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/Intrinsics.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"
36 class X86FastISel : public FastISel {
37 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
38 /// make the right decision when generating code for different targets.
39 const X86Subtarget *Subtarget;
41 /// StackPtr - Register used as the stack pointer.
45 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
46 /// floating point ops.
47 /// When SSE is available, use it for f32 operations.
48 /// When SSE2 is available, use it for f64 operations.
53 explicit X86FastISel(MachineFunction &mf,
54 MachineModuleInfo *mmi,
56 DenseMap<const Value *, unsigned> &vm,
57 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
58 DenseMap<const AllocaInst *, int> &am
60 , SmallSet<Instruction*, 8> &cil
63 : FastISel(mf, mmi, dw, vm, bm, am
68 Subtarget = &TM.getSubtarget<X86Subtarget>();
69 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
70 X86ScalarSSEf64 = Subtarget->hasSSE2();
71 X86ScalarSSEf32 = Subtarget->hasSSE1();
74 virtual bool TargetSelectInstruction(Instruction *I);
76 #include "X86GenFastISel.inc"
79 bool X86FastEmitCompare(Value *LHS, Value *RHS, MVT VT);
81 bool X86FastEmitLoad(MVT VT, const X86AddressMode &AM, unsigned &RR);
83 bool X86FastEmitStore(MVT VT, Value *Val,
84 const X86AddressMode &AM);
85 bool X86FastEmitStore(MVT VT, unsigned Val,
86 const X86AddressMode &AM);
88 bool X86FastEmitExtend(ISD::NodeType Opc, MVT DstVT, unsigned Src, MVT SrcVT,
91 bool X86SelectAddress(Value *V, X86AddressMode &AM, bool isCall);
93 bool X86SelectLoad(Instruction *I);
95 bool X86SelectStore(Instruction *I);
97 bool X86SelectCmp(Instruction *I);
99 bool X86SelectZExt(Instruction *I);
101 bool X86SelectBranch(Instruction *I);
103 bool X86SelectShift(Instruction *I);
105 bool X86SelectSelect(Instruction *I);
107 bool X86SelectTrunc(Instruction *I);
109 bool X86SelectFPExt(Instruction *I);
110 bool X86SelectFPTrunc(Instruction *I);
112 bool X86SelectExtractValue(Instruction *I);
114 bool X86VisitIntrinsicCall(CallInst &I, unsigned Intrinsic);
115 bool X86SelectCall(Instruction *I);
117 CCAssignFn *CCAssignFnForCall(unsigned CC, bool isTailCall = false);
119 const X86InstrInfo *getInstrInfo() const {
120 return getTargetMachine()->getInstrInfo();
122 const X86TargetMachine *getTargetMachine() const {
123 return static_cast<const X86TargetMachine *>(&TM);
126 unsigned TargetMaterializeConstant(Constant *C);
128 unsigned TargetMaterializeAlloca(AllocaInst *C);
130 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
131 /// computed in an SSE register, not on the X87 floating point stack.
132 bool isScalarFPTypeInSSEReg(MVT VT) const {
133 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
134 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
137 bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
140 bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
141 VT = TLI.getValueType(Ty, /*HandleUnknown=*/true);
142 if (VT == MVT::Other || !VT.isSimple())
143 // Unhandled type. Halt "fast" selection and bail.
146 // For now, require SSE/SSE2 for performing floating-point operations,
147 // since x87 requires additional work.
148 if (VT == MVT::f64 && !X86ScalarSSEf64)
150 if (VT == MVT::f32 && !X86ScalarSSEf32)
152 // Similarly, no f80 support yet.
155 // We only handle legal types. For example, on x86-32 the instruction
156 // selector contains all of the 64-bit instructions from x86-64,
157 // under the assumption that i64 won't be used if the target doesn't
159 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
162 #include "X86GenCallingConv.inc"
164 /// CCAssignFnForCall - Selects the correct CCAssignFn for a given calling
166 CCAssignFn *X86FastISel::CCAssignFnForCall(unsigned CC, bool isTaillCall) {
167 if (Subtarget->is64Bit()) {
168 if (Subtarget->isTargetWin64())
169 return CC_X86_Win64_C;
170 else if (CC == CallingConv::Fast && isTaillCall)
171 return CC_X86_64_TailCall;
176 if (CC == CallingConv::X86_FastCall)
177 return CC_X86_32_FastCall;
178 else if (CC == CallingConv::Fast)
179 return CC_X86_32_FastCC;
184 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
185 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
186 /// Return true and the result register by reference if it is possible.
187 bool X86FastISel::X86FastEmitLoad(MVT VT, const X86AddressMode &AM,
188 unsigned &ResultReg) {
189 // Get opcode and regclass of the output for the given load instruction.
191 const TargetRegisterClass *RC = NULL;
192 switch (VT.getSimpleVT()) {
193 default: return false;
196 RC = X86::GR8RegisterClass;
200 RC = X86::GR16RegisterClass;
204 RC = X86::GR32RegisterClass;
207 // Must be in x86-64 mode.
