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/CodeGen/DebugLoc.h"
51 #include "llvm/CodeGen/DwarfWriter.h"
52 #include "llvm/Analysis/DebugInfo.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 "SelectionDAGBuild.h"
60 unsigned FastISel::getRegForValue(Value *V) {
61 MVT RealVT = TLI.getValueType(V->getType(), /*AllowUnknown=*/true);
62 // Don't handle non-simple values in FastISel.
63 if (!RealVT.isSimple())
66 // Ignore illegal types. We must do this before looking up the value
67 // in ValueMap because Arguments are given virtual registers regardless
68 // of whether FastISel can handle them.
69 MVT::SimpleValueType VT = RealVT.getSimpleVT();
70 if (!TLI.isTypeLegal(VT)) {
71 // Promote MVT::i1 to a legal type though, because it's common and easy.
73 VT = TLI.getTypeToTransformTo(VT).getSimpleVT();
78 // Look up the value to see if we already have a register for it. We
79 // cache values defined by Instructions across blocks, and other values
80 // only locally. This is because Instructions already have the SSA
81 // def-dominatess-use requirement enforced.
82 if (ValueMap.count(V))
84 unsigned Reg = LocalValueMap[V];
88 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
89 if (CI->getValue().getActiveBits() <= 64)
90 Reg = FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
91 } else if (isa<AllocaInst>(V)) {
92 Reg = TargetMaterializeAlloca(cast<AllocaInst>(V));
93 } else if (isa<ConstantPointerNull>(V)) {
94 // Translate this as an integer zero so that it can be
95 // local-CSE'd with actual integer zeros.
96 Reg = getRegForValue(Constant::getNullValue(TD.getIntPtrType()));
97 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
98 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
101 const APFloat &Flt = CF->getValueAPF();
102 MVT IntVT = TLI.getPointerTy();
105 uint32_t IntBitWidth = IntVT.getSizeInBits();
107 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
108 APFloat::rmTowardZero, &isExact);
110 APInt IntVal(IntBitWidth, 2, x);
112 unsigned IntegerReg = getRegForValue(ConstantInt::get(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(TargetInstrInfo::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 MVT IdxVT = MVT::getMVT(Idx->getType(), /*HandleUnknown=*/false);
179 if (IdxVT.bitsLT(PtrVT))
180 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT.getSimpleVT(),
181 ISD::SIGN_EXTEND, IdxN);
182 else if (IdxVT.bitsGT(PtrVT))
183 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT.getSimpleVT(),
184 ISD::TRUNCATE, IdxN);
188 /// SelectBinaryOp - Select and emit code for a binary operator instruction,
189 /// which has an opcode which directly corresponds to the given ISD opcode.
191 bool FastISel::SelectBinaryOp(User *I, ISD::NodeType ISDOpcode) {
192 MVT VT = MVT::getMVT(I->getType(), /*HandleUnknown=*/true);
193 if (VT == MVT::Other || !VT.isSimple())
194 // Unhandled type. Halt "fast" selection and bail.
197 // We only handle legal types. For example, on x86-32 the instruction
198 // selector contains all of the 64-bit instructions from x86-64,
199 // under the assumption that i64 won't be used if the target doesn't
201 if (!TLI.isTypeLegal(VT)) {
202 // MVT::i1 is special. Allow AND, OR, or XOR because they
203 // don't require additional zeroing, which makes them easy.
205 (ISDOpcode == ISD::AND || ISDOpcode == ISD::OR ||
206 ISDOpcode == ISD::XOR))
207 VT = TLI.getTypeToTransformTo(VT);
212 unsigned Op0 = getRegForValue(I->getOperand(0));
214 // Unhandled operand. Halt "fast" selection and bail.
217 // Check if the second operand is a constant and handle it appropriately.
218 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
219 unsigned ResultReg = FastEmit_ri(VT.getSimpleVT(), VT.getSimpleVT(),
220 ISDOpcode, Op0, CI->getZExtValue());
221 if (ResultReg != 0) {
222 // We successfully emitted code for the given LLVM Instruction.
