1 //===-- FastISel.cpp - Implementation of the FastISel class ---------------===//
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 contains the implementation of the FastISel class.
12 // "Fast" instruction selection is designed to emit very poor code quickly.
13 // Also, it is not designed to be able to do much lowering, so most illegal
14 // types (e.g. i64 on 32-bit targets) and operations are not supported. It is
15 // also not intended to be able to do much optimization, except in a few cases
16 // where doing optimizations reduces overall compile time. For example, folding
17 // constants into immediate fields is often done, because it's cheap and it
18 // reduces the number of instructions later phases have to examine.
20 // "Fast" instruction selection is able to fail gracefully and transfer
21 // control to the SelectionDAG selector for operations that it doesn't
22 // support. In many cases, this allows us to avoid duplicating a lot of
23 // the complicated lowering logic that SelectionDAG currently has.
25 // The intended use for "fast" instruction selection is "-O0" mode
26 // compilation, where the quality of the generated code is irrelevant when
27 // weighed against the speed at which the code can be generated. Also,
28 // at -O0, the LLVM optimizers are not running, and this makes the
29 // compile time of codegen a much higher portion of the overall compile
30 // time. Despite its limitations, "fast" instruction selection is able to
31 // handle enough code on its own to provide noticeable overall speedups
34 // Basic operations are supported in a target-independent way, by reading
35 // the same instruction descriptions that the SelectionDAG selector reads,
36 // and identifying simple arithmetic operations that can be directly selected
37 // from simple operators. More complicated operations currently require
38 // target-specific code.
40 //===----------------------------------------------------------------------===//
42 #include "llvm/Function.h"
43 #include "llvm/GlobalVariable.h"
44 #include "llvm/Instructions.h"
45 #include "llvm/IntrinsicInst.h"
46 #include "llvm/Operator.h"
47 #include "llvm/CodeGen/Analysis.h"
48 #include "llvm/CodeGen/FastISel.h"
49 #include "llvm/CodeGen/FunctionLoweringInfo.h"
50 #include "llvm/CodeGen/MachineInstrBuilder.h"
51 #include "llvm/CodeGen/MachineModuleInfo.h"
52 #include "llvm/CodeGen/MachineRegisterInfo.h"
53 #include "llvm/Analysis/DebugInfo.h"
54 #include "llvm/Analysis/Loads.h"
55 #include "llvm/Target/TargetData.h"
56 #include "llvm/Target/TargetInstrInfo.h"
57 #include "llvm/Target/TargetLowering.h"
58 #include "llvm/Target/TargetMachine.h"
59 #include "llvm/Support/ErrorHandling.h"
60 #include "llvm/Support/Debug.h"
63 /// startNewBlock - Set the current block to which generated machine
64 /// instructions will be appended, and clear the local CSE map.
66 void FastISel::startNewBlock() {
67 LocalValueMap.clear();
69 // Start out as null, meaining no local-value instructions have
73 // Advance the last local value past any EH_LABEL instructions.
74 MachineBasicBlock::iterator
75 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
76 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
82 bool FastISel::hasTrivialKill(const Value *V) const {
83 // Don't consider constants or arguments to have trivial kills.
84 const Instruction *I = dyn_cast<Instruction>(V);
88 // No-op casts are trivially coalesced by fast-isel.
89 if (const CastInst *Cast = dyn_cast<CastInst>(I))
90 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
91 !hasTrivialKill(Cast->getOperand(0)))
94 // Only instructions with a single use in the same basic block are considered
95 // to have trivial kills.
96 return I->hasOneUse() &&
97 !(I->getOpcode() == Instruction::BitCast ||
98 I->getOpcode() == Instruction::PtrToInt ||
99 I->getOpcode() == Instruction::IntToPtr) &&
100 cast<Instruction>(*I->use_begin())->getParent() == I->getParent();
103 unsigned FastISel::getRegForValue(const Value *V) {
104 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
105 // Don't handle non-simple values in FastISel.
106 if (!RealVT.isSimple())
109 // Ignore illegal types. We must do this before looking up the value
110 // in ValueMap because Arguments are given virtual registers regardless
111 // of whether FastISel can handle them.
112 MVT VT = RealVT.getSimpleVT();
113 if (!TLI.isTypeLegal(VT)) {
114 // Handle integer promotions, though, because they're common and easy.
115 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
116 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
121 // Look up the value to see if we already have a register for it. We
122 // cache values defined by Instructions across blocks, and other values
123 // only locally. This is because Instructions already have the SSA
124 // def-dominates-use requirement enforced.
125 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
126 if (I != FuncInfo.ValueMap.end())
129 unsigned Reg = LocalValueMap[V];
133 // In bottom-up mode, just create the virtual register which will be used
134 // to hold the value. It will be materialized later.
