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 //===----------------------------------------------------------------------===//
14 #include "llvm/Instructions.h"
15 #include "llvm/CodeGen/FastISel.h"
16 #include "llvm/CodeGen/MachineInstrBuilder.h"
17 #include "llvm/CodeGen/MachineRegisterInfo.h"
18 #include "llvm/Target/TargetData.h"
19 #include "llvm/Target/TargetInstrInfo.h"
20 #include "llvm/Target/TargetLowering.h"
21 #include "llvm/Target/TargetMachine.h"
24 unsigned FastISel::getRegForValue(Value *V) {
25 // Look up the value to see if we already have a register for it. We
26 // cache values defined by Instructions across blocks, and other values
27 // only locally. This is because Instructions already have the SSA
28 // def-dominatess-use requirement enforced.
29 if (ValueMap.count(V))
31 unsigned Reg = LocalValueMap[V];
35 MVT::SimpleValueType VT = TLI.getValueType(V->getType()).getSimpleVT();
36 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
37 if (CI->getValue().getActiveBits() > 64)
39 // Don't cache constant materializations. To do so would require
40 // tracking what uses they dominate.
41 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
42 } else if (isa<ConstantPointerNull>(V)) {
43 Reg = FastEmit_i(VT, VT, ISD::Constant, 0);
44 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
45 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
48 const APFloat &Flt = CF->getValueAPF();
49 MVT IntVT = TLI.getPointerTy();
52 uint32_t IntBitWidth = IntVT.getSizeInBits();
53 if (Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
54 APFloat::rmTowardZero) != APFloat::opOK)
56 APInt IntVal(IntBitWidth, 2, x);
58 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
59 ISD::Constant, IntVal.getZExtValue());
62 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
66 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
67 if (!SelectOperator(CE, CE->getOpcode())) return 0;
68 Reg = LocalValueMap[CE];
69 } else if (isa<UndefValue>(V)) {
70 Reg = createResultReg(TLI.getRegClassFor(VT));
71 BuildMI(MBB, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg);
76 LocalValueMap[V] = Reg;
80 /// UpdateValueMap - Update the value map to include the new mapping for this
81 /// instruction, or insert an extra copy to get the result in a previous
82 /// determined register.
83 /// NOTE: This is only necessary because we might select a block that uses
84 /// a value before we select the block that defines the value. It might be
85 /// possible to fix this by selecting blocks in reverse postorder.
86 void FastISel::UpdateValueMap(Value* I, unsigned Reg) {
87 if (!isa<Instruction>(I)) {
88 LocalValueMap[I] = Reg;
91 if (!ValueMap.count(I))
94 TII.copyRegToReg(*MBB, MBB->end(), ValueMap[I],
95 Reg, MRI.getRegClass(Reg), MRI.getRegClass(Reg));
98 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
99 /// which has an opcode which directly corresponds to the given ISD opcode.
101 bool FastISel::SelectBinaryOp(User *I, ISD::NodeType ISDOpcode) {
102 MVT VT = MVT::getMVT(I->getType(), /*HandleUnknown=*/true);
103 if (VT == MVT::Other || !VT.isSimple())
104 // Unhandled type. Halt "fast" selection and bail.
107 // We only handle legal types. For example, on x86-32 the instruction
108 // selector contains all of the 64-bit instructions from x86-64,
109 // under the assumption that i64 won't be used if the target doesn't
111 if (!TLI.isTypeLegal(VT)) {
112 // MVT::i1 is special. Allow AND and OR (but not XOR) because they
113 // don't require additional zeroing, which makes them easy.
115 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR))
116 VT = TLI.getTypeToTransformTo(VT);
121 unsigned Op0 = getRegForValue(I->getOperand(0));
123 // Unhandled operand. Halt "fast" selection and bail.
126 // Check if the second operand is a constant and handle it appropriately.
127 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
128 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
129 ISDOpcode, Op0, CI->getZExtValue());
130 if (ResultReg != 0) {
131 // We successfully emitted code for the given LLVM Instruction.
132 UpdateValueMap(I, ResultReg);
137 // Check if the second operand is a constant float.
