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 end(), meaining no local-value instructions have
68 LastLocalValue = FuncInfo.MBB->end();
71 bool FastISel::hasTrivialKill(const Value *V) const {
72 // Don't consider constants or arguments to have trivial kills.
73 const Instruction *I = dyn_cast<Instruction>(V);
77 // No-op casts are trivially coalesced by fast-isel.
78 if (const CastInst *Cast = dyn_cast<CastInst>(I))
79 if (Cast->isNoopCast(TD.getIntPtrType(Cast->getContext())) &&
80 !hasTrivialKill(Cast->getOperand(0)))
83 // Only instructions with a single use in the same basic block are considered
84 // to have trivial kills.
85 return I->hasOneUse() &&
86 !(I->getOpcode() == Instruction::BitCast ||
87 I->getOpcode() == Instruction::PtrToInt ||
88 I->getOpcode() == Instruction::IntToPtr) &&
89 cast<Instruction>(I->use_begin())->getParent() == I->getParent();
92 unsigned FastISel::getRegForValue(const Value *V) {
93 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
94 // Don't handle non-simple values in FastISel.
95 if (!RealVT.isSimple())
98 // Ignore illegal types. We must do this before looking up the value
99 // in ValueMap because Arguments are given virtual registers regardless
100 // of whether FastISel can handle them.
101 MVT VT = RealVT.getSimpleVT();
102 if (!TLI.isTypeLegal(VT)) {
103 // Promote MVT::i1 to a legal type though, because it's common and easy.
105 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
110 // Look up the value to see if we already have a register for it. We
111 // cache values defined by Instructions across blocks, and other values
112 // only locally. This is because Instructions already have the SSA
113 // def-dominates-use requirement enforced.
114 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
115 if (I != FuncInfo.ValueMap.end())
117 unsigned Reg = LocalValueMap[V];
121 // In bottom-up mode, just create the virtual register which will be used
122 // to hold the value. It will be materialized later.
123 if (isa<Instruction>(V) &&
124 (!isa<AllocaInst>(V) ||
125 !FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V)))) {
126 Reg = createResultReg(TLI.getRegClassFor(VT));
127 FuncInfo.ValueMap[V] = Reg;
131 return materializeRegForValue(V, VT);
134 /// materializeRegForValue - Helper for getRegForVale. This function is
135 /// called when the value isn't already available in a register and must
136 /// be materialized with new instructions.
137 unsigned FastISel::materializeRegForValue(const Value *V, MVT VT) {
140 if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
141 if (CI->getValue().getActiveBits() <= 64)
142 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
143 } else if (isa<AllocaInst>(V)) {
144 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
145 } else if (isa<ConstantPointerNull>(V)) {
146 // Translate this as an integer zero so that it can be
147 // local-CSE'd with actual integer zeros.
149 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
150 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
151 // Try to emit the constant directly.
152 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
155 // Try to emit the constant by using an integer constant with a cast.
156 const APFloat &Flt = CF->getValueAPF();
157 EVT IntVT = TLI.getPointerTy();
160 uint32_t IntBitWidth = IntVT.getSizeInBits();
162 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
163 APFloat::rmTowardZero, &isExact);
165 APInt IntVal(IntBitWidth, 2, x);
167 unsigned IntegerReg =
168 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
170 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP,
171 IntegerReg, /*Kill=*/false);
174 } else if (const Operator *Op = dyn_cast<Operator>(V)) {
175 if (!SelectOperator(Op, Op->getOpcode()))
176 if (!isa<Instruction>(Op) ||
177 !TargetSelectInstruction(cast<Instruction>(Op)))
179 Reg = lookUpRegForValue(Op);
180 } else if (isa<UndefValue>(V)) {
181 Reg = createResultReg(TLI.getRegClassFor(VT));
182 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
183 TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
186 // If target-independent code couldn't handle the value, give target-specific
188 if (!Reg && isa<Constant>(V))
189 Reg = TargetMaterializeConstant(cast<Constant>(V));
191 // Don't cache constant materializations in the general ValueMap.
192 // To do so would require tracking what uses they dominate.
194 LocalValueMap[V] = Reg;
195 LastLocalValue = MRI.getVRegDef(Reg);
200 unsigned FastISel::lookUpRegForValue(const Value *V) {
201 // Look up the value to see if we already have a register for it. We
202 // cache values defined by Instructions across blocks, and other values
203 // only locally. This is because Instructions already have the SSA
204 // def-dominates-use requirement enforced.
