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.isNull() && !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);
365 // If ID is 0 then this was not an end of inlined region.
366 BuildMI(MBB, DL, II).addImm(ID);
368 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
369 ID = DW->RecordRegionEnd(cast<GlobalVariable>(REI->getContext()));
370 BuildMI(MBB, DL, II).addImm(ID);
375 case Intrinsic::dbg_func_start: {
376 if (!DW) return true;
377 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
378 Value *SP = FSI->getSubprogram();
380 if (DW->ValidDebugInfo(SP, true)) {
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 unsigned SrcFile = DW->getOrCreateSourceID(CompileUnit.getDirectory(Dir),
388 CompileUnit.getFilename(FN));
390 // Record the source line.
391 unsigned Line = Subprogram.getLineNumber();
392 unsigned LabelID = DW->RecordSourceLine(Line, 0, SrcFile);
393 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(SrcFile, Line, 0)));
394 if (!Subprogram.describes(MF.getFunction())) {
395 // This is a beginning of an inlined function.
396 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
397 BuildMI(MBB, DL, II).addImm(LabelID);
398 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
399 DW->RecordInlinedFnStart(FSI, Subprogram, LabelID,
404 // llvm.dbg.func_start also defines beginning of function scope.
405 DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
411 case Intrinsic::dbg_declare: {
412 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
413 Value *Variable = DI->getVariable();
414 if (DW && DW->ValidDebugInfo(Variable, true)) {
415 // Determine the address of the declared object.
416 Value *Address = DI->getAddress();
417 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
418 Address = BCI->getOperand(0);
419 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
420 // Don't handle byval struct arguments or VLAs, for example.
422 DenseMap<const AllocaInst*, int>::iterator SI =
423 StaticAllocaMap.find(AI);
424 if (SI == StaticAllocaMap.end()) break; // VLAs.
427 // Determine the debug globalvariable.
428 GlobalValue *GV = cast<GlobalVariable>(Variable);
430 // Build the DECLARE instruction.
431 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
432 MachineInstr *DeclareMI
433 = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
434 DIVariable DV(cast<GlobalVariable>(GV));
436 // This is a local variable
437 DW->RecordVariableScope(DV, DeclareMI);
442 case Intrinsic::eh_exception: {
443 MVT VT = TLI.getValueType(I->getType());
444 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
446 case TargetLowering::Expand: {
447 if (!MBB->isLandingPad()) {
448 // FIXME: Mark exception register as live in. Hack for PR1508.
449 unsigned Reg = TLI.getExceptionAddressRegister();
450 if (Reg) MBB->addLiveIn(Reg);
452 unsigned Reg = TLI.getExceptionAddressRegister();
453 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
454 unsigned ResultReg = createResultReg(RC);
455 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
457 assert(InsertedCopy && "Can't copy address registers!");
458 InsertedCopy = InsertedCopy;
459 UpdateValueMap(I, ResultReg);
465 case Intrinsic::eh_selector_i32:
466 case Intrinsic::eh_selector_i64: {
467 MVT VT = TLI.getValueType(I->getType());
468 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
470 case TargetLowering::Expand: {
471 MVT VT = (IID == Intrinsic::eh_selector_i32 ?
472 MVT::i32 : MVT::i64);
475 if (MBB->isLandingPad())
476 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
479 CatchInfoLost.insert(cast<CallInst>(I));
481 // FIXME: Mark exception selector register as live in. Hack for PR1508.
482 unsigned Reg = TLI.getExceptionSelectorRegister();
483 if (Reg) MBB->addLiveIn(Reg);
486 unsigned Reg = TLI.getExceptionSelectorRegister();
487 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
488 unsigned ResultReg = createResultReg(RC);
489 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
491 assert(InsertedCopy && "Can't copy address registers!");
492 InsertedCopy = InsertedCopy;
493 UpdateValueMap(I, ResultReg);
496 getRegForValue(Constant::getNullValue(I->getType()));
497 UpdateValueMap(I, ResultReg);
508 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
509 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
510 MVT DstVT = TLI.getValueType(I->getType());
512 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
513 DstVT == MVT::Other || !DstVT.isSimple())
514 // Unhandled type. Halt "fast" selection and bail.
