1 //===-- X86FastISel.cpp - X86 FastISel 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 X86-specific support for the FastISel class. Much
11 // of the target-specific code is generated by tablegen in the file
12 // X86GenFastISel.inc, which is #included here.
14 //===----------------------------------------------------------------------===//
17 #include "X86InstrBuilder.h"
18 #include "X86RegisterInfo.h"
19 #include "X86Subtarget.h"
20 #include "X86TargetMachine.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/IntrinsicInst.h"
26 #include "llvm/Operator.h"
27 #include "llvm/CodeGen/Analysis.h"
28 #include "llvm/CodeGen/FastISel.h"
29 #include "llvm/CodeGen/FunctionLoweringInfo.h"
30 #include "llvm/CodeGen/MachineConstantPool.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineRegisterInfo.h"
33 #include "llvm/Support/CallSite.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/GetElementPtrTypeIterator.h"
36 #include "llvm/Target/TargetOptions.h"
41 class X86FastISel : public FastISel {
42 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
43 /// make the right decision when generating code for different targets.
44 const X86Subtarget *Subtarget;
46 /// StackPtr - Register used as the stack pointer.
50 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
51 /// floating point ops.
52 /// When SSE is available, use it for f32 operations.
53 /// When SSE2 is available, use it for f64 operations.
58 explicit X86FastISel(FunctionLoweringInfo &funcInfo) : FastISel(funcInfo) {
59 Subtarget = &TM.getSubtarget<X86Subtarget>();
60 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
61 X86ScalarSSEf64 = Subtarget->hasSSE2();
62 X86ScalarSSEf32 = Subtarget->hasSSE1();
65 virtual bool TargetSelectInstruction(const Instruction *I);
67 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
68 /// vreg is being provided by the specified load instruction. If possible,
69 /// try to fold the load as an operand to the instruction, returning true if
71 virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
74 #include "X86GenFastISel.inc"
77 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT);
79 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
81 bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM);
82 bool X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM);
84 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
87 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
88 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
90 bool X86SelectLoad(const Instruction *I);
92 bool X86SelectStore(const Instruction *I);
94 bool X86SelectRet(const Instruction *I);
96 bool X86SelectCmp(const Instruction *I);
98 bool X86SelectZExt(const Instruction *I);
100 bool X86SelectBranch(const Instruction *I);
102 bool X86SelectShift(const Instruction *I);
104 bool X86SelectSelect(const Instruction *I);
106 bool X86SelectTrunc(const Instruction *I);
108 bool X86SelectFPExt(const Instruction *I);
109 bool X86SelectFPTrunc(const Instruction *I);
111 bool X86VisitIntrinsicCall(const IntrinsicInst &I);
112 bool X86SelectCall(const Instruction *I);
114 const X86InstrInfo *getInstrInfo() const {
115 return getTargetMachine()->getInstrInfo();
117 const X86TargetMachine *getTargetMachine() const {
118 return static_cast<const X86TargetMachine *>(&TM);
121 unsigned TargetMaterializeConstant(const Constant *C);
123 unsigned TargetMaterializeAlloca(const AllocaInst *C);
125 unsigned TargetMaterializeFloatZero(const ConstantFP *CF);
127 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
128 /// computed in an SSE register, not on the X87 floating point stack.
129 bool isScalarFPTypeInSSEReg(EVT VT) const {
130 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
131 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
134 bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
136 bool IsMemcpySmall(uint64_t Len);
138 bool TryEmitSmallMemcpy(X86AddressMode DestAM,
139 X86AddressMode SrcAM, uint64_t Len);
142 } // end anonymous namespace.
144 bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
145 EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true);
146 if (evt == MVT::Other || !evt.isSimple())
147 // Unhandled type. Halt "fast" selection and bail.
150 VT = evt.getSimpleVT();
151 // For now, require SSE/SSE2 for performing floating-point operations,
152 // since x87 requires additional work.
153 if (VT == MVT::f64 && !X86ScalarSSEf64)
155 if (VT == MVT::f32 && !X86ScalarSSEf32)
157 // Similarly, no f80 support yet.
160 // We only handle legal types. For example, on x86-32 the instruction
161 // selector contains all of the 64-bit instructions from x86-64,
162 // under the assumption that i64 won't be used if the target doesn't
164 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
167 #include "X86GenCallingConv.inc"
169 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
170 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
171 /// Return true and the result register by reference if it is possible.
172 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
173 unsigned &ResultReg) {
174 // Get opcode and regclass of the output for the given load instruction.
176 const TargetRegisterClass *RC = NULL;
177 switch (VT.getSimpleVT().SimpleTy) {
178 default: return false;
182 RC = X86::GR8RegisterClass;
186 RC = X86::GR16RegisterClass;
190 RC = X86::GR32RegisterClass;
193 // Must be in x86-64 mode.
195 RC = X86::GR64RegisterClass;
198 if (Subtarget->hasSSE1()) {
200 RC = X86::FR32RegisterClass;
203 RC = X86::RFP32RegisterClass;
207 if (Subtarget->hasSSE2()) {
209 RC = X86::FR64RegisterClass;
212 RC = X86::RFP64RegisterClass;
216 // No f80 support yet.
220 ResultReg = createResultReg(RC);
221 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
222 DL, TII.get(Opc), ResultReg), AM);
226 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
227 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
228 /// and a displacement offset, or a GlobalAddress,
229 /// i.e. V. Return true if it is possible.
231 X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) {
232 // Get opcode and regclass of the output for the given store instruction.
234 switch (VT.getSimpleVT().SimpleTy) {
235 case MVT::f80: // No f80 support yet.
236 default: return false;
238 // Mask out all but lowest bit.
239 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
240 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
241 TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
244 // FALLTHROUGH, handling i1 as i8.
245 case MVT::i8: Opc = X86::MOV8mr; break;
246 case MVT::i16: Opc = X86::MOV16mr; break;
247 case MVT::i32: Opc = X86::MOV32mr; break;
248 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
250 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
253 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
257 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
258 DL, TII.get(Opc)), AM).addReg(Val);
262 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
263 const X86AddressMode &AM) {
264 // Handle 'null' like i32/i64 0.
265 if (isa<ConstantPointerNull>(Val))
266 Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext()));
268 // If this is a store of a simple constant, fold the constant into the store.
269 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
272 switch (VT.getSimpleVT().SimpleTy) {
274 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
275 case MVT::i8: Opc = X86::MOV8mi; break;
276 case MVT::i16: Opc = X86::MOV16mi; break;
277 case MVT::i32: Opc = X86::MOV32mi; break;
279 // Must be a 32-bit sign extended value.
280 if ((int)CI->getSExtValue() == CI->getSExtValue())
281 Opc = X86::MOV64mi32;
286 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
287 DL, TII.get(Opc)), AM)
288 .addImm(Signed ? (uint64_t) CI->getSExtValue() :
294 unsigned ValReg = getRegForValue(Val);
298 return X86FastEmitStore(VT, ValReg, AM);
301 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
302 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
303 /// ISD::SIGN_EXTEND).
