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/ErrorHandling.h"
33 #include "llvm/Support/GetElementPtrTypeIterator.h"
34 #include "llvm/Target/TargetOptions.h"
39 class X86FastISel : public FastISel {
40 /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
41 /// make the right decision when generating code for different targets.
42 const X86Subtarget *Subtarget;
44 /// StackPtr - Register used as the stack pointer.
48 /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
49 /// floating point ops.
50 /// When SSE is available, use it for f32 operations.
51 /// When SSE2 is available, use it for f64 operations.
56 explicit X86FastISel(MachineFunction &mf,
57 MachineModuleInfo *mmi,
59 DenseMap<const Value *, unsigned> &vm,
60 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
61 DenseMap<const AllocaInst *, int> &am
63 , SmallSet<Instruction*, 8> &cil
66 : FastISel(mf, mmi, dw, vm, bm, am
71 Subtarget = &TM.getSubtarget<X86Subtarget>();
72 StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
73 X86ScalarSSEf64 = Subtarget->hasSSE2();
74 X86ScalarSSEf32 = Subtarget->hasSSE1();
77 virtual bool TargetSelectInstruction(Instruction *I);
79 #include "X86GenFastISel.inc"
82 bool X86FastEmitCompare(Value *LHS, Value *RHS, EVT VT);
84 bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
86 bool X86FastEmitStore(EVT VT, Value *Val,
87 const X86AddressMode &AM);
88 bool X86FastEmitStore(EVT VT, unsigned Val,
89 const X86AddressMode &AM);
91 bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
94 bool X86SelectAddress(Value *V, X86AddressMode &AM);
95 bool X86SelectCallAddress(Value *V, X86AddressMode &AM);
97 bool X86SelectLoad(Instruction *I);
99 bool X86SelectStore(Instruction *I);
101 bool X86SelectCmp(Instruction *I);
103 bool X86SelectZExt(Instruction *I);
105 bool X86SelectBranch(Instruction *I);
107 bool X86SelectShift(Instruction *I);
109 bool X86SelectSelect(Instruction *I);
111 bool X86SelectTrunc(Instruction *I);
113 bool X86SelectFPExt(Instruction *I);
114 bool X86SelectFPTrunc(Instruction *I);
116 bool X86SelectExtractValue(Instruction *I);
118 bool X86VisitIntrinsicCall(IntrinsicInst &I);
119 bool X86SelectCall(Instruction *I);
121 CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isTailCall = false);
123 const X86InstrInfo *getInstrInfo() const {
124 return getTargetMachine()->getInstrInfo();
126 const X86TargetMachine *getTargetMachine() const {
127 return static_cast<const X86TargetMachine *>(&TM);
130 unsigned TargetMaterializeConstant(Constant *C);
132 unsigned TargetMaterializeAlloca(AllocaInst *C);
134 /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
135 /// computed in an SSE register, not on the X87 floating point stack.
136 bool isScalarFPTypeInSSEReg(EVT VT) const {
137 return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
138 (VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
141 bool isTypeLegal(const Type *Ty, EVT &VT, bool AllowI1 = false);
144 } // end anonymous namespace.
146 bool X86FastISel::isTypeLegal(const Type *Ty, EVT &VT, bool AllowI1) {
147 VT = TLI.getValueType(Ty, /*HandleUnknown=*/true);
148 if (VT == MVT::Other || !VT.isSimple())
149 // Unhandled type. Halt "fast" selection and bail.
152 // For now, require SSE/SSE2 for performing floating-point operations,
153 // since x87 requires additional work.
154 if (VT == MVT::f64 && !X86ScalarSSEf64)
156 if (VT == MVT::f32 && !X86ScalarSSEf32)
158 // Similarly, no f80 support yet.
161 // We only handle legal types. For example, on x86-32 the instruction
162 // selector contains all of the 64-bit instructions from x86-64,
163 // under the assumption that i64 won't be used if the target doesn't
165 return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
168 #include "X86GenCallingConv.inc"
170 /// CCAssignFnForCall - Selects the correct CCAssignFn for a given calling
172 CCAssignFn *X86FastISel::CCAssignFnForCall(CallingConv::ID CC,
174 if (Subtarget->is64Bit()) {
175 if (Subtarget->isTargetWin64())
176 return CC_X86_Win64_C;
181 if (CC == CallingConv::X86_FastCall)
182 return CC_X86_32_FastCall;
183 else if (CC == CallingConv::Fast)
184 return CC_X86_32_FastCC;
189 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
190 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
191 /// Return true and the result register by reference if it is possible.
192 bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
193 unsigned &ResultReg) {
194 // Get opcode and regclass of the output for the given load instruction.
196 const TargetRegisterClass *RC = NULL;
197 switch (VT.getSimpleVT().SimpleTy) {
198 default: return false;
202 RC = X86::GR8RegisterClass;
206 RC = X86::GR16RegisterClass;
210 RC = X86::GR32RegisterClass;
213 // Must be in x86-64 mode.
215 RC = X86::GR64RegisterClass;
218 if (Subtarget->hasSSE1()) {
220 RC = X86::FR32RegisterClass;
223 RC = X86::RFP32RegisterClass;
227 if (Subtarget->hasSSE2()) {
229 RC = X86::FR64RegisterClass;
232 RC = X86::RFP64RegisterClass;
236 // No f80 support yet.
240 ResultReg = createResultReg(RC);
241 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
245 /// X86FastEmitStore - Emit a machine instruction to store a value Val of
246 /// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
247 /// and a displacement offset, or a GlobalAddress,
248 /// i.e. V. Return true if it is possible.
250 X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
251 const X86AddressMode &AM) {
252 // Get opcode and regclass of the output for the given store instruction.
254 switch (VT.getSimpleVT().SimpleTy) {
255 case MVT::f80: // No f80 support yet.
256 default: return false;
258 // Mask out all but lowest bit.
