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 #include "X86GenFastISel.inc"
69 bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT);
71 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
73 bool X86FastEmitStore(EVT VT, const Value *Val,
74 const X86AddressMode &AM);
75 bool X86FastEmitStore(EVT VT, unsigned Val,
76 const X86AddressMode &AM);
78 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
81 bool X86SelectAddress(const Value *V, X86AddressMode &AM);
82 bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);
84 bool X86SelectLoad(const Instruction *I);
86 bool X86SelectStore(const Instruction *I);
88 bool X86SelectRet(const Instruction *I);
90 bool X86SelectCmp(const Instruction *I);
92 bool X86SelectZExt(const Instruction *I);
94 bool X86SelectBranch(const Instruction *I);
96 bool X86SelectShift(const Instruction *I);
98 bool X86SelectSelect(const Instruction *I);
100 bool X86SelectTrunc(const Instruction *I);
102 bool X86SelectFPExt(const Instruction *I);
103 bool X86SelectFPTrunc(const Instruction *I);
105 bool X86SelectExtractValue(const Instruction *I);
107 bool X86VisitIntrinsicCall(const IntrinsicInst &I);
108 bool X86SelectCall(const Instruction *I);
110 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isTailCall = false);
111 CCAssignFn *CCAssignFnForRet(CallingConv::ID CC, bool isTailCall = false);
113 const X86InstrInfo *getInstrInfo() const {
114 return getTargetMachine()->getInstrInfo();
116 const X86TargetMachine *getTargetMachine() const {
117 return static_cast<const X86TargetMachine *>(&TM);
120 unsigned TargetMaterializeConstant(const Constant *C);
122 unsigned TargetMaterializeAlloca(const AllocaInst *C);
124 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
125 /// computed in an SSE register, not on the X87 floating point stack.
126 bool isScalarFPTypeInSSEReg(EVT VT) const {
127 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
128 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
131 bool isTypeLegal(const Type *Ty, EVT &VT, bool AllowI1 = false);
134 } // end anonymous namespace.
136 bool X86FastISel::isTypeLegal(const Type *Ty, EVT &VT, bool AllowI1) {
137 VT = TLI.getValueType(Ty, /*HandleUnknown=*/true);
138 if (VT == MVT::Other || !VT.isSimple())
139 // Unhandled type. Halt "fast" selection and bail.
142 // For now, require SSE/SSE2 for performing floating-point operations,
143 // since x87 requires additional work.
144 if (VT == MVT::f64 && !X86ScalarSSEf64)
146 if (VT == MVT::f32 && !X86ScalarSSEf32)
148 // Similarly, no f80 support yet.
151 // We only handle legal types. For example, on x86-32 the instruction
152 // selector contains all of the 64-bit instructions from x86-64,
153 // under the assumption that i64 won't be used if the target doesn't
155 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
158 #include "X86GenCallingConv.inc"
160 /// CCAssignFnForCall - Selects the correct CCAssignFn for a given calling
162 CCAssignFn *X86FastISel::CCAssignFnForCall(CallingConv::ID CC,
164 if (Subtarget->is64Bit()) {
165 if (CC == CallingConv::GHC)
166 return CC_X86_64_GHC;
167 else if (Subtarget->isTargetWin64())
168 return CC_X86_Win64_C;
173 if (CC == CallingConv::X86_FastCall)
174 return CC_X86_32_FastCall;
175 else if (CC == CallingConv::X86_ThisCall)
176 return CC_X86_32_ThisCall;
177 else if (CC == CallingConv::Fast)
178 return CC_X86_32_FastCC;
179 else if (CC == CallingConv::GHC)
180 return CC_X86_32_GHC;
185 /// CCAssignFnForRet - Selects the correct CCAssignFn for a given calling
187 CCAssignFn *X86FastISel::CCAssignFnForRet(CallingConv::ID CC,
189 if (Subtarget->is64Bit()) {
190 if (Subtarget->isTargetWin64())
191 return RetCC_X86_Win64_C;
193 return RetCC_X86_64_C;
196 return RetCC_X86_32_C;
199 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
200 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
201 /// Return true and the result register by reference if it is possible.
202 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
203 unsigned &ResultReg) {
204 // Get opcode and regclass of the output for the given load instruction.
206 const TargetRegisterClass *RC = NULL;
207 switch (VT.getSimpleVT().SimpleTy) {
208 default: return false;
212 RC = X86::GR8RegisterClass;
216 RC = X86::GR16RegisterClass;
220 RC = X86::GR32RegisterClass;
223 // Must be in x86-64 mode.
225 RC = X86::GR64RegisterClass;
228 if (Subtarget->hasSSE1()) {
230 RC = X86::FR32RegisterClass;
233 RC = X86::RFP32RegisterClass;
237 if (Subtarget->hasSSE2()) {
239 RC = X86::FR64RegisterClass;
242 RC = X86::RFP64RegisterClass;
246 // No f80 support yet.
250 ResultReg = createResultReg(RC);
251 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
252 DL, TII.get(Opc), ResultReg), AM);
256 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
257 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
258 /// and a displacement offset, or a GlobalAddress,
259 /// i.e. V. Return true if it is possible.
261 X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
262 const X86AddressMode &AM) {
263 // Get opcode and regclass of the output for the given store instruction.
265 switch (VT.getSimpleVT().SimpleTy) {
266 case MVT::f80: // No f80 support yet.
267 default: return false;
269 // Mask out all but lowest bit.
270 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
271 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
272 TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
275 // FALLTHROUGH, handling i1 as i8.
276 case MVT::i8: Opc = X86::MOV8mr; break;
277 case MVT::i16: Opc = X86::MOV16mr; break;
278 case MVT::i32: Opc = X86::MOV32mr; break;
279 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
281 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
284 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
288 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
289 DL, TII.get(Opc)), AM).addReg(Val);
293 bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
294 const X86AddressMode &AM) {
295 // Handle 'null' like i32/i64 0.
296 if (isa<ConstantPointerNull>(Val))
297 Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext()));
299 // If this is a store of a simple constant, fold the constant into the store.