138 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
139 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
141 if (ResultReg != 0) {
142 // We successfully emitted code for the given LLVM Instruction.
143 UpdateValueMap(I, ResultReg);
148 unsigned Op1 = getRegForValue(I->getOperand(1));
150 // Unhandled operand. Halt "fast" selection and bail.
153 // Now we have both operands in registers. Emit the instruction.
154 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
155 ISDOpcode, Op0, Op1);
157 // Target-specific code wasn't able to find a machine opcode for
158 // the given ISD opcode and type. Halt "fast" selection and bail.
161 // We successfully emitted code for the given LLVM Instruction.
162 UpdateValueMap(I, ResultReg);
166 bool FastISel::SelectGetElementPtr(User *I) {
167 unsigned N = getRegForValue(I->getOperand(0));
169 // Unhandled operand. Halt "fast" selection and bail.
172 const Type *Ty = I->getOperand(0)->getType();
173 MVT::SimpleValueType VT = TLI.getPointerTy().getSimpleVT();
174 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
177 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
178 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
181 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
182 // FIXME: This can be optimized by combining the add with a
184 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
186 // Unhandled operand. Halt "fast" selection and bail.
189 Ty = StTy->getElementType(Field);
191 Ty = cast<SequentialType>(Ty)->getElementType();
193 // If this is a constant subscript, handle it quickly.
194 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
195 if (CI->getZExtValue() == 0) continue;
197 TD.getABITypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
198 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
200 // Unhandled operand. Halt "fast" selection and bail.
205 // N = N + Idx * ElementSize;
206 uint64_t ElementSize = TD.getABITypeSize(Ty);
207 unsigned IdxN = getRegForValue(Idx);
209 // Unhandled operand. Halt "fast" selection and bail.
212 // If the index is smaller or larger than intptr_t, truncate or extend
214 MVT IdxVT = MVT::getMVT(Idx->getType(), /*HandleUnknown=*/false);
215 if (IdxVT.bitsLT(VT))
216 IdxN = FastEmit_r(IdxVT.getSimpleVT(), VT, ISD::SIGN_EXTEND, IdxN);
217 else if (IdxVT.bitsGT(VT))
218 IdxN = FastEmit_r(IdxVT.getSimpleVT(), VT, ISD::TRUNCATE, IdxN);
220 // Unhandled operand. Halt "fast" selection and bail.
223 if (ElementSize != 1) {
224 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
226 // Unhandled operand. Halt "fast" selection and bail.
229 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
231 // Unhandled operand. Halt "fast" selection and bail.
236 // We successfully emitted code for the given LLVM Instruction.
237 UpdateValueMap(I, N);
241 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
242 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
243 MVT DstVT = TLI.getValueType(I->getType());
245 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
246 DstVT == MVT::Other || !DstVT.isSimple() ||
247 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
248 // Unhandled type. Halt "fast" selection and bail.
251 unsigned InputReg = getRegForValue(I->getOperand(0));
253 // Unhandled operand. Halt "fast" selection and bail.
256 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
263 UpdateValueMap(I, ResultReg);
267 bool FastISel::SelectBitCast(User *I) {
268 // If the bitcast doesn't change the type, just use the operand value.
269 if (I->getType() == I->getOperand(0)->getType()) {
270 unsigned Reg = getRegForValue(I->getOperand(0));
273 UpdateValueMap(I, Reg);
277 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
278 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
279 MVT DstVT = TLI.getValueType(I->getType());
281 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
282 DstVT == MVT::Other || !DstVT.isSimple() ||
283 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
284 // Unhandled type. Halt "fast" selection and bail.
287 unsigned Op0 = getRegForValue(I->getOperand(0));
289 // Unhandled operand. Halt "fast" selection and bail.
292 // First, try to perform the bitcast by inserting a reg-reg copy.
