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 "X86ISelLowering.h"
19 #include "X86RegisterInfo.h"
20 #include "X86Subtarget.h"
21 #include "X86TargetMachine.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/CodeGen/FastISel.h"
28 #include "llvm/CodeGen/MachineConstantPool.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineRegisterInfo.h"
31 #include "llvm/Support/CallSite.h"
32 #include "llvm/Support/GetElementPtrTypeIterator.h"
33 #include "llvm/Target/TargetOptions.h"
38 class X86FastISel : public FastISel {
39 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
40 /// make the right decision when generating code for different targets.
41 const X86Subtarget *Subtarget;
43 /// StackPtr - Register used as the stack pointer.
47 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
48 /// floating point ops.
49 /// When SSE is available, use it for f32 operations.
50 /// When SSE2 is available, use it for f64 operations.
55 explicit X86FastISel(MachineFunction &mf,
56 MachineModuleInfo *mmi,
58 DenseMap<const Value *, unsigned> &vm,
59 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
60 DenseMap<const AllocaInst *, int> &am
62 , SmallSet<Instruction*, 8> &cil
65 : FastISel(mf, mmi, dw, vm, bm, am
70 Subtarget = &TM.getSubtarget<X86Subtarget>();
71 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
72 X86ScalarSSEf64 = Subtarget->hasSSE2();
73 X86ScalarSSEf32 = Subtarget->hasSSE1();
76 virtual bool TargetSelectInstruction(Instruction *I);
78 #include "X86GenFastISel.inc"
81 bool X86FastEmitCompare(Value *LHS, Value *RHS, MVT VT);
83 bool X86FastEmitLoad(MVT VT, const X86AddressMode &AM, unsigned &RR);
85 bool X86FastEmitStore(MVT VT, Value *Val,
86 const X86AddressMode &AM);
87 bool X86FastEmitStore(MVT VT, unsigned Val,
88 const X86AddressMode &AM);
90 bool X86FastEmitExtend(ISD::NodeType Opc, MVT DstVT, unsigned Src, MVT SrcVT,
93 bool X86SelectAddress(Value *V, X86AddressMode &AM);
94 bool X86SelectCallAddress(Value *V, X86AddressMode &AM);
96 bool X86SelectLoad(Instruction *I);
98 bool X86SelectStore(Instruction *I);
100 bool X86SelectCmp(Instruction *I);
102 bool X86SelectZExt(Instruction *I);
104 bool X86SelectBranch(Instruction *I);
106 bool X86SelectShift(Instruction *I);
108 bool X86SelectSelect(Instruction *I);
110 bool X86SelectTrunc(Instruction *I);
112 bool X86SelectFPExt(Instruction *I);
113 bool X86SelectFPTrunc(Instruction *I);
115 bool X86SelectExtractValue(Instruction *I);
117 bool X86VisitIntrinsicCall(IntrinsicInst &I);
118 bool X86SelectCall(Instruction *I);
120 CCAssignFn *CCAssignFnForCall(unsigned CC, bool isTailCall = false);
122 const X86InstrInfo *getInstrInfo() const {
123 return getTargetMachine()->getInstrInfo();
125 const X86TargetMachine *getTargetMachine() const {
126 return static_cast<const X86TargetMachine *>(&TM);
129 unsigned TargetMaterializeConstant(Constant *C);
131 unsigned TargetMaterializeAlloca(AllocaInst *C);
133 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
134 /// computed in an SSE register, not on the X87 floating point stack.
135 bool isScalarFPTypeInSSEReg(MVT VT) const {
136 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
137 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
140 bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
143 } // end anonymous namespace.
145 bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
146 VT = TLI.getValueType(Ty, /*HandleUnknown=*/true);
147 if (VT == MVT::Other || !VT.isSimple())
148 // Unhandled type. Halt "fast" selection and bail.
151 // For now, require SSE/SSE2 for performing floating-point operations,
152 // since x87 requires additional work.
153 if (VT == MVT::f64 && !X86ScalarSSEf64)
155 if (VT == MVT::f32 && !X86ScalarSSEf32)
157 // Similarly, no f80 support yet.
160 // We only handle legal types. For example, on x86-32 the instruction
161 // selector contains all of the 64-bit instructions from x86-64,
162 // under the assumption that i64 won't be used if the target doesn't
164 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
167 #include "X86GenCallingConv.inc"
169 /// CCAssignFnForCall - Selects the correct CCAssignFn for a given calling
171 CCAssignFn *X86FastISel::CCAssignFnForCall(unsigned CC, bool isTaillCall) {
172 if (Subtarget->is64Bit()) {
173 if (Subtarget->isTargetWin64())
174 return CC_X86_Win64_C;
179 if (CC == CallingConv::X86_FastCall)
180 return CC_X86_32_FastCall;
181 else if (CC == CallingConv::Fast)
182 return CC_X86_32_FastCC;
187 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
188 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
189 /// Return true and the result register by reference if it is possible.
190 bool X86FastISel::X86FastEmitLoad(MVT VT, const X86AddressMode &AM,
191 unsigned &ResultReg) {
192 // Get opcode and regclass of the output for the given load instruction.
194 const TargetRegisterClass *RC = NULL;
195 switch (VT.getSimpleVT()) {
196 default: return false;
199 RC = X86::GR8RegisterClass;
203 RC = X86::GR16RegisterClass;
207 RC = X86::GR32RegisterClass;
210 // Must be in x86-64 mode.
