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/CodeGen/Analysis.h"
27 #include "llvm/CodeGen/FastISel.h"
28 #include "llvm/CodeGen/FunctionLoweringInfo.h"
29 #include "llvm/CodeGen/MachineConstantPool.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/Support/CallSite.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/GetElementPtrTypeIterator.h"
35 #include "llvm/Target/TargetOptions.h"
40 class X86FastISel : public FastISel {
41 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
42 /// make the right decision when generating code for different targets.
43 const X86Subtarget *Subtarget;
45 /// StackPtr - Register used as the stack pointer.
49 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
50 /// floating point ops.
51 /// When SSE is available, use it for f32 operations.
52 /// When SSE2 is available, use it for f64 operations.
57 explicit X86FastISel(FunctionLoweringInfo &funcInfo) : FastISel(funcInfo) {
58 Subtarget = &TM.getSubtarget<X86Subtarget>();
59 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
60 X86ScalarSSEf64 = Subtarget->hasSSE2();
61 X86ScalarSSEf32 = Subtarget->hasSSE1();
64 virtual bool TargetSelectInstruction(const Instruction *I);
66 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
67 /// vreg is being provided by the specified load instruction. If possible,
68 /// try to fold the load as an operand to the instruction, returning true if
70 virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
73 #include "X86GenFastISel.inc"
76 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT);
78 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
80 bool X86FastEmitStore(EVT VT, const Value *Val,
81 const X86AddressMode &AM);
82 bool X86FastEmitStore(EVT VT, unsigned Val,
83 const X86AddressMode &AM);
85 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
88 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
89 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
91 bool X86SelectLoad(const Instruction *I);
93 bool X86SelectStore(const Instruction *I);
95 bool X86SelectRet(const Instruction *I);
97 bool X86SelectCmp(const Instruction *I);
99 bool X86SelectZExt(const Instruction *I);
101 bool X86SelectBranch(const Instruction *I);
103 bool X86SelectShift(const Instruction *I);
105 bool X86SelectSelect(const Instruction *I);
107 bool X86SelectTrunc(const Instruction *I);
109 bool X86SelectFPExt(const Instruction *I);
110 bool X86SelectFPTrunc(const Instruction *I);
112 bool X86SelectExtractValue(const Instruction *I);
114 bool X86VisitIntrinsicCall(const IntrinsicInst &I);
115 bool X86SelectCall(const Instruction *I);
117 const X86InstrInfo *getInstrInfo() const {
118 return getTargetMachine()->getInstrInfo();
120 const X86TargetMachine *getTargetMachine() const {
121 return static_cast<const X86TargetMachine *>(&TM);
124 unsigned TargetMaterializeConstant(const Constant *C);
126 unsigned TargetMaterializeAlloca(const AllocaInst *C);
128 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
129 /// computed in an SSE register, not on the X87 floating point stack.
130 bool isScalarFPTypeInSSEReg(EVT VT) const {
131 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
132 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
135 bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
138 } // end anonymous namespace.
140 bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
141 EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true);
142 if (evt == MVT::Other || !evt.isSimple())
143 // Unhandled type. Halt "fast" selection and bail.
146 VT = evt.getSimpleVT();
147 // For now, require SSE/SSE2 for performing floating-point operations,
148 // since x87 requires additional work.
149 if (VT == MVT::f64 && !X86ScalarSSEf64)
151 if (VT == MVT::f32 && !X86ScalarSSEf32)
153 // Similarly, no f80 support yet.
156 // We only handle legal types. For example, on x86-32 the instruction
157 // selector contains all of the 64-bit instructions from x86-64,
158 // under the assumption that i64 won't be used if the target doesn't
160 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
163 #include "X86GenCallingConv.inc"
165 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
166 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
167 /// Return true and the result register by reference if it is possible.
168 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
169 unsigned &ResultReg) {
170 // Get opcode and regclass of the output for the given load instruction.
172 const TargetRegisterClass *RC = NULL;
173 switch (VT.getSimpleVT().SimpleTy) {
174 default: return false;
178 RC = X86::GR8RegisterClass;
182 RC = X86::GR16RegisterClass;
186 RC = X86::GR32RegisterClass;
189 // Must be in x86-64 mode.
191 RC = X86::GR64RegisterClass;
194 if (Subtarget->hasSSE1()) {
196 RC = X86::FR32RegisterClass;
199 RC = X86::RFP32RegisterClass;
203 if (Subtarget->hasSSE2()) {
205 RC = X86::FR64RegisterClass;
208 RC = X86::RFP64RegisterClass;
212 // No f80 support yet.
216 ResultReg = createResultReg(RC);
217 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
218 DL, TII.get(Opc), ResultReg), AM);
222 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
223 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
224 /// and a displacement offset, or a GlobalAddress,
225 /// i.e. V. Return true if it is possible.
227 X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
228 const X86AddressMode &AM) {
229 // Get opcode and regclass of the output for the given store instruction.
231 switch (VT.getSimpleVT().SimpleTy) {
232 case MVT::f80: // No f80 support yet.
233 default: return false;
235 // Mask out all but lowest bit.
236 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
237 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
238 TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
241 // FALLTHROUGH, handling i1 as i8.
242 case MVT::i8: Opc = X86::MOV8mr; break;
243 case MVT::i16: Opc = X86::MOV16mr; break;
244 case MVT::i32: Opc = X86::MOV32mr; break;
245 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
247 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
250 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
254 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
255 DL, TII.get(Opc)), AM).addReg(Val);
259 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
260 const X86AddressMode &AM) {
261 // Handle 'null' like i32/i64 0.
262 if (isa<ConstantPointerNull>(Val))
263 Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext()));
265 // If this is a store of a simple constant, fold the constant into the store.
266 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
269 switch (VT.getSimpleVT().SimpleTy) {
271 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
272 case MVT::i8: Opc = X86::MOV8mi; break;
273 case MVT::i16: Opc = X86::MOV16mi; break;
274 case MVT::i32: Opc = X86::MOV32mi; break;
276 // Must be a 32-bit sign extended value.
