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 // Allow the subtarget to classify the global.
445 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
447 // If this reference is relative to the pic base, set it now.
448 if (isGlobalRelativeToPICBase(GVFlags)) {
449 // FIXME: How do we know Base.Reg is free??
450 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(&MF);
453 // Unless the ABI requires an extra load, return a direct reference to
455 if (!isGlobalStubReference(GVFlags)) {
456 if (Subtarget->isPICStyleRIPRel()) {
457 // Use rip-relative addressing if we can. Above we verified that the
458 // base and index registers are unused.
459 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
460 AM.Base.Reg = X86::RIP;
462 AM.GVOpFlags = GVFlags;
466 // Ok, we need to do a load from a stub. If we've already loaded from this
467 // stub, reuse the loaded pointer, otherwise emit the load now.
468 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
470 if (I != LocalValueMap.end() && I->second != 0) {
473 // Issue load from stub.
475 const TargetRegisterClass *RC = NULL;
476 X86AddressMode StubAM;
477 StubAM.Base.Reg = AM.Base.Reg;
479 StubAM.GVOpFlags = GVFlags;
481 if (TLI.getPointerTy() == MVT::i64) {
483 RC = X86::GR64RegisterClass;
485 if (Subtarget->isPICStyleRIPRel())
486 StubAM.Base.Reg = X86::RIP;
489 RC = X86::GR32RegisterClass;
492 LoadReg = createResultReg(RC);
493 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);
495 // Prevent loading GV stub multiple times in same MBB.
496 LocalValueMap[V] = LoadReg;
499 // Now construct the final address. Note that the Disp, Scale,
500 // and Index values may already be set here.
501 AM.Base.Reg = LoadReg;
506 // If all else fails, try to materialize the value in a register.
507 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
508 if (AM.Base.Reg == 0) {
509 AM.Base.Reg = getRegForValue(V);
510 return AM.Base.Reg != 0;
512 if (AM.IndexReg == 0) {
513 assert(AM.Scale == 1 && "Scale with no index!");
514 AM.IndexReg = getRegForValue(V);
515 return AM.IndexReg != 0;
522 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
524 bool X86FastISel::X86SelectCallAddress(Value *V, X86AddressMode &AM) {
526 unsigned Opcode = Instruction::UserOp1;
527 if (Instruction *I = dyn_cast<Instruction>(V)) {
528 Opcode = I->getOpcode();
530 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
531 Opcode = C->getOpcode();
537 case Instruction::BitCast:
538 // Look past bitcasts.
539 return X86SelectCallAddress(U->getOperand(0), AM);
541 case Instruction::IntToPtr:
542 // Look past no-op inttoptrs.
543 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
544 return X86SelectCallAddress(U->getOperand(0), AM);
547 case Instruction::PtrToInt:
548 // Look past no-op ptrtoints.
549 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
550 return X86SelectCallAddress(U->getOperand(0), AM);
554 // Handle constant address.
555 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
556 // Can't handle alternate code models yet.
557 if (TM.getCodeModel() != CodeModel::Default &&
558 TM.getCodeModel() != CodeModel::Small)
561 // RIP-relative addresses can't have additional register operands.
562 if (Subtarget->isPICStyleRIPRel() &&
563 (AM.Base.Reg != 0 || AM.IndexReg != 0))
566 // Can't handle TLS or DLLImport.
567 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
568 if (GVar->isThreadLocal() || GVar->hasDLLImportLinkage())
571 // Okay, we've committed to selecting this global. Set up the basic address.
574 // No ABI requires an extra load for anything other than DLLImport, which
575 // we rejected above. Return a direct reference to the global.
576 if (Subtarget->isPICStyleRIPRel()) {
577 // Use rip-relative addressing if we can. Above we verified that the
578 // base and index registers are unused.
579 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
580 AM.Base.Reg = X86::RIP;
581 } else if (Subtarget->isPICStyleStubPIC()) {
582 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
583 } else if (Subtarget->isPICStyleGOT()) {
584 AM.GVOpFlags = X86II::MO_GOTOFF;
590 // If all else fails, try to materialize the value in a register.