223 UpdateValueMap(I, ResultReg);
228 // Check if the second operand is a constant float.
229 if (ConstantFP *CF = dyn_cast<ConstantFP>(I->getOperand(1))) {
230 unsigned ResultReg = FastEmit_rf(VT.getSimpleVT(), VT.getSimpleVT(),
232 if (ResultReg != 0) {
233 // We successfully emitted code for the given LLVM Instruction.
234 UpdateValueMap(I, ResultReg);
239 unsigned Op1 = getRegForValue(I->getOperand(1));
241 // Unhandled operand. Halt "fast" selection and bail.
244 // Now we have both operands in registers. Emit the instruction.
245 unsigned ResultReg = FastEmit_rr(VT.getSimpleVT(), VT.getSimpleVT(),
246 ISDOpcode, Op0, Op1);
248 // Target-specific code wasn't able to find a machine opcode for
249 // the given ISD opcode and type. Halt "fast" selection and bail.
252 // We successfully emitted code for the given LLVM Instruction.
253 UpdateValueMap(I, ResultReg);
257 bool FastISel::SelectGetElementPtr(User *I) {
258 unsigned N = getRegForValue(I->getOperand(0));
260 // Unhandled operand. Halt "fast" selection and bail.
263 const Type *Ty = I->getOperand(0)->getType();
264 MVT::SimpleValueType VT = TLI.getPointerTy().getSimpleVT();
265 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
268 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
269 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
272 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
273 // FIXME: This can be optimized by combining the add with a
275 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
277 // Unhandled operand. Halt "fast" selection and bail.
280 Ty = StTy->getElementType(Field);
282 Ty = cast<SequentialType>(Ty)->getElementType();
284 // If this is a constant subscript, handle it quickly.
285 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
286 if (CI->getZExtValue() == 0) continue;
288 TD.getTypePaddedSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
289 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
291 // Unhandled operand. Halt "fast" selection and bail.
296 // N = N + Idx * ElementSize;
297 uint64_t ElementSize = TD.getTypePaddedSize(Ty);
298 unsigned IdxN = getRegForGEPIndex(Idx);
300 // Unhandled operand. Halt "fast" selection and bail.
303 if (ElementSize != 1) {
304 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
306 // Unhandled operand. Halt "fast" selection and bail.
309 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
311 // Unhandled operand. Halt "fast" selection and bail.
316 // We successfully emitted code for the given LLVM Instruction.
317 UpdateValueMap(I, N);
321 bool FastISel::SelectCall(User *I) {
322 Function *F = cast<CallInst>(I)->getCalledFunction();
323 if (!F) return false;
325 unsigned IID = F->getIntrinsicID();
328 case Intrinsic::dbg_stoppoint: {
329 DbgStopPointInst *SPI = cast<DbgStopPointInst>(I);
330 if (DW && DW->ValidDebugInfo(SPI->getContext(), true)) {
331 DICompileUnit CU(cast<GlobalVariable>(SPI->getContext()));
333 unsigned SrcFile = DW->getOrCreateSourceID(CU.getDirectory(Dir),
335 unsigned Line = SPI->getLine();
336 unsigned Col = SPI->getColumn();
337 unsigned ID = DW->RecordSourceLine(Line, Col, SrcFile);
338 unsigned Idx = MF.getOrCreateDebugLocID(SrcFile, Line, Col);
339 setCurDebugLoc(DebugLoc::get(Idx));
340 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
341 BuildMI(MBB, DL, II).addImm(ID);
345 case Intrinsic::dbg_region_start: {
346 DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I);
347 if (DW && DW->ValidDebugInfo(RSI->getContext(), true)) {
349 DW->RecordRegionStart(cast<GlobalVariable>(RSI->getContext()));
350 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
351 BuildMI(MBB, DL, II).addImm(ID);
355 case Intrinsic::dbg_region_end: {
356 DbgRegionEndInst *REI = cast<DbgRegionEndInst>(I);
357 if (DW && DW->ValidDebugInfo(REI->getContext(), true)) {
359 DISubprogram Subprogram(cast<GlobalVariable>(REI->getContext()));
360 if (!Subprogram.describes(MF.getFunction())) {
361 // This is end of an inlined function.