293 unsigned ResultReg = 0;
294 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
295 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
296 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
297 ResultReg = createResultReg(DstClass);
299 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
300 Op0, DstClass, SrcClass);
305 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
307 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
308 ISD::BIT_CONVERT, Op0);
313 UpdateValueMap(I, ResultReg);
318 FastISel::SelectInstruction(Instruction *I) {
319 return SelectOperator(I, I->getOpcode());
323 FastISel::SelectOperator(User *I, unsigned Opcode) {
325 case Instruction::Add: {
326 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
327 return SelectBinaryOp(I, Opc);
329 case Instruction::Sub: {
330 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
331 return SelectBinaryOp(I, Opc);
333 case Instruction::Mul: {
334 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
335 return SelectBinaryOp(I, Opc);
337 case Instruction::SDiv:
338 return SelectBinaryOp(I, ISD::SDIV);
339 case Instruction::UDiv:
340 return SelectBinaryOp(I, ISD::UDIV);
341 case Instruction::FDiv:
342 return SelectBinaryOp(I, ISD::FDIV);
343 case Instruction::SRem:
344 return SelectBinaryOp(I, ISD::SREM);
345 case Instruction::URem:
346 return SelectBinaryOp(I, ISD::UREM);
347 case Instruction::FRem:
348 return SelectBinaryOp(I, ISD::FREM);
349 case Instruction::Shl:
350 return SelectBinaryOp(I, ISD::SHL);
351 case Instruction::LShr:
352 return SelectBinaryOp(I, ISD::SRL);
353 case Instruction::AShr:
354 return SelectBinaryOp(I, ISD::SRA);
355 case Instruction::And:
356 return SelectBinaryOp(I, ISD::AND);
357 case Instruction::Or:
358 return SelectBinaryOp(I, ISD::OR);
359 case Instruction::Xor:
360 return SelectBinaryOp(I, ISD::XOR);
362 case Instruction::GetElementPtr:
363 return SelectGetElementPtr(I);
365 case Instruction::Br: {
366 BranchInst *BI = cast<BranchInst>(I);
368 if (BI->isUnconditional()) {
369 MachineFunction::iterator NextMBB =
370 next(MachineFunction::iterator(MBB));
371 BasicBlock *LLVMSucc = BI->getSuccessor(0);
372 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
374 if (NextMBB != MF.end() && MSucc == NextMBB) {
375 // The unconditional fall-through case, which needs no instructions.
377 // The unconditional branch case.
378 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
380 MBB->addSuccessor(MSucc);
384 // Conditional branches are not handed yet.
385 // Halt "fast" selection and bail.
389 case Instruction::Unreachable:
393 case Instruction::PHI:
394 // PHI nodes are already emitted.
397 case Instruction::BitCast:
398 return SelectBitCast(I);
400 case Instruction::FPToSI:
401 return SelectCast(I, ISD::FP_TO_SINT);
402 case Instruction::ZExt:
403 return SelectCast(I, ISD::ZERO_EXTEND);
404 case Instruction::SExt:
405 return SelectCast(I, ISD::SIGN_EXTEND);
406 case Instruction::Trunc:
407 return SelectCast(I, ISD::TRUNCATE);
408 case Instruction::SIToFP:
409 return SelectCast(I, ISD::SINT_TO_FP);
411 case Instruction::IntToPtr: // Deliberate fall-through.
412 case Instruction::PtrToInt: {
413 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
414 MVT DstVT = TLI.getValueType(I->getType());
415 if (DstVT.bitsGT(SrcVT))
416 return SelectCast(I, ISD::ZERO_EXTEND);
417 if (DstVT.bitsLT(SrcVT))
418 return SelectCast(I, ISD::TRUNCATE);
419 unsigned Reg = getRegForValue(I->getOperand(0));
420 if (Reg == 0) return false;
421 UpdateValueMap(I, Reg);
426 // Unhandled instruction. Halt "fast" selection and bail.
431 FastISel::FastISel(MachineFunction &mf,
432 DenseMap<const Value *, unsigned> &vm,
433 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm)
438 MRI(MF.getRegInfo()),
440 TD(*TM.getTargetData()),
441 TII(*TM.getInstrInfo()),
442 TLI(*TM.getTargetLowering()) {
445 FastISel::~FastISel() {}
447 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
452 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
453 ISD::NodeType, unsigned /*Op0*/) {
457 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
458 ISD::NodeType, unsigned /*Op0*/,
463 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
464 ISD::NodeType, uint64_t /*Imm*/) {
468 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
469 ISD::NodeType, ConstantFP * /*FPImm*/) {
473 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
474 ISD::NodeType, unsigned /*Op0*/,
479 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
480 ISD::NodeType, unsigned /*Op0*/,
481 ConstantFP * /*FPImm*/) {
485 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
487 unsigned /*Op0*/, unsigned /*Op1*/,
492 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
493 /// to emit an instruction with an immediate operand using FastEmit_ri.