304 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
305 unsigned Src, EVT SrcVT,
306 unsigned &ResultReg) {
307 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
308 Src, /*TODO: Kill=*/false);
317 /// X86SelectAddress - Attempt to fill in an address from the given value.
319 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
320 const User *U = NULL;
321 unsigned Opcode = Instruction::UserOp1;
322 if (const Instruction *I = dyn_cast<Instruction>(V)) {
323 // Don't walk into other basic blocks; it's possible we haven't
324 // visited them yet, so the instructions may not yet be assigned
325 // virtual registers.
326 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
327 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
328 Opcode = I->getOpcode();
331 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
332 Opcode = C->getOpcode();
336 if (const PointerType *Ty = dyn_cast<PointerType>(V->getType()))
337 if (Ty->getAddressSpace() > 255)
338 // Fast instruction selection doesn't support the special
344 case Instruction::BitCast:
345 // Look past bitcasts.
346 return X86SelectAddress(U->getOperand(0), AM);
348 case Instruction::IntToPtr:
349 // Look past no-op inttoptrs.
350 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
351 return X86SelectAddress(U->getOperand(0), AM);
354 case Instruction::PtrToInt:
355 // Look past no-op ptrtoints.
356 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
357 return X86SelectAddress(U->getOperand(0), AM);
360 case Instruction::Alloca: {
361 // Do static allocas.
362 const AllocaInst *A = cast<AllocaInst>(V);
363 DenseMap<const AllocaInst*, int>::iterator SI =
364 FuncInfo.StaticAllocaMap.find(A);
365 if (SI != FuncInfo.StaticAllocaMap.end()) {
366 AM.BaseType = X86AddressMode::FrameIndexBase;
367 AM.Base.FrameIndex = SI->second;
373 case Instruction::Add: {
374 // Adds of constants are common and easy enough.
375 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
376 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
377 // They have to fit in the 32-bit signed displacement field though.
378 if (isInt<32>(Disp)) {
379 AM.Disp = (uint32_t)Disp;
380 return X86SelectAddress(U->getOperand(0), AM);
386 case Instruction::GetElementPtr: {
387 X86AddressMode SavedAM = AM;
389 // Pattern-match simple GEPs.
390 uint64_t Disp = (int32_t)AM.Disp;
391 unsigned IndexReg = AM.IndexReg;
392 unsigned Scale = AM.Scale;
393 gep_type_iterator GTI = gep_type_begin(U);
394 // Iterate through the indices, folding what we can. Constants can be
395 // folded, and one dynamic index can be handled, if the scale is supported.
396 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
397 i != e; ++i, ++GTI) {
398 const Value *Op = *i;
399 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
400 const StructLayout *SL = TD.getStructLayout(STy);
401 Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
405 // A array/variable index is always of the form i*S where S is the
406 // constant scale size. See if we can push the scale into immediates.
407 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
409 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
410 // Constant-offset addressing.
411 Disp += CI->getSExtValue() * S;
414 if (isa<AddOperator>(Op) &&
415 (!isa<Instruction>(Op) ||
416 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
418 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
419 // An add (in the same block) with a constant operand. Fold the
422 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
423 Disp += CI->getSExtValue() * S;
424 // Iterate on the other operand.
425 Op = cast<AddOperator>(Op)->getOperand(0);
429 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
430 (S == 1 || S == 2 || S == 4 || S == 8)) {
431 // Scaled-index addressing.
433 IndexReg = getRegForGEPIndex(Op).first;
439 goto unsupported_gep;
442 // Check for displacement overflow.
443 if (!isInt<32>(Disp))
445 // Ok, the GEP indices were covered by constant-offset and scaled-index
446 // addressing. Update the address state and move on to examining the base.
447 AM.IndexReg = IndexReg;
449 AM.Disp = (uint32_t)Disp;
450 if (X86SelectAddress(U->getOperand(0), AM))
453 // If we couldn't merge the gep value into this addr mode, revert back to
454 // our address and just match the value instead of completely failing.
458 // Ok, the GEP indices weren't all covered.
463 // Handle constant address.
464 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
465 // Can't handle alternate code models or TLS yet.
466 if (TM.getCodeModel() != CodeModel::Small)
469 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
470 if (GVar->isThreadLocal())
473 // RIP-relative addresses can't have additional register operands, so if
474 // we've already folded stuff into the addressing mode, just force the
475 // global value into its own register, which we can use as the basereg.
476 if (!Subtarget->isPICStyleRIPRel() ||
477 (AM.Base.Reg == 0 && AM.IndexReg == 0)) {
478 // Okay, we've committed to selecting this global. Set up the address.
481 // Allow the subtarget to classify the global.
482 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
484 // If this reference is relative to the pic base, set it now.
485 if (isGlobalRelativeToPICBase(GVFlags)) {
486 // FIXME: How do we know Base.Reg is free??
487 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
490 // Unless the ABI requires an extra load, return a direct reference to
492 if (!isGlobalStubReference(GVFlags)) {
493 if (Subtarget->isPICStyleRIPRel()) {
494 // Use rip-relative addressing if we can. Above we verified that the
495 // base and index registers are unused.
496 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
497 AM.Base.Reg = X86::RIP;
499 AM.GVOpFlags = GVFlags;
503 // Ok, we need to do a load from a stub. If we've already loaded from
504 // this stub, reuse the loaded pointer, otherwise emit the load now.
505 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
507 if (I != LocalValueMap.end() && I->second != 0) {
510 // Issue load from stub.
512 const TargetRegisterClass *RC = NULL;
513 X86AddressMode StubAM;
514 StubAM.Base.Reg = AM.Base.Reg;
516 StubAM.GVOpFlags = GVFlags;
518 // Prepare for inserting code in the local-value area.
519 SavePoint SaveInsertPt = enterLocalValueArea();
521 if (TLI.getPointerTy() == MVT::i64) {
523 RC = X86::GR64RegisterClass;
525 if (Subtarget->isPICStyleRIPRel())
526 StubAM.Base.Reg = X86::RIP;
529 RC = X86::GR32RegisterClass;
532 LoadReg = createResultReg(RC);
533 MachineInstrBuilder LoadMI =
534 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg);
535 addFullAddress(LoadMI, StubAM);
537 // Ok, back to normal mode.
538 leaveLocalValueArea(SaveInsertPt);
540 // Prevent loading GV stub multiple times in same MBB.
541 LocalValueMap[V] = LoadReg;
544 // Now construct the final address. Note that the Disp, Scale,
545 // and Index values may already be set here.
546 AM.Base.Reg = LoadReg;
552 // If all else fails, try to materialize the value in a register.