591 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
592 if (AM.Base.Reg == 0) {
593 AM.Base.Reg = getRegForValue(V);
594 return AM.Base.Reg != 0;
596 if (AM.IndexReg == 0) {
597 assert(AM.Scale == 1 && "Scale with no index!");
598 AM.IndexReg = getRegForValue(V);
599 return AM.IndexReg != 0;
607 /// X86SelectStore - Select and emit code to implement store instructions.
608 bool X86FastISel::X86SelectStore(Instruction* I) {
610 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
614 if (!X86SelectAddress(I->getOperand(1), AM))
617 return X86FastEmitStore(VT, I->getOperand(0), AM);
620 /// X86SelectLoad - Select and emit code to implement load instructions.
622 bool X86FastISel::X86SelectLoad(Instruction *I) {
624 if (!isTypeLegal(I->getType(), VT))
628 if (!X86SelectAddress(I->getOperand(0), AM))
631 unsigned ResultReg = 0;
632 if (X86FastEmitLoad(VT, AM, ResultReg)) {
633 UpdateValueMap(I, ResultReg);
639 static unsigned X86ChooseCmpOpcode(MVT VT) {
640 switch (VT.getSimpleVT()) {
642 case MVT::i8: return X86::CMP8rr;
643 case MVT::i16: return X86::CMP16rr;
644 case MVT::i32: return X86::CMP32rr;
645 case MVT::i64: return X86::CMP64rr;
646 case MVT::f32: return X86::UCOMISSrr;
647 case MVT::f64: return X86::UCOMISDrr;
651 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
652 /// of the comparison, return an opcode that works for the compare (e.g.
653 /// CMP32ri) otherwise return 0.
654 static unsigned X86ChooseCmpImmediateOpcode(MVT VT, ConstantInt *RHSC) {
655 switch (VT.getSimpleVT()) {
656 // Otherwise, we can't fold the immediate into this comparison.
658 case MVT::i8: return X86::CMP8ri;
659 case MVT::i16: return X86::CMP16ri;
660 case MVT::i32: return X86::CMP32ri;
662 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
664 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
665 return X86::CMP64ri32;
670 bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, MVT VT) {
671 unsigned Op0Reg = getRegForValue(Op0);
672 if (Op0Reg == 0) return false;
674 // Handle 'null' like i32/i64 0.
675 if (isa<ConstantPointerNull>(Op1))
676 Op1 = Constant::getNullValue(TD.getIntPtrType());
678 // We have two options: compare with register or immediate. If the RHS of
679 // the compare is an immediate that we can fold into this compare, use
680 // CMPri, otherwise use CMPrr.
681 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
682 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
683 BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
684 .addImm(Op1C->getSExtValue());
689 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
690 if (CompareOpc == 0) return false;
692 unsigned Op1Reg = getRegForValue(Op1);
693 if (Op1Reg == 0) return false;
694 BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
699 bool X86FastISel::X86SelectCmp(Instruction *I) {
700 CmpInst *CI = cast<CmpInst>(I);
703 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
706 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
708 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
709 switch (CI->getPredicate()) {
710 case CmpInst::FCMP_OEQ: {
711 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
714 unsigned EReg = createResultReg(&X86::GR8RegClass);
715 unsigned NPReg = createResultReg(&X86::GR8RegClass);
716 BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
717 BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
719 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
720 UpdateValueMap(I, ResultReg);
723 case CmpInst::FCMP_UNE: {
724 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
727 unsigned NEReg = createResultReg(&X86::GR8RegClass);
728 unsigned PReg = createResultReg(&X86::GR8RegClass);
729 BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
730 BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
731 BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
732 UpdateValueMap(I, ResultReg);
735 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
736 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
737 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
738 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
739 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
740 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
741 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
742 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
743 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
744 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
745 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
746 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
748 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
749 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
750 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
751 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
752 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
753 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
754 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
755 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
756 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
757 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
762 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
766 // Emit a compare of Op0/Op1.
