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::Default &&
428 TM.getCodeModel() != CodeModel::Small)
431 // RIP-relative addresses can't have additional register operands.
432 if (Subtarget->isPICStyleRIPRel() &&
433 (AM.Base.Reg != 0 || AM.IndexReg != 0))
436 // Can't handle TLS yet.
437 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
438 if (GVar->isThreadLocal())
441 // Okay, we've committed to selecting this global. Set up the basic address.
444 if (TM.getRelocationModel() == Reloc::PIC_ &&
445 !Subtarget->is64Bit()) {
446 // FIXME: How do we know Base.Reg is free??
447 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(&MF);
450 // If the ABI doesn't require an extra load, return a direct reference to
452 if (!Subtarget->GVRequiresExtraLoad(GV, TM)) {
453 if (Subtarget->isPICStyleRIPRel()) {
454 // Use rip-relative addressing if we can. Above we verified that the
455 // base and index registers are unused.
456 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
457 AM.Base.Reg = X86::RIP;
458 } else if (Subtarget->isPICStyleStub() &&
459 TM.getRelocationModel() == Reloc::PIC_) {
460 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
461 } else if (Subtarget->isPICStyleGOT()) {
462 AM.GVOpFlags = X86II::MO_GOTOFF;
468 // Check to see if we've already materialized this stub loaded value into a
469 // register in this block. If so, just reuse it.
470 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
472 if (I != LocalValueMap.end() && I->second != 0) {
475 // Issue load from stub.
477 const TargetRegisterClass *RC = NULL;
478 X86AddressMode StubAM;
479 StubAM.Base.Reg = AM.Base.Reg;
482 if (TLI.getPointerTy() == MVT::i64) {
484 RC = X86::GR64RegisterClass;
486 if (Subtarget->isPICStyleRIPRel()) {
487 StubAM.GVOpFlags = X86II::MO_GOTPCREL;
488 StubAM.Base.Reg = X86::RIP;
493 RC = X86::GR32RegisterClass;
495 if (Subtarget->isPICStyleGOT())
496 StubAM.GVOpFlags = X86II::MO_GOT;
497 else if (Subtarget->isPICStyleStub()) {
498 // In darwin, we have multiple different stub types, and we have both
499 // PIC and -mdynamic-no-pic. Determine whether we have a stub
500 // reference and/or whether the reference is relative to the PIC base
502 bool IsPIC = TM.getRelocationModel() == Reloc::PIC_;
504 if (!GV->hasHiddenVisibility()) {
505 // Non-hidden $non_lazy_ptr reference.
506 StubAM.GVOpFlags = IsPIC ? X86II::MO_DARWIN_NONLAZY_PIC_BASE :
507 X86II::MO_DARWIN_NONLAZY;
509 // Hidden $non_lazy_ptr reference.
510 StubAM.GVOpFlags = IsPIC ? X86II::MO_DARWIN_HIDDEN_NONLAZY_PIC_BASE:
511 X86II::MO_DARWIN_HIDDEN_NONLAZY;
516 LoadReg = createResultReg(RC);
517 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);
519 // Prevent loading GV stub multiple times in same MBB.
520 LocalValueMap[V] = LoadReg;
523 // Now construct the final address. Note that the Disp, Scale,
524 // and Index values may already be set here.
525 AM.Base.Reg = LoadReg;
530 // If all else fails, try to materialize the value in a register.
531 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
532 if (AM.Base.Reg == 0) {
533 AM.Base.Reg = getRegForValue(V);
534 return AM.Base.Reg != 0;
536 if (AM.IndexReg == 0) {
537 assert(AM.Scale == 1 && "Scale with no index!");
538 AM.IndexReg = getRegForValue(V);
539 return AM.IndexReg != 0;
546 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
548 bool X86FastISel::X86SelectCallAddress(Value *V, X86AddressMode &AM) {
550 unsigned Opcode = Instruction::UserOp1;
551 if (Instruction *I = dyn_cast<Instruction>(V)) {
552 Opcode = I->getOpcode();
554 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
555 Opcode = C->getOpcode();
561 case Instruction::BitCast:
562 // Look past bitcasts.
563 return X86SelectCallAddress(U->getOperand(0), AM);
565 case Instruction::IntToPtr:
566 // Look past no-op inttoptrs.