259 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
261 TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
264 // FALLTHROUGH, handling i1 as i8.
265 case MVT::i8: Opc = X86::MOV8mr; break;
266 case MVT::i16: Opc = X86::MOV16mr; break;
267 case MVT::i32: Opc = X86::MOV32mr; break;
268 case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
270 Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
273 Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
277 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM).addReg(Val);
281 bool X86FastISel::X86FastEmitStore(EVT VT, Value *Val,
282 const X86AddressMode &AM) {
283 // Handle 'null' like i32/i64 0.
284 if (isa<ConstantPointerNull>(Val))
285 Val = Constant::getNullValue(TD.getIntPtrType(Val->getContext()));
287 // If this is a store of a simple constant, fold the constant into the store.
288 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
291 switch (VT.getSimpleVT().SimpleTy) {
293 case MVT::i1: Signed = false; // FALLTHROUGH to handle as i8.
294 case MVT::i8: Opc = X86::MOV8mi; break;
295 case MVT::i16: Opc = X86::MOV16mi; break;
296 case MVT::i32: Opc = X86::MOV32mi; break;
298 // Must be a 32-bit sign extended value.
299 if ((int)CI->getSExtValue() == CI->getSExtValue())
300 Opc = X86::MOV64mi32;
305 addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM)
306 .addImm(Signed ? CI->getSExtValue() :
312 unsigned ValReg = getRegForValue(Val);
316 return X86FastEmitStore(VT, ValReg, AM);
319 /// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
320 /// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
321 /// ISD::SIGN_EXTEND).
322 bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
323 unsigned Src, EVT SrcVT,
324 unsigned &ResultReg) {
325 unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc, Src);
334 /// X86SelectAddress - Attempt to fill in an address from the given value.
336 bool X86FastISel::X86SelectAddress(Value *V, X86AddressMode &AM) {
338 unsigned Opcode = Instruction::UserOp1;
339 if (Instruction *I = dyn_cast<Instruction>(V)) {
340 Opcode = I->getOpcode();
342 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
343 Opcode = C->getOpcode();
349 case Instruction::BitCast:
350 // Look past bitcasts.
351 return X86SelectAddress(U->getOperand(0), AM);
353 case Instruction::IntToPtr:
354 // Look past no-op inttoptrs.
355 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
356 return X86SelectAddress(U->getOperand(0), AM);
359 case Instruction::PtrToInt:
360 // Look past no-op ptrtoints.
361 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
362 return X86SelectAddress(U->getOperand(0), AM);
365 case Instruction::Alloca: {
366 // Do static allocas.
367 const AllocaInst *A = cast<AllocaInst>(V);
368 DenseMap<const AllocaInst*, int>::iterator SI = StaticAllocaMap.find(A);
369 if (SI != StaticAllocaMap.end()) {
370 AM.BaseType = X86AddressMode::FrameIndexBase;
371 AM.Base.FrameIndex = SI->second;
377 case Instruction::Add: {
378 // Adds of constants are common and easy enough.
379 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
380 uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
381 // They have to fit in the 32-bit signed displacement field though.
383 AM.Disp = (uint32_t)Disp;
384 return X86SelectAddress(U->getOperand(0), AM);
390 case Instruction::GetElementPtr: {
391 X86AddressMode SavedAM = AM;
393 // Pattern-match simple GEPs.
394 uint64_t Disp = (int32_t)AM.Disp;
395 unsigned IndexReg = AM.IndexReg;
396 unsigned Scale = AM.Scale;
397 gep_type_iterator GTI = gep_type_begin(U);
398 // Iterate through the indices, folding what we can. Constants can be
399 // folded, and one dynamic index can be handled, if the scale is supported.
400 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end();
401 i != e; ++i, ++GTI) {
403 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
404 const StructLayout *SL = TD.getStructLayout(STy);
405 unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
406 Disp += SL->getElementOffset(Idx);
408 uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
409 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
410 // Constant-offset addressing.
411 Disp += CI->getSExtValue() * S;
412 } else if (IndexReg == 0 &&
413 (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
414 (S == 1 || S == 2 || S == 4 || S == 8)) {
415 // Scaled-index addressing.
417 IndexReg = getRegForGEPIndex(Op);
422 goto unsupported_gep;
425 // Check for displacement overflow.
428 // Ok, the GEP indices were covered by constant-offset and scaled-index
429 // addressing. Update the address state and move on to examining the base.
430 AM.IndexReg = IndexReg;
432 AM.Disp = (uint32_t)Disp;
433 if (X86SelectAddress(U->getOperand(0), AM))
436 // If we couldn't merge the sub value into this addr mode, revert back to
437 // our address and just match the value instead of completely failing.
441 // Ok, the GEP indices weren't all covered.
446 // Handle constant address.
447 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
448 // Can't handle alternate code models yet.
449 if (TM.getCodeModel() != CodeModel::Small)
452 // RIP-relative addresses can't have additional register operands.
453 if (Subtarget->isPICStyleRIPRel() &&
454 (AM.Base.Reg != 0 || AM.IndexReg != 0))
457 // Can't handle TLS yet.
458 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
459 if (GVar->isThreadLocal())
462 // Okay, we've committed to selecting this global. Set up the basic address.
465 // Allow the subtarget to classify the global.
466 unsigned char GVFlags = Subtarget->ClassifyGlobalReference(GV, TM);
468 // If this reference is relative to the pic base, set it now.
469 if (isGlobalRelativeToPICBase(GVFlags)) {
470 // FIXME: How do we know Base.Reg is free??
471 AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(&MF);
474 // Unless the ABI requires an extra load, return a direct reference to
476 if (!isGlobalStubReference(GVFlags)) {
477 if (Subtarget->isPICStyleRIPRel()) {
478 // Use rip-relative addressing if we can. Above we verified that the
479 // base and index registers are unused.