212 RC = X86::GR64RegisterClass;
215 if (Subtarget->hasSSE1()) {
217 RC = X86::FR32RegisterClass;
220 RC = X86::RFP32RegisterClass;
224 if (Subtarget->hasSSE2()) {
226 RC = X86::FR64RegisterClass;
229 RC = X86::RFP64RegisterClass;
233 // No f80 support yet.
237 ResultReg = createResultReg(RC);
238 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
242 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
243 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
244 /// and a displacement offset, or a GlobalAddress,
245 /// i.e. V. Return true if it is possible.
247 X86FastISel::X86FastEmitStore(MVT VT, unsigned Val,
248 const X86AddressMode &AM) {
249 // Get opcode and regclass of the output for the given store instruction.
251 switch (VT.getSimpleVT()) {
252 case MVT::f80: // No f80 support yet.
253 default: return false;
254 case MVT::i8: Opc = X86::MOV8mr; break;
255 case MVT::i16: Opc = X86::MOV16mr; break;
256 case MVT::i32: Opc = X86::MOV32mr; break;
257 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
259 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
262 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
266 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM).addReg(Val);
270 bool X86FastISel::X86FastEmitStore(MVT VT, Value *Val,
271 const X86AddressMode &AM) {
272 // Handle 'null' like i32/i64 0.
273 if (isa<ConstantPointerNull>(Val))
274 Val = Constant::getNullValue(TD.getIntPtrType());
276 // If this is a store of a simple constant, fold the constant into the store.
277 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
279 switch (VT.getSimpleVT()) {
281 case MVT::i8: Opc = X86::MOV8mi; break;
282 case MVT::i16: Opc = X86::MOV16mi; break;
283 case MVT::i32: Opc = X86::MOV32mi; break;
285 // Must be a 32-bit sign extended value.
286 if ((int)CI->getSExtValue() == CI->getSExtValue())
287 Opc = X86::MOV64mi32;
292 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM)
293 .addImm(CI->getSExtValue());
298 unsigned ValReg = getRegForValue(Val);
302 return X86FastEmitStore(VT, ValReg, AM);
305 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
306 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
307 /// ISD::SIGN_EXTEND).
308 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, MVT DstVT,
309 unsigned Src, MVT SrcVT,
310 unsigned &ResultReg) {
311 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc, Src);
320 /// X86SelectAddress - Attempt to fill in an address from the given value.
322 bool X86FastISel::X86SelectAddress(Value *V, X86AddressMode &AM) {
324 unsigned Opcode = Instruction::UserOp1;
325 if (Instruction *I = dyn_cast<Instruction>(V)) {
326 Opcode = I->getOpcode();
328 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
329 Opcode = C->getOpcode();
335 case Instruction::BitCast:
336 // Look past bitcasts.
337 return X86SelectAddress(U->getOperand(0), AM);
339 case Instruction::IntToPtr:
340 // Look past no-op inttoptrs.
341 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
342 return X86SelectAddress(U->getOperand(0), AM);
345 case Instruction::PtrToInt:
346 // Look past no-op ptrtoints.
347 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
348 return X86SelectAddress(U->getOperand(0), AM);
351 case Instruction::Alloca: {
352 // Do static allocas.
353 const AllocaInst *A = cast<AllocaInst>(V);
354 DenseMap<const AllocaInst*, int>::iterator SI = StaticAllocaMap.find(A);
355 if (SI != StaticAllocaMap.end()) {
356 AM.BaseType = X86AddressMode::FrameIndexBase;
357 AM.Base.FrameIndex = SI->second;
363 case Instruction::Add: {
364 // Adds of constants are common and easy enough.
365 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
366 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
367 // They have to fit in the 32-bit signed displacement field though.
369 AM.Disp = (uint32_t)Disp;
370 return X86SelectAddress(U->getOperand(0), AM);
376 case Instruction::GetElementPtr: {
377 // Pattern-match simple GEPs.
378 uint64_t Disp = (int32_t)AM.Disp;
379 unsigned IndexReg = AM.IndexReg;
380 unsigned Scale = AM.Scale;
381 gep_type_iterator GTI = gep_type_begin(U);
382 // Iterate through the indices, folding what we can. Constants can be
383 // folded, and one dynamic index can be handled, if the scale is supported.
384 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end();
385 i != e; ++i, ++GTI) {
387 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
388 const StructLayout *SL = TD.getStructLayout(STy);
389 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
390 Disp += SL->getElementOffset(Idx);
392 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
393 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
394 // Constant-offset addressing.
395 Disp += CI->getSExtValue() * S;
396 } else if (IndexReg == 0 &&
397 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
398 (S == 1 || S == 2 || S == 4 || S == 8)) {
399 // Scaled-index addressing.
401 IndexReg = getRegForGEPIndex(Op);
406 goto unsupported_gep;
409 // Check for displacement overflow.
412 // Ok, the GEP indices were covered by constant-offset and scaled-index
413 // addressing. Update the address state and move on to examining the base.
414 AM.IndexReg = IndexReg;
416 AM.Disp = (uint32_t)Disp;
417 return X86SelectAddress(U->getOperand(0), AM);
419 // Ok, the GEP indices weren't all covered.
424 // Handle constant address.
425 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
426 // Can't handle alternate code models yet.
427 if (TM.getCodeModel() != CodeModel::Small)
430 // RIP-relative addresses can't have additional register operands.