277 if ((int)CI->getSExtValue() == CI->getSExtValue())
278 Opc = X86::MOV64mi32;
283 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
284 DL, TII.get(Opc)), AM)
285 .addImm(Signed ? (uint64_t) CI->getSExtValue() :
291 unsigned ValReg = getRegForValue(Val);
295 return X86FastEmitStore(VT, ValReg, AM);
298 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
299 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
300 /// ISD::SIGN_EXTEND).
301 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
302 unsigned Src, EVT SrcVT,
303 unsigned &ResultReg) {
304 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
305 Src, /*TODO: Kill=*/false);
314 /// X86SelectAddress - Attempt to fill in an address from the given value.
316 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
317 const User *U = NULL;
318 unsigned Opcode = Instruction::UserOp1;
319 if (const Instruction *I = dyn_cast<Instruction>(V)) {
320 // Don't walk into other basic blocks; it's possible we haven't
321 // visited them yet, so the instructions may not yet be assigned
322 // virtual registers.
323 if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
324 FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
325 Opcode = I->getOpcode();
328 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
329 Opcode = C->getOpcode();
333 if (const PointerType *Ty = dyn_cast<PointerType>(V->getType()))
334 if (Ty->getAddressSpace() > 255)
335 // Fast instruction selection doesn't support the special
341 case Instruction::BitCast:
342 // Look past bitcasts.
343 return X86SelectAddress(U->getOperand(0), AM);
345 case Instruction::IntToPtr:
346 // Look past no-op inttoptrs.
347 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
348 return X86SelectAddress(U->getOperand(0), AM);
351 case Instruction::PtrToInt:
352 // Look past no-op ptrtoints.
353 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
354 return X86SelectAddress(U->getOperand(0), AM);
357 case Instruction::Alloca: {
358 // Do static allocas.
359 const AllocaInst *A = cast<AllocaInst>(V);
360 DenseMap<const AllocaInst*, int>::iterator SI =
361 FuncInfo.StaticAllocaMap.find(A);
362 if (SI != FuncInfo.StaticAllocaMap.end()) {
363 AM.BaseType = X86AddressMode::FrameIndexBase;
364 AM.Base.FrameIndex = SI->second;
370 case Instruction::Add: {
371 // Adds of constants are common and easy enough.
372 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
373 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
374 // They have to fit in the 32-bit signed displacement field though.
375 if (isInt<32>(Disp)) {
376 AM.Disp = (uint32_t)Disp;
377 return X86SelectAddress(U->getOperand(0), AM);
383 case Instruction::GetElementPtr: {
384 X86AddressMode SavedAM = AM;
386 // Pattern-match simple GEPs.
387 uint64_t Disp = (int32_t)AM.Disp;
388 unsigned IndexReg = AM.IndexReg;
389 unsigned Scale = AM.Scale;
390 gep_type_iterator GTI = gep_type_begin(U);
391 // Iterate through the indices, folding what we can. Constants can be
392 // folded, and one dynamic index can be handled, if the scale is supported.
393 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
394 i != e; ++i, ++GTI) {
395 const Value *Op = *i;
396 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
397 const StructLayout *SL = TD.getStructLayout(STy);
398 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
399 Disp += SL->getElementOffset(Idx);
401 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
403 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
404 // Constant-offset addressing.
405 Disp += CI->getSExtValue() * S;
408 if (isa<AddOperator>(Op) &&
409 (!isa<Instruction>(Op) ||
410 FuncInfo.MBBMap[cast<Instruction>(Op)->getParent()]
412 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
413 // An add (in the same block) with a constant operand. Fold the
416 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
417 Disp += CI->getSExtValue() * S;
418 // Iterate on the other operand.
419 Op = cast<AddOperator>(Op)->getOperand(0);
423 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
424 (S == 1 || S == 2 || S == 4 || S == 8)) {
425 // Scaled-index addressing.
427 IndexReg = getRegForGEPIndex(Op).first;
433 goto unsupported_gep;
437 // Check for displacement overflow.
438 if (!isInt<32>(Disp))
440 // Ok, the GEP indices were covered by constant-offset and scaled-index
441 // addressing. Update the address state and move on to examining the base.
442 AM.IndexReg = IndexReg;
444 AM.Disp = (uint32_t)Disp;
445 if (X86SelectAddress(U->getOperand(0), AM))
448 // If we couldn't merge the sub value into this addr mode, revert back to
449 // our address and just match the value instead of completely failing.
453 // Ok, the GEP indices weren't all covered.
458 // Handle constant address.
459 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
460 // Can't handle alternate code models yet.
461 if (TM.getCodeModel() != CodeModel::Small)
464 // RIP-relative addresses can't have additional register operands.
465 if (Subtarget->isPICStyleRIPRel() &&
466 (AM.Base.Reg != 0 || AM.IndexReg != 0))
469 // Can't handle TLS yet.
470 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
471 if (GVar->isThreadLocal())
474 // Okay, we've committed to selecting this global. Set up the basic address.
477 // Allow the subtarget to classify the global.
478 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
480 // If this reference is relative to the pic base, set it now.
481 if (isGlobalRelativeToPICBase(GVFlags)) {
482 // FIXME: How do we know Base.Reg is free??
483 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
486 // Unless the ABI requires an extra load, return a direct reference to
488 if (!isGlobalStubReference(GVFlags)) {
489 if (Subtarget->isPICStyleRIPRel()) {
490 // Use rip-relative addressing if we can. Above we verified that the
491 // base and index registers are unused.
492 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
493 AM.Base.Reg = X86::RIP;
495 AM.GVOpFlags = GVFlags;
499 // Ok, we need to do a load from a stub. If we've already loaded from this
500 // stub, reuse the loaded pointer, otherwise emit the load now.
501 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
503 if (I != LocalValueMap.end() && I->second != 0) {
506 // Issue load from stub.
508 const TargetRegisterClass *RC = NULL;
509 X86AddressMode StubAM;
510 StubAM.Base.Reg = AM.Base.Reg;
512 StubAM.GVOpFlags = GVFlags;
514 // Prepare for inserting code in the local-value area.