767 if (!X86FastEmitCompare(Op0, Op1, VT))
770 BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
771 UpdateValueMap(I, ResultReg);
775 bool X86FastISel::X86SelectZExt(Instruction *I) {
776 // Handle zero-extension from i1 to i8, which is common.
777 if (I->getType() == Type::Int8Ty &&
778 I->getOperand(0)->getType() == Type::Int1Ty) {
779 unsigned ResultReg = getRegForValue(I->getOperand(0));
780 if (ResultReg == 0) return false;
781 // Set the high bits to zero.
782 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg);
783 if (ResultReg == 0) return false;
784 UpdateValueMap(I, ResultReg);
792 bool X86FastISel::X86SelectBranch(Instruction *I) {
793 // Unconditional branches are selected by tablegen-generated code.
794 // Handle a conditional branch.
795 BranchInst *BI = cast<BranchInst>(I);
796 MachineBasicBlock *TrueMBB = MBBMap[BI->getSuccessor(0)];
797 MachineBasicBlock *FalseMBB = MBBMap[BI->getSuccessor(1)];
799 // Fold the common case of a conditional branch with a comparison.
800 if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
801 if (CI->hasOneUse()) {
802 MVT VT = TLI.getValueType(CI->getOperand(0)->getType());
804 // Try to take advantage of fallthrough opportunities.
805 CmpInst::Predicate Predicate = CI->getPredicate();
806 if (MBB->isLayoutSuccessor(TrueMBB)) {
807 std::swap(TrueMBB, FalseMBB);
808 Predicate = CmpInst::getInversePredicate(Predicate);
811 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
812 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
815 case CmpInst::FCMP_OEQ:
816 std::swap(TrueMBB, FalseMBB);
817 Predicate = CmpInst::FCMP_UNE;
819 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE; break;
820 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA; break;
821 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE; break;
822 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA; break;
823 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE; break;
824 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE; break;
825 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP; break;
826 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP; break;
827 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE; break;
828 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB; break;
829 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE; break;
830 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
831 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
833 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE; break;
834 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE; break;
835 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA; break;
836 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE; break;
837 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
838 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
839 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG; break;
840 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE; break;
841 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL; break;
842 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE; break;
847 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
851 // Emit a compare of the LHS and RHS, setting the flags.
852 if (!X86FastEmitCompare(Op0, Op1, VT))
855 BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
857 if (Predicate == CmpInst::FCMP_UNE) {
858 // X86 requires a second branch to handle UNE (and OEQ,
859 // which is mapped to UNE above).
860 BuildMI(MBB, DL, TII.get(X86::JP)).addMBB(TrueMBB);
863 FastEmitBranch(FalseMBB);
864 MBB->addSuccessor(TrueMBB);
867 } else if (ExtractValueInst *EI =
868 dyn_cast<ExtractValueInst>(BI->getCondition())) {
869 // Check to see if the branch instruction is from an "arithmetic with
870 // overflow" intrinsic. The main way these intrinsics are used is:
872 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
873 // %sum = extractvalue { i32, i1 } %t, 0
874 // %obit = extractvalue { i32, i1 } %t, 1
875 // br i1 %obit, label %overflow, label %normal
877 // The %sum and %obit are converted in an ADD and a SETO/SETB before
878 // reaching the branch. Therefore, we search backwards through the MBB
879 // looking for the SETO/SETB instruction. If an instruction modifies the
880 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
881 // convert the branch into a JO/JB instruction.
882 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
883 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
884 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
885 const MachineInstr *SetMI = 0;
886 unsigned Reg = lookUpRegForValue(EI);
888 for (MachineBasicBlock::const_reverse_iterator
889 RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
890 const MachineInstr &MI = *RI;
892 if (MI.modifiesRegister(Reg)) {
893 unsigned Src, Dst, SrcSR, DstSR;
895 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
904 const TargetInstrDesc &TID = MI.getDesc();
905 if (TID.hasUnmodeledSideEffects() ||
906 TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
911 unsigned OpCode = SetMI->getOpcode();
913 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
914 BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ? X86::JO : X86::JB))
916 FastEmitBranch(FalseMBB);
917 MBB->addSuccessor(TrueMBB);
925 // Otherwise do a clumsy setcc and re-test it.