300 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
303 switch (VT.getSimpleVT().SimpleTy) {
305 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
306 case MVT::i8: Opc = X86::MOV8mi; break;
307 case MVT::i16: Opc = X86::MOV16mi; break;
308 case MVT::i32: Opc = X86::MOV32mi; break;
310 // Must be a 32-bit sign extended value.
311 if ((int)CI->getSExtValue() == CI->getSExtValue())
312 Opc = X86::MOV64mi32;
317 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
318 DL, TII.get(Opc)), AM)
319 .addImm(Signed ? (uint64_t) CI->getSExtValue() :
325 unsigned ValReg = getRegForValue(Val);
329 return X86FastEmitStore(VT, ValReg, AM);
332 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
333 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
334 /// ISD::SIGN_EXTEND).
335 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
336 unsigned Src, EVT SrcVT,
337 unsigned &ResultReg) {
338 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
339 Src, /*TODO: Kill=*/false);
348 /// X86SelectAddress - Attempt to fill in an address from the given value.
350 bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
351 const User *U = NULL;
352 unsigned Opcode = Instruction::UserOp1;
353 if (const Instruction *I = dyn_cast<Instruction>(V)) {
354 // Don't walk into other basic blocks; it's possible we haven't
355 // visited them yet, so the instructions may not yet be assigned
356 // virtual registers.
357 if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
360 Opcode = I->getOpcode();
362 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
363 Opcode = C->getOpcode();
367 if (const PointerType *Ty = dyn_cast<PointerType>(V->getType()))
368 if (Ty->getAddressSpace() > 255)
369 // Fast instruction selection doesn't support the special
375 case Instruction::BitCast:
376 // Look past bitcasts.
377 return X86SelectAddress(U->getOperand(0), AM);
379 case Instruction::IntToPtr:
380 // Look past no-op inttoptrs.
381 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
382 return X86SelectAddress(U->getOperand(0), AM);
385 case Instruction::PtrToInt:
386 // Look past no-op ptrtoints.
387 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
388 return X86SelectAddress(U->getOperand(0), AM);
391 case Instruction::Alloca: {
392 // Do static allocas.
393 const AllocaInst *A = cast<AllocaInst>(V);
394 DenseMap<const AllocaInst*, int>::iterator SI =
395 FuncInfo.StaticAllocaMap.find(A);
396 if (SI != FuncInfo.StaticAllocaMap.end()) {
397 AM.BaseType = X86AddressMode::FrameIndexBase;
398 AM.Base.FrameIndex = SI->second;
404 case Instruction::Add: {
405 // Adds of constants are common and easy enough.
406 if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
407 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
408 // They have to fit in the 32-bit signed displacement field though.
409 if (isInt<32>(Disp)) {
410 AM.Disp = (uint32_t)Disp;
411 return X86SelectAddress(U->getOperand(0), AM);
417 case Instruction::GetElementPtr: {
418 X86AddressMode SavedAM = AM;
420 // Pattern-match simple GEPs.
421 uint64_t Disp = (int32_t)AM.Disp;
422 unsigned IndexReg = AM.IndexReg;
423 unsigned Scale = AM.Scale;
424 gep_type_iterator GTI = gep_type_begin(U);
425 // Iterate through the indices, folding what we can. Constants can be
426 // folded, and one dynamic index can be handled, if the scale is supported.
427 for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
428 i != e; ++i, ++GTI) {
429 const Value *Op = *i;
430 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
431 const StructLayout *SL = TD.getStructLayout(STy);
432 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
433 Disp += SL->getElementOffset(Idx);
435 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
436 SmallVector<const Value *, 4> Worklist;
437 Worklist.push_back(Op);
439 Op = Worklist.pop_back_val();
440 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
441 // Constant-offset addressing.
442 Disp += CI->getSExtValue() * S;
443 } else if (isa<AddOperator>(Op) &&
444 isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
445 // An add with a constant operand. Fold the constant.
447 cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
448 Disp += CI->getSExtValue() * S;
449 // Add the other operand back to the work list.
450 Worklist.push_back(cast<AddOperator>(Op)->getOperand(0));
451 } else if (IndexReg == 0 &&
452 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
453 (S == 1 || S == 2 || S == 4 || S == 8)) {
454 // Scaled-index addressing.
456 IndexReg = getRegForGEPIndex(Op).first;
461 goto unsupported_gep;
462 } while (!Worklist.empty());
465 // Check for displacement overflow.
466 if (!isInt<32>(Disp))
468 // Ok, the GEP indices were covered by constant-offset and scaled-index
469 // addressing. Update the address state and move on to examining the base.
470 AM.IndexReg = IndexReg;
472 AM.Disp = (uint32_t)Disp;
473 if (X86SelectAddress(U->getOperand(0), AM))
476 // If we couldn't merge the sub value into this addr mode, revert back to
477 // our address and just match the value instead of completely failing.
481 // Ok, the GEP indices weren't all covered.
486 // Handle constant address.
487 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
488 // Can't handle alternate code models yet.
489 if (TM.getCodeModel() != CodeModel::Small)
492 // RIP-relative addresses can't have additional register operands.
493 if (Subtarget->isPICStyleRIPRel() &&
494 (AM.Base.Reg != 0 || AM.IndexReg != 0))
497 // Can't handle TLS yet.
498 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
499 if (GVar->isThreadLocal())
502 // Okay, we've committed to selecting this global. Set up the basic address.
505 // Allow the subtarget to classify the global.
506 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
508 // If this reference is relative to the pic base, set it now.
509 if (isGlobalRelativeToPICBase(GVFlags)) {
510 // FIXME: How do we know Base.Reg is free??
511 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
514 // Unless the ABI requires an extra load, return a direct reference to
516 if (!isGlobalStubReference(GVFlags)) {
517 if (Subtarget->isPICStyleRIPRel()) {
518 // Use rip-relative addressing if we can. Above we verified that the
519 // base and index registers are unused.
