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/CodeGen/FastISel.h"
47 #include "llvm/CodeGen/FunctionLoweringInfo.h"
48 #include "llvm/CodeGen/MachineInstrBuilder.h"
49 #include "llvm/CodeGen/MachineModuleInfo.h"
50 #include "llvm/CodeGen/MachineRegisterInfo.h"
51 #include "llvm/Analysis/DebugInfo.h"
52 #include "llvm/Analysis/Loads.h"
53 #include "llvm/Target/TargetData.h"
54 #include "llvm/Target/TargetInstrInfo.h"
55 #include "llvm/Target/TargetLowering.h"
56 #include "llvm/Target/TargetMachine.h"
57 #include "llvm/Support/ErrorHandling.h"
60 /// startNewBlock - Set the current block to which generated machine
61 /// instructions will be appended, and clear the local CSE map.
63 void FastISel::startNewBlock() {
64 LocalValueMap.clear();
66 // Start out as null, meaining no local-value instructions have
70 // Advance the last local value past any EH_LABEL instructions.
71 MachineBasicBlock::iterator
72 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
73 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
79 bool FastISel::hasTrivialKill(const Value *V) const {
80 // Don't consider constants or arguments to have trivial kills.
81 const Instruction *I = dyn_cast<Instruction>(V);
85 // No-op casts are trivially coalesced by fast-isel.
86 if (const CastInst *Cast = dyn_cast<CastInst>(I))
87 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
88 !hasTrivialKill(Cast->getOperand(0)))
91 // Only instructions with a single use in the same basic block are considered
92 // to have trivial kills.
93 return I->hasOneUse() &&
94 !(I->getOpcode() == Instruction::BitCast ||
95 I->getOpcode() == Instruction::PtrToInt ||
96 I->getOpcode() == Instruction::IntToPtr) &&
97 cast<Instruction>(*I->use_begin())->getParent() == I->getParent();
100 unsigned FastISel::getRegForValue(const Value *V) {
101 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
102 // Don't handle non-simple values in FastISel.
103 if (!RealVT.isSimple())
106 // Ignore illegal types. We must do this before looking up the value
107 // in ValueMap because Arguments are given virtual registers regardless
108 // of whether FastISel can handle them.
109 MVT VT = RealVT.getSimpleVT();
110 if (!TLI.isTypeLegal(VT)) {
111 // Promote MVT::i1 to a legal type though, because it's common and easy.
113 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
118 // Look up the value to see if we already have a register for it. We
119 // cache values defined by Instructions across blocks, and other values
120 // only locally. This is because Instructions already have the SSA
121 // def-dominates-use requirement enforced.
122 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
123 if (I != FuncInfo.ValueMap.end()) {
124 unsigned Reg = I->second;
127 unsigned Reg = LocalValueMap[V];
131 // In bottom-up mode, just create the virtual register which will be used
132 // to hold the value. It will be materialized later.
133 if (isa<Instruction>(V) &&
134 (!isa<AllocaInst>(V) ||
135 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
136 return FuncInfo.InitializeRegForValue(V);
138 SavePoint SaveInsertPt = enterLocalValueArea();
140 // Materialize the value in a register. Emit any instructions in the
142 Reg = materializeRegForValue(V, VT);
144 leaveLocalValueArea(SaveInsertPt);
149 /// materializeRegForValue - Helper for getRegForValue. This function is
150 /// called when the value isn't already available in a register and must
151 /// be materialized with new instructions.
152 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
155 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
156 if (CI->getValue().getActiveBits() <= 64)
157 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
158 } else if (isa<AllocaInst>(V)) {
159 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
160 } else if (isa<ConstantPointerNull>(V)) {
161 // Translate this as an integer zero so that it can be
162 // local-CSE'd with actual integer zeros.
164 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
165 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
166 // Try to emit the constant directly.
167 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
170 // Try to emit the constant by using an integer constant with a cast.