567 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
568 return X86SelectCallAddress(U->getOperand(0), AM);
571 case Instruction::PtrToInt:
572 // Look past no-op ptrtoints.
573 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
574 return X86SelectCallAddress(U->getOperand(0), AM);
578 // Handle constant address.
579 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
580 // Can't handle alternate code models yet.
581 if (TM.getCodeModel() != CodeModel::Default &&
582 TM.getCodeModel() != CodeModel::Small)
585 // RIP-relative addresses can't have additional register operands.
586 if (Subtarget->isPICStyleRIPRel() &&
587 (AM.Base.Reg != 0 || AM.IndexReg != 0))
590 // Can't handle TLS or DLLImport.
591 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
592 if (GVar->isThreadLocal() || GVar->hasDLLImportLinkage())
595 // Okay, we've committed to selecting this global. Set up the basic address.
598 // No ABI requires an extra load for anything other than DLLImport, which
599 // we rejected above. Return a direct reference to the global.
600 if (Subtarget->isPICStyleRIPRel()) {
601 // Use rip-relative addressing if we can. Above we verified that the
602 // base and index registers are unused.
603 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
604 AM.Base.Reg = X86::RIP;
605 } else if (Subtarget->isPICStyleStub() &&
606 TM.getRelocationModel() == Reloc::PIC_) {
607 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
608 } else if (Subtarget->isPICStyleGOT()) {
609 AM.GVOpFlags = X86II::MO_GOTOFF;
615 // If all else fails, try to materialize the value in a register.
616 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
617 if (AM.Base.Reg == 0) {
618 AM.Base.Reg = getRegForValue(V);
619 return AM.Base.Reg != 0;
621 if (AM.IndexReg == 0) {
622 assert(AM.Scale == 1 && "Scale with no index!");
623 AM.IndexReg = getRegForValue(V);
624 return AM.IndexReg != 0;
632 /// X86SelectStore - Select and emit code to implement store instructions.
633 bool X86FastISel::X86SelectStore(Instruction* I) {
635 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
639 if (!X86SelectAddress(I->getOperand(1), AM))
642 return X86FastEmitStore(VT, I->getOperand(0), AM);
645 /// X86SelectLoad - Select and emit code to implement load instructions.
647 bool X86FastISel::X86SelectLoad(Instruction *I) {
649 if (!isTypeLegal(I->getType(), VT))
653 if (!X86SelectAddress(I->getOperand(0), AM))
656 unsigned ResultReg = 0;
657 if (X86FastEmitLoad(VT, AM, ResultReg)) {
658 UpdateValueMap(I, ResultReg);
664 static unsigned X86ChooseCmpOpcode(MVT VT) {
665 switch (VT.getSimpleVT()) {
667 case MVT::i8: return X86::CMP8rr;
668 case MVT::i16: return X86::CMP16rr;
669 case MVT::i32: return X86::CMP32rr;
670 case MVT::i64: return X86::CMP64rr;
671 case MVT::f32: return X86::UCOMISSrr;
672 case MVT::f64: return X86::UCOMISDrr;
676 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
677 /// of the comparison, return an opcode that works for the compare (e.g.
678 /// CMP32ri) otherwise return 0.
679 static unsigned X86ChooseCmpImmediateOpcode(MVT VT, ConstantInt *RHSC) {
680 switch (VT.getSimpleVT()) {
681 // Otherwise, we can't fold the immediate into this comparison.
683 case MVT::i8: return X86::CMP8ri;
684 case MVT::i16: return X86::CMP16ri;
685 case MVT::i32: return X86::CMP32ri;
687 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
689 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
690 return X86::CMP64ri32;
695 bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, MVT VT) {
696 unsigned Op0Reg = getRegForValue(Op0);
697 if (Op0Reg == 0) return false;
699 // Handle 'null' like i32/i64 0.
700 if (isa<ConstantPointerNull>(Op1))
701 Op1 = Constant::getNullValue(TD.getIntPtrType());
703 // We have two options: compare with register or immediate. If the RHS of
704 // the compare is an immediate that we can fold into this compare, use
705 // CMPri, otherwise use CMPrr.