553 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
554 if (AM.Base.Reg == 0) {
555 AM.Base.Reg = getRegForValue(V);
556 return AM.Base.Reg != 0;
558 if (AM.IndexReg == 0) {
559 assert(AM.Scale == 1 && "Scale with no index!");
560 AM.IndexReg = getRegForValue(V);
561 return AM.IndexReg != 0;
568 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
570 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
571 const User *U = NULL;
572 unsigned Opcode = Instruction::UserOp1;
573 if (const Instruction *I = dyn_cast<Instruction>(V)) {
574 Opcode = I->getOpcode();
576 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
577 Opcode = C->getOpcode();
583 case Instruction::BitCast:
584 // Look past bitcasts.
585 return X86SelectCallAddress(U->getOperand(0), AM);
587 case Instruction::IntToPtr:
588 // Look past no-op inttoptrs.
589 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
590 return X86SelectCallAddress(U->getOperand(0), AM);
593 case Instruction::PtrToInt:
594 // Look past no-op ptrtoints.
595 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
596 return X86SelectCallAddress(U->getOperand(0), AM);
600 // Handle constant address.
601 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
602 // Can't handle alternate code models yet.
603 if (TM.getCodeModel() != CodeModel::Small)
606 // RIP-relative addresses can't have additional register operands.
607 if (Subtarget->isPICStyleRIPRel() &&
608 (AM.Base.Reg != 0 || AM.IndexReg != 0))
611 // Can't handle DLLImport.
612 if (GV->hasDLLImportLinkage())
616 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
617 if (GVar->isThreadLocal())
620 // Okay, we've committed to selecting this global. Set up the basic address.
623 // No ABI requires an extra load for anything other than DLLImport, which
624 // we rejected above. Return a direct reference to the global.
625 if (Subtarget->isPICStyleRIPRel()) {
626 // Use rip-relative addressing if we can. Above we verified that the
627 // base and index registers are unused.
628 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
629 AM.Base.Reg = X86::RIP;
630 } else if (Subtarget->isPICStyleStubPIC()) {
631 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
632 } else if (Subtarget->isPICStyleGOT()) {
633 AM.GVOpFlags = X86II::MO_GOTOFF;
639 // If all else fails, try to materialize the value in a register.
640 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
641 if (AM.Base.Reg == 0) {
642 AM.Base.Reg = getRegForValue(V);
643 return AM.Base.Reg != 0;
645 if (AM.IndexReg == 0) {
646 assert(AM.Scale == 1 && "Scale with no index!");
647 AM.IndexReg = getRegForValue(V);
648 return AM.IndexReg != 0;
656 /// X86SelectStore - Select and emit code to implement store instructions.
657 bool X86FastISel::X86SelectStore(const Instruction *I) {
659 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
663 if (!X86SelectAddress(I->getOperand(1), AM))
666 return X86FastEmitStore(VT, I->getOperand(0), AM);
669 /// X86SelectRet - Select and emit code to implement ret instructions.
670 bool X86FastISel::X86SelectRet(const Instruction *I) {
671 const ReturnInst *Ret = cast<ReturnInst>(I);
672 const Function &F = *I->getParent()->getParent();
674 if (!FuncInfo.CanLowerReturn)
677 CallingConv::ID CC = F.getCallingConv();
678 if (CC != CallingConv::C &&
679 CC != CallingConv::Fast &&
680 CC != CallingConv::X86_FastCall)
683 if (Subtarget->isTargetWin64())
686 // Don't handle popping bytes on return for now.
687 if (FuncInfo.MF->getInfo<X86MachineFunctionInfo>()
688 ->getBytesToPopOnReturn() != 0)
691 // fastcc with -tailcallopt is intended to provide a guaranteed
692 // tail call optimization. Fastisel doesn't know how to do that.
693 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
696 // Let SDISel handle vararg functions.
700 if (Ret->getNumOperands() > 0) {
701 SmallVector<ISD::OutputArg, 4> Outs;
702 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
705 // Analyze operands of the call, assigning locations to each operand.
706 SmallVector<CCValAssign, 16> ValLocs;
707 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
708 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
710 const Value *RV = Ret->getOperand(0);
711 unsigned Reg = getRegForValue(RV);
715 // Only handle a single return value for now.
716 if (ValLocs.size() != 1)
719 CCValAssign &VA = ValLocs[0];
721 // Don't bother handling odd stuff for now.
722 if (VA.getLocInfo() != CCValAssign::Full)
724 // Only handle register returns for now.
728 // The calling-convention tables for x87 returns don't tell
730 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
733 unsigned SrcReg = Reg + VA.getValNo();
734 EVT SrcVT = TLI.getValueType(RV->getType());
735 EVT DstVT = VA.getValVT();
736 // Special handling for extended integers.
737 if (SrcVT != DstVT) {
738 if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16)
741 if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
744 assert(DstVT == MVT::i32 && "X86 should always ext to i32");
746 if (SrcVT == MVT::i1) {
747 if (Outs[0].Flags.isSExt())
749 SrcReg = FastEmitZExtFromI1(MVT::i8, SrcReg, /*TODO: Kill=*/false);
752 unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND :
754 SrcReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op,
755 SrcReg, /*TODO: Kill=*/false);
759 unsigned DstReg = VA.getLocReg();
760 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
761 // Avoid a cross-class copy. This is very unlikely.
762 if (!SrcRC->contains(DstReg))
764 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
765 DstReg).addReg(SrcReg);
767 // Mark the register as live out of the function.
768 MRI.addLiveOut(VA.getLocReg());
772 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET));
776 /// X86SelectLoad - Select and emit code to implement load instructions.
778 bool X86FastISel::X86SelectLoad(const Instruction *I) {
780 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
784 if (!X86SelectAddress(I->getOperand(0), AM))
787 unsigned ResultReg = 0;
788 if (X86FastEmitLoad(VT, AM, ResultReg)) {
789 UpdateValueMap(I, ResultReg);
795 static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
796 switch (VT.getSimpleVT().SimpleTy) {
798 case MVT::i8: return X86::CMP8rr;
799 case MVT::i16: return X86::CMP16rr;
800 case MVT::i32: return X86::CMP32rr;
801 case MVT::i64: return X86::CMP64rr;
802 case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0;
803 case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0;
807 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
808 /// of the comparison, return an opcode that works for the compare (e.g.
809 /// CMP32ri) otherwise return 0.
810 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
811 switch (VT.getSimpleVT().SimpleTy) {
812 // Otherwise, we can't fold the immediate into this comparison.
814 case MVT::i8: return X86::CMP8ri;
815 case MVT::i16: return X86::CMP16ri;
816 case MVT::i32: return X86::CMP32ri;
818 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
820 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
821 return X86::CMP64ri32;
826 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
828 unsigned Op0Reg = getRegForValue(Op0);
829 if (Op0Reg == 0) return false;
831 // Handle 'null' like i32/i64 0.