135 if (isa<Instruction>(V) &&
136 (!isa<AllocaInst>(V) ||
137 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
138 return FuncInfo.InitializeRegForValue(V);
140 SavePoint SaveInsertPt = enterLocalValueArea();
142 // Materialize the value in a register. Emit any instructions in the
144 Reg = materializeRegForValue(V, VT);
146 leaveLocalValueArea(SaveInsertPt);
151 /// materializeRegForValue - Helper for getRegForValue. This function is
152 /// called when the value isn't already available in a register and must
153 /// be materialized with new instructions.
154 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
157 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
158 if (CI->getValue().getActiveBits() <= 64)
159 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
160 } else if (isa<AllocaInst>(V)) {
161 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
162 } else if (isa<ConstantPointerNull>(V)) {
163 // Translate this as an integer zero so that it can be
164 // local-CSE'd with actual integer zeros.
166 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
167 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
168 if (CF->isNullValue()) {
169 Reg = TargetMaterializeFloatZero(CF);
171 // Try to emit the constant directly.
172 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
176 // Try to emit the constant by using an integer constant with a cast.
177 const APFloat &Flt = CF->getValueAPF();
178 EVT IntVT = TLI.getPointerTy();
181 uint32_t IntBitWidth = IntVT.getSizeInBits();
183 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
184 APFloat::rmTowardZero, &isExact);
186 APInt IntVal(IntBitWidth, 2, x);
188 unsigned IntegerReg =
189 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
191 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
192 IntegerReg, /*Kill=*/false);
195 } else if (const Operator *Op = dyn_cast<Operator>(V)) {
196 if (!SelectOperator(Op, Op->getOpcode()))
197 if (!isa<Instruction>(Op) ||
198 !TargetSelectInstruction(cast<Instruction>(Op)))
200 Reg = lookUpRegForValue(Op);
201 } else if (isa<UndefValue>(V)) {
202 Reg = createResultReg(TLI.getRegClassFor(VT));
203 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
204 TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
207 // If target-independent code couldn't handle the value, give target-specific
209 if (!Reg && isa<Constant>(V))
210 Reg = TargetMaterializeConstant(cast<Constant>(V));
212 // Don't cache constant materializations in the general ValueMap.
213 // To do so would require tracking what uses they dominate.
215 LocalValueMap[V] = Reg;
216 LastLocalValue = MRI.getVRegDef(Reg);
221 unsigned FastISel::lookUpRegForValue(const Value *V) {
222 // Look up the value to see if we already have a register for it. We
223 // cache values defined by Instructions across blocks, and other values
224 // only locally. This is because Instructions already have the SSA
225 // def-dominates-use requirement enforced.
226 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
227 if (I != FuncInfo.ValueMap.end())
229 return LocalValueMap[V];
232 /// UpdateValueMap - Update the value map to include the new mapping for this
233 /// instruction, or insert an extra copy to get the result in a previous
234 /// determined register.
235 /// NOTE: This is only necessary because we might select a block that uses
236 /// a value before we select the block that defines the value. It might be
237 /// possible to fix this by selecting blocks in reverse postorder.
238 void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) {
239 if (!isa<Instruction>(I)) {
240 LocalValueMap[I] = Reg;
244 unsigned &AssignedReg = FuncInfo.ValueMap[I];
245 if (AssignedReg == 0)
246 // Use the new register.
248 else if (Reg != AssignedReg) {
249 // Arrange for uses of AssignedReg to be replaced by uses of Reg.
250 for (unsigned i = 0; i < NumRegs; i++)
251 FuncInfo.RegFixups[AssignedReg+i] = Reg+i;
257 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
258 unsigned IdxN = getRegForValue(Idx);
260 // Unhandled operand. Halt "fast" selection and bail.
261 return std::pair<unsigned, bool>(0, false);
263 bool IdxNIsKill = hasTrivialKill(Idx);
265 // If the index is smaller or larger than intptr_t, truncate or extend it.
266 MVT PtrVT = TLI.getPointerTy();
267 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
268 if (IdxVT.bitsLT(PtrVT)) {
269 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
273 else if (IdxVT.bitsGT(PtrVT)) {
274 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
278 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
281 void FastISel::recomputeInsertPt() {
282 if (getLastLocalValue()) {
283 FuncInfo.InsertPt = getLastLocalValue();
284 FuncInfo.MBB = FuncInfo.InsertPt->getParent();
287 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI();
289 // Now skip past any EH_LABELs, which must remain at the beginning.
290 while (FuncInfo.InsertPt != FuncInfo.MBB->end() &&
291 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL)
295 FastISel::SavePoint FastISel::enterLocalValueArea() {
296 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
300 SavePoint SP = { OldInsertPt, OldDL };
304 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
305 if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
306 LastLocalValue = llvm::prior(FuncInfo.InsertPt);
308 // Restore the previous insert position.