171 const APFloat &Flt = CF->getValueAPF();
172 EVT IntVT = TLI.getPointerTy();
175 uint32_t IntBitWidth = IntVT.getSizeInBits();
177 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
178 APFloat::rmTowardZero, &isExact);
180 APInt IntVal(IntBitWidth, 2, x);
182 unsigned IntegerReg =
183 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
185 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
186 IntegerReg, /*Kill=*/false);
189 } else if (const Operator *Op = dyn_cast<Operator>(V)) {
190 if (!SelectOperator(Op, Op->getOpcode()))
191 if (!isa<Instruction>(Op) ||
192 !TargetSelectInstruction(cast<Instruction>(Op)))
194 Reg = lookUpRegForValue(Op);
195 } else if (isa<UndefValue>(V)) {
196 Reg = createResultReg(TLI.getRegClassFor(VT));
197 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
198 TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
201 // If target-independent code couldn't handle the value, give target-specific
203 if (!Reg && isa<Constant>(V))
204 Reg = TargetMaterializeConstant(cast<Constant>(V));
206 // Don't cache constant materializations in the general ValueMap.
207 // To do so would require tracking what uses they dominate.
209 LocalValueMap[V] = Reg;
210 LastLocalValue = MRI.getVRegDef(Reg);
215 unsigned FastISel::lookUpRegForValue(const Value *V) {
216 // Look up the value to see if we already have a register for it. We
217 // cache values defined by Instructions across blocks, and other values
218 // only locally. This is because Instructions already have the SSA
219 // def-dominates-use requirement enforced.
220 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
221 if (I != FuncInfo.ValueMap.end())
223 return LocalValueMap[V];
226 /// UpdateValueMap - Update the value map to include the new mapping for this
227 /// instruction, or insert an extra copy to get the result in a previous
228 /// determined register.
229 /// NOTE: This is only necessary because we might select a block that uses
230 /// a value before we select the block that defines the value. It might be
231 /// possible to fix this by selecting blocks in reverse postorder.
232 unsigned FastISel::UpdateValueMap(const Value *I, unsigned Reg) {
233 if (!isa<Instruction>(I)) {
234 LocalValueMap[I] = Reg;
238 unsigned &AssignedReg = FuncInfo.ValueMap[I];
239 if (AssignedReg == 0)
240 // Use the new register.
242 else if (Reg != AssignedReg) {
243 // Arrange for uses of AssignedReg to be replaced by uses of Reg.
244 FuncInfo.RegFixups[AssignedReg] = Reg;
252 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
253 unsigned IdxN = getRegForValue(Idx);
255 // Unhandled operand. Halt "fast" selection and bail.
256 return std::pair<unsigned, bool>(0, false);
258 bool IdxNIsKill = hasTrivialKill(Idx);
260 // If the index is smaller or larger than intptr_t, truncate or extend it.
261 MVT PtrVT = TLI.getPointerTy();
262 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
263 if (IdxVT.bitsLT(PtrVT)) {
264 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
268 else if (IdxVT.bitsGT(PtrVT)) {
269 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
273 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
276 void FastISel::recomputeInsertPt() {
277 if (getLastLocalValue()) {
278 FuncInfo.InsertPt = getLastLocalValue();
279 FuncInfo.MBB = FuncInfo.InsertPt->getParent();
282 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI();
284 // Now skip past any EH_LABELs, which must remain at the beginning.
285 while (FuncInfo.InsertPt != FuncInfo.MBB->end() &&
286 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL)
290 FastISel::SavePoint FastISel::enterLocalValueArea() {
291 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
295 SavePoint SP = { OldInsertPt, OldDL };
299 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
300 if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
301 LastLocalValue = llvm::prior(FuncInfo.InsertPt);
303 // Restore the previous insert position.
304 FuncInfo.InsertPt = OldInsertPt.InsertPt;
308 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
309 /// which has an opcode which directly corresponds to the given ISD opcode.
311 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
312 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
313 if (VT == MVT::Other || !VT.isSimple())
314 // Unhandled type. Halt "fast" selection and bail.
317 // We only handle legal types. For example, on x86-32 the instruction
318 // selector contains all of the 64-bit instructions from x86-64,
319 // under the assumption that i64 won't be used if the target doesn't
321 if (!TLI.isTypeLegal(VT)) {
322 // MVT::i1 is special. Allow AND, OR, or XOR because they
323 // don't require additional zeroing, which makes them easy.