520 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
521 AM.Base.Reg = X86::RIP;
523 AM.GVOpFlags = GVFlags;
527 // Ok, we need to do a load from a stub. If we've already loaded from this
528 // stub, reuse the loaded pointer, otherwise emit the load now.
529 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
531 if (I != LocalValueMap.end() && I->second != 0) {
534 // Issue load from stub.
536 const TargetRegisterClass *RC = NULL;
537 X86AddressMode StubAM;
538 StubAM.Base.Reg = AM.Base.Reg;
540 StubAM.GVOpFlags = GVFlags;
542 if (TLI.getPointerTy() == MVT::i64) {
544 RC = X86::GR64RegisterClass;
546 if (Subtarget->isPICStyleRIPRel())
547 StubAM.Base.Reg = X86::RIP;
550 RC = X86::GR32RegisterClass;
553 LoadReg = createResultReg(RC);
554 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
555 DL, TII.get(Opc), LoadReg), StubAM);
557 // Prevent loading GV stub multiple times in same MBB.
558 LocalValueMap[V] = LoadReg;
561 // Now construct the final address. Note that the Disp, Scale,
562 // and Index values may already be set here.
563 AM.Base.Reg = LoadReg;
568 // If all else fails, try to materialize the value in a register.
569 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
570 if (AM.Base.Reg == 0) {
571 AM.Base.Reg = getRegForValue(V);
572 return AM.Base.Reg != 0;
574 if (AM.IndexReg == 0) {
575 assert(AM.Scale == 1 && "Scale with no index!");
576 AM.IndexReg = getRegForValue(V);
577 return AM.IndexReg != 0;
584 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
586 bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
587 const User *U = NULL;
588 unsigned Opcode = Instruction::UserOp1;
589 if (const Instruction *I = dyn_cast<Instruction>(V)) {
590 Opcode = I->getOpcode();
592 } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
593 Opcode = C->getOpcode();
599 case Instruction::BitCast:
600 // Look past bitcasts.
601 return X86SelectCallAddress(U->getOperand(0), AM);
603 case Instruction::IntToPtr:
604 // Look past no-op inttoptrs.
605 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
606 return X86SelectCallAddress(U->getOperand(0), AM);
609 case Instruction::PtrToInt:
610 // Look past no-op ptrtoints.
611 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
612 return X86SelectCallAddress(U->getOperand(0), AM);
616 // Handle constant address.
617 if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
618 // Can't handle alternate code models yet.
619 if (TM.getCodeModel() != CodeModel::Small)
622 // RIP-relative addresses can't have additional register operands.
623 if (Subtarget->isPICStyleRIPRel() &&
624 (AM.Base.Reg != 0 || AM.IndexReg != 0))
627 // Can't handle TLS or DLLImport.
628 if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
629 if (GVar->isThreadLocal() || GVar->hasDLLImportLinkage())
632 // Okay, we've committed to selecting this global. Set up the basic address.
635 // No ABI requires an extra load for anything other than DLLImport, which
636 // we rejected above. Return a direct reference to the global.
637 if (Subtarget->isPICStyleRIPRel()) {
638 // Use rip-relative addressing if we can. Above we verified that the
639 // base and index registers are unused.
640 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
641 AM.Base.Reg = X86::RIP;
642 } else if (Subtarget->isPICStyleStubPIC()) {
643 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
644 } else if (Subtarget->isPICStyleGOT()) {
645 AM.GVOpFlags = X86II::MO_GOTOFF;
651 // If all else fails, try to materialize the value in a register.
652 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
653 if (AM.Base.Reg == 0) {
654 AM.Base.Reg = getRegForValue(V);
655 return AM.Base.Reg != 0;
657 if (AM.IndexReg == 0) {
658 assert(AM.Scale == 1 && "Scale with no index!");
659 AM.IndexReg = getRegForValue(V);
660 return AM.IndexReg != 0;
668 /// X86SelectStore - Select and emit code to implement store instructions.
669 bool X86FastISel::X86SelectStore(const Instruction *I) {
671 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
675 if (!X86SelectAddress(I->getOperand(1), AM))
678 return X86FastEmitStore(VT, I->getOperand(0), AM);
681 /// X86SelectRet - Select and emit code to implement ret instructions.
682 bool X86FastISel::X86SelectRet(const Instruction *I) {
683 const ReturnInst *Ret = cast<ReturnInst>(I);
684 const Function &F = *I->getParent()->getParent();
686 if (!FuncInfo.CanLowerReturn)
689 CallingConv::ID CC = F.getCallingConv();
690 if (CC != CallingConv::C &&
691 CC != CallingConv::Fast &&
692 CC != CallingConv::X86_FastCall)
695 if (Subtarget->isTargetWin64())
698 // fastcc with -tailcallopt is intended to provide a guaranteed
699 // tail call optimization. Fastisel doesn't know how to do that.
700 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
703 // Let SDISel handle vararg functions.
707 if (Ret->getNumOperands() > 0) {
708 SmallVector<ISD::OutputArg, 4> Outs;
709 GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
712 // Analyze operands of the call, assigning locations to each operand.
713 SmallVector<CCValAssign, 16> ValLocs;
714 CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
715 CCInfo.AnalyzeReturn(Outs, CCAssignFnForRet(CC));
717 const Value *RV = Ret->getOperand(0);
718 unsigned Reg = getRegForValue(RV);
722 // Copy the return value into registers.
723 for (unsigned i = 0, e = ValLocs.size(); i != e; ++i) {
724 CCValAssign &VA = ValLocs[i];
726 // Don't bother handling odd stuff for now.
727 if (VA.getLocInfo() != CCValAssign::Full)
732 TargetRegisterClass* RC = TLI.getRegClassFor(VA.getValVT());
733 bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
734 VA.getLocReg(), Reg + VA.getValNo(),
736 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
738 MRI.addLiveOut(VA.getLocReg());
743 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET));
747 /// X86SelectLoad - Select and emit code to implement load instructions.