926 unsigned OpReg = getRegForValue(BI->getCondition());
927 if (OpReg == 0) return false;
929 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
930 BuildMI(MBB, DL, TII.get(X86::JNE)).addMBB(TrueMBB);
931 FastEmitBranch(FalseMBB);
932 MBB->addSuccessor(TrueMBB);
936 bool X86FastISel::X86SelectShift(Instruction *I) {
937 unsigned CReg = 0, OpReg = 0, OpImm = 0;
938 const TargetRegisterClass *RC = NULL;
939 if (I->getType() == Type::Int8Ty) {
941 RC = &X86::GR8RegClass;
942 switch (I->getOpcode()) {
943 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
944 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
945 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
946 default: return false;
948 } else if (I->getType() == Type::Int16Ty) {
950 RC = &X86::GR16RegClass;
951 switch (I->getOpcode()) {
952 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
953 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
954 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
955 default: return false;
957 } else if (I->getType() == Type::Int32Ty) {
959 RC = &X86::GR32RegClass;
960 switch (I->getOpcode()) {
961 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
962 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
963 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
964 default: return false;
966 } else if (I->getType() == Type::Int64Ty) {
968 RC = &X86::GR64RegClass;
969 switch (I->getOpcode()) {
970 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
971 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
972 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
973 default: return false;
979 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
980 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
983 unsigned Op0Reg = getRegForValue(I->getOperand(0));
984 if (Op0Reg == 0) return false;
986 // Fold immediate in shl(x,3).
987 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
988 unsigned ResultReg = createResultReg(RC);
989 BuildMI(MBB, DL, TII.get(OpImm),
990 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
991 UpdateValueMap(I, ResultReg);
995 unsigned Op1Reg = getRegForValue(I->getOperand(1));
996 if (Op1Reg == 0) return false;
997 TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);
999 // The shift instruction uses X86::CL. If we defined a super-register
1000 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
1001 // we're doing here.
1002 if (CReg != X86::CL)
1003 BuildMI(MBB, DL, TII.get(TargetInstrInfo::EXTRACT_SUBREG), X86::CL)
1004 .addReg(CReg).addImm(X86::SUBREG_8BIT);
1006 unsigned ResultReg = createResultReg(RC);
1007 BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
1008 UpdateValueMap(I, ResultReg);
1012 bool X86FastISel::X86SelectSelect(Instruction *I) {
1013 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1014 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1018 const TargetRegisterClass *RC = NULL;
1019 if (VT.getSimpleVT() == MVT::i16) {
1020 Opc = X86::CMOVE16rr;
1021 RC = &X86::GR16RegClass;
1022 } else if (VT.getSimpleVT() == MVT::i32) {
1023 Opc = X86::CMOVE32rr;
1024 RC = &X86::GR32RegClass;
1025 } else if (VT.getSimpleVT() == MVT::i64) {
1026 Opc = X86::CMOVE64rr;
1027 RC = &X86::GR64RegClass;
1032 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1033 if (Op0Reg == 0) return false;
1034 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1035 if (Op1Reg == 0) return false;
1036 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1037 if (Op2Reg == 0) return false;
1039 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
1040 unsigned ResultReg = createResultReg(RC);
1041 BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
1042 UpdateValueMap(I, ResultReg);
1046 bool X86FastISel::X86SelectFPExt(Instruction *I) {
1047 // fpext from float to double.
1048 if (Subtarget->hasSSE2() && I->getType() == Type::DoubleTy) {
1049 Value *V = I->getOperand(0);
1050 if (V->getType() == Type::FloatTy) {
1051 unsigned OpReg = getRegForValue(V);
1052 if (OpReg == 0) return false;
1053 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1054 BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
1055 UpdateValueMap(I, ResultReg);
1063 bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
1064 if (Subtarget->hasSSE2()) {
1065 if (I->getType() == Type::FloatTy) {
1066 Value *V = I->getOperand(0);
1067 if (V->getType() == Type::DoubleTy) {
1068 unsigned OpReg = getRegForValue(V);
1069 if (OpReg == 0) return false;
1070 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1071 BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
1072 UpdateValueMap(I, ResultReg);
1081 bool X86FastISel::X86SelectTrunc(Instruction *I) {
1082 if (Subtarget->is64Bit())
1083 // All other cases should be handled by the tblgen generated code.