431 if (Subtarget->isPICStyleRIPRel() &&
432 (AM.Base.Reg != 0 || AM.IndexReg != 0))
435 // Can't handle TLS yet.
436 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
437 if (GVar->isThreadLocal())
440 // Okay, we've committed to selecting this global. Set up the basic address.
443 // Allow the subtarget to classify the global.
444 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
446 // If this reference is relative to the pic base, set it now.
447 if (isGlobalRelativeToPICBase(GVFlags)) {
448 // FIXME: How do we know Base.Reg is free??
449 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(&MF);
452 // Unless the ABI requires an extra load, return a direct reference to
454 if (!isGlobalStubReference(GVFlags)) {
455 if (Subtarget->isPICStyleRIPRel()) {
456 // Use rip-relative addressing if we can. Above we verified that the
457 // base and index registers are unused.
458 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
459 AM.Base.Reg = X86::RIP;
461 AM.GVOpFlags = GVFlags;
465 // Ok, we need to do a load from a stub. If we've already loaded from this
466 // stub, reuse the loaded pointer, otherwise emit the load now.
467 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
469 if (I != LocalValueMap.end() && I->second != 0) {
472 // Issue load from stub.
474 const TargetRegisterClass *RC = NULL;
475 X86AddressMode StubAM;
476 StubAM.Base.Reg = AM.Base.Reg;
478 StubAM.GVOpFlags = GVFlags;
480 if (TLI.getPointerTy() == MVT::i64) {
482 RC = X86::GR64RegisterClass;
484 if (Subtarget->isPICStyleRIPRel())
485 StubAM.Base.Reg = X86::RIP;
488 RC = X86::GR32RegisterClass;
491 LoadReg = createResultReg(RC);
492 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);
494 // Prevent loading GV stub multiple times in same MBB.
495 LocalValueMap[V] = LoadReg;
498 // Now construct the final address. Note that the Disp, Scale,
499 // and Index values may already be set here.
500 AM.Base.Reg = LoadReg;
505 // If all else fails, try to materialize the value in a register.
506 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
507 if (AM.Base.Reg == 0) {
508 AM.Base.Reg = getRegForValue(V);
509 return AM.Base.Reg != 0;
511 if (AM.IndexReg == 0) {
512 assert(AM.Scale == 1 && "Scale with no index!");
513 AM.IndexReg = getRegForValue(V);
514 return AM.IndexReg != 0;
521 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
523 bool X86FastISel::X86SelectCallAddress(Value *V, X86AddressMode &AM) {
525 unsigned Opcode = Instruction::UserOp1;
526 if (Instruction *I = dyn_cast<Instruction>(V)) {
527 Opcode = I->getOpcode();
529 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
530 Opcode = C->getOpcode();
536 case Instruction::BitCast:
537 // Look past bitcasts.
538 return X86SelectCallAddress(U->getOperand(0), AM);
540 case Instruction::IntToPtr:
541 // Look past no-op inttoptrs.
542 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
543 return X86SelectCallAddress(U->getOperand(0), AM);
546 case Instruction::PtrToInt:
547 // Look past no-op ptrtoints.
548 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
549 return X86SelectCallAddress(U->getOperand(0), AM);
553 // Handle constant address.
554 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
555 // Can't handle alternate code models yet.
556 if (TM.getCodeModel() != CodeModel::Small)
559 // RIP-relative addresses can't have additional register operands.
560 if (Subtarget->isPICStyleRIPRel() &&
561 (AM.Base.Reg != 0 || AM.IndexReg != 0))
564 // Can't handle TLS or DLLImport.
565 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
566 if (GVar->isThreadLocal() || GVar->hasDLLImportLinkage())
569 // Okay, we've committed to selecting this global. Set up the basic address.
572 // No ABI requires an extra load for anything other than DLLImport, which
573 // we rejected above. Return a direct reference to the global.
574 if (Subtarget->isPICStyleRIPRel()) {
575 // Use rip-relative addressing if we can. Above we verified that the
576 // base and index registers are unused.
577 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
578 AM.Base.Reg = X86::RIP;
579 } else if (Subtarget->isPICStyleStubPIC()) {
580 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
581 } else if (Subtarget->isPICStyleGOT()) {
582 AM.GVOpFlags = X86II::MO_GOTOFF;
588 // If all else fails, try to materialize the value in a register.
589 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
590 if (AM.Base.Reg == 0) {
591 AM.Base.Reg = getRegForValue(V);
592 return AM.Base.Reg != 0;
594 if (AM.IndexReg == 0) {
595 assert(AM.Scale == 1 && "Scale with no index!");
596 AM.IndexReg = getRegForValue(V);
597 return AM.IndexReg != 0;
605 /// X86SelectStore - Select and emit code to implement store instructions.
606 bool X86FastISel::X86SelectStore(Instruction* I) {
608 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
612 if (!X86SelectAddress(I->getOperand(1), AM))
615 return X86FastEmitStore(VT, I->getOperand(0), AM);
618 /// X86SelectLoad - Select and emit code to implement load instructions.
620 bool X86FastISel::X86SelectLoad(Instruction *I) {
622 if (!isTypeLegal(I->getType(), VT))
626 if (!X86SelectAddress(I->getOperand(0), AM))
629 unsigned ResultReg = 0;
630 if (X86FastEmitLoad(VT, AM, ResultReg)) {
631 UpdateValueMap(I, ResultReg);
637 static unsigned X86ChooseCmpOpcode(MVT VT) {
638 switch (VT.getSimpleVT()) {
640 case MVT::i8: return X86::CMP8rr;
641 case MVT::i16: return X86::CMP16rr;
642 case MVT::i32: return X86::CMP32rr;
643 case MVT::i64: return X86::CMP64rr;
644 case MVT::f32: return X86::UCOMISSrr;
645 case MVT::f64: return X86::UCOMISDrr;
649 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
650 /// of the comparison, return an opcode that works for the compare (e.g.
