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.getTypeAllocSize(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.getTypeAllocSize(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 (DIDescriptor::ValidDebugInfo(SPI->getContext(), CodeGenOpt::None)) {
331 DICompileUnit CU(cast<GlobalVariable>(SPI->getContext()));
332 unsigned Line = SPI->getLine();
333 unsigned Col = SPI->getColumn();
334 unsigned Idx = MF.getOrCreateDebugLocID(CU.getGV(),
335 DbgScopeTrack.getCurScope(),
337 setCurDebugLoc(DebugLoc::get(Idx));
341 case Intrinsic::dbg_region_start: {
342 DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I);
343 if (!DIDescriptor::ValidDebugInfo(RSI->getContext(), CodeGenOpt::None))
346 GlobalVariable *Rgn = cast<GlobalVariable>(RSI->getContext());
347 DbgScopeTrack.EnterDebugScope(Rgn, MF);
348 if (DW && DW->ShouldEmitDwarfDebug()) {
349 unsigned ID = DW->RecordRegionStart(Rgn);
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 (!DIDescriptor::ValidDebugInfo(REI->getContext(), CodeGenOpt::None))
360 GlobalVariable *Rgn = cast<GlobalVariable>(REI->getContext());
361 DbgScopeTrack.ExitDebugScope(Rgn, MF);
362 if (DW && DW->ShouldEmitDwarfDebug()) {
364 DISubprogram Subprogram(Rgn);
365 if (!Subprogram.isNull() && !Subprogram.describes(MF.getFunction())) {
366 // This is end of an inlined function.
367 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
368 ID = DW->RecordInlinedFnEnd(Subprogram);
370 // Returned ID is 0 if this is unbalanced "end of inlined
371 // scope". This could happen if optimizer eats dbg intrinsics
372 // or "beginning of inlined scope" is not recoginized due to
373 // missing location info. In such cases, do ignore this region.end.
374 BuildMI(MBB, DL, II).addImm(ID);
376 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
377 ID = DW->RecordRegionEnd(cast<GlobalVariable>(REI->getContext()));
378 BuildMI(MBB, DL, II).addImm(ID);
383 case Intrinsic::dbg_func_start: {
384 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
385 Value *SP = FSI->getSubprogram();
386 if (!DIDescriptor::ValidDebugInfo(SP, CodeGenOpt::None))
389 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is what
390 // (most?) gdb expects.
391 DebugLoc PrevLoc = DL;
392 DISubprogram Subprogram(cast<GlobalVariable>(SP));
393 DICompileUnit CompileUnit = Subprogram.getCompileUnit();
394 DbgScopeTrack.EnterDebugScope(Subprogram.getGV(), MF);
396 if (!Subprogram.describes(MF.getFunction())) {
397 // This is a beginning of an inlined function.
399 // If llvm.dbg.func.start is seen in a new block before any
400 // llvm.dbg.stoppoint intrinsic then the location info is unknown.
401 // FIXME : Why DebugLoc is reset at the beginning of each block ?
402 if (PrevLoc.isUnknown())
404 // Record the source line.
405 unsigned Line = Subprogram.getLineNumber();
407 DebugLoc::get(MF.getOrCreateDebugLocID(CompileUnit.getGV(),
408 DbgScopeTrack.getCurScope(),
411 if (DW && DW->ShouldEmitDwarfDebug()) {
412 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
413 unsigned LabelID = DW->RecordInlinedFnStart(Subprogram,
414 DICompileUnit(PrevLocTpl.CompileUnit),
417 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
418 BuildMI(MBB, DL, II).addImm(LabelID);
421 // Record the source line.
422 unsigned Line = Subprogram.getLineNumber();
423 MF.setDefaultDebugLoc(
424 DebugLoc::get(MF.getOrCreateDebugLocID(CompileUnit.getGV(),
425 DbgScopeTrack.getCurScope(),
427 if (DW && DW->ShouldEmitDwarfDebug()) {
428 // llvm.dbg.func_start also defines beginning of function scope.
429 DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
435 case Intrinsic::dbg_declare: {
436 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
437 Value *Variable = DI->getVariable();
438 if (DIDescriptor::ValidDebugInfo(Variable, CodeGenOpt::None) &&
439 DW && DW->ShouldEmitDwarfDebug()) {
440 // Determine the address of the declared object.
441 Value *Address = DI->getAddress();
442 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
443 Address = BCI->getOperand(0);
444 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
445 // Don't handle byval struct arguments or VLAs, for example.
447 DenseMap<const AllocaInst*, int>::iterator SI =
448 StaticAllocaMap.find(AI);
449 if (SI == StaticAllocaMap.end()) break; // VLAs.
