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 (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(), Line, Col);
335 setCurDebugLoc(DebugLoc::get(Idx));
339 case Intrinsic::dbg_region_start: {
340 DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I);
341 if (DIDescriptor::ValidDebugInfo(RSI->getContext(), CodeGenOpt::None) &&
342 DW && DW->ShouldEmitDwarfDebug()) {
344 DW->RecordRegionStart(cast<GlobalVariable>(RSI->getContext()));
345 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
346 BuildMI(MBB, DL, II).addImm(ID);
350 case Intrinsic::dbg_region_end: {
351 DbgRegionEndInst *REI = cast<DbgRegionEndInst>(I);
352 if (DIDescriptor::ValidDebugInfo(REI->getContext(), CodeGenOpt::None) &&
353 DW && DW->ShouldEmitDwarfDebug()) {
355 DISubprogram Subprogram(cast<GlobalVariable>(REI->getContext()));
356 if (!Subprogram.isNull() && !Subprogram.describes(MF.getFunction())) {
357 // This is end of an inlined function.
358 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
359 ID = DW->RecordInlinedFnEnd(Subprogram);
361 // Returned ID is 0 if this is unbalanced "end of inlined
362 // scope". This could happen if optimizer eats dbg intrinsics
363 // or "beginning of inlined scope" is not recoginized due to
364 // missing location info. In such cases, do ignore this region.end.
365 BuildMI(MBB, DL, II).addImm(ID);
367 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
368 ID = DW->RecordRegionEnd(cast<GlobalVariable>(REI->getContext()),
370 BuildMI(MBB, DL, II).addImm(ID);
375 case Intrinsic::dbg_func_start: {
376 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
377 Value *SP = FSI->getSubprogram();
378 if (!DIDescriptor::ValidDebugInfo(SP, CodeGenOpt::None))
381 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is what
382 // (most?) gdb expects.
383 DebugLoc PrevLoc = DL;
384 DISubprogram Subprogram(cast<GlobalVariable>(SP));
385 DICompileUnit CompileUnit = Subprogram.getCompileUnit();
387 if (!Subprogram.describes(MF.getFunction())) {
388 // This is a beginning of an inlined function.
390 // If llvm.dbg.func.start is seen in a new block before any
391 // llvm.dbg.stoppoint intrinsic then the location info is unknown.
392 // FIXME : Why DebugLoc is reset at the beginning of each block ?
393 if (PrevLoc.isUnknown())
396 // Record the source line.
397 unsigned Line = Subprogram.getLineNumber();
398 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(
399 CompileUnit.getGV(), Line, 0)));
401 if (DW && DW->ShouldEmitDwarfDebug()) {
402 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
403 unsigned LabelID = DW->RecordInlinedFnStart(Subprogram,
404 DICompileUnit(PrevLocTpl.CompileUnit),
407 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
408 BuildMI(MBB, DL, II).addImm(LabelID);
411 // Record the source line.
412 unsigned Line = Subprogram.getLineNumber();
413 MF.setDefaultDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(
414 CompileUnit.getGV(), Line, 0)));
416 if (DW && DW->ShouldEmitDwarfDebug())
417 // llvm.dbg.func_start also defines beginning of function scope.
418 DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
423 case Intrinsic::dbg_declare: {
424 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
425 Value *Variable = DI->getVariable();
426 if (DIDescriptor::ValidDebugInfo(Variable, CodeGenOpt::None) &&
427 DW && DW->ShouldEmitDwarfDebug()) {
428 // Determine the address of the declared object.
429 Value *Address = DI->getAddress();
430 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
431 Address = BCI->getOperand(0);
432 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
433 // Don't handle byval struct arguments or VLAs, for example.
435 DenseMap<const AllocaInst*, int>::iterator SI =
436 StaticAllocaMap.find(AI);
437 if (SI == StaticAllocaMap.end()) break; // VLAs.
440 // Determine the debug globalvariable.
441 GlobalValue *GV = cast<GlobalVariable>(Variable);
443 // Build the DECLARE instruction.
444 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
445 MachineInstr *DeclareMI
446 = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
447 DIVariable DV(cast<GlobalVariable>(GV));
449 // This is a local variable
450 DW->RecordVariableScope(DV, DeclareMI);
455 case Intrinsic::eh_exception: {
456 MVT VT = TLI.getValueType(I->getType());
457 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
459 case TargetLowering::Expand: {
460 if (!MBB->isLandingPad()) {
461 // FIXME: Mark exception register as live in. Hack for PR1508.
