorVT(*DAG.getContext(),
169 Lower.getValueType().getVectorElementType(),
170 Lower.getValueType().getVectorNumElements() * 2);
172 // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
173 assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
175 // Insert the upper subvector.
176 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
178 // This is half the length of the result
179 // vector. Start inserting the upper 128
181 Lower.getValueType().getVectorNumElements(),
185 // Insert the lower subvector.
186 Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
190 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
191 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
192 bool is64Bit = Subtarget->is64Bit();
194 if (Subtarget->isTargetEnvMacho()) {
196 return new X8664_MachoTargetObjectFile();
197 return new TargetLoweringObjectFileMachO();
200 if (Subtarget->isTargetELF()) {
202 return new X8664_ELFTargetObjectFile(TM);
203 return new X8632_ELFTargetObjectFile(TM);
205 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
206 return new TargetLoweringObjectFileCOFF();
207 llvm_unreachable("unknown subtarget type");
210 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
211 : TargetLowering(TM, createTLOF(TM)) {
212 Subtarget = &TM.getSubtarget<X86Subtarget>();
213 X86ScalarSSEf64 = Subtarget->hasXMMInt();
214 X86ScalarSSEf32 = Subtarget->hasXMM();
215 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
217 RegInfo = TM.getRegisterInfo();
218 TD = getTargetData();
220 // Set up the TargetLowering object.
221 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
223 // X86 is weird, it always uses i8 for shift amounts and setcc results.
224 setBooleanContents(ZeroOrOneBooleanContent);
226 // For 64-bit since we have so many registers use the ILP scheduler, for
227 // 32-bit code use the register pressure specific scheduling.
228 if (Subtarget->is64Bit())
229 setSchedulingPreference(Sched::ILP);
231 setSchedulingPreference(Sched::RegPressure);
232 setStackPointerRegisterToSaveRestore(X86StackPtr);
234 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
235 // Setup Windows compiler runtime calls.
236 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
237 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
238 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
239 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
240 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
241 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
242 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
243 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
246 if (Subtarget->isTargetDarwin()) {
247 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
248 setUseUnderscoreSetJmp(false);
249 setUseUnderscoreLongJmp(false);
250 } else if (Subtarget->isTargetMingw()) {
251 // MS runtime is weird: it exports _setjmp, but longjmp!
252 setUseUnderscoreSetJmp(true);
253 setUseUnderscoreLongJmp(false);
255 setUseUnderscoreSetJmp(true);
256 setUseUnderscoreLongJmp(true);
259 // Set up the register classes.
260 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
261 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
262 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
263 if (Subtarget->is64Bit())
264 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
266 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
268 // We don't accept any truncstore of integer registers.
269 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
270 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
271 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
272 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
273 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
274 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
276 // SETOEQ and SETUNE require checking two conditions.
277 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
278 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
279 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
280 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
281 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
282 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
284 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
286 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
287 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
288 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
290 if (Subtarget->is64Bit()) {
291 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
292 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
293 } else if (!UseSoftFloat) {
294 // We have an algorithm for SSE2->double, and we turn this into a
295 // 64-bit FILD followed by conditional FADD for other targets.
296 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
297 // We have an algorithm for SSE2, and we turn this into a 64-bit
298 // FILD for other targets.
299 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
302 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
304 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
305 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
308 // SSE has no i16 to fp conversion, only i32
309 if (X86ScalarSSEf32) {
310 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
311 // f32 and f64 cases are Legal, f80 case is not
312 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
314 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
315 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
318 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
319 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
322 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
323 // are Legal, f80 is custom lowered.
324 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
325 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
327 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
329 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
330 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
332 if (X86ScalarSSEf32) {
333 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
334 // f32 and f64 cases are Legal, f80 case is not
335 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
337 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
338 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
341 // Handle FP_TO_UINT by promoting the destination to a larger signed
343 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
344 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
345 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
347 if (Subtarget->is64Bit()) {
348 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
349 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
350 } else if (!UseSoftFloat) {
351 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
352 // Expand FP_TO_UINT into a select.
353 // FIXME: We would like to use a Custom expander here eventually to do
354 // the optimal thing for SSE vs. the default expansion in the legalizer.
355 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
357 // With SSE3 we can use fisttpll to convert to a signed i64; without
358 // SSE, we're stuck with a fistpll.
359 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
362 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
363 if (!X86ScalarSSEf64) {
364 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
365 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
366 if (Subtarget->is64Bit()) {
367 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
368 // Without SSE, i64->f64 goes through memory.
369 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
373 // Scalar integer divide and remainder are lowered to use operations that
374 // produce two results, to match the available instructions. This exposes
375 // the two-result form to trivial CSE, which is able to combine x/y and x%y
376 // into a single instruction.
378 // Scalar integer multiply-high is also lowered to use two-result
379 // operations, to match the available instructions. However, plain multiply
380 // (low) operations are left as Legal, as there are single-result
381 // instructions for this in x86. Using the two-result multiply instructions
382 // when both high and low results are needed must be arranged by dagcombine.
383 for (unsigned i = 0, e = 4; i != e; ++i) {
385 setOperationAction(ISD::MULHS, VT, Expand);
386 setOperationAction(ISD::MULHU, VT, Expand);
387 setOperationAction(ISD::SDIV, VT, Expand);
388 setOperationAction(ISD::UDIV, VT, Expand);
389 setOperationAction(ISD::SREM, VT, Expand);
390 setOperationAction(ISD::UREM, VT, Expand);
392 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
393 setOperationAction(ISD::ADDC, VT, Custom);
394 setOperationAction(ISD::ADDE, VT, Custom);
395 setOperationAction(ISD::SUBC, VT, Custom);
396 setOperationAction(ISD::SUBE, VT, Custom);
399 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
400 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
401 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
402 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
403 if (Subtarget->is64Bit())
404 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
405 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
406 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
407 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
408 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
409 setOperationAction(ISD::FREM , MVT::f32 , Expand);
410 setOperationAction(ISD::FREM , MVT::f64 , Expand);
411 setOperationAction(ISD::FREM , MVT::f80 , Expand);
412 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
414 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
415 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
416 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
417 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
418 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
419 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
420 if (Subtarget->is64Bit()) {
421 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
422 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
425 if (Subtarget->hasPOPCNT()) {
426 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
428 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
429 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
430 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
431 if (Subtarget->is64Bit())
432 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
435 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
436 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
438 // These should be promoted to a larger select which is supported.
439 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
440 // X86 wants to expand cmov itself.
441 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
442 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
443 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
444 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
445 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
446 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
447 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
448 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
449 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
450 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
451 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
452 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
453 if (Subtarget->is64Bit()) {
454 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
455 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
457 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
460 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
461 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
462 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
463 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
464 if (Subtarget->is64Bit())
465 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
466 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
467 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
468 if (Subtarget->is64Bit()) {
469 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
470 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
471 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
472 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
473 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
475 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
476 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
477 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
478 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
479 if (Subtarget->is64Bit()) {
480 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
481 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
482 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
485 if (Subtarget->hasXMM())
486 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
488 // We may not have a libcall for MEMBARRIER so we should lower this.
489 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
491 // On X86 and X86-64, atomic operations are lowered to locked instructions.
492 // Locked instructions, in turn, have implicit fence semantics (all memory
493 // operations are flushed before issuing the locked instruction, and they
494 // are not buffered), so we can fold away the common pattern of
495 // fence-atomic-fence.
496 setShouldFoldAtomicFences(true);
498 // Expand certain atomics
499 for (unsigned i = 0, e = 4; i != e; ++i) {
501 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
502 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
505 if (!Subtarget->is64Bit()) {
506 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
507 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
508 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
509 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
510 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
511 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
512 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
515 // FIXME - use subtarget debug flags
516 if (!Subtarget->isTargetDarwin() &&
517 !Subtarget->isTargetELF() &&
518 !Subtarget->isTargetCygMing()) {
519 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
522 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
523 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
524 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
525 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
526 if (Subtarget->is64Bit()) {
527 setExceptionPointerRegister(X86::RAX);
528 setExceptionSelectorRegister(X86::RDX);
530 setExceptionPointerRegister(X86::EAX);
531 setExceptionSelectorRegister(X86::EDX);
533 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
534 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
536 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
538 setOperationAction(ISD::TRAP, MVT::Other, Legal);
540 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
541 setOperationAction(ISD::VASTART , MVT::Other, Custom);
542 setOperationAction(ISD::VAEND , MVT::Other, Expand);
543 if (Subtarget->is64Bit()) {
544 setOperationAction(ISD::VAARG , MVT::Other, Custom);
545 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
547 setOperationAction(ISD::VAARG , MVT::Other, Expand);
548 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
551 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
552 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
553 setOperationAction(ISD::DYNAMIC_STACKALLOC,
554 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
555 (Subtarget->isTargetCOFF()
556 && !Subtarget->isTargetEnvMacho()
559 if (!UseSoftFloat && X86ScalarSSEf64) {
560 // f32 and f64 use SSE.
561 // Set up the FP register classes.
562 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
563 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
565 // Use ANDPD to simulate FABS.
566 setOperationAction(ISD::FABS , MVT::f64, Custom);
567 setOperationAction(ISD::FABS , MVT::f32, Custom);
569 // Use XORP to simulate FNEG.
570 setOperationAction(ISD::FNEG , MVT::f64, Custom);
571 setOperationAction(ISD::FNEG , MVT::f32, Custom);
573 // Use ANDPD and ORPD to simulate FCOPYSIGN.
574 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
575 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
577 // Lower this to FGETSIGNx86 plus an AND.
578 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
579 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
581 // We don't support sin/cos/fmod
582 setOperationAction(ISD::FSIN , MVT::f64, Expand);
583 setOperationAction(ISD::FCOS , MVT::f64, Expand);
584 setOperationAction(ISD::FSIN , MVT::f32, Expand);
585 setOperationAction(ISD::FCOS , MVT::f32, Expand);
587 // Expand FP immediates into loads from the stack, except for the special
589 addLegalFPImmediate(APFloat(+0.0)); // xorpd
590 addLegalFPImmediate(APFloat(+0.0f)); // xorps
591 } else if (!UseSoftFloat && X86ScalarSSEf32) {
592 // Use SSE for f32, x87 for f64.
593 // Set up the FP register classes.
594 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
595 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
597 // Use ANDPS to simulate FABS.
598 setOperationAction(ISD::FABS , MVT::f32, Custom);
600 // Use XORP to simulate FNEG.
601 setOperationAction(ISD::FNEG , MVT::f32, Custom);
603 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
605 // Use ANDPS and ORPS to simulate FCOPYSIGN.
606 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
607 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
609 // We don't support sin/cos/fmod
610 setOperationAction(ISD::FSIN , MVT::f32, Expand);
611 setOperationAction(ISD::FCOS , MVT::f32, Expand);
613 // Special cases we handle for FP constants.
614 addLegalFPImmediate(APFloat(+0.0f)); // xorps
615 addLegalFPImmediate(APFloat(+0.0)); // FLD0
616 addLegalFPImmediate(APFloat(+1.0)); // FLD1
617 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
618 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
621 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
622 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
624 } else if (!UseSoftFloat) {
625 // f32 and f64 in x87.
626 // Set up the FP register classes.
627 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
628 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
630 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
631 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
632 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
633 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
636 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
637 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
639 addLegalFPImmediate(APFloat(+0.0)); // FLD0
640 addLegalFPImmediate(APFloat(+1.0)); // FLD1
641 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
642 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
643 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
644 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
645 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
646 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
649 // Long double always uses X87.
651 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
652 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
653 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
655 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
656 addLegalFPImmediate(TmpFlt); // FLD0
658 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
661 APFloat TmpFlt2(+1.0);
662 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
664 addLegalFPImmediate(TmpFlt2); // FLD1
665 TmpFlt2.changeSign();
666 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
670 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
671 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
675 // Always use a library call for pow.
676 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
677 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
678 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
680 setOperationAction(ISD::FLOG, MVT::f80, Expand);
681 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
682 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
683 setOperationAction(ISD::FEXP, MVT::f80, Expand);
684 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
686 // First set operation action for all vector types to either promote
687 // (for widening) or expand (for scalarization). Then we will selectively
688 // turn on ones that can be effectively codegen'd.
689 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
690 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
691 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
706 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
708 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
709 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
719 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
723 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
724 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
725 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
726 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
727 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
728 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
729 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
730 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
731 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
732 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
733 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
734 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
735 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
736 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
737 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
738 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
739 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
740 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
741 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
742 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
743 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
744 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
745 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
746 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
747 setTruncStoreAction((MVT::SimpleValueType)VT,
748 (MVT::SimpleValueType)InnerVT, Expand);
749 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
750 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
751 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
754 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
755 // with -msoft-float, disable use of MMX as well.
756 if (!UseSoftFloat && Subtarget->hasMMX()) {
757 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
758 // No operations on x86mmx supported, everything uses intrinsics.
761 // MMX-sized vectors (other than x86mmx) are expected to be expanded
762 // into smaller operations.
763 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
764 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
765 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
766 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
767 setOperationAction(ISD::AND, MVT::v8i8, Expand);
768 setOperationAction(ISD::AND, MVT::v4i16, Expand);
769 setOperationAction(ISD::AND, MVT::v2i32, Expand);
770 setOperationAction(ISD::AND, MVT::v1i64, Expand);
771 setOperationAction(ISD::OR, MVT::v8i8, Expand);
772 setOperationAction(ISD::OR, MVT::v4i16, Expand);
773 setOperationAction(ISD::OR, MVT::v2i32, Expand);
774 setOperationAction(ISD::OR, MVT::v1i64, Expand);
775 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
776 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
777 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
778 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
779 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
780 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
781 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
782 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
783 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
784 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
785 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
786 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
787 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
788 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
789 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
790 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
791 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
793 if (!UseSoftFloat && Subtarget->hasXMM()) {
794 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
796 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
797 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
798 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
799 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
800 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
801 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
802 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
803 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
804 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
805 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
806 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
807 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
810 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
811 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
813 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
814 // registers cannot be used even for integer operations.
815 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
816 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
817 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
818 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
820 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
821 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
822 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
823 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
824 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
825 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
826 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
827 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
828 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
829 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
830 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
831 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
832 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
833 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
834 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
835 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
837 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
838 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
839 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
840 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
842 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
843 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
848 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
849 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
850 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
851 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
852 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
854 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
855 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
856 EVT VT = (MVT::SimpleValueType)i;
857 // Do not attempt to custom lower non-power-of-2 vectors
858 if (!isPowerOf2_32(VT.getVectorNumElements()))
860 // Do not attempt to custom lower non-128-bit vectors
861 if (!VT.is128BitVector())
863 setOperationAction(ISD::BUILD_VECTOR,
864 VT.getSimpleVT().SimpleTy, Custom);
865 setOperationAction(ISD::VECTOR_SHUFFLE,
866 VT.getSimpleVT().SimpleTy, Custom);
867 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
868 VT.getSimpleVT().SimpleTy, Custom);
871 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
872 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
873 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
874 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
875 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
876 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
878 if (Subtarget->is64Bit()) {
879 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
880 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
883 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
884 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
885 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
888 // Do not attempt to promote non-128-bit vectors
889 if (!VT.is128BitVector())
892 setOperationAction(ISD::AND, SVT, Promote);
893 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
894 setOperationAction(ISD::OR, SVT, Promote);
895 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
896 setOperationAction(ISD::XOR, SVT, Promote);
897 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
898 setOperationAction(ISD::LOAD, SVT, Promote);
899 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
900 setOperationAction(ISD::SELECT, SVT, Promote);
901 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
904 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
906 // Custom lower v2i64 and v2f64 selects.
907 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
908 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
909 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
910 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
912 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
913 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
916 if (Subtarget->hasSSE41()) {
917 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
918 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
919 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
920 setOperationAction(ISD::FRINT, MVT::f32, Legal);
921 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
922 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
923 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
924 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
925 setOperationAction(ISD::FRINT, MVT::f64, Legal);
926 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
928 // FIXME: Do we need to handle scalar-to-vector here?
929 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
931 // Can turn SHL into an integer multiply.
932 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
933 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
935 // i8 and i16 vectors are custom , because the source register and source
936 // source memory operand types are not the same width. f32 vectors are
937 // custom since the immediate controlling the insert encodes additional
939 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
940 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
941 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
942 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
944 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
945 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
946 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
947 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
949 if (Subtarget->is64Bit()) {
950 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
951 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
955 if (Subtarget->hasSSE2()) {
956 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
957 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
958 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
960 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
961 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
962 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
964 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
965 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
968 if (Subtarget->hasSSE42())
969 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
971 if (!UseSoftFloat && Subtarget->hasAVX()) {
972 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
973 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
974 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
975 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
976 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
978 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
979 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
980 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
981 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
983 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
984 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
985 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
986 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
987 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
988 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
990 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
991 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
992 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
993 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
994 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
995 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
997 // Custom lower build_vector, vector_shuffle, scalar_to_vector,
998 // insert_vector_elt extract_subvector and extract_vector_elt for
1000 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1001 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1003 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
1004 // Do not attempt to custom lower non-256-bit vectors
1005 if (!isPowerOf2_32(MVT(VT).getVectorNumElements())
1006 || (MVT(VT).getSizeInBits() < 256))
1008 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1009 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1011 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1012 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1014 // Custom-lower insert_subvector and extract_subvector based on
1016 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1017 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1019 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
1020 // Do not attempt to custom lower non-256-bit vectors
1021 if (!isPowerOf2_32(MVT(VT).getVectorNumElements()))
1024 if (MVT(VT).getSizeInBits() == 128) {
1025 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1027 else if (MVT(VT).getSizeInBits() == 256) {
1028 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1032 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1033 // Don't promote loads because we need them for VPERM vector index versions.
1035 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1036 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1038 if (!isPowerOf2_32(MVT((MVT::SimpleValueType)VT).getVectorNumElements())
1039 || (MVT((MVT::SimpleValueType)VT).getSizeInBits() < 256))
1041 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
1042 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v4i64);
1043 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
1044 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v4i64);
1045 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
1046 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v4i64);
1047 //setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
1048 //AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v4i64);
1049 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
1050 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v4i64);
1054 // We want to custom lower some of our intrinsics.
1055 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1058 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1059 // handle type legalization for these operations here.
1061 // FIXME: We really should do custom legalization for addition and
1062 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1063 // than generic legalization for 64-bit multiplication-with-overflow, though.
1064 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1065 // Add/Sub/Mul with overflow operations are custom lowered.
1067 setOperationAction(ISD::SADDO, VT, Custom);
1068 setOperationAction(ISD::UADDO, VT, Custom);
1069 setOperationAction(ISD::SSUBO, VT, Custom);
1070 setOperationAction(ISD::USUBO, VT, Custom);
1071 setOperationAction(ISD::SMULO, VT, Custom);
1072 setOperationAction(ISD::UMULO, VT, Custom);
1075 // There are no 8-bit 3-address imul/mul instructions
1076 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1077 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1079 if (!Subtarget->is64Bit()) {
1080 // These libcalls are not available in 32-bit.
1081 setLibcallName(RTLIB::SHL_I128, 0);
1082 setLibcallName(RTLIB::SRL_I128, 0);
1083 setLibcallName(RTLIB::SRA_I128, 0);
1086 // We have target-specific dag combine patterns for the following nodes:
1087 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1088 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1089 setTargetDAGCombine(ISD::BUILD_VECTOR);
1090 setTargetDAGCombine(ISD::SELECT);
1091 setTargetDAGCombine(ISD::SHL);
1092 setTargetDAGCombine(ISD::SRA);
1093 setTargetDAGCombine(ISD::SRL);
1094 setTargetDAGCombine(ISD::OR);
1095 setTargetDAGCombine(ISD::AND);
1096 setTargetDAGCombine(ISD::ADD);
1097 setTargetDAGCombine(ISD::SUB);
1098 setTargetDAGCombine(ISD::STORE);
1099 setTargetDAGCombine(ISD::ZERO_EXTEND);
1100 setTargetDAGCombine(ISD::SINT_TO_FP);
1101 if (Subtarget->is64Bit())
1102 setTargetDAGCombine(ISD::MUL);
1104 computeRegisterProperties();
1106 // On Darwin, -Os means optimize for size without hurting performance,
1107 // do not reduce the limit.
1108 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1109 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1110 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1111 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1112 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1113 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1114 setPrefLoopAlignment(16);
1115 benefitFromCodePlacementOpt = true;
1117 setPrefFunctionAlignment(4);
1121 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1126 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1127 /// the desired ByVal argument alignment.
1128 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1131 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1132 if (VTy->getBitWidth() == 128)
1134 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1135 unsigned EltAlign = 0;
1136 getMaxByValAlign(ATy->getElementType(), EltAlign);
1137 if (EltAlign > MaxAlign)
1138 MaxAlign = EltAlign;
1139 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1140 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1141 unsigned EltAlign = 0;
1142 getMaxByValAlign(STy->getElementType(i), EltAlign);
1143 if (EltAlign > MaxAlign)
1144 MaxAlign = EltAlign;
1152 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1153 /// function arguments in the caller parameter area. For X86, aggregates
1154 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1155 /// are at 4-byte boundaries.
1156 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1157 if (Subtarget->is64Bit()) {
1158 // Max of 8 and alignment of type.
1159 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1166 if (Subtarget->hasXMM())
1167 getMaxByValAlign(Ty, Align);
1171 /// getOptimalMemOpType - Returns the target specific optimal type for load
1172 /// and store operations as a result of memset, memcpy, and memmove
1173 /// lowering. If DstAlign is zero that means it's safe to destination
1174 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1175 /// means there isn't a need to check it against alignment requirement,
1176 /// probably because the source does not need to be loaded. If
1177 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1178 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1179 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1180 /// constant so it does not need to be loaded.
1181 /// It returns EVT::Other if the type should be determined using generic
1182 /// target-independent logic.
1184 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1185 unsigned DstAlign, unsigned SrcAlign,
1186 bool NonScalarIntSafe,
1188 MachineFunction &MF) const {
1189 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1190 // linux. This is because the stack realignment code can't handle certain
1191 // cases like PR2962. This should be removed when PR2962 is fixed.
1192 const Function *F = MF.getFunction();
1193 if (NonScalarIntSafe &&
1194 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1196 (Subtarget->isUnalignedMemAccessFast() ||
1197 ((DstAlign == 0 || DstAlign >= 16) &&
1198 (SrcAlign == 0 || SrcAlign >= 16))) &&
1199 Subtarget->getStackAlignment() >= 16) {
1200 if (Subtarget->hasSSE2())
1202 if (Subtarget->hasSSE1())
1204 } else if (!MemcpyStrSrc && Size >= 8 &&
1205 !Subtarget->is64Bit() &&
1206 Subtarget->getStackAlignment() >= 8 &&
1207 Subtarget->hasXMMInt()) {
1208 // Do not use f64 to lower memcpy if source is string constant. It's
1209 // better to use i32 to avoid the loads.
1213 if (Subtarget->is64Bit() && Size >= 8)
1218 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1219 /// current function. The returned value is a member of the
1220 /// MachineJumpTableInfo::JTEntryKind enum.
1221 unsigned X86TargetLowering::getJumpTableEncoding() const {
1222 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1224 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1225 Subtarget->isPICStyleGOT())
1226 return MachineJumpTableInfo::EK_Custom32;
1228 // Otherwise, use the normal jump table encoding heuristics.
1229 return TargetLowering::getJumpTableEncoding();
1233 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1234 const MachineBasicBlock *MBB,
1235 unsigned uid,MCContext &Ctx) const{
1236 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1237 Subtarget->isPICStyleGOT());
1238 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1240 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1241 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1244 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1246 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1247 SelectionDAG &DAG) const {
1248 if (!Subtarget->is64Bit())
1249 // This doesn't have DebugLoc associated with it, but is not really the
1250 // same as a Register.
1251 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1255 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1256 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1258 const MCExpr *X86TargetLowering::
1259 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1260 MCContext &Ctx) const {
1261 // X86-64 uses RIP relative addressing based on the jump table label.
1262 if (Subtarget->isPICStyleRIPRel())
1263 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1265 // Otherwise, the reference is relative to the PIC base.
1266 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1269 // FIXME: Why this routine is here? Move to RegInfo!
1270 std::pair<const TargetRegisterClass*, uint8_t>
1271 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1272 const TargetRegisterClass *RRC = 0;
1274 switch (VT.getSimpleVT().SimpleTy) {
1276 return TargetLowering::findRepresentativeClass(VT);
1277 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1278 RRC = (Subtarget->is64Bit()
1279 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1282 RRC = X86::VR64RegisterClass;
1284 case MVT::f32: case MVT::f64:
1285 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1286 case MVT::v4f32: case MVT::v2f64:
1287 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1289 RRC = X86::VR128RegisterClass;
1292 return std::make_pair(RRC, Cost);
1295 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1296 unsigned &Offset) const {
1297 if (!Subtarget->isTargetLinux())
1300 if (Subtarget->is64Bit()) {
1301 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1303 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1316 //===----------------------------------------------------------------------===//
1317 // Return Value Calling Convention Implementation
1318 //===----------------------------------------------------------------------===//
1320 #include "X86GenCallingConv.inc"
1323 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1324 MachineFunction &MF, bool isVarArg,
1325 const SmallVectorImpl<ISD::OutputArg> &Outs,
1326 LLVMContext &Context) const {
1327 SmallVector<CCValAssign, 16> RVLocs;
1328 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1330 return CCInfo.CheckReturn(Outs, RetCC_X86);
1334 X86TargetLowering::LowerReturn(SDValue Chain,
1335 CallingConv::ID CallConv, bool isVarArg,
1336 const SmallVectorImpl<ISD::OutputArg> &Outs,
1337 const SmallVectorImpl<SDValue> &OutVals,
1338 DebugLoc dl, SelectionDAG &DAG) const {
1339 MachineFunction &MF = DAG.getMachineFunction();
1340 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1342 SmallVector<CCValAssign, 16> RVLocs;
1343 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1344 RVLocs, *DAG.getContext());
1345 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1347 // Add the regs to the liveout set for the function.
1348 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1349 for (unsigned i = 0; i != RVLocs.size(); ++i)
1350 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1351 MRI.addLiveOut(RVLocs[i].getLocReg());
1355 SmallVector<SDValue, 6> RetOps;
1356 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1357 // Operand #1 = Bytes To Pop
1358 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1361 // Copy the result values into the output registers.
1362 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1363 CCValAssign &VA = RVLocs[i];
1364 assert(VA.isRegLoc() && "Can only return in registers!");
1365 SDValue ValToCopy = OutVals[i];
1366 EVT ValVT = ValToCopy.getValueType();
1368 // If this is x86-64, and we disabled SSE, we can't return FP values,
1369 // or SSE or MMX vectors.
1370 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1371 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1372 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1373 report_fatal_error("SSE register return with SSE disabled");
1375 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1376 // llvm-gcc has never done it right and no one has noticed, so this
1377 // should be OK for now.
1378 if (ValVT == MVT::f64 &&
1379 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1380 report_fatal_error("SSE2 register return with SSE2 disabled");
1382 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1383 // the RET instruction and handled by the FP Stackifier.
1384 if (VA.getLocReg() == X86::ST0 ||
1385 VA.getLocReg() == X86::ST1) {
1386 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1387 // change the value to the FP stack register class.
1388 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1389 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1390 RetOps.push_back(ValToCopy);
1391 // Don't emit a copytoreg.
1395 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1396 // which is returned in RAX / RDX.
1397 if (Subtarget->is64Bit()) {
1398 if (ValVT == MVT::x86mmx) {
1399 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1400 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1401 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1403 // If we don't have SSE2 available, convert to v4f32 so the generated
1404 // register is legal.
1405 if (!Subtarget->hasSSE2())
1406 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1411 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1412 Flag = Chain.getValue(1);
1415 // The x86-64 ABI for returning structs by value requires that we copy
1416 // the sret argument into %rax for the return. We saved the argument into
1417 // a virtual register in the entry block, so now we copy the value out
1419 if (Subtarget->is64Bit() &&
1420 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1421 MachineFunction &MF = DAG.getMachineFunction();
1422 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1423 unsigned Reg = FuncInfo->getSRetReturnReg();
1425 "SRetReturnReg should have been set in LowerFormalArguments().");
1426 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1428 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1429 Flag = Chain.getValue(1);
1431 // RAX now acts like a return value.
1432 MRI.addLiveOut(X86::RAX);
1435 RetOps[0] = Chain; // Update chain.
1437 // Add the flag if we have it.
1439 RetOps.push_back(Flag);
1441 return DAG.getNode(X86ISD::RET_FLAG, dl,
1442 MVT::Other, &RetOps[0], RetOps.size());
1445 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1446 if (N->getNumValues() != 1)
1448 if (!N->hasNUsesOfValue(1, 0))
1451 SDNode *Copy = *N->use_begin();
1452 if (Copy->getOpcode() != ISD::CopyToReg &&
1453 Copy->getOpcode() != ISD::FP_EXTEND)
1456 bool HasRet = false;
1457 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1459 if (UI->getOpcode() != X86ISD::RET_FLAG)
1468 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1469 ISD::NodeType ExtendKind) const {
1471 // TODO: Is this also valid on 32-bit?
1472 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1473 ReturnMVT = MVT::i8;
1475 ReturnMVT = MVT::i32;
1477 EVT MinVT = getRegisterType(Context, ReturnMVT);
1478 return VT.bitsLT(MinVT) ? MinVT : VT;
1481 /// LowerCallResult - Lower the result values of a call into the
1482 /// appropriate copies out of appropriate physical registers.
1485 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1486 CallingConv::ID CallConv, bool isVarArg,
1487 const SmallVectorImpl<ISD::InputArg> &Ins,
1488 DebugLoc dl, SelectionDAG &DAG,
1489 SmallVectorImpl<SDValue> &InVals) const {
1491 // Assign locations to each value returned by this call.
1492 SmallVector<CCValAssign, 16> RVLocs;
1493 bool Is64Bit = Subtarget->is64Bit();
1494 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1495 getTargetMachine(), RVLocs, *DAG.getContext());
1496 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1498 // Copy all of the result registers out of their specified physreg.
1499 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1500 CCValAssign &VA = RVLocs[i];
1501 EVT CopyVT = VA.getValVT();
1503 // If this is x86-64, and we disabled SSE, we can't return FP values
1504 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1505 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1506 report_fatal_error("SSE register return with SSE disabled");
1511 // If this is a call to a function that returns an fp value on the floating
1512 // point stack, we must guarantee the the value is popped from the stack, so
1513 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1514 // if the return value is not used. We use the FpGET_ST0 instructions
1516 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1517 // If we prefer to use the value in xmm registers, copy it out as f80 and
1518 // use a truncate to move it from fp stack reg to xmm reg.
1519 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1520 bool isST0 = VA.getLocReg() == X86::ST0;
1522 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1523 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1524 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1525 SDValue Ops[] = { Chain, InFlag };
1526 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
1528 Val = Chain.getValue(0);
1530 // Round the f80 to the right size, which also moves it to the appropriate
1532 if (CopyVT != VA.getValVT())
1533 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1534 // This truncation won't change the value.
1535 DAG.getIntPtrConstant(1));
1537 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1538 CopyVT, InFlag).getValue(1);
1539 Val = Chain.getValue(0);
1541 InFlag = Chain.getValue(2);
1542 InVals.push_back(Val);
1549 //===----------------------------------------------------------------------===//
1550 // C & StdCall & Fast Calling Convention implementation
1551 //===----------------------------------------------------------------------===//
1552 // StdCall calling convention seems to be standard for many Windows' API
1553 // routines and around. It differs from C calling convention just a little:
1554 // callee should clean up the stack, not caller. Symbols should be also
1555 // decorated in some fancy way :) It doesn't support any vector arguments.
1556 // For info on fast calling convention see Fast Calling Convention (tail call)
1557 // implementation LowerX86_32FastCCCallTo.
1559 /// CallIsStructReturn - Determines whether a call uses struct return
1561 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1565 return Outs[0].Flags.isSRet();
1568 /// ArgsAreStructReturn - Determines whether a function uses struct
1569 /// return semantics.
1571 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1575 return Ins[0].Flags.isSRet();
1578 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1579 /// by "Src" to address "Dst" with size and alignment information specified by
1580 /// the specific parameter attribute. The copy will be passed as a byval
1581 /// function parameter.
1583 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1584 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1586 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1588 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1589 /*isVolatile*/false, /*AlwaysInline=*/true,
1590 MachinePointerInfo(), MachinePointerInfo());
1593 /// IsTailCallConvention - Return true if the calling convention is one that
1594 /// supports tail call optimization.
1595 static bool IsTailCallConvention(CallingConv::ID CC) {
1596 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1599 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1600 if (!CI->isTailCall())
1604 CallingConv::ID CalleeCC = CS.getCallingConv();
1605 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1611 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1612 /// a tailcall target by changing its ABI.
1613 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1614 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1618 X86TargetLowering::LowerMemArgument(SDValue Chain,
1619 CallingConv::ID CallConv,
1620 const SmallVectorImpl<ISD::InputArg> &Ins,
1621 DebugLoc dl, SelectionDAG &DAG,
1622 const CCValAssign &VA,
1623 MachineFrameInfo *MFI,
1625 // Create the nodes corresponding to a load from this parameter slot.
1626 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1627 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1628 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1631 // If value is passed by pointer we have address passed instead of the value
1633 if (VA.getLocInfo() == CCValAssign::Indirect)
1634 ValVT = VA.getLocVT();
1636 ValVT = VA.getValVT();
1638 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1639 // changed with more analysis.
1640 // In case of tail call optimization mark all arguments mutable. Since they
1641 // could be overwritten by lowering of arguments in case of a tail call.
1642 if (Flags.isByVal()) {
1643 unsigned Bytes = Flags.getByValSize();
1644 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1645 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1646 return DAG.getFrameIndex(FI, getPointerTy());
1648 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1649 VA.getLocMemOffset(), isImmutable);
1650 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1651 return DAG.getLoad(ValVT, dl, Chain, FIN,
1652 MachinePointerInfo::getFixedStack(FI),
1658 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1659 CallingConv::ID CallConv,
1661 const SmallVectorImpl<ISD::InputArg> &Ins,
1664 SmallVectorImpl<SDValue> &InVals)
1666 MachineFunction &MF = DAG.getMachineFunction();
1667 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1669 const Function* Fn = MF.getFunction();
1670 if (Fn->hasExternalLinkage() &&
1671 Subtarget->isTargetCygMing() &&
1672 Fn->getName() == "main")
1673 FuncInfo->setForceFramePointer(true);
1675 MachineFrameInfo *MFI = MF.getFrameInfo();
1676 bool Is64Bit = Subtarget->is64Bit();
1677 bool IsWin64 = Subtarget->isTargetWin64();
1679 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1680 "Var args not supported with calling convention fastcc or ghc");
1682 // Assign locations to all of the incoming arguments.
1683 SmallVector<CCValAssign, 16> ArgLocs;
1684 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1685 ArgLocs, *DAG.getContext());
1687 // Allocate shadow area for Win64
1689 CCInfo.AllocateStack(32, 8);
1692 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1694 unsigned LastVal = ~0U;
1696 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1697 CCValAssign &VA = ArgLocs[i];
1698 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1700 assert(VA.getValNo() != LastVal &&
1701 "Don't support value assigned to multiple locs yet");
1702 LastVal = VA.getValNo();
1704 if (VA.isRegLoc()) {
1705 EVT RegVT = VA.getLocVT();
1706 TargetRegisterClass *RC = NULL;
1707 if (RegVT == MVT::i32)
1708 RC = X86::GR32RegisterClass;
1709 else if (Is64Bit && RegVT == MVT::i64)
1710 RC = X86::GR64RegisterClass;
1711 else if (RegVT == MVT::f32)
1712 RC = X86::FR32RegisterClass;
1713 else if (RegVT == MVT::f64)
1714 RC = X86::FR64RegisterClass;
1715 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1716 RC = X86::VR256RegisterClass;
1717 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1718 RC = X86::VR128RegisterClass;
1719 else if (RegVT == MVT::x86mmx)
1720 RC = X86::VR64RegisterClass;
1722 llvm_unreachable("Unknown argument type!");
1724 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1725 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1727 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1728 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1730 if (VA.getLocInfo() == CCValAssign::SExt)
1731 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1732 DAG.getValueType(VA.getValVT()));
1733 else if (VA.getLocInfo() == CCValAssign::ZExt)
1734 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1735 DAG.getValueType(VA.getValVT()));
1736 else if (VA.getLocInfo() == CCValAssign::BCvt)
1737 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1739 if (VA.isExtInLoc()) {
1740 // Handle MMX values passed in XMM regs.
1741 if (RegVT.isVector()) {
1742 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1745 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1748 assert(VA.isMemLoc());
1749 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1752 // If value is passed via pointer - do a load.
1753 if (VA.getLocInfo() == CCValAssign::Indirect)
1754 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1755 MachinePointerInfo(), false, false, 0);
1757 InVals.push_back(ArgValue);
1760 // The x86-64 ABI for returning structs by value requires that we copy
1761 // the sret argument into %rax for the return. Save the argument into
1762 // a virtual register so that we can access it from the return points.
1763 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1764 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1765 unsigned Reg = FuncInfo->getSRetReturnReg();
1767 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1768 FuncInfo->setSRetReturnReg(Reg);
1770 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1771 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1774 unsigned StackSize = CCInfo.getNextStackOffset();
1775 // Align stack specially for tail calls.
1776 if (FuncIsMadeTailCallSafe(CallConv))
1777 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1779 // If the function takes variable number of arguments, make a frame index for
1780 // the start of the first vararg value... for expansion of llvm.va_start.
1782 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1783 CallConv != CallingConv::X86_ThisCall)) {
1784 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1787 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1789 // FIXME: We should really autogenerate these arrays
1790 static const unsigned GPR64ArgRegsWin64[] = {
1791 X86::RCX, X86::RDX, X86::R8, X86::R9
1793 static const unsigned GPR64ArgRegs64Bit[] = {
1794 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1796 static const unsigned XMMArgRegs64Bit[] = {
1797 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1798 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1800 const unsigned *GPR64ArgRegs;
1801 unsigned NumXMMRegs = 0;
1804 // The XMM registers which might contain var arg parameters are shadowed
1805 // in their paired GPR. So we only need to save the GPR to their home
1807 TotalNumIntRegs = 4;
1808 GPR64ArgRegs = GPR64ArgRegsWin64;
1810 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1811 GPR64ArgRegs = GPR64ArgRegs64Bit;
1813 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1815 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1818 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1819 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1820 "SSE register cannot be used when SSE is disabled!");
1821 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1822 "SSE register cannot be used when SSE is disabled!");
1823 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1824 // Kernel mode asks for SSE to be disabled, so don't push them
1826 TotalNumXMMRegs = 0;
1829 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1830 // Get to the caller-allocated home save location. Add 8 to account
1831 // for the return address.
1832 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1833 FuncInfo->setRegSaveFrameIndex(
1834 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1835 // Fixup to set vararg frame on shadow area (4 x i64).
1837 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1839 // For X86-64, if there are vararg parameters that are passed via
1840 // registers, then we must store them to their spots on the stack so they
1841 // may be loaded by deferencing the result of va_next.
1842 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1843 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1844 FuncInfo->setRegSaveFrameIndex(
1845 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1849 // Store the integer parameter registers.