452 // Determine the debug globalvariable.
453 GlobalValue *GV = cast<GlobalVariable>(Variable);
455 // Build the DECLARE instruction.
456 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
457 MachineInstr *DeclareMI
458 = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
459 DIVariable DV(cast<GlobalVariable>(GV));
461 // This is a local variable
462 DW->RecordVariableScope(DV, DeclareMI);
467 case Intrinsic::eh_exception: {
468 MVT VT = TLI.getValueType(I->getType());
469 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
471 case TargetLowering::Expand: {
472 if (!MBB->isLandingPad()) {
473 // FIXME: Mark exception register as live in. Hack for PR1508.
474 unsigned Reg = TLI.getExceptionAddressRegister();
475 if (Reg) MBB->addLiveIn(Reg);
477 unsigned Reg = TLI.getExceptionAddressRegister();
478 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
479 unsigned ResultReg = createResultReg(RC);
480 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
482 assert(InsertedCopy && "Can't copy address registers!");
483 InsertedCopy = InsertedCopy;
484 UpdateValueMap(I, ResultReg);
490 case Intrinsic::eh_selector_i32:
491 case Intrinsic::eh_selector_i64: {
492 MVT VT = TLI.getValueType(I->getType());
493 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
495 case TargetLowering::Expand: {
496 MVT VT = (IID == Intrinsic::eh_selector_i32 ?
497 MVT::i32 : MVT::i64);
500 if (MBB->isLandingPad())
501 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
504 CatchInfoLost.insert(cast<CallInst>(I));
506 // FIXME: Mark exception selector register as live in. Hack for PR1508.
507 unsigned Reg = TLI.getExceptionSelectorRegister();
508 if (Reg) MBB->addLiveIn(Reg);
511 unsigned Reg = TLI.getExceptionSelectorRegister();
512 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
513 unsigned ResultReg = createResultReg(RC);
514 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
516 assert(InsertedCopy && "Can't copy address registers!");
517 InsertedCopy = InsertedCopy;
518 UpdateValueMap(I, ResultReg);
521 getRegForValue(Constant::getNullValue(I->getType()));
522 UpdateValueMap(I, ResultReg);
533 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
534 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
535 MVT DstVT = TLI.getValueType(I->getType());
537 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
538 DstVT == MVT::Other || !DstVT.isSimple())
539 // Unhandled type. Halt "fast" selection and bail.
542 // Check if the destination type is legal. Or as a special case,
543 // it may be i1 if we're doing a truncate because that's
544 // easy and somewhat common.
545 if (!TLI.isTypeLegal(DstVT))
546 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
547 // Unhandled type. Halt "fast" selection and bail.
550 // Check if the source operand is legal. Or as a special case,
551 // it may be i1 if we're doing zero-extension because that's
552 // easy and somewhat common.
553 if (!TLI.isTypeLegal(SrcVT))
554 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
555 // Unhandled type. Halt "fast" selection and bail.
558 unsigned InputReg = getRegForValue(I->getOperand(0));
560 // Unhandled operand. Halt "fast" selection and bail.
563 // If the operand is i1, arrange for the high bits in the register to be zero.
564 if (SrcVT == MVT::i1) {
565 SrcVT = TLI.getTypeToTransformTo(SrcVT);
566 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
570 // If the result is i1, truncate to the target's type for i1 first.
571 if (DstVT == MVT::i1)
572 DstVT = TLI.getTypeToTransformTo(DstVT);
574 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
581 UpdateValueMap(I, ResultReg);
585 bool FastISel::SelectBitCast(User *I) {
586 // If the bitcast doesn't change the type, just use the operand value.
587 if (I->getType() == I->getOperand(0)->getType()) {
588 unsigned Reg = getRegForValue(I->getOperand(0));
591 UpdateValueMap(I, Reg);
595 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
596 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
597 MVT DstVT = TLI.getValueType(I->getType());
599 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
600 DstVT == MVT::Other || !DstVT.isSimple() ||
601 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
602 // Unhandled type. Halt "fast" selection and bail.
605 unsigned Op0 = getRegForValue(I->getOperand(0));
607 // Unhandled operand. Halt "fast" selection and bail.
610 // First, try to perform the bitcast by inserting a reg-reg copy.
