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/DwarfWriter.h"
51 #include "llvm/Analysis/DebugInfo.h"
52 #include "llvm/Target/TargetData.h"
53 #include "llvm/Target/TargetInstrInfo.h"
54 #include "llvm/Target/TargetLowering.h"
55 #include "llvm/Target/TargetMachine.h"
56 #include "SelectionDAGBuilder.h"
57 #include "FunctionLoweringInfo.h"
60 unsigned FastISel::getRegForValue(Value *V) {
61 EVT 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 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(V->getContext(), 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.
97 getRegForValue(Constant::getNullValue(TD.getIntPtrType(V->getContext())));
98 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) {
99 Reg = FastEmit_f(VT, VT, ISD::ConstantFP, CF);
102 const APFloat &Flt = CF->getValueAPF();
103 EVT IntVT = TLI.getPointerTy();
106 uint32_t IntBitWidth = IntVT.getSizeInBits();
108 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
109 APFloat::rmTowardZero, &isExact);
111 APInt IntVal(IntBitWidth, 2, x);
113 unsigned IntegerReg =
114 getRegForValue(ConstantInt::get(V->getContext(), IntVal));
116 Reg = FastEmit_r(IntVT.getSimpleVT(), VT, ISD::SINT_TO_FP, IntegerReg);
119 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
120 if (!SelectOperator(CE, CE->getOpcode())) return 0;
121 Reg = LocalValueMap[CE];
122 } else if (isa<UndefValue>(V)) {
123 Reg = createResultReg(TLI.getRegClassFor(VT));
124 BuildMI(MBB, DL, TII.get(TargetInstrInfo::IMPLICIT_DEF), Reg);
127 // If target-independent code couldn't handle the value, give target-specific
129 if (!Reg && isa<Constant>(V))
130 Reg = TargetMaterializeConstant(cast<Constant>(V));
132 // Don't cache constant materializations in the general ValueMap.
133 // To do so would require tracking what uses they dominate.
135 LocalValueMap[V] = Reg;
139 unsigned FastISel::lookUpRegForValue(Value *V) {
140 // Look up the value to see if we already have a register for it. We
141 // cache values defined by Instructions across blocks, and other values
142 // only locally. This is because Instructions already have the SSA
143 // def-dominatess-use requirement enforced.
144 if (ValueMap.count(V))
146 return LocalValueMap[V];
149 /// UpdateValueMap - Update the value map to include the new mapping for this
150 /// instruction, or insert an extra copy to get the result in a previous
151 /// determined register.
152 /// NOTE: This is only necessary because we might select a block that uses
153 /// a value before we select the block that defines the value. It might be
154 /// possible to fix this by selecting blocks in reverse postorder.
155 unsigned FastISel::UpdateValueMap(Value* I, unsigned Reg) {
156 if (!isa<Instruction>(I)) {
157 LocalValueMap[I] = Reg;
161 unsigned &AssignedReg = ValueMap[I];
162 if (AssignedReg == 0)
164 else if (Reg != AssignedReg) {
165 const TargetRegisterClass *RegClass = MRI.getRegClass(Reg);
166 TII.copyRegToReg(*MBB, MBB->end(), AssignedReg,
167 Reg, RegClass, RegClass);
172 unsigned FastISel::getRegForGEPIndex(Value *Idx) {
173 unsigned IdxN = getRegForValue(Idx);
175 // Unhandled operand. Halt "fast" selection and bail.
178 // If the index is smaller or larger than intptr_t, truncate or extend it.
179 MVT PtrVT = TLI.getPointerTy();
180 EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
181 if (IdxVT.bitsLT(PtrVT))
182 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, ISD::SIGN_EXTEND, IdxN);
183 else if (IdxVT.bitsGT(PtrVT))
184 IdxN = FastEmit_r(IdxVT.getSimpleVT(), PtrVT, 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 EVT VT = EVT::getEVT(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(I->getContext(), 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 VT = TLI.getPointerTy();
265 for (GetElementPtrInst::op_iterator OI = I->op_begin()+1, E = I->op_end();
268 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
269 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
272 uint64_t Offs = TD.getStructLayout(StTy)->getElementOffset(Field);
273 // FIXME: This can be optimized by combining the add with a
275 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
277 // Unhandled operand. Halt "fast" selection and bail.