309 FuncInfo.InsertPt = OldInsertPt.InsertPt;
313 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
314 /// which has an opcode which directly corresponds to the given ISD opcode.
316 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
317 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
318 if (VT == MVT::Other || !VT.isSimple())
319 // Unhandled type. Halt "fast" selection and bail.
322 // We only handle legal types. For example, on x86-32 the instruction
323 // selector contains all of the 64-bit instructions from x86-64,
324 // under the assumption that i64 won't be used if the target doesn't
326 if (!TLI.isTypeLegal(VT)) {
327 // MVT::i1 is special. Allow AND, OR, or XOR because they
328 // don't require additional zeroing, which makes them easy.
330 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
331 ISDOpcode == ISD::XOR))
332 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
337 // Check if the first operand is a constant, and handle it as "ri". At -O0,
338 // we don't have anything that canonicalizes operand order.
339 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0)))
340 if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) {
341 unsigned Op1 = getRegForValue(I->getOperand(1));
342 if (Op1 == 0) return false;
344 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
346 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1,
347 Op1IsKill, CI->getZExtValue(),
349 if (ResultReg == 0) return false;
351 // We successfully emitted code for the given LLVM Instruction.
352 UpdateValueMap(I, ResultReg);
357 unsigned Op0 = getRegForValue(I->getOperand(0));
358 if (Op0 == 0) // Unhandled operand. Halt "fast" selection and bail.
361 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
363 // Check if the second operand is a constant and handle it appropriately.
364 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
365 uint64_t Imm = CI->getZExtValue();
367 // Transform "sdiv exact X, 8" -> "sra X, 3".
368 if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) &&
369 cast<BinaryOperator>(I)->isExact() &&
370 isPowerOf2_64(Imm)) {
372 ISDOpcode = ISD::SRA;
375 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0,
376 Op0IsKill, Imm, VT.getSimpleVT());
377 if (ResultReg == 0) return false;
379 // We successfully emitted code for the given LLVM Instruction.
380 UpdateValueMap(I, ResultReg);
384 // Check if the second operand is a constant float.
385 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
386 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
387 ISDOpcode, Op0, Op0IsKill, CF);
388 if (ResultReg != 0) {
389 // We successfully emitted code for the given LLVM Instruction.
390 UpdateValueMap(I, ResultReg);
395 unsigned Op1 = getRegForValue(I->getOperand(1));
397 // Unhandled operand. Halt "fast" selection and bail.
400 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
402 // Now we have both operands in registers. Emit the instruction.
403 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
408 // Target-specific code wasn't able to find a machine opcode for
409 // the given ISD opcode and type. Halt "fast" selection and bail.
412 // We successfully emitted code for the given LLVM Instruction.
413 UpdateValueMap(I, ResultReg);
417 bool FastISel::SelectGetElementPtr(const User *I) {
418 unsigned N = getRegForValue(I->getOperand(0));
420 // Unhandled operand. Halt "fast" selection and bail.
423 bool NIsKill = hasTrivialKill(I->getOperand(0));
425 const Type *Ty = I->getOperand(0)->getType();
426 MVT VT = TLI.getPointerTy();
427 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
428 E = I->op_end(); OI != E; ++OI) {
429 const Value *Idx = *OI;
430 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
431 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
434 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
435 // FIXME: This can be optimized by combining the add with a
437 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
439 // Unhandled operand. Halt "fast" selection and bail.
443 Ty = StTy->getElementType(Field);
445 Ty = cast<SequentialType>(Ty)->getElementType();
447 // If this is a constant subscript, handle it quickly.
448 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
449 if (CI->isZero()) continue;
451 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
452 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
454 // Unhandled operand. Halt "fast" selection and bail.
460 // N = N + Idx * ElementSize;
461 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
462 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
463 unsigned IdxN = Pair.first;
464 bool IdxNIsKill = Pair.second;
466 // Unhandled operand. Halt "fast" selection and bail.
469 if (ElementSize != 1) {
470 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
472 // Unhandled operand. Halt "fast" selection and bail.
476 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
478 // Unhandled operand. Halt "fast" selection and bail.
483 // We successfully emitted code for the given LLVM Instruction.
484 UpdateValueMap(I, N);
488 bool FastISel::SelectCall(const User *I) {
489 const CallInst *Call = cast<CallInst>(I);
491 // Handle simple inline asms.
492 if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getArgOperand(0))) {
493 // Don't attempt to handle constraints.
494 if (!IA->getConstraintString().empty())
497 unsigned ExtraInfo = 0;
498 if (IA->hasSideEffects())
499 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
500 if (IA->isAlignStack())
501 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
503 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
504 TII.get(TargetOpcode::INLINEASM))
505 .addExternalSymbol(IA->getAsmString().c_str())
510 const Function *F = Call->getCalledFunction();
511 if (!F) return false;
513 // Handle selected intrinsic function calls.
