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 // Don't cache constant materializations. To do so would require
25 // tracking what uses they dominate. Non-constants, however, already
26 // have the SSA def-doms-use requirement enforced, so we can cache their
28 unsigned FastISel::getRegForValue(Value *V) {
29 if (ValueMap.count(V))
32 MVT::SimpleValueType VT = TLI.getValueType(V->getType()).getSimpleVT();
33 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
34 if (CI->getValue().getActiveBits() > 64)
36 // Don't cache constant materializations. To do so would require
37 // tracking what uses they dominate.
38 return FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
39 } else if (isa<ConstantPointerNull>(V)) {
40 return FastEmit_i(VT, VT, ISD::Constant, 0);
41 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
42 unsigned Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
45 const APFloat &Flt = CF->getValueAPF();
46 MVT IntVT = TLI.getPointerTy();
49 uint32_t IntBitWidth = IntVT.getSizeInBits();
50 if (Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
51 APFloat::rmTowardZero) != APFloat::opOK)
53 APInt IntVal(IntBitWidth, 2, x);
55 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
56 ISD::Constant, IntVal.getZExtValue());
59 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
65 } else if (isa<UndefValue>(V)) {
66 unsigned Reg = createResultReg(TLI.getRegClassFor(VT));
67 BuildMI(MBB, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg);
74 /// UpdateValueMap - Update the value map to include the new mapping for this
75 /// instruction, or insert an extra copy to get the result in a previous
76 /// determined register.
77 /// NOTE: This is only necessary because we might select a block that uses
78 /// a value before we select the block that defines the value. It might be
79 /// possible to fix this by selecting blocks in reverse postorder.
80 void FastISel::UpdateValueMap(Instruction* I, unsigned Reg) {
81 if (!ValueMap.count(I))
84 TII.copyRegToReg(*MBB, MBB->end(), ValueMap[I],
85 Reg, MRI.getRegClass(Reg), MRI.getRegClass(Reg));
88 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
89 /// which has an opcode which directly corresponds to the given ISD opcode.
91 bool FastISel::SelectBinaryOp(Instruction *I, ISD::NodeType ISDOpcode) {
92 MVT VT = MVT::getMVT(I->getType(), /*HandleUnknown=*/true);
93 if (VT == MVT::Other || !VT.isSimple())
94 // Unhandled type. Halt "fast" selection and bail.
96 // We only handle legal types. For example, on x86-32 the instruction
97 // selector contains all of the 64-bit instructions from x86-64,
98 // under the assumption that i64 won't be used if the target doesn't
100 if (!TLI.isTypeLegal(VT))
103 unsigned Op0 = getRegForValue(I->getOperand(0));
105 // Unhandled operand. Halt "fast" selection and bail.
108 // Check if the second operand is a constant and handle it appropriately.
109 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
110 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
111 ISDOpcode, Op0, CI->getZExtValue());
112 if (ResultReg != 0) {
113 // We successfully emitted code for the given LLVM Instruction.
114 UpdateValueMap(I, ResultReg);
119 // Check if the second operand is a constant float.
120 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
121 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
123 if (ResultReg != 0) {
124 // We successfully emitted code for the given LLVM Instruction.
125 UpdateValueMap(I, ResultReg);
130 unsigned Op1 = getRegForValue(I->getOperand(1));
132 // Unhandled operand. Halt "fast" selection and bail.
135 // Now we have both operands in registers. Emit the instruction.
136 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
137 ISDOpcode, Op0, Op1);
139 // Target-specific code wasn't able to find a machine opcode for
140 // the given ISD opcode and type. Halt "fast" selection and bail.
143 // We successfully emitted code for the given LLVM Instruction.
144 UpdateValueMap(I, ResultReg);
148 bool FastISel::SelectGetElementPtr(Instruction *I) {
149 unsigned N = getRegForValue(I->getOperand(0));
151 // Unhandled operand. Halt "fast" selection and bail.
154 const Type *Ty = I->getOperand(0)->getType();
155 MVT::SimpleValueType VT = TLI.getPointerTy().getSimpleVT();
156 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
159 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
160 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
163 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
164 // FIXME: This can be optimized by combining the add with a
166 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
168 // Unhandled operand. Halt "fast" selection and bail.
171 Ty = StTy->getElementType(Field);
173 Ty = cast<SequentialType>(Ty)->getElementType();
175 // If this is a constant subscript, handle it quickly.
176 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
177 if (CI->getZExtValue() == 0) continue;
179 TD.getABITypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
180 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
182 // Unhandled operand. Halt "fast" selection and bail.
187 // N = N + Idx * ElementSize;
188 uint64_t ElementSize = TD.getABITypeSize(Ty);
189 unsigned IdxN = getRegForValue(Idx);
191 // Unhandled operand. Halt "fast" selection and bail.
194 // If the index is smaller or larger than intptr_t, truncate or extend
196 MVT IdxVT = MVT::getMVT(Idx->getType(), /*HandleUnknown=*/false);
197 if (IdxVT.bitsLT(VT))
198 IdxN = FastEmit_r(IdxVT.getSimpleVT(), VT, ISD::SIGN_EXTEND, IdxN);
199 else if (IdxVT.bitsGT(VT))
200 IdxN = FastEmit_r(IdxVT.getSimpleVT(), VT, ISD::TRUNCATE, IdxN);
202 // Unhandled operand. Halt "fast" selection and bail.