1850 SmallVector<SDValue, 8> MemOps;
1851 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1853 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1854 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1855 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1856 DAG.getIntPtrConstant(Offset));
1857 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1858 X86::GR64RegisterClass);
1859 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1861 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1862 MachinePointerInfo::getFixedStack(
1863 FuncInfo->getRegSaveFrameIndex(), Offset),
1865 MemOps.push_back(Store);
1869 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1870 // Now store the XMM (fp + vector) parameter registers.
1871 SmallVector<SDValue, 11> SaveXMMOps;
1872 SaveXMMOps.push_back(Chain);
1874 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1875 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1876 SaveXMMOps.push_back(ALVal);
1878 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1879 FuncInfo->getRegSaveFrameIndex()));
1880 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1881 FuncInfo->getVarArgsFPOffset()));
1883 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1884 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1885 X86::VR128RegisterClass);
1886 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1887 SaveXMMOps.push_back(Val);
1889 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1891 &SaveXMMOps[0], SaveXMMOps.size()));
1894 if (!MemOps.empty())
1895 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1896 &MemOps[0], MemOps.size());
1900 // Some CCs need callee pop.
1901 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1902 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1904 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1905 // If this is an sret function, the return should pop the hidden pointer.
1906 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1907 FuncInfo->setBytesToPopOnReturn(4);
1911 // RegSaveFrameIndex is X86-64 only.
1912 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1913 if (CallConv == CallingConv::X86_FastCall ||
1914 CallConv == CallingConv::X86_ThisCall)
1915 // fastcc functions can't have varargs.
1916 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1923 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1924 SDValue StackPtr, SDValue Arg,
1925 DebugLoc dl, SelectionDAG &DAG,
1926 const CCValAssign &VA,
1927 ISD::ArgFlagsTy Flags) const {
1928 unsigned LocMemOffset = VA.getLocMemOffset();
1929 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1930 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1931 if (Flags.isByVal())
1932 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1934 return DAG.getStore(Chain, dl, Arg, PtrOff,
1935 MachinePointerInfo::getStack(LocMemOffset),
1939 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1940 /// optimization is performed and it is required.
1942 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1943 SDValue &OutRetAddr, SDValue Chain,
1944 bool IsTailCall, bool Is64Bit,
1945 int FPDiff, DebugLoc dl) const {
1946 // Adjust the Return address stack slot.
1947 EVT VT = getPointerTy();
1948 OutRetAddr = getReturnAddressFrameIndex(DAG);
1950 // Load the "old" Return address.
1951 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1953 return SDValue(OutRetAddr.getNode(), 1);
1956 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1957 /// optimization is performed and it is required (FPDiff!=0).
1959 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1960 SDValue Chain, SDValue RetAddrFrIdx,
1961 bool Is64Bit, int FPDiff, DebugLoc dl) {
1962 // Store the return address to the appropriate stack slot.
1963 if (!FPDiff) return Chain;
1964 // Calculate the new stack slot for the return address.
1965 int SlotSize = Is64Bit ? 8 : 4;
1966 int NewReturnAddrFI =
1967 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1968 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1969 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1970 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1971 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1977 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1978 CallingConv::ID CallConv, bool isVarArg,
1980 const SmallVectorImpl<ISD::OutputArg> &Outs,
1981 const SmallVectorImpl<SDValue> &OutVals,
1982 const SmallVectorImpl<ISD::InputArg> &Ins,
1983 DebugLoc dl, SelectionDAG &DAG,
1984 SmallVectorImpl<SDValue> &InVals) const {
1985 MachineFunction &MF = DAG.getMachineFunction();
1986 bool Is64Bit = Subtarget->is64Bit();
1987 bool IsWin64 = Subtarget->isTargetWin64();
1988 bool IsStructRet = CallIsStructReturn(Outs);
1989 bool IsSibcall = false;
1992 // Check if it's really possible to do a tail call.
1993 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1994 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1995 Outs, OutVals, Ins, DAG);
1997 // Sibcalls are automatically detected tailcalls which do not require
1999 if (!GuaranteedTailCallOpt && isTailCall)
2006 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2007 "Var args not supported with calling convention fastcc or ghc");
2009 // Analyze operands of the call, assigning locations to each operand.
2010 SmallVector<CCValAssign, 16> ArgLocs;
2011 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2012 ArgLocs, *DAG.getContext());
2014 // Allocate shadow area for Win64
2016 CCInfo.AllocateStack(32, 8);
2019 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2021 // Get a count of how many bytes are to be pushed on the stack.
2022 unsigned NumBytes = CCInfo.getNextStackOffset();
2024 // This is a sibcall. The memory operands are available in caller's
2025 // own caller's stack.
2027 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2028 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2031 if (isTailCall && !IsSibcall) {
2032 // Lower arguments at fp - stackoffset + fpdiff.
2033 unsigned NumBytesCallerPushed =
2034 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2035 FPDiff = NumBytesCallerPushed - NumBytes;
2037 // Set the delta of movement of the returnaddr stackslot.
2038 // But only set if delta is greater than previous delta.
2039 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2040 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2044 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2046 SDValue RetAddrFrIdx;
2047 // Load return address for tail calls.
2048 if (isTailCall && FPDiff)
2049 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2050 Is64Bit, FPDiff, dl);
2052 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2053 SmallVector<SDValue, 8> MemOpChains;
2056 // Walk the register/memloc assignments, inserting copies/loads. In the case
2057 // of tail call optimization arguments are handle later.
2058 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2059 CCValAssign &VA = ArgLocs[i];
2060 EVT RegVT = VA.getLocVT();
2061 SDValue Arg = OutVals[i];
2062 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2063 bool isByVal = Flags.isByVal();
2065 // Promote the value if needed.
2066 switch (VA.getLocInfo()) {
2067 default: llvm_unreachable("Unknown loc info!");
2068 case CCValAssign::Full: break;
2069 case CCValAssign::SExt:
2070 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2072 case CCValAssign::ZExt:
2073 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2075 case CCValAssign::AExt:
2076 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2077 // Special case: passing MMX values in XMM registers.
2078 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2079 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2080 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2082 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2084 case CCValAssign::BCvt:
2085 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2087 case CCValAssign::Indirect: {
2088 // Store the argument.
2089 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2090 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2091 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2092 MachinePointerInfo::getFixedStack(FI),
2099 if (VA.isRegLoc()) {
2100 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2101 if (isVarArg && IsWin64) {
2102 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2103 // shadow reg if callee is a varargs function.
2104 unsigned ShadowReg = 0;
2105 switch (VA.getLocReg()) {
2106 case X86::XMM0: ShadowReg = X86::RCX; break;
2107 case X86::XMM1: ShadowReg = X86::RDX; break;
2108 case X86::XMM2: ShadowReg = X86::R8; break;
2109 case X86::XMM3: ShadowReg = X86::R9; break;
2112 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2114 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2115 assert(VA.isMemLoc());
2116 if (StackPtr.getNode() == 0)
2117 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2118 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2119 dl, DAG, VA, Flags));
2123 if (!MemOpChains.empty())
2124 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2125 &MemOpChains[0], MemOpChains.size());
2127 // Build a sequence of copy-to-reg nodes chained together with token chain
2128 // and flag operands which copy the outgoing args into registers.
2130 // Tail call byval lowering might overwrite argument registers so in case of
2131 // tail call optimization the copies to registers are lowered later.
2133 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2134 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2135 RegsToPass[i].second, InFlag);
2136 InFlag = Chain.getValue(1);
2139 if (Subtarget->isPICStyleGOT()) {
2140 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2143 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2144 DAG.getNode(X86ISD::GlobalBaseReg,
2145 DebugLoc(), getPointerTy()),
2147 InFlag = Chain.getValue(1);
2149 // If we are tail calling and generating PIC/GOT style code load the
2150 // address of the callee into ECX. The value in ecx is used as target of
2151 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2152 // for tail calls on PIC/GOT architectures. Normally we would just put the
2153 // address of GOT into ebx and then call target@PLT. But for tail calls
2154 // ebx would be restored (since ebx is callee saved) before jumping to the
2157 // Note: The actual moving to ECX is done further down.
2158 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2159 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2160 !G->getGlobal()->hasProtectedVisibility())
2161 Callee = LowerGlobalAddress(Callee, DAG);
2162 else if (isa<ExternalSymbolSDNode>(Callee))
2163 Callee = LowerExternalSymbol(Callee, DAG);
2167 if (Is64Bit && isVarArg && !IsWin64) {
2168 // From AMD64 ABI document:
2169 // For calls that may call functions that use varargs or stdargs
2170 // (prototype-less calls or calls to functions containing ellipsis (...) in
2171 // the declaration) %al is used as hidden argument to specify the number
2172 // of SSE registers used. The contents of %al do not need to match exactly
2173 // the number of registers, but must be an ubound on the number of SSE
2174 // registers used and is in the range 0 - 8 inclusive.
2176 // Count the number of XMM registers allocated.
2177 static const unsigned XMMArgRegs[] = {
2178 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2179 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2181 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2182 assert((Subtarget->hasXMM() || !NumXMMRegs)
2183 && "SSE registers cannot be used when SSE is disabled");
2185 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2186 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2187 InFlag = Chain.getValue(1);
2191 // For tail calls lower the arguments to the 'real' stack slot.
2193 // Force all the incoming stack arguments to be loaded from the stack
2194 // before any new outgoing arguments are stored to the stack, because the
2195 // outgoing stack slots may alias the incoming argument stack slots, and
2196 // the alias isn't otherwise explicit. This is slightly more conservative
2197 // than necessary, because it means that each store effectively depends
2198 // on every argument instead of just those arguments it would clobber.
2199 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2201 SmallVector<SDValue, 8> MemOpChains2;
2204 // Do not flag preceding copytoreg stuff together with the following stuff.
2206 if (GuaranteedTailCallOpt) {
2207 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2208 CCValAssign &VA = ArgLocs[i];
2211 assert(VA.isMemLoc());
2212 SDValue Arg = OutVals[i];
2213 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2214 // Create frame index.
2215 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2216 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2217 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2218 FIN = DAG.getFrameIndex(FI, getPointerTy());
2220 if (Flags.isByVal()) {
2221 // Copy relative to framepointer.
2222 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2223 if (StackPtr.getNode() == 0)
2224 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2226 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2228 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2232 // Store relative to framepointer.
2233 MemOpChains2.push_back(
2234 DAG.getStore(ArgChain, dl, Arg, FIN,
2235 MachinePointerInfo::getFixedStack(FI),
2241 if (!MemOpChains2.empty())
2242 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2243 &MemOpChains2[0], MemOpChains2.size());
2245 // Copy arguments to their registers.
2246 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2247 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2248 RegsToPass[i].second, InFlag);
2249 InFlag = Chain.getValue(1);
2253 // Store the return address to the appropriate stack slot.
2254 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2258 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2259 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2260 // In the 64-bit large code model, we have to make all calls
2261 // through a register, since the call instruction's 32-bit
2262 // pc-relative offset may not be large enough to hold the whole
2264 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2265 // If the callee is a GlobalAddress node (quite common, every direct call
2266 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2269 // We should use extra load for direct calls to dllimported functions in
2271 const GlobalValue *GV = G->getGlobal();
2272 if (!GV->hasDLLImportLinkage()) {
2273 unsigned char OpFlags = 0;
2274 bool ExtraLoad = false;
2275 unsigned WrapperKind = ISD::DELETED_NODE;
2277 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2278 // external symbols most go through the PLT in PIC mode. If the symbol
2279 // has hidden or protected visibility, or if it is static or local, then
2280 // we don't need to use the PLT - we can directly call it.
2281 if (Subtarget->isTargetELF() &&
2282 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2283 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2284 OpFlags = X86II::MO_PLT;
2285 } else if (Subtarget->isPICStyleStubAny() &&
2286 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2287 (!Subtarget->getTargetTriple().isMacOSX() ||
2288 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2289 // PC-relative references to external symbols should go through $stub,
2290 // unless we're building with the leopard linker or later, which
2291 // automatically synthesizes these stubs.
2292 OpFlags = X86II::MO_DARWIN_STUB;
2293 } else if (Subtarget->isPICStyleRIPRel() &&
2294 isa<Function>(GV) &&
2295 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2296 // If the function is marked as non-lazy, generate an indirect call
2297 // which loads from the GOT directly. This avoids runtime overhead
2298 // at the cost of eager binding (and one extra byte of encoding).
2299 OpFlags = X86II::MO_GOTPCREL;
2300 WrapperKind = X86ISD::WrapperRIP;
2304 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2305 G->getOffset(), OpFlags);
2307 // Add a wrapper if needed.
2308 if (WrapperKind != ISD::DELETED_NODE)
2309 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2310 // Add extra indirection if needed.
2312 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2313 MachinePointerInfo::getGOT(),
2316 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2317 unsigned char OpFlags = 0;
2319 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2320 // external symbols should go through the PLT.
2321 if (Subtarget->isTargetELF() &&
2322 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2323 OpFlags = X86II::MO_PLT;
2324 } else if (Subtarget->isPICStyleStubAny() &&
2325 (!Subtarget->getTargetTriple().isMacOSX() ||
2326 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2327 // PC-relative references to external symbols should go through $stub,
2328 // unless we're building with the leopard linker or later, which
2329 // automatically synthesizes these stubs.
2330 OpFlags = X86II::MO_DARWIN_STUB;
2333 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2337 // Returns a chain & a flag for retval copy to use.
2338 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2339 SmallVector<SDValue, 8> Ops;
2341 if (!IsSibcall && isTailCall) {
2342 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2343 DAG.getIntPtrConstant(0, true), InFlag);
2344 InFlag = Chain.getValue(1);
2347 Ops.push_back(Chain);
2348 Ops.push_back(Callee);
2351 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2353 // Add argument registers to the end of the list so that they are known live
2355 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2356 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2357 RegsToPass[i].second.getValueType()));
2359 // Add an implicit use GOT pointer in EBX.
2360 if (!isTailCall && Subtarget->isPICStyleGOT())
2361 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2363 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2364 if (Is64Bit && isVarArg && !IsWin64)
2365 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2367 if (InFlag.getNode())
2368 Ops.push_back(InFlag);
2372 //// If this is the first return lowered for this function, add the regs
2373 //// to the liveout set for the function.
2374 // This isn't right, although it's probably harmless on x86; liveouts
2375 // should be computed from returns not tail calls. Consider a void
2376 // function making a tail call to a function returning int.
2377 return DAG.getNode(X86ISD::TC_RETURN, dl,
2378 NodeTys, &Ops[0], Ops.size());
2381 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2382 InFlag = Chain.getValue(1);
2384 // Create the CALLSEQ_END node.
2385 unsigned NumBytesForCalleeToPush;
2386 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2387 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2388 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2389 // If this is a call to a struct-return function, the callee
2390 // pops the hidden struct pointer, so we have to push it back.
2391 // This is common for Darwin/X86, Linux & Mingw32 targets.
2392 NumBytesForCalleeToPush = 4;
2394 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2396 // Returns a flag for retval copy to use.
2398 Chain = DAG.getCALLSEQ_END(Chain,
2399 DAG.getIntPtrConstant(NumBytes, true),
2400 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2403 InFlag = Chain.getValue(1);
2406 // Handle result values, copying them out of physregs into vregs that we
2408 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2409 Ins, dl, DAG, InVals);
2413 //===----------------------------------------------------------------------===//
2414 // Fast Calling Convention (tail call) implementation
2415 //===----------------------------------------------------------------------===//
2417 // Like std call, callee cleans arguments, convention except that ECX is
2418 // reserved for storing the tail called function address. Only 2 registers are
2419 // free for argument passing (inreg). Tail call optimization is performed
2421 // * tailcallopt is enabled
2422 // * caller/callee are fastcc
2423 // On X86_64 architecture with GOT-style position independent code only local
2424 // (within module) calls are supported at the moment.
2425 // To keep the stack aligned according to platform abi the function
2426 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2427 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2428 // If a tail called function callee has more arguments than the caller the
2429 // caller needs to make sure that there is room to move the RETADDR to. This is
2430 // achieved by reserving an area the size of the argument delta right after the
2431 // original REtADDR, but before the saved framepointer or the spilled registers
2432 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2444 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2445 /// for a 16 byte align requirement.
2447 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2448 SelectionDAG& DAG) const {
2449 MachineFunction &MF = DAG.getMachineFunction();
2450 const TargetMachine &TM = MF.getTarget();
2451 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2452 unsigned StackAlignment = TFI.getStackAlignment();
2453 uint64_t AlignMask = StackAlignment - 1;
2454 int64_t Offset = StackSize;
2455 uint64_t SlotSize = TD->getPointerSize();
2456 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2457 // Number smaller than 12 so just add the difference.
2458 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2460 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2461 Offset = ((~AlignMask) & Offset) + StackAlignment +
2462 (StackAlignment-SlotSize);
2467 /// MatchingStackOffset - Return true if the given stack call argument is
2468 /// already available in the same position (relatively) of the caller's
2469 /// incoming argument stack.
2471 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2472 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2473 const X86InstrInfo *TII) {
2474 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2476 if (Arg.getOpcode() == ISD::CopyFromReg) {
2477 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2478 if (!TargetRegisterInfo::isVirtualRegister(VR))
2480 MachineInstr *Def = MRI->getVRegDef(VR);
2483 if (!Flags.isByVal()) {
2484 if (!TII->isLoadFromStackSlot(Def, FI))
2487 unsigned Opcode = Def->getOpcode();
2488 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2489 Def->getOperand(1).isFI()) {
2490 FI = Def->getOperand(1).getIndex();
2491 Bytes = Flags.getByValSize();
2495 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2496 if (Flags.isByVal())
2497 // ByVal argument is passed in as a pointer but it's now being
2498 // dereferenced. e.g.
2499 // define @foo(%struct.X* %A) {
2500 // tail call @bar(%struct.X* byval %A)
2503 SDValue Ptr = Ld->getBasePtr();
2504 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2507 FI = FINode->getIndex();
2508 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2509 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Arg);
2510 FI = FINode->getIndex();
2511 Bytes = Flags.getByValSize();
2515 assert(FI != INT_MAX);
2516 if (!MFI->isFixedObjectIndex(FI))
2518 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2521 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2522 /// for tail call optimization. Targets which want to do tail call
2523 /// optimization should implement this function.
2525 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2526 CallingConv::ID CalleeCC,
2528 bool isCalleeStructRet,
2529 bool isCallerStructRet,
2530 const SmallVectorImpl<ISD::OutputArg> &Outs,
2531 const SmallVectorImpl<SDValue> &OutVals,
2532 const SmallVectorImpl<ISD::InputArg> &Ins,
2533 SelectionDAG& DAG) const {
2534 if (!IsTailCallConvention(CalleeCC) &&
2535 CalleeCC != CallingConv::C)
2538 // If -tailcallopt is specified, make fastcc functions tail-callable.
2539 const MachineFunction &MF = DAG.getMachineFunction();
2540 const Function *CallerF = DAG.getMachineFunction().getFunction();
2541 CallingConv::ID CallerCC = CallerF->getCallingConv();
2542 bool CCMatch = CallerCC == CalleeCC;
2544 if (GuaranteedTailCallOpt) {
2545 if (IsTailCallConvention(CalleeCC) && CCMatch)
2550 // Look for obvious safe cases to perform tail call optimization that do not
2551 // require ABI changes. This is what gcc calls sibcall.
2553 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2554 // emit a special epilogue.
2555 if (RegInfo->needsStackRealignment(MF))
2558 // Also avoid sibcall optimization if either caller or callee uses struct
2559 // return semantics.
2560 if (isCalleeStructRet || isCallerStructRet)
2563 // An stdcall caller is expected to clean up its arguments; the callee
2564 // isn't going to do that.
2565 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2568 // Do not sibcall optimize vararg calls unless all arguments are passed via
2570 if (isVarArg && !Outs.empty()) {
2572 // Optimizing for varargs on Win64 is unlikely to be safe without
2573 // additional testing.
2574 if (Subtarget->isTargetWin64())
2577 SmallVector<CCValAssign, 16> ArgLocs;
2578 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2579 getTargetMachine(), ArgLocs, *DAG.getContext());
2581 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2582 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2583 if (!ArgLocs[i].isRegLoc())
2587 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2588 // Therefore if it's not used by the call it is not safe to optimize this into
2590 bool Unused = false;
2591 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2598 SmallVector<CCValAssign, 16> RVLocs;
2599 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2600 getTargetMachine(), RVLocs, *DAG.getContext());
2601 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2602 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2603 CCValAssign &VA = RVLocs[i];
2604 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2609 // If the calling conventions do not match, then we'd better make sure the
2610 // results are returned in the same way as what the caller expects.
2612 SmallVector<CCValAssign, 16> RVLocs1;
2613 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2614 getTargetMachine(), RVLocs1, *DAG.getContext());
2615 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2617 SmallVector<CCValAssign, 16> RVLocs2;
2618 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2619 getTargetMachine(), RVLocs2, *DAG.getContext());
2620 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2622 if (RVLocs1.size() != RVLocs2.size())
2624 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2625 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2627 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2629 if (RVLocs1[i].isRegLoc()) {
2630 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2633 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2639 // If the callee takes no arguments then go on to check the results of the
2641 if (!Outs.empty()) {
2642 // Check if stack adjustment is needed. For now, do not do this if any
2643 // argument is passed on the stack.
2644 SmallVector<CCValAssign, 16> ArgLocs;
2645 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2646 getTargetMachine(), ArgLocs, *DAG.getContext());
2648 // Allocate shadow area for Win64
2649 if (Subtarget->isTargetWin64()) {
2650 CCInfo.AllocateStack(32, 8);
2653 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2654 if (CCInfo.getNextStackOffset()) {
2655 MachineFunction &MF = DAG.getMachineFunction();
2656 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2659 // Check if the arguments are already laid out in the right way as
2660 // the caller's fixed stack objects.
2661 MachineFrameInfo *MFI = MF.getFrameInfo();
2662 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2663 const X86InstrInfo *TII =
2664 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2665 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2666 CCValAssign &VA = ArgLocs[i];
2667 SDValue Arg = OutVals[i];
2668 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2669 if (VA.getLocInfo() == CCValAssign::Indirect)
2671 if (!VA.isRegLoc()) {
2672 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2679 // If the tailcall address may be in a register, then make sure it's
2680 // possible to register allocate for it. In 32-bit, the call address can
2681 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2682 // callee-saved registers are restored. These happen to be the same
2683 // registers used to pass 'inreg' arguments so watch out for those.
2684 if (!Subtarget->is64Bit() &&
2685 !isa<GlobalAddressSDNode>(Callee) &&
2686 !isa<ExternalSymbolSDNode>(Callee)) {
2687 unsigned NumInRegs = 0;
2688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2689 CCValAssign &VA = ArgLocs[i];
2692 unsigned Reg = VA.getLocReg();
2695 case X86::EAX: case X86::EDX: case X86::ECX:
2696 if (++NumInRegs == 3)
2708 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2709 return X86::createFastISel(funcInfo);
2713 //===----------------------------------------------------------------------===//
2714 // Other Lowering Hooks
2715 //===----------------------------------------------------------------------===//
2717 static bool MayFoldLoad(SDValue Op) {
2718 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2721 static bool MayFoldIntoStore(SDValue Op) {
2722 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2725 static bool isTargetShuffle(unsigned Opcode) {
2727 default: return false;
2728 case X86ISD::PSHUFD:
2729 case X86ISD::PSHUFHW:
2730 case X86ISD::PSHUFLW:
2731 case X86ISD::SHUFPD:
2732 case X86ISD::PALIGN:
2733 case X86ISD::SHUFPS:
2734 case X86ISD::MOVLHPS:
2735 case X86ISD::MOVLHPD:
2736 case X86ISD::MOVHLPS:
2737 case X86ISD::MOVLPS:
2738 case X86ISD::MOVLPD:
2739 case X86ISD::MOVSHDUP:
2740 case X86ISD::MOVSLDUP:
2741 case X86ISD::MOVDDUP:
2744 case X86ISD::UNPCKLPS:
2745 case X86ISD::UNPCKLPD:
2746 case X86ISD::VUNPCKLPS:
2747 case X86ISD::VUNPCKLPD:
2748 case X86ISD::VUNPCKLPSY:
2749 case X86ISD::VUNPCKLPDY:
2750 case X86ISD::PUNPCKLWD:
2751 case X86ISD::PUNPCKLBW:
2752 case X86ISD::PUNPCKLDQ:
2753 case X86ISD::PUNPCKLQDQ:
2754 case X86ISD::UNPCKHPS:
2755 case X86ISD::UNPCKHPD:
2756 case X86ISD::PUNPCKHWD:
2757 case X86ISD::PUNPCKHBW:
2758 case X86ISD::PUNPCKHDQ:
2759 case X86ISD::PUNPCKHQDQ:
2765 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2766 SDValue V1, SelectionDAG &DAG) {
2768 default: llvm_unreachable("Unknown x86 shuffle node");
2769 case X86ISD::MOVSHDUP:
2770 case X86ISD::MOVSLDUP:
2771 case X86ISD::MOVDDUP:
2772 return DAG.getNode(Opc, dl, VT, V1);
2778 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2779 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2781 default: llvm_unreachable("Unknown x86 shuffle node");
2782 case X86ISD::PSHUFD:
2783 case X86ISD::PSHUFHW:
2784 case X86ISD::PSHUFLW:
2785 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2791 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2792 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2794 default: llvm_unreachable("Unknown x86 shuffle node");
2795 case X86ISD::PALIGN:
2796 case X86ISD::SHUFPD:
2797 case X86ISD::SHUFPS:
2798 return DAG.getNode(Opc, dl, VT, V1, V2,
2799 DAG.getConstant(TargetMask, MVT::i8));
2804 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2805 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2807 default: llvm_unreachable("Unknown x86 shuffle node");
2808 case X86ISD::MOVLHPS:
2809 case X86ISD::MOVLHPD:
2810 case X86ISD::MOVHLPS:
2811 case X86ISD::MOVLPS:
2812 case X86ISD::MOVLPD:
2815 case X86ISD::UNPCKLPS:
2816 case X86ISD::UNPCKLPD:
2817 case X86ISD::VUNPCKLPS:
2818 case X86ISD::VUNPCKLPD:
2819 case X86ISD::VUNPCKLPSY:
2820 case X86ISD::VUNPCKLPDY:
2821 case X86ISD::PUNPCKLWD:
2822 case X86ISD::PUNPCKLBW:
2823 case X86ISD::PUNPCKLDQ:
2824 case X86ISD::PUNPCKLQDQ:
2825 case X86ISD::UNPCKHPS:
2826 case X86ISD::UNPCKHPD:
2827 case X86ISD::PUNPCKHWD:
2828 case X86ISD::PUNPCKHBW:
2829 case X86ISD::PUNPCKHDQ:
2830 case X86ISD::PUNPCKHQDQ:
2831 return DAG.getNode(Opc, dl, VT, V1, V2);
2836 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2837 MachineFunction &MF = DAG.getMachineFunction();
2838 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2839 int ReturnAddrIndex = FuncInfo->getRAIndex();
2841 if (ReturnAddrIndex == 0) {
2842 // Set up a frame object for the return address.
2843 uint64_t SlotSize = TD->getPointerSize();
2844 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2846 FuncInfo->setRAIndex(ReturnAddrIndex);
2849 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2853 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2854 bool hasSymbolicDisplacement) {
2855 // Offset should fit into 32 bit immediate field.
2856 if (!isInt<32>(Offset))
2859 // If we don't have a symbolic displacement - we don't have any extra
2861 if (!hasSymbolicDisplacement)
2864 // FIXME: Some tweaks might be needed for medium code model.
2865 if (M != CodeModel::Small && M != CodeModel::Kernel)
2868 // For small code model we assume that latest object is 16MB before end of 31
2869 // bits boundary. We may also accept pretty large negative constants knowing
2870 // that all objects are in the positive half of address space.
2871 if (M == CodeModel::Small && Offset < 16*1024*1024)
2874 // For kernel code model we know that all object resist in the negative half
2875 // of 32bits address space. We may not accept negative offsets, since they may
2876 // be just off and we may accept pretty large positive ones.
2877 if (M == CodeModel::Kernel && Offset > 0)
2883 /// isCalleePop - Determines whether the callee is required to pop its
2884 /// own arguments. Callee pop is necessary to support tail calls.
2885 bool X86::isCalleePop(CallingConv::ID CallingConv,
2886 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2890 switch (CallingConv) {
2893 case CallingConv::X86_StdCall:
2895 case CallingConv::X86_FastCall:
2897 case CallingConv::X86_ThisCall:
2899 case CallingConv::Fast:
2901 case CallingConv::GHC:
2906 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2907 /// specific condition code, returning the condition code and the LHS/RHS of the
2908 /// comparison to make.
2909 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2910 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2912 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2913 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2914 // X > -1 -> X == 0, jump !sign.
2915 RHS = DAG.getConstant(0, RHS.getValueType());
2916 return X86::COND_NS;
2917 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2918 // X < 0 -> X == 0, jump on sign.
2920 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2922 RHS = DAG.getConstant(0, RHS.getValueType());
2923 return X86::COND_LE;
2927 switch (SetCCOpcode) {
2928 default: llvm_unreachable("Invalid integer condition!");
2929 case ISD::SETEQ: return X86::COND_E;
2930 case ISD::SETGT: return X86::COND_G;
2931 case ISD::SETGE: return X86::COND_GE;
2932 case ISD::SETLT: return X86::COND_L;
2933 case ISD::SETLE: return X86::COND_LE;
2934 case ISD::SETNE: return X86::COND_NE;
2935 case ISD::SETULT: return X86::COND_B;
2936 case ISD::SETUGT: return X86::COND_A;
2937 case ISD::SETULE: return X86::COND_BE;
2938 case ISD::SETUGE: return X86::COND_AE;
2942 // First determine if it is required or is profitable to flip the operands.
2944 // If LHS is a foldable load, but RHS is not, flip the condition.
2945 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2946 !ISD::isNON_EXTLoad(RHS.getNode())) {
2947 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2948 std::swap(LHS, RHS);
2951 switch (SetCCOpcode) {
2957 std::swap(LHS, RHS);
2961 // On a floating point condition, the flags are set as follows:
2963 // 0 | 0 | 0 | X > Y
2964 // 0 | 0 | 1 | X < Y
2965 // 1 | 0 | 0 | X == Y
2966 // 1 | 1 | 1 | unordered
2967 switch (SetCCOpcode) {
2968 default: llvm_unreachable("Condcode should be pre-legalized away");
2970 case ISD::SETEQ: return X86::COND_E;
2971 case ISD::SETOLT: // flipped
2973 case ISD::SETGT: return X86::COND_A;
2974 case ISD::SETOLE: // flipped
2976 case ISD::SETGE: return X86::COND_AE;
2977 case ISD::SETUGT: // flipped
2979 case ISD::SETLT: return X86::COND_B;
2980 case ISD::SETUGE: // flipped
2982 case ISD::SETLE: return X86::COND_BE;
2984 case ISD::SETNE: return X86::COND_NE;
2985 case ISD::SETUO: return X86::COND_P;
2986 case ISD::SETO: return X86::COND_NP;
2988 case ISD::SETUNE: return X86::COND_INVALID;
2992 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2993 /// code. Current x86 isa includes the following FP cmov instructions:
2994 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2995 static bool hasFPCMov(unsigned X86CC) {
3011 /// isFPImmLegal - Returns true if the target can instruction select the
3012 /// specified FP immediate natively. If false, the legalizer will
3013 /// materialize the FP immediate as a load from a constant pool.
3014 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3015 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3016 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3022 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3023 /// the specified range (L, H].
3024 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3025 return (Val < 0) || (Val >= Low && Val < Hi);
3028 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3029 /// specified value.
3030 static bool isUndefOrEqual(int Val, int CmpVal) {
3031 if (Val < 0 || Val == CmpVal)
3036 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3037 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3038 /// the second operand.
3039 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3040 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3041 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3042 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3043 return (Mask[0] < 2 && Mask[1] < 2);
3047 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3048 SmallVector<int, 8> M;
3050 return ::isPSHUFDMask(M, N->getValueType(0));
3053 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3054 /// is suitable for input to PSHUFHW.
3055 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3056 if (VT != MVT::v8i16)
3059 // Lower quadword copied in order or undef.
3060 for (int i = 0; i != 4; ++i)
3061 if (Mask[i] >= 0 && Mask[i] != i)
3064 // Upper quadword shuffled.
3065 for (int i = 4; i != 8; ++i)
3066 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3072 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3073 SmallVector<int, 8> M;
3075 return ::isPSHUFHWMask(M, N->getValueType(0));
3078 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3079 /// is suitable for input to PSHUFLW.
3080 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3081 if (VT != MVT::v8i16)
3084 // Upper quadword copied in order.
3085 for (int i = 4; i != 8; ++i)
3086 if (Mask[i] >= 0 && Mask[i] != i)
3089 // Lower quadword shuffled.
3090 for (int i = 0; i != 4; ++i)
3097 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3098 SmallVector<int, 8> M;
3100 return ::isPSHUFLWMask(M, N->getValueType(0));
3103 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3104 /// is suitable for input to PALIGNR.
3105 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3107 int i, e = VT.getVectorNumElements();
3109 // Do not handle v2i64 / v2f64 shuffles with palignr.
3110 if (e < 4 || !hasSSSE3)
3113 for (i = 0; i != e; ++i)
3117 // All undef, not a palignr.
3121 // Determine if it's ok to perform a palignr with only the LHS, since we
3122 // don't have access to the actual shuffle elements to see if RHS is undef.
3123 bool Unary = Mask[i] < (int)e;
3124 bool NeedsUnary = false;
3126 int s = Mask[i] - i;
3128 // Check the rest of the elements to see if they are consecutive.
3129 for (++i; i != e; ++i) {
3134 Unary = Unary && (m < (int)e);
3135 NeedsUnary = NeedsUnary || (m < s);
3137 if (NeedsUnary && !Unary)
3139 if (Unary && m != ((s+i) & (e-1)))
3141 if (!Unary && m != (s+i))
3147 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3148 SmallVector<int, 8> M;
3150 return ::isPALIGNRMask(M, N->getValueType(0), true);
3153 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3154 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3155 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3156 int NumElems = VT.getVectorNumElements();
3157 if (NumElems != 2 && NumElems != 4)
3160 int Half = NumElems / 2;
3161 for (int i = 0; i < Half; ++i)
3162 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3164 for (int i = Half; i < NumElems; ++i)
3165 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3171 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3172 SmallVector<int, 8> M;
3174 return ::isSHUFPMask(M, N->getValueType(0));
3177 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3178 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3179 /// half elements to come from vector 1 (which would equal the dest.) and
3180 /// the upper half to come from vector 2.
3181 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3182 int NumElems = VT.getVectorNumElements();
3184 if (NumElems != 2 && NumElems != 4)
3187 int Half = NumElems / 2;
3188 for (int i = 0; i < Half; ++i)
3189 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3191 for (int i = Half; i < NumElems; ++i)
3192 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3197 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3198 SmallVector<int, 8> M;
3200 return isCommutedSHUFPMask(M, N->getValueType(0));
3203 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3204 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3205 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3206 if (N->getValueType(0).getVectorNumElements() != 4)
3209 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3210 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3211 isUndefOrEqual(N->getMaskElt(1), 7) &&
3212 isUndefOrEqual(N->getMaskElt(2), 2) &&
3213 isUndefOrEqual(N->getMaskElt(3), 3);
3216 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3217 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3219 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3220 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3225 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3226 isUndefOrEqual(N->getMaskElt(1), 3) &&
3227 isUndefOrEqual(N->getMaskElt(2), 2) &&
3228 isUndefOrEqual(N->getMaskElt(3), 3);
3231 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3232 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3233 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3234 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3236 if (NumElems != 2 && NumElems != 4)
3239 for (unsigned i = 0; i < NumElems/2; ++i)
3240 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3243 for (unsigned i = NumElems/2; i < NumElems; ++i)
3244 if (!isUndefOrEqual(N->getMaskElt(i), i))
3250 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3251 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3252 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3253 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3255 if ((NumElems != 2 && NumElems != 4)
3256 || N->getValueType(0).getSizeInBits() > 128)
3259 for (unsigned i = 0; i < NumElems/2; ++i)
3260 if (!isUndefOrEqual(N->getMaskElt(i), i))
3263 for (unsigned i = 0; i < NumElems/2; ++i)
3264 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3270 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3271 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3272 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3273 bool V2IsSplat = false) {
3274 int NumElts = VT.getVectorNumElements();
3275 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3278 // Handle vector lengths > 128 bits. Define a "section" as a set of
3279 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3281 unsigned NumSections = VT.getSizeInBits() / 128;
3282 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3283 unsigned NumSectionElts = NumElts / NumSections;
3286 unsigned End = NumSectionElts;
3287 for (unsigned s = 0; s < NumSections; ++s) {
3288 for (unsigned i = Start, j = s * NumSectionElts;
3292 int BitI1 = Mask[i+1];
3293 if (!isUndefOrEqual(BitI, j))
3296 if (!isUndefOrEqual(BitI1, NumElts))
3299 if (!isUndefOrEqual(BitI1, j + NumElts))
3303 // Process the next 128 bits.
3304 Start += NumSectionElts;
3305 End += NumSectionElts;
3311 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3312 SmallVector<int, 8> M;
3314 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3317 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3318 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3319 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3320 bool V2IsSplat = false) {
3321 int NumElts = VT.getVectorNumElements();
3322 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3325 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3327 int BitI1 = Mask[i+1];
3328 if (!isUndefOrEqual(BitI, j + NumElts/2))
3331 if (isUndefOrEqual(BitI1, NumElts))
3334 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3341 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3342 SmallVector<int, 8> M;
3344 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3347 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3348 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3350 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3351 int NumElems = VT.getVectorNumElements();
3352 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3355 // Handle vector lengths > 128 bits. Define a "section" as a set of
3356 // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
3358 unsigned NumSections = VT.getSizeInBits() / 128;
3359 if (NumSections == 0 ) NumSections = 1; // Handle MMX
3360 unsigned NumSectionElts = NumElems / NumSections;
3362 for (unsigned s = 0; s < NumSections; ++s) {
3363 for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
3364 i != NumSectionElts * (s + 1);
3367 int BitI1 = Mask[i+1];
3369 if (!isUndefOrEqual(BitI, j))
3371 if (!isUndefOrEqual(BitI1, j))
3379 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3380 SmallVector<int, 8> M;
3382 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3385 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3386 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3388 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3389 int NumElems = VT.getVectorNumElements();
3390 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3393 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3395 int BitI1 = Mask[i+1];
3396 if (!isUndefOrEqual(BitI, j))
3398 if (!isUndefOrEqual(BitI1, j))
3404 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3405 SmallVector<int, 8> M;
3407 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3410 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3411 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3412 /// MOVSD, and MOVD, i.e. setting the lowest element.
3413 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3414 if (VT.getVectorElementType().getSizeInBits() < 32)
3417 int NumElts = VT.getVectorNumElements();
3419 if (!isUndefOrEqual(Mask[0], NumElts))
3422 for (int i = 1; i < NumElts; ++i)
3423 if (!isUndefOrEqual(Mask[i], i))
3429 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3430 SmallVector<int, 8> M;
3432 return ::isMOVLMask(M, N->getValueType(0));
3435 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3436 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3437 /// element of vector 2 and the other elements to come from vector 1 in order.