280 Ty = StTy->getElementType(Field);
282 Ty = cast<SequentialType>(Ty)->getElementType();
284 // If this is a constant subscript, handle it quickly.
285 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
286 if (CI->getZExtValue() == 0) continue;
288 TD.getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
289 N = FastEmit_ri_(VT, ISD::ADD, N, Offs, VT);
291 // Unhandled operand. Halt "fast" selection and bail.
296 // N = N + Idx * ElementSize;
297 uint64_t ElementSize = TD.getTypeAllocSize(Ty);
298 unsigned IdxN = getRegForGEPIndex(Idx);
300 // Unhandled operand. Halt "fast" selection and bail.
303 if (ElementSize != 1) {
304 IdxN = FastEmit_ri_(VT, ISD::MUL, IdxN, ElementSize, VT);
306 // Unhandled operand. Halt "fast" selection and bail.
309 N = FastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
311 // Unhandled operand. Halt "fast" selection and bail.
316 // We successfully emitted code for the given LLVM Instruction.
317 UpdateValueMap(I, N);
321 bool FastISel::SelectCall(User *I) {
322 Function *F = cast<CallInst>(I)->getCalledFunction();
323 if (!F) return false;
325 unsigned IID = F->getIntrinsicID();
328 case Intrinsic::dbg_stoppoint:
329 case Intrinsic::dbg_region_start:
330 case Intrinsic::dbg_region_end:
331 case Intrinsic::dbg_func_start:
332 // FIXME - Remove this instructions once the dust settles.
334 case Intrinsic::dbg_declare: {
335 DbgDeclareInst *DI = cast<DbgDeclareInst>(I);
336 if (!DIDescriptor::ValidDebugInfo(DI->getVariable(), CodeGenOpt::None)||!DW
337 || !DW->ShouldEmitDwarfDebug())
340 Value *Address = DI->getAddress();
341 if (BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
342 Address = BCI->getOperand(0);
343 AllocaInst *AI = dyn_cast<AllocaInst>(Address);
344 // Don't handle byval struct arguments or VLAs, for example.
346 DenseMap<const AllocaInst*, int>::iterator SI =
347 StaticAllocaMap.find(AI);
348 if (SI == StaticAllocaMap.end()) break; // VLAs.
351 if (MDNode *Dbg = DI->getMetadata("dbg"))
352 MMI->setVariableDbgInfo(DI->getVariable(), FI, Dbg);
356 case Intrinsic::eh_exception: {
357 EVT VT = TLI.getValueType(I->getType());
358 switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) {
360 case TargetLowering::Expand: {
361 assert(MBB->isLandingPad() && "Call to eh.exception not in landing pad!");
362 unsigned Reg = TLI.getExceptionAddressRegister();
363 const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
364 unsigned ResultReg = createResultReg(RC);
365 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
367 assert(InsertedCopy && "Can't copy address registers!");
368 InsertedCopy = InsertedCopy;
369 UpdateValueMap(I, ResultReg);
375 case Intrinsic::eh_selector: {
376 EVT VT = TLI.getValueType(I->getType());
377 switch (TLI.getOperationAction(ISD::EHSELECTION, VT)) {
379 case TargetLowering::Expand: {
381 if (MBB->isLandingPad())
382 AddCatchInfo(*cast<CallInst>(I), MMI, MBB);
385 CatchInfoLost.insert(cast<CallInst>(I));
387 // FIXME: Mark exception selector register as live in. Hack for PR1508.
388 unsigned Reg = TLI.getExceptionSelectorRegister();
389 if (Reg) MBB->addLiveIn(Reg);
392 unsigned Reg = TLI.getExceptionSelectorRegister();
393 EVT SrcVT = TLI.getPointerTy();
394 const TargetRegisterClass *RC = TLI.getRegClassFor(SrcVT);
395 unsigned ResultReg = createResultReg(RC);
396 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg, Reg,
398 assert(InsertedCopy && "Can't copy address registers!");
399 InsertedCopy = InsertedCopy;
401 // Cast the register to the type of the selector.