517 // Check if the destination type is legal. Or as a special case,
518 // it may be i1 if we're doing a truncate because that's
519 // easy and somewhat common.
520 if (!TLI.isTypeLegal(DstVT))
521 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
522 // Unhandled type. Halt "fast" selection and bail.
525 // Check if the source operand is legal. Or as a special case,
526 // it may be i1 if we're doing zero-extension because that's
527 // easy and somewhat common.
528 if (!TLI.isTypeLegal(SrcVT))
529 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
530 // Unhandled type. Halt "fast" selection and bail.
533 unsigned InputReg = getRegForValue(I->getOperand(0));
535 // Unhandled operand. Halt "fast" selection and bail.
538 // If the operand is i1, arrange for the high bits in the register to be zero.
539 if (SrcVT == MVT::i1) {
540 SrcVT = TLI.getTypeToTransformTo(SrcVT);
541 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
545 // If the result is i1, truncate to the target's type for i1 first.
546 if (DstVT == MVT::i1)
547 DstVT = TLI.getTypeToTransformTo(DstVT);
549 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
556 UpdateValueMap(I, ResultReg);
560 bool FastISel::SelectBitCast(User *I) {
561 // If the bitcast doesn't change the type, just use the operand value.
562 if (I->getType() == I->getOperand(0)->getType()) {
563 unsigned Reg = getRegForValue(I->getOperand(0));
566 UpdateValueMap(I, Reg);
570 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
571 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
572 MVT DstVT = TLI.getValueType(I->getType());
574 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
575 DstVT == MVT::Other || !DstVT.isSimple() ||
576 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
577 // Unhandled type. Halt "fast" selection and bail.
580 unsigned Op0 = getRegForValue(I->getOperand(0));
582 // Unhandled operand. Halt "fast" selection and bail.
585 // First, try to perform the bitcast by inserting a reg-reg copy.
586 unsigned ResultReg = 0;
587 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
588 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
589 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
590 ResultReg = createResultReg(DstClass);
592 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
593 Op0, DstClass, SrcClass);
598 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
600 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
601 ISD::BIT_CONVERT, Op0);
606 UpdateValueMap(I, ResultReg);
611 FastISel::SelectInstruction(Instruction *I) {
612 return SelectOperator(I, I->getOpcode());
615 /// FastEmitBranch - Emit an unconditional branch to the given block,
616 /// unless it is the immediate (fall-through) successor, and update
619 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
620 MachineFunction::iterator NextMBB =
621 next(MachineFunction::iterator(MBB));
623 if (MBB->isLayoutSuccessor(MSucc)) {
624 // The unconditional fall-through case, which needs no instructions.
626 // The unconditional branch case.
627 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
629 MBB->addSuccessor(MSucc);
633 FastISel::SelectOperator(User *I, unsigned Opcode) {
635 case Instruction::Add: {
636 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
637 return SelectBinaryOp(I, Opc);
639 case Instruction::Sub: {
640 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
641 return SelectBinaryOp(I, Opc);
643 case Instruction::Mul: {
644 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
645 return SelectBinaryOp(I, Opc);
647 case Instruction::SDiv:
648 return SelectBinaryOp(I, ISD::SDIV);
649 case Instruction::UDiv:
650 return SelectBinaryOp(I, ISD::UDIV);
651 case Instruction::FDiv:
652 return SelectBinaryOp(I, ISD::FDIV);
653 case Instruction::SRem:
654 return SelectBinaryOp(I, ISD::SREM);
655 case Instruction::URem:
656 return SelectBinaryOp(I, ISD::UREM);
657 case Instruction::FRem:
658 return SelectBinaryOp(I, ISD::FREM);
659 case Instruction::Shl:
660 return SelectBinaryOp(I, ISD::SHL);
661 case Instruction::LShr:
662 return SelectBinaryOp(I, ISD::SRL);
663 case Instruction::AShr:
664 return SelectBinaryOp(I, ISD::SRA);
665 case Instruction::And:
666 return SelectBinaryOp(I, ISD::AND);
667 case Instruction::Or:
668 return SelectBinaryOp(I, ISD::OR);
669 case Instruction::Xor:
670 return SelectBinaryOp(I, ISD::XOR);
672 case Instruction::GetElementPtr:
673 return SelectGetElementPtr(I);
675 case Instruction::Br: {
676 BranchInst *BI = cast<BranchInst>(I);
678 if (BI->isUnconditional()) {
679 BasicBlock *LLVMSucc = BI->getSuccessor(0);
680 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
681 FastEmitBranch(MSucc);
685 // Conditional branches are not handed yet.