749 bool X86FastISel::X86SelectLoad(const Instruction *I) {
751 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
755 if (!X86SelectAddress(I->getOperand(0), AM))
758 unsigned ResultReg = 0;
759 if (X86FastEmitLoad(VT, AM, ResultReg)) {
760 UpdateValueMap(I, ResultReg);
766 static unsigned X86ChooseCmpOpcode(EVT VT) {
767 switch (VT.getSimpleVT().SimpleTy) {
769 case MVT::i8: return X86::CMP8rr;
770 case MVT::i16: return X86::CMP16rr;
771 case MVT::i32: return X86::CMP32rr;
772 case MVT::i64: return X86::CMP64rr;
773 case MVT::f32: return X86::UCOMISSrr;
774 case MVT::f64: return X86::UCOMISDrr;
778 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
779 /// of the comparison, return an opcode that works for the compare (e.g.
780 /// CMP32ri) otherwise return 0.
781 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
782 switch (VT.getSimpleVT().SimpleTy) {
783 // Otherwise, we can't fold the immediate into this comparison.
785 case MVT::i8: return X86::CMP8ri;
786 case MVT::i16: return X86::CMP16ri;
787 case MVT::i32: return X86::CMP32ri;
789 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
791 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
792 return X86::CMP64ri32;
797 bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
799 unsigned Op0Reg = getRegForValue(Op0);
800 if (Op0Reg == 0) return false;
802 // Handle 'null' like i32/i64 0.
803 if (isa<ConstantPointerNull>(Op1))
804 Op1 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
806 // We have two options: compare with register or immediate. If the RHS of
807 // the compare is an immediate that we can fold into this compare, use
808 // CMPri, otherwise use CMPrr.
809 if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
810 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
811 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareImmOpc))
813 .addImm(Op1C->getSExtValue());
818 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
819 if (CompareOpc == 0) return false;
821 unsigned Op1Reg = getRegForValue(Op1);
822 if (Op1Reg == 0) return false;
823 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareOpc))
830 bool X86FastISel::X86SelectCmp(const Instruction *I) {
831 const CmpInst *CI = cast<CmpInst>(I);
834 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
837 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
839 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
840 switch (CI->getPredicate()) {
841 case CmpInst::FCMP_OEQ: {
842 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
845 unsigned EReg = createResultReg(&X86::GR8RegClass);
846 unsigned NPReg = createResultReg(&X86::GR8RegClass);
847 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETEr), EReg);
848 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
849 TII.get(X86::SETNPr), NPReg);
850 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
851 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
852 UpdateValueMap(I, ResultReg);
855 case CmpInst::FCMP_UNE: {
856 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
859 unsigned NEReg = createResultReg(&X86::GR8RegClass);
860 unsigned PReg = createResultReg(&X86::GR8RegClass);
861 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
862 TII.get(X86::SETNEr), NEReg);
863 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
864 TII.get(X86::SETPr), PReg);
865 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
866 TII.get(X86::OR8rr), ResultReg)
867 .addReg(PReg).addReg(NEReg);
868 UpdateValueMap(I, ResultReg);
871 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
872 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
873 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
874 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
875 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
876 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
877 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
878 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
879 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
880 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
881 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
882 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
884 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
885 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
886 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
887 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
888 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
889 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
890 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
891 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
892 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
893 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
898 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
902 // Emit a compare of Op0/Op1.
903 if (!X86FastEmitCompare(Op0, Op1, VT))
906 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(SetCCOpc), ResultReg);
907 UpdateValueMap(I, ResultReg);
911 bool X86FastISel::X86SelectZExt(const Instruction *I) {
912 // Handle zero-extension from i1 to i8, which is common.
913 if (I->getType()->isIntegerTy(8) &&
914 I->getOperand(0)->getType()->isIntegerTy(1)) {
915 unsigned ResultReg = getRegForValue(I->getOperand(0));
916 if (ResultReg == 0) return false;
917 // Set the high bits to zero.
918 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
919 if (ResultReg == 0) return false;
920 UpdateValueMap(I, ResultReg);
928 bool X86FastISel::X86SelectBranch(const Instruction *I) {
929 // Unconditional branches are selected by tablegen-generated code.
930 // Handle a conditional branch.
931 const BranchInst *BI = cast<BranchInst>(I);
932 MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
933 MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];
935 // Fold the common case of a conditional branch with a comparison.
936 if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
937 if (CI->hasOneUse()) {
938 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
940 // Try to take advantage of fallthrough opportunities.
941 CmpInst::Predicate Predicate = CI->getPredicate();
942 if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
943 std::swap(TrueMBB, FalseMBB);
944 Predicate = CmpInst::getInversePredicate(Predicate);
947 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
948 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
951 case CmpInst::FCMP_OEQ:
952 std::swap(TrueMBB, FalseMBB);
953 Predicate = CmpInst::FCMP_UNE;
955 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
956 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
957 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
958 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
959 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
960 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
961 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
962 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
963 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
964 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
965 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
966 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
967 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
969 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
970 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
971 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
972 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
973 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
974 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
975 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
976 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
977 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
978 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
983 const Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
987 // Emit a compare of the LHS and RHS, setting the flags.
988 if (!X86FastEmitCompare(Op0, Op1, VT))
991 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BranchOpc))
994 if (Predicate == CmpInst::FCMP_UNE) {
995 // X86 requires a second branch to handle UNE (and OEQ,
996 // which is mapped to UNE above).
997 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JP_4))
1001 FastEmitBranch(FalseMBB, DL);
1002 FuncInfo.MBB->addSuccessor(TrueMBB);
1005 } else if (ExtractValueInst *EI =
1006 dyn_cast<ExtractValueInst>(BI->getCondition())) {
1007 // Check to see if the branch instruction is from an "arithmetic with
1008 // overflow" intrinsic. The main way these intrinsics are used is:
1010 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
1011 // %sum = extractvalue { i32, i1 } %t, 0
1012 // %obit = extractvalue { i32, i1 } %t, 1
1013 // br i1 %obit, label %overflow, label %normal
1015 // The %sum and %obit are converted in an ADD and a SETO/SETB before
1016 // reaching the branch. Therefore, we search backwards through the MBB
1017 // looking for the SETO/SETB instruction. If an instruction modifies the
1018 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
1019 // convert the branch into a JO/JB instruction.