3438 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3439 bool V2IsSplat = false, bool V2IsUndef = false) {
3440 int NumOps = VT.getVectorNumElements();
3441 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3444 if (!isUndefOrEqual(Mask[0], 0))
3447 for (int i = 1; i < NumOps; ++i)
3448 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3449 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3450 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3456 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3457 bool V2IsUndef = false) {
3458 SmallVector<int, 8> M;
3460 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3463 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3464 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3465 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3466 if (N->getValueType(0).getVectorNumElements() != 4)
3469 // Expect 1, 1, 3, 3
3470 for (unsigned i = 0; i < 2; ++i) {
3471 int Elt = N->getMaskElt(i);
3472 if (Elt >= 0 && Elt != 1)
3477 for (unsigned i = 2; i < 4; ++i) {
3478 int Elt = N->getMaskElt(i);
3479 if (Elt >= 0 && Elt != 3)
3484 // Don't use movshdup if it can be done with a shufps.
3485 // FIXME: verify that matching u, u, 3, 3 is what we want.
3489 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3490 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3491 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3492 if (N->getValueType(0).getVectorNumElements() != 4)
3495 // Expect 0, 0, 2, 2
3496 for (unsigned i = 0; i < 2; ++i)
3497 if (N->getMaskElt(i) > 0)
3501 for (unsigned i = 2; i < 4; ++i) {
3502 int Elt = N->getMaskElt(i);
3503 if (Elt >= 0 && Elt != 2)
3508 // Don't use movsldup if it can be done with a shufps.
3512 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3513 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3514 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3515 int e = N->getValueType(0).getVectorNumElements() / 2;
3517 for (int i = 0; i < e; ++i)
3518 if (!isUndefOrEqual(N->getMaskElt(i), i))
3520 for (int i = 0; i < e; ++i)
3521 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3526 /// isVEXTRACTF128Index - Return true if the specified
3527 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3528 /// suitable for input to VEXTRACTF128.
3529 bool X86::isVEXTRACTF128Index(SDNode *N) {
3530 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3533 // The index should be aligned on a 128-bit boundary.
3535 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3537 unsigned VL = N->getValueType(0).getVectorNumElements();
3538 unsigned VBits = N->getValueType(0).getSizeInBits();
3539 unsigned ElSize = VBits / VL;
3540 bool Result = (Index * ElSize) % 128 == 0;
3545 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3546 /// operand specifies a subvector insert that is suitable for input to
3548 bool X86::isVINSERTF128Index(SDNode *N) {
3549 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3552 // The index should be aligned on a 128-bit boundary.
3554 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3556 unsigned VL = N->getValueType(0).getVectorNumElements();
3557 unsigned VBits = N->getValueType(0).getSizeInBits();
3558 unsigned ElSize = VBits / VL;
3559 bool Result = (Index * ElSize) % 128 == 0;
3564 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3565 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3566 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3567 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3568 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3570 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3572 for (int i = 0; i < NumOperands; ++i) {
3573 int Val = SVOp->getMaskElt(NumOperands-i-1);
3574 if (Val < 0) Val = 0;
3575 if (Val >= NumOperands) Val -= NumOperands;
3577 if (i != NumOperands - 1)
3583 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3584 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3585 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3586 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3588 // 8 nodes, but we only care about the last 4.
3589 for (unsigned i = 7; i >= 4; --i) {
3590 int Val = SVOp->getMaskElt(i);
3599 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3600 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3601 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3602 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3604 // 8 nodes, but we only care about the first 4.
3605 for (int i = 3; i >= 0; --i) {
3606 int Val = SVOp->getMaskElt(i);
3615 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3616 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3617 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3618 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3619 EVT VVT = N->getValueType(0);
3620 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3624 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3625 Val = SVOp->getMaskElt(i);
3629 return (Val - i) * EltSize;
3632 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3633 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3635 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3636 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3637 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3640 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3642 EVT VecVT = N->getOperand(0).getValueType();
3643 EVT ElVT = VecVT.getVectorElementType();
3645 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3647 return Index / NumElemsPerChunk;
3650 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3651 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3653 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3654 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3655 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3658 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3660 EVT VecVT = N->getValueType(0);
3661 EVT ElVT = VecVT.getVectorElementType();
3663 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3665 return Index / NumElemsPerChunk;
3668 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3670 bool X86::isZeroNode(SDValue Elt) {
3671 return ((isa<ConstantSDNode>(Elt) &&
3672 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3673 (isa<ConstantFPSDNode>(Elt) &&
3674 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3677 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3678 /// their permute mask.
3679 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3680 SelectionDAG &DAG) {
3681 EVT VT = SVOp->getValueType(0);
3682 unsigned NumElems = VT.getVectorNumElements();
3683 SmallVector<int, 8> MaskVec;
3685 for (unsigned i = 0; i != NumElems; ++i) {
3686 int idx = SVOp->getMaskElt(i);
3688 MaskVec.push_back(idx);
3689 else if (idx < (int)NumElems)
3690 MaskVec.push_back(idx + NumElems);
3692 MaskVec.push_back(idx - NumElems);
3694 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3695 SVOp->getOperand(0), &MaskVec[0]);
3698 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3699 /// the two vector operands have swapped position.
3700 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3701 unsigned NumElems = VT.getVectorNumElements();
3702 for (unsigned i = 0; i != NumElems; ++i) {
3706 else if (idx < (int)NumElems)
3707 Mask[i] = idx + NumElems;
3709 Mask[i] = idx - NumElems;
3713 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3714 /// match movhlps. The lower half elements should come from upper half of
3715 /// V1 (and in order), and the upper half elements should come from the upper
3716 /// half of V2 (and in order).
3717 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3718 if (Op->getValueType(0).getVectorNumElements() != 4)
3720 for (unsigned i = 0, e = 2; i != e; ++i)
3721 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3723 for (unsigned i = 2; i != 4; ++i)
3724 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3729 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3730 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3732 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3733 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3735 N = N->getOperand(0).getNode();
3736 if (!ISD::isNON_EXTLoad(N))
3739 *LD = cast<LoadSDNode>(N);
3743 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3744 /// match movlp{s|d}. The lower half elements should come from lower half of
3745 /// V1 (and in order), and the upper half elements should come from the upper
3746 /// half of V2 (and in order). And since V1 will become the source of the
3747 /// MOVLP, it must be either a vector load or a scalar load to vector.
3748 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3749 ShuffleVectorSDNode *Op) {
3750 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3752 // Is V2 is a vector load, don't do this transformation. We will try to use
3753 // load folding shufps op.
3754 if (ISD::isNON_EXTLoad(V2))
3757 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3759 if (NumElems != 2 && NumElems != 4)
3761 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3762 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3764 for (unsigned i = NumElems/2; i != NumElems; ++i)
3765 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3770 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3772 static bool isSplatVector(SDNode *N) {
3773 if (N->getOpcode() != ISD::BUILD_VECTOR)
3776 SDValue SplatValue = N->getOperand(0);
3777 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3778 if (N->getOperand(i) != SplatValue)
3783 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3784 /// to an zero vector.
3785 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3786 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3787 SDValue V1 = N->getOperand(0);
3788 SDValue V2 = N->getOperand(1);
3789 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3790 for (unsigned i = 0; i != NumElems; ++i) {
3791 int Idx = N->getMaskElt(i);
3792 if (Idx >= (int)NumElems) {
3793 unsigned Opc = V2.getOpcode();
3794 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3796 if (Opc != ISD::BUILD_VECTOR ||
3797 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3799 } else if (Idx >= 0) {
3800 unsigned Opc = V1.getOpcode();
3801 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3803 if (Opc != ISD::BUILD_VECTOR ||
3804 !X86::isZeroNode(V1.getOperand(Idx)))
3811 /// getZeroVector - Returns a vector of specified type with all zero elements.
3813 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3815 assert(VT.isVector() && "Expected a vector type");
3817 // Always build SSE zero vectors as <4 x i32> bitcasted
3818 // to their dest type. This ensures they get CSE'd.
3820 if (VT.getSizeInBits() == 128) { // SSE
3821 if (HasSSE2) { // SSE2
3822 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3823 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3825 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3826 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3828 } else if (VT.getSizeInBits() == 256) { // AVX
3829 // 256-bit logic and arithmetic instructions in AVX are
3830 // all floating-point, no support for integer ops. Default
3831 // to emitting fp zeroed vectors then.
3832 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3833 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3834 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3836 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3839 /// getOnesVector - Returns a vector of specified type with all bits set.
3841 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3842 assert(VT.isVector() && "Expected a vector type");
3844 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3845 // type. This ensures they get CSE'd.
3846 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3848 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3849 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3853 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3854 /// that point to V2 points to its first element.
3855 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3856 EVT VT = SVOp->getValueType(0);
3857 unsigned NumElems = VT.getVectorNumElements();
3859 bool Changed = false;
3860 SmallVector<int, 8> MaskVec;
3861 SVOp->getMask(MaskVec);
3863 for (unsigned i = 0; i != NumElems; ++i) {
3864 if (MaskVec[i] > (int)NumElems) {
3865 MaskVec[i] = NumElems;
3870 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3871 SVOp->getOperand(1), &MaskVec[0]);
3872 return SDValue(SVOp, 0);
3875 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3876 /// operation of specified width.
3877 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3879 unsigned NumElems = VT.getVectorNumElements();
3880 SmallVector<int, 8> Mask;
3881 Mask.push_back(NumElems);
3882 for (unsigned i = 1; i != NumElems; ++i)
3884 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3887 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3888 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3890 unsigned NumElems = VT.getVectorNumElements();
3891 SmallVector<int, 8> Mask;
3892 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3894 Mask.push_back(i + NumElems);
3896 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3899 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3900 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3902 unsigned NumElems = VT.getVectorNumElements();
3903 unsigned Half = NumElems/2;
3904 SmallVector<int, 8> Mask;
3905 for (unsigned i = 0; i != Half; ++i) {
3906 Mask.push_back(i + Half);
3907 Mask.push_back(i + NumElems + Half);
3909 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3912 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3913 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3914 EVT PVT = MVT::v4f32;
3915 EVT VT = SV->getValueType(0);
3916 DebugLoc dl = SV->getDebugLoc();
3917 SDValue V1 = SV->getOperand(0);
3918 int NumElems = VT.getVectorNumElements();
3919 int EltNo = SV->getSplatIndex();
3921 // unpack elements to the correct location
3922 while (NumElems > 4) {
3923 if (EltNo < NumElems/2) {
3924 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3926 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3927 EltNo -= NumElems/2;
3932 // Perform the splat.
3933 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3934 V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
3935 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3936 return DAG.getNode(ISD::BITCAST, dl, VT, V1);
3939 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3940 /// vector of zero or undef vector. This produces a shuffle where the low
3941 /// element of V2 is swizzled into the zero/undef vector, landing at element
3942 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3943 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3944 bool isZero, bool HasSSE2,
3945 SelectionDAG &DAG) {
3946 EVT VT = V2.getValueType();
3948 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3949 unsigned NumElems = VT.getVectorNumElements();
3950 SmallVector<int, 16> MaskVec;
3951 for (unsigned i = 0; i != NumElems; ++i)
3952 // If this is the insertion idx, put the low elt of V2 here.
3953 MaskVec.push_back(i == Idx ? NumElems : i);
3954 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3957 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3958 /// element of the result of the vector shuffle.
3959 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3962 return SDValue(); // Limit search depth.
3964 SDValue V = SDValue(N, 0);
3965 EVT VT = V.getValueType();
3966 unsigned Opcode = V.getOpcode();
3968 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3969 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3970 Index = SV->getMaskElt(Index);
3973 return DAG.getUNDEF(VT.getVectorElementType());
3975 int NumElems = VT.getVectorNumElements();
3976 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3977 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3980 // Recurse into target specific vector shuffles to find scalars.
3981 if (isTargetShuffle(Opcode)) {
3982 int NumElems = VT.getVectorNumElements();
3983 SmallVector<unsigned, 16> ShuffleMask;
3987 case X86ISD::SHUFPS:
3988 case X86ISD::SHUFPD:
3989 ImmN = N->getOperand(N->getNumOperands()-1);
3990 DecodeSHUFPSMask(NumElems,
3991 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3994 case X86ISD::PUNPCKHBW:
3995 case X86ISD::PUNPCKHWD:
3996 case X86ISD::PUNPCKHDQ:
3997 case X86ISD::PUNPCKHQDQ:
3998 DecodePUNPCKHMask(NumElems, ShuffleMask);
4000 case X86ISD::UNPCKHPS:
4001 case X86ISD::UNPCKHPD:
4002 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4004 case X86ISD::PUNPCKLBW:
4005 case X86ISD::PUNPCKLWD:
4006 case X86ISD::PUNPCKLDQ:
4007 case X86ISD::PUNPCKLQDQ:
4008 DecodePUNPCKLMask(VT, ShuffleMask);
4010 case X86ISD::UNPCKLPS:
4011 case X86ISD::UNPCKLPD:
4012 case X86ISD::VUNPCKLPS:
4013 case X86ISD::VUNPCKLPD:
4014 case X86ISD::VUNPCKLPSY:
4015 case X86ISD::VUNPCKLPDY:
4016 DecodeUNPCKLPMask(VT, ShuffleMask);
4018 case X86ISD::MOVHLPS:
4019 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4021 case X86ISD::MOVLHPS:
4022 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4024 case X86ISD::PSHUFD:
4025 ImmN = N->getOperand(N->getNumOperands()-1);
4026 DecodePSHUFMask(NumElems,
4027 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4030 case X86ISD::PSHUFHW:
4031 ImmN = N->getOperand(N->getNumOperands()-1);
4032 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4035 case X86ISD::PSHUFLW:
4036 ImmN = N->getOperand(N->getNumOperands()-1);
4037 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4041 case X86ISD::MOVSD: {
4042 // The index 0 always comes from the first element of the second source,
4043 // this is why MOVSS and MOVSD are used in the first place. The other
4044 // elements come from the other positions of the first source vector.
4045 unsigned OpNum = (Index == 0) ? 1 : 0;
4046 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4050 assert("not implemented for target shuffle node");
4054 Index = ShuffleMask[Index];
4056 return DAG.getUNDEF(VT.getVectorElementType());
4058 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4059 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4063 // Actual nodes that may contain scalar elements
4064 if (Opcode == ISD::BITCAST) {
4065 V = V.getOperand(0);
4066 EVT SrcVT = V.getValueType();
4067 unsigned NumElems = VT.getVectorNumElements();
4069 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4073 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4074 return (Index == 0) ? V.getOperand(0)
4075 : DAG.getUNDEF(VT.getVectorElementType());
4077 if (V.getOpcode() == ISD::BUILD_VECTOR)
4078 return V.getOperand(Index);
4083 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4084 /// shuffle operation which come from a consecutively from a zero. The
4085 /// search can start in two different directions, from left or right.
4087 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4088 bool ZerosFromLeft, SelectionDAG &DAG) {
4091 while (i < NumElems) {
4092 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4093 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4094 if (!(Elt.getNode() &&
4095 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4103 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4104 /// MaskE correspond consecutively to elements from one of the vector operands,
4105 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4107 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4108 int OpIdx, int NumElems, unsigned &OpNum) {
4109 bool SeenV1 = false;
4110 bool SeenV2 = false;
4112 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4113 int Idx = SVOp->getMaskElt(i);
4114 // Ignore undef indicies
4123 // Only accept consecutive elements from the same vector
4124 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4128 OpNum = SeenV1 ? 0 : 1;
4132 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4133 /// logical left shift of a vector.
4134 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4135 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4136 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4137 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4138 false /* check zeros from right */, DAG);
4144 // Considering the elements in the mask that are not consecutive zeros,
4145 // check if they consecutively come from only one of the source vectors.
4147 // V1 = {X, A, B, C} 0
4149 // vector_shuffle V1, V2 <1, 2, 3, X>
4151 if (!isShuffleMaskConsecutive(SVOp,
4152 0, // Mask Start Index
4153 NumElems-NumZeros-1, // Mask End Index
4154 NumZeros, // Where to start looking in the src vector
4155 NumElems, // Number of elements in vector
4156 OpSrc)) // Which source operand ?
4161 ShVal = SVOp->getOperand(OpSrc);
4165 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4166 /// logical left shift of a vector.
4167 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4168 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4169 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4170 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4171 true /* check zeros from left */, DAG);
4177 // Considering the elements in the mask that are not consecutive zeros,
4178 // check if they consecutively come from only one of the source vectors.
4180 // 0 { A, B, X, X } = V2
4182 // vector_shuffle V1, V2 <X, X, 4, 5>
4184 if (!isShuffleMaskConsecutive(SVOp,
4185 NumZeros, // Mask Start Index
4186 NumElems-1, // Mask End Index
4187 0, // Where to start looking in the src vector
4188 NumElems, // Number of elements in vector
4189 OpSrc)) // Which source operand ?
4194 ShVal = SVOp->getOperand(OpSrc);
4198 /// isVectorShift - Returns true if the shuffle can be implemented as a
4199 /// logical left or right shift of a vector.
4200 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4201 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4202 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4203 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4209 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4211 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4212 unsigned NumNonZero, unsigned NumZero,
4214 const TargetLowering &TLI) {
4218 DebugLoc dl = Op.getDebugLoc();
4221 for (unsigned i = 0; i < 16; ++i) {
4222 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4223 if (ThisIsNonZero && First) {
4225 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4227 V = DAG.getUNDEF(MVT::v8i16);
4232 SDValue ThisElt(0, 0), LastElt(0, 0);
4233 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4234 if (LastIsNonZero) {
4235 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4236 MVT::i16, Op.getOperand(i-1));
4238 if (ThisIsNonZero) {
4239 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4240 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4241 ThisElt, DAG.getConstant(8, MVT::i8));
4243 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4247 if (ThisElt.getNode())
4248 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4249 DAG.getIntPtrConstant(i/2));
4253 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4256 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4258 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4259 unsigned NumNonZero, unsigned NumZero,
4261 const TargetLowering &TLI) {
4265 DebugLoc dl = Op.getDebugLoc();
4268 for (unsigned i = 0; i < 8; ++i) {
4269 bool isNonZero = (NonZeros & (1 << i)) != 0;
4273 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4275 V = DAG.getUNDEF(MVT::v8i16);
4278 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4279 MVT::v8i16, V, Op.getOperand(i),
4280 DAG.getIntPtrConstant(i));
4287 /// getVShift - Return a vector logical shift node.
4289 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4290 unsigned NumBits, SelectionDAG &DAG,
4291 const TargetLowering &TLI, DebugLoc dl) {
4292 EVT ShVT = MVT::v2i64;
4293 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4294 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4295 return DAG.getNode(ISD::BITCAST, dl, VT,
4296 DAG.getNode(Opc, dl, ShVT, SrcOp,
4297 DAG.getConstant(NumBits,
4298 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4302 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4303 SelectionDAG &DAG) const {
4305 // Check if the scalar load can be widened into a vector load. And if
4306 // the address is "base + cst" see if the cst can be "absorbed" into
4307 // the shuffle mask.
4308 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4309 SDValue Ptr = LD->getBasePtr();
4310 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4312 EVT PVT = LD->getValueType(0);
4313 if (PVT != MVT::i32 && PVT != MVT::f32)
4318 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4319 FI = FINode->getIndex();
4321 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4322 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4323 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4324 Offset = Ptr.getConstantOperandVal(1);
4325 Ptr = Ptr.getOperand(0);
4330 SDValue Chain = LD->getChain();
4331 // Make sure the stack object alignment is at least 16.
4332 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4333 if (DAG.InferPtrAlignment(Ptr) < 16) {
4334 if (MFI->isFixedObjectIndex(FI)) {
4335 // Can't change the alignment. FIXME: It's possible to compute
4336 // the exact stack offset and reference FI + adjust offset instead.
4337 // If someone *really* cares about this. That's the way to implement it.
4340 MFI->setObjectAlignment(FI, 16);
4344 // (Offset % 16) must be multiple of 4. Then address is then
4345 // Ptr + (Offset & ~15).
4348 if ((Offset % 16) & 3)
4350 int64_t StartOffset = Offset & ~15;
4352 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4353 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4355 int EltNo = (Offset - StartOffset) >> 2;
4356 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4357 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4358 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4359 LD->getPointerInfo().getWithOffset(StartOffset),
4361 // Canonicalize it to a v4i32 shuffle.
4362 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4363 return DAG.getNode(ISD::BITCAST, dl, VT,
4364 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4365 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4371 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4372 /// vector of type 'VT', see if the elements can be replaced by a single large
4373 /// load which has the same value as a build_vector whose operands are 'elts'.
4375 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4377 /// FIXME: we'd also like to handle the case where the last elements are zero
4378 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4379 /// There's even a handy isZeroNode for that purpose.
4380 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4381 DebugLoc &DL, SelectionDAG &DAG) {
4382 EVT EltVT = VT.getVectorElementType();
4383 unsigned NumElems = Elts.size();
4385 LoadSDNode *LDBase = NULL;
4386 unsigned LastLoadedElt = -1U;
4388 // For each element in the initializer, see if we've found a load or an undef.
4389 // If we don't find an initial load element, or later load elements are
4390 // non-consecutive, bail out.
4391 for (unsigned i = 0; i < NumElems; ++i) {
4392 SDValue Elt = Elts[i];
4394 if (!Elt.getNode() ||
4395 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4398 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4400 LDBase = cast<LoadSDNode>(Elt.getNode());
4404 if (Elt.getOpcode() == ISD::UNDEF)
4407 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4408 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4413 // If we have found an entire vector of loads and undefs, then return a large
4414 // load of the entire vector width starting at the base pointer. If we found
4415 // consecutive loads for the low half, generate a vzext_load node.
4416 if (LastLoadedElt == NumElems - 1) {
4417 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4418 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4419 LDBase->getPointerInfo(),
4420 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4421 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4422 LDBase->getPointerInfo(),
4423 LDBase->isVolatile(), LDBase->isNonTemporal(),
4424 LDBase->getAlignment());
4425 } else if (NumElems == 4 && LastLoadedElt == 1) {
4426 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4427 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4428 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4430 LDBase->getMemOperand());
4431 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4437 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4438 DebugLoc dl = Op.getDebugLoc();
4440 EVT VT = Op.getValueType();
4441 EVT ExtVT = VT.getVectorElementType();
4443 unsigned NumElems = Op.getNumOperands();
4445 // For AVX-length vectors, build the individual 128-bit pieces and
4446 // use shuffles to put them in place.
4447 if (VT.getSizeInBits() > 256 &&
4448 Subtarget->hasAVX() &&
4449 !ISD::isBuildVectorAllZeros(Op.getNode())) {
4450 SmallVector<SDValue, 8> V;
4452 for (unsigned i = 0; i < NumElems; ++i) {
4453 V[i] = Op.getOperand(i);
4456 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4458 // Build the lower subvector.
4459 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4460 // Build the upper subvector.
4461 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4464 return ConcatVectors(Lower, Upper, DAG);
4467 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4468 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4469 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4470 // is present, so AllOnes is ignored.
4471 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4472 (Op.getValueType().getSizeInBits() != 256 &&
4473 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4474 // Canonicalize this to <4 x i32> (SSE) to
4475 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4476 // eliminated on x86-32 hosts.
4477 if (Op.getValueType() == MVT::v4i32)
4480 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4481 return getOnesVector(Op.getValueType(), DAG, dl);
4482 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4485 unsigned EVTBits = ExtVT.getSizeInBits();
4487 unsigned NumZero = 0;
4488 unsigned NumNonZero = 0;
4489 unsigned NonZeros = 0;
4490 bool IsAllConstants = true;
4491 SmallSet<SDValue, 8> Values;
4492 for (unsigned i = 0; i < NumElems; ++i) {
4493 SDValue Elt = Op.getOperand(i);
4494 if (Elt.getOpcode() == ISD::UNDEF)
4497 if (Elt.getOpcode() != ISD::Constant &&
4498 Elt.getOpcode() != ISD::ConstantFP)
4499 IsAllConstants = false;
4500 if (X86::isZeroNode(Elt))
4503 NonZeros |= (1 << i);
4508 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4509 if (NumNonZero == 0)
4510 return DAG.getUNDEF(VT);
4512 // Special case for single non-zero, non-undef, element.
4513 if (NumNonZero == 1) {
4514 unsigned Idx = CountTrailingZeros_32(NonZeros);
4515 SDValue Item = Op.getOperand(Idx);
4517 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4518 // the value are obviously zero, truncate the value to i32 and do the
4519 // insertion that way. Only do this if the value is non-constant or if the
4520 // value is a constant being inserted into element 0. It is cheaper to do
4521 // a constant pool load than it is to do a movd + shuffle.
4522 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4523 (!IsAllConstants || Idx == 0)) {
4524 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4526 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4527 EVT VecVT = MVT::v4i32;
4528 unsigned VecElts = 4;
4530 // Truncate the value (which may itself be a constant) to i32, and
4531 // convert it to a vector with movd (S2V+shuffle to zero extend).
4532 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4533 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4534 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4535 Subtarget->hasSSE2(), DAG);
4537 // Now we have our 32-bit value zero extended in the low element of
4538 // a vector. If Idx != 0, swizzle it into place.
4540 SmallVector<int, 4> Mask;
4541 Mask.push_back(Idx);
4542 for (unsigned i = 1; i != VecElts; ++i)
4544 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4545 DAG.getUNDEF(Item.getValueType()),
4548 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4552 // If we have a constant or non-constant insertion into the low element of
4553 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4554 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4555 // depending on what the source datatype is.
4558 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4559 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4560 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4561 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4562 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4563 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4565 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4566 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4567 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4568 EVT MiddleVT = MVT::v4i32;
4569 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4570 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4571 Subtarget->hasSSE2(), DAG);
4572 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4576 // Is it a vector logical left shift?
4577 if (NumElems == 2 && Idx == 1 &&
4578 X86::isZeroNode(Op.getOperand(0)) &&
4579 !X86::isZeroNode(Op.getOperand(1))) {
4580 unsigned NumBits = VT.getSizeInBits();
4581 return getVShift(true, VT,
4582 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4583 VT, Op.getOperand(1)),
4584 NumBits/2, DAG, *this, dl);
4587 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4590 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4591 // is a non-constant being inserted into an element other than the low one,
4592 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4593 // movd/movss) to move this into the low element, then shuffle it into
4595 if (EVTBits == 32) {
4596 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4598 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4599 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4600 Subtarget->hasSSE2(), DAG);
4601 SmallVector<int, 8> MaskVec;
4602 for (unsigned i = 0; i < NumElems; i++)
4603 MaskVec.push_back(i == Idx ? 0 : 1);
4604 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4608 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4609 if (Values.size() == 1) {
4610 if (EVTBits == 32) {
4611 // Instead of a shuffle like this:
4612 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4613 // Check if it's possible to issue this instead.
4614 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4615 unsigned Idx = CountTrailingZeros_32(NonZeros);
4616 SDValue Item = Op.getOperand(Idx);
4617 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4618 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4623 // A vector full of immediates; various special cases are already
4624 // handled, so this is best done with a single constant-pool load.
4628 // Let legalizer expand 2-wide build_vectors.
4629 if (EVTBits == 64) {
4630 if (NumNonZero == 1) {
4631 // One half is zero or undef.
4632 unsigned Idx = CountTrailingZeros_32(NonZeros);
4633 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4634 Op.getOperand(Idx));
4635 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4636 Subtarget->hasSSE2(), DAG);
4641 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4642 if (EVTBits == 8 && NumElems == 16) {
4643 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4645 if (V.getNode()) return V;
4648 if (EVTBits == 16 && NumElems == 8) {
4649 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4651 if (V.getNode()) return V;
4654 // If element VT is == 32 bits, turn it into a number of shuffles.
4655 SmallVector<SDValue, 8> V;
4657 if (NumElems == 4 && NumZero > 0) {
4658 for (unsigned i = 0; i < 4; ++i) {
4659 bool isZero = !(NonZeros & (1 << i));
4661 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4663 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4666 for (unsigned i = 0; i < 2; ++i) {
4667 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4670 V[i] = V[i*2]; // Must be a zero vector.
4673 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4676 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4679 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4684 SmallVector<int, 8> MaskVec;
4685 bool Reverse = (NonZeros & 0x3) == 2;
4686 for (unsigned i = 0; i < 2; ++i)
4687 MaskVec.push_back(Reverse ? 1-i : i);
4688 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4689 for (unsigned i = 0; i < 2; ++i)
4690 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4691 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4694 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4695 // Check for a build vector of consecutive loads.
4696 for (unsigned i = 0; i < NumElems; ++i)
4697 V[i] = Op.getOperand(i);
4699 // Check for elements which are consecutive loads.
4700 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4704 // For SSE 4.1, use insertps to put the high elements into the low element.
4705 if (getSubtarget()->hasSSE41()) {
4707 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4708 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4710 Result = DAG.getUNDEF(VT);
4712 for (unsigned i = 1; i < NumElems; ++i) {
4713 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4714 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4715 Op.getOperand(i), DAG.getIntPtrConstant(i));
4720 // Otherwise, expand into a number of unpckl*, start by extending each of
4721 // our (non-undef) elements to the full vector width with the element in the
4722 // bottom slot of the vector (which generates no code for SSE).
4723 for (unsigned i = 0; i < NumElems; ++i) {
4724 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4725 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4727 V[i] = DAG.getUNDEF(VT);
4730 // Next, we iteratively mix elements, e.g. for v4f32:
4731 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4732 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4733 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4734 unsigned EltStride = NumElems >> 1;
4735 while (EltStride != 0) {
4736 for (unsigned i = 0; i < EltStride; ++i) {
4737 // If V[i+EltStride] is undef and this is the first round of mixing,
4738 // then it is safe to just drop this shuffle: V[i] is already in the
4739 // right place, the one element (since it's the first round) being
4740 // inserted as undef can be dropped. This isn't safe for successive
4741 // rounds because they will permute elements within both vectors.
4742 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4743 EltStride == NumElems/2)
4746 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4756 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4757 // We support concatenate two MMX registers and place them in a MMX
4758 // register. This is better than doing a stack convert.
4759 DebugLoc dl = Op.getDebugLoc();
4760 EVT ResVT = Op.getValueType();
4761 assert(Op.getNumOperands() == 2);
4762 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4763 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4765 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4766 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4767 InVec = Op.getOperand(1);
4768 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4769 unsigned NumElts = ResVT.getVectorNumElements();
4770 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4771 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4772 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4774 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4775 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4776 Mask[0] = 0; Mask[1] = 2;
4777 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4779 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4782 // v8i16 shuffles - Prefer shuffles in the following order:
4783 // 1. [all] pshuflw, pshufhw, optional move
4784 // 2. [ssse3] 1 x pshufb
4785 // 3. [ssse3] 2 x pshufb + 1 x por
4786 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4788 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4789 SelectionDAG &DAG) const {
4790 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4791 SDValue V1 = SVOp->getOperand(0);
4792 SDValue V2 = SVOp->getOperand(1);
4793 DebugLoc dl = SVOp->getDebugLoc();
4794 SmallVector<int, 8> MaskVals;
4796 // Determine if more than 1 of the words in each of the low and high quadwords
4797 // of the result come from the same quadword of one of the two inputs. Undef
4798 // mask values count as coming from any quadword, for better codegen.
4799 SmallVector<unsigned, 4> LoQuad(4);
4800 SmallVector<unsigned, 4> HiQuad(4);
4801 BitVector InputQuads(4);
4802 for (unsigned i = 0; i < 8; ++i) {
4803 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4804 int EltIdx = SVOp->getMaskElt(i);
4805 MaskVals.push_back(EltIdx);
4814 InputQuads.set(EltIdx / 4);
4817 int BestLoQuad = -1;
4818 unsigned MaxQuad = 1;
4819 for (unsigned i = 0; i < 4; ++i) {
4820 if (LoQuad[i] > MaxQuad) {
4822 MaxQuad = LoQuad[i];
4826 int BestHiQuad = -1;
4828 for (unsigned i = 0; i < 4; ++i) {
4829 if (HiQuad[i] > MaxQuad) {
4831 MaxQuad = HiQuad[i];
4835 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4836 // of the two input vectors, shuffle them into one input vector so only a
4837 // single pshufb instruction is necessary. If There are more than 2 input
4838 // quads, disable the next transformation since it does not help SSSE3.
4839 bool V1Used = InputQuads[0] || InputQuads[1];
4840 bool V2Used = InputQuads[2] || InputQuads[3];
4841 if (Subtarget->hasSSSE3()) {
4842 if (InputQuads.count() == 2 && V1Used && V2Used) {
4843 BestLoQuad = InputQuads.find_first();
4844 BestHiQuad = InputQuads.find_next(BestLoQuad);
4846 if (InputQuads.count() > 2) {
4852 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4853 // the shuffle mask. If a quad is scored as -1, that means that it contains
4854 // words from all 4 input quadwords.
4856 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4857 SmallVector<int, 8> MaskV;
4858 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4859 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4860 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4861 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4862 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4863 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4865 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4866 // source words for the shuffle, to aid later transformations.
4867 bool AllWordsInNewV = true;
4868 bool InOrder[2] = { true, true };
4869 for (unsigned i = 0; i != 8; ++i) {
4870 int idx = MaskVals[i];
4872 InOrder[i/4] = false;
4873 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4875 AllWordsInNewV = false;
4879 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4880 if (AllWordsInNewV) {
4881 for (int i = 0; i != 8; ++i) {
4882 int idx = MaskVals[i];
4885 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4886 if ((idx != i) && idx < 4)
4888 if ((idx != i) && idx > 3)
4897 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4898 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4899 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4900 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4901 unsigned TargetMask = 0;
4902 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4903 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4904 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4905 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4906 V1 = NewV.getOperand(0);
4907 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4911 // If we have SSSE3, and all words of the result are from 1 input vector,
4912 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4913 // is present, fall back to case 4.
4914 if (Subtarget->hasSSSE3()) {
4915 SmallVector<SDValue,16> pshufbMask;
4917 // If we have elements from both input vectors, set the high bit of the
4918 // shuffle mask element to zero out elements that come from V2 in the V1
4919 // mask, and elements that come from V1 in the V2 mask, so that the two
4920 // results can be OR'd together.
4921 bool TwoInputs = V1Used && V2Used;
4922 for (unsigned i = 0; i != 8; ++i) {
4923 int EltIdx = MaskVals[i] * 2;
4924 if (TwoInputs && (EltIdx >= 16)) {
4925 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4926 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4929 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4930 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4932 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4933 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4934 DAG.getNode(ISD::BUILD_VECTOR, dl,
4935 MVT::v16i8, &pshufbMask[0], 16));
4937 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4939 // Calculate the shuffle mask for the second input, shuffle it, and
4940 // OR it with the first shuffled input.
4942 for (unsigned i = 0; i != 8; ++i) {
4943 int EltIdx = MaskVals[i] * 2;
4945 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4946 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4949 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4950 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4952 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
4953 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4954 DAG.getNode(ISD::BUILD_VECTOR, dl,
4955 MVT::v16i8, &pshufbMask[0], 16));
4956 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4957 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4960 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4961 // and update MaskVals with new element order.
4962 BitVector InOrder(8);
4963 if (BestLoQuad >= 0) {
4964 SmallVector<int, 8> MaskV;
4965 for (int i = 0; i != 4; ++i) {
4966 int idx = MaskVals[i];
4968 MaskV.push_back(-1);
4970 } else if ((idx / 4) == BestLoQuad) {
4971 MaskV.push_back(idx & 3);
4974 MaskV.push_back(-1);
4977 for (unsigned i = 4; i != 8; ++i)
4979 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4982 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4983 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4985 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4989 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4990 // and update MaskVals with the new element order.
4991 if (BestHiQuad >= 0) {
4992 SmallVector<int, 8> MaskV;
4993 for (unsigned i = 0; i != 4; ++i)
4995 for (unsigned i = 4; i != 8; ++i) {
4996 int idx = MaskVals[i];
4998 MaskV.push_back(-1);
5000 } else if ((idx / 4) == BestHiQuad) {
5001 MaskV.push_back((idx & 3) + 4);
5004 MaskV.push_back(-1);
5007 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5010 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5011 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5013 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5017 // In case BestHi & BestLo were both -1, which means each quadword has a word
5018 // from each of the four input quadwords, calculate the InOrder bitvector now
5019 // before falling through to the insert/extract cleanup.
5020 if (BestLoQuad == -1 && BestHiQuad == -1) {
5022 for (int i = 0; i != 8; ++i)
5023 if (MaskVals[i] < 0 || MaskVals[i] == i)
5027 // The other elements are put in the right place using pextrw and pinsrw.
5028 for (unsigned i = 0; i != 8; ++i) {
5031 int EltIdx = MaskVals[i];
5034 SDValue ExtOp = (EltIdx < 8)
5035 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5036 DAG.getIntPtrConstant(EltIdx))
5037 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5038 DAG.getIntPtrConstant(EltIdx - 8));
5039 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5040 DAG.getIntPtrConstant(i));
5045 // v16i8 shuffles - Prefer shuffles in the following order:
5046 // 1. [ssse3] 1 x pshufb
5047 // 2. [ssse3] 2 x pshufb + 1 x por
5048 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5050 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5052 const X86TargetLowering &TLI) {
5053 SDValue V1 = SVOp->getOperand(0);
5054 SDValue V2 = SVOp->getOperand(1);
5055 DebugLoc dl = SVOp->getDebugLoc();
5056 SmallVector<int, 16> MaskVals;
5057 SVOp->getMask(MaskVals);
5059 // If we have SSSE3, case 1 is generated when all result bytes come from
5060 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5061 // present, fall back to case 3.
5062 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5065 for (unsigned i = 0; i < 16; ++i) {
5066 int EltIdx = MaskVals[i];
5075 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5076 if (TLI.getSubtarget()->hasSSSE3()) {
5077 SmallVector<SDValue,16> pshufbMask;
5079 // If all result elements are from one input vector, then only translate
5080 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5082 // Otherwise, we have elements from both input vectors, and must zero out
5083 // elements that come from V2 in the first mask, and V1 in the second mask
5084 // so that we can OR them together.
5085 bool TwoInputs = !(V1Only || V2Only);
5086 for (unsigned i = 0; i != 16; ++i) {
5087 int EltIdx = MaskVals[i];
5088 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5089 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5092 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5094 // If all the elements are from V2, assign it to V1 and return after
5095 // building the first pshufb.
5098 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5099 DAG.getNode(ISD::BUILD_VECTOR, dl,
5100 MVT::v16i8, &pshufbMask[0], 16));
5104 // Calculate the shuffle mask for the second input, shuffle it, and
5105 // OR it with the first shuffled input.
5107 for (unsigned i = 0; i != 16; ++i) {
5108 int EltIdx = MaskVals[i];
5110 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5113 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5115 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5116 DAG.getNode(ISD::BUILD_VECTOR, dl,
5117 MVT::v16i8, &pshufbMask[0], 16));
5118 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5121 // No SSSE3 - Calculate in place words and then fix all out of place words
5122 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5123 // the 16 different words that comprise the two doublequadword input vectors.
5124 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5125 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5126 SDValue NewV = V2Only ? V2 : V1;
5127 for (int i = 0; i != 8; ++i) {
5128 int Elt0 = MaskVals[i*2];
5129 int Elt1 = MaskVals[i*2+1];
5131 // This word of the result is all undef, skip it.
5132 if (Elt0 < 0 && Elt1 < 0)
5135 // This word of the result is already in the correct place, skip it.
5136 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5138 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5141 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5142 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5145 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5146 // using a single extract together, load it and store it.
