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 #define DEBUG_TYPE "isel"
43 #include "llvm/Function.h"
44 #include "llvm/GlobalVariable.h"
45 #include "llvm/Instructions.h"
46 #include "llvm/IntrinsicInst.h"
47 #include "llvm/Operator.h"
48 #include "llvm/CodeGen/Analysis.h"
49 #include "llvm/CodeGen/FastISel.h"
50 #include "llvm/CodeGen/FunctionLoweringInfo.h"
51 #include "llvm/CodeGen/MachineInstrBuilder.h"
52 #include "llvm/CodeGen/MachineModuleInfo.h"
53 #include "llvm/CodeGen/MachineRegisterInfo.h"
54 #include "llvm/Analysis/DebugInfo.h"
55 #include "llvm/Analysis/Loads.h"
56 #include "llvm/Target/TargetData.h"
57 #include "llvm/Target/TargetInstrInfo.h"
58 #include "llvm/Target/TargetLowering.h"
59 #include "llvm/Target/TargetMachine.h"
60 #include "llvm/Support/ErrorHandling.h"
61 #include "llvm/Support/Debug.h"
62 #include "llvm/ADT/Statistic.h"
65 STATISTIC(NumFastIselSuccessIndependent, "Number of insts selected by target-independent selector");
66 STATISTIC(NumFastIselSuccessTarget, "Number of insts selected by target-specific selector");
68 /// startNewBlock - Set the current block to which generated machine
69 /// instructions will be appended, and clear the local CSE map.
71 void FastISel::startNewBlock() {
72 LocalValueMap.clear();
76 // Advance the emit start point past any EH_LABEL instructions.
77 MachineBasicBlock::iterator
78 I = FuncInfo.MBB->begin(), E = FuncInfo.MBB->end();
79 while (I != E && I->getOpcode() == TargetOpcode::EH_LABEL) {
83 LastLocalValue = EmitStartPt;
86 void FastISel::flushLocalValueMap() {
87 LocalValueMap.clear();
88 LastLocalValue = EmitStartPt;
92 bool FastISel::hasTrivialKill(const Value *V) const {
93 // Don't consider constants or arguments to have trivial kills.
94 const Instruction *I = dyn_cast<Instruction>(V);
98 // No-op casts are trivially coalesced by fast-isel.
99 if (const CastInst *Cast = dyn_cast<CastInst>(I))
100 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
101 !hasTrivialKill(Cast->getOperand(0)))
104 // GEPs with all zero indices are trivially coalesced by fast-isel.
105 if (const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I))
106 if (GEP->hasAllZeroIndices() && !hasTrivialKill(GEP->getOperand(0)))
109 // Only instructions with a single use in the same basic block are considered
110 // to have trivial kills.
111 return I->hasOneUse() &&
112 !(I->getOpcode() == Instruction::BitCast ||
113 I->getOpcode() == Instruction::PtrToInt ||
114 I->getOpcode() == Instruction::IntToPtr) &&
115 cast<Instruction>(*I->use_begin())->getParent() == I->getParent();
118 unsigned FastISel::getRegForValue(const Value *V) {
119 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
120 // Don't handle non-simple values in FastISel.
121 if (!RealVT.isSimple())
124 // Ignore illegal types. We must do this before looking up the value
125 // in ValueMap because Arguments are given virtual registers regardless
126 // of whether FastISel can handle them.
127 MVT VT = RealVT.getSimpleVT();
128 if (!TLI.isTypeLegal(VT)) {
129 // Handle integer promotions, though, because they're common and easy.
130 if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
131 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
136 // Look up the value to see if we already have a register for it. We
137 // cache values defined by Instructions across blocks, and other values
138 // only locally. This is because Instructions already have the SSA
139 // def-dominates-use requirement enforced.
140 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
141 if (I != FuncInfo.ValueMap.end())
144 unsigned Reg = LocalValueMap[V];
148 // In bottom-up mode, just create the virtual register which will be used
149 // to hold the value. It will be materialized later.
150 if (isa<Instruction>(V) &&
151 (!isa<AllocaInst>(V) ||
152 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V))))
153 return FuncInfo.InitializeRegForValue(V);
155 SavePoint SaveInsertPt = enterLocalValueArea();
157 // Materialize the value in a register. Emit any instructions in the
159 Reg = materializeRegForValue(V, VT);
161 leaveLocalValueArea(SaveInsertPt);
166 /// materializeRegForValue - Helper for getRegForValue. This function is
167 /// called when the value isn't already available in a register and must
168 /// be materialized with new instructions.
169 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
172 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
173 if (CI->getValue().getActiveBits() <= 64)
174 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
175 } else if (isa<AllocaInst>(V)) {
176 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
177 } else if (isa<ConstantPointerNull>(V)) {
178 // Translate this as an integer zero so that it can be
179 // local-CSE'd with actual integer zeros.
181 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
182 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
183 if (CF->isNullValue()) {
184 Reg = TargetMaterializeFloatZero(CF);
186 // Try to emit the constant directly.
187 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
191 // Try to emit the constant by using an integer constant with a cast.