651 /// CMP32ri) otherwise return 0.
652 static unsigned X86ChooseCmpImmediateOpcode(MVT VT, ConstantInt *RHSC) {
653 switch (VT.getSimpleVT()) {
654 // Otherwise, we can't fold the immediate into this comparison.
656 case MVT::i8: return X86::CMP8ri;
657 case MVT::i16: return X86::CMP16ri;
658 case MVT::i32: return X86::CMP32ri;
660 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
662 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
663 return X86::CMP64ri32;
668 bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, MVT VT) {
669 unsigned Op0Reg = getRegForValue(Op0);
670 if (Op0Reg == 0) return false;
672 // Handle 'null' like i32/i64 0.
673 if (isa<ConstantPointerNull>(Op1))
674 Op1 = Constant::getNullValue(TD.getIntPtrType());
676 // We have two options: compare with register or immediate. If the RHS of
677 // the compare is an immediate that we can fold into this compare, use
678 // CMPri, otherwise use CMPrr.
679 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
680 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
681 BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
682 .addImm(Op1C->getSExtValue());
687 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
688 if (CompareOpc == 0) return false;
690 unsigned Op1Reg = getRegForValue(Op1);
691 if (Op1Reg == 0) return false;
692 BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
697 bool X86FastISel::X86SelectCmp(Instruction *I) {
698 CmpInst *CI = cast<CmpInst>(I);
701 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
704 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
706 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
707 switch (CI->getPredicate()) {
708 case CmpInst::FCMP_OEQ: {
709 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
712 unsigned EReg = createResultReg(&X86::GR8RegClass);
713 unsigned NPReg = createResultReg(&X86::GR8RegClass);
714 BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
715 BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
717 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
718 UpdateValueMap(I, ResultReg);
721 case CmpInst::FCMP_UNE: {
722 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
725 unsigned NEReg = createResultReg(&X86::GR8RegClass);
726 unsigned PReg = createResultReg(&X86::GR8RegClass);
727 BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
728 BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
729 BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
730 UpdateValueMap(I, ResultReg);
733 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
734 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
735 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
736 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
737 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
738 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
739 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
740 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
741 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
742 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
743 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
744 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
746 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
747 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
748 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
749 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
750 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
751 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
752 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
753 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
754 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
755 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
760 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
764 // Emit a compare of Op0/Op1.
765 if (!X86FastEmitCompare(Op0, Op1, VT))
768 BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
769 UpdateValueMap(I, ResultReg);
773 bool X86FastISel::X86SelectZExt(Instruction *I) {
774 // Handle zero-extension from i1 to i8, which is common.
775 if (I->getType() == Type::Int8Ty &&
776 I->getOperand(0)->getType() == Type::Int1Ty) {
777 unsigned ResultReg = getRegForValue(I->getOperand(0));
778 if (ResultReg == 0) return false;
779 // Set the high bits to zero.
780 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg);
781 if (ResultReg == 0) return false;
782 UpdateValueMap(I, ResultReg);
790 bool X86FastISel::X86SelectBranch(Instruction *I) {
791 // Unconditional branches are selected by tablegen-generated code.
792 // Handle a conditional branch.
793 BranchInst *BI = cast<BranchInst>(I);
794 MachineBasicBlock *TrueMBB = MBBMap[BI->getSuccessor(0)];
795 MachineBasicBlock *FalseMBB = MBBMap[BI->getSuccessor(1)];
797 // Fold the common case of a conditional branch with a comparison.
798 if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
799 if (CI->hasOneUse()) {
800 MVT VT = TLI.getValueType(CI->getOperand(0)->getType());
802 // Try to take advantage of fallthrough opportunities.
803 CmpInst::Predicate Predicate = CI->getPredicate();
804 if (MBB->isLayoutSuccessor(TrueMBB)) {
805 std::swap(TrueMBB, FalseMBB);
806 Predicate = CmpInst::getInversePredicate(Predicate);
809 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
810 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
813 case CmpInst::FCMP_OEQ:
814 std::swap(TrueMBB, FalseMBB);
815 Predicate = CmpInst::FCMP_UNE;
817 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE; break;
818 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA; break;
819 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE; break;
820 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA; break;
821 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE; break;
822 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE; break;
823 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP; break;
824 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP; break;
825 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE; break;
826 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB; break;
827 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE; break;
828 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
829 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
831 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE; break;
832 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE; break;
833 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA; break;
834 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE; break;
835 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
836 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
837 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG; break;
838 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE; break;
839 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL; break;
840 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE; break;
845 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
849 // Emit a compare of the LHS and RHS, setting the flags.
850 if (!X86FastEmitCompare(Op0, Op1, VT))
853 BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
855 if (Predicate == CmpInst::FCMP_UNE) {
856 // X86 requires a second branch to handle UNE (and OEQ,
857 // which is mapped to UNE above).