325 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
326 ISDOpcode == ISD::XOR))
327 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
332 unsigned Op0 = getRegForValue(I->getOperand(0));
334 // Unhandled operand. Halt "fast" selection and bail.
337 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
339 // Check if the second operand is a constant and handle it appropriately.
340 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
341 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
342 ISDOpcode, Op0, Op0IsKill,
344 if (ResultReg != 0) {
345 // We successfully emitted code for the given LLVM Instruction.
346 UpdateValueMap(I, ResultReg);
351 // Check if the second operand is a constant float.
352 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
353 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
354 ISDOpcode, Op0, Op0IsKill, CF);
355 if (ResultReg != 0) {
356 // We successfully emitted code for the given LLVM Instruction.
357 UpdateValueMap(I, ResultReg);
362 unsigned Op1 = getRegForValue(I->getOperand(1));
364 // Unhandled operand. Halt "fast" selection and bail.
367 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
369 // Now we have both operands in registers. Emit the instruction.
370 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
375 // Target-specific code wasn't able to find a machine opcode for
376 // the given ISD opcode and type. Halt "fast" selection and bail.
379 // We successfully emitted code for the given LLVM Instruction.
380 UpdateValueMap(I, ResultReg);
384 bool FastISel::SelectGetElementPtr(const User *I) {
385 unsigned N = getRegForValue(I->getOperand(0));
387 // Unhandled operand. Halt "fast" selection and bail.
390 bool NIsKill = hasTrivialKill(I->getOperand(0));
392 const Type *Ty = I->getOperand(0)->getType();
393 MVT VT = TLI.getPointerTy();
394 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
395 E = I->op_end(); OI != E; ++OI) {
396 const Value *Idx = *OI;
397 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
398 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
401 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
402 // FIXME: This can be optimized by combining the add with a
404 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
406 // Unhandled operand. Halt "fast" selection and bail.
410 Ty = StTy->getElementType(Field);
412 Ty = cast<SequentialType>(Ty)->getElementType();
414 // If this is a constant subscript, handle it quickly.
415 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
416 if (CI->isZero()) continue;
418 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
419 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
421 // Unhandled operand. Halt "fast" selection and bail.
427 // N = N + Idx * ElementSize;
428 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
429 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
430 unsigned IdxN = Pair.first;
431 bool IdxNIsKill = Pair.second;
433 // Unhandled operand. Halt "fast" selection and bail.
436 if (ElementSize != 1) {
437 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
439 // Unhandled operand. Halt "fast" selection and bail.
443 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
445 // Unhandled operand. Halt "fast" selection and bail.
450 // We successfully emitted code for the given LLVM Instruction.
451 UpdateValueMap(I, N);
455 bool FastISel::SelectCall(const User *I) {
456 const Function *F = cast<CallInst>(I)->getCalledFunction();
457 if (!F) return false;
459 // Handle selected intrinsic function calls.
460 unsigned IID = F->getIntrinsicID();
463 case Intrinsic::dbg_declare: {
464 const DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
465 if (!DIVariable(DI->getVariable()).Verify() ||
466 !FuncInfo.MF->getMMI().hasDebugInfo())
469 const Value *Address = DI->getAddress();
472 if (isa<UndefValue>(Address))
474 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
475 // Don't handle byval struct arguments or VLAs, for example.
477 // Building the map above is target independent. Generating DBG_VALUE
478 // inline is target dependent; do this now.
479 DenseMap<const Value *, unsigned>::iterator It =
480 FuncInfo.ValueMap.find(Address);
481 if (0 && It != FuncInfo.ValueMap.end()) {
482 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
483 TII.get(TargetOpcode::DBG_VALUE))
484 .addReg(It->second, RegState::Debug).addImm(0).addMetadata(DI->getVariable());
486 (void)TargetSelectInstruction(cast<Instruction>(I));
490 case Intrinsic::dbg_value: {
491 // This form of DBG_VALUE is target-independent.
492 const DbgValueInst *DI = cast<DbgValueInst>(I);
493 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
494 const Value *V = DI->getValue();
496 // Currently the optimizer can produce this; insert an undef to
497 // help debugging. Probably the optimizer should not do this.