362 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
363 ID = DW->RecordInlinedFnEnd(Subprogram);
364 BuildMI(MBB, DL, II).addImm(ID);
366 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
367 ID = DW->RecordRegionEnd(cast<GlobalVariable>(REI->getContext()));
368 BuildMI(MBB, DL, II).addImm(ID);
373 case Intrinsic::dbg_func_start: {
374 if (!DW) return true;
375 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
376 Value *SP = FSI->getSubprogram();
378 if (DW->ValidDebugInfo(SP, true)) {
379 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is what
380 // (most?) gdb expects.
381 DebugLoc PrevLoc = DL;
382 DISubprogram Subprogram(cast<GlobalVariable>(SP));
383 DICompileUnit CompileUnit = Subprogram.getCompileUnit();
385 unsigned SrcFile = DW->getOrCreateSourceID(CompileUnit.getDirectory(Dir),
386 CompileUnit.getFilename(FN));
388 // Record the source line.
389 unsigned Line = Subprogram.getLineNumber();
390 unsigned LabelID = DW->RecordSourceLine(Line, 0, SrcFile);
391 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(SrcFile, Line, 0)));
392 if (!Subprogram.describes(MF.getFunction())) {
393 // This is a beginning of an inlined function.
394 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
395 BuildMI(MBB, DL, II).addImm(LabelID);
396 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
397 DW->RecordInlinedFnStart(FSI, Subprogram, LabelID,
402 // llvm.dbg.func_start also defines beginning of function scope.
403 DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
409 case Intrinsic::dbg_declare: {
410 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
411 Value *Variable = DI->getVariable();
412 if (DW && DW->ValidDebugInfo(Variable, true)) {
413 // Determine the address of the declared object.
414 Value *Address = DI->getAddress();
415 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
416 Address = BCI->getOperand(0);
417 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
418 // Don't handle byval struct arguments or VLAs, for example.
420 DenseMap<const AllocaInst*, int>::iterator SI =
421 StaticAllocaMap.find(AI);
422 if (SI == StaticAllocaMap.end()) break; // VLAs.
425 // Determine the debug globalvariable.
426 GlobalValue *GV = cast<GlobalVariable>(Variable);
428 // Build the DECLARE instruction.
429 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
430 MachineInstr *DeclareMI
431 = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
432 DIVariable DV(cast<GlobalVariable>(GV));
434 // This is a local variable
435 DW->RecordVariableScope(DV, DeclareMI);
440 case Intrinsic::eh_exception: {
441 MVT VT = TLI.getValueType(I->getType());
442 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
444 case TargetLowering::Expand: {
445 if (!MBB->isLandingPad()) {
446 // FIXME: Mark exception register as live in. Hack for PR1508.
447 unsigned Reg = TLI.getExceptionAddressRegister();
448 if (Reg) MBB->addLiveIn(Reg);
450 unsigned Reg = TLI.getExceptionAddressRegister();
451 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
452 unsigned ResultReg = createResultReg(RC);
453 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
455 assert(InsertedCopy && "Can't copy address registers!");
456 InsertedCopy = InsertedCopy;
457 UpdateValueMap(I, ResultReg);
463 case Intrinsic::eh_selector_i32:
464 case Intrinsic::eh_selector_i64: {
465 MVT VT = TLI.getValueType(I->getType());
466 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
468 case TargetLowering::Expand: {
469 MVT VT = (IID == Intrinsic::eh_selector_i32 ?
470 MVT::i32 : MVT::i64);
473 if (MBB->isLandingPad())
474 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
477 CatchInfoLost.insert(cast<CallInst>(I));
479 // FIXME: Mark exception selector register as live in. Hack for PR1508.
