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/MachineInstrBuilder.h"
48 #include "llvm/CodeGen/MachineModuleInfo.h"
49 #include "llvm/CodeGen/MachineRegisterInfo.h"
50 #include "llvm/Analysis/DebugInfo.h"
51 #include "llvm/Target/TargetData.h"
52 #include "llvm/Target/TargetInstrInfo.h"
53 #include "llvm/Target/TargetLowering.h"
54 #include "llvm/Target/TargetMachine.h"
55 #include "FunctionLoweringInfo.h"
58 unsigned FastISel::getRegForValue(Value *V) {
59 EVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
60 // Don't handle non-simple values in FastISel.
61 if (!RealVT.isSimple())
64 // Ignore illegal types. We must do this before looking up the value
65 // in ValueMap because Arguments are given virtual registers regardless
66 // of whether FastISel can handle them.
67 MVT VT = RealVT.getSimpleVT();
68 if (!TLI.isTypeLegal(VT)) {
69 // Promote MVT::i1 to a legal type though, because it's common and easy.
71 VT = TLI.getTypeToTransformTo(V->getContext(), VT).getSimpleVT();
76 // Look up the value to see if we already have a register for it. We
77 // cache values defined by Instructions across blocks, and other values
78 // only locally. This is because Instructions already have the SSA
79 // def-dominates-use requirement enforced.
80 if (ValueMap.count(V))
82 unsigned Reg = LocalValueMap[V];
86 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
87 if (CI->getValue().getActiveBits() <= 64)
88 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
89 } else if (isa<AllocaInst>(V)) {
90 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
91 } else if (isa<ConstantPointerNull>(V)) {
92 // Translate this as an integer zero so that it can be
93 // local-CSE'd with actual integer zeros.
95 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
96 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
97 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
100 const APFloat &Flt = CF->getValueAPF();
101 EVT IntVT = TLI.getPointerTy();
104 uint32_t IntBitWidth = IntVT.getSizeInBits();
106 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
107 APFloat::rmTowardZero, &isExact);
109 APInt IntVal(IntBitWidth, 2, x);
111 unsigned IntegerReg =
112 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
114 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
117 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
118 if (!SelectOperator(CE, CE->getOpcode())) return 0;
119 Reg = LocalValueMap[CE];
120 } else if (isa<UndefValue>(V)) {
121 Reg = createResultReg(TLI.getRegClassFor(VT));
122 BuildMI(MBB, DL, TII.get(TargetOpcode::IMPLICIT_DEF), Reg);
125 // If target-independent code couldn't handle the value, give target-specific
127 if (!Reg && isa<Constant>(V))
128 Reg = TargetMaterializeConstant(cast<Constant>(V));
130 // Don't cache constant materializations in the general ValueMap.
131 // To do so would require tracking what uses they dominate.
133 LocalValueMap[V] = Reg;
137 unsigned FastISel::lookUpRegForValue(Value *V) {
138 // Look up the value to see if we already have a register for it. We
139 // cache values defined by Instructions across blocks, and other values
140 // only locally. This is because Instructions already have the SSA
141 // def-dominatess-use requirement enforced.
142 if (ValueMap.count(V))
144 return LocalValueMap[V];
147 /// UpdateValueMap - Update the value map to include the new mapping for this
148 /// instruction, or insert an extra copy to get the result in a previous
149 /// determined register.
150 /// NOTE: This is only necessary because we might select a block that uses
151 /// a value before we select the block that defines the value. It might be
152 /// possible to fix this by selecting blocks in reverse postorder.
153 unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) {
154 if (!isa<Instruction>(I)) {
155 LocalValueMap[I] = Reg;
159 unsigned &AssignedReg = ValueMap[I];
160 if (AssignedReg == 0)
162 else if (Reg != AssignedReg) {
163 const TargetRegisterClass *RegClass = MRI.getRegClass(Reg);
164 TII.copyRegToReg(*MBB, MBB->end(), AssignedReg,
165 Reg, RegClass, RegClass);
170 unsigned FastISel::getRegForGEPIndex(Value *Idx) {
171 unsigned IdxN = getRegForValue(Idx);
173 // Unhandled operand. Halt "fast" selection and bail.
176 // If the index is smaller or larger than intptr_t, truncate or extend it.