205 DenseMap<const Value *, unsigned>::iterator I = FuncInfo.ValueMap.find(V);
206 if (I != FuncInfo.ValueMap.end())
208 return LocalValueMap[V];
211 /// UpdateValueMap - Update the value map to include the new mapping for this
212 /// instruction, or insert an extra copy to get the result in a previous
213 /// determined register.
214 /// NOTE: This is only necessary because we might select a block that uses
215 /// a value before we select the block that defines the value. It might be
216 /// possible to fix this by selecting blocks in reverse postorder.
217 unsigned FastISel::UpdateValueMap(const Value *I, unsigned Reg) {
218 if (!isa<Instruction>(I)) {
219 LocalValueMap[I] = Reg;
223 unsigned &AssignedReg = FuncInfo.ValueMap[I];
224 if (AssignedReg == 0)
225 // Use the new register.
227 else if (Reg != AssignedReg) {
228 // We already have a register for this value. Replace uses of
229 // the existing register with uses of the new one.
230 MRI.replaceRegWith(AssignedReg, Reg);
231 // Replace uses of the existing register in PHINodesToUpdate too.
232 for (unsigned i = 0, e = FuncInfo.PHINodesToUpdate.size(); i != e; ++i)
233 if (FuncInfo.PHINodesToUpdate[i].second == AssignedReg)
234 FuncInfo.PHINodesToUpdate[i].second = Reg;
235 // And update the ValueMap.
242 std::pair<unsigned, bool> FastISel::getRegForGEPIndex(const Value *Idx) {
243 unsigned IdxN = getRegForValue(Idx);
245 // Unhandled operand. Halt "fast" selection and bail.
246 return std::pair<unsigned, bool>(0, false);
248 bool IdxNIsKill = hasTrivialKill(Idx);
250 // If the index is smaller or larger than intptr_t, truncate or extend it.
251 MVT PtrVT = TLI.getPointerTy();
252 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
253 if (IdxVT.bitsLT(PtrVT)) {
254 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND,
258 else if (IdxVT.bitsGT(PtrVT)) {
259 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE,
263 return std::pair<unsigned, bool>(IdxN, IdxNIsKill);
266 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
267 /// which has an opcode which directly corresponds to the given ISD opcode.
269 bool FastISel::SelectBinaryOp(const User *I, unsigned ISDOpcode) {
270 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
271 if (VT == MVT::Other || !VT.isSimple())
272 // Unhandled type. Halt "fast" selection and bail.
275 // We only handle legal types. For example, on x86-32 the instruction
276 // selector contains all of the 64-bit instructions from x86-64,
277 // under the assumption that i64 won't be used if the target doesn't
279 if (!TLI.isTypeLegal(VT)) {
280 // MVT::i1 is special. Allow AND, OR, or XOR because they
281 // don't require additional zeroing, which makes them easy.
283 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
284 ISDOpcode == ISD::XOR))
285 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
290 unsigned Op0 = getRegForValue(I->getOperand(0));
292 // Unhandled operand. Halt "fast" selection and bail.
295 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
297 // Check if the second operand is a constant and handle it appropriately.
298 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
299 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
300 ISDOpcode, Op0, Op0IsKill,
302 if (ResultReg != 0) {
303 // We successfully emitted code for the given LLVM Instruction.
304 UpdateValueMap(I, ResultReg);
309 // Check if the second operand is a constant float.
310 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
311 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
312 ISDOpcode, Op0, Op0IsKill, CF);
313 if (ResultReg != 0) {
314 // We successfully emitted code for the given LLVM Instruction.
315 UpdateValueMap(I, ResultReg);
320 unsigned Op1 = getRegForValue(I->getOperand(1));
322 // Unhandled operand. Halt "fast" selection and bail.
325 bool Op1IsKill = hasTrivialKill(I->getOperand(1));
327 // Now we have both operands in registers. Emit the instruction.
328 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
333 // Target-specific code wasn't able to find a machine opcode for
334 // the given ISD opcode and type. Halt "fast" selection and bail.
337 // We successfully emitted code for the given LLVM Instruction.
338 UpdateValueMap(I, ResultReg);
342 bool FastISel::SelectGetElementPtr(const User *I) {
343 unsigned N = getRegForValue(I->getOperand(0));
345 // Unhandled operand. Halt "fast" selection and bail.