209 RC = X86::GR64RegisterClass;
212 if (Subtarget->hasSSE1()) {
214 RC = X86::FR32RegisterClass;
217 RC = X86::RFP32RegisterClass;
221 if (Subtarget->hasSSE2()) {
223 RC = X86::FR64RegisterClass;
226 RC = X86::RFP64RegisterClass;
230 // No f80 support yet.
234 ResultReg = createResultReg(RC);
235 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
239 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
240 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
241 /// and a displacement offset, or a GlobalAddress,
242 /// i.e. V. Return true if it is possible.
244 X86FastISel::X86FastEmitStore(MVT VT, unsigned Val,
245 const X86AddressMode &AM) {
246 // Get opcode and regclass of the output for the given store instruction.
248 switch (VT.getSimpleVT()) {
249 case MVT::f80: // No f80 support yet.
250 default: return false;
251 case MVT::i8: Opc = X86::MOV8mr; break;
252 case MVT::i16: Opc = X86::MOV16mr; break;
253 case MVT::i32: Opc = X86::MOV32mr; break;
254 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
256 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
259 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
263 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM).addReg(Val);
267 bool X86FastISel::X86FastEmitStore(MVT VT, Value *Val,
268 const X86AddressMode &AM) {
269 // Handle 'null' like i32/i64 0.
270 if (isa<ConstantPointerNull>(Val))
271 Val = Constant::getNullValue(TD.getIntPtrType());
273 // If this is a store of a simple constant, fold the constant into the store.
274 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
276 switch (VT.getSimpleVT()) {
278 case MVT::i8: Opc = X86::MOV8mi; break;
279 case MVT::i16: Opc = X86::MOV16mi; break;
280 case MVT::i32: Opc = X86::MOV32mi; break;
282 // Must be a 32-bit sign extended value.
283 if ((int)CI->getSExtValue() == CI->getSExtValue())
284 Opc = X86::MOV64mi32;
289 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM)
290 .addImm(CI->getSExtValue());
295 unsigned ValReg = getRegForValue(Val);
299 return X86FastEmitStore(VT, ValReg, AM);
302 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
303 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
304 /// ISD::SIGN_EXTEND).
305 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, MVT DstVT,
306 unsigned Src, MVT SrcVT,
307 unsigned &ResultReg) {
308 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc, Src);
317 /// X86SelectAddress - Attempt to fill in an address from the given value.
319 bool X86FastISel::X86SelectAddress(Value *V, X86AddressMode &AM, bool isCall) {
321 unsigned Opcode = Instruction::UserOp1;
322 if (Instruction *I = dyn_cast<Instruction>(V)) {
323 Opcode = I->getOpcode();
325 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
326 Opcode = C->getOpcode();
332 case Instruction::BitCast:
333 // Look past bitcasts.
334 return X86SelectAddress(U->getOperand(0), AM, isCall);
336 case Instruction::IntToPtr:
337 // Look past no-op inttoptrs.
338 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
339 return X86SelectAddress(U->getOperand(0), AM, isCall);
342 case Instruction::PtrToInt:
343 // Look past no-op ptrtoints.
344 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
345 return X86SelectAddress(U->getOperand(0), AM, isCall);
348 case Instruction::Alloca: {
350 // Do static allocas.
351 const AllocaInst *A = cast<AllocaInst>(V);
352 DenseMap<const AllocaInst*, int>::iterator SI = StaticAllocaMap.find(A);
353 if (SI != StaticAllocaMap.end()) {
354 AM.BaseType = X86AddressMode::FrameIndexBase;
355 AM.Base.FrameIndex = SI->second;
361 case Instruction::Add: {
363 // Adds of constants are common and easy enough.
364 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
365 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
366 // They have to fit in the 32-bit signed displacement field though.
368 AM.Disp = (uint32_t)Disp;
369 return X86SelectAddress(U->getOperand(0), AM, isCall);
375 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.getTypePaddedSize(GTI.getIndexedType());
393 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
394 // Constant-offset addressing.
395 Disp += CI->getSExtValue() * S;
396 } else if (IndexReg == 0 &&
398 !getTargetMachine()->symbolicAddressesAreRIPRel()) &&
399 (S == 1 || S == 2 || S == 4 || S == 8)) {
400 // Scaled-index addressing.
402 IndexReg = getRegForGEPIndex(Op);
407 goto unsupported_gep;
410 // Check for displacement overflow.
413 // Ok, the GEP indices were covered by constant-offset and scaled-index
414 // addressing. Update the address state and move on to examining the base.
415 AM.IndexReg = IndexReg;
417 AM.Disp = (uint32_t)Disp;
418 return X86SelectAddress(U->getOperand(0), AM, isCall);
420 // Ok, the GEP indices weren't all covered.
425 // Handle constant address.
426 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
427 // Can't handle alternate code models yet.