494 /// If that fails, it materializes the immediate into a register and try
495 /// FastEmit_rr instead.
496 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
497 unsigned Op0, uint64_t Imm,
498 MVT::SimpleValueType ImmType) {
499 // First check if immediate type is legal. If not, we can't use the ri form.
500 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
503 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
504 if (MaterialReg == 0)
506 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
509 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
510 /// to emit an instruction with a floating-point immediate operand using
511 /// FastEmit_rf. If that fails, it materializes the immediate into a register
512 /// and try FastEmit_rr instead.
513 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
514 unsigned Op0, ConstantFP *FPImm,
515 MVT::SimpleValueType ImmType) {
516 // First check if immediate type is legal. If not, we can't use the rf form.
517 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
521 // Materialize the constant in a register.
522 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
523 if (MaterialReg == 0) {
524 // If the target doesn't have a way to directly enter a floating-point
525 // value into a register, use an alternate approach.
526 // TODO: The current approach only supports floating-point constants
527 // that can be constructed by conversion from integer values. This should
528 // be replaced by code that creates a load from a constant-pool entry,
529 // which will require some target-specific work.
530 const APFloat &Flt = FPImm->getValueAPF();
531 MVT IntVT = TLI.getPointerTy();
534 uint32_t IntBitWidth = IntVT.getSizeInBits();
535 if (Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
536 APFloat::rmTowardZero) != APFloat::opOK)
538 APInt IntVal(IntBitWidth, 2, x);
540 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
541 ISD::Constant, IntVal.getZExtValue());
544 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
545 ISD::SINT_TO_FP, IntegerReg);
546 if (MaterialReg == 0)
549 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
552 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
553 return MRI.createVirtualRegister(RC);
556 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
557 const TargetRegisterClass* RC) {
558 unsigned ResultReg = createResultReg(RC);
559 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
561 BuildMI(MBB, II, ResultReg);
565 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
566 const TargetRegisterClass *RC,
568 unsigned ResultReg = createResultReg(RC);
569 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
571 BuildMI(MBB, II, ResultReg).addReg(Op0);
575 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
576 const TargetRegisterClass *RC,
577 unsigned Op0, unsigned Op1) {
578 unsigned ResultReg = createResultReg(RC);
579 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
581 BuildMI(MBB, II, ResultReg).addReg(Op0).addReg(Op1);
585 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
586 const TargetRegisterClass *RC,
587 unsigned Op0, uint64_t Imm) {
588 unsigned ResultReg = createResultReg(RC);
589 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
591 BuildMI(MBB, II, ResultReg).addReg(Op0).addImm(Imm);
595 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
596 const TargetRegisterClass *RC,
597 unsigned Op0, ConstantFP *FPImm) {
598 unsigned ResultReg = createResultReg(RC);
599 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
601 BuildMI(MBB, II, ResultReg).addReg(Op0).addFPImm(FPImm);
605 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
606 const TargetRegisterClass *RC,
607 unsigned Op0, unsigned Op1, uint64_t Imm) {
608 unsigned ResultReg = createResultReg(RC);
609 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
611 BuildMI(MBB, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
615 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
616 const TargetRegisterClass *RC,
618 unsigned ResultReg = createResultReg(RC);
619 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
621 BuildMI(MBB, II, ResultReg).addImm(Imm);
625 unsigned FastISel::FastEmitInst_extractsubreg(unsigned Op0, uint32_t Idx) {
626 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
627 const TargetRegisterClass* SRC = *(RC->subregclasses_begin()+Idx-1);
629 unsigned ResultReg = createResultReg(SRC);
630 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
632 BuildMI(MBB, II, ResultReg).addReg(Op0).addImm(Idx);