177 MVT PtrVT = TLI.getPointerTy();
178 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
179 if (IdxVT.bitsLT(PtrVT))
180 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN);
181 else if (IdxVT.bitsGT(PtrVT))
182 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::TRUNCATE, IdxN);
186 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
187 /// which has an opcode which directly corresponds to the given ISD opcode.
189 bool FastISel::SelectBinaryOp(User *I, unsigned ISDOpcode) {
190 EVT VT = EVT::getEVT(I->getType(), /*HandleUnknown=*/true);
191 if (VT == MVT::Other || !VT.isSimple())
192 // Unhandled type. Halt "fast" selection and bail.
195 // We only handle legal types. For example, on x86-32 the instruction
196 // selector contains all of the 64-bit instructions from x86-64,
197 // under the assumption that i64 won't be used if the target doesn't
199 if (!TLI.isTypeLegal(VT)) {
200 // MVT::i1 is special. Allow AND, OR, or XOR because they
201 // don't require additional zeroing, which makes them easy.
203 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
204 ISDOpcode == ISD::XOR))
205 VT = TLI.getTypeToTransformTo(I->getContext(), VT);
210 unsigned Op0 = getRegForValue(I->getOperand(0));
212 // Unhandled operand. Halt "fast" selection and bail.
215 // Check if the second operand is a constant and handle it appropriately.
216 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
217 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
218 ISDOpcode, Op0, CI->getZExtValue());
219 if (ResultReg != 0) {
220 // We successfully emitted code for the given LLVM Instruction.
221 UpdateValueMap(I, ResultReg);
226 // Check if the second operand is a constant float.
227 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
228 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
230 if (ResultReg != 0) {
231 // We successfully emitted code for the given LLVM Instruction.
232 UpdateValueMap(I, ResultReg);
237 unsigned Op1 = getRegForValue(I->getOperand(1));
239 // Unhandled operand. Halt "fast" selection and bail.
242 // Now we have both operands in registers. Emit the instruction.
243 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
244 ISDOpcode, Op0, Op1);
246 // Target-specific code wasn't able to find a machine opcode for
247 // the given ISD opcode and type. Halt "fast" selection and bail.
250 // We successfully emitted code for the given LLVM Instruction.
251 UpdateValueMap(I, ResultReg);
255 bool FastISel::SelectGetElementPtr(User *I) {
256 unsigned N = getRegForValue(I->getOperand(0));
258 // Unhandled operand. Halt "fast" selection and bail.
261 const Type *Ty = I->getOperand(0)->getType();
262 MVT VT = TLI.getPointerTy();
263 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
266 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
267 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
270 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
271 // FIXME: This can be optimized by combining the add with a
273 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
275 // Unhandled operand. Halt "fast" selection and bail.
278 Ty = StTy->getElementType(Field);
280 Ty = cast<SequentialType>(Ty)->getElementType();
282 // If this is a constant subscript, handle it quickly.
283 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
284 if (CI->getZExtValue() == 0) continue;
286 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
287 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
289 // Unhandled operand. Halt "fast" selection and bail.
294 // N = N + Idx * ElementSize;
295 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
296 unsigned IdxN = getRegForGEPIndex(Idx);
298 // Unhandled operand. Halt "fast" selection and bail.
301 if (ElementSize != 1) {
302 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
304 // Unhandled operand. Halt "fast" selection and bail.
307 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
309 // Unhandled operand. Halt "fast" selection and bail.
314 // We successfully emitted code for the given LLVM Instruction.
315 UpdateValueMap(I, N);
319 bool FastISel::SelectCall(User *I) {
320 Function *F = cast<CallInst>(I)->getCalledFunction();
321 if (!F) return false;
323 unsigned IID = F->getIntrinsicID();
326 case Intrinsic::dbg_declare: {
327 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
328 if (!DIDescriptor::ValidDebugInfo(DI->getVariable(), CodeGenOpt::None) ||
329 !MMI->hasDebugInfo())
332 Value *Address = DI->getAddress();
335 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
336 // Don't handle byval struct arguments or VLAs, for example.
338 DenseMap<const AllocaInst*, int>::iterator SI =
339 StaticAllocaMap.find(AI);
340 if (SI == StaticAllocaMap.end()) break; // VLAs.