514 switch (F->getIntrinsicID()) {
516 case Intrinsic::dbg_declare: {
517 const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call);
518 if (!DIVariable(DI->getVariable()).Verify() ||
519 !FuncInfo.MF->getMMI().hasDebugInfo())
522 const Value *Address = DI->getAddress();
523 if (!Address || isa<UndefValue>(Address) || isa<AllocaInst>(Address))
528 if (const Argument *Arg = dyn_cast<Argument>(Address)) {
529 if (Arg->hasByValAttr()) {
530 // Byval arguments' frame index is recorded during argument lowering.
531 // Use this info directly.
532 Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
534 Reg = TRI.getFrameRegister(*FuncInfo.MF);
538 Reg = getRegForValue(Address);
541 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
542 TII.get(TargetOpcode::DBG_VALUE))
543 .addReg(Reg, RegState::Debug).addImm(Offset)
544 .addMetadata(DI->getVariable());
547 case Intrinsic::dbg_value: {
548 // This form of DBG_VALUE is target-independent.
549 const DbgValueInst *DI = cast<DbgValueInst>(Call);
550 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
551 const Value *V = DI->getValue();
553 // Currently the optimizer can produce this; insert an undef to
554 // help debugging. Probably the optimizer should not do this.
555 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
556 .addReg(0U).addImm(DI->getOffset())
557 .addMetadata(DI->getVariable());
558 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
559 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
560 .addImm(CI->getZExtValue()).addImm(DI->getOffset())
561 .addMetadata(DI->getVariable());
562 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
563 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
564 .addFPImm(CF).addImm(DI->getOffset())
565 .addMetadata(DI->getVariable());
566 } else if (unsigned Reg = lookUpRegForValue(V)) {
567 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
568 .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
569 .addMetadata(DI->getVariable());
571 // We can't yet handle anything else here because it would require
572 // generating code, thus altering codegen because of debug info.
573 DEBUG(dbgs() << "Dropping debug info for " << DI);
577 case Intrinsic::eh_exception: {
578 EVT VT = TLI.getValueType(Call->getType());
579 if (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)!=TargetLowering::Expand)
582 assert(FuncInfo.MBB->isLandingPad() &&
583 "Call to eh.exception not in landing pad!");
584 unsigned Reg = TLI.getExceptionAddressRegister();
585 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
586 unsigned ResultReg = createResultReg(RC);
587 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
588 ResultReg).addReg(Reg);
589 UpdateValueMap(Call, ResultReg);
592 case Intrinsic::eh_selector: {
593 EVT VT = TLI.getValueType(Call->getType());
594 if (TLI.getOperationAction(ISD::EHSELECTION, VT) != TargetLowering::Expand)
596 if (FuncInfo.MBB->isLandingPad())
597 AddCatchInfo(*Call, &FuncInfo.MF->getMMI(), FuncInfo.MBB);
600 FuncInfo.CatchInfoLost.insert(Call);
602 // FIXME: Mark exception selector register as live in. Hack for PR1508.
603 unsigned Reg = TLI.getExceptionSelectorRegister();
604 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
607 unsigned Reg = TLI.getExceptionSelectorRegister();
608 EVT SrcVT = TLI.getPointerTy();
609 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
610 unsigned ResultReg = createResultReg(RC);
611 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
612 ResultReg).addReg(Reg);
614 bool ResultRegIsKill = hasTrivialKill(Call);
616 // Cast the register to the type of the selector.
617 if (SrcVT.bitsGT(MVT::i32))
618 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
619 ResultReg, ResultRegIsKill);
620 else if (SrcVT.bitsLT(MVT::i32))
621 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
622 ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill);
624 // Unhandled operand. Halt "fast" selection and bail.
627 UpdateValueMap(Call, ResultReg);
631 case Intrinsic::objectsize: {
632 ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1));
633 unsigned long long Res = CI->isZero() ? -1ULL : 0;
634 Constant *ResCI = ConstantInt::get(Call->getType(), Res);
635 unsigned ResultReg = getRegForValue(ResCI);
638 UpdateValueMap(Call, ResultReg);
643 // An arbitrary call. Bail.
647 bool FastISel::SelectCast(const User *I, unsigned Opcode) {
648 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
649 EVT DstVT = TLI.getValueType(I->getType());
651 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
652 DstVT == MVT::Other || !DstVT.isSimple())
653 // Unhandled type. Halt "fast" selection and bail.
656 // Check if the destination type is legal.
657 if (!TLI.isTypeLegal(DstVT))
660 // Check if the source operand is legal.