402 if (SrcVT.bitsGT(MVT::i32))
403 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32, ISD::TRUNCATE,
405 else if (SrcVT.bitsLT(MVT::i32))
406 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), MVT::i32,
407 ISD::SIGN_EXTEND, ResultReg);
409 // Unhandled operand. Halt "fast" selection and bail.
412 UpdateValueMap(I, ResultReg);
415 getRegForValue(Constant::getNullValue(I->getType()));
416 UpdateValueMap(I, ResultReg);
427 bool FastISel::SelectCast(User *I, ISD::NodeType Opcode) {
428 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
429 EVT DstVT = TLI.getValueType(I->getType());
431 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
432 DstVT == MVT::Other || !DstVT.isSimple())
433 // Unhandled type. Halt "fast" selection and bail.
436 // Check if the destination type is legal. Or as a special case,
437 // it may be i1 if we're doing a truncate because that's
438 // easy and somewhat common.
439 if (!TLI.isTypeLegal(DstVT))
440 if (DstVT != MVT::i1 || Opcode != ISD::TRUNCATE)
441 // Unhandled type. Halt "fast" selection and bail.
444 // Check if the source operand is legal. Or as a special case,
445 // it may be i1 if we're doing zero-extension because that's
446 // easy and somewhat common.
447 if (!TLI.isTypeLegal(SrcVT))
448 if (SrcVT != MVT::i1 || Opcode != ISD::ZERO_EXTEND)
449 // Unhandled type. Halt "fast" selection and bail.
452 unsigned InputReg = getRegForValue(I->getOperand(0));
454 // Unhandled operand. Halt "fast" selection and bail.
457 // If the operand is i1, arrange for the high bits in the register to be zero.
458 if (SrcVT == MVT::i1) {
459 SrcVT = TLI.getTypeToTransformTo(I->getContext(), SrcVT);
460 InputReg = FastEmitZExtFromI1(SrcVT.getSimpleVT(), InputReg);
464 // If the result is i1, truncate to the target's type for i1 first.
465 if (DstVT == MVT::i1)
466 DstVT = TLI.getTypeToTransformTo(I->getContext(), DstVT);
468 unsigned ResultReg = FastEmit_r(SrcVT.getSimpleVT(),
475 UpdateValueMap(I, ResultReg);
479 bool FastISel::SelectBitCast(User *I) {
480 // If the bitcast doesn't change the type, just use the operand value.
481 if (I->getType() == I->getOperand(0)->getType()) {
482 unsigned Reg = getRegForValue(I->getOperand(0));
485 UpdateValueMap(I, Reg);
489 // Bitcasts of other values become reg-reg copies or BIT_CONVERT operators.
490 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
491 EVT DstVT = TLI.getValueType(I->getType());
493 if (SrcVT == MVT::Other || !SrcVT.isSimple() ||
494 DstVT == MVT::Other || !DstVT.isSimple() ||
495 !TLI.isTypeLegal(SrcVT) || !TLI.isTypeLegal(DstVT))
496 // Unhandled type. Halt "fast" selection and bail.
499 unsigned Op0 = getRegForValue(I->getOperand(0));
501 // Unhandled operand. Halt "fast" selection and bail.
504 // First, try to perform the bitcast by inserting a reg-reg copy.
505 unsigned ResultReg = 0;
506 if (SrcVT.getSimpleVT() == DstVT.getSimpleVT()) {
507 TargetRegisterClass* SrcClass = TLI.getRegClassFor(SrcVT);
508 TargetRegisterClass* DstClass = TLI.getRegClassFor(DstVT);
509 ResultReg = createResultReg(DstClass);
511 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
512 Op0, DstClass, SrcClass);
517 // If the reg-reg copy failed, select a BIT_CONVERT opcode.
519 ResultReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(),
520 ISD::BIT_CONVERT, Op0);
525 UpdateValueMap(I, ResultReg);
530 FastISel::SelectInstruction(Instruction *I) {
531 // First, try doing target-independent selection.
532 if (SelectOperator(I, I->getOpcode()))
535 // Next, try calling the target to attempt to handle the instruction.