498 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
499 .addReg(0U).addImm(DI->getOffset())
500 .addMetadata(DI->getVariable());
501 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
502 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
503 .addImm(CI->getZExtValue()).addImm(DI->getOffset())
504 .addMetadata(DI->getVariable());
505 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
506 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
507 .addFPImm(CF).addImm(DI->getOffset())
508 .addMetadata(DI->getVariable());
509 } else if (unsigned Reg = lookUpRegForValue(V)) {
510 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
511 .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
512 .addMetadata(DI->getVariable());
514 // We can't yet handle anything else here because it would require
515 // generating code, thus altering codegen because of debug info.
516 // Insert an undef so we can see what we dropped.
517 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
518 .addReg(0U).addImm(DI->getOffset())
519 .addMetadata(DI->getVariable());
523 case Intrinsic::eh_exception: {
524 EVT VT = TLI.getValueType(I->getType());
525 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
527 case TargetLowering::Expand: {
528 assert(FuncInfo.MBB->isLandingPad() &&
529 "Call to eh.exception not in landing pad!");
530 unsigned Reg = TLI.getExceptionAddressRegister();
531 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
532 unsigned ResultReg = createResultReg(RC);
533 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
534 ResultReg).addReg(Reg);
535 UpdateValueMap(I, ResultReg);
541 case Intrinsic::eh_selector: {
542 EVT VT = TLI.getValueType(I->getType());
543 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
545 case TargetLowering::Expand: {
546 if (FuncInfo.MBB->isLandingPad())
547 AddCatchInfo(*cast<CallInst>(I), &FuncInfo.MF->getMMI(), FuncInfo.MBB);
550 FuncInfo.CatchInfoLost.insert(cast<CallInst>(I));
552 // FIXME: Mark exception selector register as live in. Hack for PR1508.
553 unsigned Reg = TLI.getExceptionSelectorRegister();
554 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
557 unsigned Reg = TLI.getExceptionSelectorRegister();
558 EVT SrcVT = TLI.getPointerTy();
559 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
560 unsigned ResultReg = createResultReg(RC);
561 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
562 ResultReg).addReg(Reg);
564 bool ResultRegIsKill = hasTrivialKill(I);
566 // Cast the register to the type of the selector.
567 if (SrcVT.bitsGT(MVT::i32))
568 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
569 ResultReg, ResultRegIsKill);
570 else if (SrcVT.bitsLT(MVT::i32))
571 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
572 ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill);
574 // Unhandled operand. Halt "fast" selection and bail.
577 UpdateValueMap(I, ResultReg);
586 // An arbitrary call. Bail.
590 bool FastISel::SelectCast(const User *I, unsigned Opcode) {
591 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
592 EVT DstVT = TLI.getValueType(I->getType());
594 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
595 DstVT == MVT::Other || !DstVT.isSimple())
596 // Unhandled type. Halt "fast" selection and bail.
599 // Check if the destination type is legal. Or as a special case,
600 // it may be i1 if we're doing a truncate because that's
601 // easy and somewhat common.
602 if (!TLI.isTypeLegal(DstVT))
603 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
604 // Unhandled type. Halt "fast" selection and bail.
607 // Check if the source operand is legal. Or as a special case,
608 // it may be i1 if we're doing zero-extension because that's
609 // easy and somewhat common.
610 if (!TLI.isTypeLegal(SrcVT))
611 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
612 // Unhandled type. Halt "fast" selection and bail.
615 unsigned InputReg = getRegForValue(I->getOperand(0));
617 // Unhandled operand. Halt "fast" selection and bail.
620 bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
622 // If the operand is i1, arrange for the high bits in the register to be zero.
623 if (SrcVT == MVT::i1) {
624 SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
625 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg, InputRegIsKill);
628 InputRegIsKill = true;
630 // If the result is i1, truncate to the target's type for i1 first.
631 if (DstVT == MVT::i1)
632 DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
634 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
637 InputReg, InputRegIsKill);
641 UpdateValueMap(I, ResultReg);
645 bool FastISel::SelectBitCast(const User *I) {
646 // If the bitcast doesn't change the type, just use the operand value.