428 if (TM.getCodeModel() != CodeModel::Default &&
429 TM.getCodeModel() != CodeModel::Small)
432 // RIP-relative addresses can't have additional register operands.
433 if (getTargetMachine()->symbolicAddressesAreRIPRel() &&
434 (AM.Base.Reg != 0 || AM.IndexReg != 0))
437 // Can't handle TLS yet.
438 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
439 if (GVar->isThreadLocal())
442 // Set up the basic address.
445 TM.getRelocationModel() == Reloc::PIC_ &&
446 !Subtarget->is64Bit())
447 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(&MF);
449 // Emit an extra load if the ABI requires it.
450 if (Subtarget->GVRequiresExtraLoad(GV, TM, isCall)) {
451 // Check to see if we've already materialized this
452 // value in a register in this block.
453 if (unsigned Reg = LocalValueMap[V]) {
458 // Issue load from stub if necessary.
460 const TargetRegisterClass *RC = NULL;
461 if (TLI.getPointerTy() == MVT::i32) {
463 RC = X86::GR32RegisterClass;
466 RC = X86::GR64RegisterClass;
469 X86AddressMode StubAM;
470 StubAM.Base.Reg = AM.Base.Reg;
472 unsigned ResultReg = createResultReg(RC);
473 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), StubAM);
475 // Now construct the final address. Note that the Disp, Scale,
476 // and Index values may already be set here.
477 AM.Base.Reg = ResultReg;
480 // Prevent loading GV stub multiple times in same MBB.
481 LocalValueMap[V] = AM.Base.Reg;
486 // If all else fails, try to materialize the value in a register.
487 if (!AM.GV || !getTargetMachine()->symbolicAddressesAreRIPRel()) {
488 if (AM.Base.Reg == 0) {
489 AM.Base.Reg = getRegForValue(V);
490 return AM.Base.Reg != 0;
492 if (AM.IndexReg == 0) {
493 assert(AM.Scale == 1 && "Scale with no index!");
494 AM.IndexReg = getRegForValue(V);
495 return AM.IndexReg != 0;
502 /// X86SelectStore - Select and emit code to implement store instructions.
503 bool X86FastISel::X86SelectStore(Instruction* I) {
505 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
509 if (!X86SelectAddress(I->getOperand(1), AM, false))
512 return X86FastEmitStore(VT, I->getOperand(0), AM);
515 /// X86SelectLoad - Select and emit code to implement load instructions.
517 bool X86FastISel::X86SelectLoad(Instruction *I) {
519 if (!isTypeLegal(I->getType(), VT))
523 if (!X86SelectAddress(I->getOperand(0), AM, false))
526 unsigned ResultReg = 0;
527 if (X86FastEmitLoad(VT, AM, ResultReg)) {
528 UpdateValueMap(I, ResultReg);
534 static unsigned X86ChooseCmpOpcode(MVT VT) {
535 switch (VT.getSimpleVT()) {
537 case MVT::i8: return X86::CMP8rr;
538 case MVT::i16: return X86::CMP16rr;
539 case MVT::i32: return X86::CMP32rr;
540 case MVT::i64: return X86::CMP64rr;
541 case MVT::f32: return X86::UCOMISSrr;
542 case MVT::f64: return X86::UCOMISDrr;
546 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
547 /// of the comparison, return an opcode that works for the compare (e.g.
548 /// CMP32ri) otherwise return 0.
549 static unsigned X86ChooseCmpImmediateOpcode(MVT VT, ConstantInt *RHSC) {
550 switch (VT.getSimpleVT()) {
551 // Otherwise, we can't fold the immediate into this comparison.
553 case MVT::i8: return X86::CMP8ri;
554 case MVT::i16: return X86::CMP16ri;
555 case MVT::i32: return X86::CMP32ri;
557 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
559 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
560 return X86::CMP64ri32;
565 bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, MVT VT) {
566 unsigned Op0Reg = getRegForValue(Op0);
567 if (Op0Reg == 0) return false;
569 // Handle 'null' like i32/i64 0.
570 if (isa<ConstantPointerNull>(Op1))
571 Op1 = Constant::getNullValue(TD.getIntPtrType());
573 // We have two options: compare with register or immediate. If the RHS of
574 // the compare is an immediate that we can fold into this compare, use
575 // CMPri, otherwise use CMPrr.
576 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
577 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
578 BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
579 .addImm(Op1C->getSExtValue());
584 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
585 if (CompareOpc == 0) return false;
587 unsigned Op1Reg = getRegForValue(Op1);
588 if (Op1Reg == 0) return false;
589 BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
594 bool X86FastISel::X86SelectCmp(Instruction *I) {
595 CmpInst *CI = cast<CmpInst>(I);
598 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
601 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
603 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
604 switch (CI->getPredicate()) {
605 case CmpInst::FCMP_OEQ: {
606 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
609 unsigned EReg = createResultReg(&X86::GR8RegClass);
610 unsigned NPReg = createResultReg(&X86::GR8RegClass);
611 BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
612 BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
614 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
615 UpdateValueMap(I, ResultReg);
618 case CmpInst::FCMP_UNE: {
619 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
622 unsigned NEReg = createResultReg(&X86::GR8RegClass);
623 unsigned PReg = createResultReg(&X86::GR8RegClass);
624 BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
625 BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
626 BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
627 UpdateValueMap(I, ResultReg);
630 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
631 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
632 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
633 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
634 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
635 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
636 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
637 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
638 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
639 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
640 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
641 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
643 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
644 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
645 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
646 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
647 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
648 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
649 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
650 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
651 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
652 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
657 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
661 // Emit a compare of Op0/Op1.