1020 if (const IntrinsicInst *CI =
1021 dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
1022 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
1023 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
1024 const MachineInstr *SetMI = 0;
1025 unsigned Reg = getRegForValue(EI);
1027 for (MachineBasicBlock::const_reverse_iterator
1028 RI = FuncInfo.MBB->rbegin(), RE = FuncInfo.MBB->rend();
1030 const MachineInstr &MI = *RI;
1032 if (MI.definesRegister(Reg)) {
1033 unsigned Src, Dst, SrcSR, DstSR;
1035 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
1044 const TargetInstrDesc &TID = MI.getDesc();
1045 if (TID.hasUnmodeledSideEffects() ||
1046 TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
1051 unsigned OpCode = SetMI->getOpcode();
1053 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
1054 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1055 TII.get(OpCode == X86::SETOr ? X86::JO_4 : X86::JB_4))
1057 FastEmitBranch(FalseMBB, DL);
1058 FuncInfo.MBB->addSuccessor(TrueMBB);
1066 // Otherwise do a clumsy setcc and re-test it.
1067 unsigned OpReg = getRegForValue(BI->getCondition());
1068 if (OpReg == 0) return false;
1070 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1071 .addReg(OpReg).addReg(OpReg);
1072 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JNE_4))
1074 FastEmitBranch(FalseMBB, DL);
1075 FuncInfo.MBB->addSuccessor(TrueMBB);
1079 bool X86FastISel::X86SelectShift(const Instruction *I) {
1080 unsigned CReg = 0, OpReg = 0, OpImm = 0;
1081 const TargetRegisterClass *RC = NULL;
1082 if (I->getType()->isIntegerTy(8)) {
1084 RC = &X86::GR8RegClass;
1085 switch (I->getOpcode()) {
1086 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
1087 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
1088 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
1089 default: return false;
1091 } else if (I->getType()->isIntegerTy(16)) {
1093 RC = &X86::GR16RegClass;
1094 switch (I->getOpcode()) {
1095 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
1096 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
1097 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
1098 default: return false;
1100 } else if (I->getType()->isIntegerTy(32)) {
1102 RC = &X86::GR32RegClass;
1103 switch (I->getOpcode()) {
1104 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
1105 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
1106 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
1107 default: return false;
1109 } else if (I->getType()->isIntegerTy(64)) {
1111 RC = &X86::GR64RegClass;
1112 switch (I->getOpcode()) {
1113 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
1114 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
1115 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
1116 default: return false;
1122 EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1123 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1126 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1127 if (Op0Reg == 0) return false;
1129 // Fold immediate in shl(x,3).
1130 if (const ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
1131 unsigned ResultReg = createResultReg(RC);
1132 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpImm),
1133 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
1134 UpdateValueMap(I, ResultReg);
1138 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1139 if (Op1Reg == 0) return false;
1140 TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1141 CReg, Op1Reg, RC, RC, DL);
1143 // The shift instruction uses X86::CL. If we defined a super-register
1144 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
1145 // we're doing here.
1146 if (CReg != X86::CL)
1147 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1148 TII.get(TargetOpcode::EXTRACT_SUBREG), X86::CL)
1149 .addReg(CReg).addImm(X86::sub_8bit);
1151 unsigned ResultReg = createResultReg(RC);
1152 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpReg), ResultReg)
1154 UpdateValueMap(I, ResultReg);
1158 bool X86FastISel::X86SelectSelect(const Instruction *I) {
1159 EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1160 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1164 const TargetRegisterClass *RC = NULL;
1165 if (VT.getSimpleVT() == MVT::i16) {
1166 Opc = X86::CMOVE16rr;
1167 RC = &X86::GR16RegClass;
1168 } else if (VT.getSimpleVT() == MVT::i32) {
1169 Opc = X86::CMOVE32rr;
1170 RC = &X86::GR32RegClass;
1171 } else if (VT.getSimpleVT() == MVT::i64) {
1172 Opc = X86::CMOVE64rr;
1173 RC = &X86::GR64RegClass;
1178 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1179 if (Op0Reg == 0) return false;
1180 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1181 if (Op1Reg == 0) return false;
1182 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1183 if (Op2Reg == 0) return false;
1185 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
1186 .addReg(Op0Reg).addReg(Op0Reg);
1187 unsigned ResultReg = createResultReg(RC);
1188 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
1189 .addReg(Op1Reg).addReg(Op2Reg);
1190 UpdateValueMap(I, ResultReg);
1194 bool X86FastISel::X86SelectFPExt(const Instruction *I) {
1195 // fpext from float to double.
1196 if (Subtarget->hasSSE2() &&
1197 I->getType()->isDoubleTy()) {
1198 const Value *V = I->getOperand(0);
1199 if (V->getType()->isFloatTy()) {
1200 unsigned OpReg = getRegForValue(V);
1201 if (OpReg == 0) return false;
1202 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1203 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1204 TII.get(X86::CVTSS2SDrr), ResultReg)
1206 UpdateValueMap(I, ResultReg);
1214 bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
1215 if (Subtarget->hasSSE2()) {
1216 if (I->getType()->isFloatTy()) {
1217 const Value *V = I->getOperand(0);
1218 if (V->getType()->isDoubleTy()) {
1219 unsigned OpReg = getRegForValue(V);
1220 if (OpReg == 0) return false;
1221 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1222 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1223 TII.get(X86::CVTSD2SSrr), ResultReg)
1225 UpdateValueMap(I, ResultReg);
1234 bool X86FastISel::X86SelectTrunc(const Instruction *I) {
1235 if (Subtarget->is64Bit())
1236 // All other cases should be handled by the tblgen generated code.
1238 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1239 EVT DstVT = TLI.getValueType(I->getType());
1241 // This code only handles truncation to byte right now.
1242 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1243 // All other cases should be handled by the tblgen generated code.