706 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
707 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
708 BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
709 .addImm(Op1C->getSExtValue());
714 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
715 if (CompareOpc == 0) return false;
717 unsigned Op1Reg = getRegForValue(Op1);
718 if (Op1Reg == 0) return false;
719 BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
724 bool X86FastISel::X86SelectCmp(Instruction *I) {
725 CmpInst *CI = cast<CmpInst>(I);
728 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
731 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
733 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
734 switch (CI->getPredicate()) {
735 case CmpInst::FCMP_OEQ: {
736 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
739 unsigned EReg = createResultReg(&X86::GR8RegClass);
740 unsigned NPReg = createResultReg(&X86::GR8RegClass);
741 BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
742 BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
744 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
745 UpdateValueMap(I, ResultReg);
748 case CmpInst::FCMP_UNE: {
749 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
752 unsigned NEReg = createResultReg(&X86::GR8RegClass);
753 unsigned PReg = createResultReg(&X86::GR8RegClass);
754 BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
755 BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
756 BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
757 UpdateValueMap(I, ResultReg);
760 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
761 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
762 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
763 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
764 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
765 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
766 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
767 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
768 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
769 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
770 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
771 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
773 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
774 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
775 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
776 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
777 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
778 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
779 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
780 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
781 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
782 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
787 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
791 // Emit a compare of Op0/Op1.
792 if (!X86FastEmitCompare(Op0, Op1, VT))
795 BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
796 UpdateValueMap(I, ResultReg);
800 bool X86FastISel::X86SelectZExt(Instruction *I) {
801 // Handle zero-extension from i1 to i8, which is common.
802 if (I->getType() == Type::Int8Ty &&
803 I->getOperand(0)->getType() == Type::Int1Ty) {
804 unsigned ResultReg = getRegForValue(I->getOperand(0));
805 if (ResultReg == 0) return false;
806 // Set the high bits to zero.
807 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg);
808 if (ResultReg == 0) return false;
809 UpdateValueMap(I, ResultReg);
817 bool X86FastISel::X86SelectBranch(Instruction *I) {
818 // Unconditional branches are selected by tablegen-generated code.
819 // Handle a conditional branch.
820 BranchInst *BI = cast<BranchInst>(I);
821 MachineBasicBlock *TrueMBB = MBBMap[BI->getSuccessor(0)];
822 MachineBasicBlock *FalseMBB = MBBMap[BI->getSuccessor(1)];
824 // Fold the common case of a conditional branch with a comparison.
825 if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
826 if (CI->hasOneUse()) {
827 MVT VT = TLI.getValueType(CI->getOperand(0)->getType());
829 // Try to take advantage of fallthrough opportunities.
830 CmpInst::Predicate Predicate = CI->getPredicate();
831 if (MBB->isLayoutSuccessor(TrueMBB)) {
832 std::swap(TrueMBB, FalseMBB);
833 Predicate = CmpInst::getInversePredicate(Predicate);
836 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
837 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
840 case CmpInst::FCMP_OEQ:
841 std::swap(TrueMBB, FalseMBB);
842 Predicate = CmpInst::FCMP_UNE;
844 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE; break;
845 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA; break;
846 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE; break;
847 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA; break;
848 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE; break;
849 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE; break;
850 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP; break;
851 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP; break;
852 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE; break;
853 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB; break;
854 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE; break;
855 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
856 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
858 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE; break;
859 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE; break;
860 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA; break;
861 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE; break;
862 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
863 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
864 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG; break;
865 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE; break;
866 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL; break;
867 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE; break;
872 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
876 // Emit a compare of the LHS and RHS, setting the flags.
877 if (!X86FastEmitCompare(Op0, Op1, VT))
880 BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
882 if (Predicate == CmpInst::FCMP_UNE) {
883 // X86 requires a second branch to handle UNE (and OEQ,
884 // which is mapped to UNE above).
885 BuildMI(MBB, DL, TII.get(X86::JP)).addMBB(TrueMBB);
888 FastEmitBranch(FalseMBB);
889 MBB->addSuccessor(TrueMBB);
892 } else if (ExtractValueInst *EI =
893 dyn_cast<ExtractValueInst>(BI->getCondition())) {
894 // Check to see if the branch instruction is from an "arithmetic with
895 // overflow" intrinsic. The main way these intrinsics are used is:
897 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
898 // %sum = extractvalue { i32, i1 } %t, 0
899 // %obit = extractvalue { i32, i1 } %t, 1
900 // br i1 %obit, label %overflow, label %normal
902 // The %sum and %obit are converted in an ADD and a SETO/SETB before
903 // reaching the branch. Therefore, we search backwards through the MBB
904 // looking for the SETO/SETB instruction. If an instruction modifies the
905 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
906 // convert the branch into a JO/JB instruction.