647 if (I->getType() == I->getOperand(0)->getType()) {
648 unsigned Reg = getRegForValue(I->getOperand(0));
651 UpdateValueMap(I, Reg);
655 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
656 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
657 EVT DstVT = TLI.getValueType(I->getType());
659 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
660 DstVT == MVT::Other || !DstVT.isSimple() ||
661 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
662 // Unhandled type. Halt "fast" selection and bail.
665 unsigned Op0 = getRegForValue(I->getOperand(0));
667 // Unhandled operand. Halt "fast" selection and bail.
670 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
672 // First, try to perform the bitcast by inserting a reg-reg copy.
673 unsigned ResultReg = 0;
674 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
675 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
676 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
677 // Don't attempt a cross-class copy. It will likely fail.
678 if (SrcClass == DstClass) {
679 ResultReg = createResultReg(DstClass);
680 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
681 ResultReg).addReg(Op0);
685 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
687 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
688 ISD::BIT_CONVERT, Op0, Op0IsKill);
693 UpdateValueMap(I, ResultReg);
698 FastISel::SelectInstruction(const Instruction *I) {
699 // Just before the terminator instruction, insert instructions to
700 // feed PHI nodes in successor blocks.
701 if (isa<TerminatorInst>(I))
702 if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
705 DL = I->getDebugLoc();
707 // First, try doing target-independent selection.
708 if (SelectOperator(I, I->getOpcode())) {
713 // Next, try calling the target to attempt to handle the instruction.
714 if (TargetSelectInstruction(I)) {
723 /// FastEmitBranch - Emit an unconditional branch to the given block,
724 /// unless it is the immediate (fall-through) successor, and update
727 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
728 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
729 // The unconditional fall-through case, which needs no instructions.
731 // The unconditional branch case.
732 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
733 SmallVector<MachineOperand, 0>(), DL);
735 FuncInfo.MBB->addSuccessor(MSucc);
738 /// SelectFNeg - Emit an FNeg operation.
741 FastISel::SelectFNeg(const User *I) {
742 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
743 if (OpReg == 0) return false;
745 bool OpRegIsKill = hasTrivialKill(I);
747 // If the target has ISD::FNEG, use it.
748 EVT VT = TLI.getValueType(I->getType());
749 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
750 ISD::FNEG, OpReg, OpRegIsKill);
751 if (ResultReg != 0) {
752 UpdateValueMap(I, ResultReg);
756 // Bitcast the value to integer, twiddle the sign bit with xor,
757 // and then bitcast it back to floating-point.
758 if (VT.getSizeInBits() > 64) return false;
759 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
760 if (!TLI.isTypeLegal(IntVT))
763 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
764 ISD::BIT_CONVERT, OpReg, OpRegIsKill);
768 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
769 IntReg, /*Kill=*/true,
770 UINT64_C(1) << (VT.getSizeInBits()-1),
771 IntVT.getSimpleVT());
772 if (IntResultReg == 0)
775 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
776 ISD::BIT_CONVERT, IntResultReg, /*Kill=*/true);
780 UpdateValueMap(I, ResultReg);
785 FastISel::SelectOperator(const User *I, unsigned Opcode) {
787 case Instruction::Add:
788 return SelectBinaryOp(I, ISD::ADD);
789 case Instruction::FAdd:
790 return SelectBinaryOp(I, ISD::FADD);
791 case Instruction::Sub:
792 return SelectBinaryOp(I, ISD::SUB);
793 case Instruction::FSub:
794 // FNeg is currently represented in LLVM IR as a special case of FSub.