348 bool NIsKill = hasTrivialKill(I->getOperand(0));
350 const Type *Ty = I->getOperand(0)->getType();
351 MVT VT = TLI.getPointerTy();
352 for (GetElementPtrInst::const_op_iterator OI = I->op_begin()+1,
353 E = I->op_end(); OI != E; ++OI) {
354 const Value *Idx = *OI;
355 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
356 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
359 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
360 // FIXME: This can be optimized by combining the add with a
362 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
364 // Unhandled operand. Halt "fast" selection and bail.
368 Ty = StTy->getElementType(Field);
370 Ty = cast<SequentialType>(Ty)->getElementType();
372 // If this is a constant subscript, handle it quickly.
373 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
374 if (CI->isZero()) continue;
376 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
377 N = FastEmit_ri_(VT, ISD::ADD, N, NIsKill, Offs, VT);
379 // Unhandled operand. Halt "fast" selection and bail.
385 // N = N + Idx * ElementSize;
386 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
387 std::pair<unsigned, bool> Pair = getRegForGEPIndex(Idx);
388 unsigned IdxN = Pair.first;
389 bool IdxNIsKill = Pair.second;
391 // Unhandled operand. Halt "fast" selection and bail.
394 if (ElementSize != 1) {
395 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, IdxNIsKill, ElementSize, VT);
397 // Unhandled operand. Halt "fast" selection and bail.
401 N = FastEmit_rr(VT, VT, ISD::ADD, N, NIsKill, IdxN, IdxNIsKill);
403 // Unhandled operand. Halt "fast" selection and bail.
408 // We successfully emitted code for the given LLVM Instruction.
409 UpdateValueMap(I, N);
413 bool FastISel::SelectCall(const User *I) {
414 const Function *F = cast<CallInst>(I)->getCalledFunction();
415 if (!F) return false;
417 // Handle selected intrinsic function calls.
418 unsigned IID = F->getIntrinsicID();
421 case Intrinsic::dbg_declare: {
422 const DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
423 if (!DIVariable(DI->getVariable()).Verify() ||
424 !FuncInfo.MF->getMMI().hasDebugInfo())
427 const Value *Address = DI->getAddress();
430 if (isa<UndefValue>(Address))
432 const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
433 // Don't handle byval struct arguments or VLAs, for example.
434 // Note that if we have a byval struct argument, fast ISel is turned off;
435 // those are handled in SelectionDAGBuilder.
437 DenseMap<const AllocaInst*, int>::iterator SI =
438 FuncInfo.StaticAllocaMap.find(AI);
439 if (SI == FuncInfo.StaticAllocaMap.end()) break; // VLAs.
441 if (!DI->getDebugLoc().isUnknown())
442 FuncInfo.MF->getMMI().setVariableDbgInfo(DI->getVariable(),
443 FI, DI->getDebugLoc());
445 // Building the map above is target independent. Generating DBG_VALUE
446 // inline is target dependent; do this now.
447 (void)TargetSelectInstruction(cast<Instruction>(I));
450 case Intrinsic::dbg_value: {
451 // This form of DBG_VALUE is target-independent.
452 const DbgValueInst *DI = cast<DbgValueInst>(I);
453 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
454 const Value *V = DI->getValue();
456 // Currently the optimizer can produce this; insert an undef to
457 // help debugging. Probably the optimizer should not do this.
458 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
459 .addReg(0U).addImm(DI->getOffset())
460 .addMetadata(DI->getVariable());
461 } else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
462 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
463 .addImm(CI->getZExtValue()).addImm(DI->getOffset())
464 .addMetadata(DI->getVariable());
465 } else if (const ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
466 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
467 .addFPImm(CF).addImm(DI->getOffset())
468 .addMetadata(DI->getVariable());
469 } else if (unsigned Reg = lookUpRegForValue(V)) {
470 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
471 .addReg(Reg, RegState::Debug).addImm(DI->getOffset())
472 .addMetadata(DI->getVariable());
474 // We can't yet handle anything else here because it would require
475 // generating code, thus altering codegen because of debug info.
476 // Insert an undef so we can see what we dropped.