515 SavePoint SaveInsertPt = enterLocalValueArea();
517 if (TLI.getPointerTy() == MVT::i64) {
519 RC = X86::GR64RegisterClass;
521 if (Subtarget->isPICStyleRIPRel())
522 StubAM.Base.Reg = X86::RIP;
525 RC = X86::GR32RegisterClass;
528 LoadReg = createResultReg(RC);
529 MachineInstrBuilder LoadMI =
530 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg);
531 addFullAddress(LoadMI, StubAM);
533 // Ok, back to normal mode.
534 leaveLocalValueArea(SaveInsertPt);
536 // Prevent loading GV stub multiple times in same MBB.
537 LocalValueMap[V] = LoadReg;
540 // Now construct the final address. Note that the Disp, Scale,
541 // and Index values may already be set here.
542 AM.Base.Reg = LoadReg;
547 // If all else fails, try to materialize the value in a register.
548 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
549 if (AM.Base.Reg == 0) {
550 AM.Base.Reg = getRegForValue(V);
551 return AM.Base.Reg != 0;
553 if (AM.IndexReg == 0) {
554 assert(AM.Scale == 1 && "Scale with no index!");
555 AM.IndexReg = getRegForValue(V);
556 return AM.IndexReg != 0;
563 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
565 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
566 const User *U = NULL;
567 unsigned Opcode = Instruction::UserOp1;
568 if (const Instruction *I = dyn_cast<Instruction>(V)) {
569 Opcode = I->getOpcode();
571 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
572 Opcode = C->getOpcode();
578 case Instruction::BitCast:
579 // Look past bitcasts.
580 return X86SelectCallAddress(U->getOperand(0), AM);
582 case Instruction::IntToPtr:
583 // Look past no-op inttoptrs.
584 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
585 return X86SelectCallAddress(U->getOperand(0), AM);
588 case Instruction::PtrToInt:
589 // Look past no-op ptrtoints.
590 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
591 return X86SelectCallAddress(U->getOperand(0), AM);
595 // Handle constant address.
596 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
597 // Can't handle alternate code models yet.
598 if (TM.getCodeModel() != CodeModel::Small)
601 // RIP-relative addresses can't have additional register operands.
602 if (Subtarget->isPICStyleRIPRel() &&
603 (AM.Base.Reg != 0 || AM.IndexReg != 0))
606 // Can't handle DLLImport.
607 if (GV->hasDLLImportLinkage())
611 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
612 if (GVar->isThreadLocal())
615 // Okay, we've committed to selecting this global. Set up the basic address.
618 // No ABI requires an extra load for anything other than DLLImport, which
619 // we rejected above. Return a direct reference to the global.
620 if (Subtarget->isPICStyleRIPRel()) {
621 // Use rip-relative addressing if we can. Above we verified that the
622 // base and index registers are unused.
623 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
624 AM.Base.Reg = X86::RIP;
625 } else if (Subtarget->isPICStyleStubPIC()) {
626 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
627 } else if (Subtarget->isPICStyleGOT()) {
628 AM.GVOpFlags = X86II::MO_GOTOFF;
634 // If all else fails, try to materialize the value in a register.
635 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
636 if (AM.Base.Reg == 0) {
637 AM.Base.Reg = getRegForValue(V);
638 return AM.Base.Reg != 0;
640 if (AM.IndexReg == 0) {
641 assert(AM.Scale == 1 && "Scale with no index!");
642 AM.IndexReg = getRegForValue(V);
643 return AM.IndexReg != 0;
651 /// X86SelectStore - Select and emit code to implement store instructions.
652 bool X86FastISel::X86SelectStore(const Instruction *I) {
654 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
658 if (!X86SelectAddress(I->getOperand(1), AM))
661 return X86FastEmitStore(VT, I->getOperand(0), AM);
664 /// X86SelectRet - Select and emit code to implement ret instructions.
665 bool X86FastISel::X86SelectRet(const Instruction *I) {
666 const ReturnInst *Ret = cast<ReturnInst>(I);
667 const Function &F = *I->getParent()->getParent();
669 if (!FuncInfo.CanLowerReturn)
672 CallingConv::ID CC = F.getCallingConv();
673 if (CC != CallingConv::C &&
674 CC != CallingConv::Fast &&
675 CC != CallingConv::X86_FastCall)
678 if (Subtarget->isTargetWin64())
681 // Don't handle popping bytes on return for now.
682 if (FuncInfo.MF->getInfo<X86MachineFunctionInfo>()
683 ->getBytesToPopOnReturn() != 0)
686 // fastcc with -tailcallopt is intended to provide a guaranteed
687 // tail call optimization. Fastisel doesn't know how to do that.
688 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
691 // Let SDISel handle vararg functions.
695 if (Ret->getNumOperands() > 0) {
696 SmallVector<ISD::OutputArg, 4> Outs;
697 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
700 // Analyze operands of the call, assigning locations to each operand.
701 SmallVector<CCValAssign, 16> ValLocs;
702 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
703 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
705 const Value *RV = Ret->getOperand(0);
706 unsigned Reg = getRegForValue(RV);
710 // Only handle a single return value for now.
711 if (ValLocs.size() != 1)
714 CCValAssign &VA = ValLocs[0];
716 // Don't bother handling odd stuff for now.
717 if (VA.getLocInfo() != CCValAssign::Full)
719 // Only handle register returns for now.
722 // TODO: For now, don't try to handle cases where getLocInfo()
723 // says Full but the types don't match.
724 if (TLI.getValueType(RV->getType()) != VA.getValVT())
727 // The calling-convention tables for x87 returns don't tell
729 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
733 unsigned SrcReg = Reg + VA.getValNo();
734 unsigned DstReg = VA.getLocReg();
735 const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
736 // Avoid a cross-class copy. This is very unlikely.
737 if (!SrcRC->contains(DstReg))
739 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
740 DstReg).addReg(SrcReg);
742 // Mark the register as live out of the function.
743 MRI.addLiveOut(VA.getLocReg());
747 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET));
751 /// X86SelectLoad - Select and emit code to implement load instructions.