192 const APFloat &Flt = CF->getValueAPF();
193 EVT IntVT = TLI.getPointerTy();
196 uint32_t IntBitWidth = IntVT.getSizeInBits();
198 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
199 APFloat::rmTowardZero, &isExact);
201 APInt IntVal(IntBitWidth, x);
203 unsigned IntegerReg =
204 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
206 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
207 IntegerReg, /*Kill=*/false);
210 } else if (const Operator *Op = dyn_cast<Operator>(V)) {
211 if (!SelectOperator(Op, Op->getOpcode()))
212 if (!isa<Instruction>(Op) ||
213 !TargetSelectInstruction(cast<Instruction>(Op)))
215 Reg = lookUpRegForValue(Op);
216 } else if (isa<UndefValue>(V)) {
217 Reg = createResultReg(TLI.getRegClassFor(VT));
218 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
219 TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
222 // If target-independent code couldn't handle the value, give target-specific
224 if (!Reg && isa<Constant>(V))
225 Reg = TargetMaterializeConstant(cast<Constant>(V));
227 // Don't cache constant materializations in the general ValueMap.
228 // To do so would require tracking what uses they dominate.
230 LocalValueMap[V] = Reg;
231 LastLocalValue = MRI.getVRegDef(Reg);
236 unsigned FastISel::lookUpRegForValue(const Value *V) {
237 // Look up the value to see if we already have a register for it. We
238 // cache values defined by Instructions across blocks, and other values
239 // only locally. This is because Instructions already have the SSA
240 // def-dominates-use requirement enforced.
241 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
242 if (I != FuncInfo.ValueMap.end())
244 return LocalValueMap[V];
247 /// UpdateValueMap - Update the value map to include the new mapping for this
248 /// instruction, or insert an extra copy to get the result in a previous
249 /// determined register.
250 /// NOTE: This is only necessary because we might select a block that uses
251 /// a value before we select the block that defines the value. It might be
252 /// possible to fix this by selecting blocks in reverse postorder.
253 void FastISel::UpdateValueMap(const Value *I, unsigned Reg, unsigned NumRegs) {
254 if (!isa<Instruction>(I)) {
255 LocalValueMap[I] = Reg;
259 unsigned &AssignedReg = FuncInfo.ValueMap[I];
260 if (AssignedReg == 0)
261 // Use the new register.
263 else if (Reg != AssignedReg) {
264 // Arrange for uses of AssignedReg to be replaced by uses of Reg.
265 for (unsigned i = 0; i < NumRegs; i++)
266 FuncInfo.RegFixups[AssignedReg+i] = Reg+i;
272 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
273 unsigned IdxN = getRegForValue(Idx);
275 // Unhandled operand. Halt "fast" selection and bail.
276 return std::pair<unsigned, bool>(0, false);
278 bool IdxNIsKill = hasTrivialKill(Idx);
280 // If the index is smaller or larger than intptr_t, truncate or extend it.
281 MVT PtrVT = TLI.getPointerTy();
282 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
283 if (IdxVT.bitsLT(PtrVT)) {
284 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
288 else if (IdxVT.bitsGT(PtrVT)) {
289 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
293 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
296 void FastISel::recomputeInsertPt() {
297 if (getLastLocalValue()) {
298 FuncInfo.InsertPt = getLastLocalValue();
299 FuncInfo.MBB = FuncInfo.InsertPt->getParent();
302 FuncInfo.InsertPt = FuncInfo.MBB->getFirstNonPHI();
304 // Now skip past any EH_LABELs, which must remain at the beginning.
305 while (FuncInfo.InsertPt != FuncInfo.MBB->end() &&
306 FuncInfo.InsertPt->getOpcode() == TargetOpcode::EH_LABEL)
310 FastISel::SavePoint FastISel::enterLocalValueArea() {
311 MachineBasicBlock::iterator OldInsertPt = FuncInfo.InsertPt;
315 SavePoint SP = { OldInsertPt, OldDL };
319 void FastISel::leaveLocalValueArea(SavePoint OldInsertPt) {
320 if (FuncInfo.InsertPt != FuncInfo.MBB->begin())
321 LastLocalValue = llvm::prior(FuncInfo.InsertPt);
323 // Restore the previous insert position.
324 FuncInfo.InsertPt = OldInsertPt.InsertPt;
328 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
329 /// which has an opcode which directly corresponds to the given ISD opcode.
331 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
332 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
333 if (VT == MVT::Other || !VT.isSimple())
334 // Unhandled type. Halt "fast" selection and bail.
337 // We only handle legal types. For example, on x86-32 the instruction
338 // selector contains all of the 64-bit instructions from x86-64,
339 // under the assumption that i64 won't be used if the target doesn't
341 if (!TLI.isTypeLegal(VT)) {
342 // MVT::i1 is special. Allow AND, OR, or XOR because they
343 // don't require additional zeroing, which makes them easy.
345 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
346 ISDOpcode == ISD::XOR))
347 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
352 // Check if the first operand is a constant, and handle it as "ri". At -O0,
353 // we don't have anything that canonicalizes operand order.