5147 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5148 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5149 DAG.getIntPtrConstant(Elt1 / 2));
5150 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5151 DAG.getIntPtrConstant(i));
5155 // If Elt1 is defined, extract it from the appropriate source. If the
5156 // source byte is not also odd, shift the extracted word left 8 bits
5157 // otherwise clear the bottom 8 bits if we need to do an or.
5159 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5160 DAG.getIntPtrConstant(Elt1 / 2));
5161 if ((Elt1 & 1) == 0)
5162 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5164 TLI.getShiftAmountTy(InsElt.getValueType())));
5166 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5167 DAG.getConstant(0xFF00, MVT::i16));
5169 // If Elt0 is defined, extract it from the appropriate source. If the
5170 // source byte is not also even, shift the extracted word right 8 bits. If
5171 // Elt1 was also defined, OR the extracted values together before
5172 // inserting them in the result.
5174 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5175 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5176 if ((Elt0 & 1) != 0)
5177 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5179 TLI.getShiftAmountTy(InsElt0.getValueType())));
5181 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5182 DAG.getConstant(0x00FF, MVT::i16));
5183 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5186 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5187 DAG.getIntPtrConstant(i));
5189 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5192 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5193 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5194 /// done when every pair / quad of shuffle mask elements point to elements in
5195 /// the right sequence. e.g.
5196 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5198 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5199 SelectionDAG &DAG, DebugLoc dl) {
5200 EVT VT = SVOp->getValueType(0);
5201 SDValue V1 = SVOp->getOperand(0);
5202 SDValue V2 = SVOp->getOperand(1);
5203 unsigned NumElems = VT.getVectorNumElements();
5204 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5206 switch (VT.getSimpleVT().SimpleTy) {
5207 default: assert(false && "Unexpected!");
5208 case MVT::v4f32: NewVT = MVT::v2f64; break;
5209 case MVT::v4i32: NewVT = MVT::v2i64; break;
5210 case MVT::v8i16: NewVT = MVT::v4i32; break;
5211 case MVT::v16i8: NewVT = MVT::v4i32; break;
5214 int Scale = NumElems / NewWidth;
5215 SmallVector<int, 8> MaskVec;
5216 for (unsigned i = 0; i < NumElems; i += Scale) {
5218 for (int j = 0; j < Scale; ++j) {
5219 int EltIdx = SVOp->getMaskElt(i+j);
5223 StartIdx = EltIdx - (EltIdx % Scale);
5224 if (EltIdx != StartIdx + j)
5228 MaskVec.push_back(-1);
5230 MaskVec.push_back(StartIdx / Scale);
5233 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5234 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5235 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5238 /// getVZextMovL - Return a zero-extending vector move low node.
5240 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5241 SDValue SrcOp, SelectionDAG &DAG,
5242 const X86Subtarget *Subtarget, DebugLoc dl) {
5243 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5244 LoadSDNode *LD = NULL;
5245 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5246 LD = dyn_cast<LoadSDNode>(SrcOp);
5248 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5250 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5251 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5252 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5253 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5254 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5256 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5257 return DAG.getNode(ISD::BITCAST, dl, VT,
5258 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5259 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5267 return DAG.getNode(ISD::BITCAST, dl, VT,
5268 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5269 DAG.getNode(ISD::BITCAST, dl,
5273 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
5276 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5277 SDValue V1 = SVOp->getOperand(0);
5278 SDValue V2 = SVOp->getOperand(1);
5279 DebugLoc dl = SVOp->getDebugLoc();
5280 EVT VT = SVOp->getValueType(0);
5282 SmallVector<std::pair<int, int>, 8> Locs;
5284 SmallVector<int, 8> Mask1(4U, -1);
5285 SmallVector<int, 8> PermMask;
5286 SVOp->getMask(PermMask);
5290 for (unsigned i = 0; i != 4; ++i) {
5291 int Idx = PermMask[i];
5293 Locs[i] = std::make_pair(-1, -1);
5295 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5297 Locs[i] = std::make_pair(0, NumLo);
5301 Locs[i] = std::make_pair(1, NumHi);
5303 Mask1[2+NumHi] = Idx;
5309 if (NumLo <= 2 && NumHi <= 2) {
5310 // If no more than two elements come from either vector. This can be
5311 // implemented with two shuffles. First shuffle gather the elements.
5312 // The second shuffle, which takes the first shuffle as both of its
5313 // vector operands, put the elements into the right order.
5314 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5316 SmallVector<int, 8> Mask2(4U, -1);
5318 for (unsigned i = 0; i != 4; ++i) {
5319 if (Locs[i].first == -1)
5322 unsigned Idx = (i < 2) ? 0 : 4;
5323 Idx += Locs[i].first * 2 + Locs[i].second;
5328 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5329 } else if (NumLo == 3 || NumHi == 3) {
5330 // Otherwise, we must have three elements from one vector, call it X, and
5331 // one element from the other, call it Y. First, use a shufps to build an
5332 // intermediate vector with the one element from Y and the element from X
5333 // that will be in the same half in the final destination (the indexes don't
5334 // matter). Then, use a shufps to build the final vector, taking the half
5335 // containing the element from Y from the intermediate, and the other half
5338 // Normalize it so the 3 elements come from V1.
5339 CommuteVectorShuffleMask(PermMask, VT);
5343 // Find the element from V2.
5345 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5346 int Val = PermMask[HiIndex];
5353 Mask1[0] = PermMask[HiIndex];
5355 Mask1[2] = PermMask[HiIndex^1];
5357 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5360 Mask1[0] = PermMask[0];
5361 Mask1[1] = PermMask[1];
5362 Mask1[2] = HiIndex & 1 ? 6 : 4;
5363 Mask1[3] = HiIndex & 1 ? 4 : 6;
5364 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5366 Mask1[0] = HiIndex & 1 ? 2 : 0;
5367 Mask1[1] = HiIndex & 1 ? 0 : 2;
5368 Mask1[2] = PermMask[2];
5369 Mask1[3] = PermMask[3];
5374 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5378 // Break it into (shuffle shuffle_hi, shuffle_lo).
5381 SmallVector<int,8> LoMask(4U, -1);
5382 SmallVector<int,8> HiMask(4U, -1);
5384 SmallVector<int,8> *MaskPtr = &LoMask;
5385 unsigned MaskIdx = 0;
5388 for (unsigned i = 0; i != 4; ++i) {
5395 int Idx = PermMask[i];
5397 Locs[i] = std::make_pair(-1, -1);
5398 } else if (Idx < 4) {
5399 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5400 (*MaskPtr)[LoIdx] = Idx;
5403 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5404 (*MaskPtr)[HiIdx] = Idx;
5409 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5410 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5411 SmallVector<int, 8> MaskOps;
5412 for (unsigned i = 0; i != 4; ++i) {
5413 if (Locs[i].first == -1) {
5414 MaskOps.push_back(-1);
5416 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5417 MaskOps.push_back(Idx);
5420 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5423 static bool MayFoldVectorLoad(SDValue V) {
5424 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5425 V = V.getOperand(0);
5426 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5427 V = V.getOperand(0);
5433 // FIXME: the version above should always be used. Since there's
5434 // a bug where several vector shuffles can't be folded because the
5435 // DAG is not updated during lowering and a node claims to have two
5436 // uses while it only has one, use this version, and let isel match
5437 // another instruction if the load really happens to have more than
5438 // one use. Remove this version after this bug get fixed.
5439 // rdar://8434668, PR8156
5440 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5441 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5442 V = V.getOperand(0);
5443 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5444 V = V.getOperand(0);
5445 if (ISD::isNormalLoad(V.getNode()))
5450 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5451 /// a vector extract, and if both can be later optimized into a single load.
5452 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5453 /// here because otherwise a target specific shuffle node is going to be
5454 /// emitted for this shuffle, and the optimization not done.
5455 /// FIXME: This is probably not the best approach, but fix the problem
5456 /// until the right path is decided.
5458 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5459 const TargetLowering &TLI) {
5460 EVT VT = V.getValueType();
5461 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5463 // Be sure that the vector shuffle is present in a pattern like this:
5464 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5468 SDNode *N = *V.getNode()->use_begin();
5469 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5472 SDValue EltNo = N->getOperand(1);
5473 if (!isa<ConstantSDNode>(EltNo))
5476 // If the bit convert changed the number of elements, it is unsafe
5477 // to examine the mask.
5478 bool HasShuffleIntoBitcast = false;
5479 if (V.getOpcode() == ISD::BITCAST) {
5480 EVT SrcVT = V.getOperand(0).getValueType();
5481 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5483 V = V.getOperand(0);
5484 HasShuffleIntoBitcast = true;
5487 // Select the input vector, guarding against out of range extract vector.
5488 unsigned NumElems = VT.getVectorNumElements();
5489 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5490 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5491 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5493 // Skip one more bit_convert if necessary
5494 if (V.getOpcode() == ISD::BITCAST)
5495 V = V.getOperand(0);
5497 if (ISD::isNormalLoad(V.getNode())) {
5498 // Is the original load suitable?
5499 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5501 // FIXME: avoid the multi-use bug that is preventing lots of
5502 // of foldings to be detected, this is still wrong of course, but
5503 // give the temporary desired behavior, and if it happens that
5504 // the load has real more uses, during isel it will not fold, and
5505 // will generate poor code.
5506 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5509 if (!HasShuffleIntoBitcast)
5512 // If there's a bitcast before the shuffle, check if the load type and
5513 // alignment is valid.
5514 unsigned Align = LN0->getAlignment();
5516 TLI.getTargetData()->getABITypeAlignment(
5517 VT.getTypeForEVT(*DAG.getContext()));
5519 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5527 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5528 EVT VT = Op.getValueType();
5530 // Canonizalize to v2f64.
5531 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5532 return DAG.getNode(ISD::BITCAST, dl, VT,
5533 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5538 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5540 SDValue V1 = Op.getOperand(0);
5541 SDValue V2 = Op.getOperand(1);
5542 EVT VT = Op.getValueType();
5544 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5546 if (HasSSE2 && VT == MVT::v2f64)
5547 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5550 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5554 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5555 SDValue V1 = Op.getOperand(0);
5556 SDValue V2 = Op.getOperand(1);
5557 EVT VT = Op.getValueType();
5559 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5560 "unsupported shuffle type");
5562 if (V2.getOpcode() == ISD::UNDEF)
5566 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5570 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5571 SDValue V1 = Op.getOperand(0);
5572 SDValue V2 = Op.getOperand(1);
5573 EVT VT = Op.getValueType();
5574 unsigned NumElems = VT.getVectorNumElements();
5576 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5577 // operand of these instructions is only memory, so check if there's a
5578 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5580 bool CanFoldLoad = false;
5582 // Trivial case, when V2 comes from a load.
5583 if (MayFoldVectorLoad(V2))
5586 // When V1 is a load, it can be folded later into a store in isel, example:
5587 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5589 // (MOVLPSmr addr:$src1, VR128:$src2)
5590 // So, recognize this potential and also use MOVLPS or MOVLPD
5591 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5594 // Both of them can't be memory operations though.
5595 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5596 CanFoldLoad = false;
5599 if (HasSSE2 && NumElems == 2)
5600 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5603 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5606 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5607 // movl and movlp will both match v2i64, but v2i64 is never matched by
5608 // movl earlier because we make it strict to avoid messing with the movlp load
5609 // folding logic (see the code above getMOVLP call). Match it here then,
5610 // this is horrible, but will stay like this until we move all shuffle
5611 // matching to x86 specific nodes. Note that for the 1st condition all
5612 // types are matched with movsd.
5613 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5614 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5616 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5619 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5621 // Invert the operand order and use SHUFPS to match it.
5622 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5623 X86::getShuffleSHUFImmediate(SVOp), DAG);
5626 static inline unsigned getUNPCKLOpcode(EVT VT, const X86Subtarget *Subtarget) {
5627 switch(VT.getSimpleVT().SimpleTy) {
5628 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5629 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5631 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPS : X86ISD::UNPCKLPS;
5633 return Subtarget->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
5634 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5635 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5636 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5637 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5639 llvm_unreachable("Unknown type for unpckl");
5644 static inline unsigned getUNPCKHOpcode(EVT VT) {
5645 switch(VT.getSimpleVT().SimpleTy) {
5646 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5647 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5648 case MVT::v4f32: return X86ISD::UNPCKHPS;
5649 case MVT::v2f64: return X86ISD::UNPCKHPD;
5650 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5651 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5653 llvm_unreachable("Unknown type for unpckh");
5659 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5660 const TargetLowering &TLI,
5661 const X86Subtarget *Subtarget) {
5662 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5663 EVT VT = Op.getValueType();
5664 DebugLoc dl = Op.getDebugLoc();
5665 SDValue V1 = Op.getOperand(0);
5666 SDValue V2 = Op.getOperand(1);
5668 if (isZeroShuffle(SVOp))
5669 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5671 // Handle splat operations
5672 if (SVOp->isSplat()) {
5673 // Special case, this is the only place now where it's
5674 // allowed to return a vector_shuffle operation without
5675 // using a target specific node, because *hopefully* it
5676 // will be optimized away by the dag combiner.
5677 if (VT.getVectorNumElements() <= 4 &&
5678 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5681 // Handle splats by matching through known masks
5682 if (VT.getVectorNumElements() <= 4)
5685 // Canonicalize all of the remaining to v4f32.
5686 return PromoteSplat(SVOp, DAG);
5689 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5691 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5692 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5693 if (NewOp.getNode())
5694 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5695 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5696 // FIXME: Figure out a cleaner way to do this.
5697 // Try to make use of movq to zero out the top part.
5698 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5699 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5700 if (NewOp.getNode()) {
5701 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5702 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5703 DAG, Subtarget, dl);
5705 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5706 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5707 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5708 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5709 DAG, Subtarget, dl);
5716 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5717 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5718 SDValue V1 = Op.getOperand(0);
5719 SDValue V2 = Op.getOperand(1);
5720 EVT VT = Op.getValueType();
5721 DebugLoc dl = Op.getDebugLoc();
5722 unsigned NumElems = VT.getVectorNumElements();
5723 bool isMMX = VT.getSizeInBits() == 64;
5724 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5725 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5726 bool V1IsSplat = false;
5727 bool V2IsSplat = false;
5728 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5729 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5730 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5731 MachineFunction &MF = DAG.getMachineFunction();
5732 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5734 // Shuffle operations on MMX not supported.
5738 // Vector shuffle lowering takes 3 steps:
5740 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5741 // narrowing and commutation of operands should be handled.
5742 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5744 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5745 // so the shuffle can be broken into other shuffles and the legalizer can
5746 // try the lowering again.
5748 // The general ideia is that no vector_shuffle operation should be left to
5749 // be matched during isel, all of them must be converted to a target specific
5752 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5753 // narrowing and commutation of operands should be handled. The actual code
5754 // doesn't include all of those, work in progress...
5755 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5756 if (NewOp.getNode())
5759 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5760 // unpckh_undef). Only use pshufd if speed is more important than size.
5761 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5762 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5763 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
5764 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5765 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5766 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5768 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5769 RelaxedMayFoldVectorLoad(V1))
5770 return getMOVDDup(Op, dl, V1, DAG);
5772 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5773 return getMOVHighToLow(Op, dl, DAG);
5775 // Use to match splats
5776 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5777 (VT == MVT::v2f64 || VT == MVT::v2i64))
5778 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5780 if (X86::isPSHUFDMask(SVOp)) {
5781 // The actual implementation will match the mask in the if above and then
5782 // during isel it can match several different instructions, not only pshufd
5783 // as its name says, sad but true, emulate the behavior for now...
5784 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5785 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5787 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5789 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5790 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5792 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5793 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5796 if (VT == MVT::v4f32)
5797 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5801 // Check if this can be converted into a logical shift.
5802 bool isLeft = false;
5805 bool isShift = getSubtarget()->hasSSE2() &&
5806 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5807 if (isShift && ShVal.hasOneUse()) {
5808 // If the shifted value has multiple uses, it may be cheaper to use
5809 // v_set0 + movlhps or movhlps, etc.
5810 EVT EltVT = VT.getVectorElementType();
5811 ShAmt *= EltVT.getSizeInBits();
5812 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5815 if (X86::isMOVLMask(SVOp)) {
5818 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5819 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5820 if (!X86::isMOVLPMask(SVOp)) {
5821 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5822 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5824 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5825 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5829 // FIXME: fold these into legal mask.
5830 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5831 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5833 if (X86::isMOVHLPSMask(SVOp))
5834 return getMOVHighToLow(Op, dl, DAG);
5836 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5837 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5839 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5840 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5842 if (X86::isMOVLPMask(SVOp))
5843 return getMOVLP(Op, dl, DAG, HasSSE2);
5845 if (ShouldXformToMOVHLPS(SVOp) ||
5846 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5847 return CommuteVectorShuffle(SVOp, DAG);
5850 // No better options. Use a vshl / vsrl.
5851 EVT EltVT = VT.getVectorElementType();
5852 ShAmt *= EltVT.getSizeInBits();
5853 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5856 bool Commuted = false;
5857 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5858 // 1,1,1,1 -> v8i16 though.
5859 V1IsSplat = isSplatVector(V1.getNode());
5860 V2IsSplat = isSplatVector(V2.getNode());
5862 // Canonicalize the splat or undef, if present, to be on the RHS.
5863 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5864 Op = CommuteVectorShuffle(SVOp, DAG);
5865 SVOp = cast<ShuffleVectorSDNode>(Op);
5866 V1 = SVOp->getOperand(0);
5867 V2 = SVOp->getOperand(1);
5868 std::swap(V1IsSplat, V2IsSplat);
5869 std::swap(V1IsUndef, V2IsUndef);
5873 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5874 // Shuffling low element of v1 into undef, just return v1.
5877 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5878 // the instruction selector will not match, so get a canonical MOVL with
5879 // swapped operands to undo the commute.
5880 return getMOVL(DAG, dl, VT, V2, V1);
5883 if (X86::isUNPCKLMask(SVOp))
5884 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5885 dl, VT, V1, V2, DAG);
5887 if (X86::isUNPCKHMask(SVOp))
5888 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5891 // Normalize mask so all entries that point to V2 points to its first
5892 // element then try to match unpck{h|l} again. If match, return a
5893 // new vector_shuffle with the corrected mask.
5894 SDValue NewMask = NormalizeMask(SVOp, DAG);
5895 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5896 if (NSVOp != SVOp) {
5897 if (X86::isUNPCKLMask(NSVOp, true)) {
5899 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5906 // Commute is back and try unpck* again.
5907 // FIXME: this seems wrong.
5908 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5909 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5911 if (X86::isUNPCKLMask(NewSVOp))
5912 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5913 dl, VT, V2, V1, DAG);
5915 if (X86::isUNPCKHMask(NewSVOp))
5916 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5919 // Normalize the node to match x86 shuffle ops if needed
5920 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5921 return CommuteVectorShuffle(SVOp, DAG);
5923 // The checks below are all present in isShuffleMaskLegal, but they are
5924 // inlined here right now to enable us to directly emit target specific
5925 // nodes, and remove one by one until they don't return Op anymore.
5926 SmallVector<int, 16> M;
5929 if (isPALIGNRMask(M, VT, HasSSSE3))
5930 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5931 X86::getShufflePALIGNRImmediate(SVOp),
5934 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5935 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5936 if (VT == MVT::v2f64) {
5937 X86ISD::NodeType Opcode =
5938 getSubtarget()->hasAVX() ? X86ISD::VUNPCKLPD : X86ISD::UNPCKLPD;
5939 return getTargetShuffleNode(Opcode, dl, VT, V1, V1, DAG);
5941 if (VT == MVT::v2i64)
5942 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5945 if (isPSHUFHWMask(M, VT))
5946 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5947 X86::getShufflePSHUFHWImmediate(SVOp),
5950 if (isPSHUFLWMask(M, VT))
5951 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5952 X86::getShufflePSHUFLWImmediate(SVOp),
5955 if (isSHUFPMask(M, VT)) {
5956 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5957 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5958 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5960 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5961 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5965 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5966 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5967 return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
5968 dl, VT, V1, V1, DAG);
5969 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5970 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5971 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5973 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5974 if (VT == MVT::v8i16) {
5975 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5976 if (NewOp.getNode())
5980 if (VT == MVT::v16i8) {
5981 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5982 if (NewOp.getNode())
5986 // Handle all 4 wide cases with a number of shuffles.
5988 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5994 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5995 SelectionDAG &DAG) const {
5996 EVT VT = Op.getValueType();
5997 DebugLoc dl = Op.getDebugLoc();
5998 if (VT.getSizeInBits() == 8) {
5999 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6000 Op.getOperand(0), Op.getOperand(1));
6001 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6002 DAG.getValueType(VT));
6003 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6004 } else if (VT.getSizeInBits() == 16) {
6005 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6006 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6008 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6009 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6010 DAG.getNode(ISD::BITCAST, dl,
6014 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6015 Op.getOperand(0), Op.getOperand(1));
6016 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6017 DAG.getValueType(VT));
6018 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6019 } else if (VT == MVT::f32) {
6020 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6021 // the result back to FR32 register. It's only worth matching if the
6022 // result has a single use which is a store or a bitcast to i32. And in
6023 // the case of a store, it's not worth it if the index is a constant 0,
6024 // because a MOVSSmr can be used instead, which is smaller and faster.
6025 if (!Op.hasOneUse())
6027 SDNode *User = *Op.getNode()->use_begin();
6028 if ((User->getOpcode() != ISD::STORE ||
6029 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6030 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6031 (User->getOpcode() != ISD::BITCAST ||
6032 User->getValueType(0) != MVT::i32))
6034 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6035 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6038 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6039 } else if (VT == MVT::i32) {
6040 // ExtractPS works with constant index.
6041 if (isa<ConstantSDNode>(Op.getOperand(1)))
6049 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6050 SelectionDAG &DAG) const {
6051 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6054 SDValue Vec = Op.getOperand(0);
6055 EVT VecVT = Vec.getValueType();
6057 // If this is a 256-bit vector result, first extract the 128-bit
6058 // vector and then extract from the 128-bit vector.
6059 if (VecVT.getSizeInBits() > 128) {
6060 DebugLoc dl = Op.getNode()->getDebugLoc();
6061 unsigned NumElems = VecVT.getVectorNumElements();
6062 SDValue Idx = Op.getOperand(1);
6064 if (!isa<ConstantSDNode>(Idx))
6067 unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
6068 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6070 // Get the 128-bit vector.
6071 bool Upper = IdxVal >= ExtractNumElems;
6072 Vec = Extract128BitVector(Vec, Idx, DAG, dl);
6075 SDValue ScaledIdx = Idx;
6077 ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
6078 DAG.getConstant(ExtractNumElems,
6079 Idx.getValueType()));
6080 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6084 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6086 if (Subtarget->hasSSE41()) {
6087 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6092 EVT VT = Op.getValueType();
6093 DebugLoc dl = Op.getDebugLoc();
6094 // TODO: handle v16i8.
6095 if (VT.getSizeInBits() == 16) {
6096 SDValue Vec = Op.getOperand(0);
6097 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6099 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6100 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6101 DAG.getNode(ISD::BITCAST, dl,
6104 // Transform it so it match pextrw which produces a 32-bit result.
6105 EVT EltVT = MVT::i32;
6106 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6107 Op.getOperand(0), Op.getOperand(1));
6108 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6109 DAG.getValueType(VT));
6110 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6111 } else if (VT.getSizeInBits() == 32) {
6112 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6116 // SHUFPS the element to the lowest double word, then movss.
6117 int Mask[4] = { Idx, -1, -1, -1 };
6118 EVT VVT = Op.getOperand(0).getValueType();
6119 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6120 DAG.getUNDEF(VVT), Mask);
6121 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6122 DAG.getIntPtrConstant(0));
6123 } else if (VT.getSizeInBits() == 64) {
6124 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6125 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6126 // to match extract_elt for f64.
6127 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6131 // UNPCKHPD the element to the lowest double word, then movsd.
6132 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6133 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6134 int Mask[2] = { 1, -1 };
6135 EVT VVT = Op.getOperand(0).getValueType();
6136 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6137 DAG.getUNDEF(VVT), Mask);
6138 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6139 DAG.getIntPtrConstant(0));
6146 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6147 SelectionDAG &DAG) const {
6148 EVT VT = Op.getValueType();
6149 EVT EltVT = VT.getVectorElementType();
6150 DebugLoc dl = Op.getDebugLoc();
6152 SDValue N0 = Op.getOperand(0);
6153 SDValue N1 = Op.getOperand(1);
6154 SDValue N2 = Op.getOperand(2);
6156 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6157 isa<ConstantSDNode>(N2)) {
6159 if (VT == MVT::v8i16)
6160 Opc = X86ISD::PINSRW;
6161 else if (VT == MVT::v16i8)
6162 Opc = X86ISD::PINSRB;
6164 Opc = X86ISD::PINSRB;
6166 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6168 if (N1.getValueType() != MVT::i32)
6169 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6170 if (N2.getValueType() != MVT::i32)
6171 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6172 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6173 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6174 // Bits [7:6] of the constant are the source select. This will always be
6175 // zero here. The DAG Combiner may combine an extract_elt index into these
6176 // bits. For example (insert (extract, 3), 2) could be matched by putting
6177 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6178 // Bits [5:4] of the constant are the destination select. This is the
6179 // value of the incoming immediate.
6180 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6181 // combine either bitwise AND or insert of float 0.0 to set these bits.
6182 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6183 // Create this as a scalar to vector..
6184 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6185 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6186 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6187 // PINSR* works with constant index.
6194 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6195 EVT VT = Op.getValueType();
6196 EVT EltVT = VT.getVectorElementType();
6198 DebugLoc dl = Op.getDebugLoc();
6199 SDValue N0 = Op.getOperand(0);
6200 SDValue N1 = Op.getOperand(1);
6201 SDValue N2 = Op.getOperand(2);
6203 // If this is a 256-bit vector result, first insert into a 128-bit
6204 // vector and then insert into the 256-bit vector.
6205 if (VT.getSizeInBits() > 128) {
6206 if (!isa<ConstantSDNode>(N2))
6209 // Get the 128-bit vector.
6210 unsigned NumElems = VT.getVectorNumElements();
6211 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6212 bool Upper = IdxVal >= NumElems / 2;
6214 SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
6217 SDValue ScaledN2 = N2;
6219 ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
6220 DAG.getConstant(NumElems /
6221 (VT.getSizeInBits() / 128),
6222 N2.getValueType()));
6223 Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
6226 // Insert the 128-bit vector
6227 // FIXME: Why UNDEF?
6228 return Insert128BitVector(N0, Op, N2, DAG, dl);
6231 if (Subtarget->hasSSE41())
6232 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6234 if (EltVT == MVT::i8)
6237 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6238 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6239 // as its second argument.
6240 if (N1.getValueType() != MVT::i32)
6241 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6242 if (N2.getValueType() != MVT::i32)
6243 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6244 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6250 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6251 LLVMContext *Context = DAG.getContext();
6252 DebugLoc dl = Op.getDebugLoc();
6253 EVT OpVT = Op.getValueType();
6255 // If this is a 256-bit vector result, first insert into a 128-bit
6256 // vector and then insert into the 256-bit vector.
6257 if (OpVT.getSizeInBits() > 128) {
6258 // Insert into a 128-bit vector.
6259 EVT VT128 = EVT::getVectorVT(*Context,
6260 OpVT.getVectorElementType(),
6261 OpVT.getVectorNumElements() / 2);
6263 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6265 // Insert the 128-bit vector.
6266 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6267 DAG.getConstant(0, MVT::i32),
6271 if (Op.getValueType() == MVT::v1i64 &&
6272 Op.getOperand(0).getValueType() == MVT::i64)
6273 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6275 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6276 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6277 "Expected an SSE type!");
6278 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6279 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6282 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6283 // a simple subregister reference or explicit instructions to grab
6284 // upper bits of a vector.
6286 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6287 if (Subtarget->hasAVX()) {
6288 DebugLoc dl = Op.getNode()->getDebugLoc();
6289 SDValue Vec = Op.getNode()->getOperand(0);
6290 SDValue Idx = Op.getNode()->getOperand(1);
6292 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6293 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6294 return Extract128BitVector(Vec, Idx, DAG, dl);
6300 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6301 // simple superregister reference or explicit instructions to insert
6302 // the upper bits of a vector.
6304 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6305 if (Subtarget->hasAVX()) {
6306 DebugLoc dl = Op.getNode()->getDebugLoc();
6307 SDValue Vec = Op.getNode()->getOperand(0);
6308 SDValue SubVec = Op.getNode()->getOperand(1);
6309 SDValue Idx = Op.getNode()->getOperand(2);
6311 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6312 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6313 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6319 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6320 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6321 // one of the above mentioned nodes. It has to be wrapped because otherwise
6322 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6323 // be used to form addressing mode. These wrapped nodes will be selected
6326 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6327 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6329 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6331 unsigned char OpFlag = 0;
6332 unsigned WrapperKind = X86ISD::Wrapper;
6333 CodeModel::Model M = getTargetMachine().getCodeModel();
6335 if (Subtarget->isPICStyleRIPRel() &&
6336 (M == CodeModel::Small || M == CodeModel::Kernel))
6337 WrapperKind = X86ISD::WrapperRIP;
6338 else if (Subtarget->isPICStyleGOT())
6339 OpFlag = X86II::MO_GOTOFF;
6340 else if (Subtarget->isPICStyleStubPIC())
6341 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6343 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6345 CP->getOffset(), OpFlag);
6346 DebugLoc DL = CP->getDebugLoc();
6347 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6348 // With PIC, the address is actually $g + Offset.
6350 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6351 DAG.getNode(X86ISD::GlobalBaseReg,
6352 DebugLoc(), getPointerTy()),
6359 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6360 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6362 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6364 unsigned char OpFlag = 0;
6365 unsigned WrapperKind = X86ISD::Wrapper;
6366 CodeModel::Model M = getTargetMachine().getCodeModel();
6368 if (Subtarget->isPICStyleRIPRel() &&
6369 (M == CodeModel::Small || M == CodeModel::Kernel))
6370 WrapperKind = X86ISD::WrapperRIP;
6371 else if (Subtarget->isPICStyleGOT())
6372 OpFlag = X86II::MO_GOTOFF;
6373 else if (Subtarget->isPICStyleStubPIC())
6374 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6376 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6378 DebugLoc DL = JT->getDebugLoc();
6379 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6381 // With PIC, the address is actually $g + Offset.
6383 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6384 DAG.getNode(X86ISD::GlobalBaseReg,
6385 DebugLoc(), getPointerTy()),
6392 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6393 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6395 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6397 unsigned char OpFlag = 0;
6398 unsigned WrapperKind = X86ISD::Wrapper;
6399 CodeModel::Model M = getTargetMachine().getCodeModel();
6401 if (Subtarget->isPICStyleRIPRel() &&
6402 (M == CodeModel::Small || M == CodeModel::Kernel))
6403 WrapperKind = X86ISD::WrapperRIP;
6404 else if (Subtarget->isPICStyleGOT())
6405 OpFlag = X86II::MO_GOTOFF;
6406 else if (Subtarget->isPICStyleStubPIC())
6407 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6409 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6411 DebugLoc DL = Op.getDebugLoc();
6412 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6415 // With PIC, the address is actually $g + Offset.
6416 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6417 !Subtarget->is64Bit()) {
6418 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6419 DAG.getNode(X86ISD::GlobalBaseReg,
6420 DebugLoc(), getPointerTy()),
6428 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6429 // Create the TargetBlockAddressAddress node.
6430 unsigned char OpFlags =
6431 Subtarget->ClassifyBlockAddressReference();
6432 CodeModel::Model M = getTargetMachine().getCodeModel();
6433 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6434 DebugLoc dl = Op.getDebugLoc();
6435 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6436 /*isTarget=*/true, OpFlags);
6438 if (Subtarget->isPICStyleRIPRel() &&
6439 (M == CodeModel::Small || M == CodeModel::Kernel))
6440 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6442 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6444 // With PIC, the address is actually $g + Offset.
6445 if (isGlobalRelativeToPICBase(OpFlags)) {
6446 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6447 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6455 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6457 SelectionDAG &DAG) const {
6458 // Create the TargetGlobalAddress node, folding in the constant
6459 // offset if it is legal.
6460 unsigned char OpFlags =
6461 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6462 CodeModel::Model M = getTargetMachine().getCodeModel();
6464 if (OpFlags == X86II::MO_NO_FLAG &&
6465 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6466 // A direct static reference to a global.
6467 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6470 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6473 if (Subtarget->isPICStyleRIPRel() &&
6474 (M == CodeModel::Small || M == CodeModel::Kernel))
6475 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6477 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6479 // With PIC, the address is actually $g + Offset.
6480 if (isGlobalRelativeToPICBase(OpFlags)) {
6481 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6482 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6486 // For globals that require a load from a stub to get the address, emit the
6488 if (isGlobalStubReference(OpFlags))
6489 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6490 MachinePointerInfo::getGOT(), false, false, 0);
6492 // If there was a non-zero offset that we didn't fold, create an explicit
6495 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6496 DAG.getConstant(Offset, getPointerTy()));
6502 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6503 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6504 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6505 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6509 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6510 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6511 unsigned char OperandFlags) {
6512 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6513 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6514 DebugLoc dl = GA->getDebugLoc();
6515 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6516 GA->getValueType(0),
6520 SDValue Ops[] = { Chain, TGA, *InFlag };
6521 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6523 SDValue Ops[] = { Chain, TGA };
6524 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6527 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6528 MFI->setAdjustsStack(true);
6530 SDValue Flag = Chain.getValue(1);
6531 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6534 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6536 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6539 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6540 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6541 DAG.getNode(X86ISD::GlobalBaseReg,
6542 DebugLoc(), PtrVT), InFlag);
6543 InFlag = Chain.getValue(1);
6545 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6548 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6550 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6552 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6553 X86::RAX, X86II::MO_TLSGD);
6556 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6557 // "local exec" model.
6558 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6559 const EVT PtrVT, TLSModel::Model model,
6561 DebugLoc dl = GA->getDebugLoc();
6563 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6564 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6565 is64Bit ? 257 : 256));
6567 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6568 DAG.getIntPtrConstant(0),
6569 MachinePointerInfo(Ptr), false, false, 0);
6571 unsigned char OperandFlags = 0;
6572 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6574 unsigned WrapperKind = X86ISD::Wrapper;
6575 if (model == TLSModel::LocalExec) {
6576 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6577 } else if (is64Bit) {
6578 assert(model == TLSModel::InitialExec);
6579 OperandFlags = X86II::MO_GOTTPOFF;
6580 WrapperKind = X86ISD::WrapperRIP;
6582 assert(model == TLSModel::InitialExec);
6583 OperandFlags = X86II::MO_INDNTPOFF;
6586 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6588 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6589 GA->getValueType(0),
6590 GA->getOffset(), OperandFlags);
6591 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6593 if (model == TLSModel::InitialExec)
6594 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6595 MachinePointerInfo::getGOT(), false, false, 0);
6597 // The address of the thread local variable is the add of the thread
6598 // pointer with the offset of the variable.
6599 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6603 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6605 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6606 const GlobalValue *GV = GA->getGlobal();
6608 if (Subtarget->isTargetELF()) {
6609 // TODO: implement the "local dynamic" model
6610 // TODO: implement the "initial exec"model for pic executables
6612 // If GV is an alias then use the aliasee for determining
6613 // thread-localness.
6614 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6615 GV = GA->resolveAliasedGlobal(false);
6617 TLSModel::Model model
6618 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6621 case TLSModel::GeneralDynamic:
6622 case TLSModel::LocalDynamic: // not implemented
6623 if (Subtarget->is64Bit())
6624 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6625 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6627 case TLSModel::InitialExec:
6628 case TLSModel::LocalExec:
6629 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6630 Subtarget->is64Bit());
6632 } else if (Subtarget->isTargetDarwin()) {
6633 // Darwin only has one model of TLS. Lower to that.
6634 unsigned char OpFlag = 0;
6635 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6636 X86ISD::WrapperRIP : X86ISD::Wrapper;
6638 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6640 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6641 !Subtarget->is64Bit();
6643 OpFlag = X86II::MO_TLVP_PIC_BASE;
6645 OpFlag = X86II::MO_TLVP;
6646 DebugLoc DL = Op.getDebugLoc();
6647 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6648 GA->getValueType(0),
6649 GA->getOffset(), OpFlag);
6650 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6652 // With PIC32, the address is actually $g + Offset.
6654 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6655 DAG.getNode(X86ISD::GlobalBaseReg,
6656 DebugLoc(), getPointerTy()),
6659 // Lowering the machine isd will make sure everything is in the right
6661 SDValue Chain = DAG.getEntryNode();
6662 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6663 SDValue Args[] = { Chain, Offset };
6664 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6666 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6667 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6668 MFI->setAdjustsStack(true);
6670 // And our return value (tls address) is in the standard call return value
6672 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6673 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6677 "TLS not implemented for this target.");
6679 llvm_unreachable("Unreachable");
6684 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
6685 /// take a 2 x i32 value to shift plus a shift amount.
6686 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
6687 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6688 EVT VT = Op.getValueType();
6689 unsigned VTBits = VT.getSizeInBits();
6690 DebugLoc dl = Op.getDebugLoc();
6691 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6692 SDValue ShOpLo = Op.getOperand(0);
6693 SDValue ShOpHi = Op.getOperand(1);
6694 SDValue ShAmt = Op.getOperand(2);
6695 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6696 DAG.getConstant(VTBits - 1, MVT::i8))
6697 : DAG.getConstant(0, VT);
6700 if (Op.getOpcode() == ISD::SHL_PARTS) {
6701 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6702 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6704 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6705 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6708 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6709 DAG.getConstant(VTBits, MVT::i8));
6710 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6711 AndNode, DAG.getConstant(0, MVT::i8));
6714 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6715 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6716 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6718 if (Op.getOpcode() == ISD::SHL_PARTS) {
6719 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6720 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6722 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6723 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6726 SDValue Ops[2] = { Lo, Hi };
6727 return DAG.getMergeValues(Ops, 2, dl);
6730 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6731 SelectionDAG &DAG) const {
6732 EVT SrcVT = Op.getOperand(0).getValueType();
6734 if (SrcVT.isVector())
6737 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6738 "Unknown SINT_TO_FP to lower!");
6740 // These are really Legal; return the operand so the caller accepts it as
6742 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6744 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6745 Subtarget->is64Bit()) {
6749 DebugLoc dl = Op.getDebugLoc();
6750 unsigned Size = SrcVT.getSizeInBits()/8;
6751 MachineFunction &MF = DAG.getMachineFunction();
6752 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6753 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6754 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6756 MachinePointerInfo::getFixedStack(SSFI),
6758 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6761 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6763 SelectionDAG &DAG) const {
6765 DebugLoc DL = Op.getDebugLoc();
6767 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6769 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6771 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6773 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6775 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
6776 MachineMemOperand *MMO;
6778 int SSFI = FI->getIndex();
6780 DAG.getMachineFunction()
6781 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6782 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6784 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
6785 StackSlot = StackSlot.getOperand(1);
6787 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6788 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6790 Tys, Ops, array_lengthof(Ops),
6794 Chain = Result.getValue(1);
6795 SDValue InFlag = Result.getValue(2);
6797 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6798 // shouldn't be necessary except that RFP cannot be live across
6799 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6800 MachineFunction &MF = DAG.getMachineFunction();
6801 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6802 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6803 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6804 Tys = DAG.getVTList(MVT::Other);
6806 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6808 MachineMemOperand *MMO =
6809 DAG.getMachineFunction()
6810 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6811 MachineMemOperand::MOStore, SSFISize, SSFISize);
6813 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6814 Ops, array_lengthof(Ops),
6815 Op.getValueType(), MMO);
6816 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6817 MachinePointerInfo::getFixedStack(SSFI),
6824 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6825 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6826 SelectionDAG &DAG) const {
6827 // This algorithm is not obvious. Here it is in C code, more or less:
6829 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6830 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6831 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6833 // Copy ints to xmm registers.