662 if (!X86FastEmitCompare(Op0, Op1, VT))
665 BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
666 UpdateValueMap(I, ResultReg);
670 bool X86FastISel::X86SelectZExt(Instruction *I) {
671 // Special-case hack: The only i1 values we know how to produce currently
672 // set the upper bits of an i8 value to zero.
673 if (I->getType() == Type::Int8Ty &&
674 I->getOperand(0)->getType() == Type::Int1Ty) {
675 unsigned ResultReg = getRegForValue(I->getOperand(0));
676 if (ResultReg == 0) return false;
677 UpdateValueMap(I, ResultReg);
685 bool X86FastISel::X86SelectBranch(Instruction *I) {
686 // Unconditional branches are selected by tablegen-generated code.
687 // Handle a conditional branch.
688 BranchInst *BI = cast<BranchInst>(I);
689 MachineBasicBlock *TrueMBB = MBBMap[BI->getSuccessor(0)];
690 MachineBasicBlock *FalseMBB = MBBMap[BI->getSuccessor(1)];
692 // Fold the common case of a conditional branch with a comparison.
693 if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
694 if (CI->hasOneUse()) {
695 MVT VT = TLI.getValueType(CI->getOperand(0)->getType());
697 // Try to take advantage of fallthrough opportunities.
698 CmpInst::Predicate Predicate = CI->getPredicate();
699 if (MBB->isLayoutSuccessor(TrueMBB)) {
700 std::swap(TrueMBB, FalseMBB);
701 Predicate = CmpInst::getInversePredicate(Predicate);
704 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
705 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
708 case CmpInst::FCMP_OEQ:
709 std::swap(TrueMBB, FalseMBB);
710 Predicate = CmpInst::FCMP_UNE;
712 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE; break;
713 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA; break;
714 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE; break;
715 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA; break;
716 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE; break;
717 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE; break;
718 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP; break;
719 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP; break;
720 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE; break;
721 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB; break;
722 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE; break;
723 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
724 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
726 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE; break;
727 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE; break;
728 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA; break;
729 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE; break;
730 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB; break;
731 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE; break;
732 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG; break;
733 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE; break;
734 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL; break;
735 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE; break;
740 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
744 // Emit a compare of the LHS and RHS, setting the flags.
745 if (!X86FastEmitCompare(Op0, Op1, VT))
748 BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
750 if (Predicate == CmpInst::FCMP_UNE) {
751 // X86 requires a second branch to handle UNE (and OEQ,
752 // which is mapped to UNE above).
753 BuildMI(MBB, DL, TII.get(X86::JP)).addMBB(TrueMBB);
756 FastEmitBranch(FalseMBB);
757 MBB->addSuccessor(TrueMBB);
760 } else if (ExtractValueInst *EI =
761 dyn_cast<ExtractValueInst>(BI->getCondition())) {
762 // Check to see if the branch instruction is from an "arithmetic with
763 // overflow" intrinsic. The main way these intrinsics are used is:
765 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
766 // %sum = extractvalue { i32, i1 } %t, 0
767 // %obit = extractvalue { i32, i1 } %t, 1
768 // br i1 %obit, label %overflow, label %normal
770 // The %sum and %obit are converted in an ADD and a SETO/SETB before
771 // reaching the branch. Therefore, we search backwards through the MBB
772 // looking for the SETO/SETB instruction. If an instruction modifies the
773 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
774 // convert the branch into a JO/JB instruction.
776 Value *Agg = EI->getAggregateOperand();
778 if (CallInst *CI = dyn_cast<CallInst>(Agg)) {
779 Function *F = CI->getCalledFunction();
781 if (F && F->isDeclaration()) {
782 switch (F->getIntrinsicID()) {
784 case Intrinsic::sadd_with_overflow:
785 case Intrinsic::uadd_with_overflow: {
786 const MachineInstr *SetMI = 0;
787 unsigned Reg = lookUpRegForValue(EI);
789 for (MachineBasicBlock::const_reverse_iterator
790 RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
791 const MachineInstr &MI = *RI;
793 if (MI.modifiesRegister(Reg)) {
794 unsigned Src, Dst, SrcSR, DstSR;
796 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
805 const TargetInstrDesc &TID = MI.getDesc();
806 const unsigned *ImpDefs = TID.getImplicitDefs();
808 if (TID.hasUnmodeledSideEffects()) break;
810 bool ModifiesEFlags = false;
813 for (unsigned u = 0; ImpDefs[u]; ++u)
814 if (ImpDefs[u] == X86::EFLAGS) {
815 ModifiesEFlags = true;
820 if (ModifiesEFlags) break;
824 unsigned OpCode = SetMI->getOpcode();
826 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
827 BuildMI(MBB, DL, TII.get((OpCode == X86::SETOr) ?
828 X86::JO : X86::JB)).addMBB(TrueMBB);
829 FastEmitBranch(FalseMBB);
830 MBB->addSuccessor(TrueMBB);
840 // Otherwise do a clumsy setcc and re-test it.