354 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(0)))
355 if (isa<Instruction>(I) && cast<Instruction>(I)->isCommutative()) {
356 unsigned Op1 = getRegForValue(I->getOperand(1));
357 if (Op1 == 0) return false;
359 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
361 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op1,
362 Op1IsKill, CI->getZExtValue(),
364 if (ResultReg == 0) return false;
366 // We successfully emitted code for the given LLVM Instruction.
367 UpdateValueMap(I, ResultReg);
372 unsigned Op0 = getRegForValue(I->getOperand(0));
373 if (Op0 == 0) // Unhandled operand. Halt "fast" selection and bail.
376 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
378 // Check if the second operand is a constant and handle it appropriately.
379 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
380 uint64_t Imm = CI->getZExtValue();
382 // Transform "sdiv exact X, 8" -> "sra X, 3".
383 if (ISDOpcode == ISD::SDIV && isa<BinaryOperator>(I) &&
384 cast<BinaryOperator>(I)->isExact() &&
385 isPowerOf2_64(Imm)) {
387 ISDOpcode = ISD::SRA;
390 unsigned ResultReg = FastEmit_ri_(VT.getSimpleVT(), ISDOpcode, Op0,
391 Op0IsKill, Imm, VT.getSimpleVT());
392 if (ResultReg == 0) return false;
394 // We successfully emitted code for the given LLVM Instruction.
395 UpdateValueMap(I, ResultReg);
399 // Check if the second operand is a constant float.
400 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
401 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
402 ISDOpcode, Op0, Op0IsKill, CF);
403 if (ResultReg != 0) {
404 // We successfully emitted code for the given LLVM Instruction.
405 UpdateValueMap(I, ResultReg);
410 unsigned Op1 = getRegForValue(I->getOperand(1));
412 // Unhandled operand. Halt "fast" selection and bail.
415 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
417 // Now we have both operands in registers. Emit the instruction.
418 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
423 // Target-specific code wasn't able to find a machine opcode for
424 // the given ISD opcode and type. Halt "fast" selection and bail.
427 // We successfully emitted code for the given LLVM Instruction.
428 UpdateValueMap(I, ResultReg);
432 bool FastISel::SelectGetElementPtr(const User *I) {
433 unsigned N = getRegForValue(I->getOperand(0));
435 // Unhandled operand. Halt "fast" selection and bail.
438 bool NIsKill = hasTrivialKill(I->getOperand(0));
440 // Keep a running tab of the total offset to coalesce multiple N = N + Offset
441 // into a single N = N + TotalOffset.
442 uint64_t TotalOffs = 0;
443 // FIXME: What's a good SWAG number for MaxOffs?
444 uint64_t MaxOffs = 2048;
445 Type *Ty = I->getOperand(0)->getType();
446 MVT VT = TLI.getPointerTy();
447 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
448 E = I->op_end(); OI != E; ++OI) {
449 const Value *Idx = *OI;
450 if (StructType *StTy = dyn_cast<StructType>(Ty)) {
451 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
454 TotalOffs += TD.getStructLayout(StTy)->getElementOffset(Field);
455 if (TotalOffs >= MaxOffs) {
456 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
458 // Unhandled operand. Halt "fast" selection and bail.
464 Ty = StTy->getElementType(Field);
466 Ty = cast<SequentialType>(Ty)->getElementType();
468 // If this is a constant subscript, handle it quickly.
469 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
470 if (CI->isZero()) continue;
473 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
474 if (TotalOffs >= MaxOffs) {
475 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
477 // Unhandled operand. Halt "fast" selection and bail.
485 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
487 // Unhandled operand. Halt "fast" selection and bail.
493 // N = N + Idx * ElementSize;
494 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
495 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
496 unsigned IdxN = Pair.first;
497 bool IdxNIsKill = Pair.second;
499 // Unhandled operand. Halt "fast" selection and bail.
502 if (ElementSize != 1) {
503 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
505 // Unhandled operand. Halt "fast" selection and bail.
509 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
511 // Unhandled operand. Halt "fast" selection and bail.
516 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, TotalOffs, VT);
518 // Unhandled operand. Halt "fast" selection and bail.
522 // We successfully emitted code for the given LLVM Instruction.
523 UpdateValueMap(I, N);
527 bool FastISel::SelectCall(const User *I) {
528 const CallInst *Call = cast<CallInst>(I);
530 // Handle simple inline asms.
531 if (const InlineAsm *IA = dyn_cast<InlineAsm>(Call->getCalledValue())) {
532 // Don't attempt to handle constraints.
533 if (!IA->getConstraintString().empty())
536 unsigned ExtraInfo = 0;
537 if (IA->hasSideEffects())
538 ExtraInfo |= InlineAsm::Extra_HasSideEffects;
539 if (IA->isAlignStack())
540 ExtraInfo |= InlineAsm::Extra_IsAlignStack;
542 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
543 TII.get(TargetOpcode::INLINEASM))
544 .addExternalSymbol(IA->getAsmString().c_str())
549 const Function *F = Call->getCalledFunction();
550 if (!F) return false;
552 // Handle selected intrinsic function calls.