661 if (!TLI.isTypeLegal(SrcVT))
664 unsigned InputReg = getRegForValue(I->getOperand(0));
666 // Unhandled operand. Halt "fast" selection and bail.
669 bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
671 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
674 InputReg, InputRegIsKill);
678 UpdateValueMap(I, ResultReg);
682 bool FastISel::SelectBitCast(const User *I) {
683 // If the bitcast doesn't change the type, just use the operand value.
684 if (I->getType() == I->getOperand(0)->getType()) {
685 unsigned Reg = getRegForValue(I->getOperand(0));
688 UpdateValueMap(I, Reg);
692 // Bitcasts of other values become reg-reg copies or BITCAST operators.
693 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
694 EVT DstVT = TLI.getValueType(I->getType());
696 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
697 DstVT == MVT::Other || !DstVT.isSimple() ||
698 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
699 // Unhandled type. Halt "fast" selection and bail.
702 unsigned Op0 = getRegForValue(I->getOperand(0));
704 // Unhandled operand. Halt "fast" selection and bail.
707 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
709 // First, try to perform the bitcast by inserting a reg-reg copy.
710 unsigned ResultReg = 0;
711 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
712 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
713 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
714 // Don't attempt a cross-class copy. It will likely fail.
715 if (SrcClass == DstClass) {
716 ResultReg = createResultReg(DstClass);
717 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
718 ResultReg).addReg(Op0);
722 // If the reg-reg copy failed, select a BITCAST opcode.
724 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
725 ISD::BITCAST, Op0, Op0IsKill);
730 UpdateValueMap(I, ResultReg);
735 FastISel::SelectInstruction(const Instruction *I) {
736 // Just before the terminator instruction, insert instructions to
737 // feed PHI nodes in successor blocks.
738 if (isa<TerminatorInst>(I))
739 if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
742 DL = I->getDebugLoc();
744 // First, try doing target-independent selection.
745 if (SelectOperator(I, I->getOpcode())) {
750 // Next, try calling the target to attempt to handle the instruction.
751 if (TargetSelectInstruction(I)) {
760 /// FastEmitBranch - Emit an unconditional branch to the given block,
761 /// unless it is the immediate (fall-through) successor, and update
764 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
765 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
766 // The unconditional fall-through case, which needs no instructions.
768 // The unconditional branch case.
769 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
770 SmallVector<MachineOperand, 0>(), DL);
772 FuncInfo.MBB->addSuccessor(MSucc);
775 /// SelectFNeg - Emit an FNeg operation.
778 FastISel::SelectFNeg(const User *I) {
779 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
780 if (OpReg == 0) return false;
782 bool OpRegIsKill = hasTrivialKill(I);
784 // If the target has ISD::FNEG, use it.
785 EVT VT = TLI.getValueType(I->getType());
786 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
787 ISD::FNEG, OpReg, OpRegIsKill);
788 if (ResultReg != 0) {
789 UpdateValueMap(I, ResultReg);
793 // Bitcast the value to integer, twiddle the sign bit with xor,
794 // and then bitcast it back to floating-point.
795 if (VT.getSizeInBits() > 64) return false;
796 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
797 if (!TLI.isTypeLegal(IntVT))
800 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
801 ISD::BITCAST, OpReg, OpRegIsKill);
805 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
806 IntReg, /*Kill=*/true,
807 UINT64_C(1) << (VT.getSizeInBits()-1),
808 IntVT.getSimpleVT());
809 if (IntResultReg == 0)
812 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
813 ISD::BITCAST, IntResultReg, /*Kill=*/true);
817 UpdateValueMap(I, ResultReg);
822 FastISel::SelectExtractValue(const User *U) {
823 const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
827 // Make sure we only try to handle extracts with a legal result. But also
828 // allow i1 because it's easy.
829 EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true);
830 if (!RealVT.isSimple())
832 MVT VT = RealVT.getSimpleVT();
833 if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
836 const Value *Op0 = EVI->getOperand(0);
837 const Type *AggTy = Op0->getType();
839 // Get the base result register.
841 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0);
842 if (I != FuncInfo.ValueMap.end())
843 ResultReg = I->second;
844 else if (isa<Instruction>(Op0))
845 ResultReg = FuncInfo.InitializeRegForValue(Op0);
847 return false; // fast-isel can't handle aggregate constants at the moment
849 // Get the actual result register, which is an offset from the base register.
850 unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->idx_begin(), EVI->idx_end());
852 SmallVector<EVT, 4> AggValueVTs;
853 ComputeValueVTs(TLI, AggTy, AggValueVTs);
855 for (unsigned i = 0; i < VTIndex; i++)
856 ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]);
858 UpdateValueMap(EVI, ResultReg);
863 FastISel::SelectOperator(const User *I, unsigned Opcode) {
865 case Instruction::Add:
866 return SelectBinaryOp(I, ISD::ADD);
867 case Instruction::FAdd:
868 return SelectBinaryOp(I, ISD::FADD);
869 case Instruction::Sub:
870 return SelectBinaryOp(I, ISD::SUB);
871 case Instruction::FSub:
872 // FNeg is currently represented in LLVM IR as a special case of FSub.