686 // Halt "fast" selection and bail.
690 case Instruction::Unreachable:
694 case Instruction::PHI:
695 // PHI nodes are already emitted.
698 case Instruction::Alloca:
699 // FunctionLowering has the static-sized case covered.
700 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
703 // Dynamic-sized alloca is not handled yet.
706 case Instruction::Call:
707 return SelectCall(I);
709 case Instruction::BitCast:
710 return SelectBitCast(I);
712 case Instruction::FPToSI:
713 return SelectCast(I, ISD::FP_TO_SINT);
714 case Instruction::ZExt:
715 return SelectCast(I, ISD::ZERO_EXTEND);
716 case Instruction::SExt:
717 return SelectCast(I, ISD::SIGN_EXTEND);
718 case Instruction::Trunc:
719 return SelectCast(I, ISD::TRUNCATE);
720 case Instruction::SIToFP:
721 return SelectCast(I, ISD::SINT_TO_FP);
723 case Instruction::IntToPtr: // Deliberate fall-through.
724 case Instruction::PtrToInt: {
725 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
726 MVT DstVT = TLI.getValueType(I->getType());
727 if (DstVT.bitsGT(SrcVT))
728 return SelectCast(I, ISD::ZERO_EXTEND);
729 if (DstVT.bitsLT(SrcVT))
730 return SelectCast(I, ISD::TRUNCATE);
731 unsigned Reg = getRegForValue(I->getOperand(0));
732 if (Reg == 0) return false;
733 UpdateValueMap(I, Reg);
738 // Unhandled instruction. Halt "fast" selection and bail.
743 FastISel::FastISel(MachineFunction &mf,
744 MachineModuleInfo *mmi,
746 DenseMap<const Value *, unsigned> &vm,
747 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
748 DenseMap<const AllocaInst *, int> &am
750 , SmallSet<Instruction*, 8> &cil
763 MRI(MF.getRegInfo()),
764 MFI(*MF.getFrameInfo()),
765 MCP(*MF.getConstantPool()),
767 TD(*TM.getTargetData()),
768 TII(*TM.getInstrInfo()),
769 TLI(*TM.getTargetLowering()) {
772 FastISel::~FastISel() {}
774 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
779 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
780 ISD::NodeType, unsigned /*Op0*/) {
784 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
785 ISD::NodeType, unsigned /*Op0*/,
790 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
791 ISD::NodeType, uint64_t /*Imm*/) {
795 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
796 ISD::NodeType, ConstantFP * /*FPImm*/) {
800 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
801 ISD::NodeType, unsigned /*Op0*/,
806 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
807 ISD::NodeType, unsigned /*Op0*/,
808 ConstantFP * /*FPImm*/) {
812 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
814 unsigned /*Op0*/, unsigned /*Op1*/,
819 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
820 /// to emit an instruction with an immediate operand using FastEmit_ri.
821 /// If that fails, it materializes the immediate into a register and try
822 /// FastEmit_rr instead.
823 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
824 unsigned Op0, uint64_t Imm,
825 MVT::SimpleValueType ImmType) {
826 // First check if immediate type is legal. If not, we can't use the ri form.