477 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
478 .addReg(0U).addImm(DI->getOffset())
479 .addMetadata(DI->getVariable());
483 case Intrinsic::eh_exception: {
484 EVT VT = TLI.getValueType(I->getType());
485 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
487 case TargetLowering::Expand: {
488 assert(FuncInfo.MBB->isLandingPad() &&
489 "Call to eh.exception not in landing pad!");
490 unsigned Reg = TLI.getExceptionAddressRegister();
491 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
492 unsigned ResultReg = createResultReg(RC);
493 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
494 ResultReg, Reg, RC, RC, DL);
495 assert(InsertedCopy && "Can't copy address registers!");
496 InsertedCopy = InsertedCopy;
497 UpdateValueMap(I, ResultReg);
503 case Intrinsic::eh_selector: {
504 EVT VT = TLI.getValueType(I->getType());
505 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
507 case TargetLowering::Expand: {
508 if (FuncInfo.MBB->isLandingPad())
509 AddCatchInfo(*cast<CallInst>(I), &FuncInfo.MF->getMMI(), FuncInfo.MBB);
512 FuncInfo.CatchInfoLost.insert(cast<CallInst>(I));
514 // FIXME: Mark exception selector register as live in. Hack for PR1508.
515 unsigned Reg = TLI.getExceptionSelectorRegister();
516 if (Reg) FuncInfo.MBB->addLiveIn(Reg);
519 unsigned Reg = TLI.getExceptionSelectorRegister();
520 EVT SrcVT = TLI.getPointerTy();
521 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
522 unsigned ResultReg = createResultReg(RC);
523 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
524 ResultReg, Reg, RC, RC, DL);
525 assert(InsertedCopy && "Can't copy address registers!");
526 InsertedCopy = InsertedCopy;
528 bool ResultRegIsKill = hasTrivialKill(I);
530 // Cast the register to the type of the selector.
531 if (SrcVT.bitsGT(MVT::i32))
532 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
533 ResultReg, ResultRegIsKill);
534 else if (SrcVT.bitsLT(MVT::i32))
535 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
536 ISD::SIGN_EXTEND, ResultReg, ResultRegIsKill);
538 // Unhandled operand. Halt "fast" selection and bail.
541 UpdateValueMap(I, ResultReg);
550 // An arbitrary call. Bail.
554 bool FastISel::SelectCast(const User *I, unsigned Opcode) {
555 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
556 EVT DstVT = TLI.getValueType(I->getType());
558 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
559 DstVT == MVT::Other || !DstVT.isSimple())
560 // Unhandled type. Halt "fast" selection and bail.
563 // Check if the destination type is legal. Or as a special case,
564 // it may be i1 if we're doing a truncate because that's
565 // easy and somewhat common.
566 if (!TLI.isTypeLegal(DstVT))
567 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
568 // Unhandled type. Halt "fast" selection and bail.
571 // Check if the source operand is legal. Or as a special case,
572 // it may be i1 if we're doing zero-extension because that's
573 // easy and somewhat common.
574 if (!TLI.isTypeLegal(SrcVT))
575 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
576 // Unhandled type. Halt "fast" selection and bail.
579 unsigned InputReg = getRegForValue(I->getOperand(0));
581 // Unhandled operand. Halt "fast" selection and bail.
584 bool InputRegIsKill = hasTrivialKill(I->getOperand(0));
586 // If the operand is i1, arrange for the high bits in the register to be zero.
587 if (SrcVT == MVT::i1) {
588 SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
589 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg, InputRegIsKill);
592 InputRegIsKill = true;
594 // If the result is i1, truncate to the target's type for i1 first.
595 if (DstVT == MVT::i1)
596 DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
598 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
601 InputReg, InputRegIsKill);
605 UpdateValueMap(I, ResultReg);
609 bool FastISel::SelectBitCast(const User *I) {
610 // If the bitcast doesn't change the type, just use the operand value.
611 if (I->getType() == I->getOperand(0)->getType()) {
612 unsigned Reg = getRegForValue(I->getOperand(0));
615 UpdateValueMap(I, Reg);
619 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
620 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
621 EVT DstVT = TLI.getValueType(I->getType());
623 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
624 DstVT == MVT::Other || !DstVT.isSimple() ||
625 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
626 // Unhandled type. Halt "fast" selection and bail.
629 unsigned Op0 = getRegForValue(I->getOperand(0));
631 // Unhandled operand. Halt "fast" selection and bail.
634 bool Op0IsKill = hasTrivialKill(I->getOperand(0));
636 // First, try to perform the bitcast by inserting a reg-reg copy.
637 unsigned ResultReg = 0;
638 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
639 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
640 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
641 ResultReg = createResultReg(DstClass);
643 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
645 DstClass, SrcClass, DL);
650 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
652 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
653 ISD::BIT_CONVERT, Op0, Op0IsKill);
658 UpdateValueMap(I, ResultReg);
663 FastISel::SelectInstruction(const Instruction *I) {
664 // Just before the terminator instruction, insert instructions to
665 // feed PHI nodes in successor blocks.