462 unsigned Reg = TLI.getExceptionAddressRegister();
463 if (Reg) MBB->addLiveIn(Reg);
465 unsigned Reg = TLI.getExceptionAddressRegister();
466 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
467 unsigned ResultReg = createResultReg(RC);
468 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
470 assert(InsertedCopy && "Can't copy address registers!");
471 InsertedCopy = InsertedCopy;
472 UpdateValueMap(I, ResultReg);
478 case Intrinsic::eh_selector_i32:
479 case Intrinsic::eh_selector_i64: {
480 MVT VT = TLI.getValueType(I->getType());
481 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
483 case TargetLowering::Expand: {
484 MVT VT = (IID == Intrinsic::eh_selector_i32 ?
485 MVT::i32 : MVT::i64);
488 if (MBB->isLandingPad())
489 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
492 CatchInfoLost.insert(cast<CallInst>(I));
494 // FIXME: Mark exception selector register as live in. Hack for PR1508.
495 unsigned Reg = TLI.getExceptionSelectorRegister();
496 if (Reg) MBB->addLiveIn(Reg);
499 unsigned Reg = TLI.getExceptionSelectorRegister();
500 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
501 unsigned ResultReg = createResultReg(RC);
502 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
504 assert(InsertedCopy && "Can't copy address registers!");
505 InsertedCopy = InsertedCopy;
506 UpdateValueMap(I, ResultReg);
509 getRegForValue(Constant::getNullValue(I->getType()));
510 UpdateValueMap(I, ResultReg);
521 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
522 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
523 MVT DstVT = TLI.getValueType(I->getType());
525 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
526 DstVT == MVT::Other || !DstVT.isSimple())
527 // Unhandled type. Halt "fast" selection and bail.
530 // Check if the destination type is legal. Or as a special case,
531 // it may be i1 if we're doing a truncate because that's
532 // easy and somewhat common.
533 if (!TLI.isTypeLegal(DstVT))
534 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
535 // Unhandled type. Halt "fast" selection and bail.
538 // Check if the source operand is legal. Or as a special case,
539 // it may be i1 if we're doing zero-extension because that's
540 // easy and somewhat common.
541 if (!TLI.isTypeLegal(SrcVT))
542 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
543 // Unhandled type. Halt "fast" selection and bail.
546 unsigned InputReg = getRegForValue(I->getOperand(0));
548 // Unhandled operand. Halt "fast" selection and bail.
551 // If the operand is i1, arrange for the high bits in the register to be zero.
552 if (SrcVT == MVT::i1) {
553 SrcVT = TLI.getTypeToTransformTo(SrcVT);
554 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
558 // If the result is i1, truncate to the target's type for i1 first.
559 if (DstVT == MVT::i1)
560 DstVT = TLI.getTypeToTransformTo(DstVT);
562 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
569 UpdateValueMap(I, ResultReg);
573 bool FastISel::SelectBitCast(User *I) {
574 // If the bitcast doesn't change the type, just use the operand value.
575 if (I->getType() == I->getOperand(0)->getType()) {
576 unsigned Reg = getRegForValue(I->getOperand(0));
579 UpdateValueMap(I, Reg);
583 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
584 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
585 MVT DstVT = TLI.getValueType(I->getType());
587 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
588 DstVT == MVT::Other || !DstVT.isSimple() ||
589 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
590 // Unhandled type. Halt "fast" selection and bail.
593 unsigned Op0 = getRegForValue(I->getOperand(0));
595 // Unhandled operand. Halt "fast" selection and bail.
598 // First, try to perform the bitcast by inserting a reg-reg copy.
599 unsigned ResultReg = 0;
600 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
601 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
602 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
603 ResultReg = createResultReg(DstClass);
605 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
606 Op0, DstClass, SrcClass);
611 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
613 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
614 ISD::BIT_CONVERT, Op0);
619 UpdateValueMap(I, ResultReg);
624 FastISel::SelectInstruction(Instruction *I) {
625 return SelectOperator(I, I->getOpcode());
628 /// FastEmitBranch - Emit an unconditional branch to the given block,
629 /// unless it is the immediate (fall-through) successor, and update
632 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
633 MachineFunction::iterator NextMBB =
634 next(MachineFunction::iterator(MBB));
636 if (MBB->isLayoutSuccessor(MSucc)) {
637 // The unconditional fall-through case, which needs no instructions.
639 // The unconditional branch case.