841 unsigned OpReg = getRegForValue(BI->getCondition());
842 if (OpReg == 0) return false;
844 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
845 BuildMI(MBB, DL, TII.get(X86::JNE)).addMBB(TrueMBB);
846 FastEmitBranch(FalseMBB);
847 MBB->addSuccessor(TrueMBB);
851 bool X86FastISel::X86SelectShift(Instruction *I) {
852 unsigned CReg = 0, OpReg = 0, OpImm = 0;
853 const TargetRegisterClass *RC = NULL;
854 if (I->getType() == Type::Int8Ty) {
856 RC = &X86::GR8RegClass;
857 switch (I->getOpcode()) {
858 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
859 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
860 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
861 default: return false;
863 } else if (I->getType() == Type::Int16Ty) {
865 RC = &X86::GR16RegClass;
866 switch (I->getOpcode()) {
867 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
868 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
869 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
870 default: return false;
872 } else if (I->getType() == Type::Int32Ty) {
874 RC = &X86::GR32RegClass;
875 switch (I->getOpcode()) {
876 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
877 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
878 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
879 default: return false;
881 } else if (I->getType() == Type::Int64Ty) {
883 RC = &X86::GR64RegClass;
884 switch (I->getOpcode()) {
885 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
886 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
887 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
888 default: return false;
894 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
895 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
898 unsigned Op0Reg = getRegForValue(I->getOperand(0));
899 if (Op0Reg == 0) return false;
901 // Fold immediate in shl(x,3).
902 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
903 unsigned ResultReg = createResultReg(RC);
904 BuildMI(MBB, DL, TII.get(OpImm),
905 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
906 UpdateValueMap(I, ResultReg);
910 unsigned Op1Reg = getRegForValue(I->getOperand(1));
911 if (Op1Reg == 0) return false;
912 TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);
914 // The shift instruction uses X86::CL. If we defined a super-register
915 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
918 BuildMI(MBB, DL, TII.get(TargetInstrInfo::EXTRACT_SUBREG), X86::CL)
919 .addReg(CReg).addImm(X86::SUBREG_8BIT);
921 unsigned ResultReg = createResultReg(RC);
922 BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
923 UpdateValueMap(I, ResultReg);
927 bool X86FastISel::X86SelectSelect(Instruction *I) {
928 MVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
929 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
933 const TargetRegisterClass *RC = NULL;
934 if (VT.getSimpleVT() == MVT::i16) {
935 Opc = X86::CMOVE16rr;
936 RC = &X86::GR16RegClass;
937 } else if (VT.getSimpleVT() == MVT::i32) {
938 Opc = X86::CMOVE32rr;
939 RC = &X86::GR32RegClass;
940 } else if (VT.getSimpleVT() == MVT::i64) {
941 Opc = X86::CMOVE64rr;
942 RC = &X86::GR64RegClass;
947 unsigned Op0Reg = getRegForValue(I->getOperand(0));
948 if (Op0Reg == 0) return false;
949 unsigned Op1Reg = getRegForValue(I->getOperand(1));
950 if (Op1Reg == 0) return false;
951 unsigned Op2Reg = getRegForValue(I->getOperand(2));
952 if (Op2Reg == 0) return false;
954 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
955 unsigned ResultReg = createResultReg(RC);
956 BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
957 UpdateValueMap(I, ResultReg);
961 bool X86FastISel::X86SelectFPExt(Instruction *I) {
962 // fpext from float to double.
963 if (Subtarget->hasSSE2() && I->getType() == Type::DoubleTy) {
964 Value *V = I->getOperand(0);
965 if (V->getType() == Type::FloatTy) {
966 unsigned OpReg = getRegForValue(V);
967 if (OpReg == 0) return false;
968 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
969 BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
970 UpdateValueMap(I, ResultReg);
978 bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
979 if (Subtarget->hasSSE2()) {
980 if (I->getType() == Type::FloatTy) {
981 Value *V = I->getOperand(0);
982 if (V->getType() == Type::DoubleTy) {
983 unsigned OpReg = getRegForValue(V);
984 if (OpReg == 0) return false;
985 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
986 BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
987 UpdateValueMap(I, ResultReg);
996 bool X86FastISel::X86SelectTrunc(Instruction *I) {
997 if (Subtarget->is64Bit())
998 // All other cases should be handled by the tblgen generated code.