666 if (isa<TerminatorInst>(I))
667 if (!HandlePHINodesInSuccessorBlocks(I->getParent()))
670 DL = I->getDebugLoc();
672 // First, try doing target-independent selection.
673 if (SelectOperator(I, I->getOpcode())) {
678 // Next, try calling the target to attempt to handle the instruction.
679 if (TargetSelectInstruction(I)) {
688 /// FastEmitBranch - Emit an unconditional branch to the given block,
689 /// unless it is the immediate (fall-through) successor, and update
692 FastISel::FastEmitBranch(MachineBasicBlock *MSucc, DebugLoc DL) {
693 if (FuncInfo.MBB->isLayoutSuccessor(MSucc)) {
694 // The unconditional fall-through case, which needs no instructions.
696 // The unconditional branch case.
697 TII.InsertBranch(*FuncInfo.MBB, MSucc, NULL,
698 SmallVector<MachineOperand, 0>(), DL);
700 FuncInfo.MBB->addSuccessor(MSucc);
703 /// SelectFNeg - Emit an FNeg operation.
706 FastISel::SelectFNeg(const User *I) {
707 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
708 if (OpReg == 0) return false;
710 bool OpRegIsKill = hasTrivialKill(I);
712 // If the target has ISD::FNEG, use it.
713 EVT VT = TLI.getValueType(I->getType());
714 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
715 ISD::FNEG, OpReg, OpRegIsKill);
716 if (ResultReg != 0) {
717 UpdateValueMap(I, ResultReg);
721 // Bitcast the value to integer, twiddle the sign bit with xor,
722 // and then bitcast it back to floating-point.
723 if (VT.getSizeInBits() > 64) return false;
724 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
725 if (!TLI.isTypeLegal(IntVT))
728 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
729 ISD::BIT_CONVERT, OpReg, OpRegIsKill);
733 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR,
734 IntReg, /*Kill=*/true,
735 UINT64_C(1) << (VT.getSizeInBits()-1),
736 IntVT.getSimpleVT());
737 if (IntResultReg == 0)
740 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
741 ISD::BIT_CONVERT, IntResultReg, /*Kill=*/true);
745 UpdateValueMap(I, ResultReg);
750 FastISel::SelectLoad(const User *I) {
751 LoadInst *LI = const_cast<LoadInst *>(cast<LoadInst>(I));
753 // For a load from an alloca, make a limited effort to find the value
754 // already available in a register, avoiding redundant loads.
755 if (!LI->isVolatile() && isa<AllocaInst>(LI->getPointerOperand())) {
756 BasicBlock::iterator ScanFrom = LI;
757 if (const Value *V = FindAvailableLoadedValue(LI->getPointerOperand(),
758 LI->getParent(), ScanFrom)) {
759 if (!isa<Instruction>(V) ||
760 cast<Instruction>(V)->getParent() == LI->getParent() ||
761 (isa<AllocaInst>(V) && FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(V)))) {
762 unsigned ResultReg = getRegForValue(V);
763 if (ResultReg != 0) {
764 UpdateValueMap(I, ResultReg);
775 FastISel::SelectOperator(const User *I, unsigned Opcode) {
777 case Instruction::Load:
778 return SelectLoad(I);
779 case Instruction::Add:
780 return SelectBinaryOp(I, ISD::ADD);
781 case Instruction::FAdd:
782 return SelectBinaryOp(I, ISD::FADD);
783 case Instruction::Sub:
784 return SelectBinaryOp(I, ISD::SUB);
785 case Instruction::FSub:
786 // FNeg is currently represented in LLVM IR as a special case of FSub.