832 if (isa<ConstantPointerNull>(Op1))
833 Op1 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
835 // We have two options: compare with register or immediate. If the RHS of
836 // the compare is an immediate that we can fold into this compare, use
837 // CMPri, otherwise use CMPrr.
838 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
839 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
840 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareImmOpc))
842 .addImm(Op1C->getSExtValue());
847 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
848 if (CompareOpc == 0) return false;
850 unsigned Op1Reg = getRegForValue(Op1);
851 if (Op1Reg == 0) return false;
852 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareOpc))
859 bool X86FastISel::X86SelectCmp(const Instruction *I) {
860 const CmpInst *CI = cast<CmpInst>(I);
863 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
866 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
868 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
869 switch (CI->getPredicate()) {
870 case CmpInst::FCMP_OEQ: {
871 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
874 unsigned EReg = createResultReg(&X86::GR8RegClass);
875 unsigned NPReg = createResultReg(&X86::GR8RegClass);
876 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETEr), EReg);
877 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
878 TII.get(X86::SETNPr), NPReg);
879 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
880 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
881 UpdateValueMap(I, ResultReg);
884 case CmpInst::FCMP_UNE: {
885 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
888 unsigned NEReg = createResultReg(&X86::GR8RegClass);
889 unsigned PReg = createResultReg(&X86::GR8RegClass);
890 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETNEr), NEReg);
891 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETPr), PReg);
892 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::OR8rr),ResultReg)
893 .addReg(PReg).addReg(NEReg);
894 UpdateValueMap(I, ResultReg);
897 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
898 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
899 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
900 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
901 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
902 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
903 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
904 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
905 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
906 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
907 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
908 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
910 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
911 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
912 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
913 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
914 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
915 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
916 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
917 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
918 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
919 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
924 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
928 // Emit a compare of Op0/Op1.
929 if (!X86FastEmitCompare(Op0, Op1, VT))
932 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(SetCCOpc), ResultReg);
933 UpdateValueMap(I, ResultReg);
937 bool X86FastISel::X86SelectZExt(const Instruction *I) {
938 // Handle zero-extension from i1 to i8, which is common.
939 if (I->getType()->isIntegerTy(8) &&
940 I->getOperand(0)->getType()->isIntegerTy(1)) {
941 unsigned ResultReg = getRegForValue(I->getOperand(0));
942 if (ResultReg == 0) return false;
943 // Set the high bits to zero.
944 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
945 if (ResultReg == 0) return false;
946 UpdateValueMap(I, ResultReg);
954 bool X86FastISel::X86SelectBranch(const Instruction *I) {
955 // Unconditional branches are selected by tablegen-generated code.
956 // Handle a conditional branch.
957 const BranchInst *BI = cast<BranchInst>(I);
958 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
959 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
961 // Fold the common case of a conditional branch with a comparison
962 // in the same block (values defined on other blocks may not have
963 // initialized registers).
964 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
965 if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
966 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
968 // Try to take advantage of fallthrough opportunities.
969 CmpInst::Predicate Predicate = CI->getPredicate();
970 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
971 std::swap(TrueMBB, FalseMBB);
972 Predicate = CmpInst::getInversePredicate(Predicate);
975 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
976 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
979 case CmpInst::FCMP_OEQ:
980 std::swap(TrueMBB, FalseMBB);
981 Predicate = CmpInst::FCMP_UNE;
983 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
984 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
985 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
986 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
987 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
988 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
989 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
990 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
991 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
992 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
993 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
994 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
995 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
997 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
998 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
999 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
1000 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
1001 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
1002 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
1003 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
1004 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
1005 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
1006 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
1011 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
1013 std::swap(Op0, Op1);
1015 // Emit a compare of the LHS and RHS, setting the flags.
1016 if (!X86FastEmitCompare(Op0, Op1, VT))
1019 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BranchOpc))
1022 if (Predicate == CmpInst::FCMP_UNE) {
1023 // X86 requires a second branch to handle UNE (and OEQ,
1024 // which is mapped to UNE above).
1025 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JP_4))
1029 FastEmitBranch(FalseMBB, DL);
1030 FuncInfo.MBB->addSuccessor(TrueMBB);
1033 } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
1034 // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
1035 // typically happen for _Bool and C++ bools.
1037 if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
1038 isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
1039 unsigned TestOpc = 0;
1040 switch (SourceVT.SimpleTy) {
1042 case MVT::i8: TestOpc = X86::TEST8ri; break;
1043 case MVT::i16: TestOpc = X86::TEST16ri; break;
1044 case MVT::i32: TestOpc = X86::TEST32ri; break;
1045 case MVT::i64: TestOpc = X86::TEST64ri32; break;
1048 unsigned OpReg = getRegForValue(TI->getOperand(0));
1049 if (OpReg == 0) return false;
1050 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TestOpc))
1051 .addReg(OpReg).addImm(1);
1053 unsigned JmpOpc = X86::JNE_4;
1054 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
1055 std::swap(TrueMBB, FalseMBB);
1059 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(JmpOpc))
1061 FastEmitBranch(FalseMBB, DL);
1062 FuncInfo.MBB->addSuccessor(TrueMBB);
1068 // Otherwise do a clumsy setcc and re-test it.
1069 // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
1070 // in an explicit cast, so make sure to handle that correctly.
1071 unsigned OpReg = getRegForValue(BI->getCondition());
1072 if (OpReg == 0) return false;
1074 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8ri))
1075 .addReg(OpReg).addImm(1);
1076 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JNE_4))
1078 FastEmitBranch(FalseMBB, DL);
1079 FuncInfo.MBB->addSuccessor(TrueMBB);
1083 bool X86FastISel::X86SelectShift(const Instruction *I) {
1084 unsigned CReg = 0, OpReg = 0;
1085 const TargetRegisterClass *RC = NULL;
1086 if (I->getType()->isIntegerTy(8)) {
1088 RC = &X86::GR8RegClass;
1089 switch (I->getOpcode()) {
1090 case Instruction::LShr: OpReg = X86::SHR8rCL; break;
1091 case Instruction::AShr: OpReg = X86::SAR8rCL; break;
1092 case Instruction::Shl: OpReg = X86::SHL8rCL; break;
1093 default: return false;
1095 } else if (I->getType()->isIntegerTy(16)) {
1097 RC = &X86::GR16RegClass;
1098 switch (I->getOpcode()) {
1099 case Instruction::LShr: OpReg = X86::SHR16rCL; break;
1100 case Instruction::AShr: OpReg = X86::SAR16rCL; break;
1101 case Instruction::Shl: OpReg = X86::SHL16rCL; break;
1102 default: return false;
1104 } else if (I->getType()->isIntegerTy(32)) {
1106 RC = &X86::GR32RegClass;
1107 switch (I->getOpcode()) {
1108 case Instruction::LShr: OpReg = X86::SHR32rCL; break;
1109 case Instruction::AShr: OpReg = X86::SAR32rCL; break;
1110 case Instruction::Shl: OpReg = X86::SHL32rCL; break;
1111 default: return false;
1113 } else if (I->getType()->isIntegerTy(64)) {
1115 RC = &X86::GR64RegClass;
1116 switch (I->getOpcode()) {
1117 case Instruction::LShr: OpReg = X86::SHR64rCL; break;
1118 case Instruction::AShr: OpReg = X86::SAR64rCL; break;
1119 case Instruction::Shl: OpReg = X86::SHL64rCL; break;
1120 default: return false;
1127 if (!isTypeLegal(I->getType(), VT))
1130 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1131 if (Op0Reg == 0) return false;
1133 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1134 if (Op1Reg == 0) return false;
1135 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1136 CReg).addReg(Op1Reg);
1138 // The shift instruction uses X86::CL. If we defined a super-register
1139 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
1140 if (CReg != X86::CL)
1141 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1142 TII.get(TargetOpcode::KILL), X86::CL)
1143 .addReg(CReg, RegState::Kill);
1145 unsigned ResultReg = createResultReg(RC);
1146 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpReg), ResultReg)
1148 UpdateValueMap(I, ResultReg);
1152 bool X86FastISel::X86SelectSelect(const Instruction *I) {
1154 if (!isTypeLegal(I->getType(), VT))
1157 // We only use cmov here, if we don't have a cmov instruction bail.