553 switch (F->getIntrinsicID()) {
555 case Intrinsic::dbg_declare: {
556 const DbgDeclareInst *DI = cast<DbgDeclareInst>(Call);
557 if (!DIVariable(DI->getVariable()).Verify() ||
558 !FuncInfo.MF->getMMI().hasDebugInfo())
561 const Value *Address = DI->getAddress();
562 if (!Address || isa<UndefValue>(Address) || isa<AllocaInst>(Address))
567 if (const Argument *Arg = dyn_cast<Argument>(Address)) {
568 // Some arguments' frame index is recorded during argument lowering.
569 Offset = FuncInfo.getArgumentFrameIndex(Arg);
571 Reg = TRI.getFrameRegister(*FuncInfo.MF);
574 Reg = getRegForValue(Address);
577 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
578 TII.get(TargetOpcode::DBG_VALUE))
579 .addReg(Reg, RegState::Debug).addImm(Offset)
580 .addMetadata(DI->getVariable());
583 case Intrinsic::dbg_value: {
584 // This form of DBG_VALUE is target-independent.
585 const DbgValueInst *DI = cast<DbgValueInst>(Call);
586 const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
587 const Value *V = DI->getValue();
589 // Currently the optimizer can produce this; insert an undef to
590 // help debugging. Probably the optimizer should not do this.
591 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
592 .addReg(0U).addImm(DI->getOffset())
593 .addMetadata(DI->getVariable());
594 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
595 if (CI->getBitWidth() > 64)
596 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
597 .addCImm(CI).addImm(DI->getOffset())
598 .addMetadata(DI->getVariable());
600 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
601 .addImm(CI->getZExtValue()).addImm(DI->getOffset())
602 .addMetadata(DI->getVariable());
603 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
604 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
605 .addFPImm(CF).addImm(DI->getOffset())
606 .addMetadata(DI->getVariable());
607 } else if (unsigned Reg = lookUpRegForValue(V)) {
608 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
609 .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
610 .addMetadata(DI->getVariable());
612 // We can't yet handle anything else here because it would require
613 // generating code, thus altering codegen because of debug info.
614 DEBUG(dbgs() << "Dropping debug info for " << DI);
618 case Intrinsic::eh_exception: {
619 EVT VT = TLI.getValueType(Call->getType());
620 if (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)!=TargetLowering::Expand)
623 assert(FuncInfo.MBB->isLandingPad() &&
624 "Call to eh.exception not in landing pad!");
625 unsigned Reg = TLI.getExceptionAddressRegister();
626 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
627 unsigned ResultReg = createResultReg(RC);
628 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
629 ResultReg).addReg(Reg);
630 UpdateValueMap(Call, ResultReg);
633 case Intrinsic::eh_selector: {
634 EVT VT = TLI.getValueType(Call->getType());
635 if (TLI.getOperationAction(ISD::EHSELECTION, VT) != TargetLowering::Expand)
637 if (FuncInfo.MBB->isLandingPad())
638 AddCatchInfo(*Call, &FuncInfo.MF->getMMI(), FuncInfo.MBB);
641 FuncInfo.CatchInfoLost.insert(Call);
643 // FIXME: Mark exception selector register as live in. Hack for PR1508.
644 unsigned Reg = TLI.getExceptionSelectorRegister();
645 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
648 unsigned Reg = TLI.getExceptionSelectorRegister();
649 EVT SrcVT = TLI.getPointerTy();
650 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
651 unsigned ResultReg = createResultReg(RC);
652 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
653 ResultReg).addReg(Reg);
655 bool ResultRegIsKill = hasTrivialKill(Call);
657 // Cast the register to the type of the selector.
658 if (SrcVT.bitsGT(MVT::i32))
659 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
660 ResultReg, ResultRegIsKill);
661 else if (SrcVT.bitsLT(MVT::i32))
662 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
663 ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill);
665 // Unhandled operand. Halt "fast" selection and bail.
668 UpdateValueMap(Call, ResultReg);
672 case Intrinsic::objectsize: {
673 ConstantInt *CI = cast<ConstantInt>(Call->getArgOperand(1));
674 unsigned long long Res = CI->isZero() ? -1ULL : 0;
675 Constant *ResCI = ConstantInt::get(Call->getType(), Res);
676 unsigned ResultReg = getRegForValue(ResCI);
679 UpdateValueMap(Call, ResultReg);
684 // Usually, it does not make sense to initialize a value,
685 // make an unrelated function call and use the value, because
686 // it tends to be spilled on the stack. So, we move the pointer
687 // to the last local value to the beginning of the block, so that
688 // all the values which have already been materialized,
689 // appear after the call. It also makes sense to skip intrinsics
690 // since they tend to be inlined.
691 if (!isa<IntrinsicInst>(F))
692 flushLocalValueMap();
694 // An arbitrary call. Bail.
698 bool FastISel::SelectCast(const User *I, unsigned Opcode) {
699 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
700 EVT DstVT = TLI.getValueType(I->getType());
702 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
703 DstVT == MVT::Other || !DstVT.isSimple())
704 // Unhandled type. Halt "fast" selection and bail.