858 BuildMI(MBB, DL, TII.get(X86::JP)).addMBB(TrueMBB);
861 FastEmitBranch(FalseMBB);
862 MBB->addSuccessor(TrueMBB);
865 } else if (ExtractValueInst *EI =
866 dyn_cast<ExtractValueInst>(BI->getCondition())) {
867 // Check to see if the branch instruction is from an "arithmetic with
868 // overflow" intrinsic. The main way these intrinsics are used is:
870 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
871 // %sum = extractvalue { i32, i1 } %t, 0
872 // %obit = extractvalue { i32, i1 } %t, 1
873 // br i1 %obit, label %overflow, label %normal
875 // The %sum and %obit are converted in an ADD and a SETO/SETB before
876 // reaching the branch. Therefore, we search backwards through the MBB
877 // looking for the SETO/SETB instruction. If an instruction modifies the
878 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
879 // convert the branch into a JO/JB instruction.
880 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
881 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
882 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
883 const MachineInstr *SetMI = 0;
884 unsigned Reg = lookUpRegForValue(EI);
886 for (MachineBasicBlock::const_reverse_iterator
887 RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
888 const MachineInstr &MI = *RI;
890 if (MI.modifiesRegister(Reg)) {
891 unsigned Src, Dst, SrcSR, DstSR;
893 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
902 const TargetInstrDesc &TID = MI.getDesc();
903 if (TID.hasUnmodeledSideEffects() ||
904 TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
909 unsigned OpCode = SetMI->getOpcode();
911 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
912 BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ? X86::JO : X86::JB))
914 FastEmitBranch(FalseMBB);
915 MBB->addSuccessor(TrueMBB);
923 // Otherwise do a clumsy setcc and re-test it.
924 unsigned OpReg = getRegForValue(BI->getCondition());
925 if (OpReg == 0) return false;
927 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
928 BuildMI(MBB, DL, TII.get(X86::JNE)).addMBB(TrueMBB);
929 FastEmitBranch(FalseMBB);
930 MBB->addSuccessor(TrueMBB);
934 bool X86FastISel::X86SelectShift(Instruction *I) {
935 unsigned CReg = 0, OpReg = 0, OpImm = 0;
936 const TargetRegisterClass *RC = NULL;
937 if (I->getType() == Type::Int8Ty) {
939 RC = &X86::GR8RegClass;
940 switch (I->getOpcode()) {
941 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
942 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
943 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
944 default: return false;
946 } else if (I->getType() == Type::Int16Ty) {
948 RC = &X86::GR16RegClass;
949 switch (I->getOpcode()) {
950 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
951 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
952 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
953 default: return false;
955 } else if (I->getType() == Type::Int32Ty) {
957 RC = &X86::GR32RegClass;
958 switch (I->getOpcode()) {
959 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
960 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
961 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
962 default: return false;
964 } else if (I->getType() == Type::Int64Ty) {
966 RC = &X86::GR64RegClass;
967 switch (I->getOpcode()) {
968 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
969 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
970 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
971 default: return false;
977 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
978 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
981 unsigned Op0Reg = getRegForValue(I->getOperand(0));
982 if (Op0Reg == 0) return false;
984 // Fold immediate in shl(x,3).
985 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
986 unsigned ResultReg = createResultReg(RC);
987 BuildMI(MBB, DL, TII.get(OpImm),
988 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
989 UpdateValueMap(I, ResultReg);
993 unsigned Op1Reg = getRegForValue(I->getOperand(1));
994 if (Op1Reg == 0) return false;
995 TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);
997 // The shift instruction uses X86::CL. If we defined a super-register
998 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
1000 if (CReg != X86::CL)
1001 BuildMI(MBB, DL, TII.get(TargetInstrInfo::EXTRACT_SUBREG), X86::CL)
1002 .addReg(CReg).addImm(X86::SUBREG_8BIT);
1004 unsigned ResultReg = createResultReg(RC);
1005 BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
1006 UpdateValueMap(I, ResultReg);
1010 bool X86FastISel::X86SelectSelect(Instruction *I) {
1011 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1012 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1016 const TargetRegisterClass *RC = NULL;
1017 if (VT.getSimpleVT() == MVT::i16) {
1018 Opc = X86::CMOVE16rr;
1019 RC = &X86::GR16RegClass;
1020 } else if (VT.getSimpleVT() == MVT::i32) {
1021 Opc = X86::CMOVE32rr;
1022 RC = &X86::GR32RegClass;
1023 } else if (VT.getSimpleVT() == MVT::i64) {
1024 Opc = X86::CMOVE64rr;
1025 RC = &X86::GR64RegClass;
1030 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1031 if (Op0Reg == 0) return false;
1032 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1033 if (Op1Reg == 0) return false;
1034 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1035 if (Op2Reg == 0) return false;
1037 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
1038 unsigned ResultReg = createResultReg(RC);
1039 BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
1040 UpdateValueMap(I, ResultReg);
1044 bool X86FastISel::X86SelectFPExt(Instruction *I) {
1045 // fpext from float to double.
1046 if (Subtarget->hasSSE2() && I->getType() == Type::DoubleTy) {
1047 Value *V = I->getOperand(0);
1048 if (V->getType() == Type::FloatTy) {
1049 unsigned OpReg = getRegForValue(V);
1050 if (OpReg == 0) return false;
1051 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1052 BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
1053 UpdateValueMap(I, ResultReg);
1061 bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
1062 if (Subtarget->hasSSE2()) {
1063 if (I->getType() == Type::FloatTy) {
1064 Value *V = I->getOperand(0);
1065 if (V->getType() == Type::DoubleTy) {
1066 unsigned OpReg = getRegForValue(V);
1067 if (OpReg == 0) return false;
1068 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1069 BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
1070 UpdateValueMap(I, ResultReg);
1079 bool X86FastISel::X86SelectTrunc(Instruction *I) {
1080 if (Subtarget->is64Bit())
1081 // All other cases should be handled by the tblgen generated code.