611 unsigned ResultReg = 0;
612 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
613 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
614 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
615 ResultReg = createResultReg(DstClass);
617 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
618 Op0, DstClass, SrcClass);
623 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
625 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
626 ISD::BIT_CONVERT, Op0);
631 UpdateValueMap(I, ResultReg);
636 FastISel::SelectInstruction(Instruction *I) {
637 return SelectOperator(I, I->getOpcode());
640 /// FastEmitBranch - Emit an unconditional branch to the given block,
641 /// unless it is the immediate (fall-through) successor, and update
644 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
645 MachineFunction::iterator NextMBB =
646 next(MachineFunction::iterator(MBB));
648 if (MBB->isLayoutSuccessor(MSucc)) {
649 // The unconditional fall-through case, which needs no instructions.
651 // The unconditional branch case.
652 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
654 MBB->addSuccessor(MSucc);
658 FastISel::SelectOperator(User *I, unsigned Opcode) {
660 case Instruction::Add: {
661 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
662 return SelectBinaryOp(I, Opc);
664 case Instruction::Sub: {
665 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
666 return SelectBinaryOp(I, Opc);
668 case Instruction::Mul: {
669 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
670 return SelectBinaryOp(I, Opc);
672 case Instruction::SDiv:
673 return SelectBinaryOp(I, ISD::SDIV);
674 case Instruction::UDiv:
675 return SelectBinaryOp(I, ISD::UDIV);
676 case Instruction::FDiv:
677 return SelectBinaryOp(I, ISD::FDIV);
678 case Instruction::SRem:
679 return SelectBinaryOp(I, ISD::SREM);
680 case Instruction::URem:
681 return SelectBinaryOp(I, ISD::UREM);
682 case Instruction::FRem:
683 return SelectBinaryOp(I, ISD::FREM);
684 case Instruction::Shl:
685 return SelectBinaryOp(I, ISD::SHL);
686 case Instruction::LShr:
687 return SelectBinaryOp(I, ISD::SRL);
688 case Instruction::AShr:
689 return SelectBinaryOp(I, ISD::SRA);
690 case Instruction::And:
691 return SelectBinaryOp(I, ISD::AND);
692 case Instruction::Or:
693 return SelectBinaryOp(I, ISD::OR);
694 case Instruction::Xor:
695 return SelectBinaryOp(I, ISD::XOR);
697 case Instruction::GetElementPtr:
698 return SelectGetElementPtr(I);
700 case Instruction::Br: {
701 BranchInst *BI = cast<BranchInst>(I);
703 if (BI->isUnconditional()) {
704 BasicBlock *LLVMSucc = BI->getSuccessor(0);
705 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
706 FastEmitBranch(MSucc);
710 // Conditional branches are not handed yet.
711 // Halt "fast" selection and bail.
715 case Instruction::Unreachable:
719 case Instruction::PHI:
720 // PHI nodes are already emitted.
723 case Instruction::Alloca:
724 // FunctionLowering has the static-sized case covered.
725 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
728 // Dynamic-sized alloca is not handled yet.
731 case Instruction::Call:
732 return SelectCall(I);
734 case Instruction::BitCast:
735 return SelectBitCast(I);
737 case Instruction::FPToSI:
738 return SelectCast(I, ISD::FP_TO_SINT);
739 case Instruction::ZExt:
740 return SelectCast(I, ISD::ZERO_EXTEND);
741 case Instruction::SExt:
742 return SelectCast(I, ISD::SIGN_EXTEND);
743 case Instruction::Trunc:
744 return SelectCast(I, ISD::TRUNCATE);
745 case Instruction::SIToFP:
746 return SelectCast(I, ISD::SINT_TO_FP);
748 case Instruction::IntToPtr: // Deliberate fall-through.
749 case Instruction::PtrToInt: {
750 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
751 MVT DstVT = TLI.getValueType(I->getType());
752 if (DstVT.bitsGT(SrcVT))
753 return SelectCast(I, ISD::ZERO_EXTEND);
754 if (DstVT.bitsLT(SrcVT))
755 return SelectCast(I, ISD::TRUNCATE);
756 unsigned Reg = getRegForValue(I->getOperand(0));
757 if (Reg == 0) return false;
758 UpdateValueMap(I, Reg);
763 // Unhandled instruction. Halt "fast" selection and bail.
768 FastISel::FastISel(MachineFunction &mf,
769 MachineModuleInfo *mmi,
771 DenseMap<const Value *, unsigned> &vm,
772 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
773 DenseMap<const AllocaInst *, int> &am
775 , SmallSet<Instruction*, 8> &cil
788 MRI(MF.getRegInfo()),
789 MFI(*MF.getFrameInfo()),
790 MCP(*MF.getConstantPool()),
792 TD(*TM.getTargetData()),
793 TII(*TM.getInstrInfo()),
794 TLI(*TM.getTargetLowering()) {
797 FastISel::~FastISel() {}
799 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
804 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
805 ISD::NodeType, unsigned /*Op0*/) {
809 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
810 ISD::NodeType, unsigned /*Op0*/,
815 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
816 ISD::NodeType, uint64_t /*Imm*/) {
820 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
821 ISD::NodeType, ConstantFP * /*FPImm*/) {
825 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
826 ISD::NodeType, unsigned /*Op0*/,
831 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
832 ISD::NodeType, unsigned /*Op0*/,
833 ConstantFP * /*FPImm*/) {
837 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
839 unsigned /*Op0*/, unsigned /*Op1*/,
844 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
845 /// to emit an instruction with an immediate operand using FastEmit_ri.