707 // Check if the destination type is legal.
708 if (!TLI.isTypeLegal(DstVT))
711 // Check if the source operand is legal.
712 if (!TLI.isTypeLegal(SrcVT))
715 unsigned InputReg = getRegForValue(I->getOperand(0));
717 // Unhandled operand. Halt "fast" selection and bail.
720 bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
722 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
725 InputReg, InputRegIsKill);
729 UpdateValueMap(I, ResultReg);
733 bool FastISel::SelectBitCast(const User *I) {
734 // If the bitcast doesn't change the type, just use the operand value.
735 if (I->getType() == I->getOperand(0)->getType()) {
736 unsigned Reg = getRegForValue(I->getOperand(0));
739 UpdateValueMap(I, Reg);
743 // Bitcasts of other values become reg-reg copies or BITCAST operators.
744 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
745 EVT DstVT = TLI.getValueType(I->getType());
747 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
748 DstVT == MVT::Other || !DstVT.isSimple() ||
749 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
750 // Unhandled type. Halt "fast" selection and bail.
753 unsigned Op0 = getRegForValue(I->getOperand(0));
755 // Unhandled operand. Halt "fast" selection and bail.
758 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
760 // First, try to perform the bitcast by inserting a reg-reg copy.
761 unsigned ResultReg = 0;
762 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
763 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
764 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
765 // Don't attempt a cross-class copy. It will likely fail.
766 if (SrcClass == DstClass) {
767 ResultReg = createResultReg(DstClass);
768 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
769 ResultReg).addReg(Op0);
773 // If the reg-reg copy failed, select a BITCAST opcode.
775 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
776 ISD::BITCAST, Op0, Op0IsKill);
781 UpdateValueMap(I, ResultReg);
786 FastISel::SelectInstruction(const Instruction *I) {
787 // Just before the terminator instruction, insert instructions to
788 // feed PHI nodes in successor blocks.
789 if (isa<TerminatorInst>(I))
790 if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
793 DL = I->getDebugLoc();
795 // First, try doing target-independent selection.
796 if (SelectOperator(I, I->getOpcode())) {
797 ++NumFastIselSuccessIndependent;
802 // Next, try calling the target to attempt to handle the instruction.
803 if (TargetSelectInstruction(I)) {
804 ++NumFastIselSuccessTarget;
813 /// FastEmitBranch - Emit an unconditional branch to the given block,
814 /// unless it is the immediate (fall-through) successor, and update
817 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
818 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
819 // The unconditional fall-through case, which needs no instructions.
821 // The unconditional branch case.
822 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
823 SmallVector<MachineOperand, 0>(), DL);
825 FuncInfo.MBB->addSuccessor(MSucc);
828 /// SelectFNeg - Emit an FNeg operation.
831 FastISel::SelectFNeg(const User *I) {
832 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
833 if (OpReg == 0) return false;
835 bool OpRegIsKill = hasTrivialKill(I);
837 // If the target has ISD::FNEG, use it.
838 EVT VT = TLI.getValueType(I->getType());
839 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
840 ISD::FNEG, OpReg, OpRegIsKill);
841 if (ResultReg != 0) {
842 UpdateValueMap(I, ResultReg);
846 // Bitcast the value to integer, twiddle the sign bit with xor,
847 // and then bitcast it back to floating-point.
848 if (VT.getSizeInBits() > 64) return false;
849 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
850 if (!TLI.isTypeLegal(IntVT))
853 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
854 ISD::BITCAST, OpReg, OpRegIsKill);
858 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
859 IntReg, /*Kill=*/true,
860 UINT64_C(1) << (VT.getSizeInBits()-1),
861 IntVT.getSimpleVT());
862 if (IntResultReg == 0)
865 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
866 ISD::BITCAST, IntResultReg, /*Kill=*/true);
870 UpdateValueMap(I, ResultReg);
875 FastISel::SelectExtractValue(const User *U) {
876 const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(U);
880 // Make sure we only try to handle extracts with a legal result. But also
881 // allow i1 because it's easy.
882 EVT RealVT = TLI.getValueType(EVI->getType(), /*AllowUnknown=*/true);
883 if (!RealVT.isSimple())
885 MVT VT = RealVT.getSimpleVT();
886 if (!TLI.isTypeLegal(VT) && VT != MVT::i1)
889 const Value *Op0 = EVI->getOperand(0);
890 Type *AggTy = Op0->getType();
892 // Get the base result register.
894 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(Op0);
895 if (I != FuncInfo.ValueMap.end())
896 ResultReg = I->second;
897 else if (isa<Instruction>(Op0))
898 ResultReg = FuncInfo.InitializeRegForValue(Op0);
900 return false; // fast-isel can't handle aggregate constants at the moment
902 // Get the actual result register, which is an offset from the base register.
