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 "X86ShuffleDecode.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.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/CommandLine.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,
68 static SDValue Extract128BitVector(SDValue Vec,
73 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
75 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
76 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
77 /// simple subregister reference.
78 static SDValue Extract128BitVector(SDValue Vec,
82 EVT VT = Vec.getValueType();
83 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
85 EVT ElVT = VT.getVectorElementType();
87 int Factor = VT.getSizeInBits() / 128;
89 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(),
91 VT.getVectorNumElements() / Factor);
93 // Extract from UNDEF is UNDEF.
94 if (Vec.getOpcode() == ISD::UNDEF)
95 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
97 if (isa<ConstantSDNode>(Idx)) {
98 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
100 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
101 // we can match to VEXTRACTF128.
102 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
104 // This is the index of the first element of the 128-bit chunk
106 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
109 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
111 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
120 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
121 /// sets things up to match to an AVX VINSERTF128 instruction or a
122 /// simple superregister reference.
123 static SDValue Insert128BitVector(SDValue Result,
128 if (isa<ConstantSDNode>(Idx)) {
129 EVT VT = Vec.getValueType();
130 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
132 EVT ElVT = VT.getVectorElementType();
134 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
136 EVT ResultVT = Result.getValueType();
138 // Insert the relevant 128 bits.
139 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
141 // This is the index of the first element of the 128-bit chunk
143 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
146 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
148 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
156 /// Given two vectors, concat them.
157 static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
158 DebugLoc dl = Lower.getDebugLoc();
160 assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
162 EVT VT = EVT::getVectorVT(*DAG.getContext(),
163 Lower.getValueType().getVectorElementType(),
164 Lower.getValueType().getVectorNumElements() * 2);
166 // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
167 assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
169 // Insert the upper subvector.
170 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
172 // This is half the length of the result
173 // vector. Start inserting the upper 128
175 Lower.getValueType().
176 getVectorNumElements(),
180 // Insert the lower subvector.
181 Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
185 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
186 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
187 bool is64Bit = Subtarget->is64Bit();
189 if (Subtarget->isTargetEnvMacho()) {
191 return new X8664_MachoTargetObjectFile();
192 return new TargetLoweringObjectFileMachO();
195 if (Subtarget->isTargetELF()) {
197 return new X8664_ELFTargetObjectFile(TM);
198 return new X8632_ELFTargetObjectFile(TM);
200 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
201 return new TargetLoweringObjectFileCOFF();
202 llvm_unreachable("unknown subtarget type");
205 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
206 : TargetLowering(TM, createTLOF(TM)) {
207 Subtarget = &TM.getSubtarget<X86Subtarget>();
208 X86ScalarSSEf64 = Subtarget->hasXMMInt();
209 X86ScalarSSEf32 = Subtarget->hasXMM();
210 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
212 RegInfo = TM.getRegisterInfo();
213 TD = getTargetData();
215 // Set up the TargetLowering object.
216 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
218 // X86 is weird, it always uses i8 for shift amounts and setcc results.
219 setShiftAmountType(MVT::i8);
220 setBooleanContents(ZeroOrOneBooleanContent);
221 setSchedulingPreference(Sched::RegPressure);
222 setStackPointerRegisterToSaveRestore(X86StackPtr);
224 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
225 // Setup Windows compiler runtime calls.
226 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
227 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
228 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
229 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
230 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
231 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
232 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
233 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
236 if (Subtarget->isTargetDarwin()) {
237 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
238 setUseUnderscoreSetJmp(false);
239 setUseUnderscoreLongJmp(false);
240 } else if (Subtarget->isTargetMingw()) {
241 // MS runtime is weird: it exports _setjmp, but longjmp!
242 setUseUnderscoreSetJmp(true);
243 setUseUnderscoreLongJmp(false);
245 setUseUnderscoreSetJmp(true);
246 setUseUnderscoreLongJmp(true);
249 // Set up the register classes.
250 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
251 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
252 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
253 if (Subtarget->is64Bit())
254 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
256 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
258 // We don't accept any truncstore of integer registers.
259 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
260 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
261 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
262 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
263 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
264 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
266 // SETOEQ and SETUNE require checking two conditions.
267 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
268 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
269 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
270 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
271 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
272 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
274 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
276 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
277 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
278 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
280 if (Subtarget->is64Bit()) {
281 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
282 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
283 } else if (!UseSoftFloat) {
284 // We have an algorithm for SSE2->double, and we turn this into a
285 // 64-bit FILD followed by conditional FADD for other targets.
286 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
287 // We have an algorithm for SSE2, and we turn this into a 64-bit
288 // FILD for other targets.
289 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
292 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
294 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
295 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
298 // SSE has no i16 to fp conversion, only i32
299 if (X86ScalarSSEf32) {
300 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
301 // f32 and f64 cases are Legal, f80 case is not
302 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
304 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
305 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
308 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
309 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
312 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
313 // are Legal, f80 is custom lowered.
314 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
315 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
317 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
319 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
320 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
322 if (X86ScalarSSEf32) {
323 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
324 // f32 and f64 cases are Legal, f80 case is not
325 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
327 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
328 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
331 // Handle FP_TO_UINT by promoting the destination to a larger signed
333 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
334 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
335 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
337 if (Subtarget->is64Bit()) {
338 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
339 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
340 } else if (!UseSoftFloat) {
341 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
342 // Expand FP_TO_UINT into a select.
343 // FIXME: We would like to use a Custom expander here eventually to do
344 // the optimal thing for SSE vs. the default expansion in the legalizer.
345 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
347 // With SSE3 we can use fisttpll to convert to a signed i64; without
348 // SSE, we're stuck with a fistpll.
349 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
352 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
353 if (!X86ScalarSSEf64) {
354 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
355 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
356 if (Subtarget->is64Bit()) {
357 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
358 // Without SSE, i64->f64 goes through memory.
359 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
363 // Scalar integer divide and remainder are lowered to use operations that
364 // produce two results, to match the available instructions. This exposes
365 // the two-result form to trivial CSE, which is able to combine x/y and x%y
366 // into a single instruction.
368 // Scalar integer multiply-high is also lowered to use two-result
369 // operations, to match the available instructions. However, plain multiply
370 // (low) operations are left as Legal, as there are single-result
371 // instructions for this in x86. Using the two-result multiply instructions
372 // when both high and low results are needed must be arranged by dagcombine.
373 for (unsigned i = 0, e = 4; i != e; ++i) {
375 setOperationAction(ISD::MULHS, VT, Expand);
376 setOperationAction(ISD::MULHU, VT, Expand);
377 setOperationAction(ISD::SDIV, VT, Expand);
378 setOperationAction(ISD::UDIV, VT, Expand);
379 setOperationAction(ISD::SREM, VT, Expand);
380 setOperationAction(ISD::UREM, VT, Expand);
382 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
383 setOperationAction(ISD::ADDC, VT, Custom);
384 setOperationAction(ISD::ADDE, VT, Custom);
385 setOperationAction(ISD::SUBC, VT, Custom);
386 setOperationAction(ISD::SUBE, VT, Custom);
389 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
390 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
391 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
392 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
395 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
396 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
397 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
398 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
399 setOperationAction(ISD::FREM , MVT::f32 , Expand);
400 setOperationAction(ISD::FREM , MVT::f64 , Expand);
401 setOperationAction(ISD::FREM , MVT::f80 , Expand);
402 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
404 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
405 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
406 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
407 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
408 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
409 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
410 if (Subtarget->is64Bit()) {
411 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
412 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
415 if (Subtarget->hasPOPCNT()) {
416 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
418 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
419 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
420 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
421 if (Subtarget->is64Bit())
422 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
425 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
426 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
428 // These should be promoted to a larger select which is supported.
429 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
430 // X86 wants to expand cmov itself.
431 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
432 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
433 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
434 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
435 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
436 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
437 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
438 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
439 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
440 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
441 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
442 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
443 if (Subtarget->is64Bit()) {
444 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
445 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
447 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
450 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
451 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
452 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
453 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
454 if (Subtarget->is64Bit())
455 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
456 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
457 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
458 if (Subtarget->is64Bit()) {
459 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
460 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
461 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
462 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
463 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
465 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
466 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
467 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
468 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
469 if (Subtarget->is64Bit()) {
470 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
471 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
472 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
475 if (Subtarget->hasXMM())
476 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
478 // We may not have a libcall for MEMBARRIER so we should lower this.
479 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
481 // On X86 and X86-64, atomic operations are lowered to locked instructions.
482 // Locked instructions, in turn, have implicit fence semantics (all memory
483 // operations are flushed before issuing the locked instruction, and they
484 // are not buffered), so we can fold away the common pattern of
485 // fence-atomic-fence.
486 setShouldFoldAtomicFences(true);
488 // Expand certain atomics
489 for (unsigned i = 0, e = 4; i != e; ++i) {
491 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
492 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
495 if (!Subtarget->is64Bit()) {
496 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
497 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
498 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
499 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
500 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
501 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
502 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
505 // FIXME - use subtarget debug flags
506 if (!Subtarget->isTargetDarwin() &&
507 !Subtarget->isTargetELF() &&
508 !Subtarget->isTargetCygMing()) {
509 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
512 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
513 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
514 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
515 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
516 if (Subtarget->is64Bit()) {
517 setExceptionPointerRegister(X86::RAX);
518 setExceptionSelectorRegister(X86::RDX);
520 setExceptionPointerRegister(X86::EAX);
521 setExceptionSelectorRegister(X86::EDX);
523 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
524 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
526 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
528 setOperationAction(ISD::TRAP, MVT::Other, Legal);
530 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
531 setOperationAction(ISD::VASTART , MVT::Other, Custom);
532 setOperationAction(ISD::VAEND , MVT::Other, Expand);
533 if (Subtarget->is64Bit()) {
534 setOperationAction(ISD::VAARG , MVT::Other, Custom);
535 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
537 setOperationAction(ISD::VAARG , MVT::Other, Expand);
538 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
541 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
542 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
543 if (Subtarget->is64Bit())
544 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
545 if (Subtarget->isTargetCygMing() || Subtarget->isTargetWindows())
546 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
548 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
550 if (!UseSoftFloat && X86ScalarSSEf64) {
551 // f32 and f64 use SSE.
552 // Set up the FP register classes.
553 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
554 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
556 // Use ANDPD to simulate FABS.
557 setOperationAction(ISD::FABS , MVT::f64, Custom);
558 setOperationAction(ISD::FABS , MVT::f32, Custom);
560 // Use XORP to simulate FNEG.
561 setOperationAction(ISD::FNEG , MVT::f64, Custom);
562 setOperationAction(ISD::FNEG , MVT::f32, Custom);
564 // Use ANDPD and ORPD to simulate FCOPYSIGN.
565 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
566 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
568 // We don't support sin/cos/fmod
569 setOperationAction(ISD::FSIN , MVT::f64, Expand);
570 setOperationAction(ISD::FCOS , MVT::f64, Expand);
571 setOperationAction(ISD::FSIN , MVT::f32, Expand);
572 setOperationAction(ISD::FCOS , MVT::f32, Expand);
574 // Expand FP immediates into loads from the stack, except for the special
576 addLegalFPImmediate(APFloat(+0.0)); // xorpd
577 addLegalFPImmediate(APFloat(+0.0f)); // xorps
578 } else if (!UseSoftFloat && X86ScalarSSEf32) {
579 // Use SSE for f32, x87 for f64.
580 // Set up the FP register classes.
581 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
582 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
584 // Use ANDPS to simulate FABS.
585 setOperationAction(ISD::FABS , MVT::f32, Custom);
587 // Use XORP to simulate FNEG.
588 setOperationAction(ISD::FNEG , MVT::f32, Custom);
590 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
592 // Use ANDPS and ORPS to simulate FCOPYSIGN.
593 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
594 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
596 // We don't support sin/cos/fmod
597 setOperationAction(ISD::FSIN , MVT::f32, Expand);
598 setOperationAction(ISD::FCOS , MVT::f32, Expand);
600 // Special cases we handle for FP constants.
601 addLegalFPImmediate(APFloat(+0.0f)); // xorps
602 addLegalFPImmediate(APFloat(+0.0)); // FLD0
603 addLegalFPImmediate(APFloat(+1.0)); // FLD1
604 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
605 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
608 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
609 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
611 } else if (!UseSoftFloat) {
612 // f32 and f64 in x87.
613 // Set up the FP register classes.
614 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
615 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
617 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
618 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
619 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
620 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
623 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
624 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
626 addLegalFPImmediate(APFloat(+0.0)); // FLD0
627 addLegalFPImmediate(APFloat(+1.0)); // FLD1
628 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
629 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
630 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
631 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
632 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
633 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
636 // Long double always uses X87.
638 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
639 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
640 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
642 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
643 addLegalFPImmediate(TmpFlt); // FLD0
645 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
648 APFloat TmpFlt2(+1.0);
649 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
651 addLegalFPImmediate(TmpFlt2); // FLD1
652 TmpFlt2.changeSign();
653 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
657 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
658 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
662 // Always use a library call for pow.
663 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
664 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
665 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
667 setOperationAction(ISD::FLOG, MVT::f80, Expand);
668 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
669 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
670 setOperationAction(ISD::FEXP, MVT::f80, Expand);
671 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
673 // First set operation action for all vector types to either promote
674 // (for widening) or expand (for scalarization). Then we will selectively
675 // turn on ones that can be effectively codegen'd.
676 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
677 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
678 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
679 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
680 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
681 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
682 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
683 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
684 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
685 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
686 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
687 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
689 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
691 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
693 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
695 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
696 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
719 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
723 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
724 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
725 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
726 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
727 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
728 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
729 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
730 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
731 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
732 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
733 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
734 setTruncStoreAction((MVT::SimpleValueType)VT,
735 (MVT::SimpleValueType)InnerVT, Expand);
736 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
737 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
738 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
741 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
742 // with -msoft-float, disable use of MMX as well.
743 if (!UseSoftFloat && Subtarget->hasMMX()) {
744 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
745 // No operations on x86mmx supported, everything uses intrinsics.
748 // MMX-sized vectors (other than x86mmx) are expected to be expanded
749 // into smaller operations.
750 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
751 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
752 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
753 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
754 setOperationAction(ISD::AND, MVT::v8i8, Expand);
755 setOperationAction(ISD::AND, MVT::v4i16, Expand);
756 setOperationAction(ISD::AND, MVT::v2i32, Expand);
757 setOperationAction(ISD::AND, MVT::v1i64, Expand);
758 setOperationAction(ISD::OR, MVT::v8i8, Expand);
759 setOperationAction(ISD::OR, MVT::v4i16, Expand);
760 setOperationAction(ISD::OR, MVT::v2i32, Expand);
761 setOperationAction(ISD::OR, MVT::v1i64, Expand);
762 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
763 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
764 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
765 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
766 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
767 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
768 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
769 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
770 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
771 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
772 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
773 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
774 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
775 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
776 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
777 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
778 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
780 if (!UseSoftFloat && Subtarget->hasXMM()) {
781 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
783 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
784 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
785 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
786 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
787 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
788 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
789 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
790 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
791 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
793 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
794 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
797 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
798 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
800 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
801 // registers cannot be used even for integer operations.
802 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
803 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
804 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
805 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
807 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
808 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
809 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
810 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
811 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
812 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
813 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
814 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
815 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
816 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
817 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
818 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
819 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
820 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
821 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
822 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
824 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
825 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
826 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
827 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
829 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
830 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
831 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
832 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
833 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
835 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
836 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
837 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
838 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
839 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
841 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
842 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
843 EVT VT = (MVT::SimpleValueType)i;
844 // Do not attempt to custom lower non-power-of-2 vectors
845 if (!isPowerOf2_32(VT.getVectorNumElements()))
847 // Do not attempt to custom lower non-128-bit vectors
848 if (!VT.is128BitVector())
850 setOperationAction(ISD::BUILD_VECTOR,
851 VT.getSimpleVT().SimpleTy, Custom);
852 setOperationAction(ISD::VECTOR_SHUFFLE,
853 VT.getSimpleVT().SimpleTy, Custom);
854 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
855 VT.getSimpleVT().SimpleTy, Custom);
858 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
859 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
860 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
861 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
865 if (Subtarget->is64Bit()) {
866 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
867 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
870 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
871 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
872 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
875 // Do not attempt to promote non-128-bit vectors
876 if (!VT.is128BitVector())
879 setOperationAction(ISD::AND, SVT, Promote);
880 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
881 setOperationAction(ISD::OR, SVT, Promote);
882 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
883 setOperationAction(ISD::XOR, SVT, Promote);
884 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
885 setOperationAction(ISD::LOAD, SVT, Promote);
886 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
887 setOperationAction(ISD::SELECT, SVT, Promote);
888 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
891 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
893 // Custom lower v2i64 and v2f64 selects.
894 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
895 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
896 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
897 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
899 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
900 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
903 if (Subtarget->hasSSE41()) {
904 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
905 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
906 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
907 setOperationAction(ISD::FRINT, MVT::f32, Legal);
908 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
909 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
910 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
911 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
912 setOperationAction(ISD::FRINT, MVT::f64, Legal);
913 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
915 // FIXME: Do we need to handle scalar-to-vector here?
916 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
918 // Can turn SHL into an integer multiply.
919 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
920 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
922 // i8 and i16 vectors are custom , because the source register and source
923 // source memory operand types are not the same width. f32 vectors are
924 // custom since the immediate controlling the insert encodes additional
926 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
927 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
928 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
929 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
931 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
932 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
933 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
934 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
936 if (Subtarget->is64Bit()) {
937 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
938 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
942 if (Subtarget->hasSSE42())
943 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
945 if (!UseSoftFloat && Subtarget->hasAVX()) {
946 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
947 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
948 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
949 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
950 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
952 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
953 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
954 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
955 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
957 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
958 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
959 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
960 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
961 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
962 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
964 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
965 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
966 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
967 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
968 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
969 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
971 // Custom lower build_vector, vector_shuffle, scalar_to_vector,
972 // insert_vector_elt extract_subvector and extract_vector_elt for
974 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
975 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
977 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
978 // Do not attempt to custom lower non-256-bit vectors
979 if (!isPowerOf2_32(MVT(VT).getVectorNumElements())
980 || (MVT(VT).getSizeInBits() < 256))
982 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
983 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
984 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
985 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
986 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
988 // Custom-lower insert_subvector and extract_subvector based on
990 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
991 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
993 MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
994 // Do not attempt to custom lower non-256-bit vectors
995 if (!isPowerOf2_32(MVT(VT).getVectorNumElements()))
998 if (MVT(VT).getSizeInBits() == 128) {
999 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1001 else if (MVT(VT).getSizeInBits() == 256) {
1002 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1006 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1007 // Don't promote loads because we need them for VPERM vector index versions.
1009 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1010 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE;
1012 if (!isPowerOf2_32(MVT((MVT::SimpleValueType)VT).getVectorNumElements())
1013 || (MVT((MVT::SimpleValueType)VT).getSizeInBits() < 256))
1015 setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
1016 AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v4i64);
1017 setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
1018 AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v4i64);
1019 setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
1020 AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v4i64);
1021 //setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
1022 //AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v4i64);
1023 setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
1024 AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v4i64);
1028 // We want to custom lower some of our intrinsics.
1029 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1032 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1033 // handle type legalization for these operations here.
1035 // FIXME: We really should do custom legalization for addition and
1036 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1037 // than generic legalization for 64-bit multiplication-with-overflow, though.
1038 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1039 // Add/Sub/Mul with overflow operations are custom lowered.
1041 setOperationAction(ISD::SADDO, VT, Custom);
1042 setOperationAction(ISD::UADDO, VT, Custom);
1043 setOperationAction(ISD::SSUBO, VT, Custom);
1044 setOperationAction(ISD::USUBO, VT, Custom);
1045 setOperationAction(ISD::SMULO, VT, Custom);
1046 setOperationAction(ISD::UMULO, VT, Custom);
1049 // There are no 8-bit 3-address imul/mul instructions
1050 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1051 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1053 if (!Subtarget->is64Bit()) {
1054 // These libcalls are not available in 32-bit.
1055 setLibcallName(RTLIB::SHL_I128, 0);
1056 setLibcallName(RTLIB::SRL_I128, 0);
1057 setLibcallName(RTLIB::SRA_I128, 0);
1060 // We have target-specific dag combine patterns for the following nodes:
1061 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1062 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1063 setTargetDAGCombine(ISD::BUILD_VECTOR);
1064 setTargetDAGCombine(ISD::SELECT);
1065 setTargetDAGCombine(ISD::SHL);
1066 setTargetDAGCombine(ISD::SRA);
1067 setTargetDAGCombine(ISD::SRL);
1068 setTargetDAGCombine(ISD::OR);
1069 setTargetDAGCombine(ISD::AND);
1070 setTargetDAGCombine(ISD::ADD);
1071 setTargetDAGCombine(ISD::SUB);
1072 setTargetDAGCombine(ISD::STORE);
1073 setTargetDAGCombine(ISD::ZERO_EXTEND);
1074 if (Subtarget->is64Bit())
1075 setTargetDAGCombine(ISD::MUL);
1077 computeRegisterProperties();
1079 // On Darwin, -Os means optimize for size without hurting performance,
1080 // do not reduce the limit.
1081 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1082 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1083 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1084 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1085 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1086 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1087 setPrefLoopAlignment(16);
1088 benefitFromCodePlacementOpt = true;
1092 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1097 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1098 /// the desired ByVal argument alignment.
1099 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1102 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1103 if (VTy->getBitWidth() == 128)
1105 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1106 unsigned EltAlign = 0;
1107 getMaxByValAlign(ATy->getElementType(), EltAlign);
1108 if (EltAlign > MaxAlign)
1109 MaxAlign = EltAlign;
1110 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1111 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1112 unsigned EltAlign = 0;
1113 getMaxByValAlign(STy->getElementType(i), EltAlign);
1114 if (EltAlign > MaxAlign)
1115 MaxAlign = EltAlign;
1123 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1124 /// function arguments in the caller parameter area. For X86, aggregates
1125 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1126 /// are at 4-byte boundaries.
1127 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1128 if (Subtarget->is64Bit()) {
1129 // Max of 8 and alignment of type.
1130 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1137 if (Subtarget->hasXMM())
1138 getMaxByValAlign(Ty, Align);
1142 /// getOptimalMemOpType - Returns the target specific optimal type for load
1143 /// and store operations as a result of memset, memcpy, and memmove
1144 /// lowering. If DstAlign is zero that means it's safe to destination
1145 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1146 /// means there isn't a need to check it against alignment requirement,
1147 /// probably because the source does not need to be loaded. If
1148 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1149 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1150 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1151 /// constant so it does not need to be loaded.
1152 /// It returns EVT::Other if the type should be determined using generic
1153 /// target-independent logic.
1155 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1156 unsigned DstAlign, unsigned SrcAlign,
1157 bool NonScalarIntSafe,
1159 MachineFunction &MF) const {
1160 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1161 // linux. This is because the stack realignment code can't handle certain
1162 // cases like PR2962. This should be removed when PR2962 is fixed.
1163 const Function *F = MF.getFunction();
1164 if (NonScalarIntSafe &&
1165 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1167 (Subtarget->isUnalignedMemAccessFast() ||
1168 ((DstAlign == 0 || DstAlign >= 16) &&
1169 (SrcAlign == 0 || SrcAlign >= 16))) &&
1170 Subtarget->getStackAlignment() >= 16) {
1171 if (Subtarget->hasSSE2())
1173 if (Subtarget->hasSSE1())
1175 } else if (!MemcpyStrSrc && Size >= 8 &&
1176 !Subtarget->is64Bit() &&
1177 Subtarget->getStackAlignment() >= 8 &&
1178 Subtarget->hasXMMInt()) {
1179 // Do not use f64 to lower memcpy if source is string constant. It's
1180 // better to use i32 to avoid the loads.
1184 if (Subtarget->is64Bit() && Size >= 8)
1189 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1190 /// current function. The returned value is a member of the
1191 /// MachineJumpTableInfo::JTEntryKind enum.
1192 unsigned X86TargetLowering::getJumpTableEncoding() const {
1193 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1195 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1196 Subtarget->isPICStyleGOT())
1197 return MachineJumpTableInfo::EK_Custom32;
1199 // Otherwise, use the normal jump table encoding heuristics.
1200 return TargetLowering::getJumpTableEncoding();
1204 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1205 const MachineBasicBlock *MBB,
1206 unsigned uid,MCContext &Ctx) const{
1207 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1208 Subtarget->isPICStyleGOT());
1209 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1211 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1212 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1215 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1217 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1218 SelectionDAG &DAG) const {
1219 if (!Subtarget->is64Bit())
1220 // This doesn't have DebugLoc associated with it, but is not really the
1221 // same as a Register.
1222 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1226 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1227 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1229 const MCExpr *X86TargetLowering::
1230 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1231 MCContext &Ctx) const {
1232 // X86-64 uses RIP relative addressing based on the jump table label.
1233 if (Subtarget->isPICStyleRIPRel())
1234 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1236 // Otherwise, the reference is relative to the PIC base.
1237 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1240 /// getFunctionAlignment - Return the Log2 alignment of this function.
1241 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1242 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1245 // FIXME: Why this routine is here? Move to RegInfo!
1246 std::pair<const TargetRegisterClass*, uint8_t>
1247 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1248 const TargetRegisterClass *RRC = 0;
1250 switch (VT.getSimpleVT().SimpleTy) {
1252 return TargetLowering::findRepresentativeClass(VT);
1253 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1254 RRC = (Subtarget->is64Bit()
1255 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1258 RRC = X86::VR64RegisterClass;
1260 case MVT::f32: case MVT::f64:
1261 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1262 case MVT::v4f32: case MVT::v2f64:
1263 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1265 RRC = X86::VR128RegisterClass;
1268 return std::make_pair(RRC, Cost);
1271 // FIXME: Why this routine is here? Move to RegInfo!
1273 X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
1274 MachineFunction &MF) const {
1275 const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
1277 unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
1278 switch (RC->getID()) {
1281 case X86::GR32RegClassID:
1283 case X86::GR64RegClassID:
1285 case X86::VR128RegClassID:
1286 return Subtarget->is64Bit() ? 10 : 4;
1287 case X86::VR64RegClassID:
1292 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1293 unsigned &Offset) const {
1294 if (!Subtarget->isTargetLinux())
1297 if (Subtarget->is64Bit()) {
1298 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1300 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1313 //===----------------------------------------------------------------------===//
1314 // Return Value Calling Convention Implementation
1315 //===----------------------------------------------------------------------===//
1317 #include "X86GenCallingConv.inc"
1320 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1321 const SmallVectorImpl<ISD::OutputArg> &Outs,
1322 LLVMContext &Context) const {
1323 SmallVector<CCValAssign, 16> RVLocs;
1324 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1326 return CCInfo.CheckReturn(Outs, RetCC_X86);
1330 X86TargetLowering::LowerReturn(SDValue Chain,
1331 CallingConv::ID CallConv, bool isVarArg,
1332 const SmallVectorImpl<ISD::OutputArg> &Outs,
1333 const SmallVectorImpl<SDValue> &OutVals,
1334 DebugLoc dl, SelectionDAG &DAG) const {
1335 MachineFunction &MF = DAG.getMachineFunction();
1336 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1338 SmallVector<CCValAssign, 16> RVLocs;
1339 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1340 RVLocs, *DAG.getContext());
1341 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1343 // Add the regs to the liveout set for the function.
1344 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1345 for (unsigned i = 0; i != RVLocs.size(); ++i)
1346 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1347 MRI.addLiveOut(RVLocs[i].getLocReg());
1351 SmallVector<SDValue, 6> RetOps;
1352 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1353 // Operand #1 = Bytes To Pop
1354 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1357 // Copy the result values into the output registers.
1358 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1359 CCValAssign &VA = RVLocs[i];
1360 assert(VA.isRegLoc() && "Can only return in registers!");
1361 SDValue ValToCopy = OutVals[i];
1362 EVT ValVT = ValToCopy.getValueType();
1364 // If this is x86-64, and we disabled SSE, we can't return FP values,
1365 // or SSE or MMX vectors.
1366 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1367 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1368 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1369 report_fatal_error("SSE register return with SSE disabled");
1371 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1372 // llvm-gcc has never done it right and no one has noticed, so this
1373 // should be OK for now.
1374 if (ValVT == MVT::f64 &&
1375 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1376 report_fatal_error("SSE2 register return with SSE2 disabled");
1378 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1379 // the RET instruction and handled by the FP Stackifier.
1380 if (VA.getLocReg() == X86::ST0 ||
1381 VA.getLocReg() == X86::ST1) {
1382 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1383 // change the value to the FP stack register class.
1384 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1385 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1386 RetOps.push_back(ValToCopy);
1387 // Don't emit a copytoreg.
1391 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1392 // which is returned in RAX / RDX.
1393 if (Subtarget->is64Bit()) {
1394 if (ValVT == MVT::x86mmx) {
1395 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1396 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1397 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1399 // If we don't have SSE2 available, convert to v4f32 so the generated
1400 // register is legal.
1401 if (!Subtarget->hasSSE2())
1402 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1407 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1408 Flag = Chain.getValue(1);
1411 // The x86-64 ABI for returning structs by value requires that we copy
1412 // the sret argument into %rax for the return. We saved the argument into
1413 // a virtual register in the entry block, so now we copy the value out
1415 if (Subtarget->is64Bit() &&
1416 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1417 MachineFunction &MF = DAG.getMachineFunction();
1418 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1419 unsigned Reg = FuncInfo->getSRetReturnReg();
1421 "SRetReturnReg should have been set in LowerFormalArguments().");
1422 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1424 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1425 Flag = Chain.getValue(1);
1427 // RAX now acts like a return value.
1428 MRI.addLiveOut(X86::RAX);
1431 RetOps[0] = Chain; // Update chain.
1433 // Add the flag if we have it.
1435 RetOps.push_back(Flag);
1437 return DAG.getNode(X86ISD::RET_FLAG, dl,
1438 MVT::Other, &RetOps[0], RetOps.size());
1441 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1442 if (N->getNumValues() != 1)
1444 if (!N->hasNUsesOfValue(1, 0))
1447 SDNode *Copy = *N->use_begin();
1448 if (Copy->getOpcode() != ISD::CopyToReg &&
1449 Copy->getOpcode() != ISD::FP_EXTEND)
1452 bool HasRet = false;
1453 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1455 if (UI->getOpcode() != X86ISD::RET_FLAG)
1463 /// LowerCallResult - Lower the result values of a call into the
1464 /// appropriate copies out of appropriate physical registers.
1467 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1468 CallingConv::ID CallConv, bool isVarArg,
1469 const SmallVectorImpl<ISD::InputArg> &Ins,
1470 DebugLoc dl, SelectionDAG &DAG,
1471 SmallVectorImpl<SDValue> &InVals) const {
1473 // Assign locations to each value returned by this call.
1474 SmallVector<CCValAssign, 16> RVLocs;
1475 bool Is64Bit = Subtarget->is64Bit();
1476 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1477 RVLocs, *DAG.getContext());
1478 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1480 // Copy all of the result registers out of their specified physreg.
1481 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1482 CCValAssign &VA = RVLocs[i];
1483 EVT CopyVT = VA.getValVT();
1485 // If this is x86-64, and we disabled SSE, we can't return FP values
1486 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1487 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1488 report_fatal_error("SSE register return with SSE disabled");
1493 // If this is a call to a function that returns an fp value on the floating
1494 // point stack, we must guarantee the the value is popped from the stack, so
1495 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1496 // if the return value is not used. We use the FpGET_ST0 instructions
1498 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1499 // If we prefer to use the value in xmm registers, copy it out as f80 and
1500 // use a truncate to move it from fp stack reg to xmm reg.
1501 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1502 bool isST0 = VA.getLocReg() == X86::ST0;
1504 if (CopyVT == MVT::f32) Opc = isST0 ? X86::FpGET_ST0_32:X86::FpGET_ST1_32;
1505 if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
1506 if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
1507 SDValue Ops[] = { Chain, InFlag };
1508 Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
1510 Val = Chain.getValue(0);
1512 // Round the f80 to the right size, which also moves it to the appropriate
1514 if (CopyVT != VA.getValVT())
1515 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1516 // This truncation won't change the value.
1517 DAG.getIntPtrConstant(1));
1518 } else if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1519 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1520 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1521 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1522 MVT::v2i64, InFlag).getValue(1);
1523 Val = Chain.getValue(0);
1524 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1525 Val, DAG.getConstant(0, MVT::i64));
1527 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1528 MVT::i64, InFlag).getValue(1);
1529 Val = Chain.getValue(0);
1531 Val = DAG.getNode(ISD::BITCAST, dl, CopyVT, Val);
1533 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1534 CopyVT, InFlag).getValue(1);
1535 Val = Chain.getValue(0);
1537 InFlag = Chain.getValue(2);
1538 InVals.push_back(Val);
1545 //===----------------------------------------------------------------------===//
1546 // C & StdCall & Fast Calling Convention implementation
1547 //===----------------------------------------------------------------------===//
1548 // StdCall calling convention seems to be standard for many Windows' API
1549 // routines and around. It differs from C calling convention just a little:
1550 // callee should clean up the stack, not caller. Symbols should be also
1551 // decorated in some fancy way :) It doesn't support any vector arguments.
1552 // For info on fast calling convention see Fast Calling Convention (tail call)
1553 // implementation LowerX86_32FastCCCallTo.
1555 /// CallIsStructReturn - Determines whether a call uses struct return
1557 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1561 return Outs[0].Flags.isSRet();
1564 /// ArgsAreStructReturn - Determines whether a function uses struct
1565 /// return semantics.
1567 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1571 return Ins[0].Flags.isSRet();
1574 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1575 /// by "Src" to address "Dst" with size and alignment information specified by
1576 /// the specific parameter attribute. The copy will be passed as a byval
1577 /// function parameter.
1579 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1580 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1582 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1584 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1585 /*isVolatile*/false, /*AlwaysInline=*/true,
1586 MachinePointerInfo(), MachinePointerInfo());
1589 /// IsTailCallConvention - Return true if the calling convention is one that
1590 /// supports tail call optimization.
1591 static bool IsTailCallConvention(CallingConv::ID CC) {
1592 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1595 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1596 /// a tailcall target by changing its ABI.
1597 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1598 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1602 X86TargetLowering::LowerMemArgument(SDValue Chain,
1603 CallingConv::ID CallConv,
1604 const SmallVectorImpl<ISD::InputArg> &Ins,
1605 DebugLoc dl, SelectionDAG &DAG,
1606 const CCValAssign &VA,
1607 MachineFrameInfo *MFI,
1609 // Create the nodes corresponding to a load from this parameter slot.
1610 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1611 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1612 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1615 // If value is passed by pointer we have address passed instead of the value
1617 if (VA.getLocInfo() == CCValAssign::Indirect)
1618 ValVT = VA.getLocVT();
1620 ValVT = VA.getValVT();
1622 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1623 // changed with more analysis.
1624 // In case of tail call optimization mark all arguments mutable. Since they
1625 // could be overwritten by lowering of arguments in case of a tail call.
1626 if (Flags.isByVal()) {
1627 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1628 VA.getLocMemOffset(), isImmutable);
1629 return DAG.getFrameIndex(FI, getPointerTy());
1631 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1632 VA.getLocMemOffset(), isImmutable);
1633 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1634 return DAG.getLoad(ValVT, dl, Chain, FIN,
1635 MachinePointerInfo::getFixedStack(FI),
1641 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1642 CallingConv::ID CallConv,
1644 const SmallVectorImpl<ISD::InputArg> &Ins,
1647 SmallVectorImpl<SDValue> &InVals)
1649 MachineFunction &MF = DAG.getMachineFunction();
1650 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1652 const Function* Fn = MF.getFunction();
1653 if (Fn->hasExternalLinkage() &&
1654 Subtarget->isTargetCygMing() &&
1655 Fn->getName() == "main")
1656 FuncInfo->setForceFramePointer(true);
1658 MachineFrameInfo *MFI = MF.getFrameInfo();
1659 bool Is64Bit = Subtarget->is64Bit();
1660 bool IsWin64 = Subtarget->isTargetWin64();
1662 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1663 "Var args not supported with calling convention fastcc or ghc");
1665 // Assign locations to all of the incoming arguments.
1666 SmallVector<CCValAssign, 16> ArgLocs;
1667 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1668 ArgLocs, *DAG.getContext());
1669 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1671 unsigned LastVal = ~0U;
1673 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1674 CCValAssign &VA = ArgLocs[i];
1675 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1677 assert(VA.getValNo() != LastVal &&
1678 "Don't support value assigned to multiple locs yet");
1679 LastVal = VA.getValNo();
1681 if (VA.isRegLoc()) {
1682 EVT RegVT = VA.getLocVT();
1683 TargetRegisterClass *RC = NULL;
1684 if (RegVT == MVT::i32)
1685 RC = X86::GR32RegisterClass;
1686 else if (Is64Bit && RegVT == MVT::i64)
1687 RC = X86::GR64RegisterClass;
1688 else if (RegVT == MVT::f32)
1689 RC = X86::FR32RegisterClass;
1690 else if (RegVT == MVT::f64)
1691 RC = X86::FR64RegisterClass;
1692 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1693 RC = X86::VR256RegisterClass;
1694 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1695 RC = X86::VR128RegisterClass;
1696 else if (RegVT == MVT::x86mmx)
1697 RC = X86::VR64RegisterClass;
1699 llvm_unreachable("Unknown argument type!");
1701 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC, dl);
1702 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1704 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1705 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1707 if (VA.getLocInfo() == CCValAssign::SExt)
1708 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1709 DAG.getValueType(VA.getValVT()));
1710 else if (VA.getLocInfo() == CCValAssign::ZExt)
1711 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1712 DAG.getValueType(VA.getValVT()));
1713 else if (VA.getLocInfo() == CCValAssign::BCvt)
1714 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1716 if (VA.isExtInLoc()) {
1717 // Handle MMX values passed in XMM regs.
1718 if (RegVT.isVector()) {
1719 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1722 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1725 assert(VA.isMemLoc());
1726 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1729 // If value is passed via pointer - do a load.
1730 if (VA.getLocInfo() == CCValAssign::Indirect)
1731 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1732 MachinePointerInfo(), false, false, 0);
1734 InVals.push_back(ArgValue);
1737 // The x86-64 ABI for returning structs by value requires that we copy
1738 // the sret argument into %rax for the return. Save the argument into
1739 // a virtual register so that we can access it from the return points.
1740 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1741 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1742 unsigned Reg = FuncInfo->getSRetReturnReg();
1744 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1745 FuncInfo->setSRetReturnReg(Reg);
1747 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1748 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1751 unsigned StackSize = CCInfo.getNextStackOffset();
1752 // Align stack specially for tail calls.
1753 if (FuncIsMadeTailCallSafe(CallConv))
1754 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1756 // If the function takes variable number of arguments, make a frame index for
1757 // the start of the first vararg value... for expansion of llvm.va_start.