873 if (BinaryOperator::isFNeg(I))
874 return SelectFNeg(I);
875 return SelectBinaryOp(I, ISD::FSUB);
876 case Instruction::Mul:
877 return SelectBinaryOp(I, ISD::MUL);
878 case Instruction::FMul:
879 return SelectBinaryOp(I, ISD::FMUL);
880 case Instruction::SDiv:
881 return SelectBinaryOp(I, ISD::SDIV);
882 case Instruction::UDiv:
883 return SelectBinaryOp(I, ISD::UDIV);
884 case Instruction::FDiv:
885 return SelectBinaryOp(I, ISD::FDIV);
886 case Instruction::SRem:
887 return SelectBinaryOp(I, ISD::SREM);
888 case Instruction::URem:
889 return SelectBinaryOp(I, ISD::UREM);
890 case Instruction::FRem:
891 return SelectBinaryOp(I, ISD::FREM);
892 case Instruction::Shl:
893 return SelectBinaryOp(I, ISD::SHL);
894 case Instruction::LShr:
895 return SelectBinaryOp(I, ISD::SRL);
896 case Instruction::AShr:
897 return SelectBinaryOp(I, ISD::SRA);
898 case Instruction::And:
899 return SelectBinaryOp(I, ISD::AND);
900 case Instruction::Or:
901 return SelectBinaryOp(I, ISD::OR);
902 case Instruction::Xor:
903 return SelectBinaryOp(I, ISD::XOR);
905 case Instruction::GetElementPtr:
906 return SelectGetElementPtr(I);
908 case Instruction::Br: {
909 const BranchInst *BI = cast<BranchInst>(I);
911 if (BI->isUnconditional()) {
912 const BasicBlock *LLVMSucc = BI->getSuccessor(0);
913 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
914 FastEmitBranch(MSucc, BI->getDebugLoc());
918 // Conditional branches are not handed yet.
919 // Halt "fast" selection and bail.
923 case Instruction::Unreachable:
927 case Instruction::Alloca:
928 // FunctionLowering has the static-sized case covered.
929 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
932 // Dynamic-sized alloca is not handled yet.
935 case Instruction::Call:
936 return SelectCall(I);
938 case Instruction::BitCast:
939 return SelectBitCast(I);
941 case Instruction::FPToSI:
942 return SelectCast(I, ISD::FP_TO_SINT);
943 case Instruction::ZExt:
944 return SelectCast(I, ISD::ZERO_EXTEND);
945 case Instruction::SExt:
946 return SelectCast(I, ISD::SIGN_EXTEND);
947 case Instruction::Trunc:
948 return SelectCast(I, ISD::TRUNCATE);
949 case Instruction::SIToFP:
950 return SelectCast(I, ISD::SINT_TO_FP);
952 case Instruction::IntToPtr: // Deliberate fall-through.
953 case Instruction::PtrToInt: {
954 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
955 EVT DstVT = TLI.getValueType(I->getType());
956 if (DstVT.bitsGT(SrcVT))
957 return SelectCast(I, ISD::ZERO_EXTEND);
958 if (DstVT.bitsLT(SrcVT))
959 return SelectCast(I, ISD::TRUNCATE);
960 unsigned Reg = getRegForValue(I->getOperand(0));
961 if (Reg == 0) return false;
962 UpdateValueMap(I, Reg);
966 case Instruction::ExtractValue:
967 return SelectExtractValue(I);
969 case Instruction::PHI:
970 llvm_unreachable("FastISel shouldn't visit PHI nodes!");
973 // Unhandled instruction. Halt "fast" selection and bail.