827 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
830 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
831 if (MaterialReg == 0)
833 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
836 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
837 /// to emit an instruction with a floating-point immediate operand using
838 /// FastEmit_rf. If that fails, it materializes the immediate into a register
839 /// and try FastEmit_rr instead.
840 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
841 unsigned Op0, ConstantFP *FPImm,
842 MVT::SimpleValueType ImmType) {
843 // First check if immediate type is legal. If not, we can't use the rf form.
844 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
848 // Materialize the constant in a register.
849 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
850 if (MaterialReg == 0) {
851 // If the target doesn't have a way to directly enter a floating-point
852 // value into a register, use an alternate approach.
853 // TODO: The current approach only supports floating-point constants
854 // that can be constructed by conversion from integer values. This should
855 // be replaced by code that creates a load from a constant-pool entry,
856 // which will require some target-specific work.
857 const APFloat &Flt = FPImm->getValueAPF();
858 MVT IntVT = TLI.getPointerTy();
861 uint32_t IntBitWidth = IntVT.getSizeInBits();
863 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
864 APFloat::rmTowardZero, &isExact);
867 APInt IntVal(IntBitWidth, 2, x);
869 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
870 ISD::Constant, IntVal.getZExtValue());
873 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
874 ISD::SINT_TO_FP, IntegerReg);
875 if (MaterialReg == 0)
878 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
881 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
882 return MRI.createVirtualRegister(RC);
885 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
886 const TargetRegisterClass* RC) {
887 unsigned ResultReg = createResultReg(RC);
888 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
890 BuildMI(MBB, DL, II, ResultReg);
894 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
895 const TargetRegisterClass *RC,
897 unsigned ResultReg = createResultReg(RC);
898 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
900 if (II.getNumDefs() >= 1)
901 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
903 BuildMI(MBB, DL, II).addReg(Op0);
904 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
905 II.ImplicitDefs[0], RC, RC);
913 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
914 const TargetRegisterClass *RC,
915 unsigned Op0, unsigned Op1) {
916 unsigned ResultReg = createResultReg(RC);
917 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
919 if (II.getNumDefs() >= 1)
920 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
922 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
923 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
924 II.ImplicitDefs[0], RC, RC);
931 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
932 const TargetRegisterClass *RC,
933 unsigned Op0, uint64_t Imm) {
934 unsigned ResultReg = createResultReg(RC);
935 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
937 if (II.getNumDefs() >= 1)
938 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
940 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
941 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
942 II.ImplicitDefs[0], RC, RC);
949 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
950 const TargetRegisterClass *RC,
951 unsigned Op0, ConstantFP *FPImm) {
952 unsigned ResultReg = createResultReg(RC);
953 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
955 if (II.getNumDefs() >= 1)
956 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
958 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
959 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
960 II.ImplicitDefs[0], RC, RC);
967 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
968 const TargetRegisterClass *RC,
969 unsigned Op0, unsigned Op1, uint64_t Imm) {
970 unsigned ResultReg = createResultReg(RC);
971 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
973 if (II.getNumDefs() >= 1)
974 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
976 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
977 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
978 II.ImplicitDefs[0], RC, RC);
985 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
986 const TargetRegisterClass *RC,
988 unsigned ResultReg = createResultReg(RC);
989 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
991 if (II.getNumDefs() >= 1)
992 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
994 BuildMI(MBB, DL, II).addImm(Imm);
995 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
996 II.ImplicitDefs[0], RC, RC);
1003 unsigned FastISel::FastEmitInst_extractsubreg(MVT::SimpleValueType RetVT,
1004 unsigned Op0, uint32_t Idx) {
1005 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1007 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1008 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
1010 if (II.getNumDefs() >= 1)
1011 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1013 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1014 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1015 II.ImplicitDefs[0], RC, RC);
1022 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1023 /// with all but the least significant bit set to zero.
1024 unsigned FastISel::FastEmitZExtFromI1(MVT::SimpleValueType VT, unsigned Op) {
1025 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);