1759 if (!IsWin64 && (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1760 CallConv != CallingConv::X86_ThisCall))) {
1761 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1764 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1766 // FIXME: We should really autogenerate these arrays
1767 static const unsigned GPR64ArgRegsWin64[] = {
1768 X86::RCX, X86::RDX, X86::R8, X86::R9
1770 static const unsigned GPR64ArgRegs64Bit[] = {
1771 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1773 static const unsigned XMMArgRegs64Bit[] = {
1774 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1775 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1777 const unsigned *GPR64ArgRegs;
1778 unsigned NumXMMRegs = 0;
1781 // The XMM registers which might contain var arg parameters are shadowed
1782 // in their paired GPR. So we only need to save the GPR to their home
1784 TotalNumIntRegs = 4;
1785 GPR64ArgRegs = GPR64ArgRegsWin64;
1787 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1788 GPR64ArgRegs = GPR64ArgRegs64Bit;
1790 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1792 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1795 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1796 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1797 "SSE register cannot be used when SSE is disabled!");
1798 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1799 "SSE register cannot be used when SSE is disabled!");
1800 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1801 // Kernel mode asks for SSE to be disabled, so don't push them
1803 TotalNumXMMRegs = 0;
1806 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1807 // Get to the caller-allocated home save location. Add 8 to account
1808 // for the return address.
1809 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1810 FuncInfo->setRegSaveFrameIndex(
1811 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1812 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1814 // For X86-64, if there are vararg parameters that are passed via
1815 // registers, then we must store them to their spots on the stack so they
1816 // may be loaded by deferencing the result of va_next.
1817 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1818 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1819 FuncInfo->setRegSaveFrameIndex(
1820 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1824 // Store the integer parameter registers.
1825 SmallVector<SDValue, 8> MemOps;
1826 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1828 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1829 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1830 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1831 DAG.getIntPtrConstant(Offset));
1832 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1833 X86::GR64RegisterClass, dl);
1834 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1836 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1837 MachinePointerInfo::getFixedStack(
1838 FuncInfo->getRegSaveFrameIndex(), Offset),
1840 MemOps.push_back(Store);
1844 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1845 // Now store the XMM (fp + vector) parameter registers.
1846 SmallVector<SDValue, 11> SaveXMMOps;
1847 SaveXMMOps.push_back(Chain);
1849 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass, dl);
1850 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1851 SaveXMMOps.push_back(ALVal);
1853 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1854 FuncInfo->getRegSaveFrameIndex()));
1855 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1856 FuncInfo->getVarArgsFPOffset()));
1858 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1859 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1860 X86::VR128RegisterClass, dl);
1861 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1862 SaveXMMOps.push_back(Val);
1864 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1866 &SaveXMMOps[0], SaveXMMOps.size()));
1869 if (!MemOps.empty())
1870 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1871 &MemOps[0], MemOps.size());
1875 // Some CCs need callee pop.
1876 if (Subtarget->IsCalleePop(isVarArg, CallConv)) {
1877 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1879 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1880 // If this is an sret function, the return should pop the hidden pointer.
1881 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1882 FuncInfo->setBytesToPopOnReturn(4);
1886 // RegSaveFrameIndex is X86-64 only.
1887 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1888 if (CallConv == CallingConv::X86_FastCall ||
1889 CallConv == CallingConv::X86_ThisCall)
1890 // fastcc functions can't have varargs.
1891 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1898 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1899 SDValue StackPtr, SDValue Arg,
1900 DebugLoc dl, SelectionDAG &DAG,
1901 const CCValAssign &VA,
1902 ISD::ArgFlagsTy Flags) const {
1903 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1904 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1905 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1906 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1907 if (Flags.isByVal())
1908 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1910 return DAG.getStore(Chain, dl, Arg, PtrOff,
1911 MachinePointerInfo::getStack(LocMemOffset),
1915 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1916 /// optimization is performed and it is required.
1918 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1919 SDValue &OutRetAddr, SDValue Chain,
1920 bool IsTailCall, bool Is64Bit,
1921 int FPDiff, DebugLoc dl) const {
1922 // Adjust the Return address stack slot.
1923 EVT VT = getPointerTy();
1924 OutRetAddr = getReturnAddressFrameIndex(DAG);
1926 // Load the "old" Return address.
1927 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1929 return SDValue(OutRetAddr.getNode(), 1);
1932 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1933 /// optimization is performed and it is required (FPDiff!=0).
1935 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1936 SDValue Chain, SDValue RetAddrFrIdx,
1937 bool Is64Bit, int FPDiff, DebugLoc dl) {
1938 // Store the return address to the appropriate stack slot.
1939 if (!FPDiff) return Chain;
1940 // Calculate the new stack slot for the return address.
1941 int SlotSize = Is64Bit ? 8 : 4;
1942 int NewReturnAddrFI =
1943 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1944 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1945 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1946 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1947 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1953 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1954 CallingConv::ID CallConv, bool isVarArg,
1956 const SmallVectorImpl<ISD::OutputArg> &Outs,
1957 const SmallVectorImpl<SDValue> &OutVals,
1958 const SmallVectorImpl<ISD::InputArg> &Ins,
1959 DebugLoc dl, SelectionDAG &DAG,
1960 SmallVectorImpl<SDValue> &InVals) const {
1961 MachineFunction &MF = DAG.getMachineFunction();
1962 bool Is64Bit = Subtarget->is64Bit();
1963 bool IsStructRet = CallIsStructReturn(Outs);
1964 bool IsSibcall = false;
1967 // Check if it's really possible to do a tail call.
1968 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1969 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1970 Outs, OutVals, Ins, DAG);
1972 // Sibcalls are automatically detected tailcalls which do not require
1974 if (!GuaranteedTailCallOpt && isTailCall)
1981 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1982 "Var args not supported with calling convention fastcc or ghc");
1984 // Analyze operands of the call, assigning locations to each operand.
1985 SmallVector<CCValAssign, 16> ArgLocs;
1986 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1987 ArgLocs, *DAG.getContext());
1988 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
1990 // Get a count of how many bytes are to be pushed on the stack.
1991 unsigned NumBytes = CCInfo.getNextStackOffset();
1993 // This is a sibcall. The memory operands are available in caller's
1994 // own caller's stack.
1996 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1997 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2000 if (isTailCall && !IsSibcall) {
2001 // Lower arguments at fp - stackoffset + fpdiff.
2002 unsigned NumBytesCallerPushed =
2003 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2004 FPDiff = NumBytesCallerPushed - NumBytes;
2006 // Set the delta of movement of the returnaddr stackslot.
2007 // But only set if delta is greater than previous delta.
2008 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2009 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2013 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2015 SDValue RetAddrFrIdx;
2016 // Load return adress for tail calls.
2017 if (isTailCall && FPDiff)
2018 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2019 Is64Bit, FPDiff, dl);
2021 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2022 SmallVector<SDValue, 8> MemOpChains;
2025 // Walk the register/memloc assignments, inserting copies/loads. In the case
2026 // of tail call optimization arguments are handle later.
2027 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2028 CCValAssign &VA = ArgLocs[i];
2029 EVT RegVT = VA.getLocVT();
2030 SDValue Arg = OutVals[i];
2031 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2032 bool isByVal = Flags.isByVal();
2034 // Promote the value if needed.
2035 switch (VA.getLocInfo()) {
2036 default: llvm_unreachable("Unknown loc info!");
2037 case CCValAssign::Full: break;
2038 case CCValAssign::SExt:
2039 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2041 case CCValAssign::ZExt:
2042 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2044 case CCValAssign::AExt:
2045 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2046 // Special case: passing MMX values in XMM registers.
2047 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2048 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2049 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2051 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2053 case CCValAssign::BCvt:
2054 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2056 case CCValAssign::Indirect: {
2057 // Store the argument.
2058 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2059 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2060 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2061 MachinePointerInfo::getFixedStack(FI),
2068 if (VA.isRegLoc()) {
2069 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2070 if (isVarArg && Subtarget->isTargetWin64()) {
2071 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2072 // shadow reg if callee is a varargs function.
2073 unsigned ShadowReg = 0;
2074 switch (VA.getLocReg()) {
2075 case X86::XMM0: ShadowReg = X86::RCX; break;
2076 case X86::XMM1: ShadowReg = X86::RDX; break;
2077 case X86::XMM2: ShadowReg = X86::R8; break;
2078 case X86::XMM3: ShadowReg = X86::R9; break;
2081 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2083 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2084 assert(VA.isMemLoc());
2085 if (StackPtr.getNode() == 0)
2086 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2087 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2088 dl, DAG, VA, Flags));
2092 if (!MemOpChains.empty())
2093 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2094 &MemOpChains[0], MemOpChains.size());
2096 // Build a sequence of copy-to-reg nodes chained together with token chain
2097 // and flag operands which copy the outgoing args into registers.
2099 // Tail call byval lowering might overwrite argument registers so in case of
2100 // tail call optimization the copies to registers are lowered later.
2102 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2103 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2104 RegsToPass[i].second, InFlag);
2105 InFlag = Chain.getValue(1);
2108 if (Subtarget->isPICStyleGOT()) {
2109 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2112 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2113 DAG.getNode(X86ISD::GlobalBaseReg,
2114 DebugLoc(), getPointerTy()),
2116 InFlag = Chain.getValue(1);
2118 // If we are tail calling and generating PIC/GOT style code load the
2119 // address of the callee into ECX. The value in ecx is used as target of
2120 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2121 // for tail calls on PIC/GOT architectures. Normally we would just put the
2122 // address of GOT into ebx and then call target@PLT. But for tail calls
2123 // ebx would be restored (since ebx is callee saved) before jumping to the
2126 // Note: The actual moving to ECX is done further down.
2127 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2128 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2129 !G->getGlobal()->hasProtectedVisibility())
2130 Callee = LowerGlobalAddress(Callee, DAG);
2131 else if (isa<ExternalSymbolSDNode>(Callee))
2132 Callee = LowerExternalSymbol(Callee, DAG);
2136 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
2137 // From AMD64 ABI document:
2138 // For calls that may call functions that use varargs or stdargs
2139 // (prototype-less calls or calls to functions containing ellipsis (...) in
2140 // the declaration) %al is used as hidden argument to specify the number
2141 // of SSE registers used. The contents of %al do not need to match exactly
2142 // the number of registers, but must be an ubound on the number of SSE
2143 // registers used and is in the range 0 - 8 inclusive.
2145 // Count the number of XMM registers allocated.
2146 static const unsigned XMMArgRegs[] = {
2147 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2148 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2150 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2151 assert((Subtarget->hasXMM() || !NumXMMRegs)
2152 && "SSE registers cannot be used when SSE is disabled");
2154 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2155 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2156 InFlag = Chain.getValue(1);
2160 // For tail calls lower the arguments to the 'real' stack slot.
2162 // Force all the incoming stack arguments to be loaded from the stack
2163 // before any new outgoing arguments are stored to the stack, because the
2164 // outgoing stack slots may alias the incoming argument stack slots, and
2165 // the alias isn't otherwise explicit. This is slightly more conservative
2166 // than necessary, because it means that each store effectively depends
2167 // on every argument instead of just those arguments it would clobber.
2168 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2170 SmallVector<SDValue, 8> MemOpChains2;
2173 // Do not flag preceeding copytoreg stuff together with the following stuff.
2175 if (GuaranteedTailCallOpt) {
2176 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2177 CCValAssign &VA = ArgLocs[i];
2180 assert(VA.isMemLoc());
2181 SDValue Arg = OutVals[i];
2182 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2183 // Create frame index.
2184 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2185 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2186 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2187 FIN = DAG.getFrameIndex(FI, getPointerTy());
2189 if (Flags.isByVal()) {
2190 // Copy relative to framepointer.
2191 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2192 if (StackPtr.getNode() == 0)
2193 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2195 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2197 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2201 // Store relative to framepointer.
2202 MemOpChains2.push_back(
2203 DAG.getStore(ArgChain, dl, Arg, FIN,
2204 MachinePointerInfo::getFixedStack(FI),
2210 if (!MemOpChains2.empty())
2211 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2212 &MemOpChains2[0], MemOpChains2.size());
2214 // Copy arguments to their registers.
2215 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2216 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2217 RegsToPass[i].second, InFlag);
2218 InFlag = Chain.getValue(1);
2222 // Store the return address to the appropriate stack slot.
2223 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2227 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2228 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2229 // In the 64-bit large code model, we have to make all calls
2230 // through a register, since the call instruction's 32-bit
2231 // pc-relative offset may not be large enough to hold the whole
2233 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2234 // If the callee is a GlobalAddress node (quite common, every direct call
2235 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2238 // We should use extra load for direct calls to dllimported functions in
2240 const GlobalValue *GV = G->getGlobal();
2241 if (!GV->hasDLLImportLinkage()) {
2242 unsigned char OpFlags = 0;
2244 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2245 // external symbols most go through the PLT in PIC mode. If the symbol
2246 // has hidden or protected visibility, or if it is static or local, then
2247 // we don't need to use the PLT - we can directly call it.
2248 if (Subtarget->isTargetELF() &&
2249 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2250 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2251 OpFlags = X86II::MO_PLT;
2252 } else if (Subtarget->isPICStyleStubAny() &&
2253 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2254 Subtarget->getDarwinVers() < 9) {
2255 // PC-relative references to external symbols should go through $stub,
2256 // unless we're building with the leopard linker or later, which
2257 // automatically synthesizes these stubs.
2258 OpFlags = X86II::MO_DARWIN_STUB;
2261 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2262 G->getOffset(), OpFlags);
2264 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2265 unsigned char OpFlags = 0;
2267 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2268 // external symbols should go through the PLT.
2269 if (Subtarget->isTargetELF() &&
2270 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2271 OpFlags = X86II::MO_PLT;
2272 } else if (Subtarget->isPICStyleStubAny() &&
2273 Subtarget->getDarwinVers() < 9) {
2274 // PC-relative references to external symbols should go through $stub,
2275 // unless we're building with the leopard linker or later, which
2276 // automatically synthesizes these stubs.
2277 OpFlags = X86II::MO_DARWIN_STUB;
2280 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2284 // Returns a chain & a flag for retval copy to use.
2285 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2286 SmallVector<SDValue, 8> Ops;
2288 if (!IsSibcall && isTailCall) {
2289 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2290 DAG.getIntPtrConstant(0, true), InFlag);
2291 InFlag = Chain.getValue(1);
2294 Ops.push_back(Chain);
2295 Ops.push_back(Callee);
2298 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2300 // Add argument registers to the end of the list so that they are known live
2302 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2303 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2304 RegsToPass[i].second.getValueType()));
2306 // Add an implicit use GOT pointer in EBX.
2307 if (!isTailCall && Subtarget->isPICStyleGOT())
2308 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2310 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2311 if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
2312 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2314 if (InFlag.getNode())
2315 Ops.push_back(InFlag);
2319 //// If this is the first return lowered for this function, add the regs
2320 //// to the liveout set for the function.
2321 // This isn't right, although it's probably harmless on x86; liveouts
2322 // should be computed from returns not tail calls. Consider a void
2323 // function making a tail call to a function returning int.
2324 return DAG.getNode(X86ISD::TC_RETURN, dl,
2325 NodeTys, &Ops[0], Ops.size());
2328 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2329 InFlag = Chain.getValue(1);
2331 // Create the CALLSEQ_END node.
2332 unsigned NumBytesForCalleeToPush;
2333 if (Subtarget->IsCalleePop(isVarArg, CallConv))
2334 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2335 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2336 // If this is a call to a struct-return function, the callee
2337 // pops the hidden struct pointer, so we have to push it back.
2338 // This is common for Darwin/X86, Linux & Mingw32 targets.
2339 NumBytesForCalleeToPush = 4;
2341 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2343 // Returns a flag for retval copy to use.
2345 Chain = DAG.getCALLSEQ_END(Chain,
2346 DAG.getIntPtrConstant(NumBytes, true),
2347 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2350 InFlag = Chain.getValue(1);
2353 // Handle result values, copying them out of physregs into vregs that we
2355 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2356 Ins, dl, DAG, InVals);
2360 //===----------------------------------------------------------------------===//
2361 // Fast Calling Convention (tail call) implementation
2362 //===----------------------------------------------------------------------===//
2364 // Like std call, callee cleans arguments, convention except that ECX is
2365 // reserved for storing the tail called function address. Only 2 registers are
2366 // free for argument passing (inreg). Tail call optimization is performed
2368 // * tailcallopt is enabled
2369 // * caller/callee are fastcc
2370 // On X86_64 architecture with GOT-style position independent code only local
2371 // (within module) calls are supported at the moment.
2372 // To keep the stack aligned according to platform abi the function
2373 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2374 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2375 // If a tail called function callee has more arguments than the caller the
2376 // caller needs to make sure that there is room to move the RETADDR to. This is
2377 // achieved by reserving an area the size of the argument delta right after the
2378 // original REtADDR, but before the saved framepointer or the spilled registers
2379 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2391 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2392 /// for a 16 byte align requirement.
2394 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2395 SelectionDAG& DAG) const {
2396 MachineFunction &MF = DAG.getMachineFunction();
2397 const TargetMachine &TM = MF.getTarget();
2398 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2399 unsigned StackAlignment = TFI.getStackAlignment();
2400 uint64_t AlignMask = StackAlignment - 1;
2401 int64_t Offset = StackSize;
2402 uint64_t SlotSize = TD->getPointerSize();
2403 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2404 // Number smaller than 12 so just add the difference.
2405 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2407 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2408 Offset = ((~AlignMask) & Offset) + StackAlignment +
2409 (StackAlignment-SlotSize);
2414 /// MatchingStackOffset - Return true if the given stack call argument is
2415 /// already available in the same position (relatively) of the caller's
2416 /// incoming argument stack.
2418 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2419 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2420 const X86InstrInfo *TII) {
2421 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2423 if (Arg.getOpcode() == ISD::CopyFromReg) {
2424 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2425 if (!TargetRegisterInfo::isVirtualRegister(VR))
2427 MachineInstr *Def = MRI->getVRegDef(VR);
2430 if (!Flags.isByVal()) {
2431 if (!TII->isLoadFromStackSlot(Def, FI))
2434 unsigned Opcode = Def->getOpcode();
2435 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2436 Def->getOperand(1).isFI()) {
2437 FI = Def->getOperand(1).getIndex();
2438 Bytes = Flags.getByValSize();
2442 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2443 if (Flags.isByVal())
2444 // ByVal argument is passed in as a pointer but it's now being
2445 // dereferenced. e.g.
2446 // define @foo(%struct.X* %A) {
2447 // tail call @bar(%struct.X* byval %A)
2450 SDValue Ptr = Ld->getBasePtr();
2451 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2454 FI = FINode->getIndex();
2458 assert(FI != INT_MAX);
2459 if (!MFI->isFixedObjectIndex(FI))
2461 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2464 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2465 /// for tail call optimization. Targets which want to do tail call
2466 /// optimization should implement this function.
2468 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2469 CallingConv::ID CalleeCC,
2471 bool isCalleeStructRet,
2472 bool isCallerStructRet,
2473 const SmallVectorImpl<ISD::OutputArg> &Outs,
2474 const SmallVectorImpl<SDValue> &OutVals,
2475 const SmallVectorImpl<ISD::InputArg> &Ins,
2476 SelectionDAG& DAG) const {
2477 if (!IsTailCallConvention(CalleeCC) &&
2478 CalleeCC != CallingConv::C)
2481 // If -tailcallopt is specified, make fastcc functions tail-callable.
2482 const MachineFunction &MF = DAG.getMachineFunction();
2483 const Function *CallerF = DAG.getMachineFunction().getFunction();
2484 CallingConv::ID CallerCC = CallerF->getCallingConv();
2485 bool CCMatch = CallerCC == CalleeCC;
2487 if (GuaranteedTailCallOpt) {
2488 if (IsTailCallConvention(CalleeCC) && CCMatch)
2493 // Look for obvious safe cases to perform tail call optimization that do not
2494 // require ABI changes. This is what gcc calls sibcall.
2496 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2497 // emit a special epilogue.
2498 if (RegInfo->needsStackRealignment(MF))
2501 // Do not sibcall optimize vararg calls unless the call site is not passing
2503 if (isVarArg && !Outs.empty())
2506 // Also avoid sibcall optimization if either caller or callee uses struct
2507 // return semantics.
2508 if (isCalleeStructRet || isCallerStructRet)
2511 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2512 // Therefore if it's not used by the call it is not safe to optimize this into
2514 bool Unused = false;
2515 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2522 SmallVector<CCValAssign, 16> RVLocs;
2523 CCState CCInfo(CalleeCC, false, getTargetMachine(),
2524 RVLocs, *DAG.getContext());
2525 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2526 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2527 CCValAssign &VA = RVLocs[i];
2528 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2533 // If the calling conventions do not match, then we'd better make sure the
2534 // results are returned in the same way as what the caller expects.
2536 SmallVector<CCValAssign, 16> RVLocs1;
2537 CCState CCInfo1(CalleeCC, false, getTargetMachine(),
2538 RVLocs1, *DAG.getContext());
2539 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2541 SmallVector<CCValAssign, 16> RVLocs2;
2542 CCState CCInfo2(CallerCC, false, getTargetMachine(),
2543 RVLocs2, *DAG.getContext());
2544 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2546 if (RVLocs1.size() != RVLocs2.size())
2548 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2549 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2551 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2553 if (RVLocs1[i].isRegLoc()) {
2554 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2557 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2563 // If the callee takes no arguments then go on to check the results of the
2565 if (!Outs.empty()) {
2566 // Check if stack adjustment is needed. For now, do not do this if any
2567 // argument is passed on the stack.
2568 SmallVector<CCValAssign, 16> ArgLocs;
2569 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2570 ArgLocs, *DAG.getContext());
2571 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2572 if (CCInfo.getNextStackOffset()) {
2573 MachineFunction &MF = DAG.getMachineFunction();
2574 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2577 // Check if the arguments are already laid out in the right way as
2578 // the caller's fixed stack objects.
2579 MachineFrameInfo *MFI = MF.getFrameInfo();
2580 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2581 const X86InstrInfo *TII =
2582 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2583 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2584 CCValAssign &VA = ArgLocs[i];
2585 SDValue Arg = OutVals[i];
2586 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2587 if (VA.getLocInfo() == CCValAssign::Indirect)
2589 if (!VA.isRegLoc()) {
2590 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2597 // If the tailcall address may be in a register, then make sure it's
2598 // possible to register allocate for it. In 32-bit, the call address can
2599 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2600 // callee-saved registers are restored. These happen to be the same
2601 // registers used to pass 'inreg' arguments so watch out for those.
2602 if (!Subtarget->is64Bit() &&
2603 !isa<GlobalAddressSDNode>(Callee) &&
2604 !isa<ExternalSymbolSDNode>(Callee)) {
2605 unsigned NumInRegs = 0;
2606 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2607 CCValAssign &VA = ArgLocs[i];
2610 unsigned Reg = VA.getLocReg();
2613 case X86::EAX: case X86::EDX: case X86::ECX:
2614 if (++NumInRegs == 3)
2622 // An stdcall caller is expected to clean up its arguments; the callee
2623 // isn't going to do that.
2624 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2631 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2632 return X86::createFastISel(funcInfo);
2636 //===----------------------------------------------------------------------===//
2637 // Other Lowering Hooks
2638 //===----------------------------------------------------------------------===//
2640 static bool MayFoldLoad(SDValue Op) {
2641 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2644 static bool MayFoldIntoStore(SDValue Op) {
2645 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2648 static bool isTargetShuffle(unsigned Opcode) {
2650 default: return false;
2651 case X86ISD::PSHUFD:
2652 case X86ISD::PSHUFHW:
2653 case X86ISD::PSHUFLW:
2654 case X86ISD::SHUFPD:
2655 case X86ISD::PALIGN:
2656 case X86ISD::SHUFPS:
2657 case X86ISD::MOVLHPS:
2658 case X86ISD::MOVLHPD:
2659 case X86ISD::MOVHLPS:
2660 case X86ISD::MOVLPS:
2661 case X86ISD::MOVLPD:
2662 case X86ISD::MOVSHDUP:
2663 case X86ISD::MOVSLDUP:
2664 case X86ISD::MOVDDUP:
2667 case X86ISD::UNPCKLPS:
2668 case X86ISD::UNPCKLPD:
2669 case X86ISD::PUNPCKLWD:
2670 case X86ISD::PUNPCKLBW:
2671 case X86ISD::PUNPCKLDQ:
2672 case X86ISD::PUNPCKLQDQ:
2673 case X86ISD::UNPCKHPS:
2674 case X86ISD::UNPCKHPD:
2675 case X86ISD::PUNPCKHWD:
2676 case X86ISD::PUNPCKHBW:
2677 case X86ISD::PUNPCKHDQ:
2678 case X86ISD::PUNPCKHQDQ:
2684 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2685 SDValue V1, SelectionDAG &DAG) {
2687 default: llvm_unreachable("Unknown x86 shuffle node");
2688 case X86ISD::MOVSHDUP:
2689 case X86ISD::MOVSLDUP:
2690 case X86ISD::MOVDDUP:
2691 return DAG.getNode(Opc, dl, VT, V1);
2697 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2698 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2700 default: llvm_unreachable("Unknown x86 shuffle node");
2701 case X86ISD::PSHUFD:
2702 case X86ISD::PSHUFHW:
2703 case X86ISD::PSHUFLW:
2704 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2710 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2711 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2713 default: llvm_unreachable("Unknown x86 shuffle node");
2714 case X86ISD::PALIGN:
2715 case X86ISD::SHUFPD:
2716 case X86ISD::SHUFPS:
2717 return DAG.getNode(Opc, dl, VT, V1, V2,
2718 DAG.getConstant(TargetMask, MVT::i8));
2723 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2724 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2726 default: llvm_unreachable("Unknown x86 shuffle node");
2727 case X86ISD::MOVLHPS:
2728 case X86ISD::MOVLHPD:
2729 case X86ISD::MOVHLPS:
2730 case X86ISD::MOVLPS:
2731 case X86ISD::MOVLPD:
2734 case X86ISD::UNPCKLPS:
2735 case X86ISD::UNPCKLPD:
2736 case X86ISD::PUNPCKLWD:
2737 case X86ISD::PUNPCKLBW:
2738 case X86ISD::PUNPCKLDQ:
2739 case X86ISD::PUNPCKLQDQ:
2740 case X86ISD::UNPCKHPS:
2741 case X86ISD::UNPCKHPD:
2742 case X86ISD::PUNPCKHWD:
2743 case X86ISD::PUNPCKHBW:
2744 case X86ISD::PUNPCKHDQ:
2745 case X86ISD::PUNPCKHQDQ:
2746 return DAG.getNode(Opc, dl, VT, V1, V2);
2751 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2752 MachineFunction &MF = DAG.getMachineFunction();
2753 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2754 int ReturnAddrIndex = FuncInfo->getRAIndex();
2756 if (ReturnAddrIndex == 0) {
2757 // Set up a frame object for the return address.
2758 uint64_t SlotSize = TD->getPointerSize();
2759 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2761 FuncInfo->setRAIndex(ReturnAddrIndex);
2764 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2768 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2769 bool hasSymbolicDisplacement) {
2770 // Offset should fit into 32 bit immediate field.
2771 if (!isInt<32>(Offset))
2774 // If we don't have a symbolic displacement - we don't have any extra
2776 if (!hasSymbolicDisplacement)
2779 // FIXME: Some tweaks might be needed for medium code model.
2780 if (M != CodeModel::Small && M != CodeModel::Kernel)
2783 // For small code model we assume that latest object is 16MB before end of 31
2784 // bits boundary. We may also accept pretty large negative constants knowing
2785 // that all objects are in the positive half of address space.
2786 if (M == CodeModel::Small && Offset < 16*1024*1024)
2789 // For kernel code model we know that all object resist in the negative half
2790 // of 32bits address space. We may not accept negative offsets, since they may
2791 // be just off and we may accept pretty large positive ones.
2792 if (M == CodeModel::Kernel && Offset > 0)
2798 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2799 /// specific condition code, returning the condition code and the LHS/RHS of the
2800 /// comparison to make.
2801 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2802 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2804 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2805 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2806 // X > -1 -> X == 0, jump !sign.
2807 RHS = DAG.getConstant(0, RHS.getValueType());
2808 return X86::COND_NS;
2809 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2810 // X < 0 -> X == 0, jump on sign.
2812 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2814 RHS = DAG.getConstant(0, RHS.getValueType());
2815 return X86::COND_LE;
2819 switch (SetCCOpcode) {
2820 default: llvm_unreachable("Invalid integer condition!");
2821 case ISD::SETEQ: return X86::COND_E;
2822 case ISD::SETGT: return X86::COND_G;
2823 case ISD::SETGE: return X86::COND_GE;
2824 case ISD::SETLT: return X86::COND_L;
2825 case ISD::SETLE: return X86::COND_LE;
2826 case ISD::SETNE: return X86::COND_NE;
2827 case ISD::SETULT: return X86::COND_B;
2828 case ISD::SETUGT: return X86::COND_A;
2829 case ISD::SETULE: return X86::COND_BE;
2830 case ISD::SETUGE: return X86::COND_AE;
2834 // First determine if it is required or is profitable to flip the operands.
2836 // If LHS is a foldable load, but RHS is not, flip the condition.
2837 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2838 !ISD::isNON_EXTLoad(RHS.getNode())) {
2839 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2840 std::swap(LHS, RHS);
2843 switch (SetCCOpcode) {
2849 std::swap(LHS, RHS);
2853 // On a floating point condition, the flags are set as follows:
2855 // 0 | 0 | 0 | X > Y
2856 // 0 | 0 | 1 | X < Y
2857 // 1 | 0 | 0 | X == Y
2858 // 1 | 1 | 1 | unordered
2859 switch (SetCCOpcode) {
2860 default: llvm_unreachable("Condcode should be pre-legalized away");
2862 case ISD::SETEQ: return X86::COND_E;
2863 case ISD::SETOLT: // flipped
2865 case ISD::SETGT: return X86::COND_A;
2866 case ISD::SETOLE: // flipped
2868 case ISD::SETGE: return X86::COND_AE;
2869 case ISD::SETUGT: // flipped
2871 case ISD::SETLT: return X86::COND_B;
2872 case ISD::SETUGE: // flipped
2874 case ISD::SETLE: return X86::COND_BE;
2876 case ISD::SETNE: return X86::COND_NE;
2877 case ISD::SETUO: return X86::COND_P;
2878 case ISD::SETO: return X86::COND_NP;
2880 case ISD::SETUNE: return X86::COND_INVALID;
2884 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2885 /// code. Current x86 isa includes the following FP cmov instructions:
2886 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2887 static bool hasFPCMov(unsigned X86CC) {
2903 /// isFPImmLegal - Returns true if the target can instruction select the
2904 /// specified FP immediate natively. If false, the legalizer will
2905 /// materialize the FP immediate as a load from a constant pool.
2906 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2907 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2908 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2914 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2915 /// the specified range (L, H].
2916 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2917 return (Val < 0) || (Val >= Low && Val < Hi);
2920 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2921 /// specified value.
2922 static bool isUndefOrEqual(int Val, int CmpVal) {
2923 if (Val < 0 || Val == CmpVal)
2928 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2929 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2930 /// the second operand.
2931 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2932 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
2933 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2934 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2935 return (Mask[0] < 2 && Mask[1] < 2);
2939 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2940 SmallVector<int, 8> M;
2942 return ::isPSHUFDMask(M, N->getValueType(0));
2945 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2946 /// is suitable for input to PSHUFHW.
2947 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2948 if (VT != MVT::v8i16)
2951 // Lower quadword copied in order or undef.
2952 for (int i = 0; i != 4; ++i)
2953 if (Mask[i] >= 0 && Mask[i] != i)
2956 // Upper quadword shuffled.
2957 for (int i = 4; i != 8; ++i)
2958 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2964 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2965 SmallVector<int, 8> M;
2967 return ::isPSHUFHWMask(M, N->getValueType(0));
2970 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2971 /// is suitable for input to PSHUFLW.
2972 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2973 if (VT != MVT::v8i16)
2976 // Upper quadword copied in order.
2977 for (int i = 4; i != 8; ++i)
2978 if (Mask[i] >= 0 && Mask[i] != i)
2981 // Lower quadword shuffled.
2982 for (int i = 0; i != 4; ++i)
2989 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2990 SmallVector<int, 8> M;
2992 return ::isPSHUFLWMask(M, N->getValueType(0));
2995 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2996 /// is suitable for input to PALIGNR.
2997 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2999 int i, e = VT.getVectorNumElements();
3001 // Do not handle v2i64 / v2f64 shuffles with palignr.
3002 if (e < 4 || !hasSSSE3)
3005 for (i = 0; i != e; ++i)
3009 // All undef, not a palignr.
3013 // Determine if it's ok to perform a palignr with only the LHS, since we
3014 // don't have access to the actual shuffle elements to see if RHS is undef.
3015 bool Unary = Mask[i] < (int)e;
3016 bool NeedsUnary = false;
3018 int s = Mask[i] - i;
3020 // Check the rest of the elements to see if they are consecutive.
3021 for (++i; i != e; ++i) {
3026 Unary = Unary && (m < (int)e);
3027 NeedsUnary = NeedsUnary || (m < s);
3029 if (NeedsUnary && !Unary)
3031 if (Unary && m != ((s+i) & (e-1)))
3033 if (!Unary && m != (s+i))
3039 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
3040 SmallVector<int, 8> M;
3042 return ::isPALIGNRMask(M, N->getValueType(0), true);
3045 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3046 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3047 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3048 int NumElems = VT.getVectorNumElements();
3049 if (NumElems != 2 && NumElems != 4)
3052 int Half = NumElems / 2;
3053 for (int i = 0; i < Half; ++i)
3054 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3056 for (int i = Half; i < NumElems; ++i)
3057 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3063 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3064 SmallVector<int, 8> M;
3066 return ::isSHUFPMask(M, N->getValueType(0));
3069 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3070 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3071 /// half elements to come from vector 1 (which would equal the dest.) and
3072 /// the upper half to come from vector 2.
3073 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3074 int NumElems = VT.getVectorNumElements();
3076 if (NumElems != 2 && NumElems != 4)
3079 int Half = NumElems / 2;
3080 for (int i = 0; i < Half; ++i)
3081 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3083 for (int i = Half; i < NumElems; ++i)
3084 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3089 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3090 SmallVector<int, 8> M;
3092 return isCommutedSHUFPMask(M, N->getValueType(0));
3095 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3096 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3097 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3098 if (N->getValueType(0).getVectorNumElements() != 4)
3101 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3102 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3103 isUndefOrEqual(N->getMaskElt(1), 7) &&
3104 isUndefOrEqual(N->getMaskElt(2), 2) &&
3105 isUndefOrEqual(N->getMaskElt(3), 3);
3108 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3109 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3111 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3112 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3117 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3118 isUndefOrEqual(N->getMaskElt(1), 3) &&
3119 isUndefOrEqual(N->getMaskElt(2), 2) &&
3120 isUndefOrEqual(N->getMaskElt(3), 3);
3123 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3124 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3125 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3126 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3128 if (NumElems != 2 && NumElems != 4)
3131 for (unsigned i = 0; i < NumElems/2; ++i)
3132 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3135 for (unsigned i = NumElems/2; i < NumElems; ++i)
3136 if (!isUndefOrEqual(N->getMaskElt(i), i))
3142 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3143 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3144 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3145 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3147 if (NumElems != 2 && NumElems != 4)
3150 for (unsigned i = 0; i < NumElems/2; ++i)
3151 if (!isUndefOrEqual(N->getMaskElt(i), i))
3154 for (unsigned i = 0; i < NumElems/2; ++i)
3155 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3161 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3162 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3163 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3164 bool V2IsSplat = false) {
3165 int NumElts = VT.getVectorNumElements();
3166 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3169 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3171 int BitI1 = Mask[i+1];
3172 if (!isUndefOrEqual(BitI, j))
3175 if (!isUndefOrEqual(BitI1, NumElts))
3178 if (!isUndefOrEqual(BitI1, j + NumElts))
3185 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3186 SmallVector<int, 8> M;
3188 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3191 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3192 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3193 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3194 bool V2IsSplat = false) {
3195 int NumElts = VT.getVectorNumElements();
3196 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
3199 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
3201 int BitI1 = Mask[i+1];
3202 if (!isUndefOrEqual(BitI, j + NumElts/2))
3205 if (isUndefOrEqual(BitI1, NumElts))
3208 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
3215 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3216 SmallVector<int, 8> M;
3218 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3221 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3222 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3224 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3225 int NumElems = VT.getVectorNumElements();
3226 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3229 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
3231 int BitI1 = Mask[i+1];
3232 if (!isUndefOrEqual(BitI, j))
3234 if (!isUndefOrEqual(BitI1, j))
3240 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3241 SmallVector<int, 8> M;
3243 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3246 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3247 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3249 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3250 int NumElems = VT.getVectorNumElements();
3251 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3254 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3256 int BitI1 = Mask[i+1];
3257 if (!isUndefOrEqual(BitI, j))
3259 if (!isUndefOrEqual(BitI1, j))
3265 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3266 SmallVector<int, 8> M;
3268 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3271 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3272 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3273 /// MOVSD, and MOVD, i.e. setting the lowest element.
3274 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3275 if (VT.getVectorElementType().getSizeInBits() < 32)
3278 int NumElts = VT.getVectorNumElements();
3280 if (!isUndefOrEqual(Mask[0], NumElts))
3283 for (int i = 1; i < NumElts; ++i)
3284 if (!isUndefOrEqual(Mask[i], i))
3290 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3291 SmallVector<int, 8> M;
3293 return ::isMOVLMask(M, N->getValueType(0));
3296 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3297 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3298 /// element of vector 2 and the other elements to come from vector 1 in order.
3299 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3300 bool V2IsSplat = false, bool V2IsUndef = false) {
3301 int NumOps = VT.getVectorNumElements();
3302 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3305 if (!isUndefOrEqual(Mask[0], 0))
3308 for (int i = 1; i < NumOps; ++i)
3309 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3310 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3311 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3317 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3318 bool V2IsUndef = false) {
3319 SmallVector<int, 8> M;
3321 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3324 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3325 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3326 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
3327 if (N->getValueType(0).getVectorNumElements() != 4)
3330 // Expect 1, 1, 3, 3
3331 for (unsigned i = 0; i < 2; ++i) {
3332 int Elt = N->getMaskElt(i);
3333 if (Elt >= 0 && Elt != 1)
3338 for (unsigned i = 2; i < 4; ++i) {
3339 int Elt = N->getMaskElt(i);
3340 if (Elt >= 0 && Elt != 3)
3345 // Don't use movshdup if it can be done with a shufps.
3346 // FIXME: verify that matching u, u, 3, 3 is what we want.