480 unsigned Reg = TLI.getExceptionSelectorRegister();
481 if (Reg) MBB->addLiveIn(Reg);
484 unsigned Reg = TLI.getExceptionSelectorRegister();
485 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
486 unsigned ResultReg = createResultReg(RC);
487 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
489 assert(InsertedCopy && "Can't copy address registers!");
490 InsertedCopy = InsertedCopy;
491 UpdateValueMap(I, ResultReg);
494 getRegForValue(Constant::getNullValue(I->getType()));
495 UpdateValueMap(I, ResultReg);
506 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
507 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
508 MVT DstVT = TLI.getValueType(I->getType());
510 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
511 DstVT == MVT::Other || !DstVT.isSimple())
512 // Unhandled type. Halt "fast" selection and bail.
515 // Check if the destination type is legal. Or as a special case,
516 // it may be i1 if we're doing a truncate because that's
517 // easy and somewhat common.
518 if (!TLI.isTypeLegal(DstVT))
519 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
520 // Unhandled type. Halt "fast" selection and bail.
523 // Check if the source operand is legal. Or as a special case,
524 // it may be i1 if we're doing zero-extension because that's
525 // easy and somewhat common.
526 if (!TLI.isTypeLegal(SrcVT))
527 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
528 // Unhandled type. Halt "fast" selection and bail.
531 unsigned InputReg = getRegForValue(I->getOperand(0));
533 // Unhandled operand. Halt "fast" selection and bail.
536 // If the operand is i1, arrange for the high bits in the register to be zero.
537 if (SrcVT == MVT::i1) {
538 SrcVT = TLI.getTypeToTransformTo(SrcVT);
539 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
543 // If the result is i1, truncate to the target's type for i1 first.
544 if (DstVT == MVT::i1)
545 DstVT = TLI.getTypeToTransformTo(DstVT);
547 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
554 UpdateValueMap(I, ResultReg);
558 bool FastISel::SelectBitCast(User *I) {
559 // If the bitcast doesn't change the type, just use the operand value.
560 if (I->getType() == I->getOperand(0)->getType()) {
561 unsigned Reg = getRegForValue(I->getOperand(0));
564 UpdateValueMap(I, Reg);
568 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
569 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
570 MVT DstVT = TLI.getValueType(I->getType());
572 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
573 DstVT == MVT::Other || !DstVT.isSimple() ||
574 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
575 // Unhandled type. Halt "fast" selection and bail.
578 unsigned Op0 = getRegForValue(I->getOperand(0));
580 // Unhandled operand. Halt "fast" selection and bail.
583 // First, try to perform the bitcast by inserting a reg-reg copy.
584 unsigned ResultReg = 0;
585 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
586 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
587 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
588 ResultReg = createResultReg(DstClass);
590 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
591 Op0, DstClass, SrcClass);
596 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
598 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
599 ISD::BIT_CONVERT, Op0);
604 UpdateValueMap(I, ResultReg);
609 FastISel::SelectInstruction(Instruction *I) {
610 return SelectOperator(I, I->getOpcode());
613 /// FastEmitBranch - Emit an unconditional branch to the given block,
614 /// unless it is the immediate (fall-through) successor, and update
617 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
618 MachineFunction::iterator NextMBB =
619 next(MachineFunction::iterator(MBB));
621 if (MBB->isLayoutSuccessor(MSucc)) {
622 // The unconditional fall-through case, which needs no instructions.
624 // The unconditional branch case.