1085 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1086 MVT DstVT = TLI.getValueType(I->getType());
1088 // This code only handles truncation to byte right now.
1089 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1090 // All other cases should be handled by the tblgen generated code.
1092 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1093 // All other cases should be handled by the tblgen generated code.
1096 unsigned InputReg = getRegForValue(I->getOperand(0));
1098 // Unhandled operand. Halt "fast" selection and bail.
1101 // First issue a copy to GR16_ABCD or GR32_ABCD.
1102 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
1103 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1104 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1105 unsigned CopyReg = createResultReg(CopyRC);
1106 BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
1108 // Then issue an extract_subreg.
1109 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1110 CopyReg, X86::SUBREG_8BIT);
1114 UpdateValueMap(I, ResultReg);
1118 bool X86FastISel::X86SelectExtractValue(Instruction *I) {
1119 ExtractValueInst *EI = cast<ExtractValueInst>(I);
1120 Value *Agg = EI->getAggregateOperand();
1122 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1123 switch (CI->getIntrinsicID()) {
1125 case Intrinsic::sadd_with_overflow:
1126 case Intrinsic::uadd_with_overflow:
1127 // Cheat a little. We know that the registers for "add" and "seto" are
1128 // allocated sequentially. However, we only keep track of the register
1129 // for "add" in the value map. Use extractvalue's index to get the
1130 // correct register for "seto".
1131 UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
1139 bool X86FastISel::X86VisitIntrinsicCall(IntrinsicInst &I) {
1140 // FIXME: Handle more intrinsics.
1141 switch (I.getIntrinsicID()) {
1142 default: return false;
1143 case Intrinsic::sadd_with_overflow:
1144 case Intrinsic::uadd_with_overflow: {
1145 // Replace "add with overflow" intrinsics with an "add" instruction followed
1146 // by a seto/setc instruction. Later on, when the "extractvalue"
1147 // instructions are encountered, we use the fact that two registers were
1148 // created sequentially to get the correct registers for the "sum" and the
1150 const Function *Callee = I.getCalledFunction();
1152 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1155 if (!isTypeLegal(RetTy, VT))
1158 Value *Op1 = I.getOperand(1);
1159 Value *Op2 = I.getOperand(2);
1160 unsigned Reg1 = getRegForValue(Op1);
1161 unsigned Reg2 = getRegForValue(Op2);
1163 if (Reg1 == 0 || Reg2 == 0)
1164 // FIXME: Handle values *not* in registers.
1170 else if (VT == MVT::i64)
1175 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1176 BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
1177 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1179 // If the add with overflow is an intra-block value then we just want to
1180 // create temporaries for it like normal. If it is a cross-block value then
1181 // UpdateValueMap will return the cross-block register used. Since we
1182 // *really* want the value to be live in the register pair known by
1183 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1184 // the cross block case. In the non-cross-block case, we should just make
1185 // another register for the value.
1186 if (DestReg1 != ResultReg)
1187 ResultReg = DestReg1+1;
1189 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1191 unsigned Opc = X86::SETBr;
1192 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1194 BuildMI(MBB, DL, TII.get(Opc), ResultReg);
1200 bool X86FastISel::X86SelectCall(Instruction *I) {
1201 CallInst *CI = cast<CallInst>(I);
1202 Value *Callee = I->getOperand(0);
1204 // Can't handle inline asm yet.
1205 if (isa<InlineAsm>(Callee))
1208 // Handle intrinsic calls.
1209 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1210 return X86VisitIntrinsicCall(*II);
1212 // Handle only C and fastcc calling conventions for now.
1214 unsigned CC = CS.getCallingConv();
1215 if (CC != CallingConv::C &&
1216 CC != CallingConv::Fast &&
1217 CC != CallingConv::X86_FastCall)
1220 // On X86, -tailcallopt changes the fastcc ABI. FastISel doesn't
1221 // handle this for now.