753 bool X86FastISel::X86SelectLoad(const Instruction *I) {
755 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
759 if (!X86SelectAddress(I->getOperand(0), AM))
762 unsigned ResultReg = 0;
763 if (X86FastEmitLoad(VT, AM, ResultReg)) {
764 UpdateValueMap(I, ResultReg);
770 static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
771 switch (VT.getSimpleVT().SimpleTy) {
773 case MVT::i8: return X86::CMP8rr;
774 case MVT::i16: return X86::CMP16rr;
775 case MVT::i32: return X86::CMP32rr;
776 case MVT::i64: return X86::CMP64rr;
777 case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0;
778 case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0;
782 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
783 /// of the comparison, return an opcode that works for the compare (e.g.
784 /// CMP32ri) otherwise return 0.
785 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
786 switch (VT.getSimpleVT().SimpleTy) {
787 // Otherwise, we can't fold the immediate into this comparison.
789 case MVT::i8: return X86::CMP8ri;
790 case MVT::i16: return X86::CMP16ri;
791 case MVT::i32: return X86::CMP32ri;
793 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
795 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
796 return X86::CMP64ri32;
801 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
803 unsigned Op0Reg = getRegForValue(Op0);
804 if (Op0Reg == 0) return false;
806 // Handle 'null' like i32/i64 0.
807 if (isa<ConstantPointerNull>(Op1))
808 Op1 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
810 // We have two options: compare with register or immediate. If the RHS of
811 // the compare is an immediate that we can fold into this compare, use
812 // CMPri, otherwise use CMPrr.
813 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
814 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
815 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareImmOpc))
817 .addImm(Op1C->getSExtValue());
822 unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
823 if (CompareOpc == 0) return false;
825 unsigned Op1Reg = getRegForValue(Op1);
826 if (Op1Reg == 0) return false;
827 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareOpc))
834 bool X86FastISel::X86SelectCmp(const Instruction *I) {
835 const CmpInst *CI = cast<CmpInst>(I);
838 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
841 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
843 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
844 switch (CI->getPredicate()) {
845 case CmpInst::FCMP_OEQ: {
846 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
849 unsigned EReg = createResultReg(&X86::GR8RegClass);
850 unsigned NPReg = createResultReg(&X86::GR8RegClass);
851 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETEr), EReg);
852 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
853 TII.get(X86::SETNPr), NPReg);
854 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
855 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
856 UpdateValueMap(I, ResultReg);
859 case CmpInst::FCMP_UNE: {
860 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
863 unsigned NEReg = createResultReg(&X86::GR8RegClass);
864 unsigned PReg = createResultReg(&X86::GR8RegClass);
865 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
866 TII.get(X86::SETNEr), NEReg);
867 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
868 TII.get(X86::SETPr), PReg);
869 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
870 TII.get(X86::OR8rr), ResultReg)
871 .addReg(PReg).addReg(NEReg);
872 UpdateValueMap(I, ResultReg);
875 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
876 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
877 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
878 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
879 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
880 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
881 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
882 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
883 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
884 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
885 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
886 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
888 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
889 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
890 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
891 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
892 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
893 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
894 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
895 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
896 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
897 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
902 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
906 // Emit a compare of Op0/Op1.
907 if (!X86FastEmitCompare(Op0, Op1, VT))
910 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(SetCCOpc), ResultReg);
911 UpdateValueMap(I, ResultReg);
915 bool X86FastISel::X86SelectZExt(const Instruction *I) {
916 // Handle zero-extension from i1 to i8, which is common.
917 if (I->getType()->isIntegerTy(8) &&
918 I->getOperand(0)->getType()->isIntegerTy(1)) {
919 unsigned ResultReg = getRegForValue(I->getOperand(0));
920 if (ResultReg == 0) return false;
921 // Set the high bits to zero.
922 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
923 if (ResultReg == 0) return false;
924 UpdateValueMap(I, ResultReg);
932 bool X86FastISel::X86SelectBranch(const Instruction *I) {
933 // Unconditional branches are selected by tablegen-generated code.
934 // Handle a conditional branch.
935 const BranchInst *BI = cast<BranchInst>(I);
936 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
937 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
939 // Fold the common case of a conditional branch with a comparison
940 // in the same block (values defined on other blocks may not have
941 // initialized registers).
942 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
943 if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
944 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
946 // Try to take advantage of fallthrough opportunities.
947 CmpInst::Predicate Predicate = CI->getPredicate();
948 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
949 std::swap(TrueMBB, FalseMBB);
950 Predicate = CmpInst::getInversePredicate(Predicate);
953 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
954 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
957 case CmpInst::FCMP_OEQ:
958 std::swap(TrueMBB, FalseMBB);
959 Predicate = CmpInst::FCMP_UNE;
961 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
962 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
963 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
964 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
965 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
966 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
967 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
968 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
969 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
970 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
971 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
972 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
973 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
975 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
976 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
977 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
978 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
979 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
980 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
981 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
982 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
983 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
984 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
989 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
993 // Emit a compare of the LHS and RHS, setting the flags.
994 if (!X86FastEmitCompare(Op0, Op1, VT))
997 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BranchOpc))
1000 if (Predicate == CmpInst::FCMP_UNE) {
1001 // X86 requires a second branch to handle UNE (and OEQ,
1002 // which is mapped to UNE above).
1003 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JP_4))
1007 FastEmitBranch(FalseMBB, DL);
1008 FuncInfo.MBB->addSuccessor(TrueMBB);
1011 } else if (ExtractValueInst *EI =
1012 dyn_cast<ExtractValueInst>(BI->getCondition())) {
1013 // Check to see if the branch instruction is from an "arithmetic with
1014 // overflow" intrinsic. The main way these intrinsics are used is:
1016 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
1017 // %sum = extractvalue { i32, i1 } %t, 0
1018 // %obit = extractvalue { i32, i1 } %t, 1
1019 // br i1 %obit, label %overflow, label %normal
1021 // The %sum and %obit are converted in an ADD and a SETO/SETB before
1022 // reaching the branch. Therefore, we search backwards through the MBB
1023 // looking for the SETO/SETB instruction. If an instruction modifies the
1024 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
1025 // convert the branch into a JO/JB instruction.