1000 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1001 MVT DstVT = TLI.getValueType(I->getType());
1002 if (DstVT != MVT::i8)
1003 // All other cases should be handled by the tblgen generated code.
1005 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1006 // All other cases should be handled by the tblgen generated code.
1009 unsigned InputReg = getRegForValue(I->getOperand(0));
1011 // Unhandled operand. Halt "fast" selection and bail.
1014 // First issue a copy to GR16_ or GR32_.
1015 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16to16_ : X86::MOV32to32_;
1016 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1017 ? X86::GR16_RegisterClass : X86::GR32_RegisterClass;
1018 unsigned CopyReg = createResultReg(CopyRC);
1019 BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
1021 // Then issue an extract_subreg.
1022 unsigned ResultReg = FastEmitInst_extractsubreg(DstVT.getSimpleVT(),
1023 CopyReg, X86::SUBREG_8BIT);
1027 UpdateValueMap(I, ResultReg);
1031 bool X86FastISel::X86SelectExtractValue(Instruction *I) {
1032 ExtractValueInst *EI = cast<ExtractValueInst>(I);
1033 Value *Agg = EI->getAggregateOperand();
1035 if (CallInst *CI = dyn_cast<CallInst>(Agg)) {
1036 Function *F = CI->getCalledFunction();
1038 if (F && F->isDeclaration()) {
1039 switch (F->getIntrinsicID()) {
1041 case Intrinsic::sadd_with_overflow:
1042 case Intrinsic::uadd_with_overflow:
1043 // Cheat a little. We know that the registers for "add" and "seto" are
1044 // allocated sequentially. However, we only keep track of the register
1045 // for "add" in the value map. Use extractvalue's index to get the
1046 // correct register for "seto".
1047 UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
1056 bool X86FastISel::X86VisitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
1057 // FIXME: Handle more intrinsics.
1058 switch (Intrinsic) {
1059 default: return false;
1060 case Intrinsic::sadd_with_overflow:
1061 case Intrinsic::uadd_with_overflow: {
1062 // Replace "add with overflow" intrinsics with an "add" instruction followed
1063 // by a seto/setc instruction. Later on, when the "extractvalue"
1064 // instructions are encountered, we use the fact that two registers were
1065 // created sequentially to get the correct registers for the "sum" and the
1068 const Function *Callee = I.getCalledFunction();
1070 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1072 if (!isTypeLegal(RetTy, VT))
1075 Value *Op1 = I.getOperand(1);
1076 Value *Op2 = I.getOperand(2);
1077 unsigned Reg1 = getRegForValue(Op1);
1078 unsigned Reg2 = getRegForValue(Op2);
1080 if (Reg1 == 0 || Reg2 == 0)
1081 // FIXME: Handle values *not* in registers.
1088 else if (VT == MVT::i64)
1093 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1094 BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
1095 UpdateValueMap(&I, ResultReg);
1097 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1098 BuildMI(MBB, DL, TII.get((Intrinsic == Intrinsic::sadd_with_overflow) ?
1099 X86::SETOr : X86::SETBr), ResultReg);
1105 bool X86FastISel::X86SelectCall(Instruction *I) {
1106 CallInst *CI = cast<CallInst>(I);
1107 Value *Callee = I->getOperand(0);
1109 // Can't handle inline asm yet.
1110 if (isa<InlineAsm>(Callee))
1113 // Handle intrinsic calls.
1114 if (Function *F = CI->getCalledFunction())
1115 if (F->isDeclaration())
1116 if (unsigned IID = F->getIntrinsicID())
1117 return X86VisitIntrinsicCall(*CI, IID);
1119 // Handle only C and fastcc calling conventions for now.
1121 unsigned CC = CS.getCallingConv();
1122 if (CC != CallingConv::C &&
1123 CC != CallingConv::Fast &&
1124 CC != CallingConv::X86_FastCall)
1127 // Let SDISel handle vararg functions.
1128 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1129 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1130 if (FTy->isVarArg())
1133 // Handle *simple* calls for now.
1134 const Type *RetTy = CS.getType();
1136 if (RetTy == Type::VoidTy)
1137 RetVT = MVT::isVoid;
1138 else if (!isTypeLegal(RetTy, RetVT, true))
1141 // Materialize callee address in a register. FIXME: GV address can be
1142 // handled with a CALLpcrel32 instead.
1143 X86AddressMode CalleeAM;
1144 if (!X86SelectAddress(Callee, CalleeAM, true))
1146 unsigned CalleeOp = 0;
1147 GlobalValue *GV = 0;
1148 if (CalleeAM.Base.Reg != 0) {
1149 assert(CalleeAM.GV == 0);
1150 CalleeOp = CalleeAM.Base.Reg;
1151 } else if (CalleeAM.GV != 0) {
1152 assert(CalleeAM.GV != 0);
1157 // Allow calls which produce i1 results.
1158 bool AndToI1 = false;
1159 if (RetVT == MVT::i1) {
1164 // Deal with call operands first.