6834 __m128i xh = _mm_cvtsi32_si128( hi );
6835 __m128i xl = _mm_cvtsi32_si128( lo );
6837 // Combine into low half of a single xmm register.
6838 __m128i x = _mm_unpacklo_epi32( xh, xl );
6842 // Merge in appropriate exponents to give the integer bits the right
6844 x = _mm_unpacklo_epi32( x, exp );
6846 // Subtract away the biases to deal with the IEEE-754 double precision
6848 d = _mm_sub_pd( (__m128d) x, bias );
6850 // All conversions up to here are exact. The correctly rounded result is
6851 // calculated using the current rounding mode using the following
6853 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6854 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6855 // store doesn't really need to be here (except
6856 // maybe to zero the other double)
6861 DebugLoc dl = Op.getDebugLoc();
6862 LLVMContext *Context = DAG.getContext();
6864 // Build some magic constants.
6865 std::vector<Constant*> CV0;
6866 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6867 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6868 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6869 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6870 Constant *C0 = ConstantVector::get(CV0);
6871 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6873 std::vector<Constant*> CV1;
6875 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6877 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6878 Constant *C1 = ConstantVector::get(CV1);
6879 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6881 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6882 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6884 DAG.getIntPtrConstant(1)));
6885 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6886 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6888 DAG.getIntPtrConstant(0)));
6889 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6890 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6891 MachinePointerInfo::getConstantPool(),
6893 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6894 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6895 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6896 MachinePointerInfo::getConstantPool(),
6898 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6900 // Add the halves; easiest way is to swap them into another reg first.
6901 int ShufMask[2] = { 1, -1 };
6902 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6903 DAG.getUNDEF(MVT::v2f64), ShufMask);
6904 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6905 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6906 DAG.getIntPtrConstant(0));
6909 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6910 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6911 SelectionDAG &DAG) const {
6912 DebugLoc dl = Op.getDebugLoc();
6913 // FP constant to bias correct the final result.
6914 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6917 // Load the 32-bit value into an XMM register.
6918 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6919 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6921 DAG.getIntPtrConstant(0)));
6923 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6924 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6925 DAG.getIntPtrConstant(0));
6927 // Or the load with the bias.
6928 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6929 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6930 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6932 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6933 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6934 MVT::v2f64, Bias)));
6935 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6936 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
6937 DAG.getIntPtrConstant(0));
6939 // Subtract the bias.
6940 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6942 // Handle final rounding.
6943 EVT DestVT = Op.getValueType();
6945 if (DestVT.bitsLT(MVT::f64)) {
6946 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6947 DAG.getIntPtrConstant(0));
6948 } else if (DestVT.bitsGT(MVT::f64)) {
6949 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6952 // Handle final rounding.
6956 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6957 SelectionDAG &DAG) const {
6958 SDValue N0 = Op.getOperand(0);
6959 DebugLoc dl = Op.getDebugLoc();
6961 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6962 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6963 // the optimization here.
6964 if (DAG.SignBitIsZero(N0))
6965 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6967 EVT SrcVT = N0.getValueType();
6968 EVT DstVT = Op.getValueType();
6969 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6970 return LowerUINT_TO_FP_i64(Op, DAG);
6971 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6972 return LowerUINT_TO_FP_i32(Op, DAG);
6974 // Make a 64-bit buffer, and use it to build an FILD.
6975 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6976 if (SrcVT == MVT::i32) {
6977 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6978 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6979 getPointerTy(), StackSlot, WordOff);
6980 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6981 StackSlot, MachinePointerInfo(),
6983 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6984 OffsetSlot, MachinePointerInfo(),
6986 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6990 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6991 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6992 StackSlot, MachinePointerInfo(),
6994 // For i64 source, we need to add the appropriate power of 2 if the input
6995 // was negative. This is the same as the optimization in
6996 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6997 // we must be careful to do the computation in x87 extended precision, not
6998 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6999 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7000 MachineMemOperand *MMO =
7001 DAG.getMachineFunction()
7002 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7003 MachineMemOperand::MOLoad, 8, 8);
7005 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7006 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7007 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7010 APInt FF(32, 0x5F800000ULL);
7012 // Check whether the sign bit is set.
7013 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7014 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7017 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7018 SDValue FudgePtr = DAG.getConstantPool(
7019 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7022 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7023 SDValue Zero = DAG.getIntPtrConstant(0);
7024 SDValue Four = DAG.getIntPtrConstant(4);
7025 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7027 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7029 // Load the value out, extending it from f32 to f80.
7030 // FIXME: Avoid the extend by constructing the right constant pool?
7031 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7032 FudgePtr, MachinePointerInfo::getConstantPool(),
7033 MVT::f32, false, false, 4);
7034 // Extend everything to 80 bits to force it to be done on x87.
7035 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7036 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7039 std::pair<SDValue,SDValue> X86TargetLowering::
7040 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7041 DebugLoc DL = Op.getDebugLoc();
7043 EVT DstTy = Op.getValueType();
7046 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7050 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7051 DstTy.getSimpleVT() >= MVT::i16 &&
7052 "Unknown FP_TO_SINT to lower!");
7054 // These are really Legal.
7055 if (DstTy == MVT::i32 &&
7056 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7057 return std::make_pair(SDValue(), SDValue());
7058 if (Subtarget->is64Bit() &&
7059 DstTy == MVT::i64 &&
7060 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7061 return std::make_pair(SDValue(), SDValue());
7063 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7065 MachineFunction &MF = DAG.getMachineFunction();
7066 unsigned MemSize = DstTy.getSizeInBits()/8;
7067 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7068 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7073 switch (DstTy.getSimpleVT().SimpleTy) {
7074 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7075 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7076 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7077 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7080 SDValue Chain = DAG.getEntryNode();
7081 SDValue Value = Op.getOperand(0);
7082 EVT TheVT = Op.getOperand(0).getValueType();
7083 if (isScalarFPTypeInSSEReg(TheVT)) {
7084 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7085 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7086 MachinePointerInfo::getFixedStack(SSFI),
7088 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7090 Chain, StackSlot, DAG.getValueType(TheVT)
7093 MachineMemOperand *MMO =
7094 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7095 MachineMemOperand::MOLoad, MemSize, MemSize);
7096 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7098 Chain = Value.getValue(1);
7099 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7100 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7103 MachineMemOperand *MMO =
7104 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7105 MachineMemOperand::MOStore, MemSize, MemSize);
7107 // Build the FP_TO_INT*_IN_MEM
7108 SDValue Ops[] = { Chain, Value, StackSlot };
7109 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7110 Ops, 3, DstTy, MMO);
7112 return std::make_pair(FIST, StackSlot);
7115 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7116 SelectionDAG &DAG) const {
7117 if (Op.getValueType().isVector())
7120 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7121 SDValue FIST = Vals.first, StackSlot = Vals.second;
7122 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7123 if (FIST.getNode() == 0) return Op;
7126 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7127 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7130 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7131 SelectionDAG &DAG) const {
7132 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7133 SDValue FIST = Vals.first, StackSlot = Vals.second;
7134 assert(FIST.getNode() && "Unexpected failure");
7137 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7138 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7141 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7142 SelectionDAG &DAG) const {
7143 LLVMContext *Context = DAG.getContext();
7144 DebugLoc dl = Op.getDebugLoc();
7145 EVT VT = Op.getValueType();
7148 EltVT = VT.getVectorElementType();
7149 std::vector<Constant*> CV;
7150 if (EltVT == MVT::f64) {
7151 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7155 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7161 Constant *C = ConstantVector::get(CV);
7162 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7163 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7164 MachinePointerInfo::getConstantPool(),
7166 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7169 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7170 LLVMContext *Context = DAG.getContext();
7171 DebugLoc dl = Op.getDebugLoc();
7172 EVT VT = Op.getValueType();
7175 EltVT = VT.getVectorElementType();
7176 std::vector<Constant*> CV;
7177 if (EltVT == MVT::f64) {
7178 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7182 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7188 Constant *C = ConstantVector::get(CV);
7189 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7190 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7191 MachinePointerInfo::getConstantPool(),
7193 if (VT.isVector()) {
7194 return DAG.getNode(ISD::BITCAST, dl, VT,
7195 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7196 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7198 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7200 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7204 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7205 LLVMContext *Context = DAG.getContext();
7206 SDValue Op0 = Op.getOperand(0);
7207 SDValue Op1 = Op.getOperand(1);
7208 DebugLoc dl = Op.getDebugLoc();
7209 EVT VT = Op.getValueType();
7210 EVT SrcVT = Op1.getValueType();
7212 // If second operand is smaller, extend it first.
7213 if (SrcVT.bitsLT(VT)) {
7214 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7217 // And if it is bigger, shrink it first.
7218 if (SrcVT.bitsGT(VT)) {
7219 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7223 // At this point the operands and the result should have the same
7224 // type, and that won't be f80 since that is not custom lowered.
7226 // First get the sign bit of second operand.
7227 std::vector<Constant*> CV;
7228 if (SrcVT == MVT::f64) {
7229 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7230 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7232 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7233 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7234 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7235 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7237 Constant *C = ConstantVector::get(CV);
7238 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7239 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7240 MachinePointerInfo::getConstantPool(),
7242 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7244 // Shift sign bit right or left if the two operands have different types.
7245 if (SrcVT.bitsGT(VT)) {
7246 // Op0 is MVT::f32, Op1 is MVT::f64.
7247 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7248 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7249 DAG.getConstant(32, MVT::i32));
7250 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7251 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7252 DAG.getIntPtrConstant(0));
7255 // Clear first operand sign bit.
7257 if (VT == MVT::f64) {
7258 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7259 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7261 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7262 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7263 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7264 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7266 C = ConstantVector::get(CV);
7267 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7268 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7269 MachinePointerInfo::getConstantPool(),
7271 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7273 // Or the value with the sign bit.
7274 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7277 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7278 SDValue N0 = Op.getOperand(0);
7279 DebugLoc dl = Op.getDebugLoc();
7280 EVT VT = Op.getValueType();
7282 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7283 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7284 DAG.getConstant(1, VT));
7285 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7288 /// Emit nodes that will be selected as "test Op0,Op0", or something
7290 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7291 SelectionDAG &DAG) const {
7292 DebugLoc dl = Op.getDebugLoc();
7294 // CF and OF aren't always set the way we want. Determine which
7295 // of these we need.
7296 bool NeedCF = false;
7297 bool NeedOF = false;
7300 case X86::COND_A: case X86::COND_AE:
7301 case X86::COND_B: case X86::COND_BE:
7304 case X86::COND_G: case X86::COND_GE:
7305 case X86::COND_L: case X86::COND_LE:
7306 case X86::COND_O: case X86::COND_NO:
7311 // See if we can use the EFLAGS value from the operand instead of
7312 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7313 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7314 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7315 // Emit a CMP with 0, which is the TEST pattern.
7316 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7317 DAG.getConstant(0, Op.getValueType()));
7319 unsigned Opcode = 0;
7320 unsigned NumOperands = 0;
7321 switch (Op.getNode()->getOpcode()) {
7323 // Due to an isel shortcoming, be conservative if this add is likely to be
7324 // selected as part of a load-modify-store instruction. When the root node
7325 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7326 // uses of other nodes in the match, such as the ADD in this case. This
7327 // leads to the ADD being left around and reselected, with the result being
7328 // two adds in the output. Alas, even if none our users are stores, that
7329 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7330 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7331 // climbing the DAG back to the root, and it doesn't seem to be worth the
7333 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7334 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7335 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7338 if (ConstantSDNode *C =
7339 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7340 // An add of one will be selected as an INC.
7341 if (C->getAPIntValue() == 1) {
7342 Opcode = X86ISD::INC;
7347 // An add of negative one (subtract of one) will be selected as a DEC.
7348 if (C->getAPIntValue().isAllOnesValue()) {
7349 Opcode = X86ISD::DEC;
7355 // Otherwise use a regular EFLAGS-setting add.
7356 Opcode = X86ISD::ADD;
7360 // If the primary and result isn't used, don't bother using X86ISD::AND,
7361 // because a TEST instruction will be better.
7362 bool NonFlagUse = false;
7363 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7364 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7366 unsigned UOpNo = UI.getOperandNo();
7367 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7368 // Look pass truncate.
7369 UOpNo = User->use_begin().getOperandNo();
7370 User = *User->use_begin();
7373 if (User->getOpcode() != ISD::BRCOND &&
7374 User->getOpcode() != ISD::SETCC &&
7375 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7388 // Due to the ISEL shortcoming noted above, be conservative if this op is
7389 // likely to be selected as part of a load-modify-store instruction.
7390 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7391 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7392 if (UI->getOpcode() == ISD::STORE)
7395 // Otherwise use a regular EFLAGS-setting instruction.
7396 switch (Op.getNode()->getOpcode()) {
7397 default: llvm_unreachable("unexpected operator!");
7398 case ISD::SUB: Opcode = X86ISD::SUB; break;
7399 case ISD::OR: Opcode = X86ISD::OR; break;
7400 case ISD::XOR: Opcode = X86ISD::XOR; break;
7401 case ISD::AND: Opcode = X86ISD::AND; break;
7413 return SDValue(Op.getNode(), 1);
7420 // Emit a CMP with 0, which is the TEST pattern.
7421 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7422 DAG.getConstant(0, Op.getValueType()));
7424 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7425 SmallVector<SDValue, 4> Ops;
7426 for (unsigned i = 0; i != NumOperands; ++i)
7427 Ops.push_back(Op.getOperand(i));
7429 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7430 DAG.ReplaceAllUsesWith(Op, New);
7431 return SDValue(New.getNode(), 1);
7434 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7436 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7437 SelectionDAG &DAG) const {
7438 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7439 if (C->getAPIntValue() == 0)
7440 return EmitTest(Op0, X86CC, DAG);
7442 DebugLoc dl = Op0.getDebugLoc();
7443 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7446 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7447 /// if it's possible.
7448 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7449 DebugLoc dl, SelectionDAG &DAG) const {
7450 SDValue Op0 = And.getOperand(0);
7451 SDValue Op1 = And.getOperand(1);
7452 if (Op0.getOpcode() == ISD::TRUNCATE)
7453 Op0 = Op0.getOperand(0);
7454 if (Op1.getOpcode() == ISD::TRUNCATE)
7455 Op1 = Op1.getOperand(0);
7458 if (Op1.getOpcode() == ISD::SHL)
7459 std::swap(Op0, Op1);
7460 if (Op0.getOpcode() == ISD::SHL) {
7461 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7462 if (And00C->getZExtValue() == 1) {
7463 // If we looked past a truncate, check that it's only truncating away
7465 unsigned BitWidth = Op0.getValueSizeInBits();
7466 unsigned AndBitWidth = And.getValueSizeInBits();
7467 if (BitWidth > AndBitWidth) {
7468 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7469 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7470 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7474 RHS = Op0.getOperand(1);
7476 } else if (Op1.getOpcode() == ISD::Constant) {
7477 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7478 SDValue AndLHS = Op0;
7479 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7480 LHS = AndLHS.getOperand(0);
7481 RHS = AndLHS.getOperand(1);
7485 if (LHS.getNode()) {
7486 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7487 // instruction. Since the shift amount is in-range-or-undefined, we know
7488 // that doing a bittest on the i32 value is ok. We extend to i32 because
7489 // the encoding for the i16 version is larger than the i32 version.
7490 // Also promote i16 to i32 for performance / code size reason.
7491 if (LHS.getValueType() == MVT::i8 ||
7492 LHS.getValueType() == MVT::i16)
7493 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7495 // If the operand types disagree, extend the shift amount to match. Since
7496 // BT ignores high bits (like shifts) we can use anyextend.
7497 if (LHS.getValueType() != RHS.getValueType())
7498 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7500 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7501 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7502 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7503 DAG.getConstant(Cond, MVT::i8), BT);
7509 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7510 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7511 SDValue Op0 = Op.getOperand(0);
7512 SDValue Op1 = Op.getOperand(1);
7513 DebugLoc dl = Op.getDebugLoc();
7514 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7516 // Optimize to BT if possible.
7517 // Lower (X & (1 << N)) == 0 to BT(X, N).
7518 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7519 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7520 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7521 Op1.getOpcode() == ISD::Constant &&
7522 cast<ConstantSDNode>(Op1)->isNullValue() &&
7523 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7524 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7525 if (NewSetCC.getNode())
7529 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7531 if (Op1.getOpcode() == ISD::Constant &&
7532 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7533 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7534 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7536 // If the input is a setcc, then reuse the input setcc or use a new one with
7537 // the inverted condition.
7538 if (Op0.getOpcode() == X86ISD::SETCC) {
7539 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7540 bool Invert = (CC == ISD::SETNE) ^
7541 cast<ConstantSDNode>(Op1)->isNullValue();
7542 if (!Invert) return Op0;
7544 CCode = X86::GetOppositeBranchCondition(CCode);
7545 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7546 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7550 bool isFP = Op1.getValueType().isFloatingPoint();
7551 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7552 if (X86CC == X86::COND_INVALID)
7555 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7556 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7557 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7560 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7562 SDValue Op0 = Op.getOperand(0);
7563 SDValue Op1 = Op.getOperand(1);
7564 SDValue CC = Op.getOperand(2);
7565 EVT VT = Op.getValueType();
7566 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7567 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7568 DebugLoc dl = Op.getDebugLoc();
7572 EVT VT0 = Op0.getValueType();
7573 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7574 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7577 switch (SetCCOpcode) {
7580 case ISD::SETEQ: SSECC = 0; break;
7582 case ISD::SETGT: Swap = true; // Fallthrough
7584 case ISD::SETOLT: SSECC = 1; break;
7586 case ISD::SETGE: Swap = true; // Fallthrough
7588 case ISD::SETOLE: SSECC = 2; break;
7589 case ISD::SETUO: SSECC = 3; break;
7591 case ISD::SETNE: SSECC = 4; break;
7592 case ISD::SETULE: Swap = true;
7593 case ISD::SETUGE: SSECC = 5; break;
7594 case ISD::SETULT: Swap = true;
7595 case ISD::SETUGT: SSECC = 6; break;
7596 case ISD::SETO: SSECC = 7; break;
7599 std::swap(Op0, Op1);
7601 // In the two special cases we can't handle, emit two comparisons.
7603 if (SetCCOpcode == ISD::SETUEQ) {
7605 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7606 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7607 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7609 else if (SetCCOpcode == ISD::SETONE) {
7611 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7612 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7613 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7615 llvm_unreachable("Illegal FP comparison");
7617 // Handle all other FP comparisons here.
7618 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7621 // We are handling one of the integer comparisons here. Since SSE only has
7622 // GT and EQ comparisons for integer, swapping operands and multiple
7623 // operations may be required for some comparisons.
7624 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7625 bool Swap = false, Invert = false, FlipSigns = false;
7627 switch (VT.getSimpleVT().SimpleTy) {
7629 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7630 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7631 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7632 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7635 switch (SetCCOpcode) {
7637 case ISD::SETNE: Invert = true;
7638 case ISD::SETEQ: Opc = EQOpc; break;
7639 case ISD::SETLT: Swap = true;
7640 case ISD::SETGT: Opc = GTOpc; break;
7641 case ISD::SETGE: Swap = true;
7642 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7643 case ISD::SETULT: Swap = true;
7644 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7645 case ISD::SETUGE: Swap = true;
7646 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7649 std::swap(Op0, Op1);
7651 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7652 // bits of the inputs before performing those operations.
7654 EVT EltVT = VT.getVectorElementType();
7655 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7657 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7658 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7660 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7661 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7664 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7666 // If the logical-not of the result is required, perform that now.
7668 Result = DAG.getNOT(dl, Result, VT);
7673 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7674 static bool isX86LogicalCmp(SDValue Op) {
7675 unsigned Opc = Op.getNode()->getOpcode();
7676 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7678 if (Op.getResNo() == 1 &&
7679 (Opc == X86ISD::ADD ||
7680 Opc == X86ISD::SUB ||
7681 Opc == X86ISD::ADC ||
7682 Opc == X86ISD::SBB ||
7683 Opc == X86ISD::SMUL ||
7684 Opc == X86ISD::UMUL ||
7685 Opc == X86ISD::INC ||
7686 Opc == X86ISD::DEC ||
7687 Opc == X86ISD::OR ||
7688 Opc == X86ISD::XOR ||
7689 Opc == X86ISD::AND))
7692 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7698 static bool isZero(SDValue V) {
7699 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7700 return C && C->isNullValue();
7703 static bool isAllOnes(SDValue V) {
7704 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7705 return C && C->isAllOnesValue();
7708 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7709 bool addTest = true;
7710 SDValue Cond = Op.getOperand(0);
7711 SDValue Op1 = Op.getOperand(1);
7712 SDValue Op2 = Op.getOperand(2);
7713 DebugLoc DL = Op.getDebugLoc();
7716 if (Cond.getOpcode() == ISD::SETCC) {
7717 SDValue NewCond = LowerSETCC(Cond, DAG);
7718 if (NewCond.getNode())
7722 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7723 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7724 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7725 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7726 if (Cond.getOpcode() == X86ISD::SETCC &&
7727 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7728 isZero(Cond.getOperand(1).getOperand(1))) {
7729 SDValue Cmp = Cond.getOperand(1);
7731 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7733 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7734 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7735 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7737 SDValue CmpOp0 = Cmp.getOperand(0);
7738 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7739 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7741 SDValue Res = // Res = 0 or -1.
7742 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7743 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7745 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7746 Res = DAG.getNOT(DL, Res, Res.getValueType());
7748 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7749 if (N2C == 0 || !N2C->isNullValue())
7750 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7755 // Look past (and (setcc_carry (cmp ...)), 1).
7756 if (Cond.getOpcode() == ISD::AND &&
7757 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7758 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7759 if (C && C->getAPIntValue() == 1)
7760 Cond = Cond.getOperand(0);
7763 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7764 // setting operand in place of the X86ISD::SETCC.
7765 if (Cond.getOpcode() == X86ISD::SETCC ||
7766 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7767 CC = Cond.getOperand(0);
7769 SDValue Cmp = Cond.getOperand(1);
7770 unsigned Opc = Cmp.getOpcode();
7771 EVT VT = Op.getValueType();
7773 bool IllegalFPCMov = false;
7774 if (VT.isFloatingPoint() && !VT.isVector() &&
7775 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7776 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7778 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7779 Opc == X86ISD::BT) { // FIXME
7786 // Look pass the truncate.
7787 if (Cond.getOpcode() == ISD::TRUNCATE)
7788 Cond = Cond.getOperand(0);
7790 // We know the result of AND is compared against zero. Try to match
7792 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7793 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7794 if (NewSetCC.getNode()) {
7795 CC = NewSetCC.getOperand(0);
7796 Cond = NewSetCC.getOperand(1);
7803 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7804 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7807 // a < b ? -1 : 0 -> RES = ~setcc_carry
7808 // a < b ? 0 : -1 -> RES = setcc_carry
7809 // a >= b ? -1 : 0 -> RES = setcc_carry
7810 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7811 if (Cond.getOpcode() == X86ISD::CMP) {
7812 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7814 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7815 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7816 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7817 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7818 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7819 return DAG.getNOT(DL, Res, Res.getValueType());
7824 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7825 // condition is true.
7826 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7827 SDValue Ops[] = { Op2, Op1, CC, Cond };
7828 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7831 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7832 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7833 // from the AND / OR.
7834 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7835 Opc = Op.getOpcode();
7836 if (Opc != ISD::OR && Opc != ISD::AND)
7838 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7839 Op.getOperand(0).hasOneUse() &&
7840 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7841 Op.getOperand(1).hasOneUse());
7844 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7845 // 1 and that the SETCC node has a single use.
7846 static bool isXor1OfSetCC(SDValue Op) {
7847 if (Op.getOpcode() != ISD::XOR)
7849 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7850 if (N1C && N1C->getAPIntValue() == 1) {
7851 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7852 Op.getOperand(0).hasOneUse();
7857 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7858 bool addTest = true;
7859 SDValue Chain = Op.getOperand(0);
7860 SDValue Cond = Op.getOperand(1);
7861 SDValue Dest = Op.getOperand(2);
7862 DebugLoc dl = Op.getDebugLoc();
7865 if (Cond.getOpcode() == ISD::SETCC) {
7866 SDValue NewCond = LowerSETCC(Cond, DAG);
7867 if (NewCond.getNode())
7871 // FIXME: LowerXALUO doesn't handle these!!
7872 else if (Cond.getOpcode() == X86ISD::ADD ||
7873 Cond.getOpcode() == X86ISD::SUB ||
7874 Cond.getOpcode() == X86ISD::SMUL ||
7875 Cond.getOpcode() == X86ISD::UMUL)
7876 Cond = LowerXALUO(Cond, DAG);
7879 // Look pass (and (setcc_carry (cmp ...)), 1).
7880 if (Cond.getOpcode() == ISD::AND &&
7881 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7882 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7883 if (C && C->getAPIntValue() == 1)
7884 Cond = Cond.getOperand(0);
7887 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7888 // setting operand in place of the X86ISD::SETCC.
7889 if (Cond.getOpcode() == X86ISD::SETCC ||
7890 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7891 CC = Cond.getOperand(0);
7893 SDValue Cmp = Cond.getOperand(1);
7894 unsigned Opc = Cmp.getOpcode();
7895 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7896 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7900 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7904 // These can only come from an arithmetic instruction with overflow,
7905 // e.g. SADDO, UADDO.
7906 Cond = Cond.getNode()->getOperand(1);
7913 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7914 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7915 if (CondOpc == ISD::OR) {
7916 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7917 // two branches instead of an explicit OR instruction with a
7919 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7920 isX86LogicalCmp(Cmp)) {
7921 CC = Cond.getOperand(0).getOperand(0);
7922 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7923 Chain, Dest, CC, Cmp);
7924 CC = Cond.getOperand(1).getOperand(0);
7928 } else { // ISD::AND
7929 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7930 // two branches instead of an explicit AND instruction with a
7931 // separate test. However, we only do this if this block doesn't
7932 // have a fall-through edge, because this requires an explicit
7933 // jmp when the condition is false.
7934 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7935 isX86LogicalCmp(Cmp) &&
7936 Op.getNode()->hasOneUse()) {
7937 X86::CondCode CCode =
7938 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7939 CCode = X86::GetOppositeBranchCondition(CCode);
7940 CC = DAG.getConstant(CCode, MVT::i8);
7941 SDNode *User = *Op.getNode()->use_begin();
7942 // Look for an unconditional branch following this conditional branch.
7943 // We need this because we need to reverse the successors in order
7944 // to implement FCMP_OEQ.
7945 if (User->getOpcode() == ISD::BR) {
7946 SDValue FalseBB = User->getOperand(1);
7948 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7949 assert(NewBR == User);
7953 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7954 Chain, Dest, CC, Cmp);
7955 X86::CondCode CCode =
7956 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7957 CCode = X86::GetOppositeBranchCondition(CCode);
7958 CC = DAG.getConstant(CCode, MVT::i8);
7964 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7965 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7966 // It should be transformed during dag combiner except when the condition
7967 // is set by a arithmetics with overflow node.
7968 X86::CondCode CCode =
7969 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7970 CCode = X86::GetOppositeBranchCondition(CCode);
7971 CC = DAG.getConstant(CCode, MVT::i8);
7972 Cond = Cond.getOperand(0).getOperand(1);
7978 // Look pass the truncate.
7979 if (Cond.getOpcode() == ISD::TRUNCATE)
7980 Cond = Cond.getOperand(0);
7982 // We know the result of AND is compared against zero. Try to match
7984 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7985 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7986 if (NewSetCC.getNode()) {
7987 CC = NewSetCC.getOperand(0);
7988 Cond = NewSetCC.getOperand(1);
7995 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7996 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7998 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7999 Chain, Dest, CC, Cond);
8003 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8004 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8005 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8006 // that the guard pages used by the OS virtual memory manager are allocated in
8007 // correct sequence.
8009 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8010 SelectionDAG &DAG) const {
8011 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8012 "This should be used only on Windows targets");
8013 assert(!Subtarget->isTargetEnvMacho());
8014 DebugLoc dl = Op.getDebugLoc();
8017 SDValue Chain = Op.getOperand(0);
8018 SDValue Size = Op.getOperand(1);
8019 // FIXME: Ensure alignment here
8023 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8024 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8026 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8027 Flag = Chain.getValue(1);
8029 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8031 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8032 Flag = Chain.getValue(1);
8034 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8036 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8037 return DAG.getMergeValues(Ops1, 2, dl);
8040 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8041 MachineFunction &MF = DAG.getMachineFunction();
8042 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8044 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8045 DebugLoc DL = Op.getDebugLoc();
8047 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8048 // vastart just stores the address of the VarArgsFrameIndex slot into the
8049 // memory location argument.
8050 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8052 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8053 MachinePointerInfo(SV), false, false, 0);
8057 // gp_offset (0 - 6 * 8)
8058 // fp_offset (48 - 48 + 8 * 16)
8059 // overflow_arg_area (point to parameters coming in memory).
8061 SmallVector<SDValue, 8> MemOps;
8062 SDValue FIN = Op.getOperand(1);
8064 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8065 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8067 FIN, MachinePointerInfo(SV), false, false, 0);
8068 MemOps.push_back(Store);
8071 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8072 FIN, DAG.getIntPtrConstant(4));
8073 Store = DAG.getStore(Op.getOperand(0), DL,
8074 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8076 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8077 MemOps.push_back(Store);
8079 // Store ptr to overflow_arg_area
8080 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8081 FIN, DAG.getIntPtrConstant(4));
8082 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8084 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8085 MachinePointerInfo(SV, 8),
8087 MemOps.push_back(Store);
8089 // Store ptr to reg_save_area.
8090 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8091 FIN, DAG.getIntPtrConstant(8));
8092 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8094 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8095 MachinePointerInfo(SV, 16), false, false, 0);
8096 MemOps.push_back(Store);
8097 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8098 &MemOps[0], MemOps.size());
8101 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8102 assert(Subtarget->is64Bit() &&
8103 "LowerVAARG only handles 64-bit va_arg!");
8104 assert((Subtarget->isTargetLinux() ||
8105 Subtarget->isTargetDarwin()) &&
8106 "Unhandled target in LowerVAARG");
8107 assert(Op.getNode()->getNumOperands() == 4);
8108 SDValue Chain = Op.getOperand(0);
8109 SDValue SrcPtr = Op.getOperand(1);
8110 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8111 unsigned Align = Op.getConstantOperandVal(3);
8112 DebugLoc dl = Op.getDebugLoc();
8114 EVT ArgVT = Op.getNode()->getValueType(0);
8115 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8116 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8119 // Decide which area this value should be read from.
8120 // TODO: Implement the AMD64 ABI in its entirety. This simple
8121 // selection mechanism works only for the basic types.
8122 if (ArgVT == MVT::f80) {
8123 llvm_unreachable("va_arg for f80 not yet implemented");
8124 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8125 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8126 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8127 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8129 llvm_unreachable("Unhandled argument type in LowerVAARG");
8133 // Sanity Check: Make sure using fp_offset makes sense.
8134 assert(!UseSoftFloat &&
8135 !(DAG.getMachineFunction()
8136 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8137 Subtarget->hasXMM());
8140 // Insert VAARG_64 node into the DAG
8141 // VAARG_64 returns two values: Variable Argument Address, Chain
8142 SmallVector<SDValue, 11> InstOps;
8143 InstOps.push_back(Chain);
8144 InstOps.push_back(SrcPtr);
8145 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8146 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8147 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8148 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8149 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8150 VTs, &InstOps[0], InstOps.size(),
8152 MachinePointerInfo(SV),
8157 Chain = VAARG.getValue(1);
8159 // Load the next argument and return it
8160 return DAG.getLoad(ArgVT, dl,
8163 MachinePointerInfo(),
8167 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8168 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8169 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8170 SDValue Chain = Op.getOperand(0);
8171 SDValue DstPtr = Op.getOperand(1);
8172 SDValue SrcPtr = Op.getOperand(2);
8173 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8174 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8175 DebugLoc DL = Op.getDebugLoc();
8177 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8178 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8180 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8184 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8185 DebugLoc dl = Op.getDebugLoc();
8186 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8188 default: return SDValue(); // Don't custom lower most intrinsics.
8189 // Comparison intrinsics.
8190 case Intrinsic::x86_sse_comieq_ss:
8191 case Intrinsic::x86_sse_comilt_ss:
8192 case Intrinsic::x86_sse_comile_ss:
8193 case Intrinsic::x86_sse_comigt_ss:
8194 case Intrinsic::x86_sse_comige_ss:
8195 case Intrinsic::x86_sse_comineq_ss:
8196 case Intrinsic::x86_sse_ucomieq_ss:
8197 case Intrinsic::x86_sse_ucomilt_ss:
8198 case Intrinsic::x86_sse_ucomile_ss:
8199 case Intrinsic::x86_sse_ucomigt_ss:
8200 case Intrinsic::x86_sse_ucomige_ss:
8201 case Intrinsic::x86_sse_ucomineq_ss:
8202 case Intrinsic::x86_sse2_comieq_sd:
8203 case Intrinsic::x86_sse2_comilt_sd:
8204 case Intrinsic::x86_sse2_comile_sd:
8205 case Intrinsic::x86_sse2_comigt_sd:
8206 case Intrinsic::x86_sse2_comige_sd:
8207 case Intrinsic::x86_sse2_comineq_sd:
8208 case Intrinsic::x86_sse2_ucomieq_sd:
8209 case Intrinsic::x86_sse2_ucomilt_sd:
8210 case Intrinsic::x86_sse2_ucomile_sd:
8211 case Intrinsic::x86_sse2_ucomigt_sd:
8212 case Intrinsic::x86_sse2_ucomige_sd:
8213 case Intrinsic::x86_sse2_ucomineq_sd: {
8215 ISD::CondCode CC = ISD::SETCC_INVALID;
8218 case Intrinsic::x86_sse_comieq_ss:
8219 case Intrinsic::x86_sse2_comieq_sd:
8223 case Intrinsic::x86_sse_comilt_ss:
8224 case Intrinsic::x86_sse2_comilt_sd:
8228 case Intrinsic::x86_sse_comile_ss:
8229 case Intrinsic::x86_sse2_comile_sd:
8233 case Intrinsic::x86_sse_comigt_ss:
8234 case Intrinsic::x86_sse2_comigt_sd:
8238 case Intrinsic::x86_sse_comige_ss:
8239 case Intrinsic::x86_sse2_comige_sd:
8243 case Intrinsic::x86_sse_comineq_ss:
8244 case Intrinsic::x86_sse2_comineq_sd:
8248 case Intrinsic::x86_sse_ucomieq_ss:
8249 case Intrinsic::x86_sse2_ucomieq_sd:
8250 Opc = X86ISD::UCOMI;
8253 case Intrinsic::x86_sse_ucomilt_ss:
8254 case Intrinsic::x86_sse2_ucomilt_sd:
8255 Opc = X86ISD::UCOMI;
8258 case Intrinsic::x86_sse_ucomile_ss:
8259 case Intrinsic::x86_sse2_ucomile_sd:
8260 Opc = X86ISD::UCOMI;
8263 case Intrinsic::x86_sse_ucomigt_ss:
8264 case Intrinsic::x86_sse2_ucomigt_sd:
8265 Opc = X86ISD::UCOMI;
8268 case Intrinsic::x86_sse_ucomige_ss:
8269 case Intrinsic::x86_sse2_ucomige_sd:
8270 Opc = X86ISD::UCOMI;
8273 case Intrinsic::x86_sse_ucomineq_ss:
8274 case Intrinsic::x86_sse2_ucomineq_sd:
8275 Opc = X86ISD::UCOMI;
8280 SDValue LHS = Op.getOperand(1);
8281 SDValue RHS = Op.getOperand(2);
8282 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8283 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8284 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8285 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8286 DAG.getConstant(X86CC, MVT::i8), Cond);
8287 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8289 // ptest and testp intrinsics. The intrinsic these come from are designed to
8290 // return an integer value, not just an instruction so lower it to the ptest
8291 // or testp pattern and a setcc for the result.