640 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
642 MBB->addSuccessor(MSucc);
646 FastISel::SelectOperator(User *I, unsigned Opcode) {
648 case Instruction::Add: {
649 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
650 return SelectBinaryOp(I, Opc);
652 case Instruction::Sub: {
653 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
654 return SelectBinaryOp(I, Opc);
656 case Instruction::Mul: {
657 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
658 return SelectBinaryOp(I, Opc);
660 case Instruction::SDiv:
661 return SelectBinaryOp(I, ISD::SDIV);
662 case Instruction::UDiv:
663 return SelectBinaryOp(I, ISD::UDIV);
664 case Instruction::FDiv:
665 return SelectBinaryOp(I, ISD::FDIV);
666 case Instruction::SRem:
667 return SelectBinaryOp(I, ISD::SREM);
668 case Instruction::URem:
669 return SelectBinaryOp(I, ISD::UREM);
670 case Instruction::FRem:
671 return SelectBinaryOp(I, ISD::FREM);
672 case Instruction::Shl:
673 return SelectBinaryOp(I, ISD::SHL);
674 case Instruction::LShr:
675 return SelectBinaryOp(I, ISD::SRL);
676 case Instruction::AShr:
677 return SelectBinaryOp(I, ISD::SRA);
678 case Instruction::And:
679 return SelectBinaryOp(I, ISD::AND);
680 case Instruction::Or:
681 return SelectBinaryOp(I, ISD::OR);
682 case Instruction::Xor:
683 return SelectBinaryOp(I, ISD::XOR);
685 case Instruction::GetElementPtr:
686 return SelectGetElementPtr(I);
688 case Instruction::Br: {
689 BranchInst *BI = cast<BranchInst>(I);
691 if (BI->isUnconditional()) {
692 BasicBlock *LLVMSucc = BI->getSuccessor(0);
693 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
694 FastEmitBranch(MSucc);
698 // Conditional branches are not handed yet.
699 // Halt "fast" selection and bail.
703 case Instruction::Unreachable:
707 case Instruction::PHI:
708 // PHI nodes are already emitted.
711 case Instruction::Alloca:
712 // FunctionLowering has the static-sized case covered.
713 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
716 // Dynamic-sized alloca is not handled yet.
719 case Instruction::Call:
720 return SelectCall(I);
722 case Instruction::BitCast:
723 return SelectBitCast(I);
725 case Instruction::FPToSI:
726 return SelectCast(I, ISD::FP_TO_SINT);
727 case Instruction::ZExt:
728 return SelectCast(I, ISD::ZERO_EXTEND);
729 case Instruction::SExt:
730 return SelectCast(I, ISD::SIGN_EXTEND);
731 case Instruction::Trunc:
732 return SelectCast(I, ISD::TRUNCATE);
733 case Instruction::SIToFP:
734 return SelectCast(I, ISD::SINT_TO_FP);
736 case Instruction::IntToPtr: // Deliberate fall-through.
737 case Instruction::PtrToInt: {
738 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
739 MVT DstVT = TLI.getValueType(I->getType());
740 if (DstVT.bitsGT(SrcVT))
741 return SelectCast(I, ISD::ZERO_EXTEND);
742 if (DstVT.bitsLT(SrcVT))
743 return SelectCast(I, ISD::TRUNCATE);
744 unsigned Reg = getRegForValue(I->getOperand(0));
745 if (Reg == 0) return false;
746 UpdateValueMap(I, Reg);
751 // Unhandled instruction. Halt "fast" selection and bail.
756 FastISel::FastISel(MachineFunction &mf,
757 MachineModuleInfo *mmi,
759 DenseMap<const Value *, unsigned> &vm,
760 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
761 DenseMap<const AllocaInst *, int> &am
763 , SmallSet<Instruction*, 8> &cil
776 MRI(MF.getRegInfo()),
777 MFI(*MF.getFrameInfo()),
778 MCP(*MF.getConstantPool()),
780 TD(*TM.getTargetData()),
781 TII(*TM.getInstrInfo()),
782 TLI(*TM.getTargetLowering()) {
785 FastISel::~FastISel() {}
787 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
792 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
793 ISD::NodeType, unsigned /*Op0*/) {
797 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
798 ISD::NodeType, unsigned /*Op0*/,
803 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
804 ISD::NodeType, uint64_t /*Imm*/) {
808 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
809 ISD::NodeType, ConstantFP * /*FPImm*/) {
813 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
814 ISD::NodeType, unsigned /*Op0*/,
819 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
820 ISD::NodeType, unsigned /*Op0*/,
821 ConstantFP * /*FPImm*/) {
825 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
827 unsigned /*Op0*/, unsigned /*Op1*/,
832 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
833 /// to emit an instruction with an immediate operand using FastEmit_ri.