903 unsigned VTIndex = ComputeLinearIndex(AggTy, EVI->getIndices());
905 SmallVector<EVT, 4> AggValueVTs;
906 ComputeValueVTs(TLI, AggTy, AggValueVTs);
908 for (unsigned i = 0; i < VTIndex; i++)
909 ResultReg += TLI.getNumRegisters(FuncInfo.Fn->getContext(), AggValueVTs[i]);
911 UpdateValueMap(EVI, ResultReg);
916 FastISel::SelectOperator(const User *I, unsigned Opcode) {
918 case Instruction::Add:
919 return SelectBinaryOp(I, ISD::ADD);
920 case Instruction::FAdd:
921 return SelectBinaryOp(I, ISD::FADD);
922 case Instruction::Sub:
923 return SelectBinaryOp(I, ISD::SUB);
924 case Instruction::FSub:
925 // FNeg is currently represented in LLVM IR as a special case of FSub.
926 if (BinaryOperator::isFNeg(I))
927 return SelectFNeg(I);
928 return SelectBinaryOp(I, ISD::FSUB);
929 case Instruction::Mul:
930 return SelectBinaryOp(I, ISD::MUL);
931 case Instruction::FMul:
932 return SelectBinaryOp(I, ISD::FMUL);
933 case Instruction::SDiv:
934 return SelectBinaryOp(I, ISD::SDIV);
935 case Instruction::UDiv:
936 return SelectBinaryOp(I, ISD::UDIV);
937 case Instruction::FDiv:
938 return SelectBinaryOp(I, ISD::FDIV);
939 case Instruction::SRem:
940 return SelectBinaryOp(I, ISD::SREM);
941 case Instruction::URem:
942 return SelectBinaryOp(I, ISD::UREM);
943 case Instruction::FRem:
944 return SelectBinaryOp(I, ISD::FREM);
945 case Instruction::Shl:
946 return SelectBinaryOp(I, ISD::SHL);
947 case Instruction::LShr:
948 return SelectBinaryOp(I, ISD::SRL);
949 case Instruction::AShr:
950 return SelectBinaryOp(I, ISD::SRA);
951 case Instruction::And:
952 return SelectBinaryOp(I, ISD::AND);
953 case Instruction::Or:
954 return SelectBinaryOp(I, ISD::OR);
955 case Instruction::Xor:
956 return SelectBinaryOp(I, ISD::XOR);
958 case Instruction::GetElementPtr:
959 return SelectGetElementPtr(I);
961 case Instruction::Br: {
962 const BranchInst *BI = cast<BranchInst>(I);
964 if (BI->isUnconditional()) {
965 const BasicBlock *LLVMSucc = BI->getSuccessor(0);
966 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
967 FastEmitBranch(MSucc, BI->getDebugLoc());
971 // Conditional branches are not handed yet.
972 // Halt "fast" selection and bail.
976 case Instruction::Unreachable:
980 case Instruction::Alloca:
981 // FunctionLowering has the static-sized case covered.
982 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
985 // Dynamic-sized alloca is not handled yet.
988 case Instruction::Call:
989 return SelectCall(I);
991 case Instruction::BitCast:
992 return SelectBitCast(I);
994 case Instruction::FPToSI:
995 return SelectCast(I, ISD::FP_TO_SINT);
996 case Instruction::ZExt:
997 return SelectCast(I, ISD::ZERO_EXTEND);
998 case Instruction::SExt:
999 return SelectCast(I, ISD::SIGN_EXTEND);
1000 case Instruction::Trunc:
1001 return SelectCast(I, ISD::TRUNCATE);
1002 case Instruction::SIToFP:
1003 return SelectCast(I, ISD::SINT_TO_FP);
1005 case Instruction::IntToPtr: // Deliberate fall-through.
1006 case Instruction::PtrToInt: {
1007 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
1008 EVT DstVT = TLI.getValueType(I->getType());
1009 if (DstVT.bitsGT(SrcVT))
1010 return SelectCast(I, ISD::ZERO_EXTEND);
1011 if (DstVT.bitsLT(SrcVT))
1012 return SelectCast(I, ISD::TRUNCATE);
1013 unsigned Reg = getRegForValue(I->getOperand(0));
1014 if (Reg == 0) return false;
1015 UpdateValueMap(I, Reg);
1019 case Instruction::ExtractValue:
1020 return SelectExtractValue(I);
1022 case Instruction::PHI:
1023 llvm_unreachable("FastISel shouldn't visit PHI nodes!");
1026 // Unhandled instruction. Halt "fast" selection and bail.