795 if (BinaryOperator::isFNeg(I))
796 return SelectFNeg(I);
797 return SelectBinaryOp(I, ISD::FSUB);
798 case Instruction::Mul:
799 return SelectBinaryOp(I, ISD::MUL);
800 case Instruction::FMul:
801 return SelectBinaryOp(I, ISD::FMUL);
802 case Instruction::SDiv:
803 return SelectBinaryOp(I, ISD::SDIV);
804 case Instruction::UDiv:
805 return SelectBinaryOp(I, ISD::UDIV);
806 case Instruction::FDiv:
807 return SelectBinaryOp(I, ISD::FDIV);
808 case Instruction::SRem:
809 return SelectBinaryOp(I, ISD::SREM);
810 case Instruction::URem:
811 return SelectBinaryOp(I, ISD::UREM);
812 case Instruction::FRem:
813 return SelectBinaryOp(I, ISD::FREM);
814 case Instruction::Shl:
815 return SelectBinaryOp(I, ISD::SHL);
816 case Instruction::LShr:
817 return SelectBinaryOp(I, ISD::SRL);
818 case Instruction::AShr:
819 return SelectBinaryOp(I, ISD::SRA);
820 case Instruction::And:
821 return SelectBinaryOp(I, ISD::AND);
822 case Instruction::Or:
823 return SelectBinaryOp(I, ISD::OR);
824 case Instruction::Xor:
825 return SelectBinaryOp(I, ISD::XOR);
827 case Instruction::GetElementPtr:
828 return SelectGetElementPtr(I);
830 case Instruction::Br: {
831 const BranchInst *BI = cast<BranchInst>(I);
833 if (BI->isUnconditional()) {
834 const BasicBlock *LLVMSucc = BI->getSuccessor(0);
835 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
836 FastEmitBranch(MSucc, BI->getDebugLoc());
840 // Conditional branches are not handed yet.
841 // Halt "fast" selection and bail.
845 case Instruction::Unreachable:
849 case Instruction::Alloca:
850 // FunctionLowering has the static-sized case covered.
851 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
854 // Dynamic-sized alloca is not handled yet.
857 case Instruction::Call:
858 return SelectCall(I);
860 case Instruction::BitCast:
861 return SelectBitCast(I);
863 case Instruction::FPToSI:
864 return SelectCast(I, ISD::FP_TO_SINT);
865 case Instruction::ZExt:
866 return SelectCast(I, ISD::ZERO_EXTEND);
867 case Instruction::SExt:
868 return SelectCast(I, ISD::SIGN_EXTEND);
869 case Instruction::Trunc:
870 return SelectCast(I, ISD::TRUNCATE);
871 case Instruction::SIToFP:
872 return SelectCast(I, ISD::SINT_TO_FP);
874 case Instruction::IntToPtr: // Deliberate fall-through.
875 case Instruction::PtrToInt: {
876 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
877 EVT DstVT = TLI.getValueType(I->getType());
878 if (DstVT.bitsGT(SrcVT))
879 return SelectCast(I, ISD::ZERO_EXTEND);
880 if (DstVT.bitsLT(SrcVT))
881 return SelectCast(I, ISD::TRUNCATE);
882 unsigned Reg = getRegForValue(I->getOperand(0));
883 if (Reg == 0) return false;
884 UpdateValueMap(I, Reg);
888 case Instruction::PHI:
889 llvm_unreachable("FastISel shouldn't visit PHI nodes!");
892 // Unhandled instruction. Halt "fast" selection and bail.
897 FastISel::FastISel(FunctionLoweringInfo &funcInfo)
898 : FuncInfo(funcInfo),
899 MRI(FuncInfo.MF->getRegInfo()),
900 MFI(*FuncInfo.MF->getFrameInfo()),
901 MCP(*FuncInfo.MF->getConstantPool()),
902 TM(FuncInfo.MF->getTarget()),
903 TD(*TM.getTargetData()),
904 TII(*TM.getInstrInfo()),
905 TLI(*TM.getTargetLowering()),
906 TRI(*TM.getRegisterInfo()) {
909 FastISel::~FastISel() {}
911 unsigned FastISel::FastEmit_(MVT, MVT,
916 unsigned FastISel::FastEmit_r(MVT, MVT,
918 unsigned /*Op0*/, bool /*Op0IsKill*/) {
922 unsigned FastISel::FastEmit_rr(MVT, MVT,
924 unsigned /*Op0*/, bool /*Op0IsKill*/,
925 unsigned /*Op1*/, bool /*Op1IsKill*/) {
929 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
933 unsigned FastISel::FastEmit_f(MVT, MVT,
934 unsigned, const ConstantFP * /*FPImm*/) {
938 unsigned FastISel::FastEmit_ri(MVT, MVT,
940 unsigned /*Op0*/, bool /*Op0IsKill*/,
945 unsigned FastISel::FastEmit_rf(MVT, MVT,
947 unsigned /*Op0*/, bool /*Op0IsKill*/,
948 const ConstantFP * /*FPImm*/) {
952 unsigned FastISel::FastEmit_rri(MVT, MVT,
954 unsigned /*Op0*/, bool /*Op0IsKill*/,
955 unsigned /*Op1*/, bool /*Op1IsKill*/,
960 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
961 /// to emit an instruction with an immediate operand using FastEmit_ri.