1026 if (const IntrinsicInst *CI =
1027 dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
1028 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
1029 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
1030 const MachineInstr *SetMI = 0;
1031 unsigned Reg = getRegForValue(EI);
1033 for (MachineBasicBlock::const_reverse_iterator
1034 RI = FuncInfo.MBB->rbegin(), RE = FuncInfo.MBB->rend();
1036 const MachineInstr &MI = *RI;
1038 if (MI.definesRegister(Reg)) {
1040 Reg = MI.getOperand(1).getReg();
1048 const TargetInstrDesc &TID = MI.getDesc();
1049 if (TID.hasImplicitDefOfPhysReg(X86::EFLAGS) ||
1050 MI.hasUnmodeledSideEffects())
1055 unsigned OpCode = SetMI->getOpcode();
1057 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
1058 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1059 TII.get(OpCode == X86::SETOr ? X86::JO_4 : X86::JB_4))
1061 FastEmitBranch(FalseMBB, DL);
1062 FuncInfo.MBB->addSuccessor(TrueMBB);
1070 // Otherwise do a clumsy setcc and re-test it.
1071 unsigned OpReg = getRegForValue(BI->getCondition());
1072 if (OpReg == 0) return false;
1074 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1075 .addReg(OpReg).addReg(OpReg);
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, OpImm = 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; OpImm = X86::SHR8ri; break;
1091 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
1092 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; 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; OpImm = X86::SHR16ri; break;
1100 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
1101 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; 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; OpImm = X86::SHR32ri; break;
1109 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
1110 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; 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; OpImm = X86::SHR64ri; break;
1118 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
1119 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; 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 // Fold immediate in shl(x,3).
1134 if (const ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
1135 unsigned ResultReg = createResultReg(RC);
1136 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpImm),
1137 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
1138 UpdateValueMap(I, ResultReg);
1142 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1143 if (Op1Reg == 0) return false;
1144 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1145 CReg).addReg(Op1Reg);
1147 // The shift instruction uses X86::CL. If we defined a super-register
1148 // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
1149 if (CReg != X86::CL)
1150 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1151 TII.get(TargetOpcode::KILL), X86::CL)
1152 .addReg(CReg, RegState::Kill);
1154 unsigned ResultReg = createResultReg(RC);
1155 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpReg), ResultReg)
1157 UpdateValueMap(I, ResultReg);
1161 bool X86FastISel::X86SelectSelect(const Instruction *I) {
1163 if (!isTypeLegal(I->getType(), VT))
1166 // We only use cmov here, if we don't have a cmov instruction bail.
1167 if (!Subtarget->hasCMov()) return false;
1170 const TargetRegisterClass *RC = NULL;
1171 if (VT == MVT::i16) {
1172 Opc = X86::CMOVE16rr;
1173 RC = &X86::GR16RegClass;
1174 } else if (VT == MVT::i32) {
1175 Opc = X86::CMOVE32rr;
1176 RC = &X86::GR32RegClass;
1177 } else if (VT == MVT::i64) {
1178 Opc = X86::CMOVE64rr;
1179 RC = &X86::GR64RegClass;
1184 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1185 if (Op0Reg == 0) return false;
1186 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1187 if (Op1Reg == 0) return false;
1188 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1189 if (Op2Reg == 0) return false;
1191 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1192 .addReg(Op0Reg).addReg(Op0Reg);
1193 unsigned ResultReg = createResultReg(RC);
1194 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
1195 .addReg(Op1Reg).addReg(Op2Reg);
1196 UpdateValueMap(I, ResultReg);
1200 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
1201 // fpext from float to double.
1202 if (Subtarget->hasSSE2() &&
1203 I->getType()->isDoubleTy()) {
1204 const Value *V = I->getOperand(0);
1205 if (V->getType()->isFloatTy()) {
1206 unsigned OpReg = getRegForValue(V);
1207 if (OpReg == 0) return false;
1208 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1209 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1210 TII.get(X86::CVTSS2SDrr), ResultReg)
1212 UpdateValueMap(I, ResultReg);
1220 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
1221 if (Subtarget->hasSSE2()) {
1222 if (I->getType()->isFloatTy()) {
1223 const Value *V = I->getOperand(0);
1224 if (V->getType()->isDoubleTy()) {
1225 unsigned OpReg = getRegForValue(V);
1226 if (OpReg == 0) return false;
1227 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1228 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1229 TII.get(X86::CVTSD2SSrr), ResultReg)
1231 UpdateValueMap(I, ResultReg);
1240 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
1241 if (Subtarget->is64Bit())
1242 // All other cases should be handled by the tblgen generated code.
1244 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1245 EVT DstVT = TLI.getValueType(I->getType());
1247 // This code only handles truncation to byte right now.
1248 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1249 // All other cases should be handled by the tblgen generated code.
1251 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1252 // All other cases should be handled by the tblgen generated code.
1255 unsigned InputReg = getRegForValue(I->getOperand(0));
1257 // Unhandled operand. Halt "fast" selection and bail.
1260 // First issue a copy to GR16_ABCD or GR32_ABCD.
1261 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1262 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1263 unsigned CopyReg = createResultReg(CopyRC);
1264 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1265 CopyReg).addReg(InputReg);
1267 // Then issue an extract_subreg.
1268 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1269 CopyReg, /*Kill=*/true,
1274 UpdateValueMap(I, ResultReg);
1278 bool X86FastISel::X86SelectExtractValue(const Instruction *I) {
1279 const ExtractValueInst *EI = cast<ExtractValueInst>(I);
1280 const Value *Agg = EI->getAggregateOperand();
1282 if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1283 switch (CI->getIntrinsicID()) {
1285 case Intrinsic::sadd_with_overflow:
1286 case Intrinsic::uadd_with_overflow: {
1287 // Cheat a little. We know that the registers for "add" and "seto" are
1288 // allocated sequentially. However, we only keep track of the register
1289 // for "add" in the value map. Use extractvalue's index to get the
1290 // correct register for "seto".