1158 if (!Subtarget->hasCMov()) return false;
1161 const TargetRegisterClass *RC = NULL;
1162 if (VT == MVT::i16) {
1163 Opc = X86::CMOVE16rr;
1164 RC = &X86::GR16RegClass;
1165 } else if (VT == MVT::i32) {
1166 Opc = X86::CMOVE32rr;
1167 RC = &X86::GR32RegClass;
1168 } else if (VT == MVT::i64) {
1169 Opc = X86::CMOVE64rr;
1170 RC = &X86::GR64RegClass;
1175 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1176 if (Op0Reg == 0) return false;
1177 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1178 if (Op1Reg == 0) return false;
1179 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1180 if (Op2Reg == 0) return false;
1182 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1183 .addReg(Op0Reg).addReg(Op0Reg);
1184 unsigned ResultReg = createResultReg(RC);
1185 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
1186 .addReg(Op1Reg).addReg(Op2Reg);
1187 UpdateValueMap(I, ResultReg);
1191 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
1192 // fpext from float to double.
1193 if (Subtarget->hasSSE2() &&
1194 I->getType()->isDoubleTy()) {
1195 const Value *V = I->getOperand(0);
1196 if (V->getType()->isFloatTy()) {
1197 unsigned OpReg = getRegForValue(V);
1198 if (OpReg == 0) return false;
1199 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1200 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1201 TII.get(X86::CVTSS2SDrr), ResultReg)
1203 UpdateValueMap(I, ResultReg);
1211 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
1212 if (Subtarget->hasSSE2()) {
1213 if (I->getType()->isFloatTy()) {
1214 const Value *V = I->getOperand(0);
1215 if (V->getType()->isDoubleTy()) {
1216 unsigned OpReg = getRegForValue(V);
1217 if (OpReg == 0) return false;
1218 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1219 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1220 TII.get(X86::CVTSD2SSrr), ResultReg)
1222 UpdateValueMap(I, ResultReg);
1231 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
1232 if (Subtarget->is64Bit())
1233 // All other cases should be handled by the tblgen generated code.
1235 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1236 EVT DstVT = TLI.getValueType(I->getType());
1238 // This code only handles truncation to byte right now.
1239 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1240 // All other cases should be handled by the tblgen generated code.
1242 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1243 // All other cases should be handled by the tblgen generated code.
1246 unsigned InputReg = getRegForValue(I->getOperand(0));
1248 // Unhandled operand. Halt "fast" selection and bail.
1251 // First issue a copy to GR16_ABCD or GR32_ABCD.
1252 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1253 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1254 unsigned CopyReg = createResultReg(CopyRC);
1255 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1256 CopyReg).addReg(InputReg);
1258 // Then issue an extract_subreg.
1259 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1260 CopyReg, /*Kill=*/true,
1265 UpdateValueMap(I, ResultReg);
1269 bool X86FastISel::IsMemcpySmall(uint64_t Len) {
1270 return Len <= (Subtarget->is64Bit() ? 32 : 16);
1273 bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
1274 X86AddressMode SrcAM, uint64_t Len) {
1276 // Make sure we don't bloat code by inlining very large memcpy's.
1277 if (!IsMemcpySmall(Len))
1280 bool i64Legal = Subtarget->is64Bit();
1282 // We don't care about alignment here since we just emit integer accesses.
1285 if (Len >= 8 && i64Legal)
1297 bool RV = X86FastEmitLoad(VT, SrcAM, Reg);
1298 RV &= X86FastEmitStore(VT, Reg, DestAM);
1299 assert(RV && "Failed to emit load or store??");
1301 unsigned Size = VT.getSizeInBits()/8;
1303 DestAM.Disp += Size;
1310 bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
1311 // FIXME: Handle more intrinsics.
1312 switch (I.getIntrinsicID()) {
1313 default: return false;
1314 case Intrinsic::memcpy: {
1315 const MemCpyInst &MCI = cast<MemCpyInst>(I);
1316 // Don't handle volatile or variable length memcpys.
1317 if (MCI.isVolatile() || !isa<ConstantInt>(MCI.getLength()))
1320 uint64_t Len = cast<ConstantInt>(MCI.getLength())->getZExtValue();
1322 // Get the address of the dest and source addresses.
1323 X86AddressMode DestAM, SrcAM;
1324 if (!X86SelectAddress(MCI.getRawDest(), DestAM) ||
1325 !X86SelectAddress(MCI.getRawSource(), SrcAM))
1328 return TryEmitSmallMemcpy(DestAM, SrcAM, Len);
1331 case Intrinsic::stackprotector: {
1332 // Emit code inline code to store the stack guard onto the stack.
1333 EVT PtrTy = TLI.getPointerTy();
1335 const Value *Op1 = I.getArgOperand(0); // The guard's value.
1336 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
1338 // Grab the frame index.
1340 if (!X86SelectAddress(Slot, AM)) return false;
1341 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
1344 case Intrinsic::dbg_declare: {
1345 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1347 assert(DI->getAddress() && "Null address should be checked earlier!");
1348 if (!X86SelectAddress(DI->getAddress(), AM))
1350 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1351 // FIXME may need to add RegState::Debug to any registers produced,
1352 // although ESP/EBP should be the only ones at the moment.
1353 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
1354 addImm(0).addMetadata(DI->getVariable());
1357 case Intrinsic::trap: {
1358 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TRAP));
1361 case Intrinsic::sadd_with_overflow:
1362 case Intrinsic::uadd_with_overflow: {
1363 // FIXME: Should fold immediates.