962 /// If that fails, it materializes the immediate into a register and try
963 /// FastEmit_rr instead.
964 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
965 unsigned Op0, bool Op0IsKill,
966 uint64_t Imm, MVT ImmType) {
967 // First check if immediate type is legal. If not, we can't use the ri form.
968 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
971 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
972 if (MaterialReg == 0)
974 return FastEmit_rr(VT, VT, Opcode,
976 MaterialReg, /*Kill=*/true);
979 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
980 /// to emit an instruction with a floating-point immediate operand using
981 /// FastEmit_rf. If that fails, it materializes the immediate into a register
982 /// and try FastEmit_rr instead.
983 unsigned FastISel::FastEmit_rf_(MVT VT, unsigned Opcode,
984 unsigned Op0, bool Op0IsKill,
985 const ConstantFP *FPImm, MVT ImmType) {
986 // First check if immediate type is legal. If not, we can't use the rf form.
987 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, Op0IsKill, FPImm);
991 // Materialize the constant in a register.
992 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
993 if (MaterialReg == 0) {
994 // If the target doesn't have a way to directly enter a floating-point
995 // value into a register, use an alternate approach.
996 // TODO: The current approach only supports floating-point constants
997 // that can be constructed by conversion from integer values. This should
998 // be replaced by code that creates a load from a constant-pool entry,
999 // which will require some target-specific work.
1000 const APFloat &Flt = FPImm->getValueAPF();
1001 EVT IntVT = TLI.getPointerTy();
1004 uint32_t IntBitWidth = IntVT.getSizeInBits();
1006 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
1007 APFloat::rmTowardZero, &isExact);
1010 APInt IntVal(IntBitWidth, 2, x);
1012 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
1013 ISD::Constant, IntVal.getZExtValue());
1014 if (IntegerReg == 0)
1016 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
1017 ISD::SINT_TO_FP, IntegerReg, /*Kill=*/true);
1018 if (MaterialReg == 0)
1021 return FastEmit_rr(VT, VT, Opcode,
1023 MaterialReg, /*Kill=*/true);
1026 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
1027 return MRI.createVirtualRegister(RC);
1030 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
1031 const TargetRegisterClass* RC) {
1032 unsigned ResultReg = createResultReg(RC);
1033 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1035 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
1039 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
1040 const TargetRegisterClass *RC,
1041 unsigned Op0, bool Op0IsKill) {
1042 unsigned ResultReg = createResultReg(RC);
1043 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1045 if (II.getNumDefs() >= 1)
1046 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1047 .addReg(Op0, Op0IsKill * RegState::Kill);
1049 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1050 .addReg(Op0, Op0IsKill * RegState::Kill);
1051 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1052 ResultReg).addReg(II.ImplicitDefs[0]);
1058 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
1059 const TargetRegisterClass *RC,
1060 unsigned Op0, bool Op0IsKill,
1061 unsigned Op1, bool Op1IsKill) {
1062 unsigned ResultReg = createResultReg(RC);
1063 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1065 if (II.getNumDefs() >= 1)
1066 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1067 .addReg(Op0, Op0IsKill * RegState::Kill)
1068 .addReg(Op1, Op1IsKill * RegState::Kill);
1070 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1071 .addReg(Op0, Op0IsKill * RegState::Kill)
1072 .addReg(Op1, Op1IsKill * RegState::Kill);
1073 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1074 ResultReg).addReg(II.ImplicitDefs[0]);
1079 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
1080 const TargetRegisterClass *RC,
1081 unsigned Op0, bool Op0IsKill,
1083 unsigned ResultReg = createResultReg(RC);
1084 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1086 if (II.getNumDefs() >= 1)
1087 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1088 .addReg(Op0, Op0IsKill * RegState::Kill)
1091 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1092 .addReg(Op0, Op0IsKill * RegState::Kill)
1094 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1095 ResultReg).addReg(II.ImplicitDefs[0]);
1100 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
1101 const TargetRegisterClass *RC,
1102 unsigned Op0, bool Op0IsKill,
1103 const ConstantFP *FPImm) {
1104 unsigned ResultReg = createResultReg(RC);
1105 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1107 if (II.getNumDefs() >= 1)