205 if (ElementSize != 1) {
206 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
208 // Unhandled operand. Halt "fast" selection and bail.
211 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
213 // Unhandled operand. Halt "fast" selection and bail.
218 // We successfully emitted code for the given LLVM Instruction.
219 UpdateValueMap(I, N);
223 bool FastISel::SelectCast(Instruction *I, ISD::NodeType Opcode) {
224 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
225 MVT DstVT = TLI.getValueType(I->getType());
227 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
228 DstVT == MVT::Other || !DstVT.isSimple() ||
229 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
230 // Unhandled type. Halt "fast" selection and bail.
233 unsigned InputReg = getRegForValue(I->getOperand(0));
235 // Unhandled operand. Halt "fast" selection and bail.
238 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
245 UpdateValueMap(I, ResultReg);
249 bool FastISel::SelectBitCast(Instruction *I) {
250 // If the bitcast doesn't change the type, just use the operand value.
251 if (I->getType() == I->getOperand(0)->getType()) {
252 unsigned Reg = getRegForValue(I->getOperand(0));
255 UpdateValueMap(I, Reg);
259 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
260 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
261 MVT DstVT = TLI.getValueType(I->getType());
263 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
264 DstVT == MVT::Other || !DstVT.isSimple() ||
265 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
266 // Unhandled type. Halt "fast" selection and bail.
269 unsigned Op0 = getRegForValue(I->getOperand(0));
271 // Unhandled operand. Halt "fast" selection and bail.
274 // First, try to perform the bitcast by inserting a reg-reg copy.
275 unsigned ResultReg = 0;
276 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
277 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
278 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
279 ResultReg = createResultReg(DstClass);
281 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
282 Op0, DstClass, SrcClass);
287 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
289 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
290 ISD::BIT_CONVERT, Op0);
295 UpdateValueMap(I, ResultReg);
300 FastISel::SelectInstruction(Instruction *I) {
301 switch (I->getOpcode()) {
302 case Instruction::Add: {
303 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
304 return SelectBinaryOp(I, Opc);
306 case Instruction::Sub: {
307 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
308 return SelectBinaryOp(I, Opc);
310 case Instruction::Mul: {
311 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
312 return SelectBinaryOp(I, Opc);
314 case Instruction::SDiv:
315 return SelectBinaryOp(I, ISD::SDIV);
316 case Instruction::UDiv:
317 return SelectBinaryOp(I, ISD::UDIV);
318 case Instruction::FDiv:
319 return SelectBinaryOp(I, ISD::FDIV);
320 case Instruction::SRem:
321 return SelectBinaryOp(I, ISD::SREM);
322 case Instruction::URem:
323 return SelectBinaryOp(I, ISD::UREM);
324 case Instruction::FRem:
325 return SelectBinaryOp(I, ISD::FREM);
326 case Instruction::Shl:
327 return SelectBinaryOp(I, ISD::SHL);
328 case Instruction::LShr:
329 return SelectBinaryOp(I, ISD::SRL);
330 case Instruction::AShr:
331 return SelectBinaryOp(I, ISD::SRA);
332 case Instruction::And:
333 return SelectBinaryOp(I, ISD::AND);
334 case Instruction::Or:
335 return SelectBinaryOp(I, ISD::OR);
336 case Instruction::Xor:
337 return SelectBinaryOp(I, ISD::XOR);
339 case Instruction::GetElementPtr:
340 return SelectGetElementPtr(I);
342 case Instruction::Br: {
343 BranchInst *BI = cast<BranchInst>(I);
345 if (BI->isUnconditional()) {
346 MachineFunction::iterator NextMBB =
347 next(MachineFunction::iterator(MBB));
348 BasicBlock *LLVMSucc = BI->getSuccessor(0);
349 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
351 if (NextMBB != MF.end() && MSucc == NextMBB) {
352 // The unconditional fall-through case, which needs no instructions.
354 // The unconditional branch case.
355 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
357 MBB->addSuccessor(MSucc);
361 // Conditional branches are not handed yet.
362 // Halt "fast" selection and bail.
366 case Instruction::PHI:
367 // PHI nodes are already emitted.
370 case Instruction::BitCast:
371 return SelectBitCast(I);
373 case Instruction::FPToSI:
374 return SelectCast(I, ISD::FP_TO_SINT);
375 case Instruction::ZExt:
376 return SelectCast(I, ISD::ZERO_EXTEND);
377 case Instruction::SExt:
378 return SelectCast(I, ISD::SIGN_EXTEND);
379 case Instruction::Trunc:
380 return SelectCast(I, ISD::TRUNCATE);
381 case Instruction::SIToFP:
382 return SelectCast(I, ISD::SINT_TO_FP);
384 case Instruction::IntToPtr: // Deliberate fall-through.