978 FastISel::FastISel(FunctionLoweringInfo &funcInfo)
979 : FuncInfo(funcInfo),
980 MRI(FuncInfo.MF->getRegInfo()),
981 MFI(*FuncInfo.MF->getFrameInfo()),
982 MCP(*FuncInfo.MF->getConstantPool()),
983 TM(FuncInfo.MF->getTarget()),
984 TD(*TM.getTargetData()),
985 TII(*TM.getInstrInfo()),
986 TLI(*TM.getTargetLowering()),
987 TRI(*TM.getRegisterInfo()) {
990 FastISel::~FastISel() {}
992 unsigned FastISel::FastEmit_(MVT, MVT,
997 unsigned FastISel::FastEmit_r(MVT, MVT,
999 unsigned /*Op0*/, bool /*Op0IsKill*/) {
1003 unsigned FastISel::FastEmit_rr(MVT, MVT,
1005 unsigned /*Op0*/, bool /*Op0IsKill*/,
1006 unsigned /*Op1*/, bool /*Op1IsKill*/) {
1010 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
1014 unsigned FastISel::FastEmit_f(MVT, MVT,
1015 unsigned, const ConstantFP * /*FPImm*/) {
1019 unsigned FastISel::FastEmit_ri(MVT, MVT,
1021 unsigned /*Op0*/, bool /*Op0IsKill*/,
1026 unsigned FastISel::FastEmit_rf(MVT, MVT,
1028 unsigned /*Op0*/, bool /*Op0IsKill*/,
1029 const ConstantFP * /*FPImm*/) {
1033 unsigned FastISel::FastEmit_rri(MVT, MVT,
1035 unsigned /*Op0*/, bool /*Op0IsKill*/,
1036 unsigned /*Op1*/, bool /*Op1IsKill*/,
1041 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
1042 /// to emit an instruction with an immediate operand using FastEmit_ri.
1043 /// If that fails, it materializes the immediate into a register and try
1044 /// FastEmit_rr instead.
1045 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
1046 unsigned Op0, bool Op0IsKill,
1047 uint64_t Imm, MVT ImmType) {
1048 // If this is a multiply by a power of two, emit this as a shift left.
1049 if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) {
1052 } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) {
1053 // div x, 8 -> srl x, 3
1058 // Horrible hack (to be removed), check to make sure shift amounts are
1060 if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) &&
1061 Imm >= VT.getSizeInBits())
1064 // First check if immediate type is legal. If not, we can't use the ri form.
1065 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
1068 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
1069 if (MaterialReg == 0) {
1070 // This is a bit ugly/slow, but failing here means falling out of
1071 // fast-isel, which would be very slow.
1072 const IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
1073 VT.getSizeInBits());
1074 MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
1076 return FastEmit_rr(VT, VT, Opcode,
1078 MaterialReg, /*Kill=*/true);
1081 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
1082 return MRI.createVirtualRegister(RC);
1085 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
1086 const TargetRegisterClass* RC) {
1087 unsigned ResultReg = createResultReg(RC);
1088 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1090 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
1094 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
1095 const TargetRegisterClass *RC,
1096 unsigned Op0, bool Op0IsKill) {
1097 unsigned ResultReg = createResultReg(RC);
1098 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1100 if (II.getNumDefs() >= 1)
1101 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1102 .addReg(Op0, Op0IsKill * RegState::Kill);
1104 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1105 .addReg(Op0, Op0IsKill * RegState::Kill);
1106 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1107 ResultReg).addReg(II.ImplicitDefs[0]);
1113 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
1114 const TargetRegisterClass *RC,
1115 unsigned Op0, bool Op0IsKill,
1116 unsigned Op1, bool Op1IsKill) {
1117 unsigned ResultReg = createResultReg(RC);
1118 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1120 if (II.getNumDefs() >= 1)
1121 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1122 .addReg(Op0, Op0IsKill * RegState::Kill)
1123 .addReg(Op1, Op1IsKill * RegState::Kill);
1125 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1126 .addReg(Op0, Op0IsKill * RegState::Kill)
1127 .addReg(Op1, Op1IsKill * RegState::Kill);
1128 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1129 ResultReg).addReg(II.ImplicitDefs[0]);
1134 unsigned FastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
1135 const TargetRegisterClass *RC,
1136 unsigned Op0, bool Op0IsKill,
1137 unsigned Op1, bool Op1IsKill,
1138 unsigned Op2, bool Op2IsKill) {
1139 unsigned ResultReg = createResultReg(RC);
1140 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1142 if (II.getNumDefs() >= 1)
1143 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1144 .addReg(Op0, Op0IsKill * RegState::Kill)
1145 .addReg(Op1, Op1IsKill * RegState::Kill)
1146 .addReg(Op2, Op2IsKill * RegState::Kill);
1148 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1149 .addReg(Op0, Op0IsKill * RegState::Kill)
1150 .addReg(Op1, Op1IsKill * RegState::Kill)
1151 .addReg(Op2, Op2IsKill * RegState::Kill);
1152 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1153 ResultReg).addReg(II.ImplicitDefs[0]);
1158 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
1159 const TargetRegisterClass *RC,
1160 unsigned Op0, bool Op0IsKill,
1162 unsigned ResultReg = createResultReg(RC);
1163 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1165 if (II.getNumDefs() >= 1)
1166 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1167 .addReg(Op0, Op0IsKill * RegState::Kill)
1170 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1171 .addReg(Op0, Op0IsKill * RegState::Kill)
1173 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1174 ResultReg).addReg(II.ImplicitDefs[0]);
1179 unsigned FastISel::FastEmitInst_rii(unsigned MachineInstOpcode,
1180 const TargetRegisterClass *RC,
1181 unsigned Op0, bool Op0IsKill,
1182 uint64_t Imm1, uint64_t Imm2) {
1183 unsigned ResultReg = createResultReg(RC);
1184 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1186 if (II.getNumDefs() >= 1)