3350 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3351 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3352 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
3353 if (N->getValueType(0).getVectorNumElements() != 4)
3356 // Expect 0, 0, 2, 2
3357 for (unsigned i = 0; i < 2; ++i)
3358 if (N->getMaskElt(i) > 0)
3362 for (unsigned i = 2; i < 4; ++i) {
3363 int Elt = N->getMaskElt(i);
3364 if (Elt >= 0 && Elt != 2)
3369 // Don't use movsldup if it can be done with a shufps.
3373 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3374 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3375 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3376 int e = N->getValueType(0).getVectorNumElements() / 2;
3378 for (int i = 0; i < e; ++i)
3379 if (!isUndefOrEqual(N->getMaskElt(i), i))
3381 for (int i = 0; i < e; ++i)
3382 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3387 /// isVEXTRACTF128Index - Return true if the specified
3388 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3389 /// suitable for input to VEXTRACTF128.
3390 bool X86::isVEXTRACTF128Index(SDNode *N) {
3391 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3394 // The index should be aligned on a 128-bit boundary.
3396 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3398 unsigned VL = N->getValueType(0).getVectorNumElements();
3399 unsigned VBits = N->getValueType(0).getSizeInBits();
3400 unsigned ElSize = VBits / VL;
3401 bool Result = (Index * ElSize) % 128 == 0;
3406 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3407 /// operand specifies a subvector insert that is suitable for input to
3409 bool X86::isVINSERTF128Index(SDNode *N) {
3410 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3413 // The index should be aligned on a 128-bit boundary.
3415 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3417 unsigned VL = N->getValueType(0).getVectorNumElements();
3418 unsigned VBits = N->getValueType(0).getSizeInBits();
3419 unsigned ElSize = VBits / VL;
3420 bool Result = (Index * ElSize) % 128 == 0;
3425 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3426 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3427 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3428 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3429 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3431 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3433 for (int i = 0; i < NumOperands; ++i) {
3434 int Val = SVOp->getMaskElt(NumOperands-i-1);
3435 if (Val < 0) Val = 0;
3436 if (Val >= NumOperands) Val -= NumOperands;
3438 if (i != NumOperands - 1)
3444 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3445 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3446 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3447 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3449 // 8 nodes, but we only care about the last 4.
3450 for (unsigned i = 7; i >= 4; --i) {
3451 int Val = SVOp->getMaskElt(i);
3460 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3461 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3462 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3463 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3465 // 8 nodes, but we only care about the first 4.
3466 for (int i = 3; i >= 0; --i) {
3467 int Val = SVOp->getMaskElt(i);
3476 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3477 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3478 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3479 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3480 EVT VVT = N->getValueType(0);
3481 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3485 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3486 Val = SVOp->getMaskElt(i);
3490 return (Val - i) * EltSize;
3493 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3494 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3496 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3497 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3498 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3501 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3503 EVT VecVT = N->getOperand(0).getValueType();
3504 EVT ElVT = VecVT.getVectorElementType();
3506 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3508 return Index / NumElemsPerChunk;
3511 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3512 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3514 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3515 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3516 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3519 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3521 EVT VecVT = N->getValueType(0);
3522 EVT ElVT = VecVT.getVectorElementType();
3524 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3526 return Index / NumElemsPerChunk;
3529 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3531 bool X86::isZeroNode(SDValue Elt) {
3532 return ((isa<ConstantSDNode>(Elt) &&
3533 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3534 (isa<ConstantFPSDNode>(Elt) &&
3535 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3538 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3539 /// their permute mask.
3540 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3541 SelectionDAG &DAG) {
3542 EVT VT = SVOp->getValueType(0);
3543 unsigned NumElems = VT.getVectorNumElements();
3544 SmallVector<int, 8> MaskVec;
3546 for (unsigned i = 0; i != NumElems; ++i) {
3547 int idx = SVOp->getMaskElt(i);
3549 MaskVec.push_back(idx);
3550 else if (idx < (int)NumElems)
3551 MaskVec.push_back(idx + NumElems);
3553 MaskVec.push_back(idx - NumElems);
3555 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3556 SVOp->getOperand(0), &MaskVec[0]);
3559 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3560 /// the two vector operands have swapped position.
3561 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3562 unsigned NumElems = VT.getVectorNumElements();
3563 for (unsigned i = 0; i != NumElems; ++i) {
3567 else if (idx < (int)NumElems)
3568 Mask[i] = idx + NumElems;
3570 Mask[i] = idx - NumElems;
3574 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3575 /// match movhlps. The lower half elements should come from upper half of
3576 /// V1 (and in order), and the upper half elements should come from the upper
3577 /// half of V2 (and in order).
3578 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3579 if (Op->getValueType(0).getVectorNumElements() != 4)
3581 for (unsigned i = 0, e = 2; i != e; ++i)
3582 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3584 for (unsigned i = 2; i != 4; ++i)
3585 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3590 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3591 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3593 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3594 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3596 N = N->getOperand(0).getNode();
3597 if (!ISD::isNON_EXTLoad(N))
3600 *LD = cast<LoadSDNode>(N);
3604 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3605 /// match movlp{s|d}. The lower half elements should come from lower half of
3606 /// V1 (and in order), and the upper half elements should come from the upper
3607 /// half of V2 (and in order). And since V1 will become the source of the
3608 /// MOVLP, it must be either a vector load or a scalar load to vector.
3609 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3610 ShuffleVectorSDNode *Op) {
3611 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3613 // Is V2 is a vector load, don't do this transformation. We will try to use
3614 // load folding shufps op.
3615 if (ISD::isNON_EXTLoad(V2))
3618 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3620 if (NumElems != 2 && NumElems != 4)
3622 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3623 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3625 for (unsigned i = NumElems/2; i != NumElems; ++i)
3626 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3631 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3633 static bool isSplatVector(SDNode *N) {
3634 if (N->getOpcode() != ISD::BUILD_VECTOR)
3637 SDValue SplatValue = N->getOperand(0);
3638 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3639 if (N->getOperand(i) != SplatValue)
3644 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3645 /// to an zero vector.
3646 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3647 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3648 SDValue V1 = N->getOperand(0);
3649 SDValue V2 = N->getOperand(1);
3650 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3651 for (unsigned i = 0; i != NumElems; ++i) {
3652 int Idx = N->getMaskElt(i);
3653 if (Idx >= (int)NumElems) {
3654 unsigned Opc = V2.getOpcode();
3655 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3657 if (Opc != ISD::BUILD_VECTOR ||
3658 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3660 } else if (Idx >= 0) {
3661 unsigned Opc = V1.getOpcode();
3662 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3664 if (Opc != ISD::BUILD_VECTOR ||
3665 !X86::isZeroNode(V1.getOperand(Idx)))
3672 /// getZeroVector - Returns a vector of specified type with all zero elements.
3674 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3676 assert(VT.isVector() && "Expected a vector type");
3678 // Always build SSE zero vectors as <4 x i32> bitcasted
3679 // to their dest type. This ensures they get CSE'd.
3681 if (VT.getSizeInBits() == 128) { // SSE
3682 if (HasSSE2) { // SSE2
3683 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3684 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3686 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3687 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3689 } else if (VT.getSizeInBits() == 256) { // AVX
3690 // 256-bit logic and arithmetic instructions in AVX are
3691 // all floating-point, no support for integer ops. Default
3692 // to emitting fp zeroed vectors then.
3693 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3694 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3695 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3697 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3700 /// getOnesVector - Returns a vector of specified type with all bits set.
3702 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3703 assert(VT.isVector() && "Expected a vector type");
3705 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3706 // type. This ensures they get CSE'd.
3707 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3709 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3710 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3714 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3715 /// that point to V2 points to its first element.
3716 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3717 EVT VT = SVOp->getValueType(0);
3718 unsigned NumElems = VT.getVectorNumElements();
3720 bool Changed = false;
3721 SmallVector<int, 8> MaskVec;
3722 SVOp->getMask(MaskVec);
3724 for (unsigned i = 0; i != NumElems; ++i) {
3725 if (MaskVec[i] > (int)NumElems) {
3726 MaskVec[i] = NumElems;
3731 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3732 SVOp->getOperand(1), &MaskVec[0]);
3733 return SDValue(SVOp, 0);
3736 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3737 /// operation of specified width.
3738 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3740 unsigned NumElems = VT.getVectorNumElements();
3741 SmallVector<int, 8> Mask;
3742 Mask.push_back(NumElems);
3743 for (unsigned i = 1; i != NumElems; ++i)
3745 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3748 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3749 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3751 unsigned NumElems = VT.getVectorNumElements();
3752 SmallVector<int, 8> Mask;
3753 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3755 Mask.push_back(i + NumElems);
3757 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3760 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3761 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3763 unsigned NumElems = VT.getVectorNumElements();
3764 unsigned Half = NumElems/2;
3765 SmallVector<int, 8> Mask;
3766 for (unsigned i = 0; i != Half; ++i) {
3767 Mask.push_back(i + Half);
3768 Mask.push_back(i + NumElems + Half);
3770 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3773 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
3774 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
3775 EVT PVT = MVT::v4f32;
3776 EVT VT = SV->getValueType(0);
3777 DebugLoc dl = SV->getDebugLoc();
3778 SDValue V1 = SV->getOperand(0);
3779 int NumElems = VT.getVectorNumElements();
3780 int EltNo = SV->getSplatIndex();
3782 // unpack elements to the correct location
3783 while (NumElems > 4) {
3784 if (EltNo < NumElems/2) {
3785 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3787 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3788 EltNo -= NumElems/2;
3793 // Perform the splat.
3794 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3795 V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
3796 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3797 return DAG.getNode(ISD::BITCAST, dl, VT, V1);
3800 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3801 /// vector of zero or undef vector. This produces a shuffle where the low
3802 /// element of V2 is swizzled into the zero/undef vector, landing at element
3803 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3804 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3805 bool isZero, bool HasSSE2,
3806 SelectionDAG &DAG) {
3807 EVT VT = V2.getValueType();
3809 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3810 unsigned NumElems = VT.getVectorNumElements();
3811 SmallVector<int, 16> MaskVec;
3812 for (unsigned i = 0; i != NumElems; ++i)
3813 // If this is the insertion idx, put the low elt of V2 here.
3814 MaskVec.push_back(i == Idx ? NumElems : i);
3815 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3818 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
3819 /// element of the result of the vector shuffle.
3820 SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
3823 return SDValue(); // Limit search depth.
3825 SDValue V = SDValue(N, 0);
3826 EVT VT = V.getValueType();
3827 unsigned Opcode = V.getOpcode();
3829 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
3830 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
3831 Index = SV->getMaskElt(Index);
3834 return DAG.getUNDEF(VT.getVectorElementType());
3836 int NumElems = VT.getVectorNumElements();
3837 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
3838 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
3841 // Recurse into target specific vector shuffles to find scalars.
3842 if (isTargetShuffle(Opcode)) {
3843 int NumElems = VT.getVectorNumElements();
3844 SmallVector<unsigned, 16> ShuffleMask;
3848 case X86ISD::SHUFPS:
3849 case X86ISD::SHUFPD:
3850 ImmN = N->getOperand(N->getNumOperands()-1);
3851 DecodeSHUFPSMask(NumElems,
3852 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3855 case X86ISD::PUNPCKHBW:
3856 case X86ISD::PUNPCKHWD:
3857 case X86ISD::PUNPCKHDQ:
3858 case X86ISD::PUNPCKHQDQ:
3859 DecodePUNPCKHMask(NumElems, ShuffleMask);
3861 case X86ISD::UNPCKHPS:
3862 case X86ISD::UNPCKHPD:
3863 DecodeUNPCKHPMask(NumElems, ShuffleMask);
3865 case X86ISD::PUNPCKLBW:
3866 case X86ISD::PUNPCKLWD:
3867 case X86ISD::PUNPCKLDQ:
3868 case X86ISD::PUNPCKLQDQ:
3869 DecodePUNPCKLMask(NumElems, ShuffleMask);
3871 case X86ISD::UNPCKLPS:
3872 case X86ISD::UNPCKLPD:
3873 DecodeUNPCKLPMask(NumElems, ShuffleMask);
3875 case X86ISD::MOVHLPS:
3876 DecodeMOVHLPSMask(NumElems, ShuffleMask);
3878 case X86ISD::MOVLHPS:
3879 DecodeMOVLHPSMask(NumElems, ShuffleMask);
3881 case X86ISD::PSHUFD:
3882 ImmN = N->getOperand(N->getNumOperands()-1);
3883 DecodePSHUFMask(NumElems,
3884 cast<ConstantSDNode>(ImmN)->getZExtValue(),
3887 case X86ISD::PSHUFHW:
3888 ImmN = N->getOperand(N->getNumOperands()-1);
3889 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3892 case X86ISD::PSHUFLW:
3893 ImmN = N->getOperand(N->getNumOperands()-1);
3894 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
3898 case X86ISD::MOVSD: {
3899 // The index 0 always comes from the first element of the second source,
3900 // this is why MOVSS and MOVSD are used in the first place. The other
3901 // elements come from the other positions of the first source vector.
3902 unsigned OpNum = (Index == 0) ? 1 : 0;
3903 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
3907 assert("not implemented for target shuffle node");
3911 Index = ShuffleMask[Index];
3913 return DAG.getUNDEF(VT.getVectorElementType());
3915 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
3916 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
3920 // Actual nodes that may contain scalar elements
3921 if (Opcode == ISD::BITCAST) {
3922 V = V.getOperand(0);
3923 EVT SrcVT = V.getValueType();
3924 unsigned NumElems = VT.getVectorNumElements();
3926 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
3930 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
3931 return (Index == 0) ? V.getOperand(0)
3932 : DAG.getUNDEF(VT.getVectorElementType());
3934 if (V.getOpcode() == ISD::BUILD_VECTOR)
3935 return V.getOperand(Index);
3940 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
3941 /// shuffle operation which come from a consecutively from a zero. The
3942 /// search can start in two diferent directions, from left or right.
3944 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
3945 bool ZerosFromLeft, SelectionDAG &DAG) {
3948 while (i < NumElems) {
3949 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
3950 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
3951 if (!(Elt.getNode() &&
3952 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
3960 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
3961 /// MaskE correspond consecutively to elements from one of the vector operands,
3962 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
3964 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
3965 int OpIdx, int NumElems, unsigned &OpNum) {
3966 bool SeenV1 = false;
3967 bool SeenV2 = false;
3969 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
3970 int Idx = SVOp->getMaskElt(i);
3971 // Ignore undef indicies
3980 // Only accept consecutive elements from the same vector
3981 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
3985 OpNum = SeenV1 ? 0 : 1;
3989 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
3990 /// logical left shift of a vector.
3991 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3992 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3993 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
3994 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
3995 false /* check zeros from right */, DAG);
4001 // Considering the elements in the mask that are not consecutive zeros,
4002 // check if they consecutively come from only one of the source vectors.
4004 // V1 = {X, A, B, C} 0
4006 // vector_shuffle V1, V2 <1, 2, 3, X>
4008 if (!isShuffleMaskConsecutive(SVOp,
4009 0, // Mask Start Index
4010 NumElems-NumZeros-1, // Mask End Index
4011 NumZeros, // Where to start looking in the src vector
4012 NumElems, // Number of elements in vector
4013 OpSrc)) // Which source operand ?
4018 ShVal = SVOp->getOperand(OpSrc);
4022 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4023 /// logical left shift of a vector.
4024 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4025 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4026 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4027 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4028 true /* check zeros from left */, DAG);
4034 // Considering the elements in the mask that are not consecutive zeros,
4035 // check if they consecutively come from only one of the source vectors.
4037 // 0 { A, B, X, X } = V2
4039 // vector_shuffle V1, V2 <X, X, 4, 5>
4041 if (!isShuffleMaskConsecutive(SVOp,
4042 NumZeros, // Mask Start Index
4043 NumElems-1, // Mask End Index
4044 0, // Where to start looking in the src vector
4045 NumElems, // Number of elements in vector
4046 OpSrc)) // Which source operand ?
4051 ShVal = SVOp->getOperand(OpSrc);
4055 /// isVectorShift - Returns true if the shuffle can be implemented as a
4056 /// logical left or right shift of a vector.
4057 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4058 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4059 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4060 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4066 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4068 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4069 unsigned NumNonZero, unsigned NumZero,
4071 const TargetLowering &TLI) {
4075 DebugLoc dl = Op.getDebugLoc();
4078 for (unsigned i = 0; i < 16; ++i) {
4079 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4080 if (ThisIsNonZero && First) {
4082 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4084 V = DAG.getUNDEF(MVT::v8i16);
4089 SDValue ThisElt(0, 0), LastElt(0, 0);
4090 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4091 if (LastIsNonZero) {
4092 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4093 MVT::i16, Op.getOperand(i-1));
4095 if (ThisIsNonZero) {
4096 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4097 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4098 ThisElt, DAG.getConstant(8, MVT::i8));
4100 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4104 if (ThisElt.getNode())
4105 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4106 DAG.getIntPtrConstant(i/2));
4110 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4113 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4115 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4116 unsigned NumNonZero, unsigned NumZero,
4118 const TargetLowering &TLI) {
4122 DebugLoc dl = Op.getDebugLoc();
4125 for (unsigned i = 0; i < 8; ++i) {
4126 bool isNonZero = (NonZeros & (1 << i)) != 0;
4130 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4132 V = DAG.getUNDEF(MVT::v8i16);
4135 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4136 MVT::v8i16, V, Op.getOperand(i),
4137 DAG.getIntPtrConstant(i));
4144 /// getVShift - Return a vector logical shift node.
4146 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4147 unsigned NumBits, SelectionDAG &DAG,
4148 const TargetLowering &TLI, DebugLoc dl) {
4149 EVT ShVT = MVT::v2i64;
4150 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4151 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4152 return DAG.getNode(ISD::BITCAST, dl, VT,
4153 DAG.getNode(Opc, dl, ShVT, SrcOp,
4154 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
4158 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4159 SelectionDAG &DAG) const {
4161 // Check if the scalar load can be widened into a vector load. And if
4162 // the address is "base + cst" see if the cst can be "absorbed" into
4163 // the shuffle mask.
4164 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4165 SDValue Ptr = LD->getBasePtr();
4166 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4168 EVT PVT = LD->getValueType(0);
4169 if (PVT != MVT::i32 && PVT != MVT::f32)
4174 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4175 FI = FINode->getIndex();
4177 } else if (Ptr.getOpcode() == ISD::ADD &&
4178 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
4179 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4180 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4181 Offset = Ptr.getConstantOperandVal(1);
4182 Ptr = Ptr.getOperand(0);
4187 SDValue Chain = LD->getChain();
4188 // Make sure the stack object alignment is at least 16.
4189 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4190 if (DAG.InferPtrAlignment(Ptr) < 16) {
4191 if (MFI->isFixedObjectIndex(FI)) {
4192 // Can't change the alignment. FIXME: It's possible to compute
4193 // the exact stack offset and reference FI + adjust offset instead.
4194 // If someone *really* cares about this. That's the way to implement it.
4197 MFI->setObjectAlignment(FI, 16);
4201 // (Offset % 16) must be multiple of 4. Then address is then
4202 // Ptr + (Offset & ~15).
4205 if ((Offset % 16) & 3)
4207 int64_t StartOffset = Offset & ~15;
4209 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4210 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4212 int EltNo = (Offset - StartOffset) >> 2;
4213 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
4214 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
4215 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,
4216 LD->getPointerInfo().getWithOffset(StartOffset),
4218 // Canonicalize it to a v4i32 shuffle.
4219 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
4220 return DAG.getNode(ISD::BITCAST, dl, VT,
4221 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
4222 DAG.getUNDEF(MVT::v4i32),&Mask[0]));
4228 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4229 /// vector of type 'VT', see if the elements can be replaced by a single large
4230 /// load which has the same value as a build_vector whose operands are 'elts'.
4232 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4234 /// FIXME: we'd also like to handle the case where the last elements are zero
4235 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4236 /// There's even a handy isZeroNode for that purpose.
4237 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4238 DebugLoc &DL, SelectionDAG &DAG) {
4239 EVT EltVT = VT.getVectorElementType();
4240 unsigned NumElems = Elts.size();
4242 LoadSDNode *LDBase = NULL;
4243 unsigned LastLoadedElt = -1U;
4245 // For each element in the initializer, see if we've found a load or an undef.
4246 // If we don't find an initial load element, or later load elements are
4247 // non-consecutive, bail out.
4248 for (unsigned i = 0; i < NumElems; ++i) {
4249 SDValue Elt = Elts[i];
4251 if (!Elt.getNode() ||
4252 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4255 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4257 LDBase = cast<LoadSDNode>(Elt.getNode());
4261 if (Elt.getOpcode() == ISD::UNDEF)
4264 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4265 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4270 // If we have found an entire vector of loads and undefs, then return a large
4271 // load of the entire vector width starting at the base pointer. If we found
4272 // consecutive loads for the low half, generate a vzext_load node.
4273 if (LastLoadedElt == NumElems - 1) {
4274 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4275 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4276 LDBase->getPointerInfo(),
4277 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4278 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4279 LDBase->getPointerInfo(),
4280 LDBase->isVolatile(), LDBase->isNonTemporal(),
4281 LDBase->getAlignment());
4282 } else if (NumElems == 4 && LastLoadedElt == 1) {
4283 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4284 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4285 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4287 LDBase->getMemOperand());
4288 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4294 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4295 DebugLoc dl = Op.getDebugLoc();
4296 // All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
4297 // All one's are handled with pcmpeqd. In AVX, zero's are handled with
4298 // vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
4299 // is present, so AllOnes is ignored.
4300 if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
4301 (Op.getValueType().getSizeInBits() != 256 &&
4302 ISD::isBuildVectorAllOnes(Op.getNode()))) {
4303 // Canonicalize this to <4 x i32> (SSE) to
4304 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
4305 // eliminated on x86-32 hosts.
4306 if (Op.getValueType() == MVT::v4i32)
4309 if (ISD::isBuildVectorAllOnes(Op.getNode()))
4310 return getOnesVector(Op.getValueType(), DAG, dl);
4311 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4314 EVT VT = Op.getValueType();
4315 EVT ExtVT = VT.getVectorElementType();
4316 unsigned EVTBits = ExtVT.getSizeInBits();
4318 unsigned NumElems = Op.getNumOperands();
4319 unsigned NumZero = 0;
4320 unsigned NumNonZero = 0;
4321 unsigned NonZeros = 0;
4322 bool IsAllConstants = true;
4323 SmallSet<SDValue, 8> Values;
4324 for (unsigned i = 0; i < NumElems; ++i) {
4325 SDValue Elt = Op.getOperand(i);
4326 if (Elt.getOpcode() == ISD::UNDEF)
4329 if (Elt.getOpcode() != ISD::Constant &&
4330 Elt.getOpcode() != ISD::ConstantFP)
4331 IsAllConstants = false;
4332 if (X86::isZeroNode(Elt))
4335 NonZeros |= (1 << i);
4340 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4341 if (NumNonZero == 0)
4342 return DAG.getUNDEF(VT);
4344 // Special case for single non-zero, non-undef, element.
4345 if (NumNonZero == 1) {
4346 unsigned Idx = CountTrailingZeros_32(NonZeros);
4347 SDValue Item = Op.getOperand(Idx);
4349 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4350 // the value are obviously zero, truncate the value to i32 and do the
4351 // insertion that way. Only do this if the value is non-constant or if the
4352 // value is a constant being inserted into element 0. It is cheaper to do
4353 // a constant pool load than it is to do a movd + shuffle.
4354 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4355 (!IsAllConstants || Idx == 0)) {
4356 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4358 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4359 EVT VecVT = MVT::v4i32;
4360 unsigned VecElts = 4;
4362 // Truncate the value (which may itself be a constant) to i32, and
4363 // convert it to a vector with movd (S2V+shuffle to zero extend).
4364 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4365 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4366 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4367 Subtarget->hasSSE2(), DAG);
4369 // Now we have our 32-bit value zero extended in the low element of
4370 // a vector. If Idx != 0, swizzle it into place.
4372 SmallVector<int, 4> Mask;
4373 Mask.push_back(Idx);
4374 for (unsigned i = 1; i != VecElts; ++i)
4376 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4377 DAG.getUNDEF(Item.getValueType()),
4380 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4384 // If we have a constant or non-constant insertion into the low element of
4385 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4386 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4387 // depending on what the source datatype is.
4390 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4391 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4392 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4393 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4394 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4395 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4397 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4398 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4399 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4400 EVT MiddleVT = MVT::v4i32;
4401 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4402 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4403 Subtarget->hasSSE2(), DAG);
4404 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4408 // Is it a vector logical left shift?
4409 if (NumElems == 2 && Idx == 1 &&
4410 X86::isZeroNode(Op.getOperand(0)) &&
4411 !X86::isZeroNode(Op.getOperand(1))) {
4412 unsigned NumBits = VT.getSizeInBits();
4413 return getVShift(true, VT,
4414 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4415 VT, Op.getOperand(1)),
4416 NumBits/2, DAG, *this, dl);
4419 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4422 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4423 // is a non-constant being inserted into an element other than the low one,
4424 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4425 // movd/movss) to move this into the low element, then shuffle it into
4427 if (EVTBits == 32) {
4428 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4430 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4431 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4432 Subtarget->hasSSE2(), DAG);
4433 SmallVector<int, 8> MaskVec;
4434 for (unsigned i = 0; i < NumElems; i++)
4435 MaskVec.push_back(i == Idx ? 0 : 1);
4436 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4440 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4441 if (Values.size() == 1) {
4442 if (EVTBits == 32) {
4443 // Instead of a shuffle like this:
4444 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4445 // Check if it's possible to issue this instead.
4446 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4447 unsigned Idx = CountTrailingZeros_32(NonZeros);
4448 SDValue Item = Op.getOperand(Idx);
4449 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4450 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4455 // A vector full of immediates; various special cases are already
4456 // handled, so this is best done with a single constant-pool load.
4460 // Let legalizer expand 2-wide build_vectors.
4461 if (EVTBits == 64) {
4462 if (NumNonZero == 1) {
4463 // One half is zero or undef.
4464 unsigned Idx = CountTrailingZeros_32(NonZeros);
4465 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4466 Op.getOperand(Idx));
4467 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4468 Subtarget->hasSSE2(), DAG);
4473 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4474 if (EVTBits == 8 && NumElems == 16) {
4475 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4477 if (V.getNode()) return V;
4480 if (EVTBits == 16 && NumElems == 8) {
4481 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4483 if (V.getNode()) return V;
4486 // If element VT is == 32 bits, turn it into a number of shuffles.
4487 SmallVector<SDValue, 8> V;
4489 if (NumElems == 4 && NumZero > 0) {
4490 for (unsigned i = 0; i < 4; ++i) {
4491 bool isZero = !(NonZeros & (1 << i));
4493 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4495 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4498 for (unsigned i = 0; i < 2; ++i) {
4499 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4502 V[i] = V[i*2]; // Must be a zero vector.
4505 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4508 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4511 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4516 SmallVector<int, 8> MaskVec;
4517 bool Reverse = (NonZeros & 0x3) == 2;
4518 for (unsigned i = 0; i < 2; ++i)
4519 MaskVec.push_back(Reverse ? 1-i : i);
4520 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4521 for (unsigned i = 0; i < 2; ++i)
4522 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4523 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4526 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4527 // Check for a build vector of consecutive loads.
4528 for (unsigned i = 0; i < NumElems; ++i)
4529 V[i] = Op.getOperand(i);
4531 // Check for elements which are consecutive loads.
4532 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4536 // For SSE 4.1, use insertps to put the high elements into the low element.
4537 if (getSubtarget()->hasSSE41()) {
4539 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4540 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4542 Result = DAG.getUNDEF(VT);
4544 for (unsigned i = 1; i < NumElems; ++i) {
4545 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4546 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4547 Op.getOperand(i), DAG.getIntPtrConstant(i));
4552 // Otherwise, expand into a number of unpckl*, start by extending each of
4553 // our (non-undef) elements to the full vector width with the element in the
4554 // bottom slot of the vector (which generates no code for SSE).
4555 for (unsigned i = 0; i < NumElems; ++i) {
4556 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4557 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4559 V[i] = DAG.getUNDEF(VT);
4562 // Next, we iteratively mix elements, e.g. for v4f32:
4563 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4564 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4565 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4566 unsigned EltStride = NumElems >> 1;
4567 while (EltStride != 0) {
4568 for (unsigned i = 0; i < EltStride; ++i) {
4569 // If V[i+EltStride] is undef and this is the first round of mixing,
4570 // then it is safe to just drop this shuffle: V[i] is already in the
4571 // right place, the one element (since it's the first round) being
4572 // inserted as undef can be dropped. This isn't safe for successive
4573 // rounds because they will permute elements within both vectors.
4574 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4575 EltStride == NumElems/2)
4578 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
4588 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
4589 // We support concatenate two MMX registers and place them in a MMX
4590 // register. This is better than doing a stack convert.
4591 DebugLoc dl = Op.getDebugLoc();
4592 EVT ResVT = Op.getValueType();
4593 assert(Op.getNumOperands() == 2);
4594 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
4595 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
4597 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
4598 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4599 InVec = Op.getOperand(1);
4600 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4601 unsigned NumElts = ResVT.getVectorNumElements();
4602 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4603 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
4604 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
4606 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
4607 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
4608 Mask[0] = 0; Mask[1] = 2;
4609 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
4611 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
4614 // v8i16 shuffles - Prefer shuffles in the following order:
4615 // 1. [all] pshuflw, pshufhw, optional move
4616 // 2. [ssse3] 1 x pshufb
4617 // 3. [ssse3] 2 x pshufb + 1 x por
4618 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
4620 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
4621 SelectionDAG &DAG) const {
4622 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4623 SDValue V1 = SVOp->getOperand(0);
4624 SDValue V2 = SVOp->getOperand(1);
4625 DebugLoc dl = SVOp->getDebugLoc();
4626 SmallVector<int, 8> MaskVals;
4628 // Determine if more than 1 of the words in each of the low and high quadwords
4629 // of the result come from the same quadword of one of the two inputs. Undef
4630 // mask values count as coming from any quadword, for better codegen.
4631 SmallVector<unsigned, 4> LoQuad(4);
4632 SmallVector<unsigned, 4> HiQuad(4);
4633 BitVector InputQuads(4);
4634 for (unsigned i = 0; i < 8; ++i) {
4635 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
4636 int EltIdx = SVOp->getMaskElt(i);
4637 MaskVals.push_back(EltIdx);
4646 InputQuads.set(EltIdx / 4);
4649 int BestLoQuad = -1;
4650 unsigned MaxQuad = 1;
4651 for (unsigned i = 0; i < 4; ++i) {
4652 if (LoQuad[i] > MaxQuad) {
4654 MaxQuad = LoQuad[i];
4658 int BestHiQuad = -1;
4660 for (unsigned i = 0; i < 4; ++i) {
4661 if (HiQuad[i] > MaxQuad) {
4663 MaxQuad = HiQuad[i];
4667 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
4668 // of the two input vectors, shuffle them into one input vector so only a
4669 // single pshufb instruction is necessary. If There are more than 2 input
4670 // quads, disable the next transformation since it does not help SSSE3.
4671 bool V1Used = InputQuads[0] || InputQuads[1];
4672 bool V2Used = InputQuads[2] || InputQuads[3];
4673 if (Subtarget->hasSSSE3()) {
4674 if (InputQuads.count() == 2 && V1Used && V2Used) {
4675 BestLoQuad = InputQuads.find_first();
4676 BestHiQuad = InputQuads.find_next(BestLoQuad);
4678 if (InputQuads.count() > 2) {
4684 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
4685 // the shuffle mask. If a quad is scored as -1, that means that it contains
4686 // words from all 4 input quadwords.
4688 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
4689 SmallVector<int, 8> MaskV;
4690 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
4691 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
4692 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
4693 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
4694 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
4695 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
4697 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
4698 // source words for the shuffle, to aid later transformations.
4699 bool AllWordsInNewV = true;
4700 bool InOrder[2] = { true, true };
4701 for (unsigned i = 0; i != 8; ++i) {
4702 int idx = MaskVals[i];
4704 InOrder[i/4] = false;
4705 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
4707 AllWordsInNewV = false;
4711 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
4712 if (AllWordsInNewV) {
4713 for (int i = 0; i != 8; ++i) {
4714 int idx = MaskVals[i];
4717 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
4718 if ((idx != i) && idx < 4)
4720 if ((idx != i) && idx > 3)
4729 // If we've eliminated the use of V2, and the new mask is a pshuflw or
4730 // pshufhw, that's as cheap as it gets. Return the new shuffle.
4731 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
4732 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
4733 unsigned TargetMask = 0;
4734 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
4735 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
4736 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
4737 X86::getShufflePSHUFLWImmediate(NewV.getNode());
4738 V1 = NewV.getOperand(0);
4739 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
4743 // If we have SSSE3, and all words of the result are from 1 input vector,
4744 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4745 // is present, fall back to case 4.
4746 if (Subtarget->hasSSSE3()) {
4747 SmallVector<SDValue,16> pshufbMask;
4749 // If we have elements from both input vectors, set the high bit of the
4750 // shuffle mask element to zero out elements that come from V2 in the V1
4751 // mask, and elements that come from V1 in the V2 mask, so that the two
4752 // results can be OR'd together.
4753 bool TwoInputs = V1Used && V2Used;
4754 for (unsigned i = 0; i != 8; ++i) {
4755 int EltIdx = MaskVals[i] * 2;
4756 if (TwoInputs && (EltIdx >= 16)) {
4757 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4758 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4761 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4762 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4764 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
4765 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4766 DAG.getNode(ISD::BUILD_VECTOR, dl,
4767 MVT::v16i8, &pshufbMask[0], 16));
4769 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4771 // Calculate the shuffle mask for the second input, shuffle it, and
4772 // OR it with the first shuffled input.
4774 for (unsigned i = 0; i != 8; ++i) {
4775 int EltIdx = MaskVals[i] * 2;
4777 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4778 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4781 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4782 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4784 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
4785 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4786 DAG.getNode(ISD::BUILD_VECTOR, dl,
4787 MVT::v16i8, &pshufbMask[0], 16));
4788 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4789 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4792 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4793 // and update MaskVals with new element order.
4794 BitVector InOrder(8);
4795 if (BestLoQuad >= 0) {
4796 SmallVector<int, 8> MaskV;
4797 for (int i = 0; i != 4; ++i) {
4798 int idx = MaskVals[i];
4800 MaskV.push_back(-1);
4802 } else if ((idx / 4) == BestLoQuad) {
4803 MaskV.push_back(idx & 3);
4806 MaskV.push_back(-1);
4809 for (unsigned i = 4; i != 8; ++i)
4811 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4814 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4815 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
4817 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
4821 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4822 // and update MaskVals with the new element order.
4823 if (BestHiQuad >= 0) {
4824 SmallVector<int, 8> MaskV;
4825 for (unsigned i = 0; i != 4; ++i)
4827 for (unsigned i = 4; i != 8; ++i) {
4828 int idx = MaskVals[i];
4830 MaskV.push_back(-1);
4832 } else if ((idx / 4) == BestHiQuad) {
4833 MaskV.push_back((idx & 3) + 4);
4836 MaskV.push_back(-1);
4839 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4842 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
4843 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
4845 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
4849 // In case BestHi & BestLo were both -1, which means each quadword has a word
4850 // from each of the four input quadwords, calculate the InOrder bitvector now
4851 // before falling through to the insert/extract cleanup.
4852 if (BestLoQuad == -1 && BestHiQuad == -1) {
4854 for (int i = 0; i != 8; ++i)
4855 if (MaskVals[i] < 0 || MaskVals[i] == i)
4859 // The other elements are put in the right place using pextrw and pinsrw.
4860 for (unsigned i = 0; i != 8; ++i) {
4863 int EltIdx = MaskVals[i];
4866 SDValue ExtOp = (EltIdx < 8)
4867 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4868 DAG.getIntPtrConstant(EltIdx))
4869 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4870 DAG.getIntPtrConstant(EltIdx - 8));
4871 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4872 DAG.getIntPtrConstant(i));
4877 // v16i8 shuffles - Prefer shuffles in the following order:
4878 // 1. [ssse3] 1 x pshufb
4879 // 2. [ssse3] 2 x pshufb + 1 x por
4880 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4882 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4884 const X86TargetLowering &TLI) {
4885 SDValue V1 = SVOp->getOperand(0);
4886 SDValue V2 = SVOp->getOperand(1);
4887 DebugLoc dl = SVOp->getDebugLoc();
4888 SmallVector<int, 16> MaskVals;
4889 SVOp->getMask(MaskVals);
4891 // If we have SSSE3, case 1 is generated when all result bytes come from
4892 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4893 // present, fall back to case 3.
4894 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4897 for (unsigned i = 0; i < 16; ++i) {
4898 int EltIdx = MaskVals[i];
4907 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4908 if (TLI.getSubtarget()->hasSSSE3()) {
4909 SmallVector<SDValue,16> pshufbMask;
4911 // If all result elements are from one input vector, then only translate
4912 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4914 // Otherwise, we have elements from both input vectors, and must zero out
4915 // elements that come from V2 in the first mask, and V1 in the second mask
4916 // so that we can OR them together.
4917 bool TwoInputs = !(V1Only || V2Only);
4918 for (unsigned i = 0; i != 16; ++i) {
4919 int EltIdx = MaskVals[i];
4920 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4921 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4924 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4926 // If all the elements are from V2, assign it to V1 and return after
4927 // building the first pshufb.
4930 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4931 DAG.getNode(ISD::BUILD_VECTOR, dl,
4932 MVT::v16i8, &pshufbMask[0], 16));
4936 // Calculate the shuffle mask for the second input, shuffle it, and
4937 // OR it with the first shuffled input.
4939 for (unsigned i = 0; i != 16; ++i) {
4940 int EltIdx = MaskVals[i];
4942 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4945 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4947 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4948 DAG.getNode(ISD::BUILD_VECTOR, dl,
4949 MVT::v16i8, &pshufbMask[0], 16));
4950 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4953 // No SSSE3 - Calculate in place words and then fix all out of place words
4954 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4955 // the 16 different words that comprise the two doublequadword input vectors.
4956 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
4957 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
4958 SDValue NewV = V2Only ? V2 : V1;
4959 for (int i = 0; i != 8; ++i) {
4960 int Elt0 = MaskVals[i*2];
4961 int Elt1 = MaskVals[i*2+1];
4963 // This word of the result is all undef, skip it.
4964 if (Elt0 < 0 && Elt1 < 0)
4967 // This word of the result is already in the correct place, skip it.
4968 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4970 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4973 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4974 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4977 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4978 // using a single extract together, load it and store it.
4979 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4980 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4981 DAG.getIntPtrConstant(Elt1 / 2));
4982 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4983 DAG.getIntPtrConstant(i));
4987 // If Elt1 is defined, extract it from the appropriate source. If the
4988 // source byte is not also odd, shift the extracted word left 8 bits
4989 // otherwise clear the bottom 8 bits if we need to do an or.