342 if (!DI->getDebugLoc().isUnknown())
343 MMI->setVariableDbgInfo(DI->getVariable(), FI, DI->getDebugLoc());
345 // Building the map above is target independent. Generating DBG_VALUE
346 // inline is target dependent; do this now.
347 (void)TargetSelectInstruction(cast<Instruction>(I));
350 case Intrinsic::dbg_value: {
351 // This requires target support, but right now X86 is the only Fast target.
352 DbgValueInst *DI = cast<DbgValueInst>(I);
353 const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
354 Value *V = DI->getValue();
356 // Currently the optimizer can produce this; insert an undef to
357 // help debugging. Probably the optimizer should not do this.
358 BuildMI(MBB, DL, II).addReg(0U).addImm(DI->getOffset()).
359 addMetadata(DI->getVariable());
360 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
361 BuildMI(MBB, DL, II).addImm(CI->getZExtValue()).addImm(DI->getOffset()).
362 addMetadata(DI->getVariable());
363 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
364 BuildMI(MBB, DL, II).addFPImm(CF).addImm(DI->getOffset()).
365 addMetadata(DI->getVariable());
366 } else if (unsigned Reg = lookUpRegForValue(V)) {
367 BuildMI(MBB, DL, II).addReg(Reg, RegState::Debug).addImm(DI->getOffset()).
368 addMetadata(DI->getVariable());
370 // We can't yet handle anything else here because it would require
371 // generating code, thus altering codegen because of debug info.
372 // Insert an undef so we can see what we dropped.
373 BuildMI(MBB, DL, II).addReg(0U).addImm(DI->getOffset()).
374 addMetadata(DI->getVariable());
378 case Intrinsic::eh_exception: {
379 EVT VT = TLI.getValueType(I->getType());
380 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
382 case TargetLowering::Expand: {
383 assert(MBB->isLandingPad() && "Call to eh.exception not in landing pad!");
384 unsigned Reg = TLI.getExceptionAddressRegister();
385 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
386 unsigned ResultReg = createResultReg(RC);
387 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
389 assert(InsertedCopy && "Can't copy address registers!");
390 InsertedCopy = InsertedCopy;
391 UpdateValueMap(I, ResultReg);
397 case Intrinsic::eh_selector: {
398 EVT VT = TLI.getValueType(I->getType());
399 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
401 case TargetLowering::Expand: {
403 if (MBB->isLandingPad())
404 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
407 CatchInfoLost.insert(cast<CallInst>(I));
409 // FIXME: Mark exception selector register as live in. Hack for PR1508.
410 unsigned Reg = TLI.getExceptionSelectorRegister();
411 if (Reg) MBB->addLiveIn(Reg);
414 unsigned Reg = TLI.getExceptionSelectorRegister();
415 EVT SrcVT = TLI.getPointerTy();
416 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
417 unsigned ResultReg = createResultReg(RC);
418 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, Reg,
420 assert(InsertedCopy && "Can't copy address registers!");
421 InsertedCopy = InsertedCopy;
423 // Cast the register to the type of the selector.
424 if (SrcVT.bitsGT(MVT::i32))
425 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
427 else if (SrcVT.bitsLT(MVT::i32))
428 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
429 ISD::SIGN_EXTEND, ResultReg);
431 // Unhandled operand. Halt "fast" selection and bail.
434 UpdateValueMap(I, ResultReg);
437 getRegForValue(Constant::getNullValue(I->getType()));
438 UpdateValueMap(I, ResultReg);
449 bool FastISel::SelectCast(User *I, unsigned Opcode) {
450 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
451 EVT DstVT = TLI.getValueType(I->getType());
453 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
454 DstVT == MVT::Other || !DstVT.isSimple())
455 // Unhandled type. Halt "fast" selection and bail.
458 // Check if the destination type is legal. Or as a special case,
459 // it may be i1 if we're doing a truncate because that's
460 // easy and somewhat common.
461 if (!TLI.isTypeLegal(DstVT))
462 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
463 // Unhandled type. Halt "fast" selection and bail.
466 // Check if the source operand is legal. Or as a special case,
467 // it may be i1 if we're doing zero-extension because that's
468 // easy and somewhat common.
469 if (!TLI.isTypeLegal(SrcVT))
470 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
471 // Unhandled type. Halt "fast" selection and bail.