1245 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1246 // All other cases should be handled by the tblgen generated code.
1249 unsigned InputReg = getRegForValue(I->getOperand(0));
1251 // Unhandled operand. Halt "fast" selection and bail.
1254 // First issue a copy to GR16_ABCD or GR32_ABCD.
1255 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
1256 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1257 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1258 unsigned CopyReg = createResultReg(CopyRC);
1259 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CopyOpc), CopyReg)
1262 // Then issue an extract_subreg.
1263 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1264 CopyReg, /*Kill=*/true,
1269 UpdateValueMap(I, ResultReg);
1273 bool X86FastISel::X86SelectExtractValue(const Instruction *I) {
1274 const ExtractValueInst *EI = cast<ExtractValueInst>(I);
1275 const Value *Agg = EI->getAggregateOperand();
1277 if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1278 switch (CI->getIntrinsicID()) {
1280 case Intrinsic::sadd_with_overflow:
1281 case Intrinsic::uadd_with_overflow: {
1282 // Cheat a little. We know that the registers for "add" and "seto" are
1283 // allocated sequentially. However, we only keep track of the register
1284 // for "add" in the value map. Use extractvalue's index to get the
1285 // correct register for "seto".
1286 unsigned OpReg = getRegForValue(Agg);
1289 UpdateValueMap(I, OpReg + *EI->idx_begin());
1298 bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
1299 // FIXME: Handle more intrinsics.
1300 switch (I.getIntrinsicID()) {
1301 default: return false;
1302 case Intrinsic::stackprotector: {
1303 // Emit code inline code to store the stack guard onto the stack.
1304 EVT PtrTy = TLI.getPointerTy();
1306 const Value *Op1 = I.getArgOperand(0); // The guard's value.
1307 const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
1309 // Grab the frame index.
1311 if (!X86SelectAddress(Slot, AM)) return false;
1313 if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
1317 case Intrinsic::objectsize: {
1318 ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
1319 const Type *Ty = I.getCalledFunction()->getReturnType();
1321 assert(CI && "Non-constant type in Intrinsic::objectsize?");
1324 if (!isTypeLegal(Ty, VT))
1330 else if (VT == MVT::i64)
1335 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1336 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg).
1337 addImm(CI->isZero() ? -1ULL : 0);
1338 UpdateValueMap(&I, ResultReg);
1341 case Intrinsic::dbg_declare: {
1342 const DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1344 assert(DI->getAddress() && "Null address should be checked earlier!");
1345 if (!X86SelectAddress(DI->getAddress(), AM))
1347 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1348 // FIXME may need to add RegState::Debug to any registers produced,
1349 // although ESP/EBP should be the only ones at the moment.
1350 addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
1351 addImm(0).addMetadata(DI->getVariable());
1354 case Intrinsic::trap: {
1355 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TRAP));
1358 case Intrinsic::sadd_with_overflow:
1359 case Intrinsic::uadd_with_overflow: {
1360 // Replace "add with overflow" intrinsics with an "add" instruction followed
1361 // by a seto/setc instruction. Later on, when the "extractvalue"
1362 // instructions are encountered, we use the fact that two registers were
1363 // created sequentially to get the correct registers for the "sum" and the
1365 const Function *Callee = I.getCalledFunction();
1367 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1370 if (!isTypeLegal(RetTy, VT))
1373 const Value *Op1 = I.getArgOperand(0);
1374 const Value *Op2 = I.getArgOperand(1);
1375 unsigned Reg1 = getRegForValue(Op1);
1376 unsigned Reg2 = getRegForValue(Op2);
1378 if (Reg1 == 0 || Reg2 == 0)
1379 // FIXME: Handle values *not* in registers.
1385 else if (VT == MVT::i64)
1390 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1391 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
1392 .addReg(Reg1).addReg(Reg2);
1393 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1395 // If the add with overflow is an intra-block value then we just want to
1396 // create temporaries for it like normal. If it is a cross-block value then
1397 // UpdateValueMap will return the cross-block register used. Since we
1398 // *really* want the value to be live in the register pair known by
1399 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1400 // the cross block case. In the non-cross-block case, we should just make
1401 // another register for the value.
1402 if (DestReg1 != ResultReg)
1403 ResultReg = DestReg1+1;
1405 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1407 unsigned Opc = X86::SETBr;
1408 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1410 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
1416 bool X86FastISel::X86SelectCall(const Instruction *I) {
1417 const CallInst *CI = cast<CallInst>(I);
1418 const Value *Callee = CI->getCalledValue();
1420 // Can't handle inline asm yet.
1421 if (isa<InlineAsm>(Callee))
1424 // Handle intrinsic calls.
1425 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1426 return X86VisitIntrinsicCall(*II);
1428 // Handle only C and fastcc calling conventions for now.
1429 ImmutableCallSite CS(CI);
1430 CallingConv::ID CC = CS.getCallingConv();
1431 if (CC != CallingConv::C &&
1432 CC != CallingConv::Fast &&
1433 CC != CallingConv::X86_FastCall)
1436 // fastcc with -tailcallopt is intended to provide a guaranteed
1437 // tail call optimization. Fastisel doesn't know how to do that.
1438 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
1441 // Let SDISel handle vararg functions.
1442 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1443 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1444 if (FTy->isVarArg())
1447 // Fast-isel doesn't know about callee-pop yet.
1448 if (Subtarget->IsCalleePop(FTy->isVarArg(), CC))
1451 // Handle *simple* calls for now.
1452 const Type *RetTy = CS.getType();
1454 if (RetTy->isVoidTy())
1455 RetVT = MVT::isVoid;
1456 else if (!isTypeLegal(RetTy, RetVT, true))
1459 // Materialize callee address in a register. FIXME: GV address can be
1460 // handled with a CALLpcrel32 instead.
1461 X86AddressMode CalleeAM;
1462 if (!X86SelectCallAddress(Callee, CalleeAM))
1464 unsigned CalleeOp = 0;
1465 const GlobalValue *GV = 0;
1466 if (CalleeAM.GV != 0) {
1468 } else if (CalleeAM.Base.Reg != 0) {
1469 CalleeOp = CalleeAM.Base.Reg;
1473 // Allow calls which produce i1 results.