1083 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1084 MVT DstVT = TLI.getValueType(I->getType());
1086 // This code only handles truncation to byte right now.
1087 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1088 // All other cases should be handled by the tblgen generated code.
1090 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1091 // All other cases should be handled by the tblgen generated code.
1094 unsigned InputReg = getRegForValue(I->getOperand(0));
1096 // Unhandled operand. Halt "fast" selection and bail.
1099 // First issue a copy to GR16_ABCD or GR32_ABCD.
1100 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
1101 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1102 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1103 unsigned CopyReg = createResultReg(CopyRC);
1104 BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
1106 // Then issue an extract_subreg.
1107 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1108 CopyReg, X86::SUBREG_8BIT);
1112 UpdateValueMap(I, ResultReg);
1116 bool X86FastISel::X86SelectExtractValue(Instruction *I) {
1117 ExtractValueInst *EI = cast<ExtractValueInst>(I);
1118 Value *Agg = EI->getAggregateOperand();
1120 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1121 switch (CI->getIntrinsicID()) {
1123 case Intrinsic::sadd_with_overflow:
1124 case Intrinsic::uadd_with_overflow:
1125 // Cheat a little. We know that the registers for "add" and "seto" are
1126 // allocated sequentially. However, we only keep track of the register
1127 // for "add" in the value map. Use extractvalue's index to get the
1128 // correct register for "seto".
1129 UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
1137 bool X86FastISel::X86VisitIntrinsicCall(IntrinsicInst &I) {
1138 // FIXME: Handle more intrinsics.
1139 switch (I.getIntrinsicID()) {
1140 default: return false;
1141 case Intrinsic::sadd_with_overflow:
1142 case Intrinsic::uadd_with_overflow: {
1143 // Replace "add with overflow" intrinsics with an "add" instruction followed
1144 // by a seto/setc instruction. Later on, when the "extractvalue"
1145 // instructions are encountered, we use the fact that two registers were
1146 // created sequentially to get the correct registers for the "sum" and the
1148 const Function *Callee = I.getCalledFunction();
1150 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1153 if (!isTypeLegal(RetTy, VT))
1156 Value *Op1 = I.getOperand(1);
1157 Value *Op2 = I.getOperand(2);
1158 unsigned Reg1 = getRegForValue(Op1);
1159 unsigned Reg2 = getRegForValue(Op2);
1161 if (Reg1 == 0 || Reg2 == 0)
1162 // FIXME: Handle values *not* in registers.
1168 else if (VT == MVT::i64)
1173 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1174 BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
1175 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1177 // If the add with overflow is an intra-block value then we just want to
1178 // create temporaries for it like normal. If it is a cross-block value then
1179 // UpdateValueMap will return the cross-block register used. Since we
1180 // *really* want the value to be live in the register pair known by
1181 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1182 // the cross block case. In the non-cross-block case, we should just make
1183 // another register for the value.
1184 if (DestReg1 != ResultReg)
1185 ResultReg = DestReg1+1;
1187 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1189 unsigned Opc = X86::SETBr;
1190 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1192 BuildMI(MBB, DL, TII.get(Opc), ResultReg);
1198 bool X86FastISel::X86SelectCall(Instruction *I) {
1199 CallInst *CI = cast<CallInst>(I);
1200 Value *Callee = I->getOperand(0);
1202 // Can't handle inline asm yet.
1203 if (isa<InlineAsm>(Callee))
1206 // Handle intrinsic calls.
1207 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1208 return X86VisitIntrinsicCall(*II);
1210 // Handle only C and fastcc calling conventions for now.
1212 unsigned CC = CS.getCallingConv();
1213 if (CC != CallingConv::C &&
1214 CC != CallingConv::Fast &&
1215 CC != CallingConv::X86_FastCall)
1218 // On X86, -tailcallopt changes the fastcc ABI. FastISel doesn't
1219 // handle this for now.
1220 if (CC == CallingConv::Fast && PerformTailCallOpt)
1223 // Let SDISel handle vararg functions.
1224 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1225 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1226 if (FTy->isVarArg())
1229 // Handle *simple* calls for now.
1230 const Type *RetTy = CS.getType();
1232 if (RetTy == Type::VoidTy)
1233 RetVT = MVT::isVoid;
1234 else if (!isTypeLegal(RetTy, RetVT, true))
1237 // Materialize callee address in a register. FIXME: GV address can be
1238 // handled with a CALLpcrel32 instead.
1239 X86AddressMode CalleeAM;
1240 if (!X86SelectCallAddress(Callee, CalleeAM))
1242 unsigned CalleeOp = 0;
1243 GlobalValue *GV = 0;
1244 if (CalleeAM.GV != 0) {
1246 } else if (CalleeAM.Base.Reg != 0) {
1247 CalleeOp = CalleeAM.Base.Reg;
1251 // Allow calls which produce i1 results.
1252 bool AndToI1 = false;
1253 if (RetVT == MVT::i1) {
1258 // Deal with call operands first.