787 if (BinaryOperator::isFNeg(I))
788 return SelectFNeg(I);
789 return SelectBinaryOp(I, ISD::FSUB);
790 case Instruction::Mul:
791 return SelectBinaryOp(I, ISD::MUL);
792 case Instruction::FMul:
793 return SelectBinaryOp(I, ISD::FMUL);
794 case Instruction::SDiv:
795 return SelectBinaryOp(I, ISD::SDIV);
796 case Instruction::UDiv:
797 return SelectBinaryOp(I, ISD::UDIV);
798 case Instruction::FDiv:
799 return SelectBinaryOp(I, ISD::FDIV);
800 case Instruction::SRem:
801 return SelectBinaryOp(I, ISD::SREM);
802 case Instruction::URem:
803 return SelectBinaryOp(I, ISD::UREM);
804 case Instruction::FRem:
805 return SelectBinaryOp(I, ISD::FREM);
806 case Instruction::Shl:
807 return SelectBinaryOp(I, ISD::SHL);
808 case Instruction::LShr:
809 return SelectBinaryOp(I, ISD::SRL);
810 case Instruction::AShr:
811 return SelectBinaryOp(I, ISD::SRA);
812 case Instruction::And:
813 return SelectBinaryOp(I, ISD::AND);
814 case Instruction::Or:
815 return SelectBinaryOp(I, ISD::OR);
816 case Instruction::Xor:
817 return SelectBinaryOp(I, ISD::XOR);
819 case Instruction::GetElementPtr:
820 return SelectGetElementPtr(I);
822 case Instruction::Br: {
823 const BranchInst *BI = cast<BranchInst>(I);
825 if (BI->isUnconditional()) {
826 const BasicBlock *LLVMSucc = BI->getSuccessor(0);
827 MachineBasicBlock *MSucc = FuncInfo.MBBMap[LLVMSucc];
828 FastEmitBranch(MSucc, BI->getDebugLoc());
832 // Conditional branches are not handed yet.
833 // Halt "fast" selection and bail.
837 case Instruction::Unreachable:
841 case Instruction::Alloca:
842 // FunctionLowering has the static-sized case covered.
843 if (FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(I)))
846 // Dynamic-sized alloca is not handled yet.
849 case Instruction::Call:
850 return SelectCall(I);
852 case Instruction::BitCast:
853 return SelectBitCast(I);
855 case Instruction::FPToSI:
856 return SelectCast(I, ISD::FP_TO_SINT);
857 case Instruction::ZExt:
858 return SelectCast(I, ISD::ZERO_EXTEND);
859 case Instruction::SExt:
860 return SelectCast(I, ISD::SIGN_EXTEND);
861 case Instruction::Trunc:
862 return SelectCast(I, ISD::TRUNCATE);
863 case Instruction::SIToFP:
864 return SelectCast(I, ISD::SINT_TO_FP);
866 case Instruction::IntToPtr: // Deliberate fall-through.
867 case Instruction::PtrToInt: {
868 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
869 EVT DstVT = TLI.getValueType(I->getType());
870 if (DstVT.bitsGT(SrcVT))
871 return SelectCast(I, ISD::ZERO_EXTEND);
872 if (DstVT.bitsLT(SrcVT))
873 return SelectCast(I, ISD::TRUNCATE);
874 unsigned Reg = getRegForValue(I->getOperand(0));
875 if (Reg == 0) return false;
876 UpdateValueMap(I, Reg);
880 case Instruction::PHI:
881 llvm_unreachable("FastISel shouldn't visit PHI nodes!");
884 // Unhandled instruction. Halt "fast" selection and bail.
889 FastISel::FastISel(FunctionLoweringInfo &funcInfo)
890 : FuncInfo(funcInfo),
891 MRI(FuncInfo.MF->getRegInfo()),
892 MFI(*FuncInfo.MF->getFrameInfo()),
893 MCP(*FuncInfo.MF->getConstantPool()),
894 TM(FuncInfo.MF->getTarget()),
895 TD(*TM.getTargetData()),
896 TII(*TM.getInstrInfo()),
897 TLI(*TM.getTargetLowering()),
898 TRI(*TM.getRegisterInfo()) {
901 FastISel::~FastISel() {}
903 unsigned FastISel::FastEmit_(MVT, MVT,
908 unsigned FastISel::FastEmit_r(MVT, MVT,
910 unsigned /*Op0*/, bool /*Op0IsKill*/) {
914 unsigned FastISel::FastEmit_rr(MVT, MVT,
916 unsigned /*Op0*/, bool /*Op0IsKill*/,
917 unsigned /*Op1*/, bool /*Op1IsKill*/) {
921 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
925 unsigned FastISel::FastEmit_f(MVT, MVT,
926 unsigned, const ConstantFP * /*FPImm*/) {
930 unsigned FastISel::FastEmit_ri(MVT, MVT,
932 unsigned /*Op0*/, bool /*Op0IsKill*/,
937 unsigned FastISel::FastEmit_rf(MVT, MVT,
939 unsigned /*Op0*/, bool /*Op0IsKill*/,
940 const ConstantFP * /*FPImm*/) {
944 unsigned FastISel::FastEmit_rri(MVT, MVT,
946 unsigned /*Op0*/, bool /*Op0IsKill*/,
947 unsigned /*Op1*/, bool /*Op1IsKill*/,
952 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
953 /// to emit an instruction with an immediate operand using FastEmit_ri.