474 unsigned InputReg = getRegForValue(I->getOperand(0));
476 // Unhandled operand. Halt "fast" selection and bail.
479 // If the operand is i1, arrange for the high bits in the register to be zero.
480 if (SrcVT == MVT::i1) {
481 SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
482 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
486 // If the result is i1, truncate to the target's type for i1 first.
487 if (DstVT == MVT::i1)
488 DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
490 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
497 UpdateValueMap(I, ResultReg);
501 bool FastISel::SelectBitCast(User *I) {
502 // If the bitcast doesn't change the type, just use the operand value.
503 if (I->getType() == I->getOperand(0)->getType()) {
504 unsigned Reg = getRegForValue(I->getOperand(0));
507 UpdateValueMap(I, Reg);
511 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
512 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
513 EVT DstVT = TLI.getValueType(I->getType());
515 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
516 DstVT == MVT::Other || !DstVT.isSimple() ||
517 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
518 // Unhandled type. Halt "fast" selection and bail.
521 unsigned Op0 = getRegForValue(I->getOperand(0));
523 // Unhandled operand. Halt "fast" selection and bail.
526 // First, try to perform the bitcast by inserting a reg-reg copy.
527 unsigned ResultReg = 0;
528 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
529 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
530 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
531 ResultReg = createResultReg(DstClass);
533 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
534 Op0, DstClass, SrcClass);
539 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
541 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
542 ISD::BIT_CONVERT, Op0);
547 UpdateValueMap(I, ResultReg);
552 FastISel::SelectInstruction(Instruction *I) {
553 // First, try doing target-independent selection.
554 if (SelectOperator(I, I->getOpcode()))
557 // Next, try calling the target to attempt to handle the instruction.
558 if (TargetSelectInstruction(I))
564 /// FastEmitBranch - Emit an unconditional branch to the given block,
565 /// unless it is the immediate (fall-through) successor, and update
568 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
569 if (MBB->isLayoutSuccessor(MSucc)) {
570 // The unconditional fall-through case, which needs no instructions.
572 // The unconditional branch case.
573 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
575 MBB->addSuccessor(MSucc);
578 /// SelectFNeg - Emit an FNeg operation.
581 FastISel::SelectFNeg(User *I) {
582 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
583 if (OpReg == 0) return false;
585 // If the target has ISD::FNEG, use it.
586 EVT VT = TLI.getValueType(I->getType());
587 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
589 if (ResultReg != 0) {
590 UpdateValueMap(I, ResultReg);
594 // Bitcast the value to integer, twiddle the sign bit with xor,
595 // and then bitcast it back to floating-point.
596 if (VT.getSizeInBits() > 64) return false;
597 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
598 if (!TLI.isTypeLegal(IntVT))
601 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
602 ISD::BIT_CONVERT, OpReg);
606 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR, IntReg,
607 UINT64_C(1) << (VT.getSizeInBits()-1),
608 IntVT.getSimpleVT());
609 if (IntResultReg == 0)
612 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
613 ISD::BIT_CONVERT, IntResultReg);
617 UpdateValueMap(I, ResultReg);
622 FastISel::SelectOperator(User *I, unsigned Opcode) {
624 case Instruction::Add:
625 return SelectBinaryOp(I, ISD::ADD);
626 case Instruction::FAdd:
627 return SelectBinaryOp(I, ISD::FADD);
628 case Instruction::Sub:
629 return SelectBinaryOp(I, ISD::SUB);
630 case Instruction::FSub:
631 // FNeg is currently represented in LLVM IR as a special case of FSub.