4991 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4992 DAG.getIntPtrConstant(Elt1 / 2));
4993 if ((Elt1 & 1) == 0)
4994 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4995 DAG.getConstant(8, TLI.getShiftAmountTy()));
4997 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4998 DAG.getConstant(0xFF00, MVT::i16));
5000 // If Elt0 is defined, extract it from the appropriate source. If the
5001 // source byte is not also even, shift the extracted word right 8 bits. If
5002 // Elt1 was also defined, OR the extracted values together before
5003 // inserting them in the result.
5005 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5006 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5007 if ((Elt0 & 1) != 0)
5008 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5009 DAG.getConstant(8, TLI.getShiftAmountTy()));
5011 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5012 DAG.getConstant(0x00FF, MVT::i16));
5013 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5016 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5017 DAG.getIntPtrConstant(i));
5019 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5022 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5023 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5024 /// done when every pair / quad of shuffle mask elements point to elements in
5025 /// the right sequence. e.g.
5026 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5028 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5029 SelectionDAG &DAG, DebugLoc dl) {
5030 EVT VT = SVOp->getValueType(0);
5031 SDValue V1 = SVOp->getOperand(0);
5032 SDValue V2 = SVOp->getOperand(1);
5033 unsigned NumElems = VT.getVectorNumElements();
5034 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5036 switch (VT.getSimpleVT().SimpleTy) {
5037 default: assert(false && "Unexpected!");
5038 case MVT::v4f32: NewVT = MVT::v2f64; break;
5039 case MVT::v4i32: NewVT = MVT::v2i64; break;
5040 case MVT::v8i16: NewVT = MVT::v4i32; break;
5041 case MVT::v16i8: NewVT = MVT::v4i32; break;
5044 int Scale = NumElems / NewWidth;
5045 SmallVector<int, 8> MaskVec;
5046 for (unsigned i = 0; i < NumElems; i += Scale) {
5048 for (int j = 0; j < Scale; ++j) {
5049 int EltIdx = SVOp->getMaskElt(i+j);
5053 StartIdx = EltIdx - (EltIdx % Scale);
5054 if (EltIdx != StartIdx + j)
5058 MaskVec.push_back(-1);
5060 MaskVec.push_back(StartIdx / Scale);
5063 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5064 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5065 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5068 /// getVZextMovL - Return a zero-extending vector move low node.
5070 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5071 SDValue SrcOp, SelectionDAG &DAG,
5072 const X86Subtarget *Subtarget, DebugLoc dl) {
5073 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5074 LoadSDNode *LD = NULL;
5075 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5076 LD = dyn_cast<LoadSDNode>(SrcOp);
5078 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5080 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5081 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5082 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5083 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5084 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5086 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5087 return DAG.getNode(ISD::BITCAST, dl, VT,
5088 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5089 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5097 return DAG.getNode(ISD::BITCAST, dl, VT,
5098 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5099 DAG.getNode(ISD::BITCAST, dl,
5103 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
5106 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5107 SDValue V1 = SVOp->getOperand(0);
5108 SDValue V2 = SVOp->getOperand(1);
5109 DebugLoc dl = SVOp->getDebugLoc();
5110 EVT VT = SVOp->getValueType(0);
5112 SmallVector<std::pair<int, int>, 8> Locs;
5114 SmallVector<int, 8> Mask1(4U, -1);
5115 SmallVector<int, 8> PermMask;
5116 SVOp->getMask(PermMask);
5120 for (unsigned i = 0; i != 4; ++i) {
5121 int Idx = PermMask[i];
5123 Locs[i] = std::make_pair(-1, -1);
5125 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5127 Locs[i] = std::make_pair(0, NumLo);
5131 Locs[i] = std::make_pair(1, NumHi);
5133 Mask1[2+NumHi] = Idx;
5139 if (NumLo <= 2 && NumHi <= 2) {
5140 // If no more than two elements come from either vector. This can be
5141 // implemented with two shuffles. First shuffle gather the elements.
5142 // The second shuffle, which takes the first shuffle as both of its
5143 // vector operands, put the elements into the right order.
5144 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5146 SmallVector<int, 8> Mask2(4U, -1);
5148 for (unsigned i = 0; i != 4; ++i) {
5149 if (Locs[i].first == -1)
5152 unsigned Idx = (i < 2) ? 0 : 4;
5153 Idx += Locs[i].first * 2 + Locs[i].second;
5158 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5159 } else if (NumLo == 3 || NumHi == 3) {
5160 // Otherwise, we must have three elements from one vector, call it X, and
5161 // one element from the other, call it Y. First, use a shufps to build an
5162 // intermediate vector with the one element from Y and the element from X
5163 // that will be in the same half in the final destination (the indexes don't
5164 // matter). Then, use a shufps to build the final vector, taking the half
5165 // containing the element from Y from the intermediate, and the other half
5168 // Normalize it so the 3 elements come from V1.
5169 CommuteVectorShuffleMask(PermMask, VT);
5173 // Find the element from V2.
5175 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5176 int Val = PermMask[HiIndex];
5183 Mask1[0] = PermMask[HiIndex];
5185 Mask1[2] = PermMask[HiIndex^1];
5187 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5190 Mask1[0] = PermMask[0];
5191 Mask1[1] = PermMask[1];
5192 Mask1[2] = HiIndex & 1 ? 6 : 4;
5193 Mask1[3] = HiIndex & 1 ? 4 : 6;
5194 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5196 Mask1[0] = HiIndex & 1 ? 2 : 0;
5197 Mask1[1] = HiIndex & 1 ? 0 : 2;
5198 Mask1[2] = PermMask[2];
5199 Mask1[3] = PermMask[3];
5204 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5208 // Break it into (shuffle shuffle_hi, shuffle_lo).
5210 SmallVector<int,8> LoMask(4U, -1);
5211 SmallVector<int,8> HiMask(4U, -1);
5213 SmallVector<int,8> *MaskPtr = &LoMask;
5214 unsigned MaskIdx = 0;
5217 for (unsigned i = 0; i != 4; ++i) {
5224 int Idx = PermMask[i];
5226 Locs[i] = std::make_pair(-1, -1);
5227 } else if (Idx < 4) {
5228 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5229 (*MaskPtr)[LoIdx] = Idx;
5232 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5233 (*MaskPtr)[HiIdx] = Idx;
5238 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5239 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5240 SmallVector<int, 8> MaskOps;
5241 for (unsigned i = 0; i != 4; ++i) {
5242 if (Locs[i].first == -1) {
5243 MaskOps.push_back(-1);
5245 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5246 MaskOps.push_back(Idx);
5249 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5252 static bool MayFoldVectorLoad(SDValue V) {
5253 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5254 V = V.getOperand(0);
5255 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5256 V = V.getOperand(0);
5262 // FIXME: the version above should always be used. Since there's
5263 // a bug where several vector shuffles can't be folded because the
5264 // DAG is not updated during lowering and a node claims to have two
5265 // uses while it only has one, use this version, and let isel match
5266 // another instruction if the load really happens to have more than
5267 // one use. Remove this version after this bug get fixed.
5268 // rdar://8434668, PR8156
5269 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5270 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5271 V = V.getOperand(0);
5272 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5273 V = V.getOperand(0);
5274 if (ISD::isNormalLoad(V.getNode()))
5279 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5280 /// a vector extract, and if both can be later optimized into a single load.
5281 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5282 /// here because otherwise a target specific shuffle node is going to be
5283 /// emitted for this shuffle, and the optimization not done.
5284 /// FIXME: This is probably not the best approach, but fix the problem
5285 /// until the right path is decided.
5287 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5288 const TargetLowering &TLI) {
5289 EVT VT = V.getValueType();
5290 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5292 // Be sure that the vector shuffle is present in a pattern like this:
5293 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5297 SDNode *N = *V.getNode()->use_begin();
5298 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5301 SDValue EltNo = N->getOperand(1);
5302 if (!isa<ConstantSDNode>(EltNo))
5305 // If the bit convert changed the number of elements, it is unsafe
5306 // to examine the mask.
5307 bool HasShuffleIntoBitcast = false;
5308 if (V.getOpcode() == ISD::BITCAST) {
5309 EVT SrcVT = V.getOperand(0).getValueType();
5310 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5312 V = V.getOperand(0);
5313 HasShuffleIntoBitcast = true;
5316 // Select the input vector, guarding against out of range extract vector.
5317 unsigned NumElems = VT.getVectorNumElements();
5318 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5319 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5320 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5322 // Skip one more bit_convert if necessary
5323 if (V.getOpcode() == ISD::BITCAST)
5324 V = V.getOperand(0);
5326 if (ISD::isNormalLoad(V.getNode())) {
5327 // Is the original load suitable?
5328 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5330 // FIXME: avoid the multi-use bug that is preventing lots of
5331 // of foldings to be detected, this is still wrong of course, but
5332 // give the temporary desired behavior, and if it happens that
5333 // the load has real more uses, during isel it will not fold, and
5334 // will generate poor code.
5335 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5338 if (!HasShuffleIntoBitcast)
5341 // If there's a bitcast before the shuffle, check if the load type and
5342 // alignment is valid.
5343 unsigned Align = LN0->getAlignment();
5345 TLI.getTargetData()->getABITypeAlignment(
5346 VT.getTypeForEVT(*DAG.getContext()));
5348 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5356 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5357 EVT VT = Op.getValueType();
5359 // Canonizalize to v2f64.
5360 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5361 return DAG.getNode(ISD::BITCAST, dl, VT,
5362 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5367 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5369 SDValue V1 = Op.getOperand(0);
5370 SDValue V2 = Op.getOperand(1);
5371 EVT VT = Op.getValueType();
5373 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5375 if (HasSSE2 && VT == MVT::v2f64)
5376 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5379 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5383 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5384 SDValue V1 = Op.getOperand(0);
5385 SDValue V2 = Op.getOperand(1);
5386 EVT VT = Op.getValueType();
5388 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5389 "unsupported shuffle type");
5391 if (V2.getOpcode() == ISD::UNDEF)
5395 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5399 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5400 SDValue V1 = Op.getOperand(0);
5401 SDValue V2 = Op.getOperand(1);
5402 EVT VT = Op.getValueType();
5403 unsigned NumElems = VT.getVectorNumElements();
5405 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5406 // operand of these instructions is only memory, so check if there's a
5407 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5409 bool CanFoldLoad = false;
5411 // Trivial case, when V2 comes from a load.
5412 if (MayFoldVectorLoad(V2))
5415 // When V1 is a load, it can be folded later into a store in isel, example:
5416 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5418 // (MOVLPSmr addr:$src1, VR128:$src2)
5419 // So, recognize this potential and also use MOVLPS or MOVLPD
5420 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5424 if (HasSSE2 && NumElems == 2)
5425 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5428 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5431 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5432 // movl and movlp will both match v2i64, but v2i64 is never matched by
5433 // movl earlier because we make it strict to avoid messing with the movlp load
5434 // folding logic (see the code above getMOVLP call). Match it here then,
5435 // this is horrible, but will stay like this until we move all shuffle
5436 // matching to x86 specific nodes. Note that for the 1st condition all
5437 // types are matched with movsd.
5438 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5439 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5441 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5444 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5446 // Invert the operand order and use SHUFPS to match it.
5447 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5448 X86::getShuffleSHUFImmediate(SVOp), DAG);
5451 static inline unsigned getUNPCKLOpcode(EVT VT) {
5452 switch(VT.getSimpleVT().SimpleTy) {
5453 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5454 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5455 case MVT::v4f32: return X86ISD::UNPCKLPS;
5456 case MVT::v2f64: return X86ISD::UNPCKLPD;
5457 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5458 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5460 llvm_unreachable("Unknow type for unpckl");
5465 static inline unsigned getUNPCKHOpcode(EVT VT) {
5466 switch(VT.getSimpleVT().SimpleTy) {
5467 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5468 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5469 case MVT::v4f32: return X86ISD::UNPCKHPS;
5470 case MVT::v2f64: return X86ISD::UNPCKHPD;
5471 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5472 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5474 llvm_unreachable("Unknow type for unpckh");
5480 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5481 const TargetLowering &TLI,
5482 const X86Subtarget *Subtarget) {
5483 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5484 EVT VT = Op.getValueType();
5485 DebugLoc dl = Op.getDebugLoc();
5486 SDValue V1 = Op.getOperand(0);
5487 SDValue V2 = Op.getOperand(1);
5489 if (isZeroShuffle(SVOp))
5490 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5492 // Handle splat operations
5493 if (SVOp->isSplat()) {
5494 // Special case, this is the only place now where it's
5495 // allowed to return a vector_shuffle operation without
5496 // using a target specific node, because *hopefully* it
5497 // will be optimized away by the dag combiner.
5498 if (VT.getVectorNumElements() <= 4 &&
5499 CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5502 // Handle splats by matching through known masks
5503 if (VT.getVectorNumElements() <= 4)
5506 // Canonicalize all of the remaining to v4f32.
5507 return PromoteSplat(SVOp, DAG);
5510 // If the shuffle can be profitably rewritten as a narrower shuffle, then
5512 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
5513 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5514 if (NewOp.getNode())
5515 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
5516 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
5517 // FIXME: Figure out a cleaner way to do this.
5518 // Try to make use of movq to zero out the top part.
5519 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
5520 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5521 if (NewOp.getNode()) {
5522 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
5523 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
5524 DAG, Subtarget, dl);
5526 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
5527 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
5528 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
5529 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
5530 DAG, Subtarget, dl);
5537 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
5538 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5539 SDValue V1 = Op.getOperand(0);
5540 SDValue V2 = Op.getOperand(1);
5541 EVT VT = Op.getValueType();
5542 DebugLoc dl = Op.getDebugLoc();
5543 unsigned NumElems = VT.getVectorNumElements();
5544 bool isMMX = VT.getSizeInBits() == 64;
5545 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
5546 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
5547 bool V1IsSplat = false;
5548 bool V2IsSplat = false;
5549 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
5550 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
5551 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
5552 MachineFunction &MF = DAG.getMachineFunction();
5553 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
5555 // Shuffle operations on MMX not supported.
5559 // Vector shuffle lowering takes 3 steps:
5561 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
5562 // narrowing and commutation of operands should be handled.
5563 // 2) Matching of shuffles with known shuffle masks to x86 target specific
5565 // 3) Rewriting of unmatched masks into new generic shuffle operations,
5566 // so the shuffle can be broken into other shuffles and the legalizer can
5567 // try the lowering again.
5569 // The general ideia is that no vector_shuffle operation should be left to
5570 // be matched during isel, all of them must be converted to a target specific
5573 // Normalize the input vectors. Here splats, zeroed vectors, profitable
5574 // narrowing and commutation of operands should be handled. The actual code
5575 // doesn't include all of those, work in progress...
5576 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
5577 if (NewOp.getNode())
5580 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
5581 // unpckh_undef). Only use pshufd if speed is more important than size.
5582 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
5583 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5584 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5585 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
5586 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5587 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5589 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
5590 RelaxedMayFoldVectorLoad(V1))
5591 return getMOVDDup(Op, dl, V1, DAG);
5593 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
5594 return getMOVHighToLow(Op, dl, DAG);
5596 // Use to match splats
5597 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
5598 (VT == MVT::v2f64 || VT == MVT::v2i64))
5599 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5601 if (X86::isPSHUFDMask(SVOp)) {
5602 // The actual implementation will match the mask in the if above and then
5603 // during isel it can match several different instructions, not only pshufd
5604 // as its name says, sad but true, emulate the behavior for now...
5605 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
5606 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
5608 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5610 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
5611 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
5613 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5614 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
5617 if (VT == MVT::v4f32)
5618 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
5622 // Check if this can be converted into a logical shift.
5623 bool isLeft = false;
5626 bool isShift = getSubtarget()->hasSSE2() &&
5627 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
5628 if (isShift && ShVal.hasOneUse()) {
5629 // If the shifted value has multiple uses, it may be cheaper to use
5630 // v_set0 + movlhps or movhlps, etc.
5631 EVT EltVT = VT.getVectorElementType();
5632 ShAmt *= EltVT.getSizeInBits();
5633 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5636 if (X86::isMOVLMask(SVOp)) {
5639 if (ISD::isBuildVectorAllZeros(V1.getNode()))
5640 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
5641 if (!X86::isMOVLPMask(SVOp)) {
5642 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
5643 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5645 if (VT == MVT::v4i32 || VT == MVT::v4f32)
5646 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5650 // FIXME: fold these into legal mask.
5651 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
5652 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
5654 if (X86::isMOVHLPSMask(SVOp))
5655 return getMOVHighToLow(Op, dl, DAG);
5657 if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5658 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
5660 if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
5661 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
5663 if (X86::isMOVLPMask(SVOp))
5664 return getMOVLP(Op, dl, DAG, HasSSE2);
5666 if (ShouldXformToMOVHLPS(SVOp) ||
5667 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
5668 return CommuteVectorShuffle(SVOp, DAG);
5671 // No better options. Use a vshl / vsrl.
5672 EVT EltVT = VT.getVectorElementType();
5673 ShAmt *= EltVT.getSizeInBits();
5674 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
5677 bool Commuted = false;
5678 // FIXME: This should also accept a bitcast of a splat? Be careful, not
5679 // 1,1,1,1 -> v8i16 though.
5680 V1IsSplat = isSplatVector(V1.getNode());
5681 V2IsSplat = isSplatVector(V2.getNode());
5683 // Canonicalize the splat or undef, if present, to be on the RHS.
5684 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
5685 Op = CommuteVectorShuffle(SVOp, DAG);
5686 SVOp = cast<ShuffleVectorSDNode>(Op);
5687 V1 = SVOp->getOperand(0);
5688 V2 = SVOp->getOperand(1);
5689 std::swap(V1IsSplat, V2IsSplat);
5690 std::swap(V1IsUndef, V2IsUndef);
5694 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
5695 // Shuffling low element of v1 into undef, just return v1.
5698 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
5699 // the instruction selector will not match, so get a canonical MOVL with
5700 // swapped operands to undo the commute.
5701 return getMOVL(DAG, dl, VT, V2, V1);
5704 if (X86::isUNPCKLMask(SVOp))
5705 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
5707 if (X86::isUNPCKHMask(SVOp))
5708 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
5711 // Normalize mask so all entries that point to V2 points to its first
5712 // element then try to match unpck{h|l} again. If match, return a
5713 // new vector_shuffle with the corrected mask.
5714 SDValue NewMask = NormalizeMask(SVOp, DAG);
5715 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
5716 if (NSVOp != SVOp) {
5717 if (X86::isUNPCKLMask(NSVOp, true)) {
5719 } else if (X86::isUNPCKHMask(NSVOp, true)) {
5726 // Commute is back and try unpck* again.
5727 // FIXME: this seems wrong.
5728 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
5729 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
5731 if (X86::isUNPCKLMask(NewSVOp))
5732 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
5734 if (X86::isUNPCKHMask(NewSVOp))
5735 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
5738 // Normalize the node to match x86 shuffle ops if needed
5739 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
5740 return CommuteVectorShuffle(SVOp, DAG);
5742 // The checks below are all present in isShuffleMaskLegal, but they are
5743 // inlined here right now to enable us to directly emit target specific
5744 // nodes, and remove one by one until they don't return Op anymore.
5745 SmallVector<int, 16> M;
5748 if (isPALIGNRMask(M, VT, HasSSSE3))
5749 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
5750 X86::getShufflePALIGNRImmediate(SVOp),
5753 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
5754 SVOp->getSplatIndex() == 0 && V2IsUndef) {
5755 if (VT == MVT::v2f64)
5756 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
5757 if (VT == MVT::v2i64)
5758 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
5761 if (isPSHUFHWMask(M, VT))
5762 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
5763 X86::getShufflePSHUFHWImmediate(SVOp),
5766 if (isPSHUFLWMask(M, VT))
5767 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
5768 X86::getShufflePSHUFLWImmediate(SVOp),
5771 if (isSHUFPMask(M, VT)) {
5772 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
5773 if (VT == MVT::v4f32 || VT == MVT::v4i32)
5774 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
5776 if (VT == MVT::v2f64 || VT == MVT::v2i64)
5777 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
5781 if (X86::isUNPCKL_v_undef_Mask(SVOp))
5782 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5783 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
5784 if (X86::isUNPCKH_v_undef_Mask(SVOp))
5785 if (VT != MVT::v2i64 && VT != MVT::v2f64)
5786 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
5788 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
5789 if (VT == MVT::v8i16) {
5790 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
5791 if (NewOp.getNode())
5795 if (VT == MVT::v16i8) {
5796 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
5797 if (NewOp.getNode())
5801 // Handle all 4 wide cases with a number of shuffles.
5803 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
5809 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
5810 SelectionDAG &DAG) const {
5811 EVT VT = Op.getValueType();
5812 DebugLoc dl = Op.getDebugLoc();
5813 if (VT.getSizeInBits() == 8) {
5814 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
5815 Op.getOperand(0), Op.getOperand(1));
5816 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5817 DAG.getValueType(VT));
5818 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5819 } else if (VT.getSizeInBits() == 16) {
5820 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5821 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
5823 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5824 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5825 DAG.getNode(ISD::BITCAST, dl,
5829 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
5830 Op.getOperand(0), Op.getOperand(1));
5831 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
5832 DAG.getValueType(VT));
5833 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5834 } else if (VT == MVT::f32) {
5835 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
5836 // the result back to FR32 register. It's only worth matching if the
5837 // result has a single use which is a store or a bitcast to i32. And in
5838 // the case of a store, it's not worth it if the index is a constant 0,
5839 // because a MOVSSmr can be used instead, which is smaller and faster.
5840 if (!Op.hasOneUse())
5842 SDNode *User = *Op.getNode()->use_begin();
5843 if ((User->getOpcode() != ISD::STORE ||
5844 (isa<ConstantSDNode>(Op.getOperand(1)) &&
5845 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
5846 (User->getOpcode() != ISD::BITCAST ||
5847 User->getValueType(0) != MVT::i32))
5849 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5850 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
5853 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
5854 } else if (VT == MVT::i32) {
5855 // ExtractPS works with constant index.
5856 if (isa<ConstantSDNode>(Op.getOperand(1)))
5864 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
5865 SelectionDAG &DAG) const {
5866 if (!isa<ConstantSDNode>(Op.getOperand(1)))
5869 if (Subtarget->hasSSE41()) {
5870 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
5875 EVT VT = Op.getValueType();
5876 DebugLoc dl = Op.getDebugLoc();
5877 // TODO: handle v16i8.
5878 if (VT.getSizeInBits() == 16) {
5879 SDValue Vec = Op.getOperand(0);
5880 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5882 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
5883 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
5884 DAG.getNode(ISD::BITCAST, dl,
5887 // Transform it so it match pextrw which produces a 32-bit result.
5888 EVT EltVT = MVT::i32;
5889 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
5890 Op.getOperand(0), Op.getOperand(1));
5891 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
5892 DAG.getValueType(VT));
5893 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
5894 } else if (VT.getSizeInBits() == 32) {
5895 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5899 // SHUFPS the element to the lowest double word, then movss.
5900 int Mask[4] = { Idx, -1, -1, -1 };
5901 EVT VVT = Op.getOperand(0).getValueType();
5902 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5903 DAG.getUNDEF(VVT), Mask);
5904 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5905 DAG.getIntPtrConstant(0));
5906 } else if (VT.getSizeInBits() == 64) {
5907 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
5908 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
5909 // to match extract_elt for f64.
5910 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
5914 // UNPCKHPD the element to the lowest double word, then movsd.
5915 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
5916 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
5917 int Mask[2] = { 1, -1 };
5918 EVT VVT = Op.getOperand(0).getValueType();
5919 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
5920 DAG.getUNDEF(VVT), Mask);
5921 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
5922 DAG.getIntPtrConstant(0));
5929 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
5930 SelectionDAG &DAG) const {
5931 EVT VT = Op.getValueType();
5932 EVT EltVT = VT.getVectorElementType();
5933 DebugLoc dl = Op.getDebugLoc();
5935 SDValue N0 = Op.getOperand(0);
5936 SDValue N1 = Op.getOperand(1);
5937 SDValue N2 = Op.getOperand(2);
5939 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
5940 isa<ConstantSDNode>(N2)) {
5942 if (VT == MVT::v8i16)
5943 Opc = X86ISD::PINSRW;
5944 else if (VT == MVT::v16i8)
5945 Opc = X86ISD::PINSRB;
5947 Opc = X86ISD::PINSRB;
5949 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
5951 if (N1.getValueType() != MVT::i32)
5952 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5953 if (N2.getValueType() != MVT::i32)
5954 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5955 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
5956 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
5957 // Bits [7:6] of the constant are the source select. This will always be
5958 // zero here. The DAG Combiner may combine an extract_elt index into these
5959 // bits. For example (insert (extract, 3), 2) could be matched by putting
5960 // the '3' into bits [7:6] of X86ISD::INSERTPS.
5961 // Bits [5:4] of the constant are the destination select. This is the
5962 // value of the incoming immediate.
5963 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
5964 // combine either bitwise AND or insert of float 0.0 to set these bits.
5965 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
5966 // Create this as a scalar to vector..
5967 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
5968 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
5969 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
5970 // PINSR* works with constant index.
5977 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
5978 EVT VT = Op.getValueType();
5979 EVT EltVT = VT.getVectorElementType();
5981 if (Subtarget->hasSSE41())
5982 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
5984 if (EltVT == MVT::i8)
5987 DebugLoc dl = Op.getDebugLoc();
5988 SDValue N0 = Op.getOperand(0);
5989 SDValue N1 = Op.getOperand(1);
5990 SDValue N2 = Op.getOperand(2);
5992 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
5993 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
5994 // as its second argument.
5995 if (N1.getValueType() != MVT::i32)
5996 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
5997 if (N2.getValueType() != MVT::i32)
5998 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
5999 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6005 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6006 DebugLoc dl = Op.getDebugLoc();
6008 if (Op.getValueType() == MVT::v1i64 &&
6009 Op.getOperand(0).getValueType() == MVT::i64)
6010 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6012 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6013 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6014 "Expected an SSE type!");
6015 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6016 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6019 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6020 // a simple subregister reference or explicit instructions to grab
6021 // upper bits of a vector.
6023 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6024 if (Subtarget->hasAVX()) {
6030 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6031 // simple superregister reference or explicit instructions to insert
6032 // the upper bits of a vector.
6034 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6035 if (Subtarget->hasAVX()) {
6036 DebugLoc dl = Op.getNode()->getDebugLoc();
6037 SDValue Vec = Op.getNode()->getOperand(0);
6038 SDValue SubVec = Op.getNode()->getOperand(1);
6039 SDValue Idx = Op.getNode()->getOperand(2);
6041 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6042 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6049 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6050 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6051 // one of the above mentioned nodes. It has to be wrapped because otherwise
6052 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6053 // be used to form addressing mode. These wrapped nodes will be selected
6056 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6057 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6059 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6061 unsigned char OpFlag = 0;
6062 unsigned WrapperKind = X86ISD::Wrapper;
6063 CodeModel::Model M = getTargetMachine().getCodeModel();
6065 if (Subtarget->isPICStyleRIPRel() &&
6066 (M == CodeModel::Small || M == CodeModel::Kernel))
6067 WrapperKind = X86ISD::WrapperRIP;
6068 else if (Subtarget->isPICStyleGOT())
6069 OpFlag = X86II::MO_GOTOFF;
6070 else if (Subtarget->isPICStyleStubPIC())
6071 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6073 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6075 CP->getOffset(), OpFlag);
6076 DebugLoc DL = CP->getDebugLoc();
6077 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6078 // With PIC, the address is actually $g + Offset.
6080 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6081 DAG.getNode(X86ISD::GlobalBaseReg,
6082 DebugLoc(), getPointerTy()),
6089 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6090 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6092 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6094 unsigned char OpFlag = 0;
6095 unsigned WrapperKind = X86ISD::Wrapper;
6096 CodeModel::Model M = getTargetMachine().getCodeModel();
6098 if (Subtarget->isPICStyleRIPRel() &&
6099 (M == CodeModel::Small || M == CodeModel::Kernel))
6100 WrapperKind = X86ISD::WrapperRIP;
6101 else if (Subtarget->isPICStyleGOT())
6102 OpFlag = X86II::MO_GOTOFF;
6103 else if (Subtarget->isPICStyleStubPIC())
6104 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6106 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6108 DebugLoc DL = JT->getDebugLoc();
6109 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6111 // With PIC, the address is actually $g + Offset.
6113 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6114 DAG.getNode(X86ISD::GlobalBaseReg,
6115 DebugLoc(), getPointerTy()),
6122 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6123 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6125 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6127 unsigned char OpFlag = 0;
6128 unsigned WrapperKind = X86ISD::Wrapper;
6129 CodeModel::Model M = getTargetMachine().getCodeModel();
6131 if (Subtarget->isPICStyleRIPRel() &&
6132 (M == CodeModel::Small || M == CodeModel::Kernel))
6133 WrapperKind = X86ISD::WrapperRIP;
6134 else if (Subtarget->isPICStyleGOT())
6135 OpFlag = X86II::MO_GOTOFF;
6136 else if (Subtarget->isPICStyleStubPIC())
6137 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6139 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6141 DebugLoc DL = Op.getDebugLoc();
6142 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6145 // With PIC, the address is actually $g + Offset.
6146 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6147 !Subtarget->is64Bit()) {
6148 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6149 DAG.getNode(X86ISD::GlobalBaseReg,
6150 DebugLoc(), getPointerTy()),
6158 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6159 // Create the TargetBlockAddressAddress node.
6160 unsigned char OpFlags =
6161 Subtarget->ClassifyBlockAddressReference();
6162 CodeModel::Model M = getTargetMachine().getCodeModel();
6163 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6164 DebugLoc dl = Op.getDebugLoc();
6165 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6166 /*isTarget=*/true, OpFlags);
6168 if (Subtarget->isPICStyleRIPRel() &&
6169 (M == CodeModel::Small || M == CodeModel::Kernel))
6170 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6172 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6174 // With PIC, the address is actually $g + Offset.
6175 if (isGlobalRelativeToPICBase(OpFlags)) {
6176 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6177 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6185 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6187 SelectionDAG &DAG) const {
6188 // Create the TargetGlobalAddress node, folding in the constant
6189 // offset if it is legal.
6190 unsigned char OpFlags =
6191 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6192 CodeModel::Model M = getTargetMachine().getCodeModel();
6194 if (OpFlags == X86II::MO_NO_FLAG &&
6195 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6196 // A direct static reference to a global.
6197 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6200 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6203 if (Subtarget->isPICStyleRIPRel() &&
6204 (M == CodeModel::Small || M == CodeModel::Kernel))
6205 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6207 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6209 // With PIC, the address is actually $g + Offset.
6210 if (isGlobalRelativeToPICBase(OpFlags)) {
6211 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6212 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6216 // For globals that require a load from a stub to get the address, emit the
6218 if (isGlobalStubReference(OpFlags))
6219 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6220 MachinePointerInfo::getGOT(), false, false, 0);
6222 // If there was a non-zero offset that we didn't fold, create an explicit
6225 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6226 DAG.getConstant(Offset, getPointerTy()));
6232 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6233 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6234 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6235 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6239 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6240 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6241 unsigned char OperandFlags) {
6242 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6243 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6244 DebugLoc dl = GA->getDebugLoc();
6245 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6246 GA->getValueType(0),
6250 SDValue Ops[] = { Chain, TGA, *InFlag };
6251 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6253 SDValue Ops[] = { Chain, TGA };
6254 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6257 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6258 MFI->setAdjustsStack(true);
6260 SDValue Flag = Chain.getValue(1);
6261 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6264 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6266 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6269 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6270 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6271 DAG.getNode(X86ISD::GlobalBaseReg,
6272 DebugLoc(), PtrVT), InFlag);
6273 InFlag = Chain.getValue(1);
6275 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6278 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6280 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6282 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6283 X86::RAX, X86II::MO_TLSGD);
6286 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6287 // "local exec" model.
6288 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6289 const EVT PtrVT, TLSModel::Model model,
6291 DebugLoc dl = GA->getDebugLoc();
6293 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6294 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6295 is64Bit ? 257 : 256));
6297 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6298 DAG.getIntPtrConstant(0),
6299 MachinePointerInfo(Ptr), false, false, 0);
6301 unsigned char OperandFlags = 0;
6302 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6304 unsigned WrapperKind = X86ISD::Wrapper;
6305 if (model == TLSModel::LocalExec) {
6306 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6307 } else if (is64Bit) {
6308 assert(model == TLSModel::InitialExec);
6309 OperandFlags = X86II::MO_GOTTPOFF;
6310 WrapperKind = X86ISD::WrapperRIP;
6312 assert(model == TLSModel::InitialExec);
6313 OperandFlags = X86II::MO_INDNTPOFF;
6316 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6318 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6319 GA->getValueType(0),
6320 GA->getOffset(), OperandFlags);
6321 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6323 if (model == TLSModel::InitialExec)
6324 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6325 MachinePointerInfo::getGOT(), false, false, 0);
6327 // The address of the thread local variable is the add of the thread
6328 // pointer with the offset of the variable.
6329 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6333 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6335 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6336 const GlobalValue *GV = GA->getGlobal();
6338 if (Subtarget->isTargetELF()) {
6339 // TODO: implement the "local dynamic" model
6340 // TODO: implement the "initial exec"model for pic executables
6342 // If GV is an alias then use the aliasee for determining
6343 // thread-localness.
6344 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6345 GV = GA->resolveAliasedGlobal(false);
6347 TLSModel::Model model
6348 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6351 case TLSModel::GeneralDynamic:
6352 case TLSModel::LocalDynamic: // not implemented
6353 if (Subtarget->is64Bit())
6354 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6355 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6357 case TLSModel::InitialExec:
6358 case TLSModel::LocalExec:
6359 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6360 Subtarget->is64Bit());
6362 } else if (Subtarget->isTargetDarwin()) {
6363 // Darwin only has one model of TLS. Lower to that.
6364 unsigned char OpFlag = 0;
6365 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6366 X86ISD::WrapperRIP : X86ISD::Wrapper;
6368 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6370 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6371 !Subtarget->is64Bit();
6373 OpFlag = X86II::MO_TLVP_PIC_BASE;
6375 OpFlag = X86II::MO_TLVP;
6376 DebugLoc DL = Op.getDebugLoc();
6377 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6378 GA->getValueType(0),
6379 GA->getOffset(), OpFlag);
6380 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6382 // With PIC32, the address is actually $g + Offset.
6384 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6385 DAG.getNode(X86ISD::GlobalBaseReg,
6386 DebugLoc(), getPointerTy()),
6389 // Lowering the machine isd will make sure everything is in the right
6391 SDValue Chain = DAG.getEntryNode();
6392 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6393 SDValue Args[] = { Chain, Offset };
6394 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6396 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6397 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6398 MFI->setAdjustsStack(true);
6400 // And our return value (tls address) is in the standard call return value
6402 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
6403 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
6407 "TLS not implemented for this target.");
6409 llvm_unreachable("Unreachable");
6414 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
6415 /// take a 2 x i32 value to shift plus a shift amount.
6416 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
6417 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
6418 EVT VT = Op.getValueType();
6419 unsigned VTBits = VT.getSizeInBits();
6420 DebugLoc dl = Op.getDebugLoc();
6421 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
6422 SDValue ShOpLo = Op.getOperand(0);
6423 SDValue ShOpHi = Op.getOperand(1);
6424 SDValue ShAmt = Op.getOperand(2);
6425 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
6426 DAG.getConstant(VTBits - 1, MVT::i8))
6427 : DAG.getConstant(0, VT);
6430 if (Op.getOpcode() == ISD::SHL_PARTS) {
6431 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
6432 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
6434 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
6435 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
6438 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
6439 DAG.getConstant(VTBits, MVT::i8));
6440 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
6441 AndNode, DAG.getConstant(0, MVT::i8));
6444 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6445 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
6446 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
6448 if (Op.getOpcode() == ISD::SHL_PARTS) {
6449 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6450 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6452 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
6453 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
6456 SDValue Ops[2] = { Lo, Hi };
6457 return DAG.getMergeValues(Ops, 2, dl);
6460 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
6461 SelectionDAG &DAG) const {
6462 EVT SrcVT = Op.getOperand(0).getValueType();
6464 if (SrcVT.isVector())
6467 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
6468 "Unknown SINT_TO_FP to lower!");
6470 // These are really Legal; return the operand so the caller accepts it as
6472 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
6474 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
6475 Subtarget->is64Bit()) {
6479 DebugLoc dl = Op.getDebugLoc();
6480 unsigned Size = SrcVT.getSizeInBits()/8;
6481 MachineFunction &MF = DAG.getMachineFunction();
6482 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
6483 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6484 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6486 MachinePointerInfo::getFixedStack(SSFI),
6488 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
6491 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
6493 SelectionDAG &DAG) const {
6495 DebugLoc DL = Op.getDebugLoc();
6497 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
6499 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
6501 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
6503 unsigned ByteSize = SrcVT.getSizeInBits()/8;
6505 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6506 MachineMemOperand *MMO =
6507 DAG.getMachineFunction()
6508 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6509 MachineMemOperand::MOLoad, ByteSize, ByteSize);
6511 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
6512 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
6514 Tys, Ops, array_lengthof(Ops),
6518 Chain = Result.getValue(1);
6519 SDValue InFlag = Result.getValue(2);
6521 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
6522 // shouldn't be necessary except that RFP cannot be live across
6523 // multiple blocks. When stackifier is fixed, they can be uncoupled.
6524 MachineFunction &MF = DAG.getMachineFunction();
6525 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
6526 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
6527 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6528 Tys = DAG.getVTList(MVT::Other);
6530 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
6532 MachineMemOperand *MMO =
6533 DAG.getMachineFunction()
6534 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6535 MachineMemOperand::MOStore, SSFISize, SSFISize);
6537 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
6538 Ops, array_lengthof(Ops),
6539 Op.getValueType(), MMO);
6540 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
6541 MachinePointerInfo::getFixedStack(SSFI),
6548 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
6549 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
6550 SelectionDAG &DAG) const {
6551 // This algorithm is not obvious. Here it is in C code, more or less:
6553 double uint64_to_double( uint32_t hi, uint32_t lo ) {
6554 static const __m128i exp = { 0x4330000045300000ULL, 0 };
6555 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
6557 // Copy ints to xmm registers.
6558 __m128i xh = _mm_cvtsi32_si128( hi );
6559 __m128i xl = _mm_cvtsi32_si128( lo );
6561 // Combine into low half of a single xmm register.
6562 __m128i x = _mm_unpacklo_epi32( xh, xl );
6566 // Merge in appropriate exponents to give the integer bits the right
6568 x = _mm_unpacklo_epi32( x, exp );
6570 // Subtract away the biases to deal with the IEEE-754 double precision
6572 d = _mm_sub_pd( (__m128d) x, bias );
6574 // All conversions up to here are exact. The correctly rounded result is
6575 // calculated using the current rounding mode using the following
6577 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
6578 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
6579 // store doesn't really need to be here (except
6580 // maybe to zero the other double)
6585 DebugLoc dl = Op.getDebugLoc();
6586 LLVMContext *Context = DAG.getContext();
6588 // Build some magic constants.