907 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
908 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
909 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
910 const MachineInstr *SetMI = 0;
911 unsigned Reg = lookUpRegForValue(EI);
913 for (MachineBasicBlock::const_reverse_iterator
914 RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
915 const MachineInstr &MI = *RI;
917 if (MI.modifiesRegister(Reg)) {
918 unsigned Src, Dst, SrcSR, DstSR;
920 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
929 const TargetInstrDesc &TID = MI.getDesc();
930 if (TID.hasUnmodeledSideEffects() ||
931 TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
936 unsigned OpCode = SetMI->getOpcode();
938 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
939 BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ? X86::JO : X86::JB))
941 FastEmitBranch(FalseMBB);
942 MBB->addSuccessor(TrueMBB);
950 // Otherwise do a clumsy setcc and re-test it.
951 unsigned OpReg = getRegForValue(BI->getCondition());
952 if (OpReg == 0) return false;
954 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
955 BuildMI(MBB, DL, TII.get(X86::JNE)).addMBB(TrueMBB);
956 FastEmitBranch(FalseMBB);
957 MBB->addSuccessor(TrueMBB);
961 bool X86FastISel::X86SelectShift(Instruction *I) {
962 unsigned CReg = 0, OpReg = 0, OpImm = 0;
963 const TargetRegisterClass *RC = NULL;
964 if (I->getType() == Type::Int8Ty) {
966 RC = &X86::GR8RegClass;
967 switch (I->getOpcode()) {
968 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
969 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
970 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
971 default: return false;
973 } else if (I->getType() == Type::Int16Ty) {
975 RC = &X86::GR16RegClass;
976 switch (I->getOpcode()) {
977 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
978 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
979 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
980 default: return false;
982 } else if (I->getType() == Type::Int32Ty) {
984 RC = &X86::GR32RegClass;
985 switch (I->getOpcode()) {
986 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
987 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
988 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
989 default: return false;
991 } else if (I->getType() == Type::Int64Ty) {
993 RC = &X86::GR64RegClass;
994 switch (I->getOpcode()) {
995 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
996 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
997 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
998 default: return false;
1004 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1005 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1008 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1009 if (Op0Reg == 0) return false;
1011 // Fold immediate in shl(x,3).
1012 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
1013 unsigned ResultReg = createResultReg(RC);
1014 BuildMI(MBB, DL, TII.get(OpImm),
1015 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
1016 UpdateValueMap(I, ResultReg);
1020 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1021 if (Op1Reg == 0) return false;
1022 TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);
1024 // The shift instruction uses X86::CL. If we defined a super-register
1025 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
1026 // we're doing here.
1027 if (CReg != X86::CL)
1028 BuildMI(MBB, DL, TII.get(TargetInstrInfo::EXTRACT_SUBREG), X86::CL)
1029 .addReg(CReg).addImm(X86::SUBREG_8BIT);
1031 unsigned ResultReg = createResultReg(RC);
1032 BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
1033 UpdateValueMap(I, ResultReg);
1037 bool X86FastISel::X86SelectSelect(Instruction *I) {
1038 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1039 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1043 const TargetRegisterClass *RC = NULL;
1044 if (VT.getSimpleVT() == MVT::i16) {
1045 Opc = X86::CMOVE16rr;
1046 RC = &X86::GR16RegClass;
1047 } else if (VT.getSimpleVT() == MVT::i32) {
1048 Opc = X86::CMOVE32rr;
1049 RC = &X86::GR32RegClass;
1050 } else if (VT.getSimpleVT() == MVT::i64) {
1051 Opc = X86::CMOVE64rr;
1052 RC = &X86::GR64RegClass;
1057 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1058 if (Op0Reg == 0) return false;
1059 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1060 if (Op1Reg == 0) return false;
1061 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1062 if (Op2Reg == 0) return false;
1064 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
1065 unsigned ResultReg = createResultReg(RC);
1066 BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
1067 UpdateValueMap(I, ResultReg);
1071 bool X86FastISel::X86SelectFPExt(Instruction *I) {
1072 // fpext from float to double.