480 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
481 AM.Base.Reg = X86::RIP;
483 AM.GVOpFlags = GVFlags;
487 // Ok, we need to do a load from a stub. If we've already loaded from this
488 // stub, reuse the loaded pointer, otherwise emit the load now.
489 DenseMap<const Value*, unsigned>::iterator I = LocalValueMap.find(V);
491 if (I != LocalValueMap.end() && I->second != 0) {
494 // Issue load from stub.
496 const TargetRegisterClass *RC = NULL;
497 X86AddressMode StubAM;
498 StubAM.Base.Reg = AM.Base.Reg;
500 StubAM.GVOpFlags = GVFlags;
502 if (TLI.getPointerTy() == MVT::i64) {
504 RC = X86::GR64RegisterClass;
506 if (Subtarget->isPICStyleRIPRel())
507 StubAM.Base.Reg = X86::RIP;
510 RC = X86::GR32RegisterClass;
513 LoadReg = createResultReg(RC);
514 addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);
516 // Prevent loading GV stub multiple times in same MBB.
517 LocalValueMap[V] = LoadReg;
520 // Now construct the final address. Note that the Disp, Scale,
521 // and Index values may already be set here.
522 AM.Base.Reg = LoadReg;
527 // If all else fails, try to materialize the value in a register.
528 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
529 if (AM.Base.Reg == 0) {
530 AM.Base.Reg = getRegForValue(V);
531 return AM.Base.Reg != 0;
533 if (AM.IndexReg == 0) {
534 assert(AM.Scale == 1 && "Scale with no index!");
535 AM.IndexReg = getRegForValue(V);
536 return AM.IndexReg != 0;
543 /// X86SelectCallAddress - Attempt to fill in an address from the given value.
545 bool X86FastISel::X86SelectCallAddress(Value *V, X86AddressMode &AM) {
547 unsigned Opcode = Instruction::UserOp1;
548 if (Instruction *I = dyn_cast<Instruction>(V)) {
549 Opcode = I->getOpcode();
551 } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
552 Opcode = C->getOpcode();
558 case Instruction::BitCast:
559 // Look past bitcasts.
560 return X86SelectCallAddress(U->getOperand(0), AM);
562 case Instruction::IntToPtr:
563 // Look past no-op inttoptrs.
564 if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
565 return X86SelectCallAddress(U->getOperand(0), AM);
568 case Instruction::PtrToInt:
569 // Look past no-op ptrtoints.
570 if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
571 return X86SelectCallAddress(U->getOperand(0), AM);
575 // Handle constant address.
576 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
577 // Can't handle alternate code models yet.
578 if (TM.getCodeModel() != CodeModel::Small)
581 // RIP-relative addresses can't have additional register operands.
582 if (Subtarget->isPICStyleRIPRel() &&
583 (AM.Base.Reg != 0 || AM.IndexReg != 0))
586 // Can't handle TLS or DLLImport.
587 if (GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
588 if (GVar->isThreadLocal() || GVar->hasDLLImportLinkage())
591 // Okay, we've committed to selecting this global. Set up the basic address.
594 // No ABI requires an extra load for anything other than DLLImport, which
595 // we rejected above. Return a direct reference to the global.
596 if (Subtarget->isPICStyleRIPRel()) {
597 // Use rip-relative addressing if we can. Above we verified that the
598 // base and index registers are unused.
599 assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
600 AM.Base.Reg = X86::RIP;
601 } else if (Subtarget->isPICStyleStubPIC()) {
602 AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
603 } else if (Subtarget->isPICStyleGOT()) {
604 AM.GVOpFlags = X86II::MO_GOTOFF;
610 // If all else fails, try to materialize the value in a register.
611 if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
612 if (AM.Base.Reg == 0) {
613 AM.Base.Reg = getRegForValue(V);
614 return AM.Base.Reg != 0;
616 if (AM.IndexReg == 0) {
617 assert(AM.Scale == 1 && "Scale with no index!");
618 AM.IndexReg = getRegForValue(V);
619 return AM.IndexReg != 0;
627 /// X86SelectStore - Select and emit code to implement store instructions.
628 bool X86FastISel::X86SelectStore(Instruction* I) {
630 if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
634 if (!X86SelectAddress(I->getOperand(1), AM))
637 return X86FastEmitStore(VT, I->getOperand(0), AM);
640 /// X86SelectLoad - Select and emit code to implement load instructions.
642 bool X86FastISel::X86SelectLoad(Instruction *I) {
644 if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
648 if (!X86SelectAddress(I->getOperand(0), AM))
651 unsigned ResultReg = 0;
652 if (X86FastEmitLoad(VT, AM, ResultReg)) {
653 UpdateValueMap(I, ResultReg);
659 static unsigned X86ChooseCmpOpcode(EVT VT) {
660 switch (VT.getSimpleVT().SimpleTy) {
662 case MVT::i8: return X86::CMP8rr;
663 case MVT::i16: return X86::CMP16rr;
664 case MVT::i32: return X86::CMP32rr;
665 case MVT::i64: return X86::CMP64rr;
666 case MVT::f32: return X86::UCOMISSrr;
667 case MVT::f64: return X86::UCOMISDrr;
671 /// X86ChooseCmpImmediateOpcode - If we have a comparison with RHS as the RHS
672 /// of the comparison, return an opcode that works for the compare (e.g.
673 /// CMP32ri) otherwise return 0.
674 static unsigned X86ChooseCmpImmediateOpcode(EVT VT, ConstantInt *RHSC) {
675 switch (VT.getSimpleVT().SimpleTy) {
676 // Otherwise, we can't fold the immediate into this comparison.
678 case MVT::i8: return X86::CMP8ri;
679 case MVT::i16: return X86::CMP16ri;
680 case MVT::i32: return X86::CMP32ri;
682 // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
684 if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
685 return X86::CMP64ri32;
690 bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, EVT VT) {
691 unsigned Op0Reg = getRegForValue(Op0);
692 if (Op0Reg == 0) return false;
694 // Handle 'null' like i32/i64 0.