625 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
627 MBB->addSuccessor(MSucc);
631 FastISel::SelectOperator(User *I, unsigned Opcode) {
633 case Instruction::Add: {
634 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
635 return SelectBinaryOp(I, Opc);
637 case Instruction::Sub: {
638 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
639 return SelectBinaryOp(I, Opc);
641 case Instruction::Mul: {
642 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
643 return SelectBinaryOp(I, Opc);
645 case Instruction::SDiv:
646 return SelectBinaryOp(I, ISD::SDIV);
647 case Instruction::UDiv:
648 return SelectBinaryOp(I, ISD::UDIV);
649 case Instruction::FDiv:
650 return SelectBinaryOp(I, ISD::FDIV);
651 case Instruction::SRem:
652 return SelectBinaryOp(I, ISD::SREM);
653 case Instruction::URem:
654 return SelectBinaryOp(I, ISD::UREM);
655 case Instruction::FRem:
656 return SelectBinaryOp(I, ISD::FREM);
657 case Instruction::Shl:
658 return SelectBinaryOp(I, ISD::SHL);
659 case Instruction::LShr:
660 return SelectBinaryOp(I, ISD::SRL);
661 case Instruction::AShr:
662 return SelectBinaryOp(I, ISD::SRA);
663 case Instruction::And:
664 return SelectBinaryOp(I, ISD::AND);
665 case Instruction::Or:
666 return SelectBinaryOp(I, ISD::OR);
667 case Instruction::Xor:
668 return SelectBinaryOp(I, ISD::XOR);
670 case Instruction::GetElementPtr:
671 return SelectGetElementPtr(I);
673 case Instruction::Br: {
674 BranchInst *BI = cast<BranchInst>(I);
676 if (BI->isUnconditional()) {
677 BasicBlock *LLVMSucc = BI->getSuccessor(0);
678 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
679 FastEmitBranch(MSucc);
683 // Conditional branches are not handed yet.
684 // Halt "fast" selection and bail.
688 case Instruction::Unreachable:
692 case Instruction::PHI:
693 // PHI nodes are already emitted.
696 case Instruction::Alloca:
697 // FunctionLowering has the static-sized case covered.
698 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
701 // Dynamic-sized alloca is not handled yet.
704 case Instruction::Call:
705 return SelectCall(I);
707 case Instruction::BitCast:
708 return SelectBitCast(I);
710 case Instruction::FPToSI:
711 return SelectCast(I, ISD::FP_TO_SINT);
712 case Instruction::ZExt:
713 return SelectCast(I, ISD::ZERO_EXTEND);
714 case Instruction::SExt:
715 return SelectCast(I, ISD::SIGN_EXTEND);
716 case Instruction::Trunc:
717 return SelectCast(I, ISD::TRUNCATE);
718 case Instruction::SIToFP:
719 return SelectCast(I, ISD::SINT_TO_FP);
721 case Instruction::IntToPtr: // Deliberate fall-through.
722 case Instruction::PtrToInt: {
723 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
724 MVT DstVT = TLI.getValueType(I->getType());
725 if (DstVT.bitsGT(SrcVT))
726 return SelectCast(I, ISD::ZERO_EXTEND);
727 if (DstVT.bitsLT(SrcVT))
728 return SelectCast(I, ISD::TRUNCATE);
729 unsigned Reg = getRegForValue(I->getOperand(0));
730 if (Reg == 0) return false;
731 UpdateValueMap(I, Reg);
736 // Unhandled instruction. Halt "fast" selection and bail.
741 FastISel::FastISel(MachineFunction &mf,
742 MachineModuleInfo *mmi,
744 DenseMap<const Value *, unsigned> &vm,
745 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
746 DenseMap<const AllocaInst *, int> &am
748 , SmallSet<Instruction*, 8> &cil
761 MRI(MF.getRegInfo()),
762 MFI(*MF.getFrameInfo()),
763 MCP(*MF.getConstantPool()),
765 TD(*TM.getTargetData()),
766 TII(*TM.getInstrInfo()),
767 TLI(*TM.getTargetLowering()) {
770 FastISel::~FastISel() {}
772 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
777 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
778 ISD::NodeType, unsigned /*Op0*/) {
782 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
783 ISD::NodeType, unsigned /*Op0*/,
788 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
789 ISD::NodeType, uint64_t /*Imm*/) {
793 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
794 ISD::NodeType, ConstantFP * /*FPImm*/) {
798 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
799 ISD::NodeType, unsigned /*Op0*/,
804 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
805 ISD::NodeType, unsigned /*Op0*/,
806 ConstantFP * /*FPImm*/) {
810 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
812 unsigned /*Op0*/, unsigned /*Op1*/,
817 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
818 /// to emit an instruction with an immediate operand using FastEmit_ri.
819 /// If that fails, it materializes the immediate into a register and try
820 /// FastEmit_rr instead.
821 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
822 unsigned Op0, uint64_t Imm,
823 MVT::SimpleValueType ImmType) {
824 // First check if immediate type is legal. If not, we can't use the ri form.