1073 if (Subtarget->hasSSE2() && I->getType() == Type::DoubleTy) {
1074 Value *V = I->getOperand(0);
1075 if (V->getType() == Type::FloatTy) {
1076 unsigned OpReg = getRegForValue(V);
1077 if (OpReg == 0) return false;
1078 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1079 BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
1080 UpdateValueMap(I, ResultReg);
1088 bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
1089 if (Subtarget->hasSSE2()) {
1090 if (I->getType() == Type::FloatTy) {
1091 Value *V = I->getOperand(0);
1092 if (V->getType() == Type::DoubleTy) {
1093 unsigned OpReg = getRegForValue(V);
1094 if (OpReg == 0) return false;
1095 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1096 BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
1097 UpdateValueMap(I, ResultReg);
1106 bool X86FastISel::X86SelectTrunc(Instruction *I) {
1107 if (Subtarget->is64Bit())
1108 // All other cases should be handled by the tblgen generated code.
1110 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1111 MVT DstVT = TLI.getValueType(I->getType());
1113 // This code only handles truncation to byte right now.
1114 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1115 // All other cases should be handled by the tblgen generated code.
1117 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1118 // All other cases should be handled by the tblgen generated code.
1121 unsigned InputReg = getRegForValue(I->getOperand(0));
1123 // Unhandled operand. Halt "fast" selection and bail.
1126 // First issue a copy to GR16_ABCD or GR32_ABCD.
1127 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
1128 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1129 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1130 unsigned CopyReg = createResultReg(CopyRC);
1131 BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
1133 // Then issue an extract_subreg.
1134 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1135 CopyReg, X86::SUBREG_8BIT);
1139 UpdateValueMap(I, ResultReg);
1143 bool X86FastISel::X86SelectExtractValue(Instruction *I) {
1144 ExtractValueInst *EI = cast<ExtractValueInst>(I);
1145 Value *Agg = EI->getAggregateOperand();
1147 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1148 switch (CI->getIntrinsicID()) {
1150 case Intrinsic::sadd_with_overflow:
1151 case Intrinsic::uadd_with_overflow:
1152 // Cheat a little. We know that the registers for "add" and "seto" are
1153 // allocated sequentially. However, we only keep track of the register
1154 // for "add" in the value map. Use extractvalue's index to get the
1155 // correct register for "seto".
1156 UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
1164 bool X86FastISel::X86VisitIntrinsicCall(IntrinsicInst &I) {
1165 // FIXME: Handle more intrinsics.
1166 switch (I.getIntrinsicID()) {
1167 default: return false;
1168 case Intrinsic::sadd_with_overflow:
1169 case Intrinsic::uadd_with_overflow: {
1170 // Replace "add with overflow" intrinsics with an "add" instruction followed
1171 // by a seto/setc instruction. Later on, when the "extractvalue"
1172 // instructions are encountered, we use the fact that two registers were
1173 // created sequentially to get the correct registers for the "sum" and the
1175 const Function *Callee = I.getCalledFunction();
1177 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1180 if (!isTypeLegal(RetTy, VT))
1183 Value *Op1 = I.getOperand(1);
1184 Value *Op2 = I.getOperand(2);
1185 unsigned Reg1 = getRegForValue(Op1);
1186 unsigned Reg2 = getRegForValue(Op2);
1188 if (Reg1 == 0 || Reg2 == 0)
1189 // FIXME: Handle values *not* in registers.
1195 else if (VT == MVT::i64)
1200 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1201 BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
1202 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1204 // If the add with overflow is an intra-block value then we just want to
1205 // create temporaries for it like normal. If it is a cross-block value then
1206 // UpdateValueMap will return the cross-block register used. Since we
1207 // *really* want the value to be live in the register pair known by
1208 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1209 // the cross block case. In the non-cross-block case, we should just make
1210 // another register for the value.
1211 if (DestReg1 != ResultReg)
1212 ResultReg = DestReg1+1;
1214 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1216 unsigned Opc = X86::SETBr;
1217 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1219 BuildMI(MBB, DL, TII.get(Opc), ResultReg);
1225 bool X86FastISel::X86SelectCall(Instruction *I) {
1226 CallInst *CI = cast<CallInst>(I);
1227 Value *Callee = I->getOperand(0);
1229 // Can't handle inline asm yet.