8292 case Intrinsic::x86_sse41_ptestz:
8293 case Intrinsic::x86_sse41_ptestc:
8294 case Intrinsic::x86_sse41_ptestnzc:
8295 case Intrinsic::x86_avx_ptestz_256:
8296 case Intrinsic::x86_avx_ptestc_256:
8297 case Intrinsic::x86_avx_ptestnzc_256:
8298 case Intrinsic::x86_avx_vtestz_ps:
8299 case Intrinsic::x86_avx_vtestc_ps:
8300 case Intrinsic::x86_avx_vtestnzc_ps:
8301 case Intrinsic::x86_avx_vtestz_pd:
8302 case Intrinsic::x86_avx_vtestc_pd:
8303 case Intrinsic::x86_avx_vtestnzc_pd:
8304 case Intrinsic::x86_avx_vtestz_ps_256:
8305 case Intrinsic::x86_avx_vtestc_ps_256:
8306 case Intrinsic::x86_avx_vtestnzc_ps_256:
8307 case Intrinsic::x86_avx_vtestz_pd_256:
8308 case Intrinsic::x86_avx_vtestc_pd_256:
8309 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8310 bool IsTestPacked = false;
8313 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8314 case Intrinsic::x86_avx_vtestz_ps:
8315 case Intrinsic::x86_avx_vtestz_pd:
8316 case Intrinsic::x86_avx_vtestz_ps_256:
8317 case Intrinsic::x86_avx_vtestz_pd_256:
8318 IsTestPacked = true; // Fallthrough
8319 case Intrinsic::x86_sse41_ptestz:
8320 case Intrinsic::x86_avx_ptestz_256:
8322 X86CC = X86::COND_E;
8324 case Intrinsic::x86_avx_vtestc_ps:
8325 case Intrinsic::x86_avx_vtestc_pd:
8326 case Intrinsic::x86_avx_vtestc_ps_256:
8327 case Intrinsic::x86_avx_vtestc_pd_256:
8328 IsTestPacked = true; // Fallthrough
8329 case Intrinsic::x86_sse41_ptestc:
8330 case Intrinsic::x86_avx_ptestc_256:
8332 X86CC = X86::COND_B;
8334 case Intrinsic::x86_avx_vtestnzc_ps:
8335 case Intrinsic::x86_avx_vtestnzc_pd:
8336 case Intrinsic::x86_avx_vtestnzc_ps_256:
8337 case Intrinsic::x86_avx_vtestnzc_pd_256:
8338 IsTestPacked = true; // Fallthrough
8339 case Intrinsic::x86_sse41_ptestnzc:
8340 case Intrinsic::x86_avx_ptestnzc_256:
8342 X86CC = X86::COND_A;
8346 SDValue LHS = Op.getOperand(1);
8347 SDValue RHS = Op.getOperand(2);
8348 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8349 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8350 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8351 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8352 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8355 // Fix vector shift instructions where the last operand is a non-immediate
8357 case Intrinsic::x86_sse2_pslli_w:
8358 case Intrinsic::x86_sse2_pslli_d:
8359 case Intrinsic::x86_sse2_pslli_q:
8360 case Intrinsic::x86_sse2_psrli_w:
8361 case Intrinsic::x86_sse2_psrli_d:
8362 case Intrinsic::x86_sse2_psrli_q:
8363 case Intrinsic::x86_sse2_psrai_w:
8364 case Intrinsic::x86_sse2_psrai_d:
8365 case Intrinsic::x86_mmx_pslli_w:
8366 case Intrinsic::x86_mmx_pslli_d:
8367 case Intrinsic::x86_mmx_pslli_q:
8368 case Intrinsic::x86_mmx_psrli_w:
8369 case Intrinsic::x86_mmx_psrli_d:
8370 case Intrinsic::x86_mmx_psrli_q:
8371 case Intrinsic::x86_mmx_psrai_w:
8372 case Intrinsic::x86_mmx_psrai_d: {
8373 SDValue ShAmt = Op.getOperand(2);
8374 if (isa<ConstantSDNode>(ShAmt))
8377 unsigned NewIntNo = 0;
8378 EVT ShAmtVT = MVT::v4i32;
8380 case Intrinsic::x86_sse2_pslli_w:
8381 NewIntNo = Intrinsic::x86_sse2_psll_w;
8383 case Intrinsic::x86_sse2_pslli_d:
8384 NewIntNo = Intrinsic::x86_sse2_psll_d;
8386 case Intrinsic::x86_sse2_pslli_q:
8387 NewIntNo = Intrinsic::x86_sse2_psll_q;
8389 case Intrinsic::x86_sse2_psrli_w:
8390 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8392 case Intrinsic::x86_sse2_psrli_d:
8393 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8395 case Intrinsic::x86_sse2_psrli_q:
8396 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8398 case Intrinsic::x86_sse2_psrai_w:
8399 NewIntNo = Intrinsic::x86_sse2_psra_w;
8401 case Intrinsic::x86_sse2_psrai_d:
8402 NewIntNo = Intrinsic::x86_sse2_psra_d;
8405 ShAmtVT = MVT::v2i32;
8407 case Intrinsic::x86_mmx_pslli_w:
8408 NewIntNo = Intrinsic::x86_mmx_psll_w;
8410 case Intrinsic::x86_mmx_pslli_d:
8411 NewIntNo = Intrinsic::x86_mmx_psll_d;
8413 case Intrinsic::x86_mmx_pslli_q:
8414 NewIntNo = Intrinsic::x86_mmx_psll_q;
8416 case Intrinsic::x86_mmx_psrli_w:
8417 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8419 case Intrinsic::x86_mmx_psrli_d:
8420 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8422 case Intrinsic::x86_mmx_psrli_q:
8423 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8425 case Intrinsic::x86_mmx_psrai_w:
8426 NewIntNo = Intrinsic::x86_mmx_psra_w;
8428 case Intrinsic::x86_mmx_psrai_d:
8429 NewIntNo = Intrinsic::x86_mmx_psra_d;
8431 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8437 // The vector shift intrinsics with scalars uses 32b shift amounts but
8438 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8442 ShOps[1] = DAG.getConstant(0, MVT::i32);
8443 if (ShAmtVT == MVT::v4i32) {
8444 ShOps[2] = DAG.getUNDEF(MVT::i32);
8445 ShOps[3] = DAG.getUNDEF(MVT::i32);
8446 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8448 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8449 // FIXME this must be lowered to get rid of the invalid type.
8452 EVT VT = Op.getValueType();
8453 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8454 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8455 DAG.getConstant(NewIntNo, MVT::i32),
8456 Op.getOperand(1), ShAmt);
8461 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8462 SelectionDAG &DAG) const {
8463 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8464 MFI->setReturnAddressIsTaken(true);
8466 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8467 DebugLoc dl = Op.getDebugLoc();
8470 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8472 DAG.getConstant(TD->getPointerSize(),
8473 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8474 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8475 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8477 MachinePointerInfo(), false, false, 0);
8480 // Just load the return address.
8481 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8482 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8483 RetAddrFI, MachinePointerInfo(), false, false, 0);
8486 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8487 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8488 MFI->setFrameAddressIsTaken(true);
8490 EVT VT = Op.getValueType();
8491 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8492 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8493 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8494 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8496 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8497 MachinePointerInfo(),
8502 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8503 SelectionDAG &DAG) const {
8504 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8507 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8508 MachineFunction &MF = DAG.getMachineFunction();
8509 SDValue Chain = Op.getOperand(0);
8510 SDValue Offset = Op.getOperand(1);
8511 SDValue Handler = Op.getOperand(2);
8512 DebugLoc dl = Op.getDebugLoc();
8514 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8515 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8517 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8519 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8520 DAG.getIntPtrConstant(TD->getPointerSize()));
8521 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8522 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8524 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8525 MF.getRegInfo().addLiveOut(StoreAddrReg);
8527 return DAG.getNode(X86ISD::EH_RETURN, dl,
8529 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8532 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8533 SelectionDAG &DAG) const {
8534 SDValue Root = Op.getOperand(0);
8535 SDValue Trmp = Op.getOperand(1); // trampoline
8536 SDValue FPtr = Op.getOperand(2); // nested function
8537 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8538 DebugLoc dl = Op.getDebugLoc();
8540 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8542 if (Subtarget->is64Bit()) {
8543 SDValue OutChains[6];
8545 // Large code-model.
8546 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8547 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8549 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
8550 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
8552 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8554 // Load the pointer to the nested function into R11.
8555 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8556 SDValue Addr = Trmp;
8557 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8558 Addr, MachinePointerInfo(TrmpAddr),
8561 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8562 DAG.getConstant(2, MVT::i64));
8563 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8564 MachinePointerInfo(TrmpAddr, 2),
8567 // Load the 'nest' parameter value into R10.
8568 // R10 is specified in X86CallingConv.td
8569 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8570 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8571 DAG.getConstant(10, MVT::i64));
8572 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8573 Addr, MachinePointerInfo(TrmpAddr, 10),
8576 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8577 DAG.getConstant(12, MVT::i64));
8578 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8579 MachinePointerInfo(TrmpAddr, 12),
8582 // Jump to the nested function.
8583 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8584 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8585 DAG.getConstant(20, MVT::i64));
8586 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8587 Addr, MachinePointerInfo(TrmpAddr, 20),
8590 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8591 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8592 DAG.getConstant(22, MVT::i64));
8593 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8594 MachinePointerInfo(TrmpAddr, 22),
8598 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8599 return DAG.getMergeValues(Ops, 2, dl);
8601 const Function *Func =
8602 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8603 CallingConv::ID CC = Func->getCallingConv();
8608 llvm_unreachable("Unsupported calling convention");
8609 case CallingConv::C:
8610 case CallingConv::X86_StdCall: {
8611 // Pass 'nest' parameter in ECX.
8612 // Must be kept in sync with X86CallingConv.td
8615 // Check that ECX wasn't needed by an 'inreg' parameter.
8616 const FunctionType *FTy = Func->getFunctionType();
8617 const AttrListPtr &Attrs = Func->getAttributes();
8619 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8620 unsigned InRegCount = 0;
8623 for (FunctionType::param_iterator I = FTy->param_begin(),
8624 E = FTy->param_end(); I != E; ++I, ++Idx)
8625 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8626 // FIXME: should only count parameters that are lowered to integers.
8627 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8629 if (InRegCount > 2) {
8630 report_fatal_error("Nest register in use - reduce number of inreg"
8636 case CallingConv::X86_FastCall:
8637 case CallingConv::X86_ThisCall:
8638 case CallingConv::Fast:
8639 // Pass 'nest' parameter in EAX.
8640 // Must be kept in sync with X86CallingConv.td
8645 SDValue OutChains[4];
8648 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8649 DAG.getConstant(10, MVT::i32));
8650 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8652 // This is storing the opcode for MOV32ri.
8653 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8654 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8655 OutChains[0] = DAG.getStore(Root, dl,
8656 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8657 Trmp, MachinePointerInfo(TrmpAddr),
8660 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8661 DAG.getConstant(1, MVT::i32));
8662 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8663 MachinePointerInfo(TrmpAddr, 1),
8666 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8667 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8668 DAG.getConstant(5, MVT::i32));
8669 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8670 MachinePointerInfo(TrmpAddr, 5),
8673 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8674 DAG.getConstant(6, MVT::i32));
8675 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8676 MachinePointerInfo(TrmpAddr, 6),
8680 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8681 return DAG.getMergeValues(Ops, 2, dl);
8685 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8686 SelectionDAG &DAG) const {
8688 The rounding mode is in bits 11:10 of FPSR, and has the following
8695 FLT_ROUNDS, on the other hand, expects the following:
8702 To perform the conversion, we do:
8703 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8706 MachineFunction &MF = DAG.getMachineFunction();
8707 const TargetMachine &TM = MF.getTarget();
8708 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8709 unsigned StackAlignment = TFI.getStackAlignment();
8710 EVT VT = Op.getValueType();
8711 DebugLoc DL = Op.getDebugLoc();
8713 // Save FP Control Word to stack slot
8714 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8715 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8718 MachineMemOperand *MMO =
8719 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8720 MachineMemOperand::MOStore, 2, 2);
8722 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8723 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8724 DAG.getVTList(MVT::Other),
8725 Ops, 2, MVT::i16, MMO);
8727 // Load FP Control Word from stack slot
8728 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8729 MachinePointerInfo(), false, false, 0);
8731 // Transform as necessary
8733 DAG.getNode(ISD::SRL, DL, MVT::i16,
8734 DAG.getNode(ISD::AND, DL, MVT::i16,
8735 CWD, DAG.getConstant(0x800, MVT::i16)),
8736 DAG.getConstant(11, MVT::i8));
8738 DAG.getNode(ISD::SRL, DL, MVT::i16,
8739 DAG.getNode(ISD::AND, DL, MVT::i16,
8740 CWD, DAG.getConstant(0x400, MVT::i16)),
8741 DAG.getConstant(9, MVT::i8));
8744 DAG.getNode(ISD::AND, DL, MVT::i16,
8745 DAG.getNode(ISD::ADD, DL, MVT::i16,
8746 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8747 DAG.getConstant(1, MVT::i16)),
8748 DAG.getConstant(3, MVT::i16));
8751 return DAG.getNode((VT.getSizeInBits() < 16 ?
8752 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8755 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8756 EVT VT = Op.getValueType();
8758 unsigned NumBits = VT.getSizeInBits();
8759 DebugLoc dl = Op.getDebugLoc();
8761 Op = Op.getOperand(0);
8762 if (VT == MVT::i8) {
8763 // Zero extend to i32 since there is not an i8 bsr.
8765 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8768 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8769 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8770 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8772 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8775 DAG.getConstant(NumBits+NumBits-1, OpVT),
8776 DAG.getConstant(X86::COND_E, MVT::i8),
8779 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8781 // Finally xor with NumBits-1.
8782 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8785 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8789 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8790 EVT VT = Op.getValueType();
8792 unsigned NumBits = VT.getSizeInBits();
8793 DebugLoc dl = Op.getDebugLoc();
8795 Op = Op.getOperand(0);
8796 if (VT == MVT::i8) {
8798 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8801 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8802 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8803 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8805 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8808 DAG.getConstant(NumBits, OpVT),
8809 DAG.getConstant(X86::COND_E, MVT::i8),
8812 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8815 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8819 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8820 EVT VT = Op.getValueType();
8821 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8822 DebugLoc dl = Op.getDebugLoc();
8824 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8825 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8826 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8827 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8828 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8830 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8831 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8832 // return AloBlo + AloBhi + AhiBlo;
8834 SDValue A = Op.getOperand(0);
8835 SDValue B = Op.getOperand(1);
8837 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8838 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8839 A, DAG.getConstant(32, MVT::i32));
8840 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8841 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8842 B, DAG.getConstant(32, MVT::i32));
8843 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8844 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8846 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8847 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8849 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8850 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8852 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8853 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8854 AloBhi, DAG.getConstant(32, MVT::i32));
8855 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8856 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8857 AhiBlo, DAG.getConstant(32, MVT::i32));
8858 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8859 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8863 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
8865 EVT VT = Op.getValueType();
8866 DebugLoc dl = Op.getDebugLoc();
8867 SDValue R = Op.getOperand(0);
8868 SDValue Amt = Op.getOperand(1);
8870 LLVMContext *Context = DAG.getContext();
8873 if (!Subtarget->hasSSE2()) return SDValue();
8875 // Optimize shl/srl/sra with constant shift amount.
8876 if (isSplatVector(Amt.getNode())) {
8877 SDValue SclrAmt = Amt->getOperand(0);
8878 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
8879 uint64_t ShiftAmt = C->getZExtValue();
8881 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
8882 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8883 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8884 R, DAG.getConstant(ShiftAmt, MVT::i32));
8886 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
8887 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8888 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8889 R, DAG.getConstant(ShiftAmt, MVT::i32));
8891 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
8892 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8893 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8894 R, DAG.getConstant(ShiftAmt, MVT::i32));
8896 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
8897 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8898 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8899 R, DAG.getConstant(ShiftAmt, MVT::i32));
8901 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
8902 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8903 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8904 R, DAG.getConstant(ShiftAmt, MVT::i32));
8906 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
8907 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8908 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8909 R, DAG.getConstant(ShiftAmt, MVT::i32));
8911 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
8912 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8913 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8914 R, DAG.getConstant(ShiftAmt, MVT::i32));
8916 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
8917 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8918 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8919 R, DAG.getConstant(ShiftAmt, MVT::i32));
8923 // Lower SHL with variable shift amount.
8924 // Cannot lower SHL without SSE4.1 or later.
8925 if (!Subtarget->hasSSE41()) return SDValue();
8927 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
8928 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8929 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8930 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8932 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8934 std::vector<Constant*> CV(4, CI);
8935 Constant *C = ConstantVector::get(CV);
8936 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8937 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8938 MachinePointerInfo::getConstantPool(),
8941 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8942 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
8943 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8944 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8946 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
8948 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8949 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8950 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8952 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8953 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8955 std::vector<Constant*> CVM1(16, CM1);
8956 std::vector<Constant*> CVM2(16, CM2);
8957 Constant *C = ConstantVector::get(CVM1);
8958 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8959 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8960 MachinePointerInfo::getConstantPool(),
8963 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8964 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8965 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8966 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8967 DAG.getConstant(4, MVT::i32));
8968 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8970 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8972 C = ConstantVector::get(CVM2);
8973 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8974 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8975 MachinePointerInfo::getConstantPool(),
8978 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8979 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8980 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8981 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8982 DAG.getConstant(2, MVT::i32));
8983 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8985 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8987 // return pblendv(r, r+r, a);
8988 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
8989 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8995 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8996 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8997 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8998 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8999 // has only one use.
9000 SDNode *N = Op.getNode();
9001 SDValue LHS = N->getOperand(0);
9002 SDValue RHS = N->getOperand(1);
9003 unsigned BaseOp = 0;
9005 DebugLoc DL = Op.getDebugLoc();
9006 switch (Op.getOpcode()) {
9007 default: llvm_unreachable("Unknown ovf instruction!");
9009 // A subtract of one will be selected as a INC. Note that INC doesn't
9010 // set CF, so we can't do this for UADDO.
9011 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9013 BaseOp = X86ISD::INC;
9017 BaseOp = X86ISD::ADD;
9021 BaseOp = X86ISD::ADD;
9025 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9026 // set CF, so we can't do this for USUBO.
9027 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9029 BaseOp = X86ISD::DEC;
9033 BaseOp = X86ISD::SUB;
9037 BaseOp = X86ISD::SUB;
9041 BaseOp = X86ISD::SMUL;
9044 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9045 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9047 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9050 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9051 DAG.getConstant(X86::COND_O, MVT::i32),
9052 SDValue(Sum.getNode(), 2));
9054 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
9059 // Also sets EFLAGS.
9060 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9061 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9064 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9065 DAG.getConstant(Cond, MVT::i32),
9066 SDValue(Sum.getNode(), 1));
9068 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
9072 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9073 DebugLoc dl = Op.getDebugLoc();
9075 if (!Subtarget->hasSSE2()) {
9076 SDValue Chain = Op.getOperand(0);
9077 SDValue Zero = DAG.getConstant(0,
9078 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
9080 DAG.getRegister(X86::ESP, MVT::i32), // Base
9081 DAG.getTargetConstant(1, MVT::i8), // Scale
9082 DAG.getRegister(0, MVT::i32), // Index
9083 DAG.getTargetConstant(0, MVT::i32), // Disp
9084 DAG.getRegister(0, MVT::i32), // Segment.
9089 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9090 array_lengthof(Ops));
9091 return SDValue(Res, 0);
9094 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9096 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9098 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9099 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9100 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9101 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9103 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9104 if (!Op1 && !Op2 && !Op3 && Op4)
9105 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9107 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9108 if (Op1 && !Op2 && !Op3 && !Op4)
9109 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9111 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9113 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9116 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9117 EVT T = Op.getValueType();
9118 DebugLoc DL = Op.getDebugLoc();
9121 switch(T.getSimpleVT().SimpleTy) {
9123 assert(false && "Invalid value type!");
9124 case MVT::i8: Reg = X86::AL; size = 1; break;
9125 case MVT::i16: Reg = X86::AX; size = 2; break;
9126 case MVT::i32: Reg = X86::EAX; size = 4; break;
9128 assert(Subtarget->is64Bit() && "Node not type legal!");
9129 Reg = X86::RAX; size = 8;
9132 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9133 Op.getOperand(2), SDValue());
9134 SDValue Ops[] = { cpIn.getValue(0),
9137 DAG.getTargetConstant(size, MVT::i8),
9139 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9140 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9141 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9144 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9148 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9149 SelectionDAG &DAG) const {
9150 assert(Subtarget->is64Bit() && "Result not type legalized?");
9151 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9152 SDValue TheChain = Op.getOperand(0);
9153 DebugLoc dl = Op.getDebugLoc();
9154 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9155 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9156 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9158 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9159 DAG.getConstant(32, MVT::i8));
9161 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9164 return DAG.getMergeValues(Ops, 2, dl);
9167 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9168 SelectionDAG &DAG) const {
9169 EVT SrcVT = Op.getOperand(0).getValueType();
9170 EVT DstVT = Op.getValueType();
9171 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9172 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9173 assert((DstVT == MVT::i64 ||
9174 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9175 "Unexpected custom BITCAST");
9176 // i64 <=> MMX conversions are Legal.
9177 if (SrcVT==MVT::i64 && DstVT.isVector())
9179 if (DstVT==MVT::i64 && SrcVT.isVector())
9181 // MMX <=> MMX conversions are Legal.
9182 if (SrcVT.isVector() && DstVT.isVector())
9184 // All other conversions need to be expanded.
9188 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9189 SDNode *Node = Op.getNode();
9190 DebugLoc dl = Node->getDebugLoc();
9191 EVT T = Node->getValueType(0);
9192 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9193 DAG.getConstant(0, T), Node->getOperand(2));
9194 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9195 cast<AtomicSDNode>(Node)->getMemoryVT(),
9196 Node->getOperand(0),
9197 Node->getOperand(1), negOp,
9198 cast<AtomicSDNode>(Node)->getSrcValue(),
9199 cast<AtomicSDNode>(Node)->getAlignment());
9202 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9203 EVT VT = Op.getNode()->getValueType(0);
9205 // Let legalize expand this if it isn't a legal type yet.
9206 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9209 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9212 bool ExtraOp = false;
9213 switch (Op.getOpcode()) {
9214 default: assert(0 && "Invalid code");
9215 case ISD::ADDC: Opc = X86ISD::ADD; break;
9216 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9217 case ISD::SUBC: Opc = X86ISD::SUB; break;
9218 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9222 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9224 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9225 Op.getOperand(1), Op.getOperand(2));
9228 /// LowerOperation - Provide custom lowering hooks for some operations.
9230 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9231 switch (Op.getOpcode()) {
9232 default: llvm_unreachable("Should not custom lower this!");
9233 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9234 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9235 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9236 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9237 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9238 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9239 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9240 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9241 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9242 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9243 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9244 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9245 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9246 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9247 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9248 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9249 case ISD::SHL_PARTS:
9250 case ISD::SRA_PARTS:
9251 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9252 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9253 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9254 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9255 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9256 case ISD::FABS: return LowerFABS(Op, DAG);
9257 case ISD::FNEG: return LowerFNEG(Op, DAG);
9258 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9259 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9260 case ISD::SETCC: return LowerSETCC(Op, DAG);
9261 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9262 case ISD::SELECT: return LowerSELECT(Op, DAG);
9263 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9264 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9265 case ISD::VASTART: return LowerVASTART(Op, DAG);
9266 case ISD::VAARG: return LowerVAARG(Op, DAG);
9267 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9268 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9269 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9270 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9271 case ISD::FRAME_TO_ARGS_OFFSET:
9272 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9273 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9274 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9275 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9276 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9277 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9278 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9279 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9282 case ISD::SHL: return LowerShift(Op, DAG);
9288 case ISD::UMULO: return LowerXALUO(Op, DAG);
9289 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9290 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9294 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9298 void X86TargetLowering::
9299 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9300 SelectionDAG &DAG, unsigned NewOp) const {
9301 EVT T = Node->getValueType(0);
9302 DebugLoc dl = Node->getDebugLoc();
9303 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9305 SDValue Chain = Node->getOperand(0);
9306 SDValue In1 = Node->getOperand(1);
9307 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9308 Node->getOperand(2), DAG.getIntPtrConstant(0));
9309 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9310 Node->getOperand(2), DAG.getIntPtrConstant(1));
9311 SDValue Ops[] = { Chain, In1, In2L, In2H };
9312 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9314 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9315 cast<MemSDNode>(Node)->getMemOperand());
9316 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9317 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9318 Results.push_back(Result.getValue(2));
9321 /// ReplaceNodeResults - Replace a node with an illegal result type
9322 /// with a new node built out of custom code.
9323 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9324 SmallVectorImpl<SDValue>&Results,
9325 SelectionDAG &DAG) const {
9326 DebugLoc dl = N->getDebugLoc();
9327 switch (N->getOpcode()) {
9329 assert(false && "Do not know how to custom type legalize this operation!");
9335 // We don't want to expand or promote these.
9337 case ISD::FP_TO_SINT: {
9338 std::pair<SDValue,SDValue> Vals =
9339 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9340 SDValue FIST = Vals.first, StackSlot = Vals.second;
9341 if (FIST.getNode() != 0) {
9342 EVT VT = N->getValueType(0);
9343 // Return a load from the stack slot.
9344 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9345 MachinePointerInfo(), false, false, 0));
9349 case ISD::READCYCLECOUNTER: {
9350 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9351 SDValue TheChain = N->getOperand(0);
9352 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9353 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9355 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9357 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9358 SDValue Ops[] = { eax, edx };
9359 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9360 Results.push_back(edx.getValue(1));
9363 case ISD::ATOMIC_CMP_SWAP: {
9364 EVT T = N->getValueType(0);
9365 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9366 SDValue cpInL, cpInH;
9367 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9368 DAG.getConstant(0, MVT::i32));
9369 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9370 DAG.getConstant(1, MVT::i32));
9371 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9372 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9374 SDValue swapInL, swapInH;
9375 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9376 DAG.getConstant(0, MVT::i32));
9377 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9378 DAG.getConstant(1, MVT::i32));
9379 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9381 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9382 swapInL.getValue(1));
9383 SDValue Ops[] = { swapInH.getValue(0),
9385 swapInH.getValue(1) };
9386 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9387 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9388 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9390 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9391 MVT::i32, Result.getValue(1));
9392 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9393 MVT::i32, cpOutL.getValue(2));
9394 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9395 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9396 Results.push_back(cpOutH.getValue(1));
9399 case ISD::ATOMIC_LOAD_ADD:
9400 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9402 case ISD::ATOMIC_LOAD_AND:
9403 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9405 case ISD::ATOMIC_LOAD_NAND:
9406 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9408 case ISD::ATOMIC_LOAD_OR:
9409 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9411 case ISD::ATOMIC_LOAD_SUB:
9412 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9414 case ISD::ATOMIC_LOAD_XOR:
9415 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9417 case ISD::ATOMIC_SWAP:
9418 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9423 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9425 default: return NULL;
9426 case X86ISD::BSF: return "X86ISD::BSF";
9427 case X86ISD::BSR: return "X86ISD::BSR";
9428 case X86ISD::SHLD: return "X86ISD::SHLD";
9429 case X86ISD::SHRD: return "X86ISD::SHRD";
9430 case X86ISD::FAND: return "X86ISD::FAND";
9431 case X86ISD::FOR: return "X86ISD::FOR";
9432 case X86ISD::FXOR: return "X86ISD::FXOR";
9433 case X86ISD::FSRL: return "X86ISD::FSRL";
9434 case X86ISD::FILD: return "X86ISD::FILD";
9435 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9436 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9437 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9438 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9439 case X86ISD::FLD: return "X86ISD::FLD";
9440 case X86ISD::FST: return "X86ISD::FST";
9441 case X86ISD::CALL: return "X86ISD::CALL";
9442 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9443 case X86ISD::BT: return "X86ISD::BT";
9444 case X86ISD::CMP: return "X86ISD::CMP";
9445 case X86ISD::COMI: return "X86ISD::COMI";
9446 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9447 case X86ISD::SETCC: return "X86ISD::SETCC";
9448 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9449 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9450 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9451 case X86ISD::CMOV: return "X86ISD::CMOV";
9452 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9453 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9454 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9455 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9456 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9457 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9458 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9459 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9460 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9461 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9462 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9463 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9464 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9465 case X86ISD::PANDN: return "X86ISD::PANDN";
9466 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9467 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9468 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9469 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9470 case X86ISD::FMAX: return "X86ISD::FMAX";
9471 case X86ISD::FMIN: return "X86ISD::FMIN";
9472 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9473 case X86ISD::FRCP: return "X86ISD::FRCP";
9474 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9475 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9476 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9477 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9478 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9479 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9480 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9481 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9482 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9483 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9484 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9485 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9486 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9487 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9488 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9489 case X86ISD::VSHL: return "X86ISD::VSHL";
9490 case X86ISD::VSRL: return "X86ISD::VSRL";
9491 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9492 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9493 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9494 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9495 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9496 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9497 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9498 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9499 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9500 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9501 case X86ISD::ADD: return "X86ISD::ADD";
9502 case X86ISD::SUB: return "X86ISD::SUB";
9503 case X86ISD::ADC: return "X86ISD::ADC";
9504 case X86ISD::SBB: return "X86ISD::SBB";
9505 case X86ISD::SMUL: return "X86ISD::SMUL";
9506 case X86ISD::UMUL: return "X86ISD::UMUL";
9507 case X86ISD::INC: return "X86ISD::INC";
9508 case X86ISD::DEC: return "X86ISD::DEC";
9509 case X86ISD::OR: return "X86ISD::OR";
9510 case X86ISD::XOR: return "X86ISD::XOR";
9511 case X86ISD::AND: return "X86ISD::AND";
9512 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9513 case X86ISD::PTEST: return "X86ISD::PTEST";
9514 case X86ISD::TESTP: return "X86ISD::TESTP";
9515 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9516 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9517 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9518 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9519 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9520 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9521 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9522 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9523 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9524 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9525 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9526 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9527 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9528 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9529 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9530 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9531 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9532 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9533 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9534 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9535 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9536 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9537 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9538 case X86ISD::VUNPCKLPS: return "X86ISD::VUNPCKLPS";
9539 case X86ISD::VUNPCKLPD: return "X86ISD::VUNPCKLPD";
9540 case X86ISD::VUNPCKLPSY: return "X86ISD::VUNPCKLPSY";
9541 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9542 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9543 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9544 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9545 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9546 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9547 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9548 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9549 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9550 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9551 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9552 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9553 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9554 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9558 // isLegalAddressingMode - Return true if the addressing mode represented
9559 // by AM is legal for this target, for a load/store of the specified type.
9560 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9561 const Type *Ty) const {
9562 // X86 supports extremely general addressing modes.
9563 CodeModel::Model M = getTargetMachine().getCodeModel();
9564 Reloc::Model R = getTargetMachine().getRelocationModel();
9566 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9567 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9572 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9574 // If a reference to this global requires an extra load, we can't fold it.
9575 if (isGlobalStubReference(GVFlags))
9578 // If BaseGV requires a register for the PIC base, we cannot also have a
9579 // BaseReg specified.
9580 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9583 // If lower 4G is not available, then we must use rip-relative addressing.
9584 if ((M != CodeModel::Small || R != Reloc::Static) &&
9585 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9595 // These scales always work.
9600 // These scales are formed with basereg+scalereg. Only accept if there is
9605 default: // Other stuff never works.
9613 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
9614 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9616 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9617 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9618 if (NumBits1 <= NumBits2)
9623 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9624 if (!VT1.isInteger() || !VT2.isInteger())
9626 unsigned NumBits1 = VT1.getSizeInBits();
9627 unsigned NumBits2 = VT2.getSizeInBits();
9628 if (NumBits1 <= NumBits2)
9633 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
9634 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9635 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9638 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9639 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9640 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9643 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9644 // i16 instructions are longer (0x66 prefix) and potentially slower.
9645 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9648 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9649 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9650 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9651 /// are assumed to be legal.
9653 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9655 // Very little shuffling can be done for 64-bit vectors right now.
9656 if (VT.getSizeInBits() == 64)
9657 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9659 // FIXME: pshufb, blends, shifts.
9660 return (VT.getVectorNumElements() == 2 ||
9661 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9662 isMOVLMask(M, VT) ||
9663 isSHUFPMask(M, VT) ||
9664 isPSHUFDMask(M, VT) ||
9665 isPSHUFHWMask(M, VT) ||
9666 isPSHUFLWMask(M, VT) ||
9667 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9668 isUNPCKLMask(M, VT) ||
9669 isUNPCKHMask(M, VT) ||
9670 isUNPCKL_v_undef_Mask(M, VT) ||
9671 isUNPCKH_v_undef_Mask(M, VT));
9675 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9677 unsigned NumElts = VT.getVectorNumElements();
9678 // FIXME: This collection of masks seems suspect.
9681 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9682 return (isMOVLMask(Mask, VT) ||
9683 isCommutedMOVLMask(Mask, VT, true) ||
9684 isSHUFPMask(Mask, VT) ||
9685 isCommutedSHUFPMask(Mask, VT));
9690 //===----------------------------------------------------------------------===//
9691 // X86 Scheduler Hooks
9692 //===----------------------------------------------------------------------===//
9694 // private utility function
9696 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9697 MachineBasicBlock *MBB,
9704 TargetRegisterClass *RC,
9705 bool invSrc) const {
9706 // For the atomic bitwise operator, we generate
9709 // ld t1 = [bitinstr.addr]
9710 // op t2 = t1, [bitinstr.val]
9712 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9714 // fallthrough -->nextMBB
9715 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9716 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9717 MachineFunction::iterator MBBIter = MBB;
9720 /// First build the CFG
9721 MachineFunction *F = MBB->getParent();
9722 MachineBasicBlock *thisMBB = MBB;
9723 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9724 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9725 F->insert(MBBIter, newMBB);
9726 F->insert(MBBIter, nextMBB);
9728 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9729 nextMBB->splice(nextMBB->begin(), thisMBB,
9730 llvm::next(MachineBasicBlock::iterator(bInstr)),
9732 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9734 // Update thisMBB to fall through to newMBB
9735 thisMBB->addSuccessor(newMBB);
9737 // newMBB jumps to itself and fall through to nextMBB
9738 newMBB->addSuccessor(nextMBB);
9739 newMBB->addSuccessor(newMBB);
9741 // Insert instructions into newMBB based on incoming instruction
9742 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9743 "unexpected number of operands");
9744 DebugLoc dl = bInstr->getDebugLoc();
9745 MachineOperand& destOper = bInstr->getOperand(0);
9746 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9747 int numArgs = bInstr->getNumOperands() - 1;
9748 for (int i=0; i < numArgs; ++i)
9749 argOpers[i] = &bInstr->getOperand(i+1);
9751 // x86 address has 4 operands: base, index, scale, and displacement
9752 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9753 int valArgIndx = lastAddrIndx + 1;
9755 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9756 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9757 for (int i=0; i <= lastAddrIndx; ++i)
9758 (*MIB).addOperand(*argOpers[i]);
9760 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9762 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9767 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9768 assert((argOpers[valArgIndx]->isReg() ||
9769 argOpers[valArgIndx]->isImm()) &&
9771 if (argOpers[valArgIndx]->isReg())
9772 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9774 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9776 (*MIB).addOperand(*argOpers[valArgIndx]);
9778 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9781 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9782 for (int i=0; i <= lastAddrIndx; ++i)
9783 (*MIB).addOperand(*argOpers[i]);
9785 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9786 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9787 bInstr->memoperands_end());
9789 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9793 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9795 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9799 // private utility function: 64 bit atomics on 32 bit host.
9801 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9802 MachineBasicBlock *MBB,
9807 bool invSrc) const {
9808 // For the atomic bitwise operator, we generate
9809 // thisMBB (instructions are in pairs, except cmpxchg8b)
9810 // ld t1,t2 = [bitinstr.addr]
9812 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9813 // op t5, t6 <- out1, out2, [bitinstr.val]
9814 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9815 // mov ECX, EBX <- t5, t6
9816 // mov EAX, EDX <- t1, t2
9817 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9818 // mov t3, t4 <- EAX, EDX
9820 // result in out1, out2
9821 // fallthrough -->nextMBB
9823 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9824 const unsigned LoadOpc = X86::MOV32rm;
9825 const unsigned NotOpc = X86::NOT32r;
9826 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9827 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9828 MachineFunction::iterator MBBIter = MBB;
9831 /// First build the CFG
9832 MachineFunction *F = MBB->getParent();
9833 MachineBasicBlock *thisMBB = MBB;
9834 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9835 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9836 F->insert(MBBIter, newMBB);
9837 F->insert(MBBIter, nextMBB);
9839 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9840 nextMBB->splice(nextMBB->begin(), thisMBB,
9841 llvm::next(MachineBasicBlock::iterator(bInstr)),
9843 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9845 // Update thisMBB to fall through to newMBB
9846 thisMBB->addSuccessor(newMBB);
9848 // newMBB jumps to itself and fall through to nextMBB
9849 newMBB->addSuccessor(nextMBB);
9850 newMBB->addSuccessor(newMBB);
9852 DebugLoc dl = bInstr->getDebugLoc();
9853 // Insert instructions into newMBB based on incoming instruction
9854 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9855 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9856 "unexpected number of operands");
9857 MachineOperand& dest1Oper = bInstr->getOperand(0);
9858 MachineOperand& dest2Oper = bInstr->getOperand(1);
9859 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9860 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9861 argOpers[i] = &bInstr->getOperand(i+2);
9863 // We use some of the operands multiple times, so conservatively just
9864 // clear any kill flags that might be present.
9865 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9866 argOpers[i]->setIsKill(false);
9869 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9870 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9872 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9873 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9874 for (int i=0; i <= lastAddrIndx; ++i)
9875 (*MIB).addOperand(*argOpers[i]);
9876 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9877 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9878 // add 4 to displacement.
9879 for (int i=0; i <= lastAddrIndx-2; ++i)
9880 (*MIB).addOperand(*argOpers[i]);
9881 MachineOperand newOp3 = *(argOpers[3]);
9883 newOp3.setImm(newOp3.getImm()+4);
9885 newOp3.setOffset(newOp3.getOffset()+4);
9886 (*MIB).addOperand(newOp3);
9887 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9889 // t3/4 are defined later, at the bottom of the loop
9890 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9891 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9892 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9893 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9894 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9895 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9897 // The subsequent operations should be using the destination registers of
9898 //the PHI instructions.
9900 t1 = F->getRegInfo().createVirtualRegister(RC);
9901 t2 = F->getRegInfo().createVirtualRegister(RC);
9902 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9903 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9905 t1 = dest1Oper.getReg();
9906 t2 = dest2Oper.getReg();
9909 int valArgIndx = lastAddrIndx + 1;
9910 assert((argOpers[valArgIndx]->isReg() ||
9911 argOpers[valArgIndx]->isImm()) &&
9913 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9914 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9915 if (argOpers[valArgIndx]->isReg())
9916 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9918 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9919 if (regOpcL != X86::MOV32rr)
9921 (*MIB).addOperand(*argOpers[valArgIndx]);
9922 assert(argOpers[valArgIndx + 1]->isReg() ==
9923 argOpers[valArgIndx]->isReg());
9924 assert(argOpers[valArgIndx + 1]->isImm() ==
9925 argOpers[valArgIndx]->isImm());
9926 if (argOpers[valArgIndx + 1]->isReg())
9927 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9929 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9930 if (regOpcH != X86::MOV32rr)
9932 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9934 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9936 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9939 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9941 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9944 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9945 for (int i=0; i <= lastAddrIndx; ++i)
9946 (*MIB).addOperand(*argOpers[i]);
9948 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9949 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9950 bInstr->memoperands_end());
9952 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9953 MIB.addReg(X86::EAX);
9954 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9955 MIB.addReg(X86::EDX);
9958 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9960 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9964 // private utility function
9966 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9967 MachineBasicBlock *MBB,
9968 unsigned cmovOpc) const {
9969 // For the atomic min/max operator, we generate
9972 // ld t1 = [min/max.addr]
9973 // mov t2 = [min/max.val]
9975 // cmov[cond] t2 = t1
9977 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9979 // fallthrough -->nextMBB
9981 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9982 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9983 MachineFunction::iterator MBBIter = MBB;
9986 /// First build the CFG
9987 MachineFunction *F = MBB->getParent();
9988 MachineBasicBlock *thisMBB = MBB;
9989 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9990 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9991 F->insert(MBBIter, newMBB);
9992 F->insert(MBBIter, nextMBB);
9994 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9995 nextMBB->splice(nextMBB->begin(), thisMBB,
9996 llvm::next(MachineBasicBlock::iterator(mInstr)),
9998 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10000 // Update thisMBB to fall through to newMBB
10001 thisMBB->addSuccessor(newMBB);
10003 // newMBB jumps to newMBB and fall through to nextMBB
10004 newMBB->addSuccessor(nextMBB);
10005 newMBB->addSuccessor(newMBB);
10007 DebugLoc dl = mInstr->getDebugLoc();
10008 // Insert instructions into newMBB based on incoming instruction
10009 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10010 "unexpected number of operands");
10011 MachineOperand& destOper = mInstr->getOperand(0);
10012 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10013 int numArgs = mInstr->getNumOperands() - 1;
10014 for (int i=0; i < numArgs; ++i)
10015 argOpers[i] = &mInstr->getOperand(i+1);
10017 // x86 address has 4 operands: base, index, scale, and displacement
10018 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10019 int valArgIndx = lastAddrIndx + 1;
10021 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10022 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10023 for (int i=0; i <= lastAddrIndx; ++i)
10024 (*MIB).addOperand(*argOpers[i]);
10026 // We only support register and immediate values
10027 assert((argOpers[valArgIndx]->isReg() ||
10028 argOpers[valArgIndx]->isImm()) &&
10029 "invalid operand");
10031 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10032 if (argOpers[valArgIndx]->isReg())
10033 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10035 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10036 (*MIB).addOperand(*argOpers[valArgIndx]);
10038 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10041 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10046 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10047 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10051 // Cmp and exchange if none has modified the memory location
10052 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10053 for (int i=0; i <= lastAddrIndx; ++i)
10054 (*MIB).addOperand(*argOpers[i]);
10056 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10057 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10058 mInstr->memoperands_end());
10060 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10061 MIB.addReg(X86::EAX);
10064 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10066 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10070 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10071 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10072 // in the .td file.