1259 SmallVector<Value*, 8> ArgVals;
1260 SmallVector<unsigned, 8> Args;
1261 SmallVector<MVT, 8> ArgVTs;
1262 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1263 Args.reserve(CS.arg_size());
1264 ArgVals.reserve(CS.arg_size());
1265 ArgVTs.reserve(CS.arg_size());
1266 ArgFlags.reserve(CS.arg_size());
1267 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1269 unsigned Arg = getRegForValue(*i);
1272 ISD::ArgFlagsTy Flags;
1273 unsigned AttrInd = i - CS.arg_begin() + 1;
1274 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1276 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1279 // FIXME: Only handle *easy* calls for now.
1280 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1281 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1282 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1283 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1286 const Type *ArgTy = (*i)->getType();
1288 if (!isTypeLegal(ArgTy, ArgVT))
1290 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1291 Flags.setOrigAlign(OriginalAlignment);
1293 Args.push_back(Arg);
1294 ArgVals.push_back(*i);
1295 ArgVTs.push_back(ArgVT);
1296 ArgFlags.push_back(Flags);
1299 // Analyze operands of the call, assigning locations to each operand.
1300 SmallVector<CCValAssign, 16> ArgLocs;
1301 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1302 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1304 // Get a count of how many bytes are to be pushed on the stack.
1305 unsigned NumBytes = CCInfo.getNextStackOffset();
1307 // Issue CALLSEQ_START
1308 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1309 BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
1311 // Process argument: walk the register/memloc assignments, inserting
1313 SmallVector<unsigned, 4> RegArgs;
1314 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1315 CCValAssign &VA = ArgLocs[i];
1316 unsigned Arg = Args[VA.getValNo()];
1317 MVT ArgVT = ArgVTs[VA.getValNo()];
1319 // Promote the value if needed.
1320 switch (VA.getLocInfo()) {
1321 default: assert(0 && "Unknown loc info!");
1322 case CCValAssign::Full: break;
1323 case CCValAssign::SExt: {
1324 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1326 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1328 ArgVT = VA.getLocVT();
1331 case CCValAssign::ZExt: {
1332 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1334 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1336 ArgVT = VA.getLocVT();
1339 case CCValAssign::AExt: {
1340 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1343 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1346 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1349 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1350 ArgVT = VA.getLocVT();
1355 if (VA.isRegLoc()) {
1356 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1357 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
1359 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1361 RegArgs.push_back(VA.getLocReg());
1363 unsigned LocMemOffset = VA.getLocMemOffset();
1365 AM.Base.Reg = StackPtr;
1366 AM.Disp = LocMemOffset;
1367 Value *ArgVal = ArgVals[VA.getValNo()];
1369 // If this is a really simple value, emit this with the Value* version of
1370 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1371 // can cause us to reevaluate the argument.
1372 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1373 X86FastEmitStore(ArgVT, ArgVal, AM);
1375 X86FastEmitStore(ArgVT, Arg, AM);
1379 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1381 if (Subtarget->isPICStyleGOT()) {
1382 TargetRegisterClass *RC = X86::GR32RegisterClass;
1383 unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
1384 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
1385 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1390 MachineInstrBuilder MIB;
1392 // Register-indirect call.
1393 unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
1394 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);
1398 assert(GV && "Not a direct call");
1400 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
1402 // See if we need any target-specific flags on the GV operand.
1403 unsigned char OpFlags = 0;
1405 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1406 // external symbols most go through the PLT in PIC mode. If the symbol
1407 // has hidden or protected visibility, or if it is static or local, then
1408 // we don't need to use the PLT - we can directly call it.
1409 if (Subtarget->isTargetELF() &&
1410 TM.getRelocationModel() == Reloc::PIC_ &&
1411 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1412 OpFlags = X86II::MO_PLT;
1413 } else if (Subtarget->isPICStyleStubAny() &&
1414 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1415 Subtarget->getDarwinVers() < 9) {
1416 // PC-relative references to external symbols should go through $stub,
1417 // unless we're building with the leopard linker or later, which
1418 // automatically synthesizes these stubs.
1419 OpFlags = X86II::MO_DARWIN_STUB;
1423 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
1426 // Add an implicit use GOT pointer in EBX.
1427 if (Subtarget->isPICStyleGOT())
1428 MIB.addReg(X86::EBX);
1430 // Add implicit physical register uses to the call.
1431 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1432 MIB.addReg(RegArgs[i]);
1434 // Issue CALLSEQ_END
1435 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1436 BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
1438 // Now handle call return value (if any).
1439 if (RetVT.getSimpleVT() != MVT::isVoid) {
1440 SmallVector<CCValAssign, 16> RVLocs;
1441 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1442 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1444 // Copy all of the result registers out of their specified physreg.
1445 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1446 MVT CopyVT = RVLocs[0].getValVT();
1447 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1448 TargetRegisterClass *SrcRC = DstRC;
1450 // If this is a call to a function that returns an fp value on the x87 fp
1451 // stack, but where we prefer to use the value in xmm registers, copy it
1452 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1453 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1454 RVLocs[0].getLocReg() == X86::ST1) &&
1455 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1457 SrcRC = X86::RSTRegisterClass;
1458 DstRC = X86::RFP80RegisterClass;
1461 unsigned ResultReg = createResultReg(DstRC);
1462 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1463 RVLocs[0].getLocReg(), DstRC, SrcRC);
1464 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1466 if (CopyVT != RVLocs[0].getValVT()) {
1467 // Round the F80 the right size, which also moves to the appropriate xmm
1468 // register. This is accomplished by storing the F80 value in memory and
1469 // then loading it back. Ewww...