1474 bool AndToI1 = false;
1475 if (RetVT == MVT::i1) {
1480 // Deal with call operands first.
1481 SmallVector<const Value *, 8> ArgVals;
1482 SmallVector<unsigned, 8> Args;
1483 SmallVector<EVT, 8> ArgVTs;
1484 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1485 Args.reserve(CS.arg_size());
1486 ArgVals.reserve(CS.arg_size());
1487 ArgVTs.reserve(CS.arg_size());
1488 ArgFlags.reserve(CS.arg_size());
1489 for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1491 unsigned Arg = getRegForValue(*i);
1494 ISD::ArgFlagsTy Flags;
1495 unsigned AttrInd = i - CS.arg_begin() + 1;
1496 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1498 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1501 // FIXME: Only handle *easy* calls for now.
1502 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1503 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1504 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1505 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1508 const Type *ArgTy = (*i)->getType();
1510 if (!isTypeLegal(ArgTy, ArgVT))
1512 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1513 Flags.setOrigAlign(OriginalAlignment);
1515 Args.push_back(Arg);
1516 ArgVals.push_back(*i);
1517 ArgVTs.push_back(ArgVT);
1518 ArgFlags.push_back(Flags);
1521 // Analyze operands of the call, assigning locations to each operand.
1522 SmallVector<CCValAssign, 16> ArgLocs;
1523 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1525 // Allocate shadow area for Win64
1526 if (Subtarget->isTargetWin64()) {
1527 CCInfo.AllocateStack(32, 8);
1530 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1532 // Get a count of how many bytes are to be pushed on the stack.
1533 unsigned NumBytes = CCInfo.getNextStackOffset();
1535 // Issue CALLSEQ_START
1536 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1537 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
1540 // Process argument: walk the register/memloc assignments, inserting
1542 SmallVector<unsigned, 4> RegArgs;
1543 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1544 CCValAssign &VA = ArgLocs[i];
1545 unsigned Arg = Args[VA.getValNo()];
1546 EVT ArgVT = ArgVTs[VA.getValNo()];
1548 // Promote the value if needed.
1549 switch (VA.getLocInfo()) {
1550 default: llvm_unreachable("Unknown loc info!");
1551 case CCValAssign::Full: break;
1552 case CCValAssign::SExt: {
1553 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1555 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1557 ArgVT = VA.getLocVT();
1560 case CCValAssign::ZExt: {
1561 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1563 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1565 ArgVT = VA.getLocVT();
1568 case CCValAssign::AExt: {
1569 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1572 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1575 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1578 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1579 ArgVT = VA.getLocVT();
1582 case CCValAssign::BCvt: {
1583 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT().getSimpleVT(),
1584 ISD::BIT_CONVERT, Arg, /*TODO: Kill=*/false);
1585 assert(BC != 0 && "Failed to emit a bitcast!");
1587 ArgVT = VA.getLocVT();
1592 if (VA.isRegLoc()) {
1593 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1594 bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1595 VA.getLocReg(), Arg, RC, RC, DL);
1596 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1598 RegArgs.push_back(VA.getLocReg());
1600 unsigned LocMemOffset = VA.getLocMemOffset();
1602 AM.Base.Reg = StackPtr;
1603 AM.Disp = LocMemOffset;
1604 const Value *ArgVal = ArgVals[VA.getValNo()];
1606 // If this is a really simple value, emit this with the Value* version of
1607 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1608 // can cause us to reevaluate the argument.
1609 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1610 X86FastEmitStore(ArgVT, ArgVal, AM);
1612 X86FastEmitStore(ArgVT, Arg, AM);
1616 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1618 if (Subtarget->isPICStyleGOT()) {
1619 TargetRegisterClass *RC = X86::GR32RegisterClass;
1620 unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1621 bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1622 X86::EBX, Base, RC, RC, DL);
1623 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1628 MachineInstrBuilder MIB;
1630 // Register-indirect call.
1631 unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
1632 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1637 assert(GV && "Not a direct call");
1639 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
1641 // See if we need any target-specific flags on the GV operand.
1642 unsigned char OpFlags = 0;
1644 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1645 // external symbols most go through the PLT in PIC mode. If the symbol
1646 // has hidden or protected visibility, or if it is static or local, then
1647 // we don't need to use the PLT - we can directly call it.
1648 if (Subtarget->isTargetELF() &&
1649 TM.getRelocationModel() == Reloc::PIC_ &&
1650 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1651 OpFlags = X86II::MO_PLT;
1652 } else if (Subtarget->isPICStyleStubAny() &&
1653 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1654 Subtarget->getDarwinVers() < 9) {
1655 // PC-relative references to external symbols should go through $stub,
1656 // unless we're building with the leopard linker or later, which
1657 // automatically synthesizes these stubs.
1658 OpFlags = X86II::MO_DARWIN_STUB;
1662 MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
1663 .addGlobalAddress(GV, 0, OpFlags);
1666 // Add an implicit use GOT pointer in EBX.
1667 if (Subtarget->isPICStyleGOT())
1668 MIB.addReg(X86::EBX);
1670 // Add implicit physical register uses to the call.
1671 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1672 MIB.addReg(RegArgs[i]);
1674 // Issue CALLSEQ_END
1675 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1676 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
1677 .addImm(NumBytes).addImm(0);
1679 // Now handle call return value (if any).
1680 SmallVector<unsigned, 4> UsedRegs;
1681 if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
1682 SmallVector<CCValAssign, 16> RVLocs;
1683 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1684 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1686 // Copy all of the result registers out of their specified physreg.