1165 SmallVector<Value*, 8> ArgVals;
1166 SmallVector<unsigned, 8> Args;
1167 SmallVector<MVT, 8> ArgVTs;
1168 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1169 Args.reserve(CS.arg_size());
1170 ArgVals.reserve(CS.arg_size());
1171 ArgVTs.reserve(CS.arg_size());
1172 ArgFlags.reserve(CS.arg_size());
1173 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1175 unsigned Arg = getRegForValue(*i);
1178 ISD::ArgFlagsTy Flags;
1179 unsigned AttrInd = i - CS.arg_begin() + 1;
1180 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1182 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1185 // FIXME: Only handle *easy* calls for now.
1186 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1187 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1188 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1189 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1192 const Type *ArgTy = (*i)->getType();
1194 if (!isTypeLegal(ArgTy, ArgVT))
1196 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1197 Flags.setOrigAlign(OriginalAlignment);
1199 Args.push_back(Arg);
1200 ArgVals.push_back(*i);
1201 ArgVTs.push_back(ArgVT);
1202 ArgFlags.push_back(Flags);
1205 // Analyze operands of the call, assigning locations to each operand.
1206 SmallVector<CCValAssign, 16> ArgLocs;
1207 CCState CCInfo(CC, false, TM, ArgLocs);
1208 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1210 // Get a count of how many bytes are to be pushed on the stack.
1211 unsigned NumBytes = CCInfo.getNextStackOffset();
1213 // Issue CALLSEQ_START
1214 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1215 BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
1217 // Process argument: walk the register/memloc assignments, inserting
1219 SmallVector<unsigned, 4> RegArgs;
1220 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1221 CCValAssign &VA = ArgLocs[i];
1222 unsigned Arg = Args[VA.getValNo()];
1223 MVT ArgVT = ArgVTs[VA.getValNo()];
1225 // Promote the value if needed.
1226 switch (VA.getLocInfo()) {
1227 default: assert(0 && "Unknown loc info!");
1228 case CCValAssign::Full: break;
1229 case CCValAssign::SExt: {
1230 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1232 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1234 ArgVT = VA.getLocVT();
1237 case CCValAssign::ZExt: {
1238 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1240 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1242 ArgVT = VA.getLocVT();
1245 case CCValAssign::AExt: {
1246 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1249 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1252 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1255 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1256 ArgVT = VA.getLocVT();
1261 if (VA.isRegLoc()) {
1262 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1263 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
1265 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1267 RegArgs.push_back(VA.getLocReg());
1269 unsigned LocMemOffset = VA.getLocMemOffset();
1271 AM.Base.Reg = StackPtr;
1272 AM.Disp = LocMemOffset;
1273 Value *ArgVal = ArgVals[VA.getValNo()];
1275 // If this is a really simple value, emit this with the Value* version of
1276 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1277 // can cause us to reevaluate the argument.
1278 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1279 X86FastEmitStore(ArgVT, ArgVal, AM);
1281 X86FastEmitStore(ArgVT, Arg, AM);
1285 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1287 if (!Subtarget->is64Bit() &&
1288 TM.getRelocationModel() == Reloc::PIC_ &&
1289 Subtarget->isPICStyleGOT()) {
1290 TargetRegisterClass *RC = X86::GR32RegisterClass;
1291 unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
1292 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
1293 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1298 unsigned CallOpc = CalleeOp
1299 ? (Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r)
1300 : (Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32);
1301 MachineInstrBuilder MIB = CalleeOp
1302 ? BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp)
1303 : BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV);
1305 // Add an implicit use GOT pointer in EBX.
1306 if (!Subtarget->is64Bit() &&
1307 TM.getRelocationModel() == Reloc::PIC_ &&
1308 Subtarget->isPICStyleGOT())
1309 MIB.addReg(X86::EBX);
1311 // Add implicit physical register uses to the call.
1312 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1313 MIB.addReg(RegArgs[i]);
1315 // Issue CALLSEQ_END
1316 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1317 BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
1319 // Now handle call return value (if any).
1320 if (RetVT.getSimpleVT() != MVT::isVoid) {
1321 SmallVector<CCValAssign, 16> RVLocs;
1322 CCState CCInfo(CC, false, TM, RVLocs);
1323 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1325 // Copy all of the result registers out of their specified physreg.
1326 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1327 MVT CopyVT = RVLocs[0].getValVT();
1328 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1329 TargetRegisterClass *SrcRC = DstRC;
1331 // If this is a call to a function that returns an fp value on the x87 fp
1332 // stack, but where we prefer to use the value in xmm registers, copy it
1333 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1334 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1335 RVLocs[0].getLocReg() == X86::ST1) &&
1336 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1338 SrcRC = X86::RSTRegisterClass;
1339 DstRC = X86::RFP80RegisterClass;
1342 unsigned ResultReg = createResultReg(DstRC);
1343 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1344 RVLocs[0].getLocReg(), DstRC, SrcRC);
1345 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1347 if (CopyVT != RVLocs[0].getValVT()) {
1348 // Round the F80 the right size, which also moves to the appropriate xmm
1349 // register. This is accomplished by storing the F80 value in memory and
1350 // then loading it back. Ewww...
