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(), CodeGenOpt::None)) {
331 DICompileUnit CU(cast<GlobalVariable>(SPI->getContext()));
332 unsigned Line = SPI->getLine();
333 unsigned Col = SPI->getColumn();
334 unsigned ID = DW->RecordSourceLine(Line, Col, CU);
335 unsigned Idx = MF.getOrCreateDebugLocID(CU.getGV(), Line, Col);
336 setCurDebugLoc(DebugLoc::get(Idx));
337 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
338 BuildMI(MBB, DL, II).addImm(ID);
342 case Intrinsic::dbg_region_start: {
343 DbgRegionStartInst *RSI = cast<DbgRegionStartInst>(I);
344 if (DW && DW->ValidDebugInfo(RSI->getContext(), CodeGenOpt::None)) {
346 DW->RecordRegionStart(cast<GlobalVariable>(RSI->getContext()));
347 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
348 BuildMI(MBB, DL, II).addImm(ID);
352 case Intrinsic::dbg_region_end: {
353 DbgRegionEndInst *REI = cast<DbgRegionEndInst>(I);
354 if (DW && DW->ValidDebugInfo(REI->getContext(), CodeGenOpt::None)) {
356 DISubprogram Subprogram(cast<GlobalVariable>(REI->getContext()));
357 if (!Subprogram.isNull() && !Subprogram.describes(MF.getFunction())) {
358 // This is end of an inlined function.
359 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
360 ID = DW->RecordInlinedFnEnd(Subprogram);
362 // Returned ID is 0 if this is unbalanced "end of inlined
363 // scope". This could happen if optimizer eats dbg intrinsics
364 // or "beginning of inlined scope" is not recoginized due to
365 // missing location info. In such cases, do ignore this region.end.
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, 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())
395 // Record the source line.
396 unsigned Line = Subprogram.getLineNumber();
397 unsigned LabelID = DW->RecordSourceLine(Line, 0, CompileUnit);
398 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(
399 CompileUnit.getGV(), Line, 0)));
401 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DBG_LABEL);
402 BuildMI(MBB, DL, II).addImm(LabelID);
403 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
404 DW->RecordInlinedFnStart(FSI, Subprogram, LabelID,
405 DICompileUnit(PrevLocTpl.CompileUnit),
409 // Record the source line.
410 unsigned Line = Subprogram.getLineNumber();
411 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(
412 CompileUnit.getGV(), Line, 0)));
413 DW->RecordSourceLine(Line, 0, CompileUnit);
414 // llvm.dbg.func_start also defines beginning of function scope.
415 DW->RecordRegionStart(cast<GlobalVariable>(FSI->getSubprogram()));
421 case Intrinsic::dbg_declare: {
422 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
423 Value *Variable = DI->getVariable();
424 if (DW && DW->ValidDebugInfo(Variable, CodeGenOpt::None)) {
425 // Determine the address of the declared object.
426 Value *Address = DI->getAddress();
427 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
428 Address = BCI->getOperand(0);
429 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
430 // Don't handle byval struct arguments or VLAs, for example.
432 DenseMap<const AllocaInst*, int>::iterator SI =
433 StaticAllocaMap.find(AI);
434 if (SI == StaticAllocaMap.end()) break; // VLAs.
437 // Determine the debug globalvariable.
438 GlobalValue *GV = cast<GlobalVariable>(Variable);
440 // Build the DECLARE instruction.
441 const TargetInstrDesc &II = TII.get(TargetInstrInfo::DECLARE);
442 MachineInstr *DeclareMI
443 = BuildMI(MBB, DL, II).addFrameIndex(FI).addGlobalAddress(GV);
444 DIVariable DV(cast<GlobalVariable>(GV));
446 // This is a local variable
447 DW->RecordVariableScope(DV, DeclareMI);
452 case Intrinsic::eh_exception: {
453 MVT VT = TLI.getValueType(I->getType());
454 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
456 case TargetLowering::Expand: {
457 if (!MBB->isLandingPad()) {
458 // FIXME: Mark exception register as live in. Hack for PR1508.
459 unsigned Reg = TLI.getExceptionAddressRegister();
460 if (Reg) MBB->addLiveIn(Reg);
462 unsigned Reg = TLI.getExceptionAddressRegister();
463 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
464 unsigned ResultReg = createResultReg(RC);
465 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
467 assert(InsertedCopy && "Can't copy address registers!");
468 InsertedCopy = InsertedCopy;
469 UpdateValueMap(I, ResultReg);
475 case Intrinsic::eh_selector_i32:
476 case Intrinsic::eh_selector_i64: {
477 MVT VT = TLI.getValueType(I->getType());
478 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
480 case TargetLowering::Expand: {
481 MVT VT = (IID == Intrinsic::eh_selector_i32 ?
482 MVT::i32 : MVT::i64);
485 if (MBB->isLandingPad())
486 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
489 CatchInfoLost.insert(cast<CallInst>(I));
491 // FIXME: Mark exception selector register as live in. Hack for PR1508.