1687 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1688 EVT CopyVT = RVLocs[0].getValVT();
1689 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1690 TargetRegisterClass *SrcRC = DstRC;
1692 // If this is a call to a function that returns an fp value on the x87 fp
1693 // stack, but where we prefer to use the value in xmm registers, copy it
1694 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1695 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1696 RVLocs[0].getLocReg() == X86::ST1) &&
1697 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1699 SrcRC = X86::RSTRegisterClass;
1700 DstRC = X86::RFP80RegisterClass;
1703 unsigned ResultReg = createResultReg(DstRC);
1704 bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt, ResultReg,
1705 RVLocs[0].getLocReg(), DstRC, SrcRC, DL);
1706 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1708 UsedRegs.push_back(RVLocs[0].getLocReg());
1710 if (CopyVT != RVLocs[0].getValVT()) {
1711 // Round the F80 the right size, which also moves to the appropriate xmm
1712 // register. This is accomplished by storing the F80 value in memory and
1713 // then loading it back. Ewww...
1714 EVT ResVT = RVLocs[0].getValVT();
1715 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1716 unsigned MemSize = ResVT.getSizeInBits()/8;
1717 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
1718 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1721 DstRC = ResVT == MVT::f32
1722 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1723 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1724 ResultReg = createResultReg(DstRC);
1725 addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1726 TII.get(Opc), ResultReg), FI);
1730 // Mask out all but lowest bit for some call which produces an i1.
1731 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1732 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1733 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1734 ResultReg = AndResult;
1737 UpdateValueMap(I, ResultReg);
1740 // Set all unused physreg defs as dead.
1741 static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
1748 X86FastISel::TargetSelectInstruction(const Instruction *I) {
1749 switch (I->getOpcode()) {
1751 case Instruction::Load:
1752 return X86SelectLoad(I);
1753 case Instruction::Store:
1754 return X86SelectStore(I);
1755 case Instruction::Ret:
1756 return X86SelectRet(I);
1757 case Instruction::ICmp:
1758 case Instruction::FCmp:
1759 return X86SelectCmp(I);
1760 case Instruction::ZExt:
1761 return X86SelectZExt(I);
1762 case Instruction::Br:
1763 return X86SelectBranch(I);
1764 case Instruction::Call:
1765 return X86SelectCall(I);
1766 case Instruction::LShr:
1767 case Instruction::AShr:
1768 case Instruction::Shl:
1769 return X86SelectShift(I);
1770 case Instruction::Select:
1771 return X86SelectSelect(I);
1772 case Instruction::Trunc:
1773 return X86SelectTrunc(I);
1774 case Instruction::FPExt:
1775 return X86SelectFPExt(I);
1776 case Instruction::FPTrunc:
1777 return X86SelectFPTrunc(I);
1778 case Instruction::ExtractValue:
1779 return X86SelectExtractValue(I);
1780 case Instruction::IntToPtr: // Deliberate fall-through.
1781 case Instruction::PtrToInt: {
1782 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1783 EVT DstVT = TLI.getValueType(I->getType());
1784 if (DstVT.bitsGT(SrcVT))
1785 return X86SelectZExt(I);
1786 if (DstVT.bitsLT(SrcVT))
1787 return X86SelectTrunc(I);
1788 unsigned Reg = getRegForValue(I->getOperand(0));
1789 if (Reg == 0) return false;
1790 UpdateValueMap(I, Reg);
1798 unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
1800 if (!isTypeLegal(C->getType(), VT))
1803 // Get opcode and regclass of the output for the given load instruction.
1805 const TargetRegisterClass *RC = NULL;
1806 switch (VT.getSimpleVT().SimpleTy) {
1807 default: return false;
1810 RC = X86::GR8RegisterClass;
1814 RC = X86::GR16RegisterClass;
1818 RC = X86::GR32RegisterClass;
1821 // Must be in x86-64 mode.
1823 RC = X86::GR64RegisterClass;
1826 if (Subtarget->hasSSE1()) {
1828 RC = X86::FR32RegisterClass;
1830 Opc = X86::LD_Fp32m;
1831 RC = X86::RFP32RegisterClass;
1835 if (Subtarget->hasSSE2()) {
1837 RC = X86::FR64RegisterClass;
1839 Opc = X86::LD_Fp64m;
1840 RC = X86::RFP64RegisterClass;
1844 // No f80 support yet.
1848 // Materialize addresses with LEA instructions.
1849 if (isa<GlobalValue>(C)) {
1851 if (X86SelectAddress(C, AM)) {
1852 if (TLI.getPointerTy() == MVT::i32)
1856 unsigned ResultReg = createResultReg(RC);
1857 addLeaAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1858 TII.get(Opc), ResultReg), AM);
1864 // MachineConstantPool wants an explicit alignment.
1865 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1867 // Alignment of vector types. FIXME!
1868 Align = TD.getTypeAllocSize(C->getType());
1871 // x86-32 PIC requires a PIC base register for constant pools.
1872 unsigned PICBase = 0;
1873 unsigned char OpFlag = 0;
1874 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1875 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1876 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1877 } else if (Subtarget->isPICStyleGOT()) {
1878 OpFlag = X86II::MO_GOTOFF;
1879 PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
1880 } else if (Subtarget->isPICStyleRIPRel() &&
1881 TM.getCodeModel() == CodeModel::Small) {
1885 // Create the load from the constant pool.
1886 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1887 unsigned ResultReg = createResultReg(RC);
1888 addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1889 TII.get(Opc), ResultReg),
1890 MCPOffset, PICBase, OpFlag);
1895 unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
1896 // Fail on dynamic allocas. At this point, getRegForValue has already
1897 // checked its CSE maps, so if we're here trying to handle a dynamic
1898 // alloca, we're not going to succeed. X86SelectAddress has a
1899 // check for dynamic allocas, because it's called directly from
1900 // various places, but TargetMaterializeAlloca also needs a check
1901 // in order to avoid recursion between getRegForValue,
1902 // X86SelectAddrss, and TargetMaterializeAlloca.
1903 if (!FuncInfo.StaticAllocaMap.count(C))
1907 if (!X86SelectAddress(C, AM))
1909 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1910 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1911 unsigned ResultReg = createResultReg(RC);
1912 addLeaAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
1913 TII.get(Opc), ResultReg), AM);
1918 llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
1919 return new X86FastISel(funcInfo);