1365 // Replace "add with overflow" intrinsics with an "add" instruction followed
1366 // by a seto/setc instruction.
1367 const Function *Callee = I.getCalledFunction();
1369 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1372 if (!isTypeLegal(RetTy, VT))
1375 const Value *Op1 = I.getArgOperand(0);
1376 const Value *Op2 = I.getArgOperand(1);
1377 unsigned Reg1 = getRegForValue(Op1);
1378 unsigned Reg2 = getRegForValue(Op2);
1380 if (Reg1 == 0 || Reg2 == 0)
1381 // FIXME: Handle values *not* in registers.
1387 else if (VT == MVT::i64)
1392 // The call to CreateRegs builds two sequential registers, to store the
1393 // both the the returned values.
1394 unsigned ResultReg = FuncInfo.CreateRegs(I.getType());
1395 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
1396 .addReg(Reg1).addReg(Reg2);
1398 unsigned Opc = X86::SETBr;
1399 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1401 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg+1);
1403 UpdateValueMap(&I, ResultReg, 2);
1409 bool X86FastISel::X86SelectCall(const Instruction *I) {
1410 const CallInst *CI = cast<CallInst>(I);
1411 const Value *Callee = CI->getCalledValue();
1413 // Can't handle inline asm yet.
1414 if (isa<InlineAsm>(Callee))
1417 // Handle intrinsic calls.
1418 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1419 return X86VisitIntrinsicCall(*II);
1421 // Handle only C and fastcc calling conventions for now.
1422 ImmutableCallSite CS(CI);
1423 CallingConv::ID CC = CS.getCallingConv();
1424 if (CC != CallingConv::C && CC != CallingConv::Fast &&
1425 CC != CallingConv::X86_FastCall)
1428 // fastcc with -tailcallopt is intended to provide a guaranteed
1429 // tail call optimization. Fastisel doesn't know how to do that.
1430 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
1433 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1434 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1435 bool isVarArg = FTy->isVarArg();
1437 // Don't know how to handle Win64 varargs yet. Nothing special needed for
1438 // x86-32. Special handling for x86-64 is implemented.
1439 if (isVarArg && Subtarget->isTargetWin64())
1442 // Fast-isel doesn't know about callee-pop yet.
1443 if (Subtarget->IsCalleePop(isVarArg, CC))
1446 // Check whether the function can return without sret-demotion.
1447 SmallVector<ISD::OutputArg, 4> Outs;
1448 SmallVector<uint64_t, 4> Offsets;
1449 GetReturnInfo(I->getType(), CS.getAttributes().getRetAttributes(),
1450 Outs, TLI, &Offsets);
1451 bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
1452 FTy->isVarArg(), Outs, FTy->getContext());
1453 if (!CanLowerReturn)
1456 // Materialize callee address in a register. FIXME: GV address can be
1457 // handled with a CALLpcrel32 instead.
1458 X86AddressMode CalleeAM;
1459 if (!X86SelectCallAddress(Callee, CalleeAM))
1461 unsigned CalleeOp = 0;
1462 const GlobalValue *GV = 0;
1463 if (CalleeAM.GV != 0) {
1465 } else if (CalleeAM.Base.Reg != 0) {
1466 CalleeOp = CalleeAM.Base.Reg;
1470 // Deal with call operands first.
1471 SmallVector<const Value *, 8> ArgVals;
1472 SmallVector<unsigned, 8> Args;
1473 SmallVector<MVT, 8> ArgVTs;
1474 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1475 Args.reserve(CS.arg_size());
1476 ArgVals.reserve(CS.arg_size());
1477 ArgVTs.reserve(CS.arg_size());
1478 ArgFlags.reserve(CS.arg_size());
1479 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1482 ISD::ArgFlagsTy Flags;
1483 unsigned AttrInd = i - CS.arg_begin() + 1;
1484 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1486 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1489 if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) {
1490 const PointerType *Ty = cast<PointerType>(ArgVal->getType());
1491 const Type *ElementTy = Ty->getElementType();
1492 unsigned FrameSize = TD.getTypeAllocSize(ElementTy);
1493 unsigned FrameAlign = CS.getParamAlignment(AttrInd);
1495 FrameAlign = TLI.getByValTypeAlignment(ElementTy);
1497 Flags.setByValSize(FrameSize);
1498 Flags.setByValAlign(FrameAlign);
1499 if (!IsMemcpySmall(FrameSize))
1503 if (CS.paramHasAttr(AttrInd, Attribute::InReg))
1505 if (CS.paramHasAttr(AttrInd, Attribute::Nest))
1508 // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra
1509 // instruction. This is safe because it is common to all fastisel supported
1510 // calling conventions on x86.
1511 if (ConstantInt *CI = dyn_cast<ConstantInt>(ArgVal)) {
1512 if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 ||
1513 CI->getBitWidth() == 16) {
1515 ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext()));
1517 ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext()));
1523 // Passing bools around ends up doing a trunc to i1 and passing it.
1524 // Codegen this as an argument + "and 1".
1525 if (ArgVal->getType()->isIntegerTy(1) && isa<TruncInst>(ArgVal) &&
1526 cast<TruncInst>(ArgVal)->getParent() == I->getParent() &&
1527 ArgVal->hasOneUse()) {
1528 ArgVal = cast<TruncInst>(ArgVal)->getOperand(0);
1529 ArgReg = getRegForValue(ArgVal);
1530 if (ArgReg == 0) return false;
1533 if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false;
1535 ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg,
1536 ArgVal->hasOneUse(), 1);
1538 ArgReg = getRegForValue(ArgVal);
1541 if (ArgReg == 0) return false;
1543 const Type *ArgTy = ArgVal->getType();
1545 if (!isTypeLegal(ArgTy, ArgVT))
1547 if (ArgVT == MVT::x86mmx)
1549 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1550 Flags.setOrigAlign(OriginalAlignment);
1552 Args.push_back(ArgReg);
1553 ArgVals.push_back(ArgVal);
1554 ArgVTs.push_back(ArgVT);
1555 ArgFlags.push_back(Flags);
1558 // Analyze operands of the call, assigning locations to each operand.
1559 SmallVector<CCValAssign, 16> ArgLocs;
1560 CCState CCInfo(CC, isVarArg, TM, ArgLocs, I->getParent()->getContext());
1562 // Allocate shadow area for Win64
1563 if (Subtarget->isTargetWin64())
1564 CCInfo.AllocateStack(32, 8);
1566 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86);
1568 // Get a count of how many bytes are to be pushed on the stack.
1569 unsigned NumBytes = CCInfo.getNextStackOffset();
1571 // Issue CALLSEQ_START
1572 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1573 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
1576 // Process argument: walk the register/memloc assignments, inserting
1578 SmallVector<unsigned, 4> RegArgs;
1579 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1580 CCValAssign &VA = ArgLocs[i];
1581 unsigned Arg = Args[VA.getValNo()];
1582 EVT ArgVT = ArgVTs[VA.getValNo()];
1584 // Promote the value if needed.