846 /// If that fails, it materializes the immediate into a register and try
847 /// FastEmit_rr instead.
848 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
849 unsigned Op0, uint64_t Imm,
850 MVT::SimpleValueType ImmType) {
851 // First check if immediate type is legal. If not, we can't use the ri form.
852 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
855 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
856 if (MaterialReg == 0)
858 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
861 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
862 /// to emit an instruction with a floating-point immediate operand using
863 /// FastEmit_rf. If that fails, it materializes the immediate into a register
864 /// and try FastEmit_rr instead.
865 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
866 unsigned Op0, ConstantFP *FPImm,
867 MVT::SimpleValueType ImmType) {
868 // First check if immediate type is legal. If not, we can't use the rf form.
869 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
873 // Materialize the constant in a register.
874 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
875 if (MaterialReg == 0) {
876 // If the target doesn't have a way to directly enter a floating-point
877 // value into a register, use an alternate approach.
878 // TODO: The current approach only supports floating-point constants
879 // that can be constructed by conversion from integer values. This should
880 // be replaced by code that creates a load from a constant-pool entry,
881 // which will require some target-specific work.
882 const APFloat &Flt = FPImm->getValueAPF();
883 MVT IntVT = TLI.getPointerTy();
886 uint32_t IntBitWidth = IntVT.getSizeInBits();
888 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
889 APFloat::rmTowardZero, &isExact);
892 APInt IntVal(IntBitWidth, 2, x);
894 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
895 ISD::Constant, IntVal.getZExtValue());
898 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
899 ISD::SINT_TO_FP, IntegerReg);
900 if (MaterialReg == 0)
903 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
906 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
907 return MRI.createVirtualRegister(RC);
910 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
911 const TargetRegisterClass* RC) {
912 unsigned ResultReg = createResultReg(RC);
913 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
915 BuildMI(MBB, DL, II, ResultReg);
919 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
920 const TargetRegisterClass *RC,
922 unsigned ResultReg = createResultReg(RC);
923 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
925 if (II.getNumDefs() >= 1)
926 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
928 BuildMI(MBB, DL, II).addReg(Op0);
929 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
930 II.ImplicitDefs[0], RC, RC);
938 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
939 const TargetRegisterClass *RC,
940 unsigned Op0, unsigned Op1) {
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).addReg(Op1);
947 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
948 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
949 II.ImplicitDefs[0], RC, RC);
956 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
957 const TargetRegisterClass *RC,
958 unsigned Op0, 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).addImm(Imm);
965 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
966 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
967 II.ImplicitDefs[0], RC, RC);
974 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
975 const TargetRegisterClass *RC,
976 unsigned Op0, ConstantFP *FPImm) {
977 unsigned ResultReg = createResultReg(RC);
978 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
980 if (II.getNumDefs() >= 1)
981 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
983 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
984 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
985 II.ImplicitDefs[0], RC, RC);
992 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
993 const TargetRegisterClass *RC,
994 unsigned Op0, unsigned Op1, uint64_t Imm) {
995 unsigned ResultReg = createResultReg(RC);
996 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
998 if (II.getNumDefs() >= 1)
999 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
1001 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
1002 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1003 II.ImplicitDefs[0], RC, RC);
1010 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
1011 const TargetRegisterClass *RC,
1013 unsigned ResultReg = createResultReg(RC);
1014 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1016 if (II.getNumDefs() >= 1)
1017 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
1019 BuildMI(MBB, DL, II).addImm(Imm);
1020 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1021 II.ImplicitDefs[0], RC, RC);
1028 unsigned FastISel::FastEmitInst_extractsubreg(MVT::SimpleValueType RetVT,
1029 unsigned Op0, uint32_t Idx) {
1030 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1032 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1033 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
1035 if (II.getNumDefs() >= 1)
1036 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1038 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1039 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1040 II.ImplicitDefs[0], RC, RC);
1047 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1048 /// with all but the least significant bit set to zero.
1049 unsigned FastISel::FastEmitZExtFromI1(MVT::SimpleValueType VT, unsigned Op) {
1050 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);