1230 if (isa<InlineAsm>(Callee))
1233 // Handle intrinsic calls.
1234 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1235 return X86VisitIntrinsicCall(*II);
1237 // Handle only C and fastcc calling conventions for now.
1239 unsigned CC = CS.getCallingConv();
1240 if (CC != CallingConv::C &&
1241 CC != CallingConv::Fast &&
1242 CC != CallingConv::X86_FastCall)
1245 // On X86, -tailcallopt changes the fastcc ABI. FastISel doesn't
1246 // handle this for now.
1247 if (CC == CallingConv::Fast && PerformTailCallOpt)
1250 // Let SDISel handle vararg functions.
1251 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1252 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1253 if (FTy->isVarArg())
1256 // Handle *simple* calls for now.
1257 const Type *RetTy = CS.getType();
1259 if (RetTy == Type::VoidTy)
1260 RetVT = MVT::isVoid;
1261 else if (!isTypeLegal(RetTy, RetVT, true))
1264 // Materialize callee address in a register. FIXME: GV address can be
1265 // handled with a CALLpcrel32 instead.
1266 X86AddressMode CalleeAM;
1267 if (!X86SelectCallAddress(Callee, CalleeAM))
1269 unsigned CalleeOp = 0;
1270 GlobalValue *GV = 0;
1271 if (CalleeAM.GV != 0) {
1273 } else if (CalleeAM.Base.Reg != 0) {
1274 CalleeOp = CalleeAM.Base.Reg;
1278 // Allow calls which produce i1 results.
1279 bool AndToI1 = false;
1280 if (RetVT == MVT::i1) {
1285 // Deal with call operands first.
1286 SmallVector<Value*, 8> ArgVals;
1287 SmallVector<unsigned, 8> Args;
1288 SmallVector<MVT, 8> ArgVTs;
1289 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1290 Args.reserve(CS.arg_size());
1291 ArgVals.reserve(CS.arg_size());
1292 ArgVTs.reserve(CS.arg_size());
1293 ArgFlags.reserve(CS.arg_size());
1294 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1296 unsigned Arg = getRegForValue(*i);
1299 ISD::ArgFlagsTy Flags;
1300 unsigned AttrInd = i - CS.arg_begin() + 1;
1301 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1303 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1306 // FIXME: Only handle *easy* calls for now.
1307 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1308 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1309 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1310 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1313 const Type *ArgTy = (*i)->getType();
1315 if (!isTypeLegal(ArgTy, ArgVT))
1317 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1318 Flags.setOrigAlign(OriginalAlignment);
1320 Args.push_back(Arg);
1321 ArgVals.push_back(*i);
1322 ArgVTs.push_back(ArgVT);
1323 ArgFlags.push_back(Flags);
1326 // Analyze operands of the call, assigning locations to each operand.
1327 SmallVector<CCValAssign, 16> ArgLocs;
1328 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1329 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1331 // Get a count of how many bytes are to be pushed on the stack.
1332 unsigned NumBytes = CCInfo.getNextStackOffset();
1334 // Issue CALLSEQ_START
1335 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1336 BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
1338 // Process argument: walk the register/memloc assignments, inserting
1340 SmallVector<unsigned, 4> RegArgs;
1341 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1342 CCValAssign &VA = ArgLocs[i];
1343 unsigned Arg = Args[VA.getValNo()];
1344 MVT ArgVT = ArgVTs[VA.getValNo()];
1346 // Promote the value if needed.
1347 switch (VA.getLocInfo()) {
1348 default: assert(0 && "Unknown loc info!");
1349 case CCValAssign::Full: break;
1350 case CCValAssign::SExt: {
1351 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1353 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1355 ArgVT = VA.getLocVT();
1358 case CCValAssign::ZExt: {
1359 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1361 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1363 ArgVT = VA.getLocVT();
1366 case CCValAssign::AExt: {
1367 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1370 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1373 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1376 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1377 ArgVT = VA.getLocVT();
1382 if (VA.isRegLoc()) {
1383 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1384 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
1386 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1388 RegArgs.push_back(VA.getLocReg());
1390 unsigned LocMemOffset = VA.getLocMemOffset();
1392 AM.Base.Reg = StackPtr;
1393 AM.Disp = LocMemOffset;
1394 Value *ArgVal = ArgVals[VA.getValNo()];
1396 // If this is a really simple value, emit this with the Value* version of
1397 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1398 // can cause us to reevaluate the argument.