6589 std::vector<Constant*> CV0;
6590 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
6591 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
6592 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6593 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
6594 Constant *C0 = ConstantVector::get(CV0);
6595 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
6597 std::vector<Constant*> CV1;
6599 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
6601 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
6602 Constant *C1 = ConstantVector::get(CV1);
6603 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
6605 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6606 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6608 DAG.getIntPtrConstant(1)));
6609 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6610 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6612 DAG.getIntPtrConstant(0)));
6613 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
6614 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
6615 MachinePointerInfo::getConstantPool(),
6617 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
6618 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
6619 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
6620 MachinePointerInfo::getConstantPool(),
6622 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
6624 // Add the halves; easiest way is to swap them into another reg first.
6625 int ShufMask[2] = { 1, -1 };
6626 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
6627 DAG.getUNDEF(MVT::v2f64), ShufMask);
6628 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
6629 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
6630 DAG.getIntPtrConstant(0));
6633 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
6634 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
6635 SelectionDAG &DAG) const {
6636 DebugLoc dl = Op.getDebugLoc();
6637 // FP constant to bias correct the final result.
6638 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
6641 // Load the 32-bit value into an XMM register.
6642 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
6643 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6645 DAG.getIntPtrConstant(0)));
6647 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6648 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
6649 DAG.getIntPtrConstant(0));
6651 // Or the load with the bias.
6652 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
6653 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6654 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6656 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6657 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6658 MVT::v2f64, Bias)));
6659 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
6660 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
6661 DAG.getIntPtrConstant(0));
6663 // Subtract the bias.
6664 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
6666 // Handle final rounding.
6667 EVT DestVT = Op.getValueType();
6669 if (DestVT.bitsLT(MVT::f64)) {
6670 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
6671 DAG.getIntPtrConstant(0));
6672 } else if (DestVT.bitsGT(MVT::f64)) {
6673 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
6676 // Handle final rounding.
6680 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
6681 SelectionDAG &DAG) const {
6682 SDValue N0 = Op.getOperand(0);
6683 DebugLoc dl = Op.getDebugLoc();
6685 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
6686 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
6687 // the optimization here.
6688 if (DAG.SignBitIsZero(N0))
6689 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
6691 EVT SrcVT = N0.getValueType();
6692 EVT DstVT = Op.getValueType();
6693 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
6694 return LowerUINT_TO_FP_i64(Op, DAG);
6695 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
6696 return LowerUINT_TO_FP_i32(Op, DAG);
6698 // Make a 64-bit buffer, and use it to build an FILD.
6699 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
6700 if (SrcVT == MVT::i32) {
6701 SDValue WordOff = DAG.getConstant(4, getPointerTy());
6702 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
6703 getPointerTy(), StackSlot, WordOff);
6704 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6705 StackSlot, MachinePointerInfo(),
6707 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
6708 OffsetSlot, MachinePointerInfo(),
6710 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
6714 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
6715 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
6716 StackSlot, MachinePointerInfo(),
6718 // For i64 source, we need to add the appropriate power of 2 if the input
6719 // was negative. This is the same as the optimization in
6720 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
6721 // we must be careful to do the computation in x87 extended precision, not
6722 // in SSE. (The generic code can't know it's OK to do this, or how to.)
6723 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
6724 MachineMemOperand *MMO =
6725 DAG.getMachineFunction()
6726 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6727 MachineMemOperand::MOLoad, 8, 8);
6729 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
6730 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
6731 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
6734 APInt FF(32, 0x5F800000ULL);
6736 // Check whether the sign bit is set.
6737 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
6738 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
6741 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
6742 SDValue FudgePtr = DAG.getConstantPool(
6743 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
6746 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
6747 SDValue Zero = DAG.getIntPtrConstant(0);
6748 SDValue Four = DAG.getIntPtrConstant(4);
6749 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
6751 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
6753 // Load the value out, extending it from f32 to f80.
6754 // FIXME: Avoid the extend by constructing the right constant pool?
6755 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
6756 FudgePtr, MachinePointerInfo::getConstantPool(),
6757 MVT::f32, false, false, 4);
6758 // Extend everything to 80 bits to force it to be done on x87.
6759 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
6760 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
6763 std::pair<SDValue,SDValue> X86TargetLowering::
6764 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
6765 DebugLoc DL = Op.getDebugLoc();
6767 EVT DstTy = Op.getValueType();
6770 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
6774 assert(DstTy.getSimpleVT() <= MVT::i64 &&
6775 DstTy.getSimpleVT() >= MVT::i16 &&
6776 "Unknown FP_TO_SINT to lower!");
6778 // These are really Legal.
6779 if (DstTy == MVT::i32 &&
6780 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6781 return std::make_pair(SDValue(), SDValue());
6782 if (Subtarget->is64Bit() &&
6783 DstTy == MVT::i64 &&
6784 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
6785 return std::make_pair(SDValue(), SDValue());
6787 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
6789 MachineFunction &MF = DAG.getMachineFunction();
6790 unsigned MemSize = DstTy.getSizeInBits()/8;
6791 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6792 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6797 switch (DstTy.getSimpleVT().SimpleTy) {
6798 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
6799 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
6800 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
6801 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
6804 SDValue Chain = DAG.getEntryNode();
6805 SDValue Value = Op.getOperand(0);
6806 EVT TheVT = Op.getOperand(0).getValueType();
6807 if (isScalarFPTypeInSSEReg(TheVT)) {
6808 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
6809 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
6810 MachinePointerInfo::getFixedStack(SSFI),
6812 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
6814 Chain, StackSlot, DAG.getValueType(TheVT)
6817 MachineMemOperand *MMO =
6818 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6819 MachineMemOperand::MOLoad, MemSize, MemSize);
6820 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
6822 Chain = Value.getValue(1);
6823 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
6824 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6827 MachineMemOperand *MMO =
6828 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
6829 MachineMemOperand::MOStore, MemSize, MemSize);
6831 // Build the FP_TO_INT*_IN_MEM
6832 SDValue Ops[] = { Chain, Value, StackSlot };
6833 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
6834 Ops, 3, DstTy, MMO);
6836 return std::make_pair(FIST, StackSlot);
6839 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
6840 SelectionDAG &DAG) const {
6841 if (Op.getValueType().isVector())
6844 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
6845 SDValue FIST = Vals.first, StackSlot = Vals.second;
6846 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
6847 if (FIST.getNode() == 0) return Op;
6850 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6851 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6854 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
6855 SelectionDAG &DAG) const {
6856 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
6857 SDValue FIST = Vals.first, StackSlot = Vals.second;
6858 assert(FIST.getNode() && "Unexpected failure");
6861 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
6862 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
6865 SDValue X86TargetLowering::LowerFABS(SDValue Op,
6866 SelectionDAG &DAG) const {
6867 LLVMContext *Context = DAG.getContext();
6868 DebugLoc dl = Op.getDebugLoc();
6869 EVT VT = Op.getValueType();
6872 EltVT = VT.getVectorElementType();
6873 std::vector<Constant*> CV;
6874 if (EltVT == MVT::f64) {
6875 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
6879 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
6885 Constant *C = ConstantVector::get(CV);
6886 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6887 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6888 MachinePointerInfo::getConstantPool(),
6890 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
6893 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
6894 LLVMContext *Context = DAG.getContext();
6895 DebugLoc dl = Op.getDebugLoc();
6896 EVT VT = Op.getValueType();
6899 EltVT = VT.getVectorElementType();
6900 std::vector<Constant*> CV;
6901 if (EltVT == MVT::f64) {
6902 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
6906 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
6912 Constant *C = ConstantVector::get(CV);
6913 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6914 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6915 MachinePointerInfo::getConstantPool(),
6917 if (VT.isVector()) {
6918 return DAG.getNode(ISD::BITCAST, dl, VT,
6919 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
6920 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
6922 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
6924 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
6928 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
6929 LLVMContext *Context = DAG.getContext();
6930 SDValue Op0 = Op.getOperand(0);
6931 SDValue Op1 = Op.getOperand(1);
6932 DebugLoc dl = Op.getDebugLoc();
6933 EVT VT = Op.getValueType();
6934 EVT SrcVT = Op1.getValueType();
6936 // If second operand is smaller, extend it first.
6937 if (SrcVT.bitsLT(VT)) {
6938 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
6941 // And if it is bigger, shrink it first.
6942 if (SrcVT.bitsGT(VT)) {
6943 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
6947 // At this point the operands and the result should have the same
6948 // type, and that won't be f80 since that is not custom lowered.
6950 // First get the sign bit of second operand.
6951 std::vector<Constant*> CV;
6952 if (SrcVT == MVT::f64) {
6953 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
6954 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6956 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
6957 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6958 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6959 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6961 Constant *C = ConstantVector::get(CV);
6962 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6963 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
6964 MachinePointerInfo::getConstantPool(),
6966 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
6968 // Shift sign bit right or left if the two operands have different types.
6969 if (SrcVT.bitsGT(VT)) {
6970 // Op0 is MVT::f32, Op1 is MVT::f64.
6971 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
6972 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
6973 DAG.getConstant(32, MVT::i32));
6974 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
6975 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
6976 DAG.getIntPtrConstant(0));
6979 // Clear first operand sign bit.
6981 if (VT == MVT::f64) {
6982 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
6983 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
6985 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
6986 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6987 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6988 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
6990 C = ConstantVector::get(CV);
6991 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
6992 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
6993 MachinePointerInfo::getConstantPool(),
6995 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
6997 // Or the value with the sign bit.
6998 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7001 /// Emit nodes that will be selected as "test Op0,Op0", or something
7003 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7004 SelectionDAG &DAG) const {
7005 DebugLoc dl = Op.getDebugLoc();
7007 // CF and OF aren't always set the way we want. Determine which
7008 // of these we need.
7009 bool NeedCF = false;
7010 bool NeedOF = false;
7013 case X86::COND_A: case X86::COND_AE:
7014 case X86::COND_B: case X86::COND_BE:
7017 case X86::COND_G: case X86::COND_GE:
7018 case X86::COND_L: case X86::COND_LE:
7019 case X86::COND_O: case X86::COND_NO:
7024 // See if we can use the EFLAGS value from the operand instead of
7025 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7026 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7027 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7028 // Emit a CMP with 0, which is the TEST pattern.
7029 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7030 DAG.getConstant(0, Op.getValueType()));
7032 unsigned Opcode = 0;
7033 unsigned NumOperands = 0;
7034 switch (Op.getNode()->getOpcode()) {
7036 // Due to an isel shortcoming, be conservative if this add is likely to be
7037 // selected as part of a load-modify-store instruction. When the root node
7038 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7039 // uses of other nodes in the match, such as the ADD in this case. This
7040 // leads to the ADD being left around and reselected, with the result being
7041 // two adds in the output. Alas, even if none our users are stores, that
7042 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7043 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7044 // climbing the DAG back to the root, and it doesn't seem to be worth the
7046 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7047 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7048 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7051 if (ConstantSDNode *C =
7052 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7053 // An add of one will be selected as an INC.
7054 if (C->getAPIntValue() == 1) {
7055 Opcode = X86ISD::INC;
7060 // An add of negative one (subtract of one) will be selected as a DEC.
7061 if (C->getAPIntValue().isAllOnesValue()) {
7062 Opcode = X86ISD::DEC;
7068 // Otherwise use a regular EFLAGS-setting add.
7069 Opcode = X86ISD::ADD;
7073 // If the primary and result isn't used, don't bother using X86ISD::AND,
7074 // because a TEST instruction will be better.
7075 bool NonFlagUse = false;
7076 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7077 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7079 unsigned UOpNo = UI.getOperandNo();
7080 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7081 // Look pass truncate.
7082 UOpNo = User->use_begin().getOperandNo();
7083 User = *User->use_begin();
7086 if (User->getOpcode() != ISD::BRCOND &&
7087 User->getOpcode() != ISD::SETCC &&
7088 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7101 // Due to the ISEL shortcoming noted above, be conservative if this op is
7102 // likely to be selected as part of a load-modify-store instruction.
7103 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7104 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7105 if (UI->getOpcode() == ISD::STORE)
7108 // Otherwise use a regular EFLAGS-setting instruction.
7109 switch (Op.getNode()->getOpcode()) {
7110 default: llvm_unreachable("unexpected operator!");
7111 case ISD::SUB: Opcode = X86ISD::SUB; break;
7112 case ISD::OR: Opcode = X86ISD::OR; break;
7113 case ISD::XOR: Opcode = X86ISD::XOR; break;
7114 case ISD::AND: Opcode = X86ISD::AND; break;
7126 return SDValue(Op.getNode(), 1);
7133 // Emit a CMP with 0, which is the TEST pattern.
7134 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7135 DAG.getConstant(0, Op.getValueType()));
7137 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7138 SmallVector<SDValue, 4> Ops;
7139 for (unsigned i = 0; i != NumOperands; ++i)
7140 Ops.push_back(Op.getOperand(i));
7142 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7143 DAG.ReplaceAllUsesWith(Op, New);
7144 return SDValue(New.getNode(), 1);
7147 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7149 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7150 SelectionDAG &DAG) const {
7151 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7152 if (C->getAPIntValue() == 0)
7153 return EmitTest(Op0, X86CC, DAG);
7155 DebugLoc dl = Op0.getDebugLoc();
7156 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7159 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7160 /// if it's possible.
7161 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7162 DebugLoc dl, SelectionDAG &DAG) const {
7163 SDValue Op0 = And.getOperand(0);
7164 SDValue Op1 = And.getOperand(1);
7165 if (Op0.getOpcode() == ISD::TRUNCATE)
7166 Op0 = Op0.getOperand(0);
7167 if (Op1.getOpcode() == ISD::TRUNCATE)
7168 Op1 = Op1.getOperand(0);
7171 if (Op1.getOpcode() == ISD::SHL)
7172 std::swap(Op0, Op1);
7173 if (Op0.getOpcode() == ISD::SHL) {
7174 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7175 if (And00C->getZExtValue() == 1) {
7176 // If we looked past a truncate, check that it's only truncating away
7178 unsigned BitWidth = Op0.getValueSizeInBits();
7179 unsigned AndBitWidth = And.getValueSizeInBits();
7180 if (BitWidth > AndBitWidth) {
7181 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7182 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7183 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7187 RHS = Op0.getOperand(1);
7189 } else if (Op1.getOpcode() == ISD::Constant) {
7190 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7191 SDValue AndLHS = Op0;
7192 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7193 LHS = AndLHS.getOperand(0);
7194 RHS = AndLHS.getOperand(1);
7198 if (LHS.getNode()) {
7199 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7200 // instruction. Since the shift amount is in-range-or-undefined, we know
7201 // that doing a bittest on the i32 value is ok. We extend to i32 because
7202 // the encoding for the i16 version is larger than the i32 version.
7203 // Also promote i16 to i32 for performance / code size reason.
7204 if (LHS.getValueType() == MVT::i8 ||
7205 LHS.getValueType() == MVT::i16)
7206 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7208 // If the operand types disagree, extend the shift amount to match. Since
7209 // BT ignores high bits (like shifts) we can use anyextend.
7210 if (LHS.getValueType() != RHS.getValueType())
7211 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7213 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7214 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7215 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7216 DAG.getConstant(Cond, MVT::i8), BT);
7222 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7223 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7224 SDValue Op0 = Op.getOperand(0);
7225 SDValue Op1 = Op.getOperand(1);
7226 DebugLoc dl = Op.getDebugLoc();
7227 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7229 // Optimize to BT if possible.
7230 // Lower (X & (1 << N)) == 0 to BT(X, N).
7231 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7232 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7233 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7234 Op1.getOpcode() == ISD::Constant &&
7235 cast<ConstantSDNode>(Op1)->isNullValue() &&
7236 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7237 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7238 if (NewSetCC.getNode())
7242 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7244 if (Op1.getOpcode() == ISD::Constant &&
7245 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7246 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7247 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7249 // If the input is a setcc, then reuse the input setcc or use a new one with
7250 // the inverted condition.
7251 if (Op0.getOpcode() == X86ISD::SETCC) {
7252 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7253 bool Invert = (CC == ISD::SETNE) ^
7254 cast<ConstantSDNode>(Op1)->isNullValue();
7255 if (!Invert) return Op0;
7257 CCode = X86::GetOppositeBranchCondition(CCode);
7258 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7259 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7263 bool isFP = Op1.getValueType().isFloatingPoint();
7264 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7265 if (X86CC == X86::COND_INVALID)
7268 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7269 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7270 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7273 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7275 SDValue Op0 = Op.getOperand(0);
7276 SDValue Op1 = Op.getOperand(1);
7277 SDValue CC = Op.getOperand(2);
7278 EVT VT = Op.getValueType();
7279 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7280 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7281 DebugLoc dl = Op.getDebugLoc();
7285 EVT VT0 = Op0.getValueType();
7286 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7287 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7290 switch (SetCCOpcode) {
7293 case ISD::SETEQ: SSECC = 0; break;
7295 case ISD::SETGT: Swap = true; // Fallthrough
7297 case ISD::SETOLT: SSECC = 1; break;
7299 case ISD::SETGE: Swap = true; // Fallthrough
7301 case ISD::SETOLE: SSECC = 2; break;
7302 case ISD::SETUO: SSECC = 3; break;
7304 case ISD::SETNE: SSECC = 4; break;
7305 case ISD::SETULE: Swap = true;
7306 case ISD::SETUGE: SSECC = 5; break;
7307 case ISD::SETULT: Swap = true;
7308 case ISD::SETUGT: SSECC = 6; break;
7309 case ISD::SETO: SSECC = 7; break;
7312 std::swap(Op0, Op1);
7314 // In the two special cases we can't handle, emit two comparisons.
7316 if (SetCCOpcode == ISD::SETUEQ) {
7318 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7319 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7320 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7322 else if (SetCCOpcode == ISD::SETONE) {
7324 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7325 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7326 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7328 llvm_unreachable("Illegal FP comparison");
7330 // Handle all other FP comparisons here.
7331 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7334 // We are handling one of the integer comparisons here. Since SSE only has
7335 // GT and EQ comparisons for integer, swapping operands and multiple
7336 // operations may be required for some comparisons.
7337 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7338 bool Swap = false, Invert = false, FlipSigns = false;
7340 switch (VT.getSimpleVT().SimpleTy) {
7342 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7343 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7344 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7345 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7348 switch (SetCCOpcode) {
7350 case ISD::SETNE: Invert = true;
7351 case ISD::SETEQ: Opc = EQOpc; break;
7352 case ISD::SETLT: Swap = true;
7353 case ISD::SETGT: Opc = GTOpc; break;
7354 case ISD::SETGE: Swap = true;
7355 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7356 case ISD::SETULT: Swap = true;
7357 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7358 case ISD::SETUGE: Swap = true;
7359 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7362 std::swap(Op0, Op1);
7364 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7365 // bits of the inputs before performing those operations.
7367 EVT EltVT = VT.getVectorElementType();
7368 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7370 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7371 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7373 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7374 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7377 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7379 // If the logical-not of the result is required, perform that now.
7381 Result = DAG.getNOT(dl, Result, VT);
7386 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
7387 static bool isX86LogicalCmp(SDValue Op) {
7388 unsigned Opc = Op.getNode()->getOpcode();
7389 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
7391 if (Op.getResNo() == 1 &&
7392 (Opc == X86ISD::ADD ||
7393 Opc == X86ISD::SUB ||
7394 Opc == X86ISD::ADC ||
7395 Opc == X86ISD::SBB ||
7396 Opc == X86ISD::SMUL ||
7397 Opc == X86ISD::UMUL ||
7398 Opc == X86ISD::INC ||
7399 Opc == X86ISD::DEC ||
7400 Opc == X86ISD::OR ||
7401 Opc == X86ISD::XOR ||
7402 Opc == X86ISD::AND))
7405 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
7411 static bool isZero(SDValue V) {
7412 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7413 return C && C->isNullValue();
7416 static bool isAllOnes(SDValue V) {
7417 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
7418 return C && C->isAllOnesValue();
7421 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
7422 bool addTest = true;
7423 SDValue Cond = Op.getOperand(0);
7424 SDValue Op1 = Op.getOperand(1);
7425 SDValue Op2 = Op.getOperand(2);
7426 DebugLoc DL = Op.getDebugLoc();
7429 if (Cond.getOpcode() == ISD::SETCC) {
7430 SDValue NewCond = LowerSETCC(Cond, DAG);
7431 if (NewCond.getNode())
7435 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
7436 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
7437 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
7438 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
7439 if (Cond.getOpcode() == X86ISD::SETCC &&
7440 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
7441 isZero(Cond.getOperand(1).getOperand(1))) {
7442 SDValue Cmp = Cond.getOperand(1);
7444 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
7446 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
7447 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
7448 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
7450 SDValue CmpOp0 = Cmp.getOperand(0);
7451 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
7452 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
7454 SDValue Res = // Res = 0 or -1.
7455 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7456 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
7458 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
7459 Res = DAG.getNOT(DL, Res, Res.getValueType());
7461 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
7462 if (N2C == 0 || !N2C->isNullValue())
7463 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
7468 // Look past (and (setcc_carry (cmp ...)), 1).
7469 if (Cond.getOpcode() == ISD::AND &&
7470 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7471 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7472 if (C && C->getAPIntValue() == 1)
7473 Cond = Cond.getOperand(0);
7476 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7477 // setting operand in place of the X86ISD::SETCC.
7478 if (Cond.getOpcode() == X86ISD::SETCC ||
7479 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7480 CC = Cond.getOperand(0);
7482 SDValue Cmp = Cond.getOperand(1);
7483 unsigned Opc = Cmp.getOpcode();
7484 EVT VT = Op.getValueType();
7486 bool IllegalFPCMov = false;
7487 if (VT.isFloatingPoint() && !VT.isVector() &&
7488 !isScalarFPTypeInSSEReg(VT)) // FPStack?
7489 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
7491 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
7492 Opc == X86ISD::BT) { // FIXME
7499 // Look pass the truncate.
7500 if (Cond.getOpcode() == ISD::TRUNCATE)
7501 Cond = Cond.getOperand(0);
7503 // We know the result of AND is compared against zero. Try to match
7505 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7506 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
7507 if (NewSetCC.getNode()) {
7508 CC = NewSetCC.getOperand(0);
7509 Cond = NewSetCC.getOperand(1);
7516 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7517 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7520 // a < b ? -1 : 0 -> RES = ~setcc_carry
7521 // a < b ? 0 : -1 -> RES = setcc_carry
7522 // a >= b ? -1 : 0 -> RES = setcc_carry
7523 // a >= b ? 0 : -1 -> RES = ~setcc_carry
7524 if (Cond.getOpcode() == X86ISD::CMP) {
7525 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
7527 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
7528 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
7529 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
7530 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
7531 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
7532 return DAG.getNOT(DL, Res, Res.getValueType());
7537 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
7538 // condition is true.
7539 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
7540 SDValue Ops[] = { Op2, Op1, CC, Cond };
7541 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
7544 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
7545 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
7546 // from the AND / OR.
7547 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
7548 Opc = Op.getOpcode();
7549 if (Opc != ISD::OR && Opc != ISD::AND)
7551 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7552 Op.getOperand(0).hasOneUse() &&
7553 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
7554 Op.getOperand(1).hasOneUse());
7557 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
7558 // 1 and that the SETCC node has a single use.
7559 static bool isXor1OfSetCC(SDValue Op) {
7560 if (Op.getOpcode() != ISD::XOR)
7562 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
7563 if (N1C && N1C->getAPIntValue() == 1) {
7564 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
7565 Op.getOperand(0).hasOneUse();
7570 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
7571 bool addTest = true;
7572 SDValue Chain = Op.getOperand(0);
7573 SDValue Cond = Op.getOperand(1);
7574 SDValue Dest = Op.getOperand(2);
7575 DebugLoc dl = Op.getDebugLoc();
7578 if (Cond.getOpcode() == ISD::SETCC) {
7579 SDValue NewCond = LowerSETCC(Cond, DAG);
7580 if (NewCond.getNode())
7584 // FIXME: LowerXALUO doesn't handle these!!
7585 else if (Cond.getOpcode() == X86ISD::ADD ||
7586 Cond.getOpcode() == X86ISD::SUB ||
7587 Cond.getOpcode() == X86ISD::SMUL ||
7588 Cond.getOpcode() == X86ISD::UMUL)
7589 Cond = LowerXALUO(Cond, DAG);
7592 // Look pass (and (setcc_carry (cmp ...)), 1).
7593 if (Cond.getOpcode() == ISD::AND &&
7594 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
7595 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
7596 if (C && C->getAPIntValue() == 1)
7597 Cond = Cond.getOperand(0);
7600 // If condition flag is set by a X86ISD::CMP, then use it as the condition
7601 // setting operand in place of the X86ISD::SETCC.
7602 if (Cond.getOpcode() == X86ISD::SETCC ||
7603 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
7604 CC = Cond.getOperand(0);
7606 SDValue Cmp = Cond.getOperand(1);
7607 unsigned Opc = Cmp.getOpcode();
7608 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
7609 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
7613 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
7617 // These can only come from an arithmetic instruction with overflow,
7618 // e.g. SADDO, UADDO.
7619 Cond = Cond.getNode()->getOperand(1);
7626 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
7627 SDValue Cmp = Cond.getOperand(0).getOperand(1);
7628 if (CondOpc == ISD::OR) {
7629 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
7630 // two branches instead of an explicit OR instruction with a
7632 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7633 isX86LogicalCmp(Cmp)) {
7634 CC = Cond.getOperand(0).getOperand(0);
7635 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7636 Chain, Dest, CC, Cmp);
7637 CC = Cond.getOperand(1).getOperand(0);
7641 } else { // ISD::AND
7642 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
7643 // two branches instead of an explicit AND instruction with a
7644 // separate test. However, we only do this if this block doesn't
7645 // have a fall-through edge, because this requires an explicit
7646 // jmp when the condition is false.
7647 if (Cmp == Cond.getOperand(1).getOperand(1) &&
7648 isX86LogicalCmp(Cmp) &&
7649 Op.getNode()->hasOneUse()) {
7650 X86::CondCode CCode =
7651 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7652 CCode = X86::GetOppositeBranchCondition(CCode);
7653 CC = DAG.getConstant(CCode, MVT::i8);
7654 SDNode *User = *Op.getNode()->use_begin();
7655 // Look for an unconditional branch following this conditional branch.
7656 // We need this because we need to reverse the successors in order
7657 // to implement FCMP_OEQ.
7658 if (User->getOpcode() == ISD::BR) {
7659 SDValue FalseBB = User->getOperand(1);
7661 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
7662 assert(NewBR == User);
7666 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7667 Chain, Dest, CC, Cmp);
7668 X86::CondCode CCode =
7669 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
7670 CCode = X86::GetOppositeBranchCondition(CCode);
7671 CC = DAG.getConstant(CCode, MVT::i8);
7677 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
7678 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
7679 // It should be transformed during dag combiner except when the condition
7680 // is set by a arithmetics with overflow node.
7681 X86::CondCode CCode =
7682 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
7683 CCode = X86::GetOppositeBranchCondition(CCode);
7684 CC = DAG.getConstant(CCode, MVT::i8);
7685 Cond = Cond.getOperand(0).getOperand(1);
7691 // Look pass the truncate.
7692 if (Cond.getOpcode() == ISD::TRUNCATE)
7693 Cond = Cond.getOperand(0);
7695 // We know the result of AND is compared against zero. Try to match
7697 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
7698 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
7699 if (NewSetCC.getNode()) {
7700 CC = NewSetCC.getOperand(0);
7701 Cond = NewSetCC.getOperand(1);
7708 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7709 Cond = EmitTest(Cond, X86::COND_NE, DAG);
7711 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
7712 Chain, Dest, CC, Cond);
7716 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
7717 // Calls to _alloca is needed to probe the stack when allocating more than 4k
7718 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
7719 // that the guard pages used by the OS virtual memory manager are allocated in
7720 // correct sequence.
7722 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
7723 SelectionDAG &DAG) const {
7724 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
7725 "This should be used only on Windows targets");
7726 DebugLoc dl = Op.getDebugLoc();
7729 SDValue Chain = Op.getOperand(0);
7730 SDValue Size = Op.getOperand(1);
7731 // FIXME: Ensure alignment here
7735 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
7737 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
7738 Flag = Chain.getValue(1);
7740 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7742 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
7743 Flag = Chain.getValue(1);
7745 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
7747 SDValue Ops1[2] = { Chain.getValue(0), Chain };
7748 return DAG.getMergeValues(Ops1, 2, dl);
7751 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
7752 MachineFunction &MF = DAG.getMachineFunction();
7753 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
7755 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7756 DebugLoc DL = Op.getDebugLoc();
7758 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
7759 // vastart just stores the address of the VarArgsFrameIndex slot into the
7760 // memory location argument.
7761 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7763 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
7764 MachinePointerInfo(SV), false, false, 0);
7768 // gp_offset (0 - 6 * 8)
7769 // fp_offset (48 - 48 + 8 * 16)
7770 // overflow_arg_area (point to parameters coming in memory).
7772 SmallVector<SDValue, 8> MemOps;
7773 SDValue FIN = Op.getOperand(1);
7775 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
7776 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
7778 FIN, MachinePointerInfo(SV), false, false, 0);
7779 MemOps.push_back(Store);
7782 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7783 FIN, DAG.getIntPtrConstant(4));
7784 Store = DAG.getStore(Op.getOperand(0), DL,
7785 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
7787 FIN, MachinePointerInfo(SV, 4), false, false, 0);
7788 MemOps.push_back(Store);
7790 // Store ptr to overflow_arg_area
7791 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7792 FIN, DAG.getIntPtrConstant(4));
7793 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
7795 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
7796 MachinePointerInfo(SV, 8),
7798 MemOps.push_back(Store);
7800 // Store ptr to reg_save_area.
7801 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7802 FIN, DAG.getIntPtrConstant(8));
7803 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
7805 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
7806 MachinePointerInfo(SV, 16), false, false, 0);
7807 MemOps.push_back(Store);
7808 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
7809 &MemOps[0], MemOps.size());
7812 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
7813 assert(Subtarget->is64Bit() &&
7814 "LowerVAARG only handles 64-bit va_arg!");
7815 assert((Subtarget->isTargetLinux() ||
7816 Subtarget->isTargetDarwin()) &&
7817 "Unhandled target in LowerVAARG");
7818 assert(Op.getNode()->getNumOperands() == 4);
7819 SDValue Chain = Op.getOperand(0);
7820 SDValue SrcPtr = Op.getOperand(1);
7821 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
7822 unsigned Align = Op.getConstantOperandVal(3);
7823 DebugLoc dl = Op.getDebugLoc();
7825 EVT ArgVT = Op.getNode()->getValueType(0);
7826 const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
7827 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
7830 // Decide which area this value should be read from.
7831 // TODO: Implement the AMD64 ABI in its entirety. This simple
7832 // selection mechanism works only for the basic types.
7833 if (ArgVT == MVT::f80) {
7834 llvm_unreachable("va_arg for f80 not yet implemented");
7835 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
7836 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
7837 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
7838 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
7840 llvm_unreachable("Unhandled argument type in LowerVAARG");
7844 // Sanity Check: Make sure using fp_offset makes sense.
7845 assert(!UseSoftFloat &&
7846 !(DAG.getMachineFunction()
7847 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
7848 Subtarget->hasXMM());
7851 // Insert VAARG_64 node into the DAG
7852 // VAARG_64 returns two values: Variable Argument Address, Chain
7853 SmallVector<SDValue, 11> InstOps;
7854 InstOps.push_back(Chain);
7855 InstOps.push_back(SrcPtr);
7856 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
7857 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
7858 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
7859 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
7860 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
7861 VTs, &InstOps[0], InstOps.size(),
7863 MachinePointerInfo(SV),
7868 Chain = VAARG.getValue(1);
7870 // Load the next argument and return it
7871 return DAG.getLoad(ArgVT, dl,
7874 MachinePointerInfo(),
7878 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
7879 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
7880 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
7881 SDValue Chain = Op.getOperand(0);
7882 SDValue DstPtr = Op.getOperand(1);
7883 SDValue SrcPtr = Op.getOperand(2);
7884 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
7885 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7886 DebugLoc DL = Op.getDebugLoc();
7888 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
7889 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
7891 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
7895 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
7896 DebugLoc dl = Op.getDebugLoc();
7897 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7899 default: return SDValue(); // Don't custom lower most intrinsics.
7900 // Comparison intrinsics.
7901 case Intrinsic::x86_sse_comieq_ss:
7902 case Intrinsic::x86_sse_comilt_ss:
7903 case Intrinsic::x86_sse_comile_ss:
7904 case Intrinsic::x86_sse_comigt_ss:
7905 case Intrinsic::x86_sse_comige_ss:
7906 case Intrinsic::x86_sse_comineq_ss:
7907 case Intrinsic::x86_sse_ucomieq_ss:
7908 case Intrinsic::x86_sse_ucomilt_ss:
7909 case Intrinsic::x86_sse_ucomile_ss:
7910 case Intrinsic::x86_sse_ucomigt_ss:
7911 case Intrinsic::x86_sse_ucomige_ss:
7912 case Intrinsic::x86_sse_ucomineq_ss:
7913 case Intrinsic::x86_sse2_comieq_sd:
7914 case Intrinsic::x86_sse2_comilt_sd:
7915 case Intrinsic::x86_sse2_comile_sd:
7916 case Intrinsic::x86_sse2_comigt_sd:
7917 case Intrinsic::x86_sse2_comige_sd:
7918 case Intrinsic::x86_sse2_comineq_sd:
7919 case Intrinsic::x86_sse2_ucomieq_sd:
7920 case Intrinsic::x86_sse2_ucomilt_sd:
7921 case Intrinsic::x86_sse2_ucomile_sd:
7922 case Intrinsic::x86_sse2_ucomigt_sd:
7923 case Intrinsic::x86_sse2_ucomige_sd:
7924 case Intrinsic::x86_sse2_ucomineq_sd: {
7926 ISD::CondCode CC = ISD::SETCC_INVALID;
7929 case Intrinsic::x86_sse_comieq_ss:
7930 case Intrinsic::x86_sse2_comieq_sd:
7934 case Intrinsic::x86_sse_comilt_ss:
7935 case Intrinsic::x86_sse2_comilt_sd:
7939 case Intrinsic::x86_sse_comile_ss:
7940 case Intrinsic::x86_sse2_comile_sd:
7944 case Intrinsic::x86_sse_comigt_ss:
7945 case Intrinsic::x86_sse2_comigt_sd:
7949 case Intrinsic::x86_sse_comige_ss:
7950 case Intrinsic::x86_sse2_comige_sd:
7954 case Intrinsic::x86_sse_comineq_ss:
7955 case Intrinsic::x86_sse2_comineq_sd:
7959 case Intrinsic::x86_sse_ucomieq_ss:
7960 case Intrinsic::x86_sse2_ucomieq_sd:
7961 Opc = X86ISD::UCOMI;
7964 case Intrinsic::x86_sse_ucomilt_ss:
7965 case Intrinsic::x86_sse2_ucomilt_sd:
7966 Opc = X86ISD::UCOMI;
7969 case Intrinsic::x86_sse_ucomile_ss:
7970 case Intrinsic::x86_sse2_ucomile_sd:
7971 Opc = X86ISD::UCOMI;
7974 case Intrinsic::x86_sse_ucomigt_ss:
7975 case Intrinsic::x86_sse2_ucomigt_sd:
7976 Opc = X86ISD::UCOMI;
7979 case Intrinsic::x86_sse_ucomige_ss:
7980 case Intrinsic::x86_sse2_ucomige_sd:
7981 Opc = X86ISD::UCOMI;
7984 case Intrinsic::x86_sse_ucomineq_ss:
7985 case Intrinsic::x86_sse2_ucomineq_sd:
7986 Opc = X86ISD::UCOMI;
7991 SDValue LHS = Op.getOperand(1);
7992 SDValue RHS = Op.getOperand(2);
7993 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
7994 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
7995 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
7996 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7997 DAG.getConstant(X86CC, MVT::i8), Cond);
7998 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8000 // ptest and testp intrinsics. The intrinsic these come from are designed to
8001 // return an integer value, not just an instruction so lower it to the ptest
8002 // or testp pattern and a setcc for the result.