954 /// If that fails, it materializes the immediate into a register and try
955 /// FastEmit_rr instead.
956 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
957 unsigned Op0, bool Op0IsKill,
958 uint64_t Imm, MVT ImmType) {
959 // First check if immediate type is legal. If not, we can't use the ri form.
960 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Op0IsKill, Imm);
963 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
964 if (MaterialReg == 0)
966 return FastEmit_rr(VT, VT, Opcode,
968 MaterialReg, /*Kill=*/true);
971 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
972 /// to emit an instruction with a floating-point immediate operand using
973 /// FastEmit_rf. If that fails, it materializes the immediate into a register
974 /// and try FastEmit_rr instead.
975 unsigned FastISel::FastEmit_rf_(MVT VT, unsigned Opcode,
976 unsigned Op0, bool Op0IsKill,
977 const ConstantFP *FPImm, MVT ImmType) {
978 // First check if immediate type is legal. If not, we can't use the rf form.
979 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, Op0IsKill, FPImm);
983 // Materialize the constant in a register.
984 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
985 if (MaterialReg == 0) {
986 // If the target doesn't have a way to directly enter a floating-point
987 // value into a register, use an alternate approach.
988 // TODO: The current approach only supports floating-point constants
989 // that can be constructed by conversion from integer values. This should
990 // be replaced by code that creates a load from a constant-pool entry,
991 // which will require some target-specific work.
992 const APFloat &Flt = FPImm->getValueAPF();
993 EVT IntVT = TLI.getPointerTy();
996 uint32_t IntBitWidth = IntVT.getSizeInBits();
998 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
999 APFloat::rmTowardZero, &isExact);
1002 APInt IntVal(IntBitWidth, 2, x);
1004 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
1005 ISD::Constant, IntVal.getZExtValue());
1006 if (IntegerReg == 0)
1008 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
1009 ISD::SINT_TO_FP, IntegerReg, /*Kill=*/true);
1010 if (MaterialReg == 0)
1013 return FastEmit_rr(VT, VT, Opcode,
1015 MaterialReg, /*Kill=*/true);
1018 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
1019 return MRI.createVirtualRegister(RC);
1022 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
1023 const TargetRegisterClass* RC) {
1024 unsigned ResultReg = createResultReg(RC);
1025 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1027 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg);
1031 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
1032 const TargetRegisterClass *RC,
1033 unsigned Op0, bool Op0IsKill) {
1034 unsigned ResultReg = createResultReg(RC);
1035 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1037 if (II.getNumDefs() >= 1)
1038 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1039 .addReg(Op0, Op0IsKill * RegState::Kill);
1041 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1042 .addReg(Op0, Op0IsKill * RegState::Kill);
1043 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1044 ResultReg, II.ImplicitDefs[0],
1053 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
1054 const TargetRegisterClass *RC,
1055 unsigned Op0, bool Op0IsKill,
1056 unsigned Op1, bool Op1IsKill) {
1057 unsigned ResultReg = createResultReg(RC);
1058 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1060 if (II.getNumDefs() >= 1)
1061 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1062 .addReg(Op0, Op0IsKill * RegState::Kill)
1063 .addReg(Op1, Op1IsKill * RegState::Kill);
1065 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1066 .addReg(Op0, Op0IsKill * RegState::Kill)
1067 .addReg(Op1, Op1IsKill * RegState::Kill);
1068 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1069 ResultReg, II.ImplicitDefs[0],
1077 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
1078 const TargetRegisterClass *RC,
1079 unsigned Op0, bool Op0IsKill,
1081 unsigned ResultReg = createResultReg(RC);
1082 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1084 if (II.getNumDefs() >= 1)
1085 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1086 .addReg(Op0, Op0IsKill * RegState::Kill)
1089 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1090 .addReg(Op0, Op0IsKill * RegState::Kill)
1092 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1093 ResultReg, II.ImplicitDefs[0],
1101 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
1102 const TargetRegisterClass *RC,
1103 unsigned Op0, bool Op0IsKill,
1104 const ConstantFP *FPImm) {
1105 unsigned ResultReg = createResultReg(RC);
1106 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1108 if (II.getNumDefs() >= 1)
1109 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1110 .addReg(Op0, Op0IsKill * RegState::Kill)
1113 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1114 .addReg(Op0, Op0IsKill * RegState::Kill)