1291 unsigned OpReg = getRegForValue(Agg);
1294 UpdateValueMap(I, OpReg + *EI->idx_begin());
1303 bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
1304 // FIXME: Handle more intrinsics.
1305 switch (I.getIntrinsicID()) {
1306 default: return false;
1307 case Intrinsic::stackprotector: {
1308 // Emit code inline code to store the stack guard onto the stack.
1309 EVT PtrTy = TLI.getPointerTy();
1311 const Value *Op1 = I.getArgOperand(0); // The guard's value.
1312 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
1314 // Grab the frame index.
1316 if (!X86SelectAddress(Slot, AM)) return false;
1318 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
1322 case Intrinsic::objectsize: {
1323 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
1324 const Type *Ty = I.getCalledFunction()->getReturnType();
1326 assert(CI && "Non-constant type in Intrinsic::objectsize?");
1329 if (!isTypeLegal(Ty, VT))
1335 else if (VT == MVT::i64)
1340 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1341 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg).
1342 addImm(CI->isZero() ? -1ULL : 0);
1343 UpdateValueMap(&I, ResultReg);
1346 case Intrinsic::dbg_declare: {
1347 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1349 assert(DI->getAddress() && "Null address should be checked earlier!");
1350 if (!X86SelectAddress(DI->getAddress(), AM))
1352 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1353 // FIXME may need to add RegState::Debug to any registers produced,
1354 // although ESP/EBP should be the only ones at the moment.
1355 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
1356 addImm(0).addMetadata(DI->getVariable());
1359 case Intrinsic::trap: {
1360 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TRAP));
1363 case Intrinsic::sadd_with_overflow:
1364 case Intrinsic::uadd_with_overflow: {
1365 // Replace "add with overflow" intrinsics with an "add" instruction followed
1366 // by a seto/setc instruction. Later on, when the "extractvalue"
1367 // instructions are encountered, we use the fact that two registers were
1368 // created sequentially to get the correct registers for the "sum" and the
1370 const Function *Callee = I.getCalledFunction();
1372 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1375 if (!isTypeLegal(RetTy, VT))
1378 const Value *Op1 = I.getArgOperand(0);
1379 const Value *Op2 = I.getArgOperand(1);
1380 unsigned Reg1 = getRegForValue(Op1);
1381 unsigned Reg2 = getRegForValue(Op2);
1383 if (Reg1 == 0 || Reg2 == 0)
1384 // FIXME: Handle values *not* in registers.
1390 else if (VT == MVT::i64)
1395 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1396 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
1397 .addReg(Reg1).addReg(Reg2);
1398 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1400 // If the add with overflow is an intra-block value then we just want to
1401 // create temporaries for it like normal. If it is a cross-block value then
1402 // UpdateValueMap will return the cross-block register used. Since we
1403 // *really* want the value to be live in the register pair known by
1404 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1405 // the cross block case. In the non-cross-block case, we should just make
1406 // another register for the value.
1407 if (DestReg1 != ResultReg)
1408 ResultReg = DestReg1+1;
1410 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1412 unsigned Opc = X86::SETBr;
1413 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1415 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
1421 bool X86FastISel::X86SelectCall(const Instruction *I) {
1422 const CallInst *CI = cast<CallInst>(I);
1423 const Value *Callee = CI->getCalledValue();
1425 // Can't handle inline asm yet.
1426 if (isa<InlineAsm>(Callee))
1429 // Handle intrinsic calls.
1430 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1431 return X86VisitIntrinsicCall(*II);
1433 // Handle only C and fastcc calling conventions for now.
1434 ImmutableCallSite CS(CI);
1435 CallingConv::ID CC = CS.getCallingConv();
1436 if (CC != CallingConv::C &&
1437 CC != CallingConv::Fast &&
1438 CC != CallingConv::X86_FastCall)
1441 // fastcc with -tailcallopt is intended to provide a guaranteed
1442 // tail call optimization. Fastisel doesn't know how to do that.
1443 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
1446 // Let SDISel handle vararg functions.
1447 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1448 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1449 if (FTy->isVarArg())
1452 // Fast-isel doesn't know about callee-pop yet.
1453 if (Subtarget->IsCalleePop(FTy->isVarArg(), CC))
1456 // Handle *simple* calls for now.
1457 const Type *RetTy = CS.getType();
1459 if (RetTy->isVoidTy())
1460 RetVT = MVT::isVoid;
1461 else if (!isTypeLegal(RetTy, RetVT, true))
1464 // Materialize callee address in a register. FIXME: GV address can be
1465 // handled with a CALLpcrel32 instead.
1466 X86AddressMode CalleeAM;
1467 if (!X86SelectCallAddress(Callee, CalleeAM))
1469 unsigned CalleeOp = 0;
1470 const GlobalValue *GV = 0;
1471 if (CalleeAM.GV != 0) {
1473 } else if (CalleeAM.Base.Reg != 0) {
1474 CalleeOp = CalleeAM.Base.Reg;
1478 // Allow calls which produce i1 results.
1479 bool AndToI1 = false;
1480 if (RetVT == MVT::i1) {
1485 // Deal with call operands first.
1486 SmallVector<const Value *, 8> ArgVals;
1487 SmallVector<unsigned, 8> Args;
1488 SmallVector<MVT, 8> ArgVTs;
1489 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1490 Args.reserve(CS.arg_size());
1491 ArgVals.reserve(CS.arg_size());
1492 ArgVTs.reserve(CS.arg_size());
1493 ArgFlags.reserve(CS.arg_size());
1494 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1496 unsigned Arg = getRegForValue(*i);
1499 ISD::ArgFlagsTy Flags;
1500 unsigned AttrInd = i - CS.arg_begin() + 1;
1501 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1503 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1506 // FIXME: Only handle *easy* calls for now.