8003 case Intrinsic::x86_sse41_ptestz:
8004 case Intrinsic::x86_sse41_ptestc:
8005 case Intrinsic::x86_sse41_ptestnzc:
8006 case Intrinsic::x86_avx_ptestz_256:
8007 case Intrinsic::x86_avx_ptestc_256:
8008 case Intrinsic::x86_avx_ptestnzc_256:
8009 case Intrinsic::x86_avx_vtestz_ps:
8010 case Intrinsic::x86_avx_vtestc_ps:
8011 case Intrinsic::x86_avx_vtestnzc_ps:
8012 case Intrinsic::x86_avx_vtestz_pd:
8013 case Intrinsic::x86_avx_vtestc_pd:
8014 case Intrinsic::x86_avx_vtestnzc_pd:
8015 case Intrinsic::x86_avx_vtestz_ps_256:
8016 case Intrinsic::x86_avx_vtestc_ps_256:
8017 case Intrinsic::x86_avx_vtestnzc_ps_256:
8018 case Intrinsic::x86_avx_vtestz_pd_256:
8019 case Intrinsic::x86_avx_vtestc_pd_256:
8020 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8021 bool IsTestPacked = false;
8024 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8025 case Intrinsic::x86_avx_vtestz_ps:
8026 case Intrinsic::x86_avx_vtestz_pd:
8027 case Intrinsic::x86_avx_vtestz_ps_256:
8028 case Intrinsic::x86_avx_vtestz_pd_256:
8029 IsTestPacked = true; // Fallthrough
8030 case Intrinsic::x86_sse41_ptestz:
8031 case Intrinsic::x86_avx_ptestz_256:
8033 X86CC = X86::COND_E;
8035 case Intrinsic::x86_avx_vtestc_ps:
8036 case Intrinsic::x86_avx_vtestc_pd:
8037 case Intrinsic::x86_avx_vtestc_ps_256:
8038 case Intrinsic::x86_avx_vtestc_pd_256:
8039 IsTestPacked = true; // Fallthrough
8040 case Intrinsic::x86_sse41_ptestc:
8041 case Intrinsic::x86_avx_ptestc_256:
8043 X86CC = X86::COND_B;
8045 case Intrinsic::x86_avx_vtestnzc_ps:
8046 case Intrinsic::x86_avx_vtestnzc_pd:
8047 case Intrinsic::x86_avx_vtestnzc_ps_256:
8048 case Intrinsic::x86_avx_vtestnzc_pd_256:
8049 IsTestPacked = true; // Fallthrough
8050 case Intrinsic::x86_sse41_ptestnzc:
8051 case Intrinsic::x86_avx_ptestnzc_256:
8053 X86CC = X86::COND_A;
8057 SDValue LHS = Op.getOperand(1);
8058 SDValue RHS = Op.getOperand(2);
8059 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8060 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8061 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8062 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8063 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8066 // Fix vector shift instructions where the last operand is a non-immediate
8068 case Intrinsic::x86_sse2_pslli_w:
8069 case Intrinsic::x86_sse2_pslli_d:
8070 case Intrinsic::x86_sse2_pslli_q:
8071 case Intrinsic::x86_sse2_psrli_w:
8072 case Intrinsic::x86_sse2_psrli_d:
8073 case Intrinsic::x86_sse2_psrli_q:
8074 case Intrinsic::x86_sse2_psrai_w:
8075 case Intrinsic::x86_sse2_psrai_d:
8076 case Intrinsic::x86_mmx_pslli_w:
8077 case Intrinsic::x86_mmx_pslli_d:
8078 case Intrinsic::x86_mmx_pslli_q:
8079 case Intrinsic::x86_mmx_psrli_w:
8080 case Intrinsic::x86_mmx_psrli_d:
8081 case Intrinsic::x86_mmx_psrli_q:
8082 case Intrinsic::x86_mmx_psrai_w:
8083 case Intrinsic::x86_mmx_psrai_d: {
8084 SDValue ShAmt = Op.getOperand(2);
8085 if (isa<ConstantSDNode>(ShAmt))
8088 unsigned NewIntNo = 0;
8089 EVT ShAmtVT = MVT::v4i32;
8091 case Intrinsic::x86_sse2_pslli_w:
8092 NewIntNo = Intrinsic::x86_sse2_psll_w;
8094 case Intrinsic::x86_sse2_pslli_d:
8095 NewIntNo = Intrinsic::x86_sse2_psll_d;
8097 case Intrinsic::x86_sse2_pslli_q:
8098 NewIntNo = Intrinsic::x86_sse2_psll_q;
8100 case Intrinsic::x86_sse2_psrli_w:
8101 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8103 case Intrinsic::x86_sse2_psrli_d:
8104 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8106 case Intrinsic::x86_sse2_psrli_q:
8107 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8109 case Intrinsic::x86_sse2_psrai_w:
8110 NewIntNo = Intrinsic::x86_sse2_psra_w;
8112 case Intrinsic::x86_sse2_psrai_d:
8113 NewIntNo = Intrinsic::x86_sse2_psra_d;
8116 ShAmtVT = MVT::v2i32;
8118 case Intrinsic::x86_mmx_pslli_w:
8119 NewIntNo = Intrinsic::x86_mmx_psll_w;
8121 case Intrinsic::x86_mmx_pslli_d:
8122 NewIntNo = Intrinsic::x86_mmx_psll_d;
8124 case Intrinsic::x86_mmx_pslli_q:
8125 NewIntNo = Intrinsic::x86_mmx_psll_q;
8127 case Intrinsic::x86_mmx_psrli_w:
8128 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8130 case Intrinsic::x86_mmx_psrli_d:
8131 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8133 case Intrinsic::x86_mmx_psrli_q:
8134 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8136 case Intrinsic::x86_mmx_psrai_w:
8137 NewIntNo = Intrinsic::x86_mmx_psra_w;
8139 case Intrinsic::x86_mmx_psrai_d:
8140 NewIntNo = Intrinsic::x86_mmx_psra_d;
8142 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8148 // The vector shift intrinsics with scalars uses 32b shift amounts but
8149 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8153 ShOps[1] = DAG.getConstant(0, MVT::i32);
8154 if (ShAmtVT == MVT::v4i32) {
8155 ShOps[2] = DAG.getUNDEF(MVT::i32);
8156 ShOps[3] = DAG.getUNDEF(MVT::i32);
8157 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8159 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8160 // FIXME this must be lowered to get rid of the invalid type.
8163 EVT VT = Op.getValueType();
8164 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8165 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8166 DAG.getConstant(NewIntNo, MVT::i32),
8167 Op.getOperand(1), ShAmt);
8172 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8173 SelectionDAG &DAG) const {
8174 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8175 MFI->setReturnAddressIsTaken(true);
8177 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8178 DebugLoc dl = Op.getDebugLoc();
8181 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8183 DAG.getConstant(TD->getPointerSize(),
8184 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8185 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8186 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8188 MachinePointerInfo(), false, false, 0);
8191 // Just load the return address.
8192 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8193 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8194 RetAddrFI, MachinePointerInfo(), false, false, 0);
8197 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8198 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8199 MFI->setFrameAddressIsTaken(true);
8201 EVT VT = Op.getValueType();
8202 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8203 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8204 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8205 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8207 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8208 MachinePointerInfo(),
8213 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8214 SelectionDAG &DAG) const {
8215 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8218 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8219 MachineFunction &MF = DAG.getMachineFunction();
8220 SDValue Chain = Op.getOperand(0);
8221 SDValue Offset = Op.getOperand(1);
8222 SDValue Handler = Op.getOperand(2);
8223 DebugLoc dl = Op.getDebugLoc();
8225 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8226 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8228 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8230 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8231 DAG.getIntPtrConstant(TD->getPointerSize()));
8232 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8233 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8235 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8236 MF.getRegInfo().addLiveOut(StoreAddrReg);
8238 return DAG.getNode(X86ISD::EH_RETURN, dl,
8240 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8243 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8244 SelectionDAG &DAG) const {
8245 SDValue Root = Op.getOperand(0);
8246 SDValue Trmp = Op.getOperand(1); // trampoline
8247 SDValue FPtr = Op.getOperand(2); // nested function
8248 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8249 DebugLoc dl = Op.getDebugLoc();
8251 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8253 if (Subtarget->is64Bit()) {
8254 SDValue OutChains[6];
8256 // Large code-model.
8257 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8258 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8260 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
8261 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
8263 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8265 // Load the pointer to the nested function into R11.
8266 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8267 SDValue Addr = Trmp;
8268 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8269 Addr, MachinePointerInfo(TrmpAddr),
8272 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8273 DAG.getConstant(2, MVT::i64));
8274 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8275 MachinePointerInfo(TrmpAddr, 2),
8278 // Load the 'nest' parameter value into R10.
8279 // R10 is specified in X86CallingConv.td
8280 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8281 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8282 DAG.getConstant(10, MVT::i64));
8283 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8284 Addr, MachinePointerInfo(TrmpAddr, 10),
8287 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8288 DAG.getConstant(12, MVT::i64));
8289 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8290 MachinePointerInfo(TrmpAddr, 12),
8293 // Jump to the nested function.
8294 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8295 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8296 DAG.getConstant(20, MVT::i64));
8297 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8298 Addr, MachinePointerInfo(TrmpAddr, 20),
8301 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8302 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8303 DAG.getConstant(22, MVT::i64));
8304 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8305 MachinePointerInfo(TrmpAddr, 22),
8309 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8310 return DAG.getMergeValues(Ops, 2, dl);
8312 const Function *Func =
8313 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8314 CallingConv::ID CC = Func->getCallingConv();
8319 llvm_unreachable("Unsupported calling convention");
8320 case CallingConv::C:
8321 case CallingConv::X86_StdCall: {
8322 // Pass 'nest' parameter in ECX.
8323 // Must be kept in sync with X86CallingConv.td
8326 // Check that ECX wasn't needed by an 'inreg' parameter.
8327 const FunctionType *FTy = Func->getFunctionType();
8328 const AttrListPtr &Attrs = Func->getAttributes();
8330 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8331 unsigned InRegCount = 0;
8334 for (FunctionType::param_iterator I = FTy->param_begin(),
8335 E = FTy->param_end(); I != E; ++I, ++Idx)
8336 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8337 // FIXME: should only count parameters that are lowered to integers.
8338 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8340 if (InRegCount > 2) {
8341 report_fatal_error("Nest register in use - reduce number of inreg"
8347 case CallingConv::X86_FastCall:
8348 case CallingConv::X86_ThisCall:
8349 case CallingConv::Fast:
8350 // Pass 'nest' parameter in EAX.
8351 // Must be kept in sync with X86CallingConv.td
8356 SDValue OutChains[4];
8359 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8360 DAG.getConstant(10, MVT::i32));
8361 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8363 // This is storing the opcode for MOV32ri.
8364 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8365 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
8366 OutChains[0] = DAG.getStore(Root, dl,
8367 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8368 Trmp, MachinePointerInfo(TrmpAddr),
8371 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8372 DAG.getConstant(1, MVT::i32));
8373 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8374 MachinePointerInfo(TrmpAddr, 1),
8377 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8378 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8379 DAG.getConstant(5, MVT::i32));
8380 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
8381 MachinePointerInfo(TrmpAddr, 5),
8384 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8385 DAG.getConstant(6, MVT::i32));
8386 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
8387 MachinePointerInfo(TrmpAddr, 6),
8391 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
8392 return DAG.getMergeValues(Ops, 2, dl);
8396 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
8397 SelectionDAG &DAG) const {
8399 The rounding mode is in bits 11:10 of FPSR, and has the following
8406 FLT_ROUNDS, on the other hand, expects the following:
8413 To perform the conversion, we do:
8414 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
8417 MachineFunction &MF = DAG.getMachineFunction();
8418 const TargetMachine &TM = MF.getTarget();
8419 const TargetFrameLowering &TFI = *TM.getFrameLowering();
8420 unsigned StackAlignment = TFI.getStackAlignment();
8421 EVT VT = Op.getValueType();
8422 DebugLoc DL = Op.getDebugLoc();
8424 // Save FP Control Word to stack slot
8425 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
8426 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
8429 MachineMemOperand *MMO =
8430 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
8431 MachineMemOperand::MOStore, 2, 2);
8433 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
8434 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
8435 DAG.getVTList(MVT::Other),
8436 Ops, 2, MVT::i16, MMO);
8438 // Load FP Control Word from stack slot
8439 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
8440 MachinePointerInfo(), false, false, 0);
8442 // Transform as necessary
8444 DAG.getNode(ISD::SRL, DL, MVT::i16,
8445 DAG.getNode(ISD::AND, DL, MVT::i16,
8446 CWD, DAG.getConstant(0x800, MVT::i16)),
8447 DAG.getConstant(11, MVT::i8));
8449 DAG.getNode(ISD::SRL, DL, MVT::i16,
8450 DAG.getNode(ISD::AND, DL, MVT::i16,
8451 CWD, DAG.getConstant(0x400, MVT::i16)),
8452 DAG.getConstant(9, MVT::i8));
8455 DAG.getNode(ISD::AND, DL, MVT::i16,
8456 DAG.getNode(ISD::ADD, DL, MVT::i16,
8457 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
8458 DAG.getConstant(1, MVT::i16)),
8459 DAG.getConstant(3, MVT::i16));
8462 return DAG.getNode((VT.getSizeInBits() < 16 ?
8463 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
8466 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
8467 EVT VT = Op.getValueType();
8469 unsigned NumBits = VT.getSizeInBits();
8470 DebugLoc dl = Op.getDebugLoc();
8472 Op = Op.getOperand(0);
8473 if (VT == MVT::i8) {
8474 // Zero extend to i32 since there is not an i8 bsr.
8476 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8479 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
8480 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8481 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
8483 // If src is zero (i.e. bsr sets ZF), returns NumBits.
8486 DAG.getConstant(NumBits+NumBits-1, OpVT),
8487 DAG.getConstant(X86::COND_E, MVT::i8),
8490 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8492 // Finally xor with NumBits-1.
8493 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
8496 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8500 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
8501 EVT VT = Op.getValueType();
8503 unsigned NumBits = VT.getSizeInBits();
8504 DebugLoc dl = Op.getDebugLoc();
8506 Op = Op.getOperand(0);
8507 if (VT == MVT::i8) {
8509 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
8512 // Issue a bsf (scan bits forward) which also sets EFLAGS.
8513 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
8514 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
8516 // If src is zero (i.e. bsf sets ZF), returns NumBits.
8519 DAG.getConstant(NumBits, OpVT),
8520 DAG.getConstant(X86::COND_E, MVT::i8),
8523 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
8526 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
8530 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
8531 EVT VT = Op.getValueType();
8532 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
8533 DebugLoc dl = Op.getDebugLoc();
8535 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
8536 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
8537 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
8538 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
8539 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
8541 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
8542 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
8543 // return AloBlo + AloBhi + AhiBlo;
8545 SDValue A = Op.getOperand(0);
8546 SDValue B = Op.getOperand(1);
8548 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8549 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8550 A, DAG.getConstant(32, MVT::i32));
8551 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8552 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8553 B, DAG.getConstant(32, MVT::i32));
8554 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8555 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8557 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8558 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8560 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8561 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
8563 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8564 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8565 AloBhi, DAG.getConstant(32, MVT::i32));
8566 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8567 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8568 AhiBlo, DAG.getConstant(32, MVT::i32));
8569 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
8570 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
8574 SDValue X86TargetLowering::LowerSHL(SDValue Op, SelectionDAG &DAG) const {
8575 EVT VT = Op.getValueType();
8576 DebugLoc dl = Op.getDebugLoc();
8577 SDValue R = Op.getOperand(0);
8579 LLVMContext *Context = DAG.getContext();
8581 assert(Subtarget->hasSSE41() && "Cannot lower SHL without SSE4.1 or later");
8583 if (VT == MVT::v4i32) {
8584 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8585 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8586 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
8588 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
8590 std::vector<Constant*> CV(4, CI);
8591 Constant *C = ConstantVector::get(CV);
8592 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8593 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8594 MachinePointerInfo::getConstantPool(),
8597 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
8598 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
8599 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
8600 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
8602 if (VT == MVT::v16i8) {
8604 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8605 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8606 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
8608 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
8609 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
8611 std::vector<Constant*> CVM1(16, CM1);
8612 std::vector<Constant*> CVM2(16, CM2);
8613 Constant *C = ConstantVector::get(CVM1);
8614 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8615 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8616 MachinePointerInfo::getConstantPool(),
8619 // r = pblendv(r, psllw(r & (char16)15, 4), a);
8620 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8621 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8622 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8623 DAG.getConstant(4, MVT::i32));
8624 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8626 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8628 C = ConstantVector::get(CVM2);
8629 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8630 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8631 MachinePointerInfo::getConstantPool(),
8634 // r = pblendv(r, psllw(r & (char16)63, 2), a);
8635 M = DAG.getNode(ISD::AND, dl, VT, R, M);
8636 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8637 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
8638 DAG.getConstant(2, MVT::i32));
8639 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
8641 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
8643 // return pblendv(r, r+r, a);
8644 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
8645 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
8651 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
8652 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
8653 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
8654 // looks for this combo and may remove the "setcc" instruction if the "setcc"
8655 // has only one use.
8656 SDNode *N = Op.getNode();
8657 SDValue LHS = N->getOperand(0);
8658 SDValue RHS = N->getOperand(1);
8659 unsigned BaseOp = 0;
8661 DebugLoc DL = Op.getDebugLoc();
8662 switch (Op.getOpcode()) {
8663 default: llvm_unreachable("Unknown ovf instruction!");
8665 // A subtract of one will be selected as a INC. Note that INC doesn't
8666 // set CF, so we can't do this for UADDO.
8667 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8668 if (C->getAPIntValue() == 1) {
8669 BaseOp = X86ISD::INC;
8673 BaseOp = X86ISD::ADD;
8677 BaseOp = X86ISD::ADD;
8681 // A subtract of one will be selected as a DEC. Note that DEC doesn't
8682 // set CF, so we can't do this for USUBO.
8683 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
8684 if (C->getAPIntValue() == 1) {
8685 BaseOp = X86ISD::DEC;
8689 BaseOp = X86ISD::SUB;
8693 BaseOp = X86ISD::SUB;
8697 BaseOp = X86ISD::SMUL;
8700 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
8701 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
8703 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
8706 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8707 DAG.getConstant(X86::COND_O, MVT::i32),
8708 SDValue(Sum.getNode(), 2));
8710 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8715 // Also sets EFLAGS.
8716 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
8717 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
8720 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
8721 DAG.getConstant(Cond, MVT::i32),
8722 SDValue(Sum.getNode(), 1));
8724 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
8728 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
8729 DebugLoc dl = Op.getDebugLoc();
8731 if (!Subtarget->hasSSE2()) {
8732 SDValue Chain = Op.getOperand(0);
8733 SDValue Zero = DAG.getConstant(0,
8734 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8736 DAG.getRegister(X86::ESP, MVT::i32), // Base
8737 DAG.getTargetConstant(1, MVT::i8), // Scale
8738 DAG.getRegister(0, MVT::i32), // Index
8739 DAG.getTargetConstant(0, MVT::i32), // Disp
8740 DAG.getRegister(0, MVT::i32), // Segment.
8745 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
8746 array_lengthof(Ops));
8747 return SDValue(Res, 0);
8750 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
8752 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
8754 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
8755 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
8756 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
8757 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
8759 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
8760 if (!Op1 && !Op2 && !Op3 && Op4)
8761 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
8763 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
8764 if (Op1 && !Op2 && !Op3 && !Op4)
8765 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
8767 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
8769 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
8772 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
8773 EVT T = Op.getValueType();
8774 DebugLoc DL = Op.getDebugLoc();
8777 switch(T.getSimpleVT().SimpleTy) {
8779 assert(false && "Invalid value type!");
8780 case MVT::i8: Reg = X86::AL; size = 1; break;
8781 case MVT::i16: Reg = X86::AX; size = 2; break;
8782 case MVT::i32: Reg = X86::EAX; size = 4; break;
8784 assert(Subtarget->is64Bit() && "Node not type legal!");
8785 Reg = X86::RAX; size = 8;
8788 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
8789 Op.getOperand(2), SDValue());
8790 SDValue Ops[] = { cpIn.getValue(0),
8793 DAG.getTargetConstant(size, MVT::i8),
8795 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
8796 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
8797 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
8800 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
8804 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
8805 SelectionDAG &DAG) const {
8806 assert(Subtarget->is64Bit() && "Result not type legalized?");
8807 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
8808 SDValue TheChain = Op.getOperand(0);
8809 DebugLoc dl = Op.getDebugLoc();
8810 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
8811 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
8812 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
8814 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
8815 DAG.getConstant(32, MVT::i8));
8817 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
8820 return DAG.getMergeValues(Ops, 2, dl);
8823 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
8824 SelectionDAG &DAG) const {
8825 EVT SrcVT = Op.getOperand(0).getValueType();
8826 EVT DstVT = Op.getValueType();
8827 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
8828 Subtarget->hasMMX() && "Unexpected custom BITCAST");
8829 assert((DstVT == MVT::i64 ||
8830 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
8831 "Unexpected custom BITCAST");
8832 // i64 <=> MMX conversions are Legal.
8833 if (SrcVT==MVT::i64 && DstVT.isVector())
8835 if (DstVT==MVT::i64 && SrcVT.isVector())
8837 // MMX <=> MMX conversions are Legal.
8838 if (SrcVT.isVector() && DstVT.isVector())
8840 // All other conversions need to be expanded.
8844 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
8845 SDNode *Node = Op.getNode();
8846 DebugLoc dl = Node->getDebugLoc();
8847 EVT T = Node->getValueType(0);
8848 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
8849 DAG.getConstant(0, T), Node->getOperand(2));
8850 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
8851 cast<AtomicSDNode>(Node)->getMemoryVT(),
8852 Node->getOperand(0),
8853 Node->getOperand(1), negOp,
8854 cast<AtomicSDNode>(Node)->getSrcValue(),
8855 cast<AtomicSDNode>(Node)->getAlignment());
8858 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
8859 EVT VT = Op.getNode()->getValueType(0);
8861 // Let legalize expand this if it isn't a legal type yet.
8862 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
8865 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
8868 bool ExtraOp = false;
8869 switch (Op.getOpcode()) {
8870 default: assert(0 && "Invalid code");
8871 case ISD::ADDC: Opc = X86ISD::ADD; break;
8872 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
8873 case ISD::SUBC: Opc = X86ISD::SUB; break;
8874 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
8878 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
8880 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
8881 Op.getOperand(1), Op.getOperand(2));
8884 /// LowerOperation - Provide custom lowering hooks for some operations.
8886 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
8887 switch (Op.getOpcode()) {
8888 default: llvm_unreachable("Should not custom lower this!");
8889 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
8890 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
8891 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
8892 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
8893 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
8894 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
8895 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
8896 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
8897 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
8898 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
8899 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
8900 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
8901 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
8902 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
8903 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
8904 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
8905 case ISD::SHL_PARTS:
8906 case ISD::SRA_PARTS:
8907 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
8908 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
8909 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
8910 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
8911 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
8912 case ISD::FABS: return LowerFABS(Op, DAG);
8913 case ISD::FNEG: return LowerFNEG(Op, DAG);
8914 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
8915 case ISD::SETCC: return LowerSETCC(Op, DAG);
8916 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
8917 case ISD::SELECT: return LowerSELECT(Op, DAG);
8918 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
8919 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
8920 case ISD::VASTART: return LowerVASTART(Op, DAG);
8921 case ISD::VAARG: return LowerVAARG(Op, DAG);
8922 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
8923 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
8924 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
8925 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
8926 case ISD::FRAME_TO_ARGS_OFFSET:
8927 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
8928 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
8929 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
8930 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
8931 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
8932 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
8933 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
8934 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
8935 case ISD::SHL: return LowerSHL(Op, DAG);
8941 case ISD::UMULO: return LowerXALUO(Op, DAG);
8942 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
8943 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
8947 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
8951 void X86TargetLowering::
8952 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
8953 SelectionDAG &DAG, unsigned NewOp) const {
8954 EVT T = Node->getValueType(0);
8955 DebugLoc dl = Node->getDebugLoc();
8956 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
8958 SDValue Chain = Node->getOperand(0);
8959 SDValue In1 = Node->getOperand(1);
8960 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8961 Node->getOperand(2), DAG.getIntPtrConstant(0));
8962 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
8963 Node->getOperand(2), DAG.getIntPtrConstant(1));
8964 SDValue Ops[] = { Chain, In1, In2L, In2H };
8965 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
8967 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
8968 cast<MemSDNode>(Node)->getMemOperand());
8969 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
8970 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
8971 Results.push_back(Result.getValue(2));
8974 /// ReplaceNodeResults - Replace a node with an illegal result type
8975 /// with a new node built out of custom code.
8976 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
8977 SmallVectorImpl<SDValue>&Results,
8978 SelectionDAG &DAG) const {
8979 DebugLoc dl = N->getDebugLoc();
8980 switch (N->getOpcode()) {
8982 assert(false && "Do not know how to custom type legalize this operation!");
8988 // We don't want to expand or promote these.
8990 case ISD::FP_TO_SINT: {
8991 std::pair<SDValue,SDValue> Vals =
8992 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
8993 SDValue FIST = Vals.first, StackSlot = Vals.second;
8994 if (FIST.getNode() != 0) {
8995 EVT VT = N->getValueType(0);
8996 // Return a load from the stack slot.
8997 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
8998 MachinePointerInfo(), false, false, 0));
9002 case ISD::READCYCLECOUNTER: {
9003 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9004 SDValue TheChain = N->getOperand(0);
9005 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9006 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9008 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9010 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9011 SDValue Ops[] = { eax, edx };
9012 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9013 Results.push_back(edx.getValue(1));
9016 case ISD::ATOMIC_CMP_SWAP: {
9017 EVT T = N->getValueType(0);
9018 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9019 SDValue cpInL, cpInH;
9020 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9021 DAG.getConstant(0, MVT::i32));
9022 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9023 DAG.getConstant(1, MVT::i32));
9024 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9025 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9027 SDValue swapInL, swapInH;
9028 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9029 DAG.getConstant(0, MVT::i32));
9030 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9031 DAG.getConstant(1, MVT::i32));
9032 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9034 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9035 swapInL.getValue(1));
9036 SDValue Ops[] = { swapInH.getValue(0),
9038 swapInH.getValue(1) };
9039 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9040 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9041 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9043 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9044 MVT::i32, Result.getValue(1));
9045 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9046 MVT::i32, cpOutL.getValue(2));
9047 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9048 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9049 Results.push_back(cpOutH.getValue(1));
9052 case ISD::ATOMIC_LOAD_ADD:
9053 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9055 case ISD::ATOMIC_LOAD_AND:
9056 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9058 case ISD::ATOMIC_LOAD_NAND:
9059 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9061 case ISD::ATOMIC_LOAD_OR:
9062 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9064 case ISD::ATOMIC_LOAD_SUB:
9065 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9067 case ISD::ATOMIC_LOAD_XOR:
9068 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9070 case ISD::ATOMIC_SWAP:
9071 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9076 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9078 default: return NULL;
9079 case X86ISD::BSF: return "X86ISD::BSF";
9080 case X86ISD::BSR: return "X86ISD::BSR";
9081 case X86ISD::SHLD: return "X86ISD::SHLD";
9082 case X86ISD::SHRD: return "X86ISD::SHRD";
9083 case X86ISD::FAND: return "X86ISD::FAND";
9084 case X86ISD::FOR: return "X86ISD::FOR";
9085 case X86ISD::FXOR: return "X86ISD::FXOR";
9086 case X86ISD::FSRL: return "X86ISD::FSRL";
9087 case X86ISD::FILD: return "X86ISD::FILD";
9088 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9089 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9090 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9091 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9092 case X86ISD::FLD: return "X86ISD::FLD";
9093 case X86ISD::FST: return "X86ISD::FST";
9094 case X86ISD::CALL: return "X86ISD::CALL";
9095 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9096 case X86ISD::BT: return "X86ISD::BT";
9097 case X86ISD::CMP: return "X86ISD::CMP";
9098 case X86ISD::COMI: return "X86ISD::COMI";
9099 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9100 case X86ISD::SETCC: return "X86ISD::SETCC";
9101 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9102 case X86ISD::CMOV: return "X86ISD::CMOV";
9103 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9104 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9105 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9106 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9107 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9108 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9109 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9110 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9111 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9112 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9113 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9114 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9115 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9116 case X86ISD::PANDN: return "X86ISD::PANDN";
9117 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9118 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9119 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9120 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9121 case X86ISD::FMAX: return "X86ISD::FMAX";
9122 case X86ISD::FMIN: return "X86ISD::FMIN";
9123 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9124 case X86ISD::FRCP: return "X86ISD::FRCP";
9125 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9126 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9127 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9128 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9129 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9130 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9131 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9132 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9133 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9134 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9135 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9136 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9137 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9138 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9139 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9140 case X86ISD::VSHL: return "X86ISD::VSHL";
9141 case X86ISD::VSRL: return "X86ISD::VSRL";
9142 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9143 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9144 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9145 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9146 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9147 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9148 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9149 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9150 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9151 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9152 case X86ISD::ADD: return "X86ISD::ADD";
9153 case X86ISD::SUB: return "X86ISD::SUB";
9154 case X86ISD::ADC: return "X86ISD::ADC";
9155 case X86ISD::SBB: return "X86ISD::SBB";
9156 case X86ISD::SMUL: return "X86ISD::SMUL";
9157 case X86ISD::UMUL: return "X86ISD::UMUL";
9158 case X86ISD::INC: return "X86ISD::INC";
9159 case X86ISD::DEC: return "X86ISD::DEC";
9160 case X86ISD::OR: return "X86ISD::OR";
9161 case X86ISD::XOR: return "X86ISD::XOR";
9162 case X86ISD::AND: return "X86ISD::AND";
9163 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9164 case X86ISD::PTEST: return "X86ISD::PTEST";
9165 case X86ISD::TESTP: return "X86ISD::TESTP";
9166 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9167 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9168 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9169 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9170 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9171 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9172 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9173 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9174 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9175 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9176 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9177 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9178 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9179 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9180 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9181 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9182 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9183 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9184 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9185 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9186 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9187 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9188 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9189 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9190 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9191 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9192 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9193 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9194 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9195 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9196 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9197 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9198 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9199 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9200 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9201 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9205 // isLegalAddressingMode - Return true if the addressing mode represented
9206 // by AM is legal for this target, for a load/store of the specified type.
9207 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9208 const Type *Ty) const {
9209 // X86 supports extremely general addressing modes.
9210 CodeModel::Model M = getTargetMachine().getCodeModel();
9211 Reloc::Model R = getTargetMachine().getRelocationModel();
9213 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9214 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
9219 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
9221 // If a reference to this global requires an extra load, we can't fold it.
9222 if (isGlobalStubReference(GVFlags))
9225 // If BaseGV requires a register for the PIC base, we cannot also have a
9226 // BaseReg specified.
9227 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
9230 // If lower 4G is not available, then we must use rip-relative addressing.
9231 if ((M != CodeModel::Small || R != Reloc::Static) &&
9232 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
9242 // These scales always work.
9247 // These scales are formed with basereg+scalereg. Only accept if there is
9252 default: // Other stuff never works.
9260 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
9261 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
9263 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
9264 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
9265 if (NumBits1 <= NumBits2)
9270 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
9271 if (!VT1.isInteger() || !VT2.isInteger())
9273 unsigned NumBits1 = VT1.getSizeInBits();
9274 unsigned NumBits2 = VT2.getSizeInBits();
9275 if (NumBits1 <= NumBits2)
9280 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
9281 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9282 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
9285 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
9286 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
9287 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
9290 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
9291 // i16 instructions are longer (0x66 prefix) and potentially slower.
9292 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
9295 /// isShuffleMaskLegal - Targets can use this to indicate that they only
9296 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
9297 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
9298 /// are assumed to be legal.
9300 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
9302 // Very little shuffling can be done for 64-bit vectors right now.
9303 if (VT.getSizeInBits() == 64)
9304 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
9306 // FIXME: pshufb, blends, shifts.
9307 return (VT.getVectorNumElements() == 2 ||
9308 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
9309 isMOVLMask(M, VT) ||
9310 isSHUFPMask(M, VT) ||
9311 isPSHUFDMask(M, VT) ||
9312 isPSHUFHWMask(M, VT) ||
9313 isPSHUFLWMask(M, VT) ||
9314 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
9315 isUNPCKLMask(M, VT) ||
9316 isUNPCKHMask(M, VT) ||
9317 isUNPCKL_v_undef_Mask(M, VT) ||
9318 isUNPCKH_v_undef_Mask(M, VT));
9322 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
9324 unsigned NumElts = VT.getVectorNumElements();
9325 // FIXME: This collection of masks seems suspect.
9328 if (NumElts == 4 && VT.getSizeInBits() == 128) {
9329 return (isMOVLMask(Mask, VT) ||
9330 isCommutedMOVLMask(Mask, VT, true) ||
9331 isSHUFPMask(Mask, VT) ||
9332 isCommutedSHUFPMask(Mask, VT));
9337 //===----------------------------------------------------------------------===//
9338 // X86 Scheduler Hooks
9339 //===----------------------------------------------------------------------===//
9341 // private utility function
9343 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
9344 MachineBasicBlock *MBB,
9351 TargetRegisterClass *RC,
9352 bool invSrc) const {
9353 // For the atomic bitwise operator, we generate
9356 // ld t1 = [bitinstr.addr]
9357 // op t2 = t1, [bitinstr.val]
9359 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9361 // fallthrough -->nextMBB
9362 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9363 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9364 MachineFunction::iterator MBBIter = MBB;
9367 /// First build the CFG
9368 MachineFunction *F = MBB->getParent();
9369 MachineBasicBlock *thisMBB = MBB;
9370 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9371 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9372 F->insert(MBBIter, newMBB);
9373 F->insert(MBBIter, nextMBB);
9375 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9376 nextMBB->splice(nextMBB->begin(), thisMBB,
9377 llvm::next(MachineBasicBlock::iterator(bInstr)),
9379 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9381 // Update thisMBB to fall through to newMBB
9382 thisMBB->addSuccessor(newMBB);
9384 // newMBB jumps to itself and fall through to nextMBB
9385 newMBB->addSuccessor(nextMBB);
9386 newMBB->addSuccessor(newMBB);
9388 // Insert instructions into newMBB based on incoming instruction
9389 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9390 "unexpected number of operands");
9391 DebugLoc dl = bInstr->getDebugLoc();
9392 MachineOperand& destOper = bInstr->getOperand(0);
9393 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9394 int numArgs = bInstr->getNumOperands() - 1;
9395 for (int i=0; i < numArgs; ++i)
9396 argOpers[i] = &bInstr->getOperand(i+1);
9398 // x86 address has 4 operands: base, index, scale, and displacement
9399 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9400 int valArgIndx = lastAddrIndx + 1;
9402 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9403 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
9404 for (int i=0; i <= lastAddrIndx; ++i)
9405 (*MIB).addOperand(*argOpers[i]);
9407 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
9409 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
9414 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9415 assert((argOpers[valArgIndx]->isReg() ||
9416 argOpers[valArgIndx]->isImm()) &&
9418 if (argOpers[valArgIndx]->isReg())
9419 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
9421 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
9423 (*MIB).addOperand(*argOpers[valArgIndx]);
9425 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
9428 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
9429 for (int i=0; i <= lastAddrIndx; ++i)
9430 (*MIB).addOperand(*argOpers[i]);
9432 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9433 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9434 bInstr->memoperands_end());
9436 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9440 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9442 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9446 // private utility function: 64 bit atomics on 32 bit host.
9448 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
9449 MachineBasicBlock *MBB,
9454 bool invSrc) const {
9455 // For the atomic bitwise operator, we generate
9456 // thisMBB (instructions are in pairs, except cmpxchg8b)
9457 // ld t1,t2 = [bitinstr.addr]
9459 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
9460 // op t5, t6 <- out1, out2, [bitinstr.val]
9461 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
9462 // mov ECX, EBX <- t5, t6
9463 // mov EAX, EDX <- t1, t2
9464 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
9465 // mov t3, t4 <- EAX, EDX
9467 // result in out1, out2
9468 // fallthrough -->nextMBB
9470 const TargetRegisterClass *RC = X86::GR32RegisterClass;
9471 const unsigned LoadOpc = X86::MOV32rm;
9472 const unsigned NotOpc = X86::NOT32r;
9473 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9474 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9475 MachineFunction::iterator MBBIter = MBB;
9478 /// First build the CFG
9479 MachineFunction *F = MBB->getParent();
9480 MachineBasicBlock *thisMBB = MBB;
9481 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9482 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9483 F->insert(MBBIter, newMBB);
9484 F->insert(MBBIter, nextMBB);
9486 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9487 nextMBB->splice(nextMBB->begin(), thisMBB,
9488 llvm::next(MachineBasicBlock::iterator(bInstr)),
9490 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9492 // Update thisMBB to fall through to newMBB
9493 thisMBB->addSuccessor(newMBB);
9495 // newMBB jumps to itself and fall through to nextMBB
9496 newMBB->addSuccessor(nextMBB);
9497 newMBB->addSuccessor(newMBB);
9499 DebugLoc dl = bInstr->getDebugLoc();
9500 // Insert instructions into newMBB based on incoming instruction
9501 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
9502 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
9503 "unexpected number of operands");
9504 MachineOperand& dest1Oper = bInstr->getOperand(0);
9505 MachineOperand& dest2Oper = bInstr->getOperand(1);
9506 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9507 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
9508 argOpers[i] = &bInstr->getOperand(i+2);
9510 // We use some of the operands multiple times, so conservatively just
9511 // clear any kill flags that might be present.
9512 if (argOpers[i]->isReg() && argOpers[i]->isUse())
9513 argOpers[i]->setIsKill(false);
9516 // x86 address has 5 operands: base, index, scale, displacement, and segment.
9517 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9519 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
9520 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
9521 for (int i=0; i <= lastAddrIndx; ++i)
9522 (*MIB).addOperand(*argOpers[i]);
9523 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
9524 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
9525 // add 4 to displacement.
9526 for (int i=0; i <= lastAddrIndx-2; ++i)
9527 (*MIB).addOperand(*argOpers[i]);
9528 MachineOperand newOp3 = *(argOpers[3]);
9530 newOp3.setImm(newOp3.getImm()+4);
9532 newOp3.setOffset(newOp3.getOffset()+4);
9533 (*MIB).addOperand(newOp3);
9534 (*MIB).addOperand(*argOpers[lastAddrIndx]);
9536 // t3/4 are defined later, at the bottom of the loop
9537 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
9538 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
9539 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
9540 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
9541 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
9542 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
9544 // The subsequent operations should be using the destination registers of
9545 //the PHI instructions.
9547 t1 = F->getRegInfo().createVirtualRegister(RC);
9548 t2 = F->getRegInfo().createVirtualRegister(RC);
9549 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
9550 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
9552 t1 = dest1Oper.getReg();
9553 t2 = dest2Oper.getReg();
9556 int valArgIndx = lastAddrIndx + 1;
9557 assert((argOpers[valArgIndx]->isReg() ||
9558 argOpers[valArgIndx]->isImm()) &&
9560 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
9561 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
9562 if (argOpers[valArgIndx]->isReg())
9563 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
9565 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
9566 if (regOpcL != X86::MOV32rr)
9568 (*MIB).addOperand(*argOpers[valArgIndx]);
9569 assert(argOpers[valArgIndx + 1]->isReg() ==
9570 argOpers[valArgIndx]->isReg());
9571 assert(argOpers[valArgIndx + 1]->isImm() ==
9572 argOpers[valArgIndx]->isImm());
9573 if (argOpers[valArgIndx + 1]->isReg())
9574 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
9576 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
9577 if (regOpcH != X86::MOV32rr)
9579 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
9581 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9583 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
9586 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
9588 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
9591 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
9592 for (int i=0; i <= lastAddrIndx; ++i)
9593 (*MIB).addOperand(*argOpers[i]);
9595 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9596 (*MIB).setMemRefs(bInstr->memoperands_begin(),
9597 bInstr->memoperands_end());
9599 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
9600 MIB.addReg(X86::EAX);
9601 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
9602 MIB.addReg(X86::EDX);
9605 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9607 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
9611 // private utility function
9613 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
9614 MachineBasicBlock *MBB,
9615 unsigned cmovOpc) const {
9616 // For the atomic min/max operator, we generate
9619 // ld t1 = [min/max.addr]
9620 // mov t2 = [min/max.val]
9622 // cmov[cond] t2 = t1
9624 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
9626 // fallthrough -->nextMBB
9628 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9629 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9630 MachineFunction::iterator MBBIter = MBB;
9633 /// First build the CFG
9634 MachineFunction *F = MBB->getParent();
9635 MachineBasicBlock *thisMBB = MBB;
9636 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
9637 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
9638 F->insert(MBBIter, newMBB);
9639 F->insert(MBBIter, nextMBB);
9641 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
9642 nextMBB->splice(nextMBB->begin(), thisMBB,
9643 llvm::next(MachineBasicBlock::iterator(mInstr)),
9645 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9647 // Update thisMBB to fall through to newMBB
9648 thisMBB->addSuccessor(newMBB);
9650 // newMBB jumps to newMBB and fall through to nextMBB
9651 newMBB->addSuccessor(nextMBB);
9652 newMBB->addSuccessor(newMBB);
9654 DebugLoc dl = mInstr->getDebugLoc();
9655 // Insert instructions into newMBB based on incoming instruction
9656 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
9657 "unexpected number of operands");
9658 MachineOperand& destOper = mInstr->getOperand(0);
9659 MachineOperand* argOpers[2 + X86::AddrNumOperands];
9660 int numArgs = mInstr->getNumOperands() - 1;
9661 for (int i=0; i < numArgs; ++i)
9662 argOpers[i] = &mInstr->getOperand(i+1);
9664 // x86 address has 4 operands: base, index, scale, and displacement
9665 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
9666 int valArgIndx = lastAddrIndx + 1;
9668 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9669 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
9670 for (int i=0; i <= lastAddrIndx; ++i)
9671 (*MIB).addOperand(*argOpers[i]);
9673 // We only support register and immediate values
9674 assert((argOpers[valArgIndx]->isReg() ||
9675 argOpers[valArgIndx]->isImm()) &&
9678 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9679 if (argOpers[valArgIndx]->isReg())
9680 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
9682 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
9683 (*MIB).addOperand(*argOpers[valArgIndx]);
9685 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
9688 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
9693 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
9694 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
9698 // Cmp and exchange if none has modified the memory location
9699 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
9700 for (int i=0; i <= lastAddrIndx; ++i)
9701 (*MIB).addOperand(*argOpers[i]);
9703 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
9704 (*MIB).setMemRefs(mInstr->memoperands_begin(),
9705 mInstr->memoperands_end());
9707 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
9708 MIB.addReg(X86::EAX);
9711 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
9713 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
9717 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
9718 // or XMM0_V32I8 in AVX all of this code can be replaced with that
9721 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
9722 unsigned numArgs, bool memArg) const {
9723 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
9724 "Target must have SSE4.2 or AVX features enabled");
9726 DebugLoc dl = MI->getDebugLoc();
9727 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9729 if (!Subtarget->hasAVX()) {
9731 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
9733 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
9736 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
9738 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
9741 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
9742 for (unsigned i = 0; i < numArgs; ++i) {
9743 MachineOperand &Op = MI->getOperand(i+1);
9744 if (!(Op.isReg() && Op.isImplicit()))
9747 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
9750 MI->eraseFromParent();
9755 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
9756 DebugLoc dl = MI->getDebugLoc();
9757 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9759 // Address into RAX/EAX, other two args into ECX, EDX.