1222 if (CC == CallingConv::Fast && PerformTailCallOpt)
1225 // Let SDISel handle vararg functions.
1226 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1227 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1228 if (FTy->isVarArg())
1231 // Handle *simple* calls for now.
1232 const Type *RetTy = CS.getType();
1234 if (RetTy == Type::VoidTy)
1235 RetVT = MVT::isVoid;
1236 else if (!isTypeLegal(RetTy, RetVT, true))
1239 // Materialize callee address in a register. FIXME: GV address can be
1240 // handled with a CALLpcrel32 instead.
1241 X86AddressMode CalleeAM;
1242 if (!X86SelectCallAddress(Callee, CalleeAM))
1244 unsigned CalleeOp = 0;
1245 GlobalValue *GV = 0;
1246 if (CalleeAM.GV != 0) {
1248 } else if (CalleeAM.Base.Reg != 0) {
1249 CalleeOp = CalleeAM.Base.Reg;
1253 // Allow calls which produce i1 results.
1254 bool AndToI1 = false;
1255 if (RetVT == MVT::i1) {
1260 // Deal with call operands first.
1261 SmallVector<Value*, 8> ArgVals;
1262 SmallVector<unsigned, 8> Args;
1263 SmallVector<MVT, 8> ArgVTs;
1264 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1265 Args.reserve(CS.arg_size());
1266 ArgVals.reserve(CS.arg_size());
1267 ArgVTs.reserve(CS.arg_size());
1268 ArgFlags.reserve(CS.arg_size());
1269 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1271 unsigned Arg = getRegForValue(*i);
1274 ISD::ArgFlagsTy Flags;
1275 unsigned AttrInd = i - CS.arg_begin() + 1;
1276 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1278 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1281 // FIXME: Only handle *easy* calls for now.
1282 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1283 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1284 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1285 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1288 const Type *ArgTy = (*i)->getType();
1290 if (!isTypeLegal(ArgTy, ArgVT))
1292 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1293 Flags.setOrigAlign(OriginalAlignment);
1295 Args.push_back(Arg);
1296 ArgVals.push_back(*i);
1297 ArgVTs.push_back(ArgVT);
1298 ArgFlags.push_back(Flags);
1301 // Analyze operands of the call, assigning locations to each operand.
1302 SmallVector<CCValAssign, 16> ArgLocs;
1303 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1304 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1306 // Get a count of how many bytes are to be pushed on the stack.
1307 unsigned NumBytes = CCInfo.getNextStackOffset();
1309 // Issue CALLSEQ_START
1310 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1311 BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
1313 // Process argument: walk the register/memloc assignments, inserting
1315 SmallVector<unsigned, 4> RegArgs;
1316 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1317 CCValAssign &VA = ArgLocs[i];
1318 unsigned Arg = Args[VA.getValNo()];
1319 MVT ArgVT = ArgVTs[VA.getValNo()];
1321 // Promote the value if needed.
1322 switch (VA.getLocInfo()) {
1323 default: assert(0 && "Unknown loc info!");
1324 case CCValAssign::Full: break;
1325 case CCValAssign::SExt: {
1326 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1328 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1330 ArgVT = VA.getLocVT();
1333 case CCValAssign::ZExt: {
1334 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1336 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1338 ArgVT = VA.getLocVT();
1341 case CCValAssign::AExt: {
1342 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1345 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1348 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1351 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1352 ArgVT = VA.getLocVT();
1357 if (VA.isRegLoc()) {
1358 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1359 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
1361 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1363 RegArgs.push_back(VA.getLocReg());
1365 unsigned LocMemOffset = VA.getLocMemOffset();
1367 AM.Base.Reg = StackPtr;
1368 AM.Disp = LocMemOffset;
1369 Value *ArgVal = ArgVals[VA.getValNo()];
1371 // If this is a really simple value, emit this with the Value* version of
1372 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1373 // can cause us to reevaluate the argument.