385 case Instruction::PtrToInt: {
386 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
387 MVT DstVT = TLI.getValueType(I->getType());
388 if (DstVT.bitsGT(SrcVT))
389 return SelectCast(I, ISD::ZERO_EXTEND);
390 if (DstVT.bitsLT(SrcVT))
391 return SelectCast(I, ISD::TRUNCATE);
392 unsigned Reg = getRegForValue(I->getOperand(0));
393 if (Reg == 0) return false;
394 UpdateValueMap(I, Reg);
399 // Unhandled instruction. Halt "fast" selection and bail.
404 FastISel::FastISel(MachineFunction &mf,
405 DenseMap<const Value *, unsigned> &vm,
406 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm)
411 MRI(MF.getRegInfo()),
413 TD(*TM.getTargetData()),
414 TII(*TM.getInstrInfo()),
415 TLI(*TM.getTargetLowering()) {
418 FastISel::~FastISel() {}
420 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
425 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
426 ISD::NodeType, unsigned /*Op0*/) {
430 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
431 ISD::NodeType, unsigned /*Op0*/,
436 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
437 ISD::NodeType, uint64_t /*Imm*/) {
441 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
442 ISD::NodeType, ConstantFP * /*FPImm*/) {
446 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
447 ISD::NodeType, unsigned /*Op0*/,
452 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
453 ISD::NodeType, unsigned /*Op0*/,
454 ConstantFP * /*FPImm*/) {
458 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
460 unsigned /*Op0*/, unsigned /*Op1*/,
465 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
466 /// to emit an instruction with an immediate operand using FastEmit_ri.
467 /// If that fails, it materializes the immediate into a register and try
468 /// FastEmit_rr instead.
469 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
470 unsigned Op0, uint64_t Imm,
471 MVT::SimpleValueType ImmType) {
472 // First check if immediate type is legal. If not, we can't use the ri form.
473 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
476 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
477 if (MaterialReg == 0)
479 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
482 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
483 /// to emit an instruction with a floating-point immediate operand using
484 /// FastEmit_rf. If that fails, it materializes the immediate into a register
485 /// and try FastEmit_rr instead.
486 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
487 unsigned Op0, ConstantFP *FPImm,
488 MVT::SimpleValueType ImmType) {
489 // First check if immediate type is legal. If not, we can't use the rf form.
490 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
494 // Materialize the constant in a register.
495 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
496 if (MaterialReg == 0) {
497 // If the target doesn't have a way to directly enter a floating-point
498 // value into a register, use an alternate approach.
499 // TODO: The current approach only supports floating-point constants
500 // that can be constructed by conversion from integer values. This should
501 // be replaced by code that creates a load from a constant-pool entry,
502 // which will require some target-specific work.
503 const APFloat &Flt = FPImm->getValueAPF();
504 MVT IntVT = TLI.getPointerTy();
507 uint32_t IntBitWidth = IntVT.getSizeInBits();
508 if (Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
509 APFloat::rmTowardZero) != APFloat::opOK)
511 APInt IntVal(IntBitWidth, 2, x);
513 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
514 ISD::Constant, IntVal.getZExtValue());
517 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
518 ISD::SINT_TO_FP, IntegerReg);
519 if (MaterialReg == 0)
522 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
525 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
526 return MRI.createVirtualRegister(RC);
529 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
530 const TargetRegisterClass* RC) {
531 unsigned ResultReg = createResultReg(RC);
532 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
534 BuildMI(MBB, II, ResultReg);
538 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
539 const TargetRegisterClass *RC,
541 unsigned ResultReg = createResultReg(RC);
542 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
544 BuildMI(MBB, II, ResultReg).addReg(Op0);
548 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
549 const TargetRegisterClass *RC,
550 unsigned Op0, unsigned Op1) {
551 unsigned ResultReg = createResultReg(RC);
552 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
554 BuildMI(MBB, II, ResultReg).addReg(Op0).addReg(Op1);
558 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
559 const TargetRegisterClass *RC,
560 unsigned Op0, uint64_t Imm) {
561 unsigned ResultReg = createResultReg(RC);
562 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
564 BuildMI(MBB, II, ResultReg).addReg(Op0).addImm(Imm);
568 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
569 const TargetRegisterClass *RC,
570 unsigned Op0, ConstantFP *FPImm) {
571 unsigned ResultReg = createResultReg(RC);
572 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
574 BuildMI(MBB, II, ResultReg).addReg(Op0).addFPImm(FPImm);
578 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
579 const TargetRegisterClass *RC,
580 unsigned Op0, unsigned Op1, uint64_t Imm) {
581 unsigned ResultReg = createResultReg(RC);
582 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
584 BuildMI(MBB, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
588 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
589 const TargetRegisterClass *RC,
591 unsigned ResultReg = createResultReg(RC);
592 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
594 BuildMI(MBB, II, ResultReg).addImm(Imm);
598 unsigned FastISel::FastEmitInst_extractsubreg(unsigned Op0, uint32_t Idx) {
599 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
600 const TargetRegisterClass* SRC = *(RC->subregclasses_begin()+Idx-1);
602 unsigned ResultReg = createResultReg(SRC);
603 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
605 BuildMI(MBB, II, ResultReg).addReg(Op0).addImm(Idx);