492 unsigned Reg = TLI.getExceptionSelectorRegister();
493 if (Reg) MBB->addLiveIn(Reg);
496 unsigned Reg = TLI.getExceptionSelectorRegister();
497 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
498 unsigned ResultReg = createResultReg(RC);
499 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
501 assert(InsertedCopy && "Can't copy address registers!");
502 InsertedCopy = InsertedCopy;
503 UpdateValueMap(I, ResultReg);
506 getRegForValue(Constant::getNullValue(I->getType()));
507 UpdateValueMap(I, ResultReg);
518 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
519 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
520 MVT DstVT = TLI.getValueType(I->getType());
522 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
523 DstVT == MVT::Other || !DstVT.isSimple())
524 // Unhandled type. Halt "fast" selection and bail.
527 // Check if the destination type is legal. Or as a special case,
528 // it may be i1 if we're doing a truncate because that's
529 // easy and somewhat common.
530 if (!TLI.isTypeLegal(DstVT))
531 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
532 // Unhandled type. Halt "fast" selection and bail.
535 // Check if the source operand is legal. Or as a special case,
536 // it may be i1 if we're doing zero-extension because that's
537 // easy and somewhat common.
538 if (!TLI.isTypeLegal(SrcVT))
539 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
540 // Unhandled type. Halt "fast" selection and bail.
543 unsigned InputReg = getRegForValue(I->getOperand(0));
545 // Unhandled operand. Halt "fast" selection and bail.
548 // If the operand is i1, arrange for the high bits in the register to be zero.
549 if (SrcVT == MVT::i1) {
550 SrcVT = TLI.getTypeToTransformTo(SrcVT);
551 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
555 // If the result is i1, truncate to the target's type for i1 first.
556 if (DstVT == MVT::i1)
557 DstVT = TLI.getTypeToTransformTo(DstVT);
559 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
566 UpdateValueMap(I, ResultReg);
570 bool FastISel::SelectBitCast(User *I) {
571 // If the bitcast doesn't change the type, just use the operand value.
572 if (I->getType() == I->getOperand(0)->getType()) {
573 unsigned Reg = getRegForValue(I->getOperand(0));
576 UpdateValueMap(I, Reg);
580 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
581 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
582 MVT DstVT = TLI.getValueType(I->getType());
584 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
585 DstVT == MVT::Other || !DstVT.isSimple() ||
586 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
587 // Unhandled type. Halt "fast" selection and bail.
590 unsigned Op0 = getRegForValue(I->getOperand(0));
592 // Unhandled operand. Halt "fast" selection and bail.
595 // First, try to perform the bitcast by inserting a reg-reg copy.
596 unsigned ResultReg = 0;
597 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
598 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
599 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
600 ResultReg = createResultReg(DstClass);
602 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
603 Op0, DstClass, SrcClass);
608 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
610 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
611 ISD::BIT_CONVERT, Op0);
616 UpdateValueMap(I, ResultReg);
621 FastISel::SelectInstruction(Instruction *I) {
622 return SelectOperator(I, I->getOpcode());
625 /// FastEmitBranch - Emit an unconditional branch to the given block,
626 /// unless it is the immediate (fall-through) successor, and update
629 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
630 MachineFunction::iterator NextMBB =
631 next(MachineFunction::iterator(MBB));
633 if (MBB->isLayoutSuccessor(MSucc)) {
634 // The unconditional fall-through case, which needs no instructions.
636 // The unconditional branch case.