695 if (isa<ConstantPointerNull>(Op1))
696 Op1 = Constant::getNullValue(TD.getIntPtrType(Op0->getContext()));
698 // We have two options: compare with register or immediate. If the RHS of
699 // the compare is an immediate that we can fold into this compare, use
700 // CMPri, otherwise use CMPrr.
701 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
702 if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
703 BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
704 .addImm(Op1C->getSExtValue());
709 unsigned CompareOpc = X86ChooseCmpOpcode(VT);
710 if (CompareOpc == 0) return false;
712 unsigned Op1Reg = getRegForValue(Op1);
713 if (Op1Reg == 0) return false;
714 BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
719 bool X86FastISel::X86SelectCmp(Instruction *I) {
720 CmpInst *CI = cast<CmpInst>(I);
723 if (!isTypeLegal(I->getOperand(0)->getType(), VT))
726 unsigned ResultReg = createResultReg(&X86::GR8RegClass);
728 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
729 switch (CI->getPredicate()) {
730 case CmpInst::FCMP_OEQ: {
731 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
734 unsigned EReg = createResultReg(&X86::GR8RegClass);
735 unsigned NPReg = createResultReg(&X86::GR8RegClass);
736 BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
737 BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
739 TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
740 UpdateValueMap(I, ResultReg);
743 case CmpInst::FCMP_UNE: {
744 if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
747 unsigned NEReg = createResultReg(&X86::GR8RegClass);
748 unsigned PReg = createResultReg(&X86::GR8RegClass);
749 BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
750 BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
751 BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
752 UpdateValueMap(I, ResultReg);
755 case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
756 case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
757 case CmpInst::FCMP_OLT: SwapArgs = true; SetCCOpc = X86::SETAr; break;
758 case CmpInst::FCMP_OLE: SwapArgs = true; SetCCOpc = X86::SETAEr; break;
759 case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
760 case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
761 case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr; break;
762 case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
763 case CmpInst::FCMP_UGT: SwapArgs = true; SetCCOpc = X86::SETBr; break;
764 case CmpInst::FCMP_UGE: SwapArgs = true; SetCCOpc = X86::SETBEr; break;
765 case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
766 case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
768 case CmpInst::ICMP_EQ: SwapArgs = false; SetCCOpc = X86::SETEr; break;
769 case CmpInst::ICMP_NE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
770 case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr; break;
771 case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
772 case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr; break;
773 case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
774 case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr; break;
775 case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
776 case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr; break;
777 case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
782 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
786 // Emit a compare of Op0/Op1.
787 if (!X86FastEmitCompare(Op0, Op1, VT))
790 BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
791 UpdateValueMap(I, ResultReg);
795 bool X86FastISel::X86SelectZExt(Instruction *I) {
796 // Handle zero-extension from i1 to i8, which is common.
797 if (I->getType()->isIntegerTy(8) &&
798 I->getOperand(0)->getType()->isIntegerTy(1)) {
799 unsigned ResultReg = getRegForValue(I->getOperand(0));
800 if (ResultReg == 0) return false;
801 // Set the high bits to zero.
802 ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg);
803 if (ResultReg == 0) return false;
804 UpdateValueMap(I, ResultReg);
812 bool X86FastISel::X86SelectBranch(Instruction *I) {
813 // Unconditional branches are selected by tablegen-generated code.
814 // Handle a conditional branch.
815 BranchInst *BI = cast<BranchInst>(I);
816 MachineBasicBlock *TrueMBB = MBBMap[BI->getSuccessor(0)];
817 MachineBasicBlock *FalseMBB = MBBMap[BI->getSuccessor(1)];
819 // Fold the common case of a conditional branch with a comparison.
820 if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
821 if (CI->hasOneUse()) {
822 EVT VT = TLI.getValueType(CI->getOperand(0)->getType());
824 // Try to take advantage of fallthrough opportunities.
825 CmpInst::Predicate Predicate = CI->getPredicate();
826 if (MBB->isLayoutSuccessor(TrueMBB)) {
827 std::swap(TrueMBB, FalseMBB);
828 Predicate = CmpInst::getInversePredicate(Predicate);
831 bool SwapArgs; // false -> compare Op0, Op1. true -> compare Op1, Op0.
832 unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
835 case CmpInst::FCMP_OEQ:
836 std::swap(TrueMBB, FalseMBB);
837 Predicate = CmpInst::FCMP_UNE;
839 case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
840 case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
841 case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
842 case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA_4; break;
843 case CmpInst::FCMP_OLE: SwapArgs = true; BranchOpc = X86::JAE_4; break;
844 case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
845 case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
846 case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4; break;
847 case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
848 case CmpInst::FCMP_UGT: SwapArgs = true; BranchOpc = X86::JB_4; break;
849 case CmpInst::FCMP_UGE: SwapArgs = true; BranchOpc = X86::JBE_4; break;
850 case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
851 case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
853 case CmpInst::ICMP_EQ: SwapArgs = false; BranchOpc = X86::JE_4; break;
854 case CmpInst::ICMP_NE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
855 case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4; break;
856 case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
857 case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4; break;
858 case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
859 case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4; break;
860 case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
861 case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4; break;
862 case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
867 Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
871 // Emit a compare of the LHS and RHS, setting the flags.
872 if (!X86FastEmitCompare(Op0, Op1, VT))
875 BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
877 if (Predicate == CmpInst::FCMP_UNE) {
878 // X86 requires a second branch to handle UNE (and OEQ,
879 // which is mapped to UNE above).