1031 FastISel::FastISel(FunctionLoweringInfo &funcInfo)
1032 : FuncInfo(funcInfo),
1033 MRI(FuncInfo.MF->getRegInfo()),
1034 MFI(*FuncInfo.MF->getFrameInfo()),
1035 MCP(*FuncInfo.MF->getConstantPool()),
1036 TM(FuncInfo.MF->getTarget()),
1037 TD(*TM.getTargetData()),
1038 TII(*TM.getInstrInfo()),
1039 TLI(*TM.getTargetLowering()),
1040 TRI(*TM.getRegisterInfo()) {
1043 FastISel::~FastISel() {}
1045 unsigned FastISel::FastEmit_(MVT, MVT,
1050 unsigned FastISel::FastEmit_r(MVT, MVT,
1052 unsigned /*Op0*/, bool /*Op0IsKill*/) {
1056 unsigned FastISel::FastEmit_rr(MVT, MVT,
1058 unsigned /*Op0*/, bool /*Op0IsKill*/,
1059 unsigned /*Op1*/, bool /*Op1IsKill*/) {
1063 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
1067 unsigned FastISel::FastEmit_f(MVT, MVT,
1068 unsigned, const ConstantFP * /*FPImm*/) {
1072 unsigned FastISel::FastEmit_ri(MVT, MVT,
1074 unsigned /*Op0*/, bool /*Op0IsKill*/,
1079 unsigned FastISel::FastEmit_rf(MVT, MVT,
1081 unsigned /*Op0*/, bool /*Op0IsKill*/,
1082 const ConstantFP * /*FPImm*/) {
1086 unsigned FastISel::FastEmit_rri(MVT, MVT,
1088 unsigned /*Op0*/, bool /*Op0IsKill*/,
1089 unsigned /*Op1*/, bool /*Op1IsKill*/,
1094 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
1095 /// to emit an instruction with an immediate operand using FastEmit_ri.
1096 /// If that fails, it materializes the immediate into a register and try
1097 /// FastEmit_rr instead.
1098 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
1099 unsigned Op0, bool Op0IsKill,
1100 uint64_t Imm, MVT ImmType) {
1101 // If this is a multiply by a power of two, emit this as a shift left.
1102 if (Opcode == ISD::MUL && isPowerOf2_64(Imm)) {
1105 } else if (Opcode == ISD::UDIV && isPowerOf2_64(Imm)) {
1106 // div x, 8 -> srl x, 3
1111 // Horrible hack (to be removed), check to make sure shift amounts are
1113 if ((Opcode == ISD::SHL || Opcode == ISD::SRA || Opcode == ISD::SRL) &&
1114 Imm >= VT.getSizeInBits())
1117 // First check if immediate type is legal. If not, we can't use the ri form.
1118 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
1121 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
1122 if (MaterialReg == 0) {
1123 // This is a bit ugly/slow, but failing here means falling out of
1124 // fast-isel, which would be very slow.
1125 IntegerType *ITy = IntegerType::get(FuncInfo.Fn->getContext(),
1126 VT.getSizeInBits());
1127 MaterialReg = getRegForValue(ConstantInt::get(ITy, Imm));
1129 return FastEmit_rr(VT, VT, Opcode,
1131 MaterialReg, /*Kill=*/true);
1134 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
1135 return MRI.createVirtualRegister(RC);
1138 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
1139 const TargetRegisterClass* RC) {
1140 unsigned ResultReg = createResultReg(RC);
1141 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1143 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
1147 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
1148 const TargetRegisterClass *RC,
1149 unsigned Op0, bool Op0IsKill) {
1150 unsigned ResultReg = createResultReg(RC);
1151 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1153 if (II.getNumDefs() >= 1)
1154 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1155 .addReg(Op0, Op0IsKill * RegState::Kill);
1157 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1158 .addReg(Op0, Op0IsKill * RegState::Kill);
1159 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1160 ResultReg).addReg(II.ImplicitDefs[0]);
1166 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
1167 const TargetRegisterClass *RC,
1168 unsigned Op0, bool Op0IsKill,
1169 unsigned Op1, bool Op1IsKill) {
1170 unsigned ResultReg = createResultReg(RC);
1171 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1173 if (II.getNumDefs() >= 1)
1174 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1175 .addReg(Op0, Op0IsKill * RegState::Kill)
1176 .addReg(Op1, Op1IsKill * RegState::Kill);
1178 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1179 .addReg(Op0, Op0IsKill * RegState::Kill)
1180 .addReg(Op1, Op1IsKill * RegState::Kill);
1181 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1182 ResultReg).addReg(II.ImplicitDefs[0]);
1187 unsigned FastISel::FastEmitInst_rrr(unsigned MachineInstOpcode,
1188 const TargetRegisterClass *RC,
1189 unsigned Op0, bool Op0IsKill,
1190 unsigned Op1, bool Op1IsKill,
1191 unsigned Op2, bool Op2IsKill) {
1192 unsigned ResultReg = createResultReg(RC);
1193 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1195 if (II.getNumDefs() >= 1)
1196 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1197 .addReg(Op0, Op0IsKill * RegState::Kill)
1198 .addReg(Op1, Op1IsKill * RegState::Kill)
1199 .addReg(Op2, Op2IsKill * RegState::Kill);
1201 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1202 .addReg(Op0, Op0IsKill * RegState::Kill)
1203 .addReg(Op1, Op1IsKill * RegState::Kill)
1204 .addReg(Op2, Op2IsKill * RegState::Kill);
1205 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1206 ResultReg).addReg(II.ImplicitDefs[0]);
1211 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
1212 const TargetRegisterClass *RC,
1213 unsigned Op0, bool Op0IsKill,
1215 unsigned ResultReg = createResultReg(RC);
1216 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1218 if (II.getNumDefs() >= 1)
1219 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1220 .addReg(Op0, Op0IsKill * RegState::Kill)
1223 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1224 .addReg(Op0, Op0IsKill * RegState::Kill)
1226 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1227 ResultReg).addReg(II.ImplicitDefs[0]);
1232 unsigned FastISel::FastEmitInst_rii(unsigned MachineInstOpcode,
1233 const TargetRegisterClass *RC,
1234 unsigned Op0, bool Op0IsKill,
1235 uint64_t Imm1, uint64_t Imm2) {
1236 unsigned ResultReg = createResultReg(RC);
1237 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1239 if (II.getNumDefs() >= 1)