10073 MachineBasicBlock *
10074 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10075 unsigned numArgs, bool memArg) const {
10076 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10077 "Target must have SSE4.2 or AVX features enabled");
10079 DebugLoc dl = MI->getDebugLoc();
10080 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10082 if (!Subtarget->hasAVX()) {
10084 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10086 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10089 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10091 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10094 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10095 for (unsigned i = 0; i < numArgs; ++i) {
10096 MachineOperand &Op = MI->getOperand(i+1);
10097 if (!(Op.isReg() && Op.isImplicit()))
10098 MIB.addOperand(Op);
10100 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10101 .addReg(X86::XMM0);
10103 MI->eraseFromParent();
10107 MachineBasicBlock *
10108 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10109 DebugLoc dl = MI->getDebugLoc();
10110 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10112 // Address into RAX/EAX, other two args into ECX, EDX.
10113 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10114 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10115 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10116 for (int i = 0; i < X86::AddrNumOperands; ++i)
10117 MIB.addOperand(MI->getOperand(i));
10119 unsigned ValOps = X86::AddrNumOperands;
10120 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10121 .addReg(MI->getOperand(ValOps).getReg());
10122 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10123 .addReg(MI->getOperand(ValOps+1).getReg());
10125 // The instruction doesn't actually take any operands though.
10126 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10128 MI->eraseFromParent(); // The pseudo is gone now.
10132 MachineBasicBlock *
10133 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10134 DebugLoc dl = MI->getDebugLoc();
10135 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10137 // First arg in ECX, the second in EAX.
10138 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10139 .addReg(MI->getOperand(0).getReg());
10140 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10141 .addReg(MI->getOperand(1).getReg());
10143 // The instruction doesn't actually take any operands though.
10144 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10146 MI->eraseFromParent(); // The pseudo is gone now.
10150 MachineBasicBlock *
10151 X86TargetLowering::EmitVAARG64WithCustomInserter(
10153 MachineBasicBlock *MBB) const {
10154 // Emit va_arg instruction on X86-64.
10156 // Operands to this pseudo-instruction:
10157 // 0 ) Output : destination address (reg)
10158 // 1-5) Input : va_list address (addr, i64mem)
10159 // 6 ) ArgSize : Size (in bytes) of vararg type
10160 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10161 // 8 ) Align : Alignment of type
10162 // 9 ) EFLAGS (implicit-def)
10164 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10165 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10167 unsigned DestReg = MI->getOperand(0).getReg();
10168 MachineOperand &Base = MI->getOperand(1);
10169 MachineOperand &Scale = MI->getOperand(2);
10170 MachineOperand &Index = MI->getOperand(3);
10171 MachineOperand &Disp = MI->getOperand(4);
10172 MachineOperand &Segment = MI->getOperand(5);
10173 unsigned ArgSize = MI->getOperand(6).getImm();
10174 unsigned ArgMode = MI->getOperand(7).getImm();
10175 unsigned Align = MI->getOperand(8).getImm();
10177 // Memory Reference
10178 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10179 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10180 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10182 // Machine Information
10183 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10184 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10185 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10186 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10187 DebugLoc DL = MI->getDebugLoc();
10189 // struct va_list {
10192 // i64 overflow_area (address)
10193 // i64 reg_save_area (address)
10195 // sizeof(va_list) = 24
10196 // alignment(va_list) = 8
10198 unsigned TotalNumIntRegs = 6;
10199 unsigned TotalNumXMMRegs = 8;
10200 bool UseGPOffset = (ArgMode == 1);
10201 bool UseFPOffset = (ArgMode == 2);
10202 unsigned MaxOffset = TotalNumIntRegs * 8 +
10203 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10205 /* Align ArgSize to a multiple of 8 */
10206 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10207 bool NeedsAlign = (Align > 8);
10209 MachineBasicBlock *thisMBB = MBB;
10210 MachineBasicBlock *overflowMBB;
10211 MachineBasicBlock *offsetMBB;
10212 MachineBasicBlock *endMBB;
10214 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10215 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10216 unsigned OffsetReg = 0;
10218 if (!UseGPOffset && !UseFPOffset) {
10219 // If we only pull from the overflow region, we don't create a branch.
10220 // We don't need to alter control flow.
10221 OffsetDestReg = 0; // unused
10222 OverflowDestReg = DestReg;
10225 overflowMBB = thisMBB;
10228 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10229 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10230 // If not, pull from overflow_area. (branch to overflowMBB)
10235 // offsetMBB overflowMBB
10240 // Registers for the PHI in endMBB
10241 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10242 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10244 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10245 MachineFunction *MF = MBB->getParent();
10246 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10247 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10248 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10250 MachineFunction::iterator MBBIter = MBB;
10253 // Insert the new basic blocks
10254 MF->insert(MBBIter, offsetMBB);
10255 MF->insert(MBBIter, overflowMBB);
10256 MF->insert(MBBIter, endMBB);
10258 // Transfer the remainder of MBB and its successor edges to endMBB.
10259 endMBB->splice(endMBB->begin(), thisMBB,
10260 llvm::next(MachineBasicBlock::iterator(MI)),
10262 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10264 // Make offsetMBB and overflowMBB successors of thisMBB
10265 thisMBB->addSuccessor(offsetMBB);
10266 thisMBB->addSuccessor(overflowMBB);
10268 // endMBB is a successor of both offsetMBB and overflowMBB
10269 offsetMBB->addSuccessor(endMBB);
10270 overflowMBB->addSuccessor(endMBB);
10272 // Load the offset value into a register
10273 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10274 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10278 .addDisp(Disp, UseFPOffset ? 4 : 0)
10279 .addOperand(Segment)
10280 .setMemRefs(MMOBegin, MMOEnd);
10282 // Check if there is enough room left to pull this argument.
10283 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10285 .addImm(MaxOffset + 8 - ArgSizeA8);
10287 // Branch to "overflowMBB" if offset >= max
10288 // Fall through to "offsetMBB" otherwise
10289 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10290 .addMBB(overflowMBB);
10293 // In offsetMBB, emit code to use the reg_save_area.
10295 assert(OffsetReg != 0);
10297 // Read the reg_save_area address.
10298 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10299 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10304 .addOperand(Segment)
10305 .setMemRefs(MMOBegin, MMOEnd);
10307 // Zero-extend the offset
10308 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10309 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10312 .addImm(X86::sub_32bit);
10314 // Add the offset to the reg_save_area to get the final address.
10315 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10316 .addReg(OffsetReg64)
10317 .addReg(RegSaveReg);
10319 // Compute the offset for the next argument
10320 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10321 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10323 .addImm(UseFPOffset ? 16 : 8);
10325 // Store it back into the va_list.
10326 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10330 .addDisp(Disp, UseFPOffset ? 4 : 0)
10331 .addOperand(Segment)
10332 .addReg(NextOffsetReg)
10333 .setMemRefs(MMOBegin, MMOEnd);
10336 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10341 // Emit code to use overflow area
10344 // Load the overflow_area address into a register.
10345 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10346 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10351 .addOperand(Segment)
10352 .setMemRefs(MMOBegin, MMOEnd);
10354 // If we need to align it, do so. Otherwise, just copy the address
10355 // to OverflowDestReg.
10357 // Align the overflow address
10358 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10359 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10361 // aligned_addr = (addr + (align-1)) & ~(align-1)
10362 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10363 .addReg(OverflowAddrReg)
10366 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10368 .addImm(~(uint64_t)(Align-1));
10370 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10371 .addReg(OverflowAddrReg);
10374 // Compute the next overflow address after this argument.
10375 // (the overflow address should be kept 8-byte aligned)
10376 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10377 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10378 .addReg(OverflowDestReg)
10379 .addImm(ArgSizeA8);
10381 // Store the new overflow address.
10382 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10387 .addOperand(Segment)
10388 .addReg(NextAddrReg)
10389 .setMemRefs(MMOBegin, MMOEnd);
10391 // If we branched, emit the PHI to the front of endMBB.
10393 BuildMI(*endMBB, endMBB->begin(), DL,
10394 TII->get(X86::PHI), DestReg)
10395 .addReg(OffsetDestReg).addMBB(offsetMBB)
10396 .addReg(OverflowDestReg).addMBB(overflowMBB);
10399 // Erase the pseudo instruction
10400 MI->eraseFromParent();
10405 MachineBasicBlock *
10406 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10408 MachineBasicBlock *MBB) const {
10409 // Emit code to save XMM registers to the stack. The ABI says that the
10410 // number of registers to save is given in %al, so it's theoretically
10411 // possible to do an indirect jump trick to avoid saving all of them,
10412 // however this code takes a simpler approach and just executes all
10413 // of the stores if %al is non-zero. It's less code, and it's probably
10414 // easier on the hardware branch predictor, and stores aren't all that
10415 // expensive anyway.
10417 // Create the new basic blocks. One block contains all the XMM stores,
10418 // and one block is the final destination regardless of whether any
10419 // stores were performed.
10420 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10421 MachineFunction *F = MBB->getParent();
10422 MachineFunction::iterator MBBIter = MBB;
10424 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10425 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10426 F->insert(MBBIter, XMMSaveMBB);
10427 F->insert(MBBIter, EndMBB);
10429 // Transfer the remainder of MBB and its successor edges to EndMBB.
10430 EndMBB->splice(EndMBB->begin(), MBB,
10431 llvm::next(MachineBasicBlock::iterator(MI)),
10433 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10435 // The original block will now fall through to the XMM save block.
10436 MBB->addSuccessor(XMMSaveMBB);
10437 // The XMMSaveMBB will fall through to the end block.
10438 XMMSaveMBB->addSuccessor(EndMBB);
10440 // Now add the instructions.
10441 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10442 DebugLoc DL = MI->getDebugLoc();
10444 unsigned CountReg = MI->getOperand(0).getReg();
10445 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10446 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10448 if (!Subtarget->isTargetWin64()) {
10449 // If %al is 0, branch around the XMM save block.
10450 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10451 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10452 MBB->addSuccessor(EndMBB);
10455 // In the XMM save block, save all the XMM argument registers.
10456 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10457 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10458 MachineMemOperand *MMO =
10459 F->getMachineMemOperand(
10460 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10461 MachineMemOperand::MOStore,
10462 /*Size=*/16, /*Align=*/16);
10463 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10464 .addFrameIndex(RegSaveFrameIndex)
10465 .addImm(/*Scale=*/1)
10466 .addReg(/*IndexReg=*/0)
10467 .addImm(/*Disp=*/Offset)
10468 .addReg(/*Segment=*/0)
10469 .addReg(MI->getOperand(i).getReg())
10470 .addMemOperand(MMO);
10473 MI->eraseFromParent(); // The pseudo instruction is gone now.
10478 MachineBasicBlock *
10479 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10480 MachineBasicBlock *BB) const {
10481 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10482 DebugLoc DL = MI->getDebugLoc();
10484 // To "insert" a SELECT_CC instruction, we actually have to insert the
10485 // diamond control-flow pattern. The incoming instruction knows the
10486 // destination vreg to set, the condition code register to branch on, the
10487 // true/false values to select between, and a branch opcode to use.
10488 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10489 MachineFunction::iterator It = BB;
10495 // cmpTY ccX, r1, r2
10497 // fallthrough --> copy0MBB
10498 MachineBasicBlock *thisMBB = BB;
10499 MachineFunction *F = BB->getParent();
10500 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10501 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10502 F->insert(It, copy0MBB);
10503 F->insert(It, sinkMBB);
10505 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10506 // live into the sink and copy blocks.
10507 const MachineFunction *MF = BB->getParent();
10508 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10509 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10511 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10512 const MachineOperand &MO = MI->getOperand(I);
10513 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10514 unsigned Reg = MO.getReg();
10515 if (Reg != X86::EFLAGS) continue;
10516 copy0MBB->addLiveIn(Reg);
10517 sinkMBB->addLiveIn(Reg);
10520 // Transfer the remainder of BB and its successor edges to sinkMBB.
10521 sinkMBB->splice(sinkMBB->begin(), BB,
10522 llvm::next(MachineBasicBlock::iterator(MI)),
10524 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10526 // Add the true and fallthrough blocks as its successors.
10527 BB->addSuccessor(copy0MBB);
10528 BB->addSuccessor(sinkMBB);
10530 // Create the conditional branch instruction.
10532 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10533 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10536 // %FalseValue = ...
10537 // # fallthrough to sinkMBB
10538 copy0MBB->addSuccessor(sinkMBB);
10541 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10543 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10544 TII->get(X86::PHI), MI->getOperand(0).getReg())
10545 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10546 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10548 MI->eraseFromParent(); // The pseudo instruction is gone now.
10552 MachineBasicBlock *
10553 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10554 MachineBasicBlock *BB) const {
10555 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10556 DebugLoc DL = MI->getDebugLoc();
10558 assert(!Subtarget->isTargetEnvMacho());
10560 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10561 // non-trivial part is impdef of ESP.
10563 if (Subtarget->isTargetWin64()) {
10564 if (Subtarget->isTargetCygMing()) {
10565 // ___chkstk(Mingw64):
10566 // Clobbers R10, R11, RAX and EFLAGS.
10568 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10569 .addExternalSymbol("___chkstk")
10570 .addReg(X86::RAX, RegState::Implicit)
10571 .addReg(X86::RSP, RegState::Implicit)
10572 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
10573 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
10574 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10576 // __chkstk(MSVCRT): does not update stack pointer.
10577 // Clobbers R10, R11 and EFLAGS.
10578 // FIXME: RAX(allocated size) might be reused and not killed.
10579 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10580 .addExternalSymbol("__chkstk")
10581 .addReg(X86::RAX, RegState::Implicit)
10582 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10583 // RAX has the offset to subtracted from RSP.
10584 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
10589 const char *StackProbeSymbol =
10590 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10592 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10593 .addExternalSymbol(StackProbeSymbol)
10594 .addReg(X86::EAX, RegState::Implicit)
10595 .addReg(X86::ESP, RegState::Implicit)
10596 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10597 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10598 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10601 MI->eraseFromParent(); // The pseudo instruction is gone now.
10605 MachineBasicBlock *
10606 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10607 MachineBasicBlock *BB) const {
10608 // This is pretty easy. We're taking the value that we received from
10609 // our load from the relocation, sticking it in either RDI (x86-64)
10610 // or EAX and doing an indirect call. The return value will then
10611 // be in the normal return register.
10612 const X86InstrInfo *TII
10613 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10614 DebugLoc DL = MI->getDebugLoc();
10615 MachineFunction *F = BB->getParent();
10617 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10618 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10620 if (Subtarget->is64Bit()) {
10621 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10622 TII->get(X86::MOV64rm), X86::RDI)
10624 .addImm(0).addReg(0)
10625 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10626 MI->getOperand(3).getTargetFlags())
10628 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10629 addDirectMem(MIB, X86::RDI);
10630 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10631 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10632 TII->get(X86::MOV32rm), X86::EAX)
10634 .addImm(0).addReg(0)
10635 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10636 MI->getOperand(3).getTargetFlags())
10638 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10639 addDirectMem(MIB, X86::EAX);
10641 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10642 TII->get(X86::MOV32rm), X86::EAX)
10643 .addReg(TII->getGlobalBaseReg(F))
10644 .addImm(0).addReg(0)
10645 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10646 MI->getOperand(3).getTargetFlags())
10648 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10649 addDirectMem(MIB, X86::EAX);
10652 MI->eraseFromParent(); // The pseudo instruction is gone now.
10656 MachineBasicBlock *
10657 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10658 MachineBasicBlock *BB) const {
10659 switch (MI->getOpcode()) {
10660 default: assert(false && "Unexpected instr type to insert");
10661 case X86::TAILJMPd64:
10662 case X86::TAILJMPr64:
10663 case X86::TAILJMPm64:
10664 assert(!"TAILJMP64 would not be touched here.");
10665 case X86::TCRETURNdi64:
10666 case X86::TCRETURNri64:
10667 case X86::TCRETURNmi64:
10668 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10669 // On AMD64, additional defs should be added before register allocation.
10670 if (!Subtarget->isTargetWin64()) {
10671 MI->addRegisterDefined(X86::RSI);
10672 MI->addRegisterDefined(X86::RDI);
10673 MI->addRegisterDefined(X86::XMM6);
10674 MI->addRegisterDefined(X86::XMM7);
10675 MI->addRegisterDefined(X86::XMM8);
10676 MI->addRegisterDefined(X86::XMM9);
10677 MI->addRegisterDefined(X86::XMM10);
10678 MI->addRegisterDefined(X86::XMM11);
10679 MI->addRegisterDefined(X86::XMM12);
10680 MI->addRegisterDefined(X86::XMM13);
10681 MI->addRegisterDefined(X86::XMM14);
10682 MI->addRegisterDefined(X86::XMM15);
10685 case X86::WIN_ALLOCA:
10686 return EmitLoweredWinAlloca(MI, BB);
10687 case X86::TLSCall_32:
10688 case X86::TLSCall_64:
10689 return EmitLoweredTLSCall(MI, BB);
10690 case X86::CMOV_GR8:
10691 case X86::CMOV_FR32:
10692 case X86::CMOV_FR64:
10693 case X86::CMOV_V4F32:
10694 case X86::CMOV_V2F64:
10695 case X86::CMOV_V2I64:
10696 case X86::CMOV_GR16:
10697 case X86::CMOV_GR32:
10698 case X86::CMOV_RFP32:
10699 case X86::CMOV_RFP64:
10700 case X86::CMOV_RFP80:
10701 return EmitLoweredSelect(MI, BB);
10703 case X86::FP32_TO_INT16_IN_MEM:
10704 case X86::FP32_TO_INT32_IN_MEM:
10705 case X86::FP32_TO_INT64_IN_MEM:
10706 case X86::FP64_TO_INT16_IN_MEM:
10707 case X86::FP64_TO_INT32_IN_MEM:
10708 case X86::FP64_TO_INT64_IN_MEM:
10709 case X86::FP80_TO_INT16_IN_MEM:
10710 case X86::FP80_TO_INT32_IN_MEM:
10711 case X86::FP80_TO_INT64_IN_MEM: {
10712 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10713 DebugLoc DL = MI->getDebugLoc();
10715 // Change the floating point control register to use "round towards zero"
10716 // mode when truncating to an integer value.
10717 MachineFunction *F = BB->getParent();
10718 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10719 addFrameReference(BuildMI(*BB, MI, DL,
10720 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10722 // Load the old value of the high byte of the control word...
10724 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10725 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10728 // Set the high part to be round to zero...
10729 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10732 // Reload the modified control word now...
10733 addFrameReference(BuildMI(*BB, MI, DL,
10734 TII->get(X86::FLDCW16m)), CWFrameIdx);
10736 // Restore the memory image of control word to original value
10737 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10740 // Get the X86 opcode to use.
10742 switch (MI->getOpcode()) {
10743 default: llvm_unreachable("illegal opcode!");
10744 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10745 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10746 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10747 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10748 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10749 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10750 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10751 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10752 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10756 MachineOperand &Op = MI->getOperand(0);
10758 AM.BaseType = X86AddressMode::RegBase;
10759 AM.Base.Reg = Op.getReg();
10761 AM.BaseType = X86AddressMode::FrameIndexBase;
10762 AM.Base.FrameIndex = Op.getIndex();
10764 Op = MI->getOperand(1);
10766 AM.Scale = Op.getImm();
10767 Op = MI->getOperand(2);
10769 AM.IndexReg = Op.getImm();
10770 Op = MI->getOperand(3);
10771 if (Op.isGlobal()) {
10772 AM.GV = Op.getGlobal();
10774 AM.Disp = Op.getImm();
10776 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10777 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10779 // Reload the original control word now.
10780 addFrameReference(BuildMI(*BB, MI, DL,
10781 TII->get(X86::FLDCW16m)), CWFrameIdx);
10783 MI->eraseFromParent(); // The pseudo instruction is gone now.
10786 // String/text processing lowering.
10787 case X86::PCMPISTRM128REG:
10788 case X86::VPCMPISTRM128REG:
10789 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10790 case X86::PCMPISTRM128MEM:
10791 case X86::VPCMPISTRM128MEM:
10792 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10793 case X86::PCMPESTRM128REG:
10794 case X86::VPCMPESTRM128REG:
10795 return EmitPCMP(MI, BB, 5, false /* in mem */);
10796 case X86::PCMPESTRM128MEM:
10797 case X86::VPCMPESTRM128MEM:
10798 return EmitPCMP(MI, BB, 5, true /* in mem */);
10800 // Thread synchronization.
10802 return EmitMonitor(MI, BB);
10804 return EmitMwait(MI, BB);
10806 // Atomic Lowering.
10807 case X86::ATOMAND32:
10808 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10809 X86::AND32ri, X86::MOV32rm,
10811 X86::NOT32r, X86::EAX,
10812 X86::GR32RegisterClass);
10813 case X86::ATOMOR32:
10814 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10815 X86::OR32ri, X86::MOV32rm,
10817 X86::NOT32r, X86::EAX,
10818 X86::GR32RegisterClass);
10819 case X86::ATOMXOR32:
10820 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10821 X86::XOR32ri, X86::MOV32rm,
10823 X86::NOT32r, X86::EAX,
10824 X86::GR32RegisterClass);
10825 case X86::ATOMNAND32:
10826 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10827 X86::AND32ri, X86::MOV32rm,
10829 X86::NOT32r, X86::EAX,
10830 X86::GR32RegisterClass, true);
10831 case X86::ATOMMIN32:
10832 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10833 case X86::ATOMMAX32:
10834 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10835 case X86::ATOMUMIN32:
10836 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10837 case X86::ATOMUMAX32:
10838 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10840 case X86::ATOMAND16:
10841 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10842 X86::AND16ri, X86::MOV16rm,
10844 X86::NOT16r, X86::AX,
10845 X86::GR16RegisterClass);
10846 case X86::ATOMOR16:
10847 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10848 X86::OR16ri, X86::MOV16rm,
10850 X86::NOT16r, X86::AX,
10851 X86::GR16RegisterClass);
10852 case X86::ATOMXOR16:
10853 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10854 X86::XOR16ri, X86::MOV16rm,
10856 X86::NOT16r, X86::AX,
10857 X86::GR16RegisterClass);
10858 case X86::ATOMNAND16:
10859 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10860 X86::AND16ri, X86::MOV16rm,
10862 X86::NOT16r, X86::AX,
10863 X86::GR16RegisterClass, true);
10864 case X86::ATOMMIN16:
10865 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10866 case X86::ATOMMAX16:
10867 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10868 case X86::ATOMUMIN16:
10869 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10870 case X86::ATOMUMAX16:
10871 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10873 case X86::ATOMAND8:
10874 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10875 X86::AND8ri, X86::MOV8rm,
10877 X86::NOT8r, X86::AL,
10878 X86::GR8RegisterClass);
10880 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
10881 X86::OR8ri, X86::MOV8rm,
10883 X86::NOT8r, X86::AL,
10884 X86::GR8RegisterClass);
10885 case X86::ATOMXOR8:
10886 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
10887 X86::XOR8ri, X86::MOV8rm,
10889 X86::NOT8r, X86::AL,
10890 X86::GR8RegisterClass);
10891 case X86::ATOMNAND8:
10892 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10893 X86::AND8ri, X86::MOV8rm,
10895 X86::NOT8r, X86::AL,
10896 X86::GR8RegisterClass, true);
10897 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
10898 // This group is for 64-bit host.
10899 case X86::ATOMAND64:
10900 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10901 X86::AND64ri32, X86::MOV64rm,
10903 X86::NOT64r, X86::RAX,
10904 X86::GR64RegisterClass);
10905 case X86::ATOMOR64:
10906 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
10907 X86::OR64ri32, X86::MOV64rm,
10909 X86::NOT64r, X86::RAX,
10910 X86::GR64RegisterClass);
10911 case X86::ATOMXOR64:
10912 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
10913 X86::XOR64ri32, X86::MOV64rm,
10915 X86::NOT64r, X86::RAX,
10916 X86::GR64RegisterClass);
10917 case X86::ATOMNAND64:
10918 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10919 X86::AND64ri32, X86::MOV64rm,
10921 X86::NOT64r, X86::RAX,
10922 X86::GR64RegisterClass, true);
10923 case X86::ATOMMIN64:
10924 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
10925 case X86::ATOMMAX64:
10926 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
10927 case X86::ATOMUMIN64:
10928 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
10929 case X86::ATOMUMAX64:
10930 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
10932 // This group does 64-bit operations on a 32-bit host.
10933 case X86::ATOMAND6432:
10934 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10935 X86::AND32rr, X86::AND32rr,
10936 X86::AND32ri, X86::AND32ri,
10938 case X86::ATOMOR6432:
10939 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10940 X86::OR32rr, X86::OR32rr,
10941 X86::OR32ri, X86::OR32ri,
10943 case X86::ATOMXOR6432:
10944 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10945 X86::XOR32rr, X86::XOR32rr,
10946 X86::XOR32ri, X86::XOR32ri,
10948 case X86::ATOMNAND6432:
10949 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10950 X86::AND32rr, X86::AND32rr,
10951 X86::AND32ri, X86::AND32ri,
10953 case X86::ATOMADD6432:
10954 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10955 X86::ADD32rr, X86::ADC32rr,
10956 X86::ADD32ri, X86::ADC32ri,
10958 case X86::ATOMSUB6432:
10959 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10960 X86::SUB32rr, X86::SBB32rr,
10961 X86::SUB32ri, X86::SBB32ri,
10963 case X86::ATOMSWAP6432:
10964 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10965 X86::MOV32rr, X86::MOV32rr,
10966 X86::MOV32ri, X86::MOV32ri,
10968 case X86::VASTART_SAVE_XMM_REGS:
10969 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
10971 case X86::VAARG_64:
10972 return EmitVAARG64WithCustomInserter(MI, BB);
10976 //===----------------------------------------------------------------------===//
10977 // X86 Optimization Hooks
10978 //===----------------------------------------------------------------------===//
10980 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10984 const SelectionDAG &DAG,
10985 unsigned Depth) const {
10986 unsigned Opc = Op.getOpcode();
10987 assert((Opc >= ISD::BUILTIN_OP_END ||
10988 Opc == ISD::INTRINSIC_WO_CHAIN ||
10989 Opc == ISD::INTRINSIC_W_CHAIN ||
10990 Opc == ISD::INTRINSIC_VOID) &&
10991 "Should use MaskedValueIsZero if you don't know whether Op"
10992 " is a target node!");
10994 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11008 // These nodes' second result is a boolean.
11009 if (Op.getResNo() == 0)
11012 case X86ISD::SETCC:
11013 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11014 Mask.getBitWidth() - 1);
11019 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11020 unsigned Depth) const {
11021 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11022 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11023 return Op.getValueType().getScalarType().getSizeInBits();
11029 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11030 /// node is a GlobalAddress + offset.
11031 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11032 const GlobalValue* &GA,
11033 int64_t &Offset) const {
11034 if (N->getOpcode() == X86ISD::Wrapper) {
11035 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11036 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11037 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11041 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11044 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
11045 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
11046 /// if the load addresses are consecutive, non-overlapping, and in the right
11048 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11049 TargetLowering::DAGCombinerInfo &DCI) {
11050 DebugLoc dl = N->getDebugLoc();
11051 EVT VT = N->getValueType(0);
11053 if (VT.getSizeInBits() != 128)
11056 // Don't create instructions with illegal types after legalize types has run.
11057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11058 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11061 SmallVector<SDValue, 16> Elts;
11062 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11063 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11065 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11068 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11069 /// generation and convert it from being a bunch of shuffles and extracts
11070 /// to a simple store and scalar loads to extract the elements.
11071 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11072 const TargetLowering &TLI) {
11073 SDValue InputVector = N->getOperand(0);
11075 // Only operate on vectors of 4 elements, where the alternative shuffling
11076 // gets to be more expensive.
11077 if (InputVector.getValueType() != MVT::v4i32)
11080 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11081 // single use which is a sign-extend or zero-extend, and all elements are
11083 SmallVector<SDNode *, 4> Uses;
11084 unsigned ExtractedElements = 0;
11085 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11086 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11087 if (UI.getUse().getResNo() != InputVector.getResNo())
11090 SDNode *Extract = *UI;
11091 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11094 if (Extract->getValueType(0) != MVT::i32)
11096 if (!Extract->hasOneUse())
11098 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11099 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11101 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11104 // Record which element was extracted.
11105 ExtractedElements |=
11106 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11108 Uses.push_back(Extract);
11111 // If not all the elements were used, this may not be worthwhile.
11112 if (ExtractedElements != 15)
11115 // Ok, we've now decided to do the transformation.
11116 DebugLoc dl = InputVector.getDebugLoc();
11118 // Store the value to a temporary stack slot.
11119 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11120 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11121 MachinePointerInfo(), false, false, 0);
11123 // Replace each use (extract) with a load of the appropriate element.
11124 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11125 UE = Uses.end(); UI != UE; ++UI) {
11126 SDNode *Extract = *UI;
11128 // cOMpute the element's address.
11129 SDValue Idx = Extract->getOperand(1);
11131 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11132 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11133 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11135 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11136 StackPtr, OffsetVal);
11138 // Load the scalar.
11139 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11140 ScalarAddr, MachinePointerInfo(),
11143 // Replace the exact with the load.
11144 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11147 // The replacement was made in place; don't return anything.
11151 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11152 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11153 const X86Subtarget *Subtarget) {
11154 DebugLoc DL = N->getDebugLoc();
11155 SDValue Cond = N->getOperand(0);
11156 // Get the LHS/RHS of the select.
11157 SDValue LHS = N->getOperand(1);
11158 SDValue RHS = N->getOperand(2);
11160 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11161 // instructions match the semantics of the common C idiom x<y?x:y but not
11162 // x<=y?x:y, because of how they handle negative zero (which can be
11163 // ignored in unsafe-math mode).
11164 if (Subtarget->hasSSE2() &&
11165 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11166 Cond.getOpcode() == ISD::SETCC) {
11167 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11169 unsigned Opcode = 0;
11170 // Check for x CC y ? x : y.
11171 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11172 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11176 // Converting this to a min would handle NaNs incorrectly, and swapping
11177 // the operands would cause it to handle comparisons between positive
11178 // and negative zero incorrectly.
11179 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11180 if (!UnsafeFPMath &&
11181 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11183 std::swap(LHS, RHS);
11185 Opcode = X86ISD::FMIN;
11188 // Converting this to a min would handle comparisons between positive
11189 // and negative zero incorrectly.
11190 if (!UnsafeFPMath &&
11191 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11193 Opcode = X86ISD::FMIN;
11196 // Converting this to a min would handle both negative zeros and NaNs
11197 // incorrectly, but we can swap the operands to fix both.
11198 std::swap(LHS, RHS);
11202 Opcode = X86ISD::FMIN;
11206 // Converting this to a max would handle comparisons between positive
11207 // and negative zero incorrectly.
11208 if (!UnsafeFPMath &&
11209 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11211 Opcode = X86ISD::FMAX;
11214 // Converting this to a max would handle NaNs incorrectly, and swapping
11215 // the operands would cause it to handle comparisons between positive
11216 // and negative zero incorrectly.
11217 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11218 if (!UnsafeFPMath &&
11219 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11221 std::swap(LHS, RHS);
11223 Opcode = X86ISD::FMAX;
11226 // Converting this to a max would handle both negative zeros and NaNs
11227 // incorrectly, but we can swap the operands to fix both.
11228 std::swap(LHS, RHS);
11232 Opcode = X86ISD::FMAX;
11235 // Check for x CC y ? y : x -- a min/max with reversed arms.
11236 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11237 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11241 // Converting this to a min would handle comparisons between positive
11242 // and negative zero incorrectly, and swapping the operands would
11243 // cause it to handle NaNs incorrectly.
11244 if (!UnsafeFPMath &&
11245 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11246 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11248 std::swap(LHS, RHS);
11250 Opcode = X86ISD::FMIN;
11253 // Converting this to a min would handle NaNs incorrectly.
11254 if (!UnsafeFPMath &&
11255 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11257 Opcode = X86ISD::FMIN;
11260 // Converting this to a min would handle both negative zeros and NaNs
11261 // incorrectly, but we can swap the operands to fix both.
11262 std::swap(LHS, RHS);
11266 Opcode = X86ISD::FMIN;
11270 // Converting this to a max would handle NaNs incorrectly.
11271 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11273 Opcode = X86ISD::FMAX;
11276 // Converting this to a max would handle comparisons between positive
11277 // and negative zero incorrectly, and swapping the operands would
11278 // cause it to handle NaNs incorrectly.
11279 if (!UnsafeFPMath &&
11280 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11281 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11283 std::swap(LHS, RHS);
11285 Opcode = X86ISD::FMAX;
11288 // Converting this to a max would handle both negative zeros and NaNs
11289 // incorrectly, but we can swap the operands to fix both.
11290 std::swap(LHS, RHS);
11294 Opcode = X86ISD::FMAX;
11300 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11303 // If this is a select between two integer constants, try to do some
11305 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11306 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11307 // Don't do this for crazy integer types.
11308 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11309 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11310 // so that TrueC (the true value) is larger than FalseC.
11311 bool NeedsCondInvert = false;
11313 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11314 // Efficiently invertible.
11315 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11316 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11317 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11318 NeedsCondInvert = true;
11319 std::swap(TrueC, FalseC);
11322 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11323 if (FalseC->getAPIntValue() == 0 &&
11324 TrueC->getAPIntValue().isPowerOf2()) {
11325 if (NeedsCondInvert) // Invert the condition if needed.
11326 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11327 DAG.getConstant(1, Cond.getValueType()));
11329 // Zero extend the condition if needed.
11330 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11332 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11333 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11334 DAG.getConstant(ShAmt, MVT::i8));
11337 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11338 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11339 if (NeedsCondInvert) // Invert the condition if needed.
11340 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11341 DAG.getConstant(1, Cond.getValueType()));
11343 // Zero extend the condition if needed.
11344 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11345 FalseC->getValueType(0), Cond);
11346 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11347 SDValue(FalseC, 0));
11350 // Optimize cases that will turn into an LEA instruction. This requires
11351 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11352 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11353 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11354 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11356 bool isFastMultiplier = false;
11358 switch ((unsigned char)Diff) {
11360 case 1: // result = add base, cond
11361 case 2: // result = lea base( , cond*2)
11362 case 3: // result = lea base(cond, cond*2)
11363 case 4: // result = lea base( , cond*4)
11364 case 5: // result = lea base(cond, cond*4)
11365 case 8: // result = lea base( , cond*8)
11366 case 9: // result = lea base(cond, cond*8)
11367 isFastMultiplier = true;
11372 if (isFastMultiplier) {
11373 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11374 if (NeedsCondInvert) // Invert the condition if needed.
11375 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11376 DAG.getConstant(1, Cond.getValueType()));
11378 // Zero extend the condition if needed.
11379 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11381 // Scale the condition by the difference.
11383 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11384 DAG.getConstant(Diff, Cond.getValueType()));
11386 // Add the base if non-zero.
11387 if (FalseC->getAPIntValue() != 0)
11388 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11389 SDValue(FalseC, 0));
11399 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11400 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11401 TargetLowering::DAGCombinerInfo &DCI) {
11402 DebugLoc DL = N->getDebugLoc();
11404 // If the flag operand isn't dead, don't touch this CMOV.
11405 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11408 SDValue FalseOp = N->getOperand(0);
11409 SDValue TrueOp = N->getOperand(1);
11410 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11411 SDValue Cond = N->getOperand(3);
11412 if (CC == X86::COND_E || CC == X86::COND_NE) {
11413 switch (Cond.getOpcode()) {
11417 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11418 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11419 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11423 // If this is a select between two integer constants, try to do some
11424 // optimizations. Note that the operands are ordered the opposite of SELECT
11426 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11427 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11428 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11429 // larger than FalseC (the false value).
11430 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11431 CC = X86::GetOppositeBranchCondition(CC);
11432 std::swap(TrueC, FalseC);
11435 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11436 // This is efficient for any integer data type (including i8/i16) and
11438 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11439 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11440 DAG.getConstant(CC, MVT::i8), Cond);
11442 // Zero extend the condition if needed.
11443 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11445 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11446 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11447 DAG.getConstant(ShAmt, MVT::i8));
11448 if (N->getNumValues() == 2) // Dead flag value?
11449 return DCI.CombineTo(N, Cond, SDValue());
11453 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11454 // for any integer data type, including i8/i16.
11455 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11456 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11457 DAG.getConstant(CC, MVT::i8), Cond);
11459 // Zero extend the condition if needed.
11460 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11461 FalseC->getValueType(0), Cond);
11462 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11463 SDValue(FalseC, 0));
11465 if (N->getNumValues() == 2) // Dead flag value?
11466 return DCI.CombineTo(N, Cond, SDValue());
11470 // Optimize cases that will turn into an LEA instruction. This requires
11471 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11472 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11473 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11474 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11476 bool isFastMultiplier = false;
11478 switch ((unsigned char)Diff) {
11480 case 1: // result = add base, cond
11481 case 2: // result = lea base( , cond*2)
11482 case 3: // result = lea base(cond, cond*2)
11483 case 4: // result = lea base( , cond*4)
11484 case 5: // result = lea base(cond, cond*4)
11485 case 8: // result = lea base( , cond*8)
11486 case 9: // result = lea base(cond, cond*8)
11487 isFastMultiplier = true;
11492 if (isFastMultiplier) {
11493 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11494 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11495 DAG.getConstant(CC, MVT::i8), Cond);
11496 // Zero extend the condition if needed.
11497 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11499 // Scale the condition by the difference.
11501 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11502 DAG.getConstant(Diff, Cond.getValueType()));
11504 // Add the base if non-zero.
11505 if (FalseC->getAPIntValue() != 0)
11506 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11507 SDValue(FalseC, 0));
11508 if (N->getNumValues() == 2) // Dead flag value?
11509 return DCI.CombineTo(N, Cond, SDValue());
11519 /// PerformMulCombine - Optimize a single multiply with constant into two
11520 /// in order to implement it with two cheaper instructions, e.g.
11521 /// LEA + SHL, LEA + LEA.
11522 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11523 TargetLowering::DAGCombinerInfo &DCI) {
11524 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11527 EVT VT = N->getValueType(0);
11528 if (VT != MVT::i64)
11531 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11534 uint64_t MulAmt = C->getZExtValue();
11535 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11538 uint64_t MulAmt1 = 0;
11539 uint64_t MulAmt2 = 0;
11540 if ((MulAmt % 9) == 0) {
11542 MulAmt2 = MulAmt / 9;
11543 } else if ((MulAmt % 5) == 0) {
11545 MulAmt2 = MulAmt / 5;
11546 } else if ((MulAmt % 3) == 0) {
11548 MulAmt2 = MulAmt / 3;
11551 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11552 DebugLoc DL = N->getDebugLoc();
11554 if (isPowerOf2_64(MulAmt2) &&
11555 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11556 // If second multiplifer is pow2, issue it first. We want the multiply by
11557 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11559 std::swap(MulAmt1, MulAmt2);
11562 if (isPowerOf2_64(MulAmt1))
11563 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11564 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11566 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11567 DAG.getConstant(MulAmt1, VT));
11569 if (isPowerOf2_64(MulAmt2))
11570 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11571 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11573 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11574 DAG.getConstant(MulAmt2, VT));
11576 // Do not add new nodes to DAG combiner worklist.