1108 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1109 .addReg(Op0, Op0IsKill * RegState::Kill)
1112 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1113 .addReg(Op0, Op0IsKill * RegState::Kill)
1115 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1116 ResultReg).addReg(II.ImplicitDefs[0]);
1121 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
1122 const TargetRegisterClass *RC,
1123 unsigned Op0, bool Op0IsKill,
1124 unsigned Op1, bool Op1IsKill,
1126 unsigned ResultReg = createResultReg(RC);
1127 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1129 if (II.getNumDefs() >= 1)
1130 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1131 .addReg(Op0, Op0IsKill * RegState::Kill)
1132 .addReg(Op1, Op1IsKill * RegState::Kill)
1135 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1136 .addReg(Op0, Op0IsKill * RegState::Kill)
1137 .addReg(Op1, Op1IsKill * RegState::Kill)
1139 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1140 ResultReg).addReg(II.ImplicitDefs[0]);
1145 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1146 const TargetRegisterClass *RC,
1148 unsigned ResultReg = createResultReg(RC);
1149 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1151 if (II.getNumDefs() >= 1)
1152 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
1154 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
1155 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1156 ResultReg).addReg(II.ImplicitDefs[0]);
1161 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
1162 unsigned Op0, bool Op0IsKill,
1164 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1165 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
1166 "Cannot yet extract from physregs");
1167 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
1168 DL, TII.get(TargetOpcode::COPY), ResultReg)
1169 .addReg(Op0, getKillRegState(Op0IsKill), Idx);
1173 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1174 /// with all but the least significant bit set to zero.
1175 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
1176 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
1179 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
1180 /// Emit code to ensure constants are copied into registers when needed.
1181 /// Remember the virtual registers that need to be added to the Machine PHI
1182 /// nodes as input. We cannot just directly add them, because expansion
1183 /// might result in multiple MBB's for one BB. As such, the start of the
1184 /// BB might correspond to a different MBB than the end.
1185 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
1186 const TerminatorInst *TI = LLVMBB->getTerminator();
1188 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
1189 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
1191 // Check successor nodes' PHI nodes that expect a constant to be available
1193 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
1194 const BasicBlock *SuccBB = TI->getSuccessor(succ);
1195 if (!isa<PHINode>(SuccBB->begin())) continue;
1196 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
1198 // If this terminator has multiple identical successors (common for
1199 // switches), only handle each succ once.
1200 if (!SuccsHandled.insert(SuccMBB)) continue;
1202 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
1204 // At this point we know that there is a 1-1 correspondence between LLVM PHI
1205 // nodes and Machine PHI nodes, but the incoming operands have not been
1207 for (BasicBlock::const_iterator I = SuccBB->begin();
1208 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1210 // Ignore dead phi's.
1211 if (PN->use_empty()) continue;
1213 // Only handle legal types. Two interesting things to note here. First,
1214 // by bailing out early, we may leave behind some dead instructions,
1215 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
1216 // own moves. Second, this check is necessary becuase FastISel doesn't
1217 // use CreateRegs to create registers, so it always creates
1218 // exactly one register for each non-void instruction.
1219 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
1220 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
1223 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
1225 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1230 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
1232 // Set the DebugLoc for the copy. Prefer the location of the operand
1233 // if there is one; use the location of the PHI otherwise.
1234 DL = PN->getDebugLoc();
1235 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
1236 DL = Inst->getDebugLoc();
1238 unsigned Reg = getRegForValue(PHIOp);
1240 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1243 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));