1240 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1241 .addReg(Op0, Op0IsKill * RegState::Kill)
1245 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1246 .addReg(Op0, Op0IsKill * RegState::Kill)
1249 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1250 ResultReg).addReg(II.ImplicitDefs[0]);
1255 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
1256 const TargetRegisterClass *RC,
1257 unsigned Op0, bool Op0IsKill,
1258 const ConstantFP *FPImm) {
1259 unsigned ResultReg = createResultReg(RC);
1260 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1262 if (II.getNumDefs() >= 1)
1263 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1264 .addReg(Op0, Op0IsKill * RegState::Kill)
1267 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1268 .addReg(Op0, Op0IsKill * RegState::Kill)
1270 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1271 ResultReg).addReg(II.ImplicitDefs[0]);
1276 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
1277 const TargetRegisterClass *RC,
1278 unsigned Op0, bool Op0IsKill,
1279 unsigned Op1, bool Op1IsKill,
1281 unsigned ResultReg = createResultReg(RC);
1282 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1284 if (II.getNumDefs() >= 1)
1285 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1286 .addReg(Op0, Op0IsKill * RegState::Kill)
1287 .addReg(Op1, Op1IsKill * RegState::Kill)
1290 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1291 .addReg(Op0, Op0IsKill * RegState::Kill)
1292 .addReg(Op1, Op1IsKill * RegState::Kill)
1294 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1295 ResultReg).addReg(II.ImplicitDefs[0]);
1300 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1301 const TargetRegisterClass *RC,
1303 unsigned ResultReg = createResultReg(RC);
1304 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1306 if (II.getNumDefs() >= 1)
1307 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
1309 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
1310 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1311 ResultReg).addReg(II.ImplicitDefs[0]);
1316 unsigned FastISel::FastEmitInst_ii(unsigned MachineInstOpcode,
1317 const TargetRegisterClass *RC,
1318 uint64_t Imm1, uint64_t Imm2) {
1319 unsigned ResultReg = createResultReg(RC);
1320 const MCInstrDesc &II = TII.get(MachineInstOpcode);
1322 if (II.getNumDefs() >= 1)
1323 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1324 .addImm(Imm1).addImm(Imm2);
1326 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm1).addImm(Imm2);
1327 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
1328 ResultReg).addReg(II.ImplicitDefs[0]);
1333 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
1334 unsigned Op0, bool Op0IsKill,
1336 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1337 assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
1338 "Cannot yet extract from physregs");
1339 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
1340 DL, TII.get(TargetOpcode::COPY), ResultReg)
1341 .addReg(Op0, getKillRegState(Op0IsKill), Idx);
1345 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1346 /// with all but the least significant bit set to zero.
1347 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
1348 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
1351 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
1352 /// Emit code to ensure constants are copied into registers when needed.
1353 /// Remember the virtual registers that need to be added to the Machine PHI
1354 /// nodes as input. We cannot just directly add them, because expansion
1355 /// might result in multiple MBB's for one BB. As such, the start of the
1356 /// BB might correspond to a different MBB than the end.
1357 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
1358 const TerminatorInst *TI = LLVMBB->getTerminator();
1360 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
1361 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
1363 // Check successor nodes' PHI nodes that expect a constant to be available
1365 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
1366 const BasicBlock *SuccBB = TI->getSuccessor(succ);
1367 if (!isa<PHINode>(SuccBB->begin())) continue;
1368 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
1370 // If this terminator has multiple identical successors (common for
1371 // switches), only handle each succ once.
1372 if (!SuccsHandled.insert(SuccMBB)) continue;
1374 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
1376 // At this point we know that there is a 1-1 correspondence between LLVM PHI
1377 // nodes and Machine PHI nodes, but the incoming operands have not been
1379 for (BasicBlock::const_iterator I = SuccBB->begin();
1380 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1382 // Ignore dead phi's.
1383 if (PN->use_empty()) continue;
1385 // Only handle legal types. Two interesting things to note here. First,
1386 // by bailing out early, we may leave behind some dead instructions,
1387 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
1388 // own moves. Second, this check is necessary because FastISel doesn't
1389 // use CreateRegs to create registers, so it always creates
1390 // exactly one register for each non-void instruction.
1391 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
1392 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
1395 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
1397 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1402 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
1404 // Set the DebugLoc for the copy. Prefer the location of the operand
1405 // if there is one; use the location of the PHI otherwise.
1406 DL = PN->getDebugLoc();
1407 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
1408 DL = Inst->getDebugLoc();
1410 unsigned Reg = getRegForValue(PHIOp);
1412 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1415 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));