880 BuildMI(MBB, DL, TII.get(X86::JP_4)).addMBB(TrueMBB);
883 FastEmitBranch(FalseMBB);
884 MBB->addSuccessor(TrueMBB);
887 } else if (ExtractValueInst *EI =
888 dyn_cast<ExtractValueInst>(BI->getCondition())) {
889 // Check to see if the branch instruction is from an "arithmetic with
890 // overflow" intrinsic. The main way these intrinsics are used is:
892 // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
893 // %sum = extractvalue { i32, i1 } %t, 0
894 // %obit = extractvalue { i32, i1 } %t, 1
895 // br i1 %obit, label %overflow, label %normal
897 // The %sum and %obit are converted in an ADD and a SETO/SETB before
898 // reaching the branch. Therefore, we search backwards through the MBB
899 // looking for the SETO/SETB instruction. If an instruction modifies the
900 // EFLAGS register before we reach the SETO/SETB instruction, then we can't
901 // convert the branch into a JO/JB instruction.
902 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
903 if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
904 CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
905 const MachineInstr *SetMI = 0;
906 unsigned Reg = lookUpRegForValue(EI);
908 for (MachineBasicBlock::const_reverse_iterator
909 RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
910 const MachineInstr &MI = *RI;
912 if (MI.modifiesRegister(Reg)) {
913 unsigned Src, Dst, SrcSR, DstSR;
915 if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
924 const TargetInstrDesc &TID = MI.getDesc();
925 if (TID.hasUnmodeledSideEffects() ||
926 TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
931 unsigned OpCode = SetMI->getOpcode();
933 if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
934 BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ?
935 X86::JO_4 : X86::JB_4))
937 FastEmitBranch(FalseMBB);
938 MBB->addSuccessor(TrueMBB);
946 // Otherwise do a clumsy setcc and re-test it.
947 unsigned OpReg = getRegForValue(BI->getCondition());
948 if (OpReg == 0) return false;
950 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
951 BuildMI(MBB, DL, TII.get(X86::JNE_4)).addMBB(TrueMBB);
952 FastEmitBranch(FalseMBB);
953 MBB->addSuccessor(TrueMBB);
957 bool X86FastISel::X86SelectShift(Instruction *I) {
958 unsigned CReg = 0, OpReg = 0, OpImm = 0;
959 const TargetRegisterClass *RC = NULL;
960 if (I->getType()->isIntegerTy(8)) {
962 RC = &X86::GR8RegClass;
963 switch (I->getOpcode()) {
964 case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
965 case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
966 case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
967 default: return false;
969 } else if (I->getType()->isIntegerTy(16)) {
971 RC = &X86::GR16RegClass;
972 switch (I->getOpcode()) {
973 case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
974 case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
975 case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
976 default: return false;
978 } else if (I->getType()->isIntegerTy(32)) {
980 RC = &X86::GR32RegClass;
981 switch (I->getOpcode()) {
982 case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
983 case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
984 case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
985 default: return false;
987 } else if (I->getType()->isIntegerTy(64)) {
989 RC = &X86::GR64RegClass;
990 switch (I->getOpcode()) {
991 case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
992 case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
993 case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
994 default: return false;
1000 EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1001 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1004 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1005 if (Op0Reg == 0) return false;
1007 // Fold immediate in shl(x,3).
1008 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
1009 unsigned ResultReg = createResultReg(RC);
1010 BuildMI(MBB, DL, TII.get(OpImm),
1011 ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
1012 UpdateValueMap(I, ResultReg);
1016 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1017 if (Op1Reg == 0) return false;
1018 TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);
1020 // The shift instruction uses X86::CL. If we defined a super-register
1021 // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
1022 // we're doing here.
1023 if (CReg != X86::CL)
1024 BuildMI(MBB, DL, TII.get(TargetOpcode::EXTRACT_SUBREG), X86::CL)
1025 .addReg(CReg).addImm(X86::SUBREG_8BIT);
1027 unsigned ResultReg = createResultReg(RC);
1028 BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
1029 UpdateValueMap(I, ResultReg);
1033 bool X86FastISel::X86SelectSelect(Instruction *I) {
1034 EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
1035 if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
1039 const TargetRegisterClass *RC = NULL;
1040 if (VT.getSimpleVT() == MVT::i16) {
1041 Opc = X86::CMOVE16rr;
1042 RC = &X86::GR16RegClass;
1043 } else if (VT.getSimpleVT() == MVT::i32) {
1044 Opc = X86::CMOVE32rr;
1045 RC = &X86::GR32RegClass;
1046 } else if (VT.getSimpleVT() == MVT::i64) {
1047 Opc = X86::CMOVE64rr;
1048 RC = &X86::GR64RegClass;
1053 unsigned Op0Reg = getRegForValue(I->getOperand(0));
1054 if (Op0Reg == 0) return false;
1055 unsigned Op1Reg = getRegForValue(I->getOperand(1));
1056 if (Op1Reg == 0) return false;
1057 unsigned Op2Reg = getRegForValue(I->getOperand(2));
1058 if (Op2Reg == 0) return false;
1060 BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
1061 unsigned ResultReg = createResultReg(RC);
1062 BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
1063 UpdateValueMap(I, ResultReg);
1067 bool X86FastISel::X86SelectFPExt(Instruction *I) {
1068 // fpext from float to double.
1069 if (Subtarget->hasSSE2() &&
1070 I->getType()->isDoubleTy()) {
1071 Value *V = I->getOperand(0);
1072 if (V->getType()->isFloatTy()) {
1073 unsigned OpReg = getRegForValue(V);
1074 if (OpReg == 0) return false;
1075 unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
1076 BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
1077 UpdateValueMap(I, ResultReg);
1085 bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
1086 if (Subtarget->hasSSE2()) {
1087 if (I->getType()->isFloatTy()) {
1088 Value *V = I->getOperand(0);
1089 if (V->getType()->isDoubleTy()) {
1090 unsigned OpReg = getRegForValue(V);
1091 if (OpReg == 0) return false;
1092 unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
1093 BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
1094 UpdateValueMap(I, ResultReg);
1103 bool X86FastISel::X86SelectTrunc(Instruction *I) {
1104 if (Subtarget->is64Bit())
1105 // All other cases should be handled by the tblgen generated code.