1585 switch (VA.getLocInfo()) {
1586 default: llvm_unreachable("Unknown loc info!");
1587 case CCValAssign::Full: break;
1588 case CCValAssign::SExt: {
1589 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
1590 "Unexpected extend");
1591 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1593 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1594 ArgVT = VA.getLocVT();
1597 case CCValAssign::ZExt: {
1598 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
1599 "Unexpected extend");
1600 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1602 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1603 ArgVT = VA.getLocVT();
1606 case CCValAssign::AExt: {
1607 assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
1608 "Unexpected extend");
1609 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1612 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1615 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1618 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1619 ArgVT = VA.getLocVT();
1622 case CCValAssign::BCvt: {
1623 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(),
1624 ISD::BITCAST, Arg, /*TODO: Kill=*/false);
1625 assert(BC != 0 && "Failed to emit a bitcast!");
1627 ArgVT = VA.getLocVT();
1632 if (VA.isRegLoc()) {
1633 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1634 VA.getLocReg()).addReg(Arg);
1635 RegArgs.push_back(VA.getLocReg());
1637 unsigned LocMemOffset = VA.getLocMemOffset();
1639 AM.Base.Reg = StackPtr;
1640 AM.Disp = LocMemOffset;
1641 const Value *ArgVal = ArgVals[VA.getValNo()];
1642 ISD::ArgFlagsTy Flags = ArgFlags[VA.getValNo()];
1644 if (Flags.isByVal()) {
1645 X86AddressMode SrcAM;
1646 SrcAM.Base.Reg = Arg;
1647 bool Res = TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize());
1648 assert(Res && "memcpy length already checked!"); (void)Res;
1649 } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
1650 // If this is a really simple value, emit this with the Value* version
1651 //of X86FastEmitStore. If it isn't simple, we don't want to do this,
1652 // as it can cause us to reevaluate the argument.
1653 X86FastEmitStore(ArgVT, ArgVal, AM);
1655 X86FastEmitStore(ArgVT, Arg, AM);
1660 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1662 if (Subtarget->isPICStyleGOT()) {
1663 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1664 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1665 X86::EBX).addReg(Base);
1668 if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) {
1669 // Count the number of XMM registers allocated.
1670 static const unsigned XMMArgRegs[] = {
1671 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1672 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1674 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1675 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::MOV8ri),
1676 X86::AL).addImm(NumXMMRegs);
1680 MachineInstrBuilder MIB;
1682 // Register-indirect call.
1684 if (Subtarget->isTargetWin64())
1685 CallOpc = X86::WINCALL64r;
1686 else if (Subtarget->is64Bit())
1687 CallOpc = X86::CALL64r;
1689 CallOpc = X86::CALL32r;
1690 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1695 assert(GV && "Not a direct call");
1697 if (Subtarget->isTargetWin64())
1698 CallOpc = X86::WINCALL64pcrel32;
1699 else if (Subtarget->is64Bit())
1700 CallOpc = X86::CALL64pcrel32;
1702 CallOpc = X86::CALLpcrel32;
1704 // See if we need any target-specific flags on the GV operand.
1705 unsigned char OpFlags = 0;
1707 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1708 // external symbols most go through the PLT in PIC mode. If the symbol
1709 // has hidden or protected visibility, or if it is static or local, then
1710 // we don't need to use the PLT - we can directly call it.
1711 if (Subtarget->isTargetELF() &&
1712 TM.getRelocationModel() == Reloc::PIC_ &&
1713 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1714 OpFlags = X86II::MO_PLT;
1715 } else if (Subtarget->isPICStyleStubAny() &&
1716 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1717 (!Subtarget->getTargetTriple().isMacOSX() ||
1718 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
1719 // PC-relative references to external symbols should go through $stub,
1720 // unless we're building with the leopard linker or later, which
1721 // automatically synthesizes these stubs.
1722 OpFlags = X86II::MO_DARWIN_STUB;
1726 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1727 .addGlobalAddress(GV, 0, OpFlags);
1730 // Add an implicit use GOT pointer in EBX.
1731 if (Subtarget->isPICStyleGOT())
1732 MIB.addReg(X86::EBX);
1734 if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64())
1735 MIB.addReg(X86::AL);
1737 // Add implicit physical register uses to the call.
1738 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1739 MIB.addReg(RegArgs[i]);
1741 // Issue CALLSEQ_END
1742 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1743 unsigned NumBytesCallee = 0;
1744 if (!Subtarget->is64Bit() && CS.paramHasAttr(1, Attribute::StructRet))
1746 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
1747 .addImm(NumBytes).addImm(NumBytesCallee);
1749 // Build info for return calling conv lowering code.
1750 // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo.
1751 SmallVector<ISD::InputArg, 32> Ins;
1752 SmallVector<EVT, 4> RetTys;
1753 ComputeValueVTs(TLI, I->getType(), RetTys);
1754 for (unsigned i = 0, e = RetTys.size(); i != e; ++i) {
1756 EVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT);
1757 unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT);
1758 for (unsigned j = 0; j != NumRegs; ++j) {
1759 ISD::InputArg MyFlags;
1760 MyFlags.VT = RegisterVT.getSimpleVT();
1761 MyFlags.Used = !CS.getInstruction()->use_empty();
1762 if (CS.paramHasAttr(0, Attribute::SExt))
1763 MyFlags.Flags.setSExt();
1764 if (CS.paramHasAttr(0, Attribute::ZExt))
1765 MyFlags.Flags.setZExt();
1766 if (CS.paramHasAttr(0, Attribute::InReg))
1767 MyFlags.Flags.setInReg();
1768 Ins.push_back(MyFlags);
1772 // Now handle call return values.
1773 SmallVector<unsigned, 4> UsedRegs;
1774 SmallVector<CCValAssign, 16> RVLocs;
1775 CCState CCRetInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1776 unsigned ResultReg = FuncInfo.CreateRegs(I->getType());
1777 CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
1778 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1779 EVT CopyVT = RVLocs[i].getValVT();
1780 unsigned CopyReg = ResultReg + i;
1782 // If this is a call to a function that returns an fp value on the x87 fp
1783 // stack, but where we prefer to use the value in xmm registers, copy it
1784 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1785 if ((RVLocs[i].getLocReg() == X86::ST0 ||
1786 RVLocs[i].getLocReg() == X86::ST1) &&
1787 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1789 CopyReg = createResultReg(X86::RFP80RegisterClass);
1792 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1793 CopyReg).addReg(RVLocs[i].getLocReg());
1794 UsedRegs.push_back(RVLocs[i].getLocReg());
1796 if (CopyVT != RVLocs[i].getValVT()) {
1797 // Round the F80 the right size, which also moves to the appropriate xmm
1798 // register. This is accomplished by storing the F80 value in memory and
1799 // then loading it back. Ewww...