1470 MVT ResVT = RVLocs[0].getValVT();
1471 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1472 unsigned MemSize = ResVT.getSizeInBits()/8;
1473 int FI = MFI.CreateStackObject(MemSize, MemSize);
1474 addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
1475 DstRC = ResVT == MVT::f32
1476 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1477 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1478 ResultReg = createResultReg(DstRC);
1479 addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
1483 // Mask out all but lowest bit for some call which produces an i1.
1484 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1486 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1487 ResultReg = AndResult;
1490 UpdateValueMap(I, ResultReg);
1498 X86FastISel::TargetSelectInstruction(Instruction *I) {
1499 switch (I->getOpcode()) {
1501 case Instruction::Load:
1502 return X86SelectLoad(I);
1503 case Instruction::Store:
1504 return X86SelectStore(I);
1505 case Instruction::ICmp:
1506 case Instruction::FCmp:
1507 return X86SelectCmp(I);
1508 case Instruction::ZExt:
1509 return X86SelectZExt(I);
1510 case Instruction::Br:
1511 return X86SelectBranch(I);
1512 case Instruction::Call:
1513 return X86SelectCall(I);
1514 case Instruction::LShr:
1515 case Instruction::AShr:
1516 case Instruction::Shl:
1517 return X86SelectShift(I);
1518 case Instruction::Select:
1519 return X86SelectSelect(I);
1520 case Instruction::Trunc:
1521 return X86SelectTrunc(I);
1522 case Instruction::FPExt:
1523 return X86SelectFPExt(I);
1524 case Instruction::FPTrunc:
1525 return X86SelectFPTrunc(I);
1526 case Instruction::ExtractValue:
1527 return X86SelectExtractValue(I);
1528 case Instruction::IntToPtr: // Deliberate fall-through.
1529 case Instruction::PtrToInt: {
1530 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1531 MVT DstVT = TLI.getValueType(I->getType());
1532 if (DstVT.bitsGT(SrcVT))
1533 return X86SelectZExt(I);
1534 if (DstVT.bitsLT(SrcVT))
1535 return X86SelectTrunc(I);
1536 unsigned Reg = getRegForValue(I->getOperand(0));
1537 if (Reg == 0) return false;
1538 UpdateValueMap(I, Reg);
1546 unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
1548 if (!isTypeLegal(C->getType(), VT))
1551 // Get opcode and regclass of the output for the given load instruction.
1553 const TargetRegisterClass *RC = NULL;
1554 switch (VT.getSimpleVT()) {
1555 default: return false;
1558 RC = X86::GR8RegisterClass;
1562 RC = X86::GR16RegisterClass;
1566 RC = X86::GR32RegisterClass;
1569 // Must be in x86-64 mode.
1571 RC = X86::GR64RegisterClass;
1574 if (Subtarget->hasSSE1()) {
1576 RC = X86::FR32RegisterClass;
1578 Opc = X86::LD_Fp32m;
1579 RC = X86::RFP32RegisterClass;
1583 if (Subtarget->hasSSE2()) {
1585 RC = X86::FR64RegisterClass;
1587 Opc = X86::LD_Fp64m;
1588 RC = X86::RFP64RegisterClass;
1592 // No f80 support yet.
1596 // Materialize addresses with LEA instructions.
1597 if (isa<GlobalValue>(C)) {
1599 if (X86SelectAddress(C, AM)) {
1600 if (TLI.getPointerTy() == MVT::i32)
1604 unsigned ResultReg = createResultReg(RC);
1605 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1611 // MachineConstantPool wants an explicit alignment.
1612 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1614 // Alignment of vector types. FIXME!
1615 Align = TD.getTypeAllocSize(C->getType());
1618 // x86-32 PIC requires a PIC base register for constant pools.
1619 unsigned PICBase = 0;
1620 unsigned char OpFlag = 0;
1621 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1622 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1623 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1624 } else if (Subtarget->isPICStyleGOT()) {
1625 OpFlag = X86II::MO_GOTOFF;
1626 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1627 } else if (Subtarget->isPICStyleRIPRel() &&
1628 TM.getCodeModel() == CodeModel::Small) {
1632 // Create the load from the constant pool.
1633 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1634 unsigned ResultReg = createResultReg(RC);
1635 addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
1636 MCPOffset, PICBase, OpFlag);
1641 unsigned X86FastISel::TargetMaterializeAlloca(AllocaInst *C) {
1642 // Fail on dynamic allocas. At this point, getRegForValue has already
1643 // checked its CSE maps, so if we're here trying to handle a dynamic
1644 // alloca, we're not going to succeed. X86SelectAddress has a
1645 // check for dynamic allocas, because it's called directly from
1646 // various places, but TargetMaterializeAlloca also needs a check
1647 // in order to avoid recursion between getRegForValue,
1648 // X86SelectAddrss, and TargetMaterializeAlloca.
1649 if (!StaticAllocaMap.count(C))
1653 if (!X86SelectAddress(C, AM))
1655 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1656 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1657 unsigned ResultReg = createResultReg(RC);
1658 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1663 llvm::FastISel *X86::createFastISel(MachineFunction &mf,
1664 MachineModuleInfo *mmi,
1666 DenseMap<const Value *, unsigned> &vm,
1667 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
1668 DenseMap<const AllocaInst *, int> &am
1670 , SmallSet<Instruction*, 8> &cil
1673 return new X86FastISel(mf, mmi, dw, vm, bm, am