1107 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1108 EVT DstVT = TLI.getValueType(I->getType());
1110 // This code only handles truncation to byte right now.
1111 if (DstVT != MVT::i8 && DstVT != MVT::i1)
1112 // All other cases should be handled by the tblgen generated code.
1114 if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
1115 // All other cases should be handled by the tblgen generated code.
1118 unsigned InputReg = getRegForValue(I->getOperand(0));
1120 // Unhandled operand. Halt "fast" selection and bail.
1123 // First issue a copy to GR16_ABCD or GR32_ABCD.
1124 unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
1125 const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
1126 ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
1127 unsigned CopyReg = createResultReg(CopyRC);
1128 BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
1130 // Then issue an extract_subreg.
1131 unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
1132 CopyReg, X86::SUBREG_8BIT);
1136 UpdateValueMap(I, ResultReg);
1140 bool X86FastISel::X86SelectExtractValue(Instruction *I) {
1141 ExtractValueInst *EI = cast<ExtractValueInst>(I);
1142 Value *Agg = EI->getAggregateOperand();
1144 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
1145 switch (CI->getIntrinsicID()) {
1147 case Intrinsic::sadd_with_overflow:
1148 case Intrinsic::uadd_with_overflow:
1149 // Cheat a little. We know that the registers for "add" and "seto" are
1150 // allocated sequentially. However, we only keep track of the register
1151 // for "add" in the value map. Use extractvalue's index to get the
1152 // correct register for "seto".
1153 UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
1161 bool X86FastISel::X86VisitIntrinsicCall(IntrinsicInst &I) {
1162 // FIXME: Handle more intrinsics.
1163 switch (I.getIntrinsicID()) {
1164 default: return false;
1165 case Intrinsic::dbg_declare: {
1166 DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
1168 assert(DI->getAddress() && "Null address should be checked earlier!");
1169 if (!X86SelectAddress(DI->getAddress(), AM))
1171 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
1172 // FIXME may need to add RegState::Debug to any registers produced,
1173 // although ESP/EBP should be the only ones at the moment.
1174 addFullAddress(BuildMI(MBB, DL, II), AM).addImm(0).
1175 addMetadata(DI->getVariable());
1178 case Intrinsic::trap: {
1179 BuildMI(MBB, DL, TII.get(X86::TRAP));
1182 case Intrinsic::sadd_with_overflow:
1183 case Intrinsic::uadd_with_overflow: {
1184 // Replace "add with overflow" intrinsics with an "add" instruction followed
1185 // by a seto/setc instruction. Later on, when the "extractvalue"
1186 // instructions are encountered, we use the fact that two registers were
1187 // created sequentially to get the correct registers for the "sum" and the
1189 const Function *Callee = I.getCalledFunction();
1191 cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
1194 if (!isTypeLegal(RetTy, VT))
1197 Value *Op1 = I.getOperand(1);
1198 Value *Op2 = I.getOperand(2);
1199 unsigned Reg1 = getRegForValue(Op1);
1200 unsigned Reg2 = getRegForValue(Op2);
1202 if (Reg1 == 0 || Reg2 == 0)
1203 // FIXME: Handle values *not* in registers.
1209 else if (VT == MVT::i64)
1214 unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
1215 BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
1216 unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
1218 // If the add with overflow is an intra-block value then we just want to
1219 // create temporaries for it like normal. If it is a cross-block value then
1220 // UpdateValueMap will return the cross-block register used. Since we
1221 // *really* want the value to be live in the register pair known by
1222 // UpdateValueMap, we have to use DestReg1+1 as the destination register in
1223 // the cross block case. In the non-cross-block case, we should just make
1224 // another register for the value.
1225 if (DestReg1 != ResultReg)
1226 ResultReg = DestReg1+1;
1228 ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
1230 unsigned Opc = X86::SETBr;
1231 if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
1233 BuildMI(MBB, DL, TII.get(Opc), ResultReg);
1239 bool X86FastISel::X86SelectCall(Instruction *I) {
1240 CallInst *CI = cast<CallInst>(I);
1241 Value *Callee = I->getOperand(0);
1243 // Can't handle inline asm yet.
1244 if (isa<InlineAsm>(Callee))
1247 // Handle intrinsic calls.
1248 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
1249 return X86VisitIntrinsicCall(*II);
1251 // Handle only C and fastcc calling conventions for now.
1253 CallingConv::ID CC = CS.getCallingConv();
1254 if (CC != CallingConv::C &&
1255 CC != CallingConv::Fast &&
1256 CC != CallingConv::X86_FastCall)
1259 // fastcc with -tailcallopt is intended to provide a guaranteed
1260 // tail call optimization. Fastisel doesn't know how to do that.
1261 if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
1264 // Let SDISel handle vararg functions.
1265 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
1266 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
1267 if (FTy->isVarArg())
1270 // Handle *simple* calls for now.
1271 const Type *RetTy = CS.getType();
1273 if (RetTy->isVoidTy())
1274 RetVT = MVT::isVoid;
1275 else if (!isTypeLegal(RetTy, RetVT, true))
1278 // Materialize callee address in a register. FIXME: GV address can be
1279 // handled with a CALLpcrel32 instead.
1280 X86AddressMode CalleeAM;
1281 if (!X86SelectCallAddress(Callee, CalleeAM))
1283 unsigned CalleeOp = 0;
1284 GlobalValue *GV = 0;
1285 if (CalleeAM.GV != 0) {
1287 } else if (CalleeAM.Base.Reg != 0) {
1288 CalleeOp = CalleeAM.Base.Reg;
1292 // Allow calls which produce i1 results.
1293 bool AndToI1 = false;
1294 if (RetVT == MVT::i1) {
1299 // Deal with call operands first.