1399 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1400 X86FastEmitStore(ArgVT, ArgVal, AM);
1402 X86FastEmitStore(ArgVT, Arg, AM);
1406 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1408 if (Subtarget->isPICStyleGOT()) {
1409 TargetRegisterClass *RC = X86::GR32RegisterClass;
1410 unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
1411 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
1412 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1417 MachineInstrBuilder MIB;
1419 // Register-indirect call.
1420 unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
1421 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);
1425 assert(GV && "Not a direct call");
1427 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
1429 // See if we need any target-specific flags on the GV operand.
1430 unsigned char OpFlags = 0;
1432 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1433 // external symbols most go through the PLT in PIC mode. If the symbol
1434 // has hidden or protected visibility, or if it is static or local, then
1435 // we don't need to use the PLT - we can directly call it.
1436 if (Subtarget->isTargetELF() &&
1437 TM.getRelocationModel() == Reloc::PIC_ &&
1438 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1439 OpFlags = X86II::MO_PLT;
1440 } else if (Subtarget->isPICStyleStub() &&
1441 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1442 Subtarget->getDarwinVers() < 9) {
1443 // PC-relative references to external symbols should go through $stub,
1444 // unless we're building with the leopard linker or later, which
1445 // automatically synthesizes these stubs.
1446 OpFlags = X86II::MO_DARWIN_STUB;
1450 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
1453 // Add an implicit use GOT pointer in EBX.
1454 if (Subtarget->isPICStyleGOT())
1455 MIB.addReg(X86::EBX);
1457 // Add implicit physical register uses to the call.
1458 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1459 MIB.addReg(RegArgs[i]);
1461 // Issue CALLSEQ_END
1462 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1463 BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
1465 // Now handle call return value (if any).
1466 if (RetVT.getSimpleVT() != MVT::isVoid) {
1467 SmallVector<CCValAssign, 16> RVLocs;
1468 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1469 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1471 // Copy all of the result registers out of their specified physreg.
1472 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1473 MVT CopyVT = RVLocs[0].getValVT();
1474 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1475 TargetRegisterClass *SrcRC = DstRC;
1477 // If this is a call to a function that returns an fp value on the x87 fp
1478 // stack, but where we prefer to use the value in xmm registers, copy it
1479 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1480 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1481 RVLocs[0].getLocReg() == X86::ST1) &&
1482 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1484 SrcRC = X86::RSTRegisterClass;
1485 DstRC = X86::RFP80RegisterClass;
1488 unsigned ResultReg = createResultReg(DstRC);
1489 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1490 RVLocs[0].getLocReg(), DstRC, SrcRC);
1491 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1493 if (CopyVT != RVLocs[0].getValVT()) {
1494 // Round the F80 the right size, which also moves to the appropriate xmm
1495 // register. This is accomplished by storing the F80 value in memory and
1496 // then loading it back. Ewww...
1497 MVT ResVT = RVLocs[0].getValVT();
1498 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1499 unsigned MemSize = ResVT.getSizeInBits()/8;
1500 int FI = MFI.CreateStackObject(MemSize, MemSize);
1501 addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
1502 DstRC = ResVT == MVT::f32
1503 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1504 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1505 ResultReg = createResultReg(DstRC);
1506 addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
1510 // Mask out all but lowest bit for some call which produces an i1.
1511 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1513 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1514 ResultReg = AndResult;
1517 UpdateValueMap(I, ResultReg);
1525 X86FastISel::TargetSelectInstruction(Instruction *I) {
1526 switch (I->getOpcode()) {
1528 case Instruction::Load:
1529 return X86SelectLoad(I);
1530 case Instruction::Store:
1531 return X86SelectStore(I);
1532 case Instruction::ICmp:
1533 case Instruction::FCmp:
1534 return X86SelectCmp(I);
1535 case Instruction::ZExt:
1536 return X86SelectZExt(I);
1537 case Instruction::Br:
1538 return X86SelectBranch(I);
1539 case Instruction::Call:
1540 return X86SelectCall(I);
1541 case Instruction::LShr:
1542 case Instruction::AShr:
1543 case Instruction::Shl:
1544 return X86SelectShift(I);
1545 case Instruction::Select:
1546 return X86SelectSelect(I);
1547 case Instruction::Trunc:
1548 return X86SelectTrunc(I);
1549 case Instruction::FPExt:
1550 return X86SelectFPExt(I);
1551 case Instruction::FPTrunc:
1552 return X86SelectFPTrunc(I);
1553 case Instruction::ExtractValue:
1554 return X86SelectExtractValue(I);
1555 case Instruction::IntToPtr: // Deliberate fall-through.