637 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
639 MBB->addSuccessor(MSucc);
643 FastISel::SelectOperator(User *I, unsigned Opcode) {
645 case Instruction::Add: {
646 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FADD : ISD::ADD;
647 return SelectBinaryOp(I, Opc);
649 case Instruction::Sub: {
650 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FSUB : ISD::SUB;
651 return SelectBinaryOp(I, Opc);
653 case Instruction::Mul: {
654 ISD::NodeType Opc = I->getType()->isFPOrFPVector() ? ISD::FMUL : ISD::MUL;
655 return SelectBinaryOp(I, Opc);
657 case Instruction::SDiv:
658 return SelectBinaryOp(I, ISD::SDIV);
659 case Instruction::UDiv:
660 return SelectBinaryOp(I, ISD::UDIV);
661 case Instruction::FDiv:
662 return SelectBinaryOp(I, ISD::FDIV);
663 case Instruction::SRem:
664 return SelectBinaryOp(I, ISD::SREM);
665 case Instruction::URem:
666 return SelectBinaryOp(I, ISD::UREM);
667 case Instruction::FRem:
668 return SelectBinaryOp(I, ISD::FREM);
669 case Instruction::Shl:
670 return SelectBinaryOp(I, ISD::SHL);
671 case Instruction::LShr:
672 return SelectBinaryOp(I, ISD::SRL);
673 case Instruction::AShr:
674 return SelectBinaryOp(I, ISD::SRA);
675 case Instruction::And:
676 return SelectBinaryOp(I, ISD::AND);
677 case Instruction::Or:
678 return SelectBinaryOp(I, ISD::OR);
679 case Instruction::Xor:
680 return SelectBinaryOp(I, ISD::XOR);
682 case Instruction::GetElementPtr:
683 return SelectGetElementPtr(I);
685 case Instruction::Br: {
686 BranchInst *BI = cast<BranchInst>(I);
688 if (BI->isUnconditional()) {
689 BasicBlock *LLVMSucc = BI->getSuccessor(0);
690 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
691 FastEmitBranch(MSucc);
695 // Conditional branches are not handed yet.
696 // Halt "fast" selection and bail.
700 case Instruction::Unreachable:
704 case Instruction::PHI:
705 // PHI nodes are already emitted.
708 case Instruction::Alloca:
709 // FunctionLowering has the static-sized case covered.
710 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
713 // Dynamic-sized alloca is not handled yet.
716 case Instruction::Call:
717 return SelectCall(I);
719 case Instruction::BitCast:
720 return SelectBitCast(I);
722 case Instruction::FPToSI:
723 return SelectCast(I, ISD::FP_TO_SINT);
724 case Instruction::ZExt:
725 return SelectCast(I, ISD::ZERO_EXTEND);
726 case Instruction::SExt:
727 return SelectCast(I, ISD::SIGN_EXTEND);
728 case Instruction::Trunc:
729 return SelectCast(I, ISD::TRUNCATE);
730 case Instruction::SIToFP:
731 return SelectCast(I, ISD::SINT_TO_FP);
733 case Instruction::IntToPtr: // Deliberate fall-through.
734 case Instruction::PtrToInt: {
735 MVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
736 MVT DstVT = TLI.getValueType(I->getType());
737 if (DstVT.bitsGT(SrcVT))
738 return SelectCast(I, ISD::ZERO_EXTEND);
739 if (DstVT.bitsLT(SrcVT))
740 return SelectCast(I, ISD::TRUNCATE);
741 unsigned Reg = getRegForValue(I->getOperand(0));
742 if (Reg == 0) return false;
743 UpdateValueMap(I, Reg);
748 // Unhandled instruction. Halt "fast" selection and bail.
753 FastISel::FastISel(MachineFunction &mf,
754 MachineModuleInfo *mmi,
756 DenseMap<const Value *, unsigned> &vm,
757 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
758 DenseMap<const AllocaInst *, int> &am
760 , SmallSet<Instruction*, 8> &cil
773 MRI(MF.getRegInfo()),
774 MFI(*MF.getFrameInfo()),
775 MCP(*MF.getConstantPool()),
777 TD(*TM.getTargetData()),
778 TII(*TM.getInstrInfo()),
779 TLI(*TM.getTargetLowering()) {
782 FastISel::~FastISel() {}
784 unsigned FastISel::FastEmit_(MVT::SimpleValueType, MVT::SimpleValueType,
789 unsigned FastISel::FastEmit_r(MVT::SimpleValueType, MVT::SimpleValueType,
790 ISD::NodeType, unsigned /*Op0*/) {
794 unsigned FastISel::FastEmit_rr(MVT::SimpleValueType, MVT::SimpleValueType,
795 ISD::NodeType, unsigned /*Op0*/,
800 unsigned FastISel::FastEmit_i(MVT::SimpleValueType, MVT::SimpleValueType,
801 ISD::NodeType, uint64_t /*Imm*/) {
805 unsigned FastISel::FastEmit_f(MVT::SimpleValueType, MVT::SimpleValueType,
806 ISD::NodeType, ConstantFP * /*FPImm*/) {
810 unsigned FastISel::FastEmit_ri(MVT::SimpleValueType, MVT::SimpleValueType,
811 ISD::NodeType, unsigned /*Op0*/,
816 unsigned FastISel::FastEmit_rf(MVT::SimpleValueType, MVT::SimpleValueType,
817 ISD::NodeType, unsigned /*Op0*/,
818 ConstantFP * /*FPImm*/) {
822 unsigned FastISel::FastEmit_rri(MVT::SimpleValueType, MVT::SimpleValueType,
824 unsigned /*Op0*/, unsigned /*Op1*/,
829 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
830 /// to emit an instruction with an immediate operand using FastEmit_ri.