1300 SmallVector<Value*, 8> ArgVals;
1301 SmallVector<unsigned, 8> Args;
1302 SmallVector<EVT, 8> ArgVTs;
1303 SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
1304 Args.reserve(CS.arg_size());
1305 ArgVals.reserve(CS.arg_size());
1306 ArgVTs.reserve(CS.arg_size());
1307 ArgFlags.reserve(CS.arg_size());
1308 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
1310 unsigned Arg = getRegForValue(*i);
1313 ISD::ArgFlagsTy Flags;
1314 unsigned AttrInd = i - CS.arg_begin() + 1;
1315 if (CS.paramHasAttr(AttrInd, Attribute::SExt))
1317 if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
1320 // FIXME: Only handle *easy* calls for now.
1321 if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
1322 CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
1323 CS.paramHasAttr(AttrInd, Attribute::Nest) ||
1324 CS.paramHasAttr(AttrInd, Attribute::ByVal))
1327 const Type *ArgTy = (*i)->getType();
1329 if (!isTypeLegal(ArgTy, ArgVT))
1331 unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
1332 Flags.setOrigAlign(OriginalAlignment);
1334 Args.push_back(Arg);
1335 ArgVals.push_back(*i);
1336 ArgVTs.push_back(ArgVT);
1337 ArgFlags.push_back(Flags);
1340 // Analyze operands of the call, assigning locations to each operand.
1341 SmallVector<CCValAssign, 16> ArgLocs;
1342 CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
1343 CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));
1345 // Get a count of how many bytes are to be pushed on the stack.
1346 unsigned NumBytes = CCInfo.getNextStackOffset();
1348 // Issue CALLSEQ_START
1349 unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
1350 BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
1352 // Process argument: walk the register/memloc assignments, inserting
1354 SmallVector<unsigned, 4> RegArgs;
1355 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1356 CCValAssign &VA = ArgLocs[i];
1357 unsigned Arg = Args[VA.getValNo()];
1358 EVT ArgVT = ArgVTs[VA.getValNo()];
1360 // Promote the value if needed.
1361 switch (VA.getLocInfo()) {
1362 default: llvm_unreachable("Unknown loc info!");
1363 case CCValAssign::Full: break;
1364 case CCValAssign::SExt: {
1365 bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1367 assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
1369 ArgVT = VA.getLocVT();
1372 case CCValAssign::ZExt: {
1373 bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1375 assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
1377 ArgVT = VA.getLocVT();
1380 case CCValAssign::AExt: {
1381 bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
1384 Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
1387 Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
1390 assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
1391 ArgVT = VA.getLocVT();
1394 case CCValAssign::BCvt: {
1395 unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT().getSimpleVT(),
1396 ISD::BIT_CONVERT, Arg);
1397 assert(BC != 0 && "Failed to emit a bitcast!");
1399 ArgVT = VA.getLocVT();
1404 if (VA.isRegLoc()) {
1405 TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
1406 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
1408 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1410 RegArgs.push_back(VA.getLocReg());
1412 unsigned LocMemOffset = VA.getLocMemOffset();
1414 AM.Base.Reg = StackPtr;
1415 AM.Disp = LocMemOffset;
1416 Value *ArgVal = ArgVals[VA.getValNo()];
1418 // If this is a really simple value, emit this with the Value* version of
1419 // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
1420 // can cause us to reevaluate the argument.
1421 if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
1422 X86FastEmitStore(ArgVT, ArgVal, AM);
1424 X86FastEmitStore(ArgVT, Arg, AM);
1428 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1430 if (Subtarget->isPICStyleGOT()) {
1431 TargetRegisterClass *RC = X86::GR32RegisterClass;
1432 unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
1433 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
1434 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1439 MachineInstrBuilder MIB;
1441 // Register-indirect call.
1442 unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
1443 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);
1447 assert(GV && "Not a direct call");
1449 Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;
1451 // See if we need any target-specific flags on the GV operand.
1452 unsigned char OpFlags = 0;
1454 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1455 // external symbols most go through the PLT in PIC mode. If the symbol
1456 // has hidden or protected visibility, or if it is static or local, then
1457 // we don't need to use the PLT - we can directly call it.
1458 if (Subtarget->isTargetELF() &&
1459 TM.getRelocationModel() == Reloc::PIC_ &&
1460 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1461 OpFlags = X86II::MO_PLT;
1462 } else if (Subtarget->isPICStyleStubAny() &&
1463 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1464 Subtarget->getDarwinVers() < 9) {
1465 // PC-relative references to external symbols should go through $stub,
1466 // unless we're building with the leopard linker or later, which
1467 // automatically synthesizes these stubs.
1468 OpFlags = X86II::MO_DARWIN_STUB;
1472 MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
1475 // Add an implicit use GOT pointer in EBX.
1476 if (Subtarget->isPICStyleGOT())
1477 MIB.addReg(X86::EBX);
1479 // Add implicit physical register uses to the call.
1480 for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
1481 MIB.addReg(RegArgs[i]);
1483 // Issue CALLSEQ_END
1484 unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
1485 BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
1487 // Now handle call return value (if any).
1488 if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
1489 SmallVector<CCValAssign, 16> RVLocs;
1490 CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
1491 CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
1493 // Copy all of the result registers out of their specified physreg.
1494 assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
1495 EVT CopyVT = RVLocs[0].getValVT();
1496 TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
1497 TargetRegisterClass *SrcRC = DstRC;
1499 // If this is a call to a function that returns an fp value on the x87 fp
1500 // stack, but where we prefer to use the value in xmm registers, copy it
1501 // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1502 if ((RVLocs[0].getLocReg() == X86::ST0 ||
1503 RVLocs[0].getLocReg() == X86::ST1) &&
1504 isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
1506 SrcRC = X86::RSTRegisterClass;
1507 DstRC = X86::RFP80RegisterClass;
1510 unsigned ResultReg = createResultReg(DstRC);
1511 bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1512 RVLocs[0].getLocReg(), DstRC, SrcRC);
1513 assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
1515 if (CopyVT != RVLocs[0].getValVT()) {
1516 // Round the F80 the right size, which also moves to the appropriate xmm
1517 // register. This is accomplished by storing the F80 value in memory and
1518 // then loading it back. Ewww...