1556 case Instruction::PtrToInt: {
1557 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1558 MVT DstVT = TLI.getValueType(I->getType());
1559 if (DstVT.bitsGT(SrcVT))
1560 return X86SelectZExt(I);
1561 if (DstVT.bitsLT(SrcVT))
1562 return X86SelectTrunc(I);
1563 unsigned Reg = getRegForValue(I->getOperand(0));
1564 if (Reg == 0) return false;
1565 UpdateValueMap(I, Reg);
1573 unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
1575 if (!isTypeLegal(C->getType(), VT))
1578 // Get opcode and regclass of the output for the given load instruction.
1580 const TargetRegisterClass *RC = NULL;
1581 switch (VT.getSimpleVT()) {
1582 default: return false;
1585 RC = X86::GR8RegisterClass;
1589 RC = X86::GR16RegisterClass;
1593 RC = X86::GR32RegisterClass;
1596 // Must be in x86-64 mode.
1598 RC = X86::GR64RegisterClass;
1601 if (Subtarget->hasSSE1()) {
1603 RC = X86::FR32RegisterClass;
1605 Opc = X86::LD_Fp32m;
1606 RC = X86::RFP32RegisterClass;
1610 if (Subtarget->hasSSE2()) {
1612 RC = X86::FR64RegisterClass;
1614 Opc = X86::LD_Fp64m;
1615 RC = X86::RFP64RegisterClass;
1619 // No f80 support yet.
1623 // Materialize addresses with LEA instructions.
1624 if (isa<GlobalValue>(C)) {
1626 if (X86SelectAddress(C, AM)) {
1627 if (TLI.getPointerTy() == MVT::i32)
1631 unsigned ResultReg = createResultReg(RC);
1632 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1638 // MachineConstantPool wants an explicit alignment.
1639 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1641 // Alignment of vector types. FIXME!
1642 Align = TD.getTypeAllocSize(C->getType());
1645 // x86-32 PIC requires a PIC base register for constant pools.
1646 unsigned PICBase = 0;
1647 unsigned char OpFlag = 0;
1648 if (Subtarget->isPICStyleStub() &&
1649 TM.getRelocationModel() == Reloc::PIC_) { // Not dynamic-no-pic
1650 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1651 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1652 } else if (Subtarget->isPICStyleGOT()) {
1653 OpFlag = X86II::MO_GOTOFF;
1654 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1655 } else if (Subtarget->isPICStyleRIPRel() &&
1656 TM.getCodeModel() == CodeModel::Small) {
1660 // Create the load from the constant pool.
1661 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1662 unsigned ResultReg = createResultReg(RC);
1663 addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
1664 MCPOffset, PICBase, OpFlag);
1669 unsigned X86FastISel::TargetMaterializeAlloca(AllocaInst *C) {
1670 // Fail on dynamic allocas. At this point, getRegForValue has already
1671 // checked its CSE maps, so if we're here trying to handle a dynamic
1672 // alloca, we're not going to succeed. X86SelectAddress has a
1673 // check for dynamic allocas, because it's called directly from
1674 // various places, but TargetMaterializeAlloca also needs a check
1675 // in order to avoid recursion between getRegForValue,
1676 // X86SelectAddrss, and TargetMaterializeAlloca.
1677 if (!StaticAllocaMap.count(C))
1681 if (!X86SelectAddress(C, AM))
1683 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1684 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1685 unsigned ResultReg = createResultReg(RC);
1686 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1691 llvm::FastISel *X86::createFastISel(MachineFunction &mf,
1692 MachineModuleInfo *mmi,
1694 DenseMap<const Value *, unsigned> &vm,
1695 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
1696 DenseMap<const AllocaInst *, int> &am
1698 , SmallSet<Instruction*, 8> &cil
1701 return new X86FastISel(mf, mmi, dw, vm, bm, am