9760 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
9761 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
9762 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
9763 for (int i = 0; i < X86::AddrNumOperands; ++i)
9764 MIB.addOperand(MI->getOperand(i));
9766 unsigned ValOps = X86::AddrNumOperands;
9767 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9768 .addReg(MI->getOperand(ValOps).getReg());
9769 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
9770 .addReg(MI->getOperand(ValOps+1).getReg());
9772 // The instruction doesn't actually take any operands though.
9773 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
9775 MI->eraseFromParent(); // The pseudo is gone now.
9780 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
9781 DebugLoc dl = MI->getDebugLoc();
9782 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9784 // First arg in ECX, the second in EAX.
9785 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
9786 .addReg(MI->getOperand(0).getReg());
9787 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
9788 .addReg(MI->getOperand(1).getReg());
9790 // The instruction doesn't actually take any operands though.
9791 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
9793 MI->eraseFromParent(); // The pseudo is gone now.
9798 X86TargetLowering::EmitVAARG64WithCustomInserter(
9800 MachineBasicBlock *MBB) const {
9801 // Emit va_arg instruction on X86-64.
9803 // Operands to this pseudo-instruction:
9804 // 0 ) Output : destination address (reg)
9805 // 1-5) Input : va_list address (addr, i64mem)
9806 // 6 ) ArgSize : Size (in bytes) of vararg type
9807 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
9808 // 8 ) Align : Alignment of type
9809 // 9 ) EFLAGS (implicit-def)
9811 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
9812 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
9814 unsigned DestReg = MI->getOperand(0).getReg();
9815 MachineOperand &Base = MI->getOperand(1);
9816 MachineOperand &Scale = MI->getOperand(2);
9817 MachineOperand &Index = MI->getOperand(3);
9818 MachineOperand &Disp = MI->getOperand(4);
9819 MachineOperand &Segment = MI->getOperand(5);
9820 unsigned ArgSize = MI->getOperand(6).getImm();
9821 unsigned ArgMode = MI->getOperand(7).getImm();
9822 unsigned Align = MI->getOperand(8).getImm();
9825 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
9826 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
9827 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
9829 // Machine Information
9830 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
9831 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
9832 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
9833 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
9834 DebugLoc DL = MI->getDebugLoc();
9839 // i64 overflow_area (address)
9840 // i64 reg_save_area (address)
9842 // sizeof(va_list) = 24
9843 // alignment(va_list) = 8
9845 unsigned TotalNumIntRegs = 6;
9846 unsigned TotalNumXMMRegs = 8;
9847 bool UseGPOffset = (ArgMode == 1);
9848 bool UseFPOffset = (ArgMode == 2);
9849 unsigned MaxOffset = TotalNumIntRegs * 8 +
9850 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
9852 /* Align ArgSize to a multiple of 8 */
9853 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
9854 bool NeedsAlign = (Align > 8);
9856 MachineBasicBlock *thisMBB = MBB;
9857 MachineBasicBlock *overflowMBB;
9858 MachineBasicBlock *offsetMBB;
9859 MachineBasicBlock *endMBB;
9861 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
9862 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
9863 unsigned OffsetReg = 0;
9865 if (!UseGPOffset && !UseFPOffset) {
9866 // If we only pull from the overflow region, we don't create a branch.
9867 // We don't need to alter control flow.
9868 OffsetDestReg = 0; // unused
9869 OverflowDestReg = DestReg;
9872 overflowMBB = thisMBB;
9875 // First emit code to check if gp_offset (or fp_offset) is below the bound.
9876 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
9877 // If not, pull from overflow_area. (branch to overflowMBB)
9882 // offsetMBB overflowMBB
9887 // Registers for the PHI in endMBB
9888 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
9889 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
9891 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
9892 MachineFunction *MF = MBB->getParent();
9893 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9894 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9895 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
9897 MachineFunction::iterator MBBIter = MBB;
9900 // Insert the new basic blocks
9901 MF->insert(MBBIter, offsetMBB);
9902 MF->insert(MBBIter, overflowMBB);
9903 MF->insert(MBBIter, endMBB);
9905 // Transfer the remainder of MBB and its successor edges to endMBB.
9906 endMBB->splice(endMBB->begin(), thisMBB,
9907 llvm::next(MachineBasicBlock::iterator(MI)),
9909 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
9911 // Make offsetMBB and overflowMBB successors of thisMBB
9912 thisMBB->addSuccessor(offsetMBB);
9913 thisMBB->addSuccessor(overflowMBB);
9915 // endMBB is a successor of both offsetMBB and overflowMBB
9916 offsetMBB->addSuccessor(endMBB);
9917 overflowMBB->addSuccessor(endMBB);
9919 // Load the offset value into a register
9920 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
9921 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
9925 .addDisp(Disp, UseFPOffset ? 4 : 0)
9926 .addOperand(Segment)
9927 .setMemRefs(MMOBegin, MMOEnd);
9929 // Check if there is enough room left to pull this argument.
9930 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
9932 .addImm(MaxOffset + 8 - ArgSizeA8);
9934 // Branch to "overflowMBB" if offset >= max
9935 // Fall through to "offsetMBB" otherwise
9936 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
9937 .addMBB(overflowMBB);
9940 // In offsetMBB, emit code to use the reg_save_area.
9942 assert(OffsetReg != 0);
9944 // Read the reg_save_area address.
9945 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
9946 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
9951 .addOperand(Segment)
9952 .setMemRefs(MMOBegin, MMOEnd);
9954 // Zero-extend the offset
9955 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
9956 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
9959 .addImm(X86::sub_32bit);
9961 // Add the offset to the reg_save_area to get the final address.
9962 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
9963 .addReg(OffsetReg64)
9964 .addReg(RegSaveReg);
9966 // Compute the offset for the next argument
9967 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
9968 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
9970 .addImm(UseFPOffset ? 16 : 8);
9972 // Store it back into the va_list.
9973 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
9977 .addDisp(Disp, UseFPOffset ? 4 : 0)
9978 .addOperand(Segment)
9979 .addReg(NextOffsetReg)
9980 .setMemRefs(MMOBegin, MMOEnd);
9983 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
9988 // Emit code to use overflow area
9991 // Load the overflow_area address into a register.
9992 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
9993 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
9998 .addOperand(Segment)
9999 .setMemRefs(MMOBegin, MMOEnd);
10001 // If we need to align it, do so. Otherwise, just copy the address
10002 // to OverflowDestReg.
10004 // Align the overflow address
10005 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10006 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10008 // aligned_addr = (addr + (align-1)) & ~(align-1)
10009 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10010 .addReg(OverflowAddrReg)
10013 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10015 .addImm(~(uint64_t)(Align-1));
10017 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10018 .addReg(OverflowAddrReg);
10021 // Compute the next overflow address after this argument.
10022 // (the overflow address should be kept 8-byte aligned)
10023 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10024 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10025 .addReg(OverflowDestReg)
10026 .addImm(ArgSizeA8);
10028 // Store the new overflow address.
10029 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10034 .addOperand(Segment)
10035 .addReg(NextAddrReg)
10036 .setMemRefs(MMOBegin, MMOEnd);
10038 // If we branched, emit the PHI to the front of endMBB.
10040 BuildMI(*endMBB, endMBB->begin(), DL,
10041 TII->get(X86::PHI), DestReg)
10042 .addReg(OffsetDestReg).addMBB(offsetMBB)
10043 .addReg(OverflowDestReg).addMBB(overflowMBB);
10046 // Erase the pseudo instruction
10047 MI->eraseFromParent();
10052 MachineBasicBlock *
10053 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10055 MachineBasicBlock *MBB) const {
10056 // Emit code to save XMM registers to the stack. The ABI says that the
10057 // number of registers to save is given in %al, so it's theoretically
10058 // possible to do an indirect jump trick to avoid saving all of them,
10059 // however this code takes a simpler approach and just executes all
10060 // of the stores if %al is non-zero. It's less code, and it's probably
10061 // easier on the hardware branch predictor, and stores aren't all that
10062 // expensive anyway.
10064 // Create the new basic blocks. One block contains all the XMM stores,
10065 // and one block is the final destination regardless of whether any
10066 // stores were performed.
10067 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10068 MachineFunction *F = MBB->getParent();
10069 MachineFunction::iterator MBBIter = MBB;
10071 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10072 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10073 F->insert(MBBIter, XMMSaveMBB);
10074 F->insert(MBBIter, EndMBB);
10076 // Transfer the remainder of MBB and its successor edges to EndMBB.
10077 EndMBB->splice(EndMBB->begin(), MBB,
10078 llvm::next(MachineBasicBlock::iterator(MI)),
10080 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10082 // The original block will now fall through to the XMM save block.
10083 MBB->addSuccessor(XMMSaveMBB);
10084 // The XMMSaveMBB will fall through to the end block.
10085 XMMSaveMBB->addSuccessor(EndMBB);
10087 // Now add the instructions.
10088 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10089 DebugLoc DL = MI->getDebugLoc();
10091 unsigned CountReg = MI->getOperand(0).getReg();
10092 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10093 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10095 if (!Subtarget->isTargetWin64()) {
10096 // If %al is 0, branch around the XMM save block.
10097 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10098 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10099 MBB->addSuccessor(EndMBB);
10102 // In the XMM save block, save all the XMM argument registers.
10103 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10104 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10105 MachineMemOperand *MMO =
10106 F->getMachineMemOperand(
10107 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10108 MachineMemOperand::MOStore,
10109 /*Size=*/16, /*Align=*/16);
10110 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10111 .addFrameIndex(RegSaveFrameIndex)
10112 .addImm(/*Scale=*/1)
10113 .addReg(/*IndexReg=*/0)
10114 .addImm(/*Disp=*/Offset)
10115 .addReg(/*Segment=*/0)
10116 .addReg(MI->getOperand(i).getReg())
10117 .addMemOperand(MMO);
10120 MI->eraseFromParent(); // The pseudo instruction is gone now.
10125 MachineBasicBlock *
10126 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10127 MachineBasicBlock *BB) const {
10128 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10129 DebugLoc DL = MI->getDebugLoc();
10131 // To "insert" a SELECT_CC instruction, we actually have to insert the
10132 // diamond control-flow pattern. The incoming instruction knows the
10133 // destination vreg to set, the condition code register to branch on, the
10134 // true/false values to select between, and a branch opcode to use.
10135 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10136 MachineFunction::iterator It = BB;
10142 // cmpTY ccX, r1, r2
10144 // fallthrough --> copy0MBB
10145 MachineBasicBlock *thisMBB = BB;
10146 MachineFunction *F = BB->getParent();
10147 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10148 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10149 F->insert(It, copy0MBB);
10150 F->insert(It, sinkMBB);
10152 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10153 // live into the sink and copy blocks.
10154 const MachineFunction *MF = BB->getParent();
10155 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10156 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10158 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10159 const MachineOperand &MO = MI->getOperand(I);
10160 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10161 unsigned Reg = MO.getReg();
10162 if (Reg != X86::EFLAGS) continue;
10163 copy0MBB->addLiveIn(Reg);
10164 sinkMBB->addLiveIn(Reg);
10167 // Transfer the remainder of BB and its successor edges to sinkMBB.
10168 sinkMBB->splice(sinkMBB->begin(), BB,
10169 llvm::next(MachineBasicBlock::iterator(MI)),
10171 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10173 // Add the true and fallthrough blocks as its successors.
10174 BB->addSuccessor(copy0MBB);
10175 BB->addSuccessor(sinkMBB);
10177 // Create the conditional branch instruction.
10179 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10180 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10183 // %FalseValue = ...
10184 // # fallthrough to sinkMBB
10185 copy0MBB->addSuccessor(sinkMBB);
10188 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10190 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10191 TII->get(X86::PHI), MI->getOperand(0).getReg())
10192 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10193 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10195 MI->eraseFromParent(); // The pseudo instruction is gone now.
10199 MachineBasicBlock *
10200 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10201 MachineBasicBlock *BB) const {
10202 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10203 DebugLoc DL = MI->getDebugLoc();
10205 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10206 // non-trivial part is impdef of ESP.
10207 // FIXME: The code should be tweaked as soon as we'll try to do codegen for
10210 const char *StackProbeSymbol =
10211 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
10213 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
10214 .addExternalSymbol(StackProbeSymbol)
10215 .addReg(X86::EAX, RegState::Implicit)
10216 .addReg(X86::ESP, RegState::Implicit)
10217 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
10218 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
10219 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
10221 MI->eraseFromParent(); // The pseudo instruction is gone now.
10225 MachineBasicBlock *
10226 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
10227 MachineBasicBlock *BB) const {
10228 // This is pretty easy. We're taking the value that we received from
10229 // our load from the relocation, sticking it in either RDI (x86-64)
10230 // or EAX and doing an indirect call. The return value will then
10231 // be in the normal return register.
10232 const X86InstrInfo *TII
10233 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
10234 DebugLoc DL = MI->getDebugLoc();
10235 MachineFunction *F = BB->getParent();
10237 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
10238 assert(MI->getOperand(3).isGlobal() && "This should be a global");
10240 if (Subtarget->is64Bit()) {
10241 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10242 TII->get(X86::MOV64rm), X86::RDI)
10244 .addImm(0).addReg(0)
10245 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10246 MI->getOperand(3).getTargetFlags())
10248 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
10249 addDirectMem(MIB, X86::RDI);
10250 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
10251 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10252 TII->get(X86::MOV32rm), X86::EAX)
10254 .addImm(0).addReg(0)
10255 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10256 MI->getOperand(3).getTargetFlags())
10258 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10259 addDirectMem(MIB, X86::EAX);
10261 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
10262 TII->get(X86::MOV32rm), X86::EAX)
10263 .addReg(TII->getGlobalBaseReg(F))
10264 .addImm(0).addReg(0)
10265 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
10266 MI->getOperand(3).getTargetFlags())
10268 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
10269 addDirectMem(MIB, X86::EAX);
10272 MI->eraseFromParent(); // The pseudo instruction is gone now.
10276 MachineBasicBlock *
10277 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
10278 MachineBasicBlock *BB) const {
10279 switch (MI->getOpcode()) {
10280 default: assert(false && "Unexpected instr type to insert");
10281 case X86::TAILJMPd64:
10282 case X86::TAILJMPr64:
10283 case X86::TAILJMPm64:
10284 assert(!"TAILJMP64 would not be touched here.");
10285 case X86::TCRETURNdi64:
10286 case X86::TCRETURNri64:
10287 case X86::TCRETURNmi64:
10288 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
10289 // On AMD64, additional defs should be added before register allocation.
10290 if (!Subtarget->isTargetWin64()) {
10291 MI->addRegisterDefined(X86::RSI);
10292 MI->addRegisterDefined(X86::RDI);
10293 MI->addRegisterDefined(X86::XMM6);
10294 MI->addRegisterDefined(X86::XMM7);
10295 MI->addRegisterDefined(X86::XMM8);
10296 MI->addRegisterDefined(X86::XMM9);
10297 MI->addRegisterDefined(X86::XMM10);
10298 MI->addRegisterDefined(X86::XMM11);
10299 MI->addRegisterDefined(X86::XMM12);
10300 MI->addRegisterDefined(X86::XMM13);
10301 MI->addRegisterDefined(X86::XMM14);
10302 MI->addRegisterDefined(X86::XMM15);
10305 case X86::WIN_ALLOCA:
10306 return EmitLoweredWinAlloca(MI, BB);
10307 case X86::TLSCall_32:
10308 case X86::TLSCall_64:
10309 return EmitLoweredTLSCall(MI, BB);
10310 case X86::CMOV_GR8:
10311 case X86::CMOV_FR32:
10312 case X86::CMOV_FR64:
10313 case X86::CMOV_V4F32:
10314 case X86::CMOV_V2F64:
10315 case X86::CMOV_V2I64:
10316 case X86::CMOV_GR16:
10317 case X86::CMOV_GR32:
10318 case X86::CMOV_RFP32:
10319 case X86::CMOV_RFP64:
10320 case X86::CMOV_RFP80:
10321 return EmitLoweredSelect(MI, BB);
10323 case X86::FP32_TO_INT16_IN_MEM:
10324 case X86::FP32_TO_INT32_IN_MEM:
10325 case X86::FP32_TO_INT64_IN_MEM:
10326 case X86::FP64_TO_INT16_IN_MEM:
10327 case X86::FP64_TO_INT32_IN_MEM:
10328 case X86::FP64_TO_INT64_IN_MEM:
10329 case X86::FP80_TO_INT16_IN_MEM:
10330 case X86::FP80_TO_INT32_IN_MEM:
10331 case X86::FP80_TO_INT64_IN_MEM: {
10332 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10333 DebugLoc DL = MI->getDebugLoc();
10335 // Change the floating point control register to use "round towards zero"
10336 // mode when truncating to an integer value.
10337 MachineFunction *F = BB->getParent();
10338 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
10339 addFrameReference(BuildMI(*BB, MI, DL,
10340 TII->get(X86::FNSTCW16m)), CWFrameIdx);
10342 // Load the old value of the high byte of the control word...
10344 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
10345 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
10348 // Set the high part to be round to zero...
10349 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
10352 // Reload the modified control word now...
10353 addFrameReference(BuildMI(*BB, MI, DL,
10354 TII->get(X86::FLDCW16m)), CWFrameIdx);
10356 // Restore the memory image of control word to original value
10357 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
10360 // Get the X86 opcode to use.
10362 switch (MI->getOpcode()) {
10363 default: llvm_unreachable("illegal opcode!");
10364 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
10365 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
10366 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
10367 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
10368 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
10369 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
10370 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
10371 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
10372 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
10376 MachineOperand &Op = MI->getOperand(0);
10378 AM.BaseType = X86AddressMode::RegBase;
10379 AM.Base.Reg = Op.getReg();
10381 AM.BaseType = X86AddressMode::FrameIndexBase;
10382 AM.Base.FrameIndex = Op.getIndex();
10384 Op = MI->getOperand(1);
10386 AM.Scale = Op.getImm();
10387 Op = MI->getOperand(2);
10389 AM.IndexReg = Op.getImm();
10390 Op = MI->getOperand(3);
10391 if (Op.isGlobal()) {
10392 AM.GV = Op.getGlobal();
10394 AM.Disp = Op.getImm();
10396 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
10397 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
10399 // Reload the original control word now.
10400 addFrameReference(BuildMI(*BB, MI, DL,
10401 TII->get(X86::FLDCW16m)), CWFrameIdx);
10403 MI->eraseFromParent(); // The pseudo instruction is gone now.
10406 // String/text processing lowering.
10407 case X86::PCMPISTRM128REG:
10408 case X86::VPCMPISTRM128REG:
10409 return EmitPCMP(MI, BB, 3, false /* in-mem */);
10410 case X86::PCMPISTRM128MEM:
10411 case X86::VPCMPISTRM128MEM:
10412 return EmitPCMP(MI, BB, 3, true /* in-mem */);
10413 case X86::PCMPESTRM128REG:
10414 case X86::VPCMPESTRM128REG:
10415 return EmitPCMP(MI, BB, 5, false /* in mem */);
10416 case X86::PCMPESTRM128MEM:
10417 case X86::VPCMPESTRM128MEM:
10418 return EmitPCMP(MI, BB, 5, true /* in mem */);
10420 // Thread synchronization.
10422 return EmitMonitor(MI, BB);
10424 return EmitMwait(MI, BB);
10426 // Atomic Lowering.
10427 case X86::ATOMAND32:
10428 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10429 X86::AND32ri, X86::MOV32rm,
10431 X86::NOT32r, X86::EAX,
10432 X86::GR32RegisterClass);
10433 case X86::ATOMOR32:
10434 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
10435 X86::OR32ri, X86::MOV32rm,
10437 X86::NOT32r, X86::EAX,
10438 X86::GR32RegisterClass);
10439 case X86::ATOMXOR32:
10440 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
10441 X86::XOR32ri, X86::MOV32rm,
10443 X86::NOT32r, X86::EAX,
10444 X86::GR32RegisterClass);
10445 case X86::ATOMNAND32:
10446 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
10447 X86::AND32ri, X86::MOV32rm,
10449 X86::NOT32r, X86::EAX,
10450 X86::GR32RegisterClass, true);
10451 case X86::ATOMMIN32:
10452 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
10453 case X86::ATOMMAX32:
10454 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
10455 case X86::ATOMUMIN32:
10456 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
10457 case X86::ATOMUMAX32:
10458 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
10460 case X86::ATOMAND16:
10461 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10462 X86::AND16ri, X86::MOV16rm,
10464 X86::NOT16r, X86::AX,
10465 X86::GR16RegisterClass);
10466 case X86::ATOMOR16:
10467 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
10468 X86::OR16ri, X86::MOV16rm,
10470 X86::NOT16r, X86::AX,
10471 X86::GR16RegisterClass);
10472 case X86::ATOMXOR16:
10473 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
10474 X86::XOR16ri, X86::MOV16rm,
10476 X86::NOT16r, X86::AX,
10477 X86::GR16RegisterClass);
10478 case X86::ATOMNAND16:
10479 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
10480 X86::AND16ri, X86::MOV16rm,
10482 X86::NOT16r, X86::AX,
10483 X86::GR16RegisterClass, true);
10484 case X86::ATOMMIN16:
10485 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
10486 case X86::ATOMMAX16:
10487 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
10488 case X86::ATOMUMIN16:
10489 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
10490 case X86::ATOMUMAX16:
10491 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
10493 case X86::ATOMAND8:
10494 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10495 X86::AND8ri, X86::MOV8rm,
10497 X86::NOT8r, X86::AL,
10498 X86::GR8RegisterClass);
10500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
10501 X86::OR8ri, X86::MOV8rm,
10503 X86::NOT8r, X86::AL,
10504 X86::GR8RegisterClass);
10505 case X86::ATOMXOR8:
10506 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
10507 X86::XOR8ri, X86::MOV8rm,
10509 X86::NOT8r, X86::AL,
10510 X86::GR8RegisterClass);
10511 case X86::ATOMNAND8:
10512 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
10513 X86::AND8ri, X86::MOV8rm,
10515 X86::NOT8r, X86::AL,
10516 X86::GR8RegisterClass, true);
10517 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
10518 // This group is for 64-bit host.
10519 case X86::ATOMAND64:
10520 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10521 X86::AND64ri32, X86::MOV64rm,
10523 X86::NOT64r, X86::RAX,
10524 X86::GR64RegisterClass);
10525 case X86::ATOMOR64:
10526 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
10527 X86::OR64ri32, X86::MOV64rm,
10529 X86::NOT64r, X86::RAX,
10530 X86::GR64RegisterClass);
10531 case X86::ATOMXOR64:
10532 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
10533 X86::XOR64ri32, X86::MOV64rm,
10535 X86::NOT64r, X86::RAX,
10536 X86::GR64RegisterClass);
10537 case X86::ATOMNAND64:
10538 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
10539 X86::AND64ri32, X86::MOV64rm,
10541 X86::NOT64r, X86::RAX,
10542 X86::GR64RegisterClass, true);
10543 case X86::ATOMMIN64:
10544 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
10545 case X86::ATOMMAX64:
10546 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
10547 case X86::ATOMUMIN64:
10548 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
10549 case X86::ATOMUMAX64:
10550 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
10552 // This group does 64-bit operations on a 32-bit host.
10553 case X86::ATOMAND6432:
10554 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10555 X86::AND32rr, X86::AND32rr,
10556 X86::AND32ri, X86::AND32ri,
10558 case X86::ATOMOR6432:
10559 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10560 X86::OR32rr, X86::OR32rr,
10561 X86::OR32ri, X86::OR32ri,
10563 case X86::ATOMXOR6432:
10564 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10565 X86::XOR32rr, X86::XOR32rr,
10566 X86::XOR32ri, X86::XOR32ri,
10568 case X86::ATOMNAND6432:
10569 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10570 X86::AND32rr, X86::AND32rr,
10571 X86::AND32ri, X86::AND32ri,
10573 case X86::ATOMADD6432:
10574 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10575 X86::ADD32rr, X86::ADC32rr,
10576 X86::ADD32ri, X86::ADC32ri,
10578 case X86::ATOMSUB6432:
10579 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10580 X86::SUB32rr, X86::SBB32rr,
10581 X86::SUB32ri, X86::SBB32ri,
10583 case X86::ATOMSWAP6432:
10584 return EmitAtomicBit6432WithCustomInserter(MI, BB,
10585 X86::MOV32rr, X86::MOV32rr,
10586 X86::MOV32ri, X86::MOV32ri,
10588 case X86::VASTART_SAVE_XMM_REGS:
10589 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
10591 case X86::VAARG_64:
10592 return EmitVAARG64WithCustomInserter(MI, BB);
10596 //===----------------------------------------------------------------------===//
10597 // X86 Optimization Hooks
10598 //===----------------------------------------------------------------------===//
10600 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
10604 const SelectionDAG &DAG,
10605 unsigned Depth) const {
10606 unsigned Opc = Op.getOpcode();
10607 assert((Opc >= ISD::BUILTIN_OP_END ||
10608 Opc == ISD::INTRINSIC_WO_CHAIN ||
10609 Opc == ISD::INTRINSIC_W_CHAIN ||
10610 Opc == ISD::INTRINSIC_VOID) &&
10611 "Should use MaskedValueIsZero if you don't know whether Op"
10612 " is a target node!");
10614 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
10628 // These nodes' second result is a boolean.
10629 if (Op.getResNo() == 0)
10632 case X86ISD::SETCC:
10633 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
10634 Mask.getBitWidth() - 1);
10639 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
10640 unsigned Depth) const {
10641 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
10642 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
10643 return Op.getValueType().getScalarType().getSizeInBits();
10649 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
10650 /// node is a GlobalAddress + offset.
10651 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
10652 const GlobalValue* &GA,
10653 int64_t &Offset) const {
10654 if (N->getOpcode() == X86ISD::Wrapper) {
10655 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
10656 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
10657 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
10661 return TargetLowering::isGAPlusOffset(N, GA, Offset);
10664 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
10665 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
10666 /// if the load addresses are consecutive, non-overlapping, and in the right
10668 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
10669 TargetLowering::DAGCombinerInfo &DCI) {
10670 DebugLoc dl = N->getDebugLoc();
10671 EVT VT = N->getValueType(0);
10673 if (VT.getSizeInBits() != 128)
10676 // Don't create instructions with illegal types after legalize types has run.
10677 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
10678 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
10681 SmallVector<SDValue, 16> Elts;
10682 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
10683 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
10685 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
10688 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
10689 /// generation and convert it from being a bunch of shuffles and extracts
10690 /// to a simple store and scalar loads to extract the elements.
10691 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
10692 const TargetLowering &TLI) {
10693 SDValue InputVector = N->getOperand(0);
10695 // Only operate on vectors of 4 elements, where the alternative shuffling
10696 // gets to be more expensive.
10697 if (InputVector.getValueType() != MVT::v4i32)
10700 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
10701 // single use which is a sign-extend or zero-extend, and all elements are
10703 SmallVector<SDNode *, 4> Uses;
10704 unsigned ExtractedElements = 0;
10705 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
10706 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
10707 if (UI.getUse().getResNo() != InputVector.getResNo())
10710 SDNode *Extract = *UI;
10711 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10714 if (Extract->getValueType(0) != MVT::i32)
10716 if (!Extract->hasOneUse())
10718 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
10719 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
10721 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
10724 // Record which element was extracted.
10725 ExtractedElements |=
10726 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
10728 Uses.push_back(Extract);
10731 // If not all the elements were used, this may not be worthwhile.
10732 if (ExtractedElements != 15)
10735 // Ok, we've now decided to do the transformation.
10736 DebugLoc dl = InputVector.getDebugLoc();
10738 // Store the value to a temporary stack slot.
10739 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
10740 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
10741 MachinePointerInfo(), false, false, 0);
10743 // Replace each use (extract) with a load of the appropriate element.
10744 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
10745 UE = Uses.end(); UI != UE; ++UI) {
10746 SDNode *Extract = *UI;
10748 // Compute the element's address.
10749 SDValue Idx = Extract->getOperand(1);
10751 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
10752 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
10753 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
10755 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(),
10756 StackPtr, OffsetVal);
10758 // Load the scalar.
10759 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
10760 ScalarAddr, MachinePointerInfo(),
10763 // Replace the exact with the load.
10764 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
10767 // The replacement was made in place; don't return anything.
10771 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
10772 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
10773 const X86Subtarget *Subtarget) {
10774 DebugLoc DL = N->getDebugLoc();
10775 SDValue Cond = N->getOperand(0);
10776 // Get the LHS/RHS of the select.
10777 SDValue LHS = N->getOperand(1);
10778 SDValue RHS = N->getOperand(2);
10780 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
10781 // instructions match the semantics of the common C idiom x<y?x:y but not
10782 // x<=y?x:y, because of how they handle negative zero (which can be
10783 // ignored in unsafe-math mode).
10784 if (Subtarget->hasSSE2() &&
10785 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
10786 Cond.getOpcode() == ISD::SETCC) {
10787 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
10789 unsigned Opcode = 0;
10790 // Check for x CC y ? x : y.
10791 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
10792 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
10796 // Converting this to a min would handle NaNs incorrectly, and swapping
10797 // the operands would cause it to handle comparisons between positive
10798 // and negative zero incorrectly.
10799 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10800 if (!UnsafeFPMath &&
10801 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10803 std::swap(LHS, RHS);
10805 Opcode = X86ISD::FMIN;
10808 // Converting this to a min would handle comparisons between positive
10809 // and negative zero incorrectly.
10810 if (!UnsafeFPMath &&
10811 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
10813 Opcode = X86ISD::FMIN;
10816 // Converting this to a min would handle both negative zeros and NaNs
10817 // incorrectly, but we can swap the operands to fix both.
10818 std::swap(LHS, RHS);
10822 Opcode = X86ISD::FMIN;
10826 // Converting this to a max would handle comparisons between positive
10827 // and negative zero incorrectly.
10828 if (!UnsafeFPMath &&
10829 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
10831 Opcode = X86ISD::FMAX;
10834 // Converting this to a max would handle NaNs incorrectly, and swapping
10835 // the operands would cause it to handle comparisons between positive
10836 // and negative zero incorrectly.
10837 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
10838 if (!UnsafeFPMath &&
10839 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
10841 std::swap(LHS, RHS);
10843 Opcode = X86ISD::FMAX;
10846 // Converting this to a max would handle both negative zeros and NaNs
10847 // incorrectly, but we can swap the operands to fix both.
10848 std::swap(LHS, RHS);
10852 Opcode = X86ISD::FMAX;
10855 // Check for x CC y ? y : x -- a min/max with reversed arms.
10856 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
10857 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
10861 // Converting this to a min would handle comparisons between positive
10862 // and negative zero incorrectly, and swapping the operands would
10863 // cause it to handle NaNs incorrectly.
10864 if (!UnsafeFPMath &&
10865 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
10866 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10868 std::swap(LHS, RHS);
10870 Opcode = X86ISD::FMIN;
10873 // Converting this to a min would handle NaNs incorrectly.
10874 if (!UnsafeFPMath &&
10875 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
10877 Opcode = X86ISD::FMIN;
10880 // Converting this to a min would handle both negative zeros and NaNs
10881 // incorrectly, but we can swap the operands to fix both.
10882 std::swap(LHS, RHS);
10886 Opcode = X86ISD::FMIN;
10890 // Converting this to a max would handle NaNs incorrectly.
10891 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10893 Opcode = X86ISD::FMAX;
10896 // Converting this to a max would handle comparisons between positive
10897 // and negative zero incorrectly, and swapping the operands would
10898 // cause it to handle NaNs incorrectly.
10899 if (!UnsafeFPMath &&
10900 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
10901 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
10903 std::swap(LHS, RHS);
10905 Opcode = X86ISD::FMAX;
10908 // Converting this to a max would handle both negative zeros and NaNs
10909 // incorrectly, but we can swap the operands to fix both.
10910 std::swap(LHS, RHS);
10914 Opcode = X86ISD::FMAX;
10920 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
10923 // If this is a select between two integer constants, try to do some
10925 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
10926 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
10927 // Don't do this for crazy integer types.
10928 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
10929 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
10930 // so that TrueC (the true value) is larger than FalseC.
10931 bool NeedsCondInvert = false;
10933 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
10934 // Efficiently invertible.
10935 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
10936 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
10937 isa<ConstantSDNode>(Cond.getOperand(1))))) {
10938 NeedsCondInvert = true;
10939 std::swap(TrueC, FalseC);
10942 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
10943 if (FalseC->getAPIntValue() == 0 &&
10944 TrueC->getAPIntValue().isPowerOf2()) {
10945 if (NeedsCondInvert) // Invert the condition if needed.
10946 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10947 DAG.getConstant(1, Cond.getValueType()));
10949 // Zero extend the condition if needed.
10950 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
10952 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
10953 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
10954 DAG.getConstant(ShAmt, MVT::i8));
10957 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
10958 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
10959 if (NeedsCondInvert) // Invert the condition if needed.
10960 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10961 DAG.getConstant(1, Cond.getValueType()));
10963 // Zero extend the condition if needed.
10964 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
10965 FalseC->getValueType(0), Cond);
10966 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
10967 SDValue(FalseC, 0));
10970 // Optimize cases that will turn into an LEA instruction. This requires
10971 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
10972 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
10973 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
10974 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
10976 bool isFastMultiplier = false;
10978 switch ((unsigned char)Diff) {
10980 case 1: // result = add base, cond
10981 case 2: // result = lea base( , cond*2)
10982 case 3: // result = lea base(cond, cond*2)
10983 case 4: // result = lea base( , cond*4)
10984 case 5: // result = lea base(cond, cond*4)
10985 case 8: // result = lea base( , cond*8)
10986 case 9: // result = lea base(cond, cond*8)
10987 isFastMultiplier = true;
10992 if (isFastMultiplier) {
10993 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
10994 if (NeedsCondInvert) // Invert the condition if needed.
10995 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
10996 DAG.getConstant(1, Cond.getValueType()));
10998 // Zero extend the condition if needed.
10999 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11001 // Scale the condition by the difference.
11003 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11004 DAG.getConstant(Diff, Cond.getValueType()));
11006 // Add the base if non-zero.
11007 if (FalseC->getAPIntValue() != 0)
11008 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11009 SDValue(FalseC, 0));
11019 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11020 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11021 TargetLowering::DAGCombinerInfo &DCI) {
11022 DebugLoc DL = N->getDebugLoc();
11024 // If the flag operand isn't dead, don't touch this CMOV.
11025 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11028 // If this is a select between two integer constants, try to do some
11029 // optimizations. Note that the operands are ordered the opposite of SELECT
11031 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
11032 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
11033 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11034 // larger than FalseC (the false value).
11035 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11037 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11038 CC = X86::GetOppositeBranchCondition(CC);
11039 std::swap(TrueC, FalseC);
11042 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11043 // This is efficient for any integer data type (including i8/i16) and
11045 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11046 SDValue Cond = N->getOperand(3);
11047 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11048 DAG.getConstant(CC, MVT::i8), Cond);
11050 // Zero extend the condition if needed.
11051 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11053 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11054 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11055 DAG.getConstant(ShAmt, MVT::i8));
11056 if (N->getNumValues() == 2) // Dead flag value?
11057 return DCI.CombineTo(N, Cond, SDValue());
11061 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11062 // for any integer data type, including i8/i16.
11063 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11064 SDValue Cond = N->getOperand(3);
11065 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11066 DAG.getConstant(CC, MVT::i8), Cond);
11068 // Zero extend the condition if needed.
11069 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11070 FalseC->getValueType(0), Cond);
11071 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11072 SDValue(FalseC, 0));
11074 if (N->getNumValues() == 2) // Dead flag value?
11075 return DCI.CombineTo(N, Cond, SDValue());
11079 // Optimize cases that will turn into an LEA instruction. This requires
11080 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11081 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11082 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11083 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11085 bool isFastMultiplier = false;
11087 switch ((unsigned char)Diff) {
11089 case 1: // result = add base, cond
11090 case 2: // result = lea base( , cond*2)
11091 case 3: // result = lea base(cond, cond*2)
11092 case 4: // result = lea base( , cond*4)
11093 case 5: // result = lea base(cond, cond*4)
11094 case 8: // result = lea base( , cond*8)
11095 case 9: // result = lea base(cond, cond*8)
11096 isFastMultiplier = true;
11101 if (isFastMultiplier) {
11102 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11103 SDValue Cond = N->getOperand(3);
11104 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11105 DAG.getConstant(CC, MVT::i8), Cond);
11106 // Zero extend the condition if needed.
11107 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11109 // Scale the condition by the difference.
11111 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11112 DAG.getConstant(Diff, Cond.getValueType()));
11114 // Add the base if non-zero.
11115 if (FalseC->getAPIntValue() != 0)
11116 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11117 SDValue(FalseC, 0));
11118 if (N->getNumValues() == 2) // Dead flag value?