1507 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1508 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1509 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1510 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1513 const Type *ArgTy = (*i)->getType();
1515 if (!isTypeLegal(ArgTy, ArgVT))
1517 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1518 Flags.setOrigAlign(OriginalAlignment);
1520 Args.push_back(Arg);
1521 ArgVals.push_back(*i);
1522 ArgVTs.push_back(ArgVT);
1523 ArgFlags.push_back(Flags);
1526 // Analyze operands of the call, assigning locations to each operand.
1527 SmallVector<CCValAssign, 16> ArgLocs;
1528 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1530 // Allocate shadow area for Win64
1531 if (Subtarget->isTargetWin64()) {
1532 CCInfo.AllocateStack(32, 8);
1535 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86);
1537 // Get a count of how many bytes are to be pushed on the stack.
1538 unsigned NumBytes = CCInfo.getNextStackOffset();
1540 // Issue CALLSEQ_START
1541 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1542 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
1545 // Process argument: walk the register/memloc assignments, inserting
1547 SmallVector<unsigned, 4> RegArgs;
1548 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1549 CCValAssign &VA = ArgLocs[i];
1550 unsigned Arg = Args[VA.getValNo()];
1551 EVT ArgVT = ArgVTs[VA.getValNo()];
1553 // Promote the value if needed.
1554 switch (VA.getLocInfo()) {
1555 default: llvm_unreachable("Unknown loc info!");
1556 case CCValAssign::Full: break;
1557 case CCValAssign::SExt: {
1558 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1560 assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
1561 ArgVT = VA.getLocVT();
1564 case CCValAssign::ZExt: {
1565 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1567 assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
1568 ArgVT = VA.getLocVT();
1571 case CCValAssign::AExt: {
1572 // We don't handle MMX parameters yet.
1573 if (VA.getLocVT().isVector() && VA.getLocVT().getSizeInBits() == 128)
1575 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1578 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1581 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1584 assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
1585 ArgVT = VA.getLocVT();
1588 case CCValAssign::BCvt: {
1589 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT(),
1590 ISD::BITCAST, Arg, /*TODO: Kill=*/false);
1591 assert(BC != 0 && "Failed to emit a bitcast!");
1593 ArgVT = VA.getLocVT();
1598 if (VA.isRegLoc()) {
1599 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1600 VA.getLocReg()).addReg(Arg);
1601 RegArgs.push_back(VA.getLocReg());
1603 unsigned LocMemOffset = VA.getLocMemOffset();
1605 AM.Base.Reg = StackPtr;
1606 AM.Disp = LocMemOffset;
1607 const Value *ArgVal = ArgVals[VA.getValNo()];
1609 // If this is a really simple value, emit this with the Value* version of
1610 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1611 // can cause us to reevaluate the argument.
1612 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1613 X86FastEmitStore(ArgVT, ArgVal, AM);
1615 X86FastEmitStore(ArgVT, Arg, AM);
1619 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1621 if (Subtarget->isPICStyleGOT()) {
1622 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1623 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1624 X86::EBX).addReg(Base);
1628 MachineInstrBuilder MIB;
1630 // Register-indirect call.
1632 if (Subtarget->isTargetWin64())
1633 CallOpc = X86::WINCALL64r;
1634 else if (Subtarget->is64Bit())
1635 CallOpc = X86::CALL64r;
1637 CallOpc = X86::CALL32r;
1638 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1643 assert(GV && "Not a direct call");
1645 if (Subtarget->isTargetWin64())
1646 CallOpc = X86::WINCALL64pcrel32;
1647 else if (Subtarget->is64Bit())
1648 CallOpc = X86::CALL64pcrel32;
1650 CallOpc = X86::CALLpcrel32;
1652 // See if we need any target-specific flags on the GV operand.
1653 unsigned char OpFlags = 0;
1655 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1656 // external symbols most go through the PLT in PIC mode. If the symbol
1657 // has hidden or protected visibility, or if it is static or local, then
1658 // we don't need to use the PLT - we can directly call it.
1659 if (Subtarget->isTargetELF() &&
1660 TM.getRelocationModel() == Reloc::PIC_ &&
1661 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1662 OpFlags = X86II::MO_PLT;
1663 } else if (Subtarget->isPICStyleStubAny() &&
1664 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1665 Subtarget->getDarwinVers() < 9) {
1666 // PC-relative references to external symbols should go through $stub,
1667 // unless we're building with the leopard linker or later, which
1668 // automatically synthesizes these stubs.
1669 OpFlags = X86II::MO_DARWIN_STUB;
1673 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1674 .addGlobalAddress(GV, 0, OpFlags);
1677 // Add an implicit use GOT pointer in EBX.
1678 if (Subtarget->isPICStyleGOT())
1679 MIB.addReg(X86::EBX);
1681 // Add implicit physical register uses to the call.
1682 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1683 MIB.addReg(RegArgs[i]);
1685 // Issue CALLSEQ_END
1686 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1687 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
1688 .addImm(NumBytes).addImm(0);
1690 // Now handle call return value (if any).
1691 SmallVector<unsigned, 4> UsedRegs;
1692 if (RetVT != MVT::isVoid) {
1693 SmallVector<CCValAssign, 16> RVLocs;
1694 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1695 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1697 // Copy all of the result registers out of their specified physreg.
1698 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1699 EVT CopyVT = RVLocs[0].getValVT();
1700 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1702 // If this is a call to a function that returns an fp value on the x87 fp
1703 // stack, but where we prefer to use the value in xmm registers, copy it
1704 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1705 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1706 RVLocs[0].getLocReg() == X86::ST1) &&
1707 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1709 DstRC = X86::RFP80RegisterClass;
1712 unsigned ResultReg = createResultReg(DstRC);
1713 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1714 ResultReg).addReg(RVLocs[0].getLocReg());
1715 UsedRegs.push_back(RVLocs[0].getLocReg());
1717 if (CopyVT != RVLocs[0].getValVT()) {
1718 // Round the F80 the right size, which also moves to the appropriate xmm
1719 // register. This is accomplished by storing the F80 value in memory and
1720 // then loading it back. Ewww...