1374 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1375 X86FastEmitStore(ArgVT, ArgVal, AM);
1377 X86FastEmitStore(ArgVT, Arg, AM);
1381 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1383 if (Subtarget->isPICStyleGOT()) {
1384 TargetRegisterClass *RC = X86::GR32RegisterClass;
1385 unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
1386 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
1387 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1392 MachineInstrBuilder MIB;
1394 // Register-indirect call.
1395 unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
1396 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);
1400 assert(GV && "Not a direct call");
1402 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
1404 // See if we need any target-specific flags on the GV operand.
1405 unsigned char OpFlags = 0;
1407 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1408 // external symbols most go through the PLT in PIC mode. If the symbol
1409 // has hidden or protected visibility, or if it is static or local, then
1410 // we don't need to use the PLT - we can directly call it.
1411 if (Subtarget->isTargetELF() &&
1412 TM.getRelocationModel() == Reloc::PIC_ &&
1413 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1414 OpFlags = X86II::MO_PLT;
1415 } else if (Subtarget->isPICStyleStubAny() &&
1416 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1417 Subtarget->getDarwinVers() < 9) {
1418 // PC-relative references to external symbols should go through $stub,
1419 // unless we're building with the leopard linker or later, which
1420 // automatically synthesizes these stubs.
1421 OpFlags = X86II::MO_DARWIN_STUB;
1425 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
1428 // Add an implicit use GOT pointer in EBX.
1429 if (Subtarget->isPICStyleGOT())
1430 MIB.addReg(X86::EBX);
1432 // Add implicit physical register uses to the call.
1433 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1434 MIB.addReg(RegArgs[i]);
1436 // Issue CALLSEQ_END
1437 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1438 BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
1440 // Now handle call return value (if any).
1441 if (RetVT.getSimpleVT() != MVT::isVoid) {
1442 SmallVector<CCValAssign, 16> RVLocs;
1443 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1444 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1446 // Copy all of the result registers out of their specified physreg.
1447 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1448 MVT CopyVT = RVLocs[0].getValVT();
1449 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1450 TargetRegisterClass *SrcRC = DstRC;
1452 // If this is a call to a function that returns an fp value on the x87 fp
1453 // stack, but where we prefer to use the value in xmm registers, copy it
1454 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1455 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1456 RVLocs[0].getLocReg() == X86::ST1) &&
1457 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1459 SrcRC = X86::RSTRegisterClass;
1460 DstRC = X86::RFP80RegisterClass;
1463 unsigned ResultReg = createResultReg(DstRC);
1464 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1465 RVLocs[0].getLocReg(), DstRC, SrcRC);
1466 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1468 if (CopyVT != RVLocs[0].getValVT()) {
1469 // Round the F80 the right size, which also moves to the appropriate xmm
1470 // register. This is accomplished by storing the F80 value in memory and
1471 // then loading it back. Ewww...
1472 MVT ResVT = RVLocs[0].getValVT();
1473 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1474 unsigned MemSize = ResVT.getSizeInBits()/8;
1475 int FI = MFI.CreateStackObject(MemSize, MemSize);
1476 addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
1477 DstRC = ResVT == MVT::f32
1478 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1479 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1480 ResultReg = createResultReg(DstRC);
1481 addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
1485 // Mask out all but lowest bit for some call which produces an i1.
1486 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1488 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1489 ResultReg = AndResult;
1492 UpdateValueMap(I, ResultReg);
1500 X86FastISel::TargetSelectInstruction(Instruction *I) {
1501 switch (I->getOpcode()) {
1503 case Instruction::Load:
1504 return X86SelectLoad(I);
1505 case Instruction::Store:
1506 return X86SelectStore(I);
1507 case Instruction::ICmp:
1508 case Instruction::FCmp:
1509 return X86SelectCmp(I);
1510 case Instruction::ZExt:
1511 return X86SelectZExt(I);
1512 case Instruction::Br:
1513 return X86SelectBranch(I);
1514 case Instruction::Call:
1515 return X86SelectCall(I);
1516 case Instruction::LShr:
1517 case Instruction::AShr:
1518 case Instruction::Shl:
1519 return X86SelectShift(I);
1520 case Instruction::Select:
1521 return X86SelectSelect(I);
1522 case Instruction::Trunc:
1523 return X86SelectTrunc(I);
1524 case Instruction::FPExt:
1525 return X86SelectFPExt(I);
1526 case Instruction::FPTrunc:
1527 return X86SelectFPTrunc(I);
1528 case Instruction::ExtractValue:
1529 return X86SelectExtractValue(I);
1530 case Instruction::IntToPtr: // Deliberate fall-through.