1351 MVT ResVT = RVLocs[0].getValVT();
1352 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1353 unsigned MemSize = ResVT.getSizeInBits()/8;
1354 int FI = MFI.CreateStackObject(MemSize, MemSize);
1355 addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
1356 DstRC = ResVT == MVT::f32
1357 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1358 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1359 ResultReg = createResultReg(DstRC);
1360 addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
1364 // Mask out all but lowest bit for some call which produces an i1.
1365 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1367 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1368 ResultReg = AndResult;
1371 UpdateValueMap(I, ResultReg);
1379 X86FastISel::TargetSelectInstruction(Instruction *I) {
1380 switch (I->getOpcode()) {
1382 case Instruction::Load:
1383 return X86SelectLoad(I);
1384 case Instruction::Store:
1385 return X86SelectStore(I);
1386 case Instruction::ICmp:
1387 case Instruction::FCmp:
1388 return X86SelectCmp(I);
1389 case Instruction::ZExt:
1390 return X86SelectZExt(I);
1391 case Instruction::Br:
1392 return X86SelectBranch(I);
1393 case Instruction::Call:
1394 return X86SelectCall(I);
1395 case Instruction::LShr:
1396 case Instruction::AShr:
1397 case Instruction::Shl:
1398 return X86SelectShift(I);
1399 case Instruction::Select:
1400 return X86SelectSelect(I);
1401 case Instruction::Trunc:
1402 return X86SelectTrunc(I);
1403 case Instruction::FPExt:
1404 return X86SelectFPExt(I);
1405 case Instruction::FPTrunc:
1406 return X86SelectFPTrunc(I);
1407 case Instruction::ExtractValue:
1408 return X86SelectExtractValue(I);
1414 unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
1416 if (!isTypeLegal(C->getType(), VT))
1419 // Get opcode and regclass of the output for the given load instruction.
1421 const TargetRegisterClass *RC = NULL;
1422 switch (VT.getSimpleVT()) {
1423 default: return false;
1426 RC = X86::GR8RegisterClass;
1430 RC = X86::GR16RegisterClass;
1434 RC = X86::GR32RegisterClass;
1437 // Must be in x86-64 mode.
1439 RC = X86::GR64RegisterClass;
1442 if (Subtarget->hasSSE1()) {
1444 RC = X86::FR32RegisterClass;
1446 Opc = X86::LD_Fp32m;
1447 RC = X86::RFP32RegisterClass;
1451 if (Subtarget->hasSSE2()) {
1453 RC = X86::FR64RegisterClass;
1455 Opc = X86::LD_Fp64m;
1456 RC = X86::RFP64RegisterClass;
1460 // No f80 support yet.
1464 // Materialize addresses with LEA instructions.
1465 if (isa<GlobalValue>(C)) {
1467 if (X86SelectAddress(C, AM, false)) {
1468 if (TLI.getPointerTy() == MVT::i32)
1472 unsigned ResultReg = createResultReg(RC);
1473 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1479 // MachineConstantPool wants an explicit alignment.
1480 unsigned Align = TD.getPreferredTypeAlignmentShift(C->getType());
1482 // Alignment of vector types. FIXME!
1483 Align = TD.getTypePaddedSize(C->getType());
1484 Align = Log2_64(Align);
1487 // x86-32 PIC requires a PIC base register for constant pools.
1488 unsigned PICBase = 0;
1489 if (TM.getRelocationModel() == Reloc::PIC_ &&
1490 !Subtarget->is64Bit())
1491 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1493 // Create the load from the constant pool.
1494 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1495 unsigned ResultReg = createResultReg(RC);
1496 addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), MCPOffset,
1502 unsigned X86FastISel::TargetMaterializeAlloca(AllocaInst *C) {
1503 // Fail on dynamic allocas. At this point, getRegForValue has already
1504 // checked its CSE maps, so if we're here trying to handle a dynamic
1505 // alloca, we're not going to succeed. X86SelectAddress has a
1506 // check for dynamic allocas, because it's called directly from
1507 // various places, but TargetMaterializeAlloca also needs a check
1508 // in order to avoid recursion between getRegForValue,
1509 // X86SelectAddrss, and TargetMaterializeAlloca.
1510 if (!StaticAllocaMap.count(C))
1514 if (!X86SelectAddress(C, AM, false))
1516 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1517 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1518 unsigned ResultReg = createResultReg(RC);
1519 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1524 llvm::FastISel *X86::createFastISel(MachineFunction &mf,
1525 MachineModuleInfo *mmi,
1527 DenseMap<const Value *, unsigned> &vm,
1528 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
1529 DenseMap<const AllocaInst *, int> &am
1531 , SmallSet<Instruction*, 8> &cil
1534 return new X86FastISel(mf, mmi, dw, vm, bm, am