1721 EVT ResVT = RVLocs[0].getValVT();
1722 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1723 unsigned MemSize = ResVT.getSizeInBits()/8;
1724 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
1725 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1728 DstRC = ResVT == MVT::f32
1729 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1730 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1731 ResultReg = createResultReg(DstRC);
1732 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1733 TII.get(Opc), ResultReg), FI);
1737 // Mask out all but lowest bit for some call which produces an i1.
1738 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1739 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1740 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1741 ResultReg = AndResult;
1744 UpdateValueMap(I, ResultReg);
1747 // Set all unused physreg defs as dead.
1748 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1755 X86FastISel::TargetSelectInstruction(const Instruction *I) {
1756 switch (I->getOpcode()) {
1758 case Instruction::Load:
1759 return X86SelectLoad(I);
1760 case Instruction::Store:
1761 return X86SelectStore(I);
1762 case Instruction::Ret:
1763 return X86SelectRet(I);
1764 case Instruction::ICmp:
1765 case Instruction::FCmp:
1766 return X86SelectCmp(I);
1767 case Instruction::ZExt:
1768 return X86SelectZExt(I);
1769 case Instruction::Br:
1770 return X86SelectBranch(I);
1771 case Instruction::Call:
1772 return X86SelectCall(I);
1773 case Instruction::LShr:
1774 case Instruction::AShr:
1775 case Instruction::Shl:
1776 return X86SelectShift(I);
1777 case Instruction::Select:
1778 return X86SelectSelect(I);
1779 case Instruction::Trunc:
1780 return X86SelectTrunc(I);
1781 case Instruction::FPExt:
1782 return X86SelectFPExt(I);
1783 case Instruction::FPTrunc:
1784 return X86SelectFPTrunc(I);
1785 case Instruction::ExtractValue:
1786 return X86SelectExtractValue(I);
1787 case Instruction::IntToPtr: // Deliberate fall-through.
1788 case Instruction::PtrToInt: {
1789 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1790 EVT DstVT = TLI.getValueType(I->getType());
1791 if (DstVT.bitsGT(SrcVT))
1792 return X86SelectZExt(I);
1793 if (DstVT.bitsLT(SrcVT))
1794 return X86SelectTrunc(I);
1795 unsigned Reg = getRegForValue(I->getOperand(0));
1796 if (Reg == 0) return false;
1797 UpdateValueMap(I, Reg);
1805 unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
1807 if (!isTypeLegal(C->getType(), VT))
1810 // Get opcode and regclass of the output for the given load instruction.
1812 const TargetRegisterClass *RC = NULL;
1813 switch (VT.SimpleTy) {
1814 default: return false;
1817 RC = X86::GR8RegisterClass;
1821 RC = X86::GR16RegisterClass;
1825 RC = X86::GR32RegisterClass;
1828 // Must be in x86-64 mode.
1830 RC = X86::GR64RegisterClass;
1833 if (Subtarget->hasSSE1()) {
1835 RC = X86::FR32RegisterClass;
1837 Opc = X86::LD_Fp32m;
1838 RC = X86::RFP32RegisterClass;
1842 if (Subtarget->hasSSE2()) {
1844 RC = X86::FR64RegisterClass;
1846 Opc = X86::LD_Fp64m;
1847 RC = X86::RFP64RegisterClass;
1851 // No f80 support yet.
1855 // Materialize addresses with LEA instructions.
1856 if (isa<GlobalValue>(C)) {
1858 if (X86SelectAddress(C, AM)) {
1859 if (TLI.getPointerTy() == MVT::i32)
1863 unsigned ResultReg = createResultReg(RC);
1864 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1865 TII.get(Opc), ResultReg), AM);
1871 // MachineConstantPool wants an explicit alignment.
1872 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1874 // Alignment of vector types. FIXME!
1875 Align = TD.getTypeAllocSize(C->getType());
1878 // x86-32 PIC requires a PIC base register for constant pools.
1879 unsigned PICBase = 0;
1880 unsigned char OpFlag = 0;
1881 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1882 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1883 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1884 } else if (Subtarget->isPICStyleGOT()) {
1885 OpFlag = X86II::MO_GOTOFF;
1886 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1887 } else if (Subtarget->isPICStyleRIPRel() &&
1888 TM.getCodeModel() == CodeModel::Small) {
1892 // Create the load from the constant pool.
1893 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1894 unsigned ResultReg = createResultReg(RC);
1895 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1896 TII.get(Opc), ResultReg),
1897 MCPOffset, PICBase, OpFlag);
1902 unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
1903 // Fail on dynamic allocas. At this point, getRegForValue has already
1904 // checked its CSE maps, so if we're here trying to handle a dynamic
1905 // alloca, we're not going to succeed. X86SelectAddress has a
1906 // check for dynamic allocas, because it's called directly from
1907 // various places, but TargetMaterializeAlloca also needs a check
1908 // in order to avoid recursion between getRegForValue,
1909 // X86SelectAddrss, and TargetMaterializeAlloca.
1910 if (!FuncInfo.StaticAllocaMap.count(C))
1914 if (!X86SelectAddress(C, AM))
1916 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1917 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1918 unsigned ResultReg = createResultReg(RC);
1919 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1920 TII.get(Opc), ResultReg), AM);
1924 /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
1925 /// vreg is being provided by the specified load instruction. If possible,
1926 /// try to fold the load as an operand to the instruction, returning true if
1928 bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
1929 const LoadInst *LI) {
1931 if (!X86SelectAddress(LI->getOperand(0), AM))
1934 X86InstrInfo &XII = (X86InstrInfo&)TII;
1936 unsigned Size = TD.getTypeAllocSize(LI->getType());
1937 unsigned Alignment = LI->getAlignment();
1939 SmallVector<MachineOperand, 8> AddrOps;
1940 AM.getFullAddress(AddrOps);
1942 MachineInstr *Result =
1943 XII.foldMemoryOperandImpl(*FuncInfo.MF, MI, OpNo, AddrOps, Size, Alignment);
1944 if (Result == 0) return false;
1946 FuncInfo.MBB->insert(FuncInfo.InsertPt, Result);
1947 MI->eraseFromParent();
1953 llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
1954 return new X86FastISel(funcInfo);