825 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
828 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
829 if (MaterialReg == 0)
831 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
834 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
835 /// to emit an instruction with a floating-point immediate operand using
836 /// FastEmit_rf. If that fails, it materializes the immediate into a register
837 /// and try FastEmit_rr instead.
838 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
839 unsigned Op0, ConstantFP *FPImm,
840 MVT::SimpleValueType ImmType) {
841 // First check if immediate type is legal. If not, we can't use the rf form.
842 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
846 // Materialize the constant in a register.
847 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
848 if (MaterialReg == 0) {
849 // If the target doesn't have a way to directly enter a floating-point
850 // value into a register, use an alternate approach.
851 // TODO: The current approach only supports floating-point constants
852 // that can be constructed by conversion from integer values. This should
853 // be replaced by code that creates a load from a constant-pool entry,
854 // which will require some target-specific work.
855 const APFloat &Flt = FPImm->getValueAPF();
856 MVT IntVT = TLI.getPointerTy();
859 uint32_t IntBitWidth = IntVT.getSizeInBits();
861 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
862 APFloat::rmTowardZero, &isExact);
865 APInt IntVal(IntBitWidth, 2, x);
867 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
868 ISD::Constant, IntVal.getZExtValue());
871 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
872 ISD::SINT_TO_FP, IntegerReg);
873 if (MaterialReg == 0)
876 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
879 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
880 return MRI.createVirtualRegister(RC);
883 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
884 const TargetRegisterClass* RC) {
885 unsigned ResultReg = createResultReg(RC);
886 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
888 BuildMI(MBB, DL, II, ResultReg);
892 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
893 const TargetRegisterClass *RC,
895 unsigned ResultReg = createResultReg(RC);
896 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
898 if (II.getNumDefs() >= 1)
899 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
901 BuildMI(MBB, DL, II).addReg(Op0);
902 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
903 II.ImplicitDefs[0], RC, RC);
911 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
912 const TargetRegisterClass *RC,
913 unsigned Op0, unsigned Op1) {
914 unsigned ResultReg = createResultReg(RC);
915 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
917 if (II.getNumDefs() >= 1)
918 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
920 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
921 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
922 II.ImplicitDefs[0], RC, RC);
929 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
930 const TargetRegisterClass *RC,
931 unsigned Op0, uint64_t Imm) {
932 unsigned ResultReg = createResultReg(RC);
933 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
935 if (II.getNumDefs() >= 1)
936 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
938 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
939 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
940 II.ImplicitDefs[0], RC, RC);
947 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
948 const TargetRegisterClass *RC,
949 unsigned Op0, ConstantFP *FPImm) {
950 unsigned ResultReg = createResultReg(RC);
951 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
953 if (II.getNumDefs() >= 1)
954 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
956 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
957 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
958 II.ImplicitDefs[0], RC, RC);
965 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
966 const TargetRegisterClass *RC,
967 unsigned Op0, unsigned Op1, uint64_t Imm) {
968 unsigned ResultReg = createResultReg(RC);
969 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
971 if (II.getNumDefs() >= 1)
972 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
974 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
975 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
976 II.ImplicitDefs[0], RC, RC);
983 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
984 const TargetRegisterClass *RC,
986 unsigned ResultReg = createResultReg(RC);
987 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
989 if (II.getNumDefs() >= 1)
990 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
992 BuildMI(MBB, DL, II).addImm(Imm);
993 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
994 II.ImplicitDefs[0], RC, RC);
1001 unsigned FastISel::FastEmitInst_extractsubreg(MVT::SimpleValueType RetVT,
1002 unsigned Op0, uint32_t Idx) {
1003 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1005 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1006 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
1008 if (II.getNumDefs() >= 1)
1009 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1011 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1012 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1013 II.ImplicitDefs[0], RC, RC);
1020 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1021 /// with all but the least significant bit set to zero.
1022 unsigned FastISel::FastEmitZExtFromI1(MVT::SimpleValueType VT, unsigned Op) {
1023 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);