1519 EVT ResVT = RVLocs[0].getValVT();
1520 unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
1521 unsigned MemSize = ResVT.getSizeInBits()/8;
1522 int FI = MFI.CreateStackObject(MemSize, MemSize, false);
1523 addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
1524 DstRC = ResVT == MVT::f32
1525 ? X86::FR32RegisterClass : X86::FR64RegisterClass;
1526 Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
1527 ResultReg = createResultReg(DstRC);
1528 addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
1532 // Mask out all but lowest bit for some call which produces an i1.
1533 unsigned AndResult = createResultReg(X86::GR8RegisterClass);
1535 TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
1536 ResultReg = AndResult;
1539 UpdateValueMap(I, ResultReg);
1547 X86FastISel::TargetSelectInstruction(Instruction *I) {
1548 switch (I->getOpcode()) {
1550 case Instruction::Load:
1551 return X86SelectLoad(I);
1552 case Instruction::Store:
1553 return X86SelectStore(I);
1554 case Instruction::ICmp:
1555 case Instruction::FCmp:
1556 return X86SelectCmp(I);
1557 case Instruction::ZExt:
1558 return X86SelectZExt(I);
1559 case Instruction::Br:
1560 return X86SelectBranch(I);
1561 case Instruction::Call:
1562 return X86SelectCall(I);
1563 case Instruction::LShr:
1564 case Instruction::AShr:
1565 case Instruction::Shl:
1566 return X86SelectShift(I);
1567 case Instruction::Select:
1568 return X86SelectSelect(I);
1569 case Instruction::Trunc:
1570 return X86SelectTrunc(I);
1571 case Instruction::FPExt:
1572 return X86SelectFPExt(I);
1573 case Instruction::FPTrunc:
1574 return X86SelectFPTrunc(I);
1575 case Instruction::ExtractValue:
1576 return X86SelectExtractValue(I);
1577 case Instruction::IntToPtr: // Deliberate fall-through.
1578 case Instruction::PtrToInt: {
1579 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1580 EVT DstVT = TLI.getValueType(I->getType());
1581 if (DstVT.bitsGT(SrcVT))
1582 return X86SelectZExt(I);
1583 if (DstVT.bitsLT(SrcVT))
1584 return X86SelectTrunc(I);
1585 unsigned Reg = getRegForValue(I->getOperand(0));
1586 if (Reg == 0) return false;
1587 UpdateValueMap(I, Reg);
1595 unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
1597 if (!isTypeLegal(C->getType(), VT))
1600 // Get opcode and regclass of the output for the given load instruction.
1602 const TargetRegisterClass *RC = NULL;
1603 switch (VT.getSimpleVT().SimpleTy) {
1604 default: return false;
1607 RC = X86::GR8RegisterClass;
1611 RC = X86::GR16RegisterClass;
1615 RC = X86::GR32RegisterClass;
1618 // Must be in x86-64 mode.
1620 RC = X86::GR64RegisterClass;
1623 if (Subtarget->hasSSE1()) {
1625 RC = X86::FR32RegisterClass;
1627 Opc = X86::LD_Fp32m;
1628 RC = X86::RFP32RegisterClass;
1632 if (Subtarget->hasSSE2()) {
1634 RC = X86::FR64RegisterClass;
1636 Opc = X86::LD_Fp64m;
1637 RC = X86::RFP64RegisterClass;
1641 // No f80 support yet.
1645 // Materialize addresses with LEA instructions.
1646 if (isa<GlobalValue>(C)) {
1648 if (X86SelectAddress(C, AM)) {
1649 if (TLI.getPointerTy() == MVT::i32)
1653 unsigned ResultReg = createResultReg(RC);
1654 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1660 // MachineConstantPool wants an explicit alignment.
1661 unsigned Align = TD.getPrefTypeAlignment(C->getType());
1663 // Alignment of vector types. FIXME!
1664 Align = TD.getTypeAllocSize(C->getType());
1667 // x86-32 PIC requires a PIC base register for constant pools.
1668 unsigned PICBase = 0;
1669 unsigned char OpFlag = 0;
1670 if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
1671 OpFlag = X86II::MO_PIC_BASE_OFFSET;
1672 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1673 } else if (Subtarget->isPICStyleGOT()) {
1674 OpFlag = X86II::MO_GOTOFF;
1675 PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
1676 } else if (Subtarget->isPICStyleRIPRel() &&
1677 TM.getCodeModel() == CodeModel::Small) {
1681 // Create the load from the constant pool.
1682 unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
1683 unsigned ResultReg = createResultReg(RC);
1684 addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
1685 MCPOffset, PICBase, OpFlag);
1690 unsigned X86FastISel::TargetMaterializeAlloca(AllocaInst *C) {
1691 // Fail on dynamic allocas. At this point, getRegForValue has already
1692 // checked its CSE maps, so if we're here trying to handle a dynamic
1693 // alloca, we're not going to succeed. X86SelectAddress has a
1694 // check for dynamic allocas, because it's called directly from
1695 // various places, but TargetMaterializeAlloca also needs a check
1696 // in order to avoid recursion between getRegForValue,
1697 // X86SelectAddrss, and TargetMaterializeAlloca.
1698 if (!StaticAllocaMap.count(C))
1702 if (!X86SelectAddress(C, AM))
1704 unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
1705 TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
1706 unsigned ResultReg = createResultReg(RC);
1707 addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
1712 llvm::FastISel *X86::createFastISel(MachineFunction &mf,
1713 MachineModuleInfo *mmi,
1715 DenseMap<const Value *, unsigned> &vm,
1716 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
1717 DenseMap<const AllocaInst *, int> &am
1719 , SmallSet<Instruction*, 8> &cil
1722 return new X86FastISel(mf, mmi, dw, vm, bm, am