11119 return DCI.CombineTo(N, Cond, SDValue());
11129 /// PerformMulCombine - Optimize a single multiply with constant into two
11130 /// in order to implement it with two cheaper instructions, e.g.
11131 /// LEA + SHL, LEA + LEA.
11132 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
11133 TargetLowering::DAGCombinerInfo &DCI) {
11134 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
11137 EVT VT = N->getValueType(0);
11138 if (VT != MVT::i64)
11141 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
11144 uint64_t MulAmt = C->getZExtValue();
11145 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
11148 uint64_t MulAmt1 = 0;
11149 uint64_t MulAmt2 = 0;
11150 if ((MulAmt % 9) == 0) {
11152 MulAmt2 = MulAmt / 9;
11153 } else if ((MulAmt % 5) == 0) {
11155 MulAmt2 = MulAmt / 5;
11156 } else if ((MulAmt % 3) == 0) {
11158 MulAmt2 = MulAmt / 3;
11161 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
11162 DebugLoc DL = N->getDebugLoc();
11164 if (isPowerOf2_64(MulAmt2) &&
11165 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
11166 // If second multiplifer is pow2, issue it first. We want the multiply by
11167 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
11169 std::swap(MulAmt1, MulAmt2);
11172 if (isPowerOf2_64(MulAmt1))
11173 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
11174 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
11176 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
11177 DAG.getConstant(MulAmt1, VT));
11179 if (isPowerOf2_64(MulAmt2))
11180 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
11181 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
11183 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
11184 DAG.getConstant(MulAmt2, VT));
11186 // Do not add new nodes to DAG combiner worklist.
11187 DCI.CombineTo(N, NewMul, false);
11192 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
11193 SDValue N0 = N->getOperand(0);
11194 SDValue N1 = N->getOperand(1);
11195 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
11196 EVT VT = N0.getValueType();
11198 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
11199 // since the result of setcc_c is all zero's or all ones.
11200 if (N1C && N0.getOpcode() == ISD::AND &&
11201 N0.getOperand(1).getOpcode() == ISD::Constant) {
11202 SDValue N00 = N0.getOperand(0);
11203 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
11204 ((N00.getOpcode() == ISD::ANY_EXTEND ||
11205 N00.getOpcode() == ISD::ZERO_EXTEND) &&
11206 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
11207 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
11208 APInt ShAmt = N1C->getAPIntValue();
11209 Mask = Mask.shl(ShAmt);
11211 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
11212 N00, DAG.getConstant(Mask, VT));
11219 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
11221 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
11222 const X86Subtarget *Subtarget) {
11223 EVT VT = N->getValueType(0);
11224 if (!VT.isVector() && VT.isInteger() &&
11225 N->getOpcode() == ISD::SHL)
11226 return PerformSHLCombine(N, DAG);
11228 // On X86 with SSE2 support, we can transform this to a vector shift if
11229 // all elements are shifted by the same amount. We can't do this in legalize
11230 // because the a constant vector is typically transformed to a constant pool
11231 // so we have no knowledge of the shift amount.
11232 if (!Subtarget->hasSSE2())
11235 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
11238 SDValue ShAmtOp = N->getOperand(1);
11239 EVT EltVT = VT.getVectorElementType();
11240 DebugLoc DL = N->getDebugLoc();
11241 SDValue BaseShAmt = SDValue();
11242 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
11243 unsigned NumElts = VT.getVectorNumElements();
11245 for (; i != NumElts; ++i) {
11246 SDValue Arg = ShAmtOp.getOperand(i);
11247 if (Arg.getOpcode() == ISD::UNDEF) continue;
11251 for (; i != NumElts; ++i) {
11252 SDValue Arg = ShAmtOp.getOperand(i);
11253 if (Arg.getOpcode() == ISD::UNDEF) continue;
11254 if (Arg != BaseShAmt) {
11258 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
11259 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
11260 SDValue InVec = ShAmtOp.getOperand(0);
11261 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
11262 unsigned NumElts = InVec.getValueType().getVectorNumElements();
11264 for (; i != NumElts; ++i) {
11265 SDValue Arg = InVec.getOperand(i);
11266 if (Arg.getOpcode() == ISD::UNDEF) continue;
11270 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
11271 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
11272 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
11273 if (C->getZExtValue() == SplatIdx)
11274 BaseShAmt = InVec.getOperand(1);
11277 if (BaseShAmt.getNode() == 0)
11278 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
11279 DAG.getIntPtrConstant(0));
11283 // The shift amount is an i32.
11284 if (EltVT.bitsGT(MVT::i32))
11285 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
11286 else if (EltVT.bitsLT(MVT::i32))
11287 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
11289 // The shift amount is identical so we can do a vector shift.
11290 SDValue ValOp = N->getOperand(0);
11291 switch (N->getOpcode()) {
11293 llvm_unreachable("Unknown shift opcode!");
11296 if (VT == MVT::v2i64)
11297 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11298 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
11300 if (VT == MVT::v4i32)
11301 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11302 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
11304 if (VT == MVT::v8i16)
11305 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11306 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
11310 if (VT == MVT::v4i32)
11311 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11312 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
11314 if (VT == MVT::v8i16)
11315 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11316 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
11320 if (VT == MVT::v2i64)
11321 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11322 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
11324 if (VT == MVT::v4i32)
11325 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11326 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
11328 if (VT == MVT::v8i16)
11329 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
11330 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
11338 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
11339 TargetLowering::DAGCombinerInfo &DCI,
11340 const X86Subtarget *Subtarget) {
11341 if (DCI.isBeforeLegalizeOps())
11344 // Want to form PANDN nodes, in the hopes of then easily combining them with
11345 // OR and AND nodes to form PBLEND/PSIGN.
11346 EVT VT = N->getValueType(0);
11347 if (VT != MVT::v2i64)
11350 SDValue N0 = N->getOperand(0);
11351 SDValue N1 = N->getOperand(1);
11352 DebugLoc DL = N->getDebugLoc();
11354 // Check LHS for vnot
11355 if (N0.getOpcode() == ISD::XOR &&
11356 ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
11357 return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
11359 // Check RHS for vnot
11360 if (N1.getOpcode() == ISD::XOR &&
11361 ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
11362 return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
11367 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
11368 TargetLowering::DAGCombinerInfo &DCI,
11369 const X86Subtarget *Subtarget) {
11370 if (DCI.isBeforeLegalizeOps())
11373 EVT VT = N->getValueType(0);
11374 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
11377 SDValue N0 = N->getOperand(0);
11378 SDValue N1 = N->getOperand(1);
11380 // look for psign/blend
11381 if (Subtarget->hasSSSE3()) {
11382 if (VT == MVT::v2i64) {
11383 // Canonicalize pandn to RHS
11384 if (N0.getOpcode() == X86ISD::PANDN)
11386 // or (and (m, x), (pandn m, y))
11387 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
11388 SDValue Mask = N1.getOperand(0);
11389 SDValue X = N1.getOperand(1);
11391 if (N0.getOperand(0) == Mask)
11392 Y = N0.getOperand(1);
11393 if (N0.getOperand(1) == Mask)
11394 Y = N0.getOperand(0);
11396 // Check to see if the mask appeared in both the AND and PANDN and
11400 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
11401 if (Mask.getOpcode() != ISD::BITCAST ||
11402 X.getOpcode() != ISD::BITCAST ||
11403 Y.getOpcode() != ISD::BITCAST)
11406 // Look through mask bitcast.
11407 Mask = Mask.getOperand(0);
11408 EVT MaskVT = Mask.getValueType();
11410 // Validate that the Mask operand is a vector sra node. The sra node
11411 // will be an intrinsic.
11412 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
11415 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
11416 // there is no psrai.b
11417 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
11418 case Intrinsic::x86_sse2_psrai_w:
11419 case Intrinsic::x86_sse2_psrai_d:
11421 default: return SDValue();
11424 // Check that the SRA is all signbits.
11425 SDValue SraC = Mask.getOperand(2);
11426 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
11427 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
11428 if ((SraAmt + 1) != EltBits)
11431 DebugLoc DL = N->getDebugLoc();
11433 // Now we know we at least have a plendvb with the mask val. See if
11434 // we can form a psignb/w/d.
11435 // psign = x.type == y.type == mask.type && y = sub(0, x);
11436 X = X.getOperand(0);
11437 Y = Y.getOperand(0);
11438 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
11439 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
11440 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
11443 case 8: Opc = X86ISD::PSIGNB; break;
11444 case 16: Opc = X86ISD::PSIGNW; break;
11445 case 32: Opc = X86ISD::PSIGND; break;
11449 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
11450 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
11453 // PBLENDVB only available on SSE 4.1
11454 if (!Subtarget->hasSSE41())
11457 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
11458 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
11459 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
11460 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
11461 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
11466 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
11467 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
11469 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
11471 if (!N0.hasOneUse() || !N1.hasOneUse())
11474 SDValue ShAmt0 = N0.getOperand(1);
11475 if (ShAmt0.getValueType() != MVT::i8)
11477 SDValue ShAmt1 = N1.getOperand(1);
11478 if (ShAmt1.getValueType() != MVT::i8)
11480 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
11481 ShAmt0 = ShAmt0.getOperand(0);
11482 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
11483 ShAmt1 = ShAmt1.getOperand(0);
11485 DebugLoc DL = N->getDebugLoc();
11486 unsigned Opc = X86ISD::SHLD;
11487 SDValue Op0 = N0.getOperand(0);
11488 SDValue Op1 = N1.getOperand(0);
11489 if (ShAmt0.getOpcode() == ISD::SUB) {
11490 Opc = X86ISD::SHRD;
11491 std::swap(Op0, Op1);
11492 std::swap(ShAmt0, ShAmt1);
11495 unsigned Bits = VT.getSizeInBits();
11496 if (ShAmt1.getOpcode() == ISD::SUB) {
11497 SDValue Sum = ShAmt1.getOperand(0);
11498 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
11499 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
11500 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
11501 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
11502 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
11503 return DAG.getNode(Opc, DL, VT,
11505 DAG.getNode(ISD::TRUNCATE, DL,
11508 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
11509 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
11511 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
11512 return DAG.getNode(Opc, DL, VT,
11513 N0.getOperand(0), N1.getOperand(0),
11514 DAG.getNode(ISD::TRUNCATE, DL,
11521 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
11522 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
11523 const X86Subtarget *Subtarget) {
11524 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
11525 // the FP state in cases where an emms may be missing.
11526 // A preferable solution to the general problem is to figure out the right
11527 // places to insert EMMS. This qualifies as a quick hack.
11529 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
11530 StoreSDNode *St = cast<StoreSDNode>(N);
11531 EVT VT = St->getValue().getValueType();
11532 if (VT.getSizeInBits() != 64)
11535 const Function *F = DAG.getMachineFunction().getFunction();
11536 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
11537 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
11538 && Subtarget->hasSSE2();
11539 if ((VT.isVector() ||
11540 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
11541 isa<LoadSDNode>(St->getValue()) &&
11542 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
11543 St->getChain().hasOneUse() && !St->isVolatile()) {
11544 SDNode* LdVal = St->getValue().getNode();
11545 LoadSDNode *Ld = 0;
11546 int TokenFactorIndex = -1;
11547 SmallVector<SDValue, 8> Ops;
11548 SDNode* ChainVal = St->getChain().getNode();
11549 // Must be a store of a load. We currently handle two cases: the load
11550 // is a direct child, and it's under an intervening TokenFactor. It is
11551 // possible to dig deeper under nested TokenFactors.
11552 if (ChainVal == LdVal)
11553 Ld = cast<LoadSDNode>(St->getChain());
11554 else if (St->getValue().hasOneUse() &&
11555 ChainVal->getOpcode() == ISD::TokenFactor) {
11556 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
11557 if (ChainVal->getOperand(i).getNode() == LdVal) {
11558 TokenFactorIndex = i;
11559 Ld = cast<LoadSDNode>(St->getValue());
11561 Ops.push_back(ChainVal->getOperand(i));
11565 if (!Ld || !ISD::isNormalLoad(Ld))
11568 // If this is not the MMX case, i.e. we are just turning i64 load/store
11569 // into f64 load/store, avoid the transformation if there are multiple
11570 // uses of the loaded value.
11571 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
11574 DebugLoc LdDL = Ld->getDebugLoc();
11575 DebugLoc StDL = N->getDebugLoc();
11576 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
11577 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
11579 if (Subtarget->is64Bit() || F64IsLegal) {
11580 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
11581 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
11582 Ld->getPointerInfo(), Ld->isVolatile(),
11583 Ld->isNonTemporal(), Ld->getAlignment());
11584 SDValue NewChain = NewLd.getValue(1);
11585 if (TokenFactorIndex != -1) {
11586 Ops.push_back(NewChain);
11587 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11590 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
11591 St->getPointerInfo(),
11592 St->isVolatile(), St->isNonTemporal(),
11593 St->getAlignment());
11596 // Otherwise, lower to two pairs of 32-bit loads / stores.
11597 SDValue LoAddr = Ld->getBasePtr();
11598 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
11599 DAG.getConstant(4, MVT::i32));
11601 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
11602 Ld->getPointerInfo(),
11603 Ld->isVolatile(), Ld->isNonTemporal(),
11604 Ld->getAlignment());
11605 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
11606 Ld->getPointerInfo().getWithOffset(4),
11607 Ld->isVolatile(), Ld->isNonTemporal(),
11608 MinAlign(Ld->getAlignment(), 4));
11610 SDValue NewChain = LoLd.getValue(1);
11611 if (TokenFactorIndex != -1) {
11612 Ops.push_back(LoLd);
11613 Ops.push_back(HiLd);
11614 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
11618 LoAddr = St->getBasePtr();
11619 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
11620 DAG.getConstant(4, MVT::i32));
11622 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
11623 St->getPointerInfo(),
11624 St->isVolatile(), St->isNonTemporal(),
11625 St->getAlignment());
11626 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
11627 St->getPointerInfo().getWithOffset(4),
11629 St->isNonTemporal(),
11630 MinAlign(St->getAlignment(), 4));
11631 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
11636 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
11637 /// X86ISD::FXOR nodes.
11638 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
11639 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
11640 // F[X]OR(0.0, x) -> x
11641 // F[X]OR(x, 0.0) -> x
11642 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11643 if (C->getValueAPF().isPosZero())
11644 return N->getOperand(1);
11645 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11646 if (C->getValueAPF().isPosZero())
11647 return N->getOperand(0);
11651 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
11652 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
11653 // FAND(0.0, x) -> 0.0
11654 // FAND(x, 0.0) -> 0.0
11655 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
11656 if (C->getValueAPF().isPosZero())
11657 return N->getOperand(0);
11658 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
11659 if (C->getValueAPF().isPosZero())
11660 return N->getOperand(1);
11664 static SDValue PerformBTCombine(SDNode *N,
11666 TargetLowering::DAGCombinerInfo &DCI) {
11667 // BT ignores high bits in the bit index operand.
11668 SDValue Op1 = N->getOperand(1);
11669 if (Op1.hasOneUse()) {
11670 unsigned BitWidth = Op1.getValueSizeInBits();
11671 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
11672 APInt KnownZero, KnownOne;
11673 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
11674 !DCI.isBeforeLegalizeOps());
11675 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11676 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
11677 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
11678 DCI.CommitTargetLoweringOpt(TLO);
11683 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
11684 SDValue Op = N->getOperand(0);
11685 if (Op.getOpcode() == ISD::BITCAST)
11686 Op = Op.getOperand(0);
11687 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
11688 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
11689 VT.getVectorElementType().getSizeInBits() ==
11690 OpVT.getVectorElementType().getSizeInBits()) {
11691 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
11696 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
11697 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
11698 // (and (i32 x86isd::setcc_carry), 1)
11699 // This eliminates the zext. This transformation is necessary because
11700 // ISD::SETCC is always legalized to i8.
11701 DebugLoc dl = N->getDebugLoc();
11702 SDValue N0 = N->getOperand(0);
11703 EVT VT = N->getValueType(0);
11704 if (N0.getOpcode() == ISD::AND &&
11706 N0.getOperand(0).hasOneUse()) {
11707 SDValue N00 = N0.getOperand(0);
11708 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
11710 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
11711 if (!C || C->getZExtValue() != 1)
11713 return DAG.getNode(ISD::AND, dl, VT,
11714 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
11715 N00.getOperand(0), N00.getOperand(1)),
11716 DAG.getConstant(1, VT));
11722 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
11723 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
11724 unsigned X86CC = N->getConstantOperandVal(0);
11725 SDValue EFLAG = N->getOperand(1);
11726 DebugLoc DL = N->getDebugLoc();
11728 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
11729 // a zext and produces an all-ones bit which is more useful than 0/1 in some
11731 if (X86CC == X86::COND_B)
11732 return DAG.getNode(ISD::AND, DL, MVT::i8,
11733 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
11734 DAG.getConstant(X86CC, MVT::i8), EFLAG),
11735 DAG.getConstant(1, MVT::i8));
11740 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
11741 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
11742 X86TargetLowering::DAGCombinerInfo &DCI) {
11743 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
11744 // the result is either zero or one (depending on the input carry bit).
11745 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
11746 if (X86::isZeroNode(N->getOperand(0)) &&
11747 X86::isZeroNode(N->getOperand(1)) &&
11748 // We don't have a good way to replace an EFLAGS use, so only do this when
11750 SDValue(N, 1).use_empty()) {
11751 DebugLoc DL = N->getDebugLoc();
11752 EVT VT = N->getValueType(0);
11753 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
11754 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
11755 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
11756 DAG.getConstant(X86::COND_B,MVT::i8),
11758 DAG.getConstant(1, VT));
11759 return DCI.CombineTo(N, Res1, CarryOut);
11765 // fold (add Y, (sete X, 0)) -> adc 0, Y
11766 // (add Y, (setne X, 0)) -> sbb -1, Y
11767 // (sub (sete X, 0), Y) -> sbb 0, Y
11768 // (sub (setne X, 0), Y) -> adc -1, Y
11769 static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
11770 DebugLoc DL = N->getDebugLoc();
11772 // Look through ZExts.
11773 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
11774 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
11777 SDValue SetCC = Ext.getOperand(0);
11778 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
11781 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
11782 if (CC != X86::COND_E && CC != X86::COND_NE)
11785 SDValue Cmp = SetCC.getOperand(1);
11786 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
11787 !X86::isZeroNode(Cmp.getOperand(1)) ||
11788 !Cmp.getOperand(0).getValueType().isInteger())
11791 SDValue CmpOp0 = Cmp.getOperand(0);
11792 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
11793 DAG.getConstant(1, CmpOp0.getValueType()));
11795 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
11796 if (CC == X86::COND_NE)
11797 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
11798 DL, OtherVal.getValueType(), OtherVal,
11799 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
11800 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
11801 DL, OtherVal.getValueType(), OtherVal,
11802 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
11805 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
11806 DAGCombinerInfo &DCI) const {
11807 SelectionDAG &DAG = DCI.DAG;
11808 switch (N->getOpcode()) {
11810 case ISD::EXTRACT_VECTOR_ELT:
11811 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
11812 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
11813 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
11815 case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
11816 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
11817 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
11820 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
11821 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
11822 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
11823 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
11825 case X86ISD::FOR: return PerformFORCombine(N, DAG);
11826 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
11827 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
11828 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
11829 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
11830 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
11831 case X86ISD::SHUFPS: // Handle all target specific shuffles
11832 case X86ISD::SHUFPD:
11833 case X86ISD::PALIGN:
11834 case X86ISD::PUNPCKHBW:
11835 case X86ISD::PUNPCKHWD:
11836 case X86ISD::PUNPCKHDQ:
11837 case X86ISD::PUNPCKHQDQ:
11838 case X86ISD::UNPCKHPS:
11839 case X86ISD::UNPCKHPD:
11840 case X86ISD::PUNPCKLBW:
11841 case X86ISD::PUNPCKLWD:
11842 case X86ISD::PUNPCKLDQ:
11843 case X86ISD::PUNPCKLQDQ:
11844 case X86ISD::UNPCKLPS:
11845 case X86ISD::UNPCKLPD:
11846 case X86ISD::MOVHLPS:
11847 case X86ISD::MOVLHPS:
11848 case X86ISD::PSHUFD:
11849 case X86ISD::PSHUFHW:
11850 case X86ISD::PSHUFLW:
11851 case X86ISD::MOVSS:
11852 case X86ISD::MOVSD:
11853 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
11859 /// isTypeDesirableForOp - Return true if the target has native support for
11860 /// the specified value type and it is 'desirable' to use the type for the
11861 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
11862 /// instruction encodings are longer and some i16 instructions are slow.
11863 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
11864 if (!isTypeLegal(VT))
11866 if (VT != MVT::i16)
11873 case ISD::SIGN_EXTEND:
11874 case ISD::ZERO_EXTEND:
11875 case ISD::ANY_EXTEND:
11888 /// IsDesirableToPromoteOp - This method query the target whether it is
11889 /// beneficial for dag combiner to promote the specified node. If true, it
11890 /// should return the desired promotion type by reference.
11891 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
11892 EVT VT = Op.getValueType();
11893 if (VT != MVT::i16)
11896 bool Promote = false;
11897 bool Commute = false;
11898 switch (Op.getOpcode()) {
11901 LoadSDNode *LD = cast<LoadSDNode>(Op);
11902 // If the non-extending load has a single use and it's not live out, then it
11903 // might be folded.
11904 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
11905 Op.hasOneUse()*/) {
11906 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
11907 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
11908 // The only case where we'd want to promote LOAD (rather then it being
11909 // promoted as an operand is when it's only use is liveout.
11910 if (UI->getOpcode() != ISD::CopyToReg)
11917 case ISD::SIGN_EXTEND:
11918 case ISD::ZERO_EXTEND:
11919 case ISD::ANY_EXTEND:
11924 SDValue N0 = Op.getOperand(0);
11925 // Look out for (store (shl (load), x)).
11926 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
11939 SDValue N0 = Op.getOperand(0);
11940 SDValue N1 = Op.getOperand(1);
11941 if (!Commute && MayFoldLoad(N1))
11943 // Avoid disabling potential load folding opportunities.
11944 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
11946 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
11956 //===----------------------------------------------------------------------===//
11957 // X86 Inline Assembly Support
11958 //===----------------------------------------------------------------------===//
11960 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
11961 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
11963 std::string AsmStr = IA->getAsmString();
11965 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
11966 SmallVector<StringRef, 4> AsmPieces;
11967 SplitString(AsmStr, AsmPieces, ";\n");
11969 switch (AsmPieces.size()) {
11970 default: return false;
11972 AsmStr = AsmPieces[0];
11974 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
11976 // FIXME: this should verify that we are targetting a 486 or better. If not,
11977 // we will turn this bswap into something that will be lowered to logical ops
11978 // instead of emitting the bswap asm. For now, we don't support 486 or lower
11979 // so don't worry about this.
11981 if (AsmPieces.size() == 2 &&
11982 (AsmPieces[0] == "bswap" ||
11983 AsmPieces[0] == "bswapq" ||
11984 AsmPieces[0] == "bswapl") &&
11985 (AsmPieces[1] == "$0" ||
11986 AsmPieces[1] == "${0:q}")) {
11987 // No need to check constraints, nothing other than the equivalent of
11988 // "=r,0" would be valid here.
11989 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
11990 if (!Ty || Ty->getBitWidth() % 16 != 0)
11992 return IntrinsicLowering::LowerToByteSwap(CI);
11994 // rorw $$8, ${0:w} --> llvm.bswap.i16
11995 if (CI->getType()->isIntegerTy(16) &&
11996 AsmPieces.size() == 3 &&
11997 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
11998 AsmPieces[1] == "$$8," &&
11999 AsmPieces[2] == "${0:w}" &&
12000 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12002 const std::string &ConstraintsStr = IA->getConstraintString();
12003 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12004 std::sort(AsmPieces.begin(), AsmPieces.end());
12005 if (AsmPieces.size() == 4 &&
12006 AsmPieces[0] == "~{cc}" &&
12007 AsmPieces[1] == "~{dirflag}" &&
12008 AsmPieces[2] == "~{flags}" &&
12009 AsmPieces[3] == "~{fpsr}") {
12010 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12011 if (!Ty || Ty->getBitWidth() % 16 != 0)
12013 return IntrinsicLowering::LowerToByteSwap(CI);
12018 if (CI->getType()->isIntegerTy(32) &&
12019 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
12020 SmallVector<StringRef, 4> Words;
12021 SplitString(AsmPieces[0], Words, " \t,");
12022 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12023 Words[2] == "${0:w}") {
12025 SplitString(AsmPieces[1], Words, " \t,");
12026 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
12027 Words[2] == "$0") {
12029 SplitString(AsmPieces[2], Words, " \t,");
12030 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
12031 Words[2] == "${0:w}") {
12033 const std::string &ConstraintsStr = IA->getConstraintString();
12034 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
12035 std::sort(AsmPieces.begin(), AsmPieces.end());
12036 if (AsmPieces.size() == 4 &&
12037 AsmPieces[0] == "~{cc}" &&
12038 AsmPieces[1] == "~{dirflag}" &&
12039 AsmPieces[2] == "~{flags}" &&
12040 AsmPieces[3] == "~{fpsr}") {
12041 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12042 if (!Ty || Ty->getBitWidth() % 16 != 0)
12044 return IntrinsicLowering::LowerToByteSwap(CI);
12051 if (CI->getType()->isIntegerTy(64)) {
12052 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
12053 if (Constraints.size() >= 2 &&
12054 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
12055 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
12056 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
12057 SmallVector<StringRef, 4> Words;
12058 SplitString(AsmPieces[0], Words, " \t");
12059 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
12061 SplitString(AsmPieces[1], Words, " \t");
12062 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
12064 SplitString(AsmPieces[2], Words, " \t,");
12065 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
12066 Words[2] == "%edx") {
12067 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
12068 if (!Ty || Ty->getBitWidth() % 16 != 0)
12070 return IntrinsicLowering::LowerToByteSwap(CI);
12083 /// getConstraintType - Given a constraint letter, return the type of
12084 /// constraint it is for this target.
12085 X86TargetLowering::ConstraintType
12086 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
12087 if (Constraint.size() == 1) {
12088 switch (Constraint[0]) {
12098 return C_RegisterClass;
12122 return TargetLowering::getConstraintType(Constraint);
12125 /// Examine constraint type and operand type and determine a weight value.
12126 /// This object must already have been set up with the operand type
12127 /// and the current alternative constraint selected.
12128 TargetLowering::ConstraintWeight
12129 X86TargetLowering::getSingleConstraintMatchWeight(
12130 AsmOperandInfo &info, const char *constraint) const {
12131 ConstraintWeight weight = CW_Invalid;
12132 Value *CallOperandVal = info.CallOperandVal;
12133 // If we don't have a value, we can't do a match,
12134 // but allow it at the lowest weight.
12135 if (CallOperandVal == NULL)
12137 const Type *type = CallOperandVal->getType();
12138 // Look at the constraint type.
12139 switch (*constraint) {
12141 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
12152 if (CallOperandVal->getType()->isIntegerTy())
12153 weight = CW_SpecificReg;
12158 if (type->isFloatingPointTy())
12159 weight = CW_SpecificReg;
12162 if (type->isX86_MMXTy() && Subtarget->hasMMX())
12163 weight = CW_SpecificReg;
12167 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
12168 weight = CW_Register;
12171 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
12172 if (C->getZExtValue() <= 31)
12173 weight = CW_Constant;
12177 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12178 if (C->getZExtValue() <= 63)
12179 weight = CW_Constant;
12183 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12184 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
12185 weight = CW_Constant;
12189 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12190 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
12191 weight = CW_Constant;
12195 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12196 if (C->getZExtValue() <= 3)
12197 weight = CW_Constant;
12201 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12202 if (C->getZExtValue() <= 0xff)
12203 weight = CW_Constant;
12208 if (dyn_cast<ConstantFP>(CallOperandVal)) {
12209 weight = CW_Constant;
12213 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12214 if ((C->getSExtValue() >= -0x80000000LL) &&
12215 (C->getSExtValue() <= 0x7fffffffLL))
12216 weight = CW_Constant;
12220 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
12221 if (C->getZExtValue() <= 0xffffffff)
12222 weight = CW_Constant;
12229 /// LowerXConstraint - try to replace an X constraint, which matches anything,
12230 /// with another that has more specific requirements based on the type of the
12231 /// corresponding operand.
12232 const char *X86TargetLowering::
12233 LowerXConstraint(EVT ConstraintVT) const {
12234 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
12235 // 'f' like normal targets.
12236 if (ConstraintVT.isFloatingPoint()) {
12237 if (Subtarget->hasXMMInt())
12239 if (Subtarget->hasXMM())
12243 return TargetLowering::LowerXConstraint(ConstraintVT);
12246 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
12247 /// vector. If it is invalid, don't add anything to Ops.
12248 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
12250 std::vector<SDValue>&Ops,
12251 SelectionDAG &DAG) const {
12252 SDValue Result(0, 0);
12254 switch (Constraint) {
12257 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12258 if (C->getZExtValue() <= 31) {
12259 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12265 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12266 if (C->getZExtValue() <= 63) {
12267 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12273 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12274 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
12275 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12281 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12282 if (C->getZExtValue() <= 255) {
12283 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12289 // 32-bit signed value
12290 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12291 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12292 C->getSExtValue())) {
12293 // Widen to 64 bits here to get it sign extended.
12294 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
12297 // FIXME gcc accepts some relocatable values here too, but only in certain
12298 // memory models; it's complicated.
12303 // 32-bit unsigned value
12304 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
12305 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
12306 C->getZExtValue())) {
12307 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
12311 // FIXME gcc accepts some relocatable values here too, but only in certain
12312 // memory models; it's complicated.
12316 // Literal immediates are always ok.
12317 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
12318 // Widen to 64 bits here to get it sign extended.
12319 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
12323 // In any sort of PIC mode addresses need to be computed at runtime by
12324 // adding in a register or some sort of table lookup. These can't
12325 // be used as immediates.
12326 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
12329 // If we are in non-pic codegen mode, we allow the address of a global (with
12330 // an optional displacement) to be used with 'i'.
12331 GlobalAddressSDNode *GA = 0;
12332 int64_t Offset = 0;
12334 // Match either (GA), (GA+C), (GA+C1+C2), etc.
12336 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
12337 Offset += GA->getOffset();
12339 } else if (Op.getOpcode() == ISD::ADD) {
12340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12341 Offset += C->getZExtValue();
12342 Op = Op.getOperand(0);
12345 } else if (Op.getOpcode() == ISD::SUB) {
12346 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
12347 Offset += -C->getZExtValue();
12348 Op = Op.getOperand(0);
12353 // Otherwise, this isn't something we can handle, reject it.
12357 const GlobalValue *GV = GA->getGlobal();
12358 // If we require an extra load to get this address, as in PIC mode, we
12359 // can't accept it.
12360 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
12361 getTargetMachine())))
12364 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
12365 GA->getValueType(0), Offset);
12370 if (Result.getNode()) {
12371 Ops.push_back(Result);
12374 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
12377 std::vector<unsigned> X86TargetLowering::
12378 getRegClassForInlineAsmConstraint(const std::string &Constraint,
12380 if (Constraint.size() == 1) {
12381 // FIXME: not handling fp-stack yet!
12382 switch (Constraint[0]) { // GCC X86 Constraint Letters
12383 default: break; // Unknown constraint letter
12384 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
12385 if (Subtarget->is64Bit()) {
12386 if (VT == MVT::i32)
12387 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
12388 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
12389 X86::R10D,X86::R11D,X86::R12D,
12390 X86::R13D,X86::R14D,X86::R15D,
12391 X86::EBP, X86::ESP, 0);
12392 else if (VT == MVT::i16)
12393 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
12394 X86::SI, X86::DI, X86::R8W,X86::R9W,
12395 X86::R10W,X86::R11W,X86::R12W,
12396 X86::R13W,X86::R14W,X86::R15W,
12397 X86::BP, X86::SP, 0);
12398 else if (VT == MVT::i8)
12399 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
12400 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
12401 X86::R10B,X86::R11B,X86::R12B,
12402 X86::R13B,X86::R14B,X86::R15B,
12403 X86::BPL, X86::SPL, 0);
12405 else if (VT == MVT::i64)
12406 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
12407 X86::RSI, X86::RDI, X86::R8, X86::R9,
12408 X86::R10, X86::R11, X86::R12,
12409 X86::R13, X86::R14, X86::R15,
12410 X86::RBP, X86::RSP, 0);
12414 // 32-bit fallthrough
12415 case 'Q': // Q_REGS
12416 if (VT == MVT::i32)
12417 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
12418 else if (VT == MVT::i16)
12419 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
12420 else if (VT == MVT::i8)
12421 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
12422 else if (VT == MVT::i64)
12423 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
12428 return std::vector<unsigned>();
12431 std::pair<unsigned, const TargetRegisterClass*>
12432 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
12434 // First, see if this is a constraint that directly corresponds to an LLVM
12436 if (Constraint.size() == 1) {
12437 // GCC Constraint Letters
12438 switch (Constraint[0]) {
12440 case 'r': // GENERAL_REGS
12441 case 'l': // INDEX_REGS
12443 return std::make_pair(0U, X86::GR8RegisterClass);
12444 if (VT == MVT::i16)
12445 return std::make_pair(0U, X86::GR16RegisterClass);
12446 if (VT == MVT::i32 || !Subtarget->is64Bit())
12447 return std::make_pair(0U, X86::GR32RegisterClass);
12448 return std::make_pair(0U, X86::GR64RegisterClass);
12449 case 'R': // LEGACY_REGS
12451 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
12452 if (VT == MVT::i16)
12453 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
12454 if (VT == MVT::i32 || !Subtarget->is64Bit())
12455 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
12456 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
12457 case 'f': // FP Stack registers.
12458 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
12459 // value to the correct fpstack register class.
12460 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
12461 return std::make_pair(0U, X86::RFP32RegisterClass);
12462 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
12463 return std::make_pair(0U, X86::RFP64RegisterClass);
12464 return std::make_pair(0U, X86::RFP80RegisterClass);
12465 case 'y': // MMX_REGS if MMX allowed.
12466 if (!Subtarget->hasMMX()) break;
12467 return std::make_pair(0U, X86::VR64RegisterClass);
12468 case 'Y': // SSE_REGS if SSE2 allowed
12469 if (!Subtarget->hasXMMInt()) break;
12471 case 'x': // SSE_REGS if SSE1 allowed
12472 if (!Subtarget->hasXMM()) break;
12474 switch (VT.getSimpleVT().SimpleTy) {
12476 // Scalar SSE types.
12479 return std::make_pair(0U, X86::FR32RegisterClass);
12482 return std::make_pair(0U, X86::FR64RegisterClass);
12490 return std::make_pair(0U, X86::VR128RegisterClass);
12496 // Use the default implementation in TargetLowering to convert the register
12497 // constraint into a member of a register class.
12498 std::pair<unsigned, const TargetRegisterClass*> Res;
12499 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
12501 // Not found as a standard register?
12502 if (Res.second == 0) {
12503 // Map st(0) -> st(7) -> ST0
12504 if (Constraint.size() == 7 && Constraint[0] == '{' &&
12505 tolower(Constraint[1]) == 's' &&
12506 tolower(Constraint[2]) == 't' &&
12507 Constraint[3] == '(' &&
12508 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
12509 Constraint[5] == ')' &&
12510 Constraint[6] == '}') {
12512 Res.first = X86::ST0+Constraint[4]-'0';
12513 Res.second = X86::RFP80RegisterClass;
12517 // GCC allows "st(0)" to be called just plain "st".
12518 if (StringRef("{st}").equals_lower(Constraint)) {
12519 Res.first = X86::ST0;
12520 Res.second = X86::RFP80RegisterClass;
12525 if (StringRef("{flags}").equals_lower(Constraint)) {
12526 Res.first = X86::EFLAGS;
12527 Res.second = X86::CCRRegisterClass;
12531 // 'A' means EAX + EDX.
12532 if (Constraint == "A") {
12533 Res.first = X86::EAX;
12534 Res.second = X86::GR32_ADRegisterClass;
12540 // Otherwise, check to see if this is a register class of the wrong value
12541 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
12542 // turn into {ax},{dx}.
12543 if (Res.second->hasType(VT))
12544 return Res; // Correct type already, nothing to do.
12546 // All of the single-register GCC register classes map their values onto
12547 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
12548 // really want an 8-bit or 32-bit register, map to the appropriate register
12549 // class and return the appropriate register.
12550 if (Res.second == X86::GR16RegisterClass) {
12551 if (VT == MVT::i8) {
12552 unsigned DestReg = 0;
12553 switch (Res.first) {
12555 case X86::AX: DestReg = X86::AL; break;
12556 case X86::DX: DestReg = X86::DL; break;
12557 case X86::CX: DestReg = X86::CL; break;
12558 case X86::BX: DestReg = X86::BL; break;
12561 Res.first = DestReg;
12562 Res.second = X86::GR8RegisterClass;
12564 } else if (VT == MVT::i32) {
12565 unsigned DestReg = 0;
12566 switch (Res.first) {
12568 case X86::AX: DestReg = X86::EAX; break;
12569 case X86::DX: DestReg = X86::EDX; break;
12570 case X86::CX: DestReg = X86::ECX; break;
12571 case X86::BX: DestReg = X86::EBX; break;
12572 case X86::SI: DestReg = X86::ESI; break;
12573 case X86::DI: DestReg = X86::EDI; break;
12574 case X86::BP: DestReg = X86::EBP; break;
12575 case X86::SP: DestReg = X86::ESP; break;
12578 Res.first = DestReg;
12579 Res.second = X86::GR32RegisterClass;
12581 } else if (VT == MVT::i64) {
12582 unsigned DestReg = 0;
12583 switch (Res.first) {
12585 case X86::AX: DestReg = X86::RAX; break;
12586 case X86::DX: DestReg = X86::RDX; break;
12587 case X86::CX: DestReg = X86::RCX; break;
12588 case X86::BX: DestReg = X86::RBX; break;
12589 case X86::SI: DestReg = X86::RSI; break;
12590 case X86::DI: DestReg = X86::RDI; break;
12591 case X86::BP: DestReg = X86::RBP; break;
12592 case X86::SP: DestReg = X86::RSP; break;
12595 Res.first = DestReg;
12596 Res.second = X86::GR64RegisterClass;
12599 } else if (Res.second == X86::FR32RegisterClass ||
12600 Res.second == X86::FR64RegisterClass ||
12601 Res.second == X86::VR128RegisterClass) {
12602 // Handle references to XMM physical registers that got mapped into the
12603 // wrong class. This can happen with constraints like {xmm0} where the
12604 // target independent register mapper will just pick the first match it can
12605 // find, ignoring the required type.
12606 if (VT == MVT::f32)
12607 Res.second = X86::FR32RegisterClass;
12608 else if (VT == MVT::f64)
12609 Res.second = X86::FR64RegisterClass;
12610 else if (X86::VR128RegisterClass->hasType(VT))
12611 Res.second = X86::VR128RegisterClass;