11577 DCI.CombineTo(N, NewMul, false);
11582 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11583 SDValue N0 = N->getOperand(0);
11584 SDValue N1 = N->getOperand(1);
11585 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11586 EVT VT = N0.getValueType();
11588 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11589 // since the result of setcc_c is all zero's or all ones.
11590 if (N1C && N0.getOpcode() == ISD::AND &&
11591 N0.getOperand(1).getOpcode() == ISD::Constant) {
11592 SDValue N00 = N0.getOperand(0);
11593 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11594 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11595 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11596 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11597 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11598 APInt ShAmt = N1C->getAPIntValue();
11599 Mask = Mask.shl(ShAmt);
11601 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11602 N00, DAG.getConstant(Mask, VT));
11609 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11611 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11612 const X86Subtarget *Subtarget) {
11613 EVT VT = N->getValueType(0);
11614 if (!VT.isVector() && VT.isInteger() &&
11615 N->getOpcode() == ISD::SHL)
11616 return PerformSHLCombine(N, DAG);
11618 // On X86 with SSE2 support, we can transform this to a vector shift if
11619 // all elements are shifted by the same amount. We can't do this in legalize
11620 // because the a constant vector is typically transformed to a constant pool
11621 // so we have no knowledge of the shift amount.
11622 if (!Subtarget->hasSSE2())
11625 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11628 SDValue ShAmtOp = N->getOperand(1);
11629 EVT EltVT = VT.getVectorElementType();
11630 DebugLoc DL = N->getDebugLoc();
11631 SDValue BaseShAmt = SDValue();
11632 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11633 unsigned NumElts = VT.getVectorNumElements();
11635 for (; i != NumElts; ++i) {
11636 SDValue Arg = ShAmtOp.getOperand(i);
11637 if (Arg.getOpcode() == ISD::UNDEF) continue;
11641 for (; i != NumElts; ++i) {
11642 SDValue Arg = ShAmtOp.getOperand(i);
11643 if (Arg.getOpcode() == ISD::UNDEF) continue;
11644 if (Arg != BaseShAmt) {
11648 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11649 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11650 SDValue InVec = ShAmtOp.getOperand(0);
11651 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11652 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11654 for (; i != NumElts; ++i) {
11655 SDValue Arg = InVec.getOperand(i);
11656 if (Arg.getOpcode() == ISD::UNDEF) continue;
11660 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11661 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11662 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11663 if (C->getZExtValue() == SplatIdx)
11664 BaseShAmt = InVec.getOperand(1);
11667 if (BaseShAmt.getNode() == 0)
11668 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11669 DAG.getIntPtrConstant(0));
11673 // The shift amount is an i32.
11674 if (EltVT.bitsGT(MVT::i32))
11675 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11676 else if (EltVT.bitsLT(MVT::i32))
11677 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11679 // The shift amount is identical so we can do a vector shift.
11680 SDValue ValOp = N->getOperand(0);
11681 switch (N->getOpcode()) {
11683 llvm_unreachable("Unknown shift opcode!");
11686 if (VT == MVT::v2i64)
11687 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11688 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11690 if (VT == MVT::v4i32)
11691 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11692 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11694 if (VT == MVT::v8i16)
11695 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11696 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11700 if (VT == MVT::v4i32)
11701 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11702 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11704 if (VT == MVT::v8i16)
11705 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11706 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11710 if (VT == MVT::v2i64)
11711 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11712 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11714 if (VT == MVT::v4i32)
11715 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11716 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11718 if (VT == MVT::v8i16)
11719 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11720 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11728 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
11729 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
11730 // and friends. Likewise for OR -> CMPNEQSS.
11731 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
11732 TargetLowering::DAGCombinerInfo &DCI,
11733 const X86Subtarget *Subtarget) {
11736 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
11737 // we're requiring SSE2 for both.
11738 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
11739 SDValue N0 = N->getOperand(0);
11740 SDValue N1 = N->getOperand(1);
11741 SDValue CMP0 = N0->getOperand(1);
11742 SDValue CMP1 = N1->getOperand(1);
11743 DebugLoc DL = N->getDebugLoc();
11745 // The SETCCs should both refer to the same CMP.
11746 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
11749 SDValue CMP00 = CMP0->getOperand(0);
11750 SDValue CMP01 = CMP0->getOperand(1);
11751 EVT VT = CMP00.getValueType();
11753 if (VT == MVT::f32 || VT == MVT::f64) {
11754 bool ExpectingFlags = false;
11755 // Check for any users that want flags:
11756 for (SDNode::use_iterator UI = N->use_begin(),
11758 !ExpectingFlags && UI != UE; ++UI)
11759 switch (UI->getOpcode()) {
11764 ExpectingFlags = true;
11766 case ISD::CopyToReg:
11767 case ISD::SIGN_EXTEND:
11768 case ISD::ZERO_EXTEND:
11769 case ISD::ANY_EXTEND:
11773 if (!ExpectingFlags) {
11774 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
11775 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
11777 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
11778 X86::CondCode tmp = cc0;
11783 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
11784 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
11785 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
11786 X86ISD::NodeType NTOperator = is64BitFP ?
11787 X86ISD::FSETCCsd : X86ISD::FSETCCss;
11788 // FIXME: need symbolic constants for these magic numbers.
11789 // See X86ATTInstPrinter.cpp:printSSECC().
11790 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
11791 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
11792 DAG.getConstant(x86cc, MVT::i8));
11793 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
11795 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
11796 DAG.getConstant(1, MVT::i32));
11797 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
11798 return OneBitOfTruth;
11806 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11807 TargetLowering::DAGCombinerInfo &DCI,
11808 const X86Subtarget *Subtarget) {
11809 if (DCI.isBeforeLegalizeOps())
11812 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
11816 // Want to form PANDN nodes, in the hopes of then easily combining them with
11817 // OR and AND nodes to form PBLEND/PSIGN.
11818 EVT VT = N->getValueType(0);
11819 if (VT != MVT::v2i64)
11822 SDValue N0 = N->getOperand(0);
11823 SDValue N1 = N->getOperand(1);
11824 DebugLoc DL = N->getDebugLoc();
11826 // Check LHS for vnot
11827 if (N0.getOpcode() == ISD::XOR &&
11828 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
11829 return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
11831 // Check RHS for vnot
11832 if (N1.getOpcode() == ISD::XOR &&
11833 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
11834 return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
11839 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
11840 TargetLowering::DAGCombinerInfo &DCI,
11841 const X86Subtarget *Subtarget) {
11842 if (DCI.isBeforeLegalizeOps())
11845 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
11849 EVT VT = N->getValueType(0);
11850 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
11853 SDValue N0 = N->getOperand(0);
11854 SDValue N1 = N->getOperand(1);
11856 // look for psign/blend
11857 if (Subtarget->hasSSSE3()) {
11858 if (VT == MVT::v2i64) {
11859 // Canonicalize pandn to RHS
11860 if (N0.getOpcode() == X86ISD::PANDN)
11862 // or (and (m, x), (pandn m, y))
11863 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
11864 SDValue Mask = N1.getOperand(0);
11865 SDValue X = N1.getOperand(1);
11867 if (N0.getOperand(0) == Mask)
11868 Y = N0.getOperand(1);
11869 if (N0.getOperand(1) == Mask)
11870 Y = N0.getOperand(0);
11872 // Check to see if the mask appeared in both the AND and PANDN and
11876 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
11877 if (Mask.getOpcode() != ISD::BITCAST ||
11878 X.getOpcode() != ISD::BITCAST ||
11879 Y.getOpcode() != ISD::BITCAST)
11882 // Look through mask bitcast.
11883 Mask = Mask.getOperand(0);
11884 EVT MaskVT = Mask.getValueType();
11886 // Validate that the Mask operand is a vector sra node. The sra node
11887 // will be an intrinsic.
11888 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
11891 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
11892 // there is no psrai.b
11893 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
11894 case Intrinsic::x86_sse2_psrai_w:
11895 case Intrinsic::x86_sse2_psrai_d:
11897 default: return SDValue();
11900 // Check that the SRA is all signbits.
11901 SDValue SraC = Mask.getOperand(2);
11902 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
11903 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
11904 if ((SraAmt + 1) != EltBits)
11907 DebugLoc DL = N->getDebugLoc();
11909 // Now we know we at least have a plendvb with the mask val. See if
11910 // we can form a psignb/w/d.
11911 // psign = x.type == y.type == mask.type && y = sub(0, x);
11912 X = X.getOperand(0);
11913 Y = Y.getOperand(0);
11914 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
11915 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
11916 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
11919 case 8: Opc = X86ISD::PSIGNB; break;
11920 case 16: Opc = X86ISD::PSIGNW; break;
11921 case 32: Opc = X86ISD::PSIGND; break;
11925 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
11926 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
11929 // PBLENDVB only available on SSE 4.1
11930 if (!Subtarget->hasSSE41())
11933 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
11934 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
11935 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
11936 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
11937 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
11942 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
11943 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
11945 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
11947 if (!N0.hasOneUse() || !N1.hasOneUse())
11950 SDValue ShAmt0 = N0.getOperand(1);
11951 if (ShAmt0.getValueType() != MVT::i8)
11953 SDValue ShAmt1 = N1.getOperand(1);
11954 if (ShAmt1.getValueType() != MVT::i8)
11956 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
11957 ShAmt0 = ShAmt0.getOperand(0);
11958 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
11959 ShAmt1 = ShAmt1.getOperand(0);
11961 DebugLoc DL = N->getDebugLoc();
11962 unsigned Opc = X86ISD::SHLD;
11963 SDValue Op0 = N0.getOperand(0);
11964 SDValue Op1 = N1.getOperand(0);
11965 if (ShAmt0.getOpcode() == ISD::SUB) {
11966 Opc = X86ISD::SHRD;
11967 std::swap(Op0, Op1);
11968 std::swap(ShAmt0, ShAmt1);
11971 unsigned Bits = VT.getSizeInBits();
11972 if (ShAmt1.getOpcode() == ISD::SUB) {
11973 SDValue Sum = ShAmt1.getOperand(0);
11974 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
11975 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
11976 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
11977 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
11978 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
11979 return DAG.getNode(Opc, DL, VT,
11981 DAG.getNode(ISD::TRUNCATE, DL,
11984 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
11985 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
11987 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
11988 return DAG.getNode(Opc, DL, VT,
11989 N0.getOperand(0), N1.getOperand(0),
11990 DAG.getNode(ISD::TRUNCATE, DL,
11997 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
11998 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
11999 const X86Subtarget *Subtarget) {
12000 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12001 // the FP state in cases where an emms may be missing.
12002 // A preferable solution to the general problem is to figure out the right
12003 // places to insert EMMS. This qualifies as a quick hack.
12005 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12006 StoreSDNode *St = cast<StoreSDNode>(N);
12007 EVT VT = St->getValue().getValueType();
12008 if (VT.getSizeInBits() != 64)
12011 const Function *F = DAG.getMachineFunction().getFunction();
12012 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12013 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12014 && Subtarget->hasSSE2();
12015 if ((VT.isVector() ||
12016 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12017 isa<LoadSDNode>(St->getValue()) &&
12018 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12019 St->getChain().hasOneUse() && !St->isVolatile()) {
12020 SDNode* LdVal = St->getValue().getNode();
12021 LoadSDNode *Ld = 0;
12022 int TokenFactorIndex = -1;
12023 SmallVector<SDValue, 8> Ops;
12024 SDNode* ChainVal = St->getChain().getNode();
12025 // Must be a store of a load. We currently handle two cases: the load
12026 // is a direct child, and it's under an intervening TokenFactor. It is
12027 // possible to dig deeper under nested TokenFactors.
12028 if (ChainVal == LdVal)
12029 Ld = cast<LoadSDNode>(St->getChain());
12030 else if (St->getValue().hasOneUse() &&
12031 ChainVal->getOpcode() == ISD::TokenFactor) {
12032 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12033 if (ChainVal->getOperand(i).getNode() == LdVal) {
12034 TokenFactorIndex = i;
12035 Ld = cast<LoadSDNode>(St->getValue());
12037 Ops.push_back(ChainVal->getOperand(i));
12041 if (!Ld || !ISD::isNormalLoad(Ld))
12044 // If this is not the MMX case, i.e. we are just turning i64 load/store
12045 // into f64 load/store, avoid the transformation if there are multiple
12046 // uses of the loaded value.
12047 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12050 DebugLoc LdDL = Ld->getDebugLoc();
12051 DebugLoc StDL = N->getDebugLoc();
12052 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12053 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12055 if (Subtarget->is64Bit() || F64IsLegal) {
12056 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12057 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12058 Ld->getPointerInfo(), Ld->isVolatile(),
12059 Ld->isNonTemporal(), Ld->getAlignment());
12060 SDValue NewChain = NewLd.getValue(1);
12061 if (TokenFactorIndex != -1) {
12062 Ops.push_back(NewChain);
12063 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12066 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12067 St->getPointerInfo(),
12068 St->isVolatile(), St->isNonTemporal(),
12069 St->getAlignment());
12072 // Otherwise, lower to two pairs of 32-bit loads / stores.
12073 SDValue LoAddr = Ld->getBasePtr();
12074 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12075 DAG.getConstant(4, MVT::i32));
12077 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12078 Ld->getPointerInfo(),
12079 Ld->isVolatile(), Ld->isNonTemporal(),
12080 Ld->getAlignment());
12081 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12082 Ld->getPointerInfo().getWithOffset(4),
12083 Ld->isVolatile(), Ld->isNonTemporal(),
12084 MinAlign(Ld->getAlignment(), 4));
12086 SDValue NewChain = LoLd.getValue(1);
12087 if (TokenFactorIndex != -1) {
12088 Ops.push_back(LoLd);
12089 Ops.push_back(HiLd);
12090 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12094 LoAddr = St->getBasePtr();
12095 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12096 DAG.getConstant(4, MVT::i32));
12098 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12099 St->getPointerInfo(),
12100 St->isVolatile(), St->isNonTemporal(),
12101 St->getAlignment());
12102 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12103 St->getPointerInfo().getWithOffset(4),
12105 St->isNonTemporal(),
12106 MinAlign(St->getAlignment(), 4));
12107 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12112 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12113 /// X86ISD::FXOR nodes.
12114 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12115 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12116 // F[X]OR(0.0, x) -> x
12117 // F[X]OR(x, 0.0) -> x
12118 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12119 if (C->getValueAPF().isPosZero())
12120 return N->getOperand(1);
12121 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12122 if (C->getValueAPF().isPosZero())
12123 return N->getOperand(0);
12127 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12128 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12129 // FAND(0.0, x) -> 0.0
12130 // FAND(x, 0.0) -> 0.0
12131 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12132 if (C->getValueAPF().isPosZero())
12133 return N->getOperand(0);
12134 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12135 if (C->getValueAPF().isPosZero())
12136 return N->getOperand(1);
12140 static SDValue PerformBTCombine(SDNode *N,
12142 TargetLowering::DAGCombinerInfo &DCI) {
12143 // BT ignores high bits in the bit index operand.
12144 SDValue Op1 = N->getOperand(1);
12145 if (Op1.hasOneUse()) {
12146 unsigned BitWidth = Op1.getValueSizeInBits();
12147 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12148 APInt KnownZero, KnownOne;
12149 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12150 !DCI.isBeforeLegalizeOps());
12151 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12152 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12153 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12154 DCI.CommitTargetLoweringOpt(TLO);
12159 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12160 SDValue Op = N->getOperand(0);
12161 if (Op.getOpcode() == ISD::BITCAST)
12162 Op = Op.getOperand(0);
12163 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12164 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12165 VT.getVectorElementType().getSizeInBits() ==
12166 OpVT.getVectorElementType().getSizeInBits()) {
12167 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12172 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12173 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12174 // (and (i32 x86isd::setcc_carry), 1)
12175 // This eliminates the zext. This transformation is necessary because
12176 // ISD::SETCC is always legalized to i8.
12177 DebugLoc dl = N->getDebugLoc();
12178 SDValue N0 = N->getOperand(0);
12179 EVT VT = N->getValueType(0);
12180 if (N0.getOpcode() == ISD::AND &&
12182 N0.getOperand(0).hasOneUse()) {
12183 SDValue N00 = N0.getOperand(0);
12184 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12186 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12187 if (!C || C->getZExtValue() != 1)
12189 return DAG.getNode(ISD::AND, dl, VT,
12190 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12191 N00.getOperand(0), N00.getOperand(1)),
12192 DAG.getConstant(1, VT));
12198 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12199 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12200 unsigned X86CC = N->getConstantOperandVal(0);
12201 SDValue EFLAG = N->getOperand(1);
12202 DebugLoc DL = N->getDebugLoc();
12204 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12205 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12207 if (X86CC == X86::COND_B)
12208 return DAG.getNode(ISD::AND, DL, MVT::i8,
12209 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12210 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12211 DAG.getConstant(1, MVT::i8));
12216 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12217 const X86TargetLowering *XTLI) {
12218 SDValue Op0 = N->getOperand(0);
12219 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12220 // a 32-bit target where SSE doesn't support i64->FP operations.
12221 if (Op0.getOpcode() == ISD::LOAD) {
12222 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12223 EVT VT = Ld->getValueType(0);
12224 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12225 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12226 !XTLI->getSubtarget()->is64Bit() &&
12227 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12228 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12229 Ld->getChain(), Op0, DAG);
12230 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12237 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12238 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12239 X86TargetLowering::DAGCombinerInfo &DCI) {
12240 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12241 // the result is either zero or one (depending on the input carry bit).
12242 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12243 if (X86::isZeroNode(N->getOperand(0)) &&
12244 X86::isZeroNode(N->getOperand(1)) &&
12245 // We don't have a good way to replace an EFLAGS use, so only do this when
12247 SDValue(N, 1).use_empty()) {
12248 DebugLoc DL = N->getDebugLoc();
12249 EVT VT = N->getValueType(0);
12250 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12251 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12252 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12253 DAG.getConstant(X86::COND_B,MVT::i8),
12255 DAG.getConstant(1, VT));
12256 return DCI.CombineTo(N, Res1, CarryOut);
12262 // fold (add Y, (sete X, 0)) -> adc 0, Y
12263 // (add Y, (setne X, 0)) -> sbb -1, Y
12264 // (sub (sete X, 0), Y) -> sbb 0, Y
12265 // (sub (setne X, 0), Y) -> adc -1, Y
12266 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
12267 DebugLoc DL = N->getDebugLoc();
12269 // Look through ZExts.
12270 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12271 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12274 SDValue SetCC = Ext.getOperand(0);
12275 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12278 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12279 if (CC != X86::COND_E && CC != X86::COND_NE)
12282 SDValue Cmp = SetCC.getOperand(1);
12283 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12284 !X86::isZeroNode(Cmp.getOperand(1)) ||
12285 !Cmp.getOperand(0).getValueType().isInteger())
12288 SDValue CmpOp0 = Cmp.getOperand(0);
12289 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12290 DAG.getConstant(1, CmpOp0.getValueType()));
12292 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12293 if (CC == X86::COND_NE)
12294 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12295 DL, OtherVal.getValueType(), OtherVal,
12296 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12297 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12298 DL, OtherVal.getValueType(), OtherVal,
12299 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12302 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12303 DAGCombinerInfo &DCI) const {
12304 SelectionDAG &DAG = DCI.DAG;
12305 switch (N->getOpcode()) {
12307 case ISD::EXTRACT_VECTOR_ELT:
12308 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12309 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12310 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12312 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
12313 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12314 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12317 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12318 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12319 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12320 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12321 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12323 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12324 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12325 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12326 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12327 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12328 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12329 case X86ISD::SHUFPS: // Handle all target specific shuffles
12330 case X86ISD::SHUFPD:
12331 case X86ISD::PALIGN:
12332 case X86ISD::PUNPCKHBW:
12333 case X86ISD::PUNPCKHWD:
12334 case X86ISD::PUNPCKHDQ:
12335 case X86ISD::PUNPCKHQDQ:
12336 case X86ISD::UNPCKHPS:
12337 case X86ISD::UNPCKHPD:
12338 case X86ISD::PUNPCKLBW:
12339 case X86ISD::PUNPCKLWD:
12340 case X86ISD::PUNPCKLDQ:
12341 case X86ISD::PUNPCKLQDQ:
12342 case X86ISD::UNPCKLPS:
12343 case X86ISD::UNPCKLPD:
12344 case X86ISD::VUNPCKLPS:
12345 case X86ISD::VUNPCKLPD:
12346 case X86ISD::VUNPCKLPSY:
12347 case X86ISD::VUNPCKLPDY:
12348 case X86ISD::MOVHLPS:
12349 case X86ISD::MOVLHPS:
12350 case X86ISD::PSHUFD:
12351 case X86ISD::PSHUFHW:
12352 case X86ISD::PSHUFLW:
12353 case X86ISD::MOVSS:
12354 case X86ISD::MOVSD:
12355 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12361 /// isTypeDesirableForOp - Return true if the target has native support for
12362 /// the specified value type and it is 'desirable' to use the type for the
12363 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12364 /// instruction encodings are longer and some i16 instructions are slow.
12365 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12366 if (!isTypeLegal(VT))
12368 if (VT != MVT::i16)
12375 case ISD::SIGN_EXTEND:
12376 case ISD::ZERO_EXTEND:
12377 case ISD::ANY_EXTEND:
12390 /// IsDesirableToPromoteOp - This method query the target whether it is
12391 /// beneficial for dag combiner to promote the specified node. If true, it
12392 /// should return the desired promotion type by reference.
12393 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12394 EVT VT = Op.getValueType();
12395 if (VT != MVT::i16)
12398 bool Promote = false;
12399 bool Commute = false;
12400 switch (Op.getOpcode()) {
12403 LoadSDNode *LD = cast<LoadSDNode>(Op);
12404 // If the non-extending load has a single use and it's not live out, then it
12405 // might be folded.
12406 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12407 Op.hasOneUse()*/) {
12408 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12409 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12410 // The only case where we'd want to promote LOAD (rather then it being
12411 // promoted as an operand is when it's only use is liveout.
12412 if (UI->getOpcode() != ISD::CopyToReg)
12419 case ISD::SIGN_EXTEND:
12420 case ISD::ZERO_EXTEND:
12421 case ISD::ANY_EXTEND:
12426 SDValue N0 = Op.getOperand(0);
12427 // Look out for (store (shl (load), x)).
12428 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12441 SDValue N0 = Op.getOperand(0);
12442 SDValue N1 = Op.getOperand(1);
12443 if (!Commute && MayFoldLoad(N1))
12445 // Avoid disabling potential load folding opportunities.
12446 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12448 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
12458 //===----------------------------------------------------------------------===//
12459 // X86 Inline Assembly Support
12460 //===----------------------------------------------------------------------===//
12462 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
12463 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
12465 std::string AsmStr = IA->getAsmString();
12467 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
12468 SmallVector<StringRef, 4> AsmPieces;
12469 SplitString(AsmStr, AsmPieces, ";\n");
12471 switch (AsmPieces.size()) {
12472 default: return false;
12474 AsmStr = AsmPieces[0];
12476 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
12478 // FIXME: this should verify that we are targeting a 486 or better. If not,
12479 // we will turn this bswap into something that will be lowered to logical ops
12480 // instead of emitting the bswap asm. For now, we don't support 486 or lower
12481 // so don't worry about this.
12483 if (AsmPieces.size() == 2 &&
12484 (AsmPieces[0] == "bswap" ||
12485 AsmPieces[0] == "bswapq" ||
12486 AsmPieces[0] == "bswapl") &&
12487 (AsmPieces[1] == "$0" ||
12488 AsmPieces[1] == "${0:q}")) {
12489 // No need to check constraints, nothing other than the equivalent of
12490 // "=r,0" would be valid here.
12491 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12492 if (!Ty || Ty->getBitWidth() % 16 != 0)
12494 return IntrinsicLowering::LowerToByteSwap(CI);
12496 // rorw $$8, ${0:w} --> llvm.bswap.i16
12497 if (CI->getType()->isIntegerTy(16) &&
12498 AsmPieces.size() == 3 &&
12499 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
12500 AsmPieces[1] == "$$8," &&
12501 AsmPieces[2] == "${0:w}" &&
12502 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12504 const std::string &ConstraintsStr = IA->getConstraintString();
12505 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12506 std::sort(AsmPieces.begin(), AsmPieces.end());
12507 if (AsmPieces.size() == 4 &&
12508 AsmPieces[0] == "~{cc}" &&
12509 AsmPieces[1] == "~{dirflag}" &&
12510 AsmPieces[2] == "~{flags}" &&
12511 AsmPieces[3] == "~{fpsr}") {
12512 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12513 if (!Ty || Ty->getBitWidth() % 16 != 0)
12515 return IntrinsicLowering::LowerToByteSwap(CI);
12520 if (CI->getType()->isIntegerTy(32) &&
12521 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12522 SmallVector<StringRef, 4> Words;
12523 SplitString(AsmPieces[0], Words, " \t,");
12524 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12525 Words[2] == "${0:w}") {
12527 SplitString(AsmPieces[1], Words, " \t,");
12528 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12529 Words[2] == "$0") {
12531 SplitString(AsmPieces[2], Words, " \t,");
12532 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12533 Words[2] == "${0:w}") {
12535 const std::string &ConstraintsStr = IA->getConstraintString();
12536 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12537 std::sort(AsmPieces.begin(), AsmPieces.end());
12538 if (AsmPieces.size() == 4 &&
12539 AsmPieces[0] == "~{cc}" &&
12540 AsmPieces[1] == "~{dirflag}" &&
12541 AsmPieces[2] == "~{flags}" &&
12542 AsmPieces[3] == "~{fpsr}") {
12543 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12544 if (!Ty || Ty->getBitWidth() % 16 != 0)
12546 return IntrinsicLowering::LowerToByteSwap(CI);
12553 if (CI->getType()->isIntegerTy(64)) {
12554 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12555 if (Constraints.size() >= 2 &&
12556 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12557 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12558 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12559 SmallVector<StringRef, 4> Words;
12560 SplitString(AsmPieces[0], Words, " \t");
12561 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12563 SplitString(AsmPieces[1], Words, " \t");
12564 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12566 SplitString(AsmPieces[2], Words, " \t,");
12567 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12568 Words[2] == "%edx") {
12569 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12570 if (!Ty || Ty->getBitWidth() % 16 != 0)
12572 return IntrinsicLowering::LowerToByteSwap(CI);
12585 /// getConstraintType - Given a constraint letter, return the type of
12586 /// constraint it is for this target.
12587 X86TargetLowering::ConstraintType
12588 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12589 if (Constraint.size() == 1) {
12590 switch (Constraint[0]) {
12600 return C_RegisterClass;
12624 return TargetLowering::getConstraintType(Constraint);
12627 /// Examine constraint type and operand type and determine a weight value.
12628 /// This object must already have been set up with the operand type
12629 /// and the current alternative constraint selected.
12630 TargetLowering::ConstraintWeight
12631 X86TargetLowering::getSingleConstraintMatchWeight(
12632 AsmOperandInfo &info, const char *constraint) const {
12633 ConstraintWeight weight = CW_Invalid;
12634 Value *CallOperandVal = info.CallOperandVal;
12635 // If we don't have a value, we can't do a match,
12636 // but allow it at the lowest weight.
12637 if (CallOperandVal == NULL)
12639 const Type *type = CallOperandVal->getType();
12640 // Look at the constraint type.
12641 switch (*constraint) {
12643 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12654 if (CallOperandVal->getType()->isIntegerTy())
12655 weight = CW_SpecificReg;
12660 if (type->isFloatingPointTy())
12661 weight = CW_SpecificReg;
12664 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12665 weight = CW_SpecificReg;
12669 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12670 weight = CW_Register;
12673 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12674 if (C->getZExtValue() <= 31)
12675 weight = CW_Constant;
12679 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12680 if (C->getZExtValue() <= 63)
12681 weight = CW_Constant;
12685 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12686 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12687 weight = CW_Constant;
12691 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12692 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12693 weight = CW_Constant;
12697 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12698 if (C->getZExtValue() <= 3)
12699 weight = CW_Constant;
12703 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12704 if (C->getZExtValue() <= 0xff)
12705 weight = CW_Constant;
12710 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12711 weight = CW_Constant;
12715 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12716 if ((C->getSExtValue() >= -0x80000000LL) &&
12717 (C->getSExtValue() <= 0x7fffffffLL))
12718 weight = CW_Constant;
12722 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12723 if (C->getZExtValue() <= 0xffffffff)
12724 weight = CW_Constant;
12731 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12732 /// with another that has more specific requirements based on the type of the
12733 /// corresponding operand.
12734 const char *X86TargetLowering::
12735 LowerXConstraint(EVT ConstraintVT) const {
12736 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12737 // 'f' like normal targets.
12738 if (ConstraintVT.isFloatingPoint()) {
12739 if (Subtarget->hasXMMInt())
12741 if (Subtarget->hasXMM())
12745 return TargetLowering::LowerXConstraint(ConstraintVT);
12748 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12749 /// vector. If it is invalid, don't add anything to Ops.
12750 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12751 std::string &Constraint,
12752 std::vector<SDValue>&Ops,
12753 SelectionDAG &DAG) const {
12754 SDValue Result(0, 0);
12756 // Only support length 1 constraints for now.
12757 if (Constraint.length() > 1) return;
12759 char ConstraintLetter = Constraint[0];
12760 switch (ConstraintLetter) {
12763 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12764 if (C->getZExtValue() <= 31) {
12765 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12771 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12772 if (C->getZExtValue() <= 63) {
12773 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12779 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12780 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12781 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12787 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12788 if (C->getZExtValue() <= 255) {
12789 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12795 // 32-bit signed value
12796 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12797 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12798 C->getSExtValue())) {
12799 // Widen to 64 bits here to get it sign extended.
12800 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12803 // FIXME gcc accepts some relocatable values here too, but only in certain
12804 // memory models; it's complicated.
12809 // 32-bit unsigned value
12810 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12811 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12812 C->getZExtValue())) {
12813 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12817 // FIXME gcc accepts some relocatable values here too, but only in certain
12818 // memory models; it's complicated.
12822 // Literal immediates are always ok.
12823 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
12824 // Widen to 64 bits here to get it sign extended.
12825 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
12829 // In any sort of PIC mode addresses need to be computed at runtime by
12830 // adding in a register or some sort of table lookup. These can't
12831 // be used as immediates.
12832 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
12835 // If we are in non-pic codegen mode, we allow the address of a global (with
12836 // an optional displacement) to be used with 'i'.
12837 GlobalAddressSDNode *GA = 0;
12838 int64_t Offset = 0;
12840 // Match either (GA), (GA+C), (GA+C1+C2), etc.
12842 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
12843 Offset += GA->getOffset();
12845 } else if (Op.getOpcode() == ISD::ADD) {
12846 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12847 Offset += C->getZExtValue();
12848 Op = Op.getOperand(0);
12851 } else if (Op.getOpcode() == ISD::SUB) {
12852 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12853 Offset += -C->getZExtValue();
12854 Op = Op.getOperand(0);
12859 // Otherwise, this isn't something we can handle, reject it.
12863 const GlobalValue *GV = GA->getGlobal();
12864 // If we require an extra load to get this address, as in PIC mode, we
12865 // can't accept it.
12866 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
12867 getTargetMachine())))
12870 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
12871 GA->getValueType(0), Offset);
12876 if (Result.getNode()) {
12877 Ops.push_back(Result);
12880 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
12883 std::vector<unsigned> X86TargetLowering::
12884 getRegClassForInlineAsmConstraint(const std::string &Constraint,
12886 if (Constraint.size() == 1) {
12887 // FIXME: not handling fp-stack yet!
12888 switch (Constraint[0]) { // GCC X86 Constraint Letters
12889 default: break; // Unknown constraint letter
12890 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
12891 if (Subtarget->is64Bit()) {
12892 if (VT == MVT::i32)
12893 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
12894 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
12895 X86::R10D,X86::R11D,X86::R12D,
12896 X86::R13D,X86::R14D,X86::R15D,
12897 X86::EBP, X86::ESP, 0);
12898 else if (VT == MVT::i16)
12899 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
12900 X86::SI, X86::DI, X86::R8W,X86::R9W,
12901 X86::R10W,X86::R11W,X86::R12W,
12902 X86::R13W,X86::R14W,X86::R15W,
12903 X86::BP, X86::SP, 0);
12904 else if (VT == MVT::i8)
12905 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
12906 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
12907 X86::R10B,X86::R11B,X86::R12B,
12908 X86::R13B,X86::R14B,X86::R15B,
12909 X86::BPL, X86::SPL, 0);
12911 else if (VT == MVT::i64)
12912 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
12913 X86::RSI, X86::RDI, X86::R8, X86::R9,
12914 X86::R10, X86::R11, X86::R12,
12915 X86::R13, X86::R14, X86::R15,
12916 X86::RBP, X86::RSP, 0);
12920 // 32-bit fallthrough
12921 case 'Q': // Q_REGS
12922 if (VT == MVT::i32)
12923 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
12924 else if (VT == MVT::i16)
12925 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
12926 else if (VT == MVT::i8)
12927 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
12928 else if (VT == MVT::i64)
12929 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
12934 return std::vector<unsigned>();
12937 std::pair<unsigned, const TargetRegisterClass*>
12938 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
12940 // First, see if this is a constraint that directly corresponds to an LLVM
12942 if (Constraint.size() == 1) {
12943 // GCC Constraint Letters
12944 switch (Constraint[0]) {
12946 case 'r': // GENERAL_REGS
12947 case 'l': // INDEX_REGS
12949 return std::make_pair(0U, X86::GR8RegisterClass);
12950 if (VT == MVT::i16)
12951 return std::make_pair(0U, X86::GR16RegisterClass);
12952 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
12953 return std::make_pair(0U, X86::GR32RegisterClass);
12954 return std::make_pair(0U, X86::GR64RegisterClass);
12955 case 'R': // LEGACY_REGS
12957 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
12958 if (VT == MVT::i16)
12959 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
12960 if (VT == MVT::i32 || !Subtarget->is64Bit())
12961 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
12962 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
12963 case 'f': // FP Stack registers.
12964 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
12965 // value to the correct fpstack register class.
12966 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
12967 return std::make_pair(0U, X86::RFP32RegisterClass);
12968 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
12969 return std::make_pair(0U, X86::RFP64RegisterClass);
12970 return std::make_pair(0U, X86::RFP80RegisterClass);
12971 case 'y': // MMX_REGS if MMX allowed.
12972 if (!Subtarget->hasMMX()) break;
12973 return std::make_pair(0U, X86::VR64RegisterClass);
12974 case 'Y': // SSE_REGS if SSE2 allowed
12975 if (!Subtarget->hasXMMInt()) break;
12977 case 'x': // SSE_REGS if SSE1 allowed
12978 if (!Subtarget->hasXMM()) break;
12980 switch (VT.getSimpleVT().SimpleTy) {
12982 // Scalar SSE types.
12985 return std::make_pair(0U, X86::FR32RegisterClass);
12988 return std::make_pair(0U, X86::FR64RegisterClass);
12996 return std::make_pair(0U, X86::VR128RegisterClass);
13002 // Use the default implementation in TargetLowering to convert the register
13003 // constraint into a member of a register class.
13004 std::pair<unsigned, const TargetRegisterClass*> Res;
13005 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13007 // Not found as a standard register?
13008 if (Res.second == 0) {
13009 // Map st(0) -> st(7) -> ST0
13010 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13011 tolower(Constraint[1]) == 's' &&
13012 tolower(Constraint[2]) == 't' &&
13013 Constraint[3] == '(' &&
13014 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13015 Constraint[5] == ')' &&
13016 Constraint[6] == '}') {
13018 Res.first = X86::ST0+Constraint[4]-'0';
13019 Res.second = X86::RFP80RegisterClass;
13023 // GCC allows "st(0)" to be called just plain "st".
13024 if (StringRef("{st}").equals_lower(Constraint)) {
13025 Res.first = X86::ST0;
13026 Res.second = X86::RFP80RegisterClass;
13031 if (StringRef("{flags}").equals_lower(Constraint)) {
13032 Res.first = X86::EFLAGS;
13033 Res.second = X86::CCRRegisterClass;
13037 // 'A' means EAX + EDX.
13038 if (Constraint == "A") {
13039 Res.first = X86::EAX;
13040 Res.second = X86::GR32_ADRegisterClass;
13046 // Otherwise, check to see if this is a register class of the wrong value
13047 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13048 // turn into {ax},{dx}.
13049 if (Res.second->hasType(VT))
13050 return Res; // Correct type already, nothing to do.
13052 // All of the single-register GCC register classes map their values onto
13053 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13054 // really want an 8-bit or 32-bit register, map to the appropriate register
13055 // class and return the appropriate register.
13056 if (Res.second == X86::GR16RegisterClass) {
13057 if (VT == MVT::i8) {
13058 unsigned DestReg = 0;
13059 switch (Res.first) {
13061 case X86::AX: DestReg = X86::AL; break;
13062 case X86::DX: DestReg = X86::DL; break;
13063 case X86::CX: DestReg = X86::CL; break;
13064 case X86::BX: DestReg = X86::BL; break;
13067 Res.first = DestReg;
13068 Res.second = X86::GR8RegisterClass;
13070 } else if (VT == MVT::i32) {
13071 unsigned DestReg = 0;
13072 switch (Res.first) {
13074 case X86::AX: DestReg = X86::EAX; break;
13075 case X86::DX: DestReg = X86::EDX; break;
13076 case X86::CX: DestReg = X86::ECX; break;
13077 case X86::BX: DestReg = X86::EBX; break;
13078 case X86::SI: DestReg = X86::ESI; break;
13079 case X86::DI: DestReg = X86::EDI; break;
13080 case X86::BP: DestReg = X86::EBP; break;
13081 case X86::SP: DestReg = X86::ESP; break;
13084 Res.first = DestReg;
13085 Res.second = X86::GR32RegisterClass;
13087 } else if (VT == MVT::i64) {
13088 unsigned DestReg = 0;
13089 switch (Res.first) {
13091 case X86::AX: DestReg = X86::RAX; break;
13092 case X86::DX: DestReg = X86::RDX; break;
13093 case X86::CX: DestReg = X86::RCX; break;
13094 case X86::BX: DestReg = X86::RBX; break;
13095 case X86::SI: DestReg = X86::RSI; break;
13096 case X86::DI: DestReg = X86::RDI; break;
13097 case X86::BP: DestReg = X86::RBP; break;
13098 case X86::SP: DestReg = X86::RSP; break;
13101 Res.first = DestReg;
13102 Res.second = X86::GR64RegisterClass;
13105 } else if (Res.second == X86::FR32RegisterClass ||
13106 Res.second == X86::FR64RegisterClass ||
13107 Res.second == X86::VR128RegisterClass) {
13108 // Handle references to XMM physical registers that got mapped into the
13109 // wrong class. This can happen with constraints like {xmm0} where the
13110 // target independent register mapper will just pick the first match it can
13111 // find, ignoring the required type.
13112 if (VT == MVT::f32)
13113 Res.second = X86::FR32RegisterClass;
13114 else if (VT == MVT::f64)
13115 Res.second = X86::FR64RegisterClass;
13116 else if (X86::VR128RegisterClass->hasType(VT))
13117 Res.second = X86::VR128RegisterClass;