1800 EVT ResVT = RVLocs[i].getValVT();
1801 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1802 unsigned MemSize = ResVT.getSizeInBits()/8;
1803 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
1804 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1807 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1808 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1809 TII.get(Opc), ResultReg + i), FI);
1814 UpdateValueMap(I, ResultReg, RVLocs.size());
1816 // Set all unused physreg defs as dead.
1817 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1824 X86FastISel::TargetSelectInstruction(const Instruction *I) {
1825 switch (I->getOpcode()) {
1827 case Instruction::Load:
1828 return X86SelectLoad(I);
1829 case Instruction::Store:
1830 return X86SelectStore(I);
1831 case Instruction::Ret:
1832 return X86SelectRet(I);
1833 case Instruction::ICmp:
1834 case Instruction::FCmp:
1835 return X86SelectCmp(I);
1836 case Instruction::ZExt:
1837 return X86SelectZExt(I);
1838 case Instruction::Br:
1839 return X86SelectBranch(I);
1840 case Instruction::Call:
1841 return X86SelectCall(I);
1842 case Instruction::LShr:
1843 case Instruction::AShr:
1844 case Instruction::Shl:
1845 return X86SelectShift(I);
1846 case Instruction::Select:
1847 return X86SelectSelect(I);
1848 case Instruction::Trunc:
1849 return X86SelectTrunc(I);
1850 case Instruction::FPExt:
1851 return X86SelectFPExt(I);
1852 case Instruction::FPTrunc:
1853 return X86SelectFPTrunc(I);
1854 case Instruction::IntToPtr: // Deliberate fall-through.
1855 case Instruction::PtrToInt: {
1856 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1857 EVT DstVT = TLI.getValueType(I->getType());
1858 if (DstVT.bitsGT(SrcVT))
1859 return X86SelectZExt(I);
1860 if (DstVT.bitsLT(SrcVT))
1861 return X86SelectTrunc(I);
1862 unsigned Reg = getRegForValue(I->getOperand(0));
1863 if (Reg == 0) return false;
1864 UpdateValueMap(I, Reg);
1872 unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
1874 if (!isTypeLegal(C->getType(), VT))
1877 // Get opcode and regclass of the output for the given load instruction.
1879 const TargetRegisterClass *RC = NULL;
1880 switch (VT.SimpleTy) {
1881 default: return false;
1884 RC = X86::GR8RegisterClass;
1888 RC = X86::GR16RegisterClass;
1892 RC = X86::GR32RegisterClass;
1895 // Must be in x86-64 mode.
1897 RC = X86::GR64RegisterClass;
1900 if (Subtarget->hasSSE1()) {
1902 RC = X86::FR32RegisterClass;
1904 Opc = X86::LD_Fp32m;
1905 RC = X86::RFP32RegisterClass;
1909 if (Subtarget->hasSSE2()) {
1911 RC = X86::FR64RegisterClass;
1913 Opc = X86::LD_Fp64m;
1914 RC = X86::RFP64RegisterClass;
1918 // No f80 support yet.
1922 // Materialize addresses with LEA instructions.
1923 if (isa<GlobalValue>(C)) {
1925 if (X86SelectAddress(C, AM)) {
1926 // If the expression is just a basereg, then we're done, otherwise we need
1928 if (AM.BaseType == X86AddressMode::RegBase &&
1929 AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == 0)
1932 Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
1933 unsigned ResultReg = createResultReg(RC);
1934 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1935 TII.get(Opc), ResultReg), AM);
1941 // MachineConstantPool wants an explicit alignment.
1942 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1944 // Alignment of vector types. FIXME!
1945 Align = TD.getTypeAllocSize(C->getType());
1948 // x86-32 PIC requires a PIC base register for constant pools.
1949 unsigned PICBase = 0;
1950 unsigned char OpFlag = 0;
1951 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1952 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1953 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1954 } else if (Subtarget->isPICStyleGOT()) {
1955 OpFlag = X86II::MO_GOTOFF;
1956 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1957 } else if (Subtarget->isPICStyleRIPRel() &&
1958 TM.getCodeModel() == CodeModel::Small) {
1962 // Create the load from the constant pool.
1963 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1964 unsigned ResultReg = createResultReg(RC);
1965 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1966 TII.get(Opc), ResultReg),
1967 MCPOffset, PICBase, OpFlag);
1972 unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
1973 // Fail on dynamic allocas. At this point, getRegForValue has already
1974 // checked its CSE maps, so if we're here trying to handle a dynamic
1975 // alloca, we're not going to succeed. X86SelectAddress has a
1976 // check for dynamic allocas, because it's called directly from
1977 // various places, but TargetMaterializeAlloca also needs a check
1978 // in order to avoid recursion between getRegForValue,
1979 // X86SelectAddrss, and TargetMaterializeAlloca.
1980 if (!FuncInfo.StaticAllocaMap.count(C))
1984 if (!X86SelectAddress(C, AM))
1986 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1987 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1988 unsigned ResultReg = createResultReg(RC);
1989 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1990 TII.get(Opc), ResultReg), AM);
1994 unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) {
1996 if (!isTypeLegal(CF->getType(), VT))
1999 // Get opcode and regclass for the given zero.
2001 const TargetRegisterClass *RC = NULL;
2002 switch (VT.SimpleTy) {
2003 default: return false;
2005 if (Subtarget->hasSSE1()) {
2006 Opc = X86::FsFLD0SS;
2007 RC = X86::FR32RegisterClass;
2009 Opc = X86::LD_Fp032;
2010 RC = X86::RFP32RegisterClass;
2014 if (Subtarget->hasSSE2()) {
2015 Opc = X86::FsFLD0SD;
2016 RC = X86::FR64RegisterClass;
2018 Opc = X86::LD_Fp064;
2019 RC = X86::RFP64RegisterClass;
2023 // No f80 support yet.
2027 unsigned ResultReg = createResultReg(RC);
2028 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
2033 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
2034 /// vreg is being provided by the specified load instruction. If possible,
2035 /// try to fold the load as an operand to the instruction, returning true if
2037 bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
2038 const LoadInst *LI) {
2040 if (!X86SelectAddress(LI->getOperand(0), AM))
2043 X86InstrInfo &XII = (X86InstrInfo&)TII;
2045 unsigned Size = TD.getTypeAllocSize(LI->getType());
2046 unsigned Alignment = LI->getAlignment();
2048 SmallVector<MachineOperand, 8> AddrOps;
2049 AM.getFullAddress(AddrOps);
2051 MachineInstr *Result =
2052 XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment);
2053 if (Result == 0) return false;
2055 FuncInfo.MBB->insert(FuncInfo.InsertPt, Result);
2056 MI->eraseFromParent();
2062 llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
2063 return new X86FastISel(funcInfo);