1531 case Instruction::PtrToInt: {
1532 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1533 MVT DstVT = TLI.getValueType(I->getType());
1534 if (DstVT.bitsGT(SrcVT))
1535 return X86SelectZExt(I);
1536 if (DstVT.bitsLT(SrcVT))
1537 return X86SelectTrunc(I);
1538 unsigned Reg = getRegForValue(I->getOperand(0));
1539 if (Reg == 0) return false;
1540 UpdateValueMap(I, Reg);
1548 unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
1550 if (!isTypeLegal(C->getType(), VT))
1553 // Get opcode and regclass of the output for the given load instruction.
1555 const TargetRegisterClass *RC = NULL;
1556 switch (VT.getSimpleVT()) {
1557 default: return false;
1560 RC = X86::GR8RegisterClass;
1564 RC = X86::GR16RegisterClass;
1568 RC = X86::GR32RegisterClass;
1571 // Must be in x86-64 mode.
1573 RC = X86::GR64RegisterClass;
1576 if (Subtarget->hasSSE1()) {
1578 RC = X86::FR32RegisterClass;
1580 Opc = X86::LD_Fp32m;
1581 RC = X86::RFP32RegisterClass;
1585 if (Subtarget->hasSSE2()) {
1587 RC = X86::FR64RegisterClass;
1589 Opc = X86::LD_Fp64m;
1590 RC = X86::RFP64RegisterClass;
1594 // No f80 support yet.
1598 // Materialize addresses with LEA instructions.
1599 if (isa<GlobalValue>(C)) {
1601 if (X86SelectAddress(C, AM)) {
1602 if (TLI.getPointerTy() == MVT::i32)
1606 unsigned ResultReg = createResultReg(RC);
1607 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1613 // MachineConstantPool wants an explicit alignment.
1614 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1616 // Alignment of vector types. FIXME!
1617 Align = TD.getTypeAllocSize(C->getType());
1620 // x86-32 PIC requires a PIC base register for constant pools.
1621 unsigned PICBase = 0;
1622 unsigned char OpFlag = 0;
1623 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1624 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1625 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1626 } else if (Subtarget->isPICStyleGOT()) {
1627 OpFlag = X86II::MO_GOTOFF;
1628 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1629 } else if (Subtarget->isPICStyleRIPRel() &&
1630 TM.getCodeModel() == CodeModel::Small) {
1634 // Create the load from the constant pool.
1635 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1636 unsigned ResultReg = createResultReg(RC);
1637 addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
1638 MCPOffset, PICBase, OpFlag);
1643 unsigned X86FastISel::TargetMaterializeAlloca(AllocaInst *C) {
1644 // Fail on dynamic allocas. At this point, getRegForValue has already
1645 // checked its CSE maps, so if we're here trying to handle a dynamic
1646 // alloca, we're not going to succeed. X86SelectAddress has a
1647 // check for dynamic allocas, because it's called directly from
1648 // various places, but TargetMaterializeAlloca also needs a check
1649 // in order to avoid recursion between getRegForValue,
1650 // X86SelectAddrss, and TargetMaterializeAlloca.
1651 if (!StaticAllocaMap.count(C))
1655 if (!X86SelectAddress(C, AM))
1657 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1658 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1659 unsigned ResultReg = createResultReg(RC);
1660 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1665 llvm::FastISel *X86::createFastISel(MachineFunction &mf,
1666 MachineModuleInfo *mmi,
1668 DenseMap<const Value *, unsigned> &vm,
1669 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
1670 DenseMap<const AllocaInst *, int> &am
1672 , SmallSet<Instruction*, 8> &cil
1675 return new X86FastISel(mf, mmi, dw, vm, bm, am