1116 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1117 ResultReg, II.ImplicitDefs[0],
1125 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
1126 const TargetRegisterClass *RC,
1127 unsigned Op0, bool Op0IsKill,
1128 unsigned Op1, bool Op1IsKill,
1130 unsigned ResultReg = createResultReg(RC);
1131 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1133 if (II.getNumDefs() >= 1)
1134 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1135 .addReg(Op0, Op0IsKill * RegState::Kill)
1136 .addReg(Op1, Op1IsKill * RegState::Kill)
1139 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1140 .addReg(Op0, Op0IsKill * RegState::Kill)
1141 .addReg(Op1, Op1IsKill * RegState::Kill)
1143 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1144 ResultReg, II.ImplicitDefs[0],
1152 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1153 const TargetRegisterClass *RC,
1155 unsigned ResultReg = createResultReg(RC);
1156 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1158 if (II.getNumDefs() >= 1)
1159 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg).addImm(Imm);
1161 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II).addImm(Imm);
1162 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1163 ResultReg, II.ImplicitDefs[0],
1171 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
1172 unsigned Op0, bool Op0IsKill,
1174 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1176 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1177 const TargetInstrDesc &II = TII.get(TargetOpcode::EXTRACT_SUBREG);
1179 if (II.getNumDefs() >= 1)
1180 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
1181 .addReg(Op0, Op0IsKill * RegState::Kill)
1184 BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
1185 .addReg(Op0, Op0IsKill * RegState::Kill)
1187 bool InsertedCopy = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
1188 ResultReg, II.ImplicitDefs[0],
1196 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1197 /// with all but the least significant bit set to zero.
1198 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op0, bool Op0IsKill) {
1199 return FastEmit_ri(VT, VT, ISD::AND, Op0, Op0IsKill, 1);
1202 /// HandlePHINodesInSuccessorBlocks - Handle PHI nodes in successor blocks.
1203 /// Emit code to ensure constants are copied into registers when needed.
1204 /// Remember the virtual registers that need to be added to the Machine PHI
1205 /// nodes as input. We cannot just directly add them, because expansion
1206 /// might result in multiple MBB's for one BB. As such, the start of the
1207 /// BB might correspond to a different MBB than the end.
1208 bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
1209 const TerminatorInst *TI = LLVMBB->getTerminator();
1211 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
1212 unsigned OrigNumPHINodesToUpdate = FuncInfo.PHINodesToUpdate.size();
1214 // Check successor nodes' PHI nodes that expect a constant to be available
1216 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
1217 const BasicBlock *SuccBB = TI->getSuccessor(succ);
1218 if (!isa<PHINode>(SuccBB->begin())) continue;
1219 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
1221 // If this terminator has multiple identical successors (common for
1222 // switches), only handle each succ once.
1223 if (!SuccsHandled.insert(SuccMBB)) continue;
1225 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
1227 // At this point we know that there is a 1-1 correspondence between LLVM PHI
1228 // nodes and Machine PHI nodes, but the incoming operands have not been
1230 for (BasicBlock::const_iterator I = SuccBB->begin();
1231 const PHINode *PN = dyn_cast<PHINode>(I); ++I) {
1233 // Ignore dead phi's.
1234 if (PN->use_empty()) continue;
1236 // Only handle legal types. Two interesting things to note here. First,
1237 // by bailing out early, we may leave behind some dead instructions,
1238 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
1239 // own moves. Second, this check is necessary becuase FastISel doesn't
1240 // use CreateRegs to create registers, so it always creates
1241 // exactly one register for each non-void instruction.
1242 EVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
1243 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
1246 VT = TLI.getTypeToTransformTo(LLVMBB->getContext(), VT);
1248 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1253 const Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
1255 // Set the DebugLoc for the copy. Prefer the location of the operand
1256 // if there is one; use the location of the PHI otherwise.
1257 DL = PN->getDebugLoc();
1258 if (const Instruction *Inst = dyn_cast<Instruction>(PHIOp))
1259 DL = Inst->getDebugLoc();
1261 unsigned Reg = getRegForValue(PHIOp);
1263 FuncInfo.PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
1266 FuncInfo.PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));