834 /// If that fails, it materializes the immediate into a register and try
835 /// FastEmit_rr instead.
836 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
837 unsigned Op0, uint64_t Imm,
838 MVT::SimpleValueType ImmType) {
839 // First check if immediate type is legal. If not, we can't use the ri form.
840 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
843 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
844 if (MaterialReg == 0)
846 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
849 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
850 /// to emit an instruction with a floating-point immediate operand using
851 /// FastEmit_rf. If that fails, it materializes the immediate into a register
852 /// and try FastEmit_rr instead.
853 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
854 unsigned Op0, ConstantFP *FPImm,
855 MVT::SimpleValueType ImmType) {
856 // First check if immediate type is legal. If not, we can't use the rf form.
857 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
861 // Materialize the constant in a register.
862 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
863 if (MaterialReg == 0) {
864 // If the target doesn't have a way to directly enter a floating-point
865 // value into a register, use an alternate approach.
866 // TODO: The current approach only supports floating-point constants
867 // that can be constructed by conversion from integer values. This should
868 // be replaced by code that creates a load from a constant-pool entry,
869 // which will require some target-specific work.
870 const APFloat &Flt = FPImm->getValueAPF();
871 MVT IntVT = TLI.getPointerTy();
874 uint32_t IntBitWidth = IntVT.getSizeInBits();
876 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
877 APFloat::rmTowardZero, &isExact);
880 APInt IntVal(IntBitWidth, 2, x);
882 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
883 ISD::Constant, IntVal.getZExtValue());
886 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
887 ISD::SINT_TO_FP, IntegerReg);
888 if (MaterialReg == 0)
891 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
894 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
895 return MRI.createVirtualRegister(RC);
898 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
899 const TargetRegisterClass* RC) {
900 unsigned ResultReg = createResultReg(RC);
901 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
903 BuildMI(MBB, DL, II, ResultReg);
907 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
908 const TargetRegisterClass *RC,
910 unsigned ResultReg = createResultReg(RC);
911 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
913 if (II.getNumDefs() >= 1)
914 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
916 BuildMI(MBB, DL, II).addReg(Op0);
917 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
918 II.ImplicitDefs[0], RC, RC);
926 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
927 const TargetRegisterClass *RC,
928 unsigned Op0, unsigned Op1) {
929 unsigned ResultReg = createResultReg(RC);
930 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
932 if (II.getNumDefs() >= 1)
933 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
935 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
936 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
937 II.ImplicitDefs[0], RC, RC);
944 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
945 const TargetRegisterClass *RC,
946 unsigned Op0, uint64_t Imm) {
947 unsigned ResultReg = createResultReg(RC);
948 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
950 if (II.getNumDefs() >= 1)
951 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
953 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
954 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
955 II.ImplicitDefs[0], RC, RC);
962 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
963 const TargetRegisterClass *RC,
964 unsigned Op0, ConstantFP *FPImm) {
965 unsigned ResultReg = createResultReg(RC);
966 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
968 if (II.getNumDefs() >= 1)
969 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
971 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
972 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
973 II.ImplicitDefs[0], RC, RC);
980 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
981 const TargetRegisterClass *RC,
982 unsigned Op0, unsigned Op1, uint64_t Imm) {
983 unsigned ResultReg = createResultReg(RC);
984 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
986 if (II.getNumDefs() >= 1)
987 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
989 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
990 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
991 II.ImplicitDefs[0], RC, RC);
998 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
999 const TargetRegisterClass *RC,
1001 unsigned ResultReg = createResultReg(RC);
1002 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1004 if (II.getNumDefs() >= 1)
1005 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
1007 BuildMI(MBB, DL, II).addImm(Imm);
1008 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1009 II.ImplicitDefs[0], RC, RC);
1016 unsigned FastISel::FastEmitInst_extractsubreg(MVT::SimpleValueType RetVT,
1017 unsigned Op0, uint32_t Idx) {
1018 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1020 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1021 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
1023 if (II.getNumDefs() >= 1)
1024 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1026 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1027 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1028 II.ImplicitDefs[0], RC, RC);
1035 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1036 /// with all but the least significant bit set to zero.
1037 unsigned FastISel::FastEmitZExtFromI1(MVT::SimpleValueType VT, unsigned Op) {
1038 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);