632 if (BinaryOperator::isFNeg(I))
633 return SelectFNeg(I);
634 return SelectBinaryOp(I, ISD::FSUB);
635 case Instruction::Mul:
636 return SelectBinaryOp(I, ISD::MUL);
637 case Instruction::FMul:
638 return SelectBinaryOp(I, ISD::FMUL);
639 case Instruction::SDiv:
640 return SelectBinaryOp(I, ISD::SDIV);
641 case Instruction::UDiv:
642 return SelectBinaryOp(I, ISD::UDIV);
643 case Instruction::FDiv:
644 return SelectBinaryOp(I, ISD::FDIV);
645 case Instruction::SRem:
646 return SelectBinaryOp(I, ISD::SREM);
647 case Instruction::URem:
648 return SelectBinaryOp(I, ISD::UREM);
649 case Instruction::FRem:
650 return SelectBinaryOp(I, ISD::FREM);
651 case Instruction::Shl:
652 return SelectBinaryOp(I, ISD::SHL);
653 case Instruction::LShr:
654 return SelectBinaryOp(I, ISD::SRL);
655 case Instruction::AShr:
656 return SelectBinaryOp(I, ISD::SRA);
657 case Instruction::And:
658 return SelectBinaryOp(I, ISD::AND);
659 case Instruction::Or:
660 return SelectBinaryOp(I, ISD::OR);
661 case Instruction::Xor:
662 return SelectBinaryOp(I, ISD::XOR);
664 case Instruction::GetElementPtr:
665 return SelectGetElementPtr(I);
667 case Instruction::Br: {
668 BranchInst *BI = cast<BranchInst>(I);
670 if (BI->isUnconditional()) {
671 BasicBlock *LLVMSucc = BI->getSuccessor(0);
672 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
673 FastEmitBranch(MSucc);
677 // Conditional branches are not handed yet.
678 // Halt "fast" selection and bail.
682 case Instruction::Unreachable:
686 case Instruction::PHI:
687 // PHI nodes are already emitted.
690 case Instruction::Alloca:
691 // FunctionLowering has the static-sized case covered.
692 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
695 // Dynamic-sized alloca is not handled yet.
698 case Instruction::Call:
699 return SelectCall(I);
701 case Instruction::BitCast:
702 return SelectBitCast(I);
704 case Instruction::FPToSI:
705 return SelectCast(I, ISD::FP_TO_SINT);
706 case Instruction::ZExt:
707 return SelectCast(I, ISD::ZERO_EXTEND);
708 case Instruction::SExt:
709 return SelectCast(I, ISD::SIGN_EXTEND);
710 case Instruction::Trunc:
711 return SelectCast(I, ISD::TRUNCATE);
712 case Instruction::SIToFP:
713 return SelectCast(I, ISD::SINT_TO_FP);
715 case Instruction::IntToPtr: // Deliberate fall-through.
716 case Instruction::PtrToInt: {
717 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
718 EVT DstVT = TLI.getValueType(I->getType());
719 if (DstVT.bitsGT(SrcVT))
720 return SelectCast(I, ISD::ZERO_EXTEND);
721 if (DstVT.bitsLT(SrcVT))
722 return SelectCast(I, ISD::TRUNCATE);
723 unsigned Reg = getRegForValue(I->getOperand(0));
724 if (Reg == 0) return false;
725 UpdateValueMap(I, Reg);
730 // Unhandled instruction. Halt "fast" selection and bail.
735 FastISel::FastISel(MachineFunction &mf,
736 MachineModuleInfo *mmi,
737 DenseMap<const Value *, unsigned> &vm,
738 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
739 DenseMap<const AllocaInst *, int> &am
741 , SmallSet<Instruction*, 8> &cil
753 MRI(MF.getRegInfo()),
754 MFI(*MF.getFrameInfo()),
755 MCP(*MF.getConstantPool()),
757 TD(*TM.getTargetData()),
758 TII(*TM.getInstrInfo()),
759 TLI(*TM.getTargetLowering()) {
762 FastISel::~FastISel() {}
764 unsigned FastISel::FastEmit_(MVT, MVT,
769 unsigned FastISel::FastEmit_r(MVT, MVT,
770 unsigned, unsigned /*Op0*/) {
774 unsigned FastISel::FastEmit_rr(MVT, MVT,
775 unsigned, unsigned /*Op0*/,
780 unsigned FastISel::FastEmit_i(MVT, MVT, unsigned, uint64_t /*Imm*/) {
784 unsigned FastISel::FastEmit_f(MVT, MVT,
785 unsigned, ConstantFP * /*FPImm*/) {
789 unsigned FastISel::FastEmit_ri(MVT, MVT,
790 unsigned, unsigned /*Op0*/,
795 unsigned FastISel::FastEmit_rf(MVT, MVT,
796 unsigned, unsigned /*Op0*/,
797 ConstantFP * /*FPImm*/) {
801 unsigned FastISel::FastEmit_rri(MVT, MVT,
803 unsigned /*Op0*/, unsigned /*Op1*/,
808 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
809 /// to emit an instruction with an immediate operand using FastEmit_ri.