1187 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1188 .addReg(Op0, Op0IsKill * RegState::Kill)
1192 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1193 .addReg(Op0, Op0IsKill * RegState::Kill)
1196 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1197 ResultReg).addReg(II.ImplicitDefs[0]);
1202 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
1203 const TargetRegisterClass *RC,
1204 unsigned Op0, bool Op0IsKill,
1205 const ConstantFP *FPImm) {
1206 unsigned ResultReg = createResultReg(RC);
1207 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1209 if (II.getNumDefs() >= 1)
1210 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1211 .addReg(Op0, Op0IsKill * RegState::Kill)
1214 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1215 .addReg(Op0, Op0IsKill * RegState::Kill)
1217 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1218 ResultReg).addReg(II.ImplicitDefs[0]);
1223 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
1224 const TargetRegisterClass *RC,
1225 unsigned Op0, bool Op0IsKill,
1226 unsigned Op1, bool Op1IsKill,
1228 unsigned ResultReg = createResultReg(RC);
1229 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1231 if (II.getNumDefs() >= 1)
1232 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1233 .addReg(Op0, Op0IsKill * RegState::Kill)
1234 .addReg(Op1, Op1IsKill * RegState::Kill)
1237 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1238 .addReg(Op0, Op0IsKill * RegState::Kill)
1239 .addReg(Op1, Op1IsKill * RegState::Kill)
1241 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1242 ResultReg).addReg(II.ImplicitDefs[0]);
1247 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1248 const TargetRegisterClass *RC,
1250 unsigned ResultReg = createResultReg(RC);
1251 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1253 if (II.getNumDefs() >= 1)
1254 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
1256 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
1257 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1258 ResultReg).addReg(II.ImplicitDefs[0]);
1263 unsigned FastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
1264 const TargetRegisterClass *RC,
1265 uint64_t Imm1, uint64_t Imm2) {
1266 unsigned ResultReg = createResultReg(RC);
1267 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1269 if (II.getNumDefs() >= 1)
1270 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1271 .addImm(Imm1).addImm(Imm2);
1273 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2);
1274 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1275 ResultReg).addReg(II.ImplicitDefs[0]);
1280 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
1281 unsigned Op0, bool Op0IsKill,
1283 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1284 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
1285 "Cannot yet extract from physregs");
1286 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
1287 DL, TII.get(TargetOpcode::COPY), ResultReg)
1288 .addReg(Op0, getKillRegState(Op0IsKill), Idx);
1292 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1293 /// with all but the least significant bit set to zero.
1294 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
1295 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
1298 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
1299 /// Emit code to ensure constants are copied into registers when needed.
1300 /// Remember the virtual registers that need to be added to the Machine PHI
1301 /// nodes as input. We cannot just directly add them, because expansion
1302 /// might result in multiple MBB's for one BB. As such, the start of the
1303 /// BB might correspond to a different MBB than the end.
1304 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
1305 const TerminatorInst *TI = LLVMBB->getTerminator();
1307 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
1308 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
1310 // Check successor nodes' PHI nodes that expect a constant to be available
1312 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
1313 const BasicBlock *SuccBB = TI->getSuccessor(succ);
1314 if (!isa<PHINode>(SuccBB->begin())) continue;
1315 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
1317 // If this terminator has multiple identical successors (common for
1318 // switches), only handle each succ once.
1319 if (!SuccsHandled.insert(SuccMBB)) continue;
1321 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
1323 // At this point we know that there is a 1-1 correspondence between LLVM PHI
1324 // nodes and Machine PHI nodes, but the incoming operands have not been
1326 for (BasicBlock::const_iterator I = SuccBB->begin();
1327 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1329 // Ignore dead phi's.
1330 if (PN->use_empty()) continue;
1332 // Only handle legal types. Two interesting things to note here. First,
1333 // by bailing out early, we may leave behind some dead instructions,
1334 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
1335 // own moves. Second, this check is necessary because FastISel doesn't
1336 // use CreateRegs to create registers, so it always creates
1337 // exactly one register for each non-void instruction.
1338 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
1339 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
1342 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
1344 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1349 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
1351 // Set the DebugLoc for the copy. Prefer the location of the operand
1352 // if there is one; use the location of the PHI otherwise.
1353 DL = PN->getDebugLoc();
1354 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
1355 DL = Inst->getDebugLoc();
1357 unsigned Reg = getRegForValue(PHIOp);
1359 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1362 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));