831 /// If that fails, it materializes the immediate into a register and try
832 /// FastEmit_rr instead.
833 unsigned FastISel::FastEmit_ri_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
834 unsigned Op0, uint64_t Imm,
835 MVT::SimpleValueType ImmType) {
836 // First check if immediate type is legal. If not, we can't use the ri form.
837 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
840 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
841 if (MaterialReg == 0)
843 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
846 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
847 /// to emit an instruction with a floating-point immediate operand using
848 /// FastEmit_rf. If that fails, it materializes the immediate into a register
849 /// and try FastEmit_rr instead.
850 unsigned FastISel::FastEmit_rf_(MVT::SimpleValueType VT, ISD::NodeType Opcode,
851 unsigned Op0, ConstantFP *FPImm,
852 MVT::SimpleValueType ImmType) {
853 // First check if immediate type is legal. If not, we can't use the rf form.
854 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
858 // Materialize the constant in a register.
859 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
860 if (MaterialReg == 0) {
861 // If the target doesn't have a way to directly enter a floating-point
862 // value into a register, use an alternate approach.
863 // TODO: The current approach only supports floating-point constants
864 // that can be constructed by conversion from integer values. This should
865 // be replaced by code that creates a load from a constant-pool entry,
866 // which will require some target-specific work.
867 const APFloat &Flt = FPImm->getValueAPF();
868 MVT IntVT = TLI.getPointerTy();
871 uint32_t IntBitWidth = IntVT.getSizeInBits();
873 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
874 APFloat::rmTowardZero, &isExact);
877 APInt IntVal(IntBitWidth, 2, x);
879 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
880 ISD::Constant, IntVal.getZExtValue());
883 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
884 ISD::SINT_TO_FP, IntegerReg);
885 if (MaterialReg == 0)
888 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
891 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
892 return MRI.createVirtualRegister(RC);
895 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
896 const TargetRegisterClass* RC) {
897 unsigned ResultReg = createResultReg(RC);
898 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
900 BuildMI(MBB, DL, II, ResultReg);
904 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
905 const TargetRegisterClass *RC,
907 unsigned ResultReg = createResultReg(RC);
908 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
910 if (II.getNumDefs() >= 1)
911 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
913 BuildMI(MBB, DL, II).addReg(Op0);
914 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
915 II.ImplicitDefs[0], RC, RC);
923 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
924 const TargetRegisterClass *RC,
925 unsigned Op0, unsigned Op1) {
926 unsigned ResultReg = createResultReg(RC);
927 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
929 if (II.getNumDefs() >= 1)
930 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
932 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
933 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
934 II.ImplicitDefs[0], RC, RC);
941 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
942 const TargetRegisterClass *RC,
943 unsigned Op0, uint64_t Imm) {
944 unsigned ResultReg = createResultReg(RC);
945 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
947 if (II.getNumDefs() >= 1)
948 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
950 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
951 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
952 II.ImplicitDefs[0], RC, RC);
959 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
960 const TargetRegisterClass *RC,
961 unsigned Op0, ConstantFP *FPImm) {
962 unsigned ResultReg = createResultReg(RC);
963 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
965 if (II.getNumDefs() >= 1)
966 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
968 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
969 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
970 II.ImplicitDefs[0], RC, RC);
977 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
978 const TargetRegisterClass *RC,
979 unsigned Op0, unsigned Op1, uint64_t Imm) {
980 unsigned ResultReg = createResultReg(RC);
981 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
983 if (II.getNumDefs() >= 1)
984 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
986 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
987 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
988 II.ImplicitDefs[0], RC, RC);
995 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
996 const TargetRegisterClass *RC,
998 unsigned ResultReg = createResultReg(RC);
999 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
1001 if (II.getNumDefs() >= 1)
1002 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
1004 BuildMI(MBB, DL, II).addImm(Imm);
1005 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1006 II.ImplicitDefs[0], RC, RC);
1013 unsigned FastISel::FastEmitInst_extractsubreg(MVT::SimpleValueType RetVT,
1014 unsigned Op0, uint32_t Idx) {
1015 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
1017 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
1018 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
1020 if (II.getNumDefs() >= 1)
1021 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
1023 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
1024 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
1025 II.ImplicitDefs[0], RC, RC);
1032 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
1033 /// with all but the least significant bit set to zero.
1034 unsigned FastISel::FastEmitZExtFromI1(MVT::SimpleValueType VT, unsigned Op) {
1035 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);