810 /// If that fails, it materializes the immediate into a register and try
811 /// FastEmit_rr instead.
812 unsigned FastISel::FastEmit_ri_(MVT VT, unsigned Opcode,
813 unsigned Op0, uint64_t Imm,
815 // First check if immediate type is legal. If not, we can't use the ri form.
816 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
819 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
820 if (MaterialReg == 0)
822 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
825 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
826 /// to emit an instruction with a floating-point immediate operand using
827 /// FastEmit_rf. If that fails, it materializes the immediate into a register
828 /// and try FastEmit_rr instead.
829 unsigned FastISel::FastEmit_rf_(MVT VT, unsigned Opcode,
830 unsigned Op0, ConstantFP *FPImm,
832 // First check if immediate type is legal. If not, we can't use the rf form.
833 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
837 // Materialize the constant in a register.
838 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
839 if (MaterialReg == 0) {
840 // If the target doesn't have a way to directly enter a floating-point
841 // value into a register, use an alternate approach.
842 // TODO: The current approach only supports floating-point constants
843 // that can be constructed by conversion from integer values. This should
844 // be replaced by code that creates a load from a constant-pool entry,
845 // which will require some target-specific work.
846 const APFloat &Flt = FPImm->getValueAPF();
847 EVT IntVT = TLI.getPointerTy();
850 uint32_t IntBitWidth = IntVT.getSizeInBits();
852 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
853 APFloat::rmTowardZero, &isExact);
856 APInt IntVal(IntBitWidth, 2, x);
858 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
859 ISD::Constant, IntVal.getZExtValue());
862 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
863 ISD::SINT_TO_FP, IntegerReg);
864 if (MaterialReg == 0)
867 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
870 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
871 return MRI.createVirtualRegister(RC);
874 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
875 const TargetRegisterClass* RC) {
876 unsigned ResultReg = createResultReg(RC);
877 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
879 BuildMI(MBB, DL, II, ResultReg);
883 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
884 const TargetRegisterClass *RC,
886 unsigned ResultReg = createResultReg(RC);
887 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
889 if (II.getNumDefs() >= 1)
890 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
892 BuildMI(MBB, DL, II).addReg(Op0);
893 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
894 II.ImplicitDefs[0], RC, RC);
902 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
903 const TargetRegisterClass *RC,
904 unsigned Op0, unsigned Op1) {
905 unsigned ResultReg = createResultReg(RC);
906 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
908 if (II.getNumDefs() >= 1)
909 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
911 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
912 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
913 II.ImplicitDefs[0], RC, RC);
920 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
921 const TargetRegisterClass *RC,
922 unsigned Op0, uint64_t Imm) {
923 unsigned ResultReg = createResultReg(RC);
924 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
926 if (II.getNumDefs() >= 1)
927 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
929 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
930 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
931 II.ImplicitDefs[0], RC, RC);
938 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
939 const TargetRegisterClass *RC,
940 unsigned Op0, ConstantFP *FPImm) {
941 unsigned ResultReg = createResultReg(RC);
942 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
944 if (II.getNumDefs() >= 1)
945 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
947 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
948 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
949 II.ImplicitDefs[0], RC, RC);
956 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
957 const TargetRegisterClass *RC,
958 unsigned Op0, unsigned Op1, uint64_t Imm) {
959 unsigned ResultReg = createResultReg(RC);
960 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
962 if (II.getNumDefs() >= 1)
963 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
965 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
966 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
967 II.ImplicitDefs[0], RC, RC);
974 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
975 const TargetRegisterClass *RC,
977 unsigned ResultReg = createResultReg(RC);
978 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
980 if (II.getNumDefs() >= 1)
981 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
983 BuildMI(MBB, DL, II).addImm(Imm);
984 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
985 II.ImplicitDefs[0], RC, RC);
992 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
993 unsigned Op0, uint32_t Idx) {
994 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
996 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
997 const TargetInstrDesc &II = TII.get(TargetOpcode::EXTRACT_SUBREG);
999 if (II.getNumDefs() >= 1)
1000 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1002 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1003 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1004 II.ImplicitDefs[0], RC, RC);
1011 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1012 /// with all but the least significant bit set to zero.
1013 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op) {
1014 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);