536 if (TargetSelectInstruction(I))
542 /// FastEmitBranch - Emit an unconditional branch to the given block,
543 /// unless it is the immediate (fall-through) successor, and update
546 FastISel::FastEmitBranch(MachineBasicBlock *MSucc) {
547 if (MBB->isLayoutSuccessor(MSucc)) {
548 // The unconditional fall-through case, which needs no instructions.
550 // The unconditional branch case.
551 TII.InsertBranch(*MBB, MSucc, NULL, SmallVector<MachineOperand, 0>());
553 MBB->addSuccessor(MSucc);
556 /// SelectFNeg - Emit an FNeg operation.
559 FastISel::SelectFNeg(User *I) {
560 unsigned OpReg = getRegForValue(BinaryOperator::getFNegArgument(I));
561 if (OpReg == 0) return false;
563 // If the target has ISD::FNEG, use it.
564 EVT VT = TLI.getValueType(I->getType());
565 unsigned ResultReg = FastEmit_r(VT.getSimpleVT(), VT.getSimpleVT(),
567 if (ResultReg != 0) {
568 UpdateValueMap(I, ResultReg);
572 // Bitcast the value to integer, twiddle the sign bit with xor,
573 // and then bitcast it back to floating-point.
574 if (VT.getSizeInBits() > 64) return false;
575 EVT IntVT = EVT::getIntegerVT(I->getContext(), VT.getSizeInBits());
576 if (!TLI.isTypeLegal(IntVT))
579 unsigned IntReg = FastEmit_r(VT.getSimpleVT(), IntVT.getSimpleVT(),
580 ISD::BIT_CONVERT, OpReg);
584 unsigned IntResultReg = FastEmit_ri_(IntVT.getSimpleVT(), ISD::XOR, IntReg,
585 UINT64_C(1) << (VT.getSizeInBits()-1),
586 IntVT.getSimpleVT());
587 if (IntResultReg == 0)
590 ResultReg = FastEmit_r(IntVT.getSimpleVT(), VT.getSimpleVT(),
591 ISD::BIT_CONVERT, IntResultReg);
595 UpdateValueMap(I, ResultReg);
600 FastISel::SelectOperator(User *I, unsigned Opcode) {
602 case Instruction::Add:
603 return SelectBinaryOp(I, ISD::ADD);
604 case Instruction::FAdd:
605 return SelectBinaryOp(I, ISD::FADD);
606 case Instruction::Sub:
607 return SelectBinaryOp(I, ISD::SUB);
608 case Instruction::FSub:
609 // FNeg is currently represented in LLVM IR as a special case of FSub.
610 if (BinaryOperator::isFNeg(I))
611 return SelectFNeg(I);
612 return SelectBinaryOp(I, ISD::FSUB);
613 case Instruction::Mul:
614 return SelectBinaryOp(I, ISD::MUL);
615 case Instruction::FMul:
616 return SelectBinaryOp(I, ISD::FMUL);
617 case Instruction::SDiv:
618 return SelectBinaryOp(I, ISD::SDIV);
619 case Instruction::UDiv:
620 return SelectBinaryOp(I, ISD::UDIV);
621 case Instruction::FDiv:
622 return SelectBinaryOp(I, ISD::FDIV);
623 case Instruction::SRem:
624 return SelectBinaryOp(I, ISD::SREM);
625 case Instruction::URem:
626 return SelectBinaryOp(I, ISD::UREM);
627 case Instruction::FRem:
628 return SelectBinaryOp(I, ISD::FREM);
629 case Instruction::Shl:
630 return SelectBinaryOp(I, ISD::SHL);
631 case Instruction::LShr:
632 return SelectBinaryOp(I, ISD::SRL);
633 case Instruction::AShr:
634 return SelectBinaryOp(I, ISD::SRA);
635 case Instruction::And:
636 return SelectBinaryOp(I, ISD::AND);
637 case Instruction::Or:
638 return SelectBinaryOp(I, ISD::OR);
639 case Instruction::Xor:
640 return SelectBinaryOp(I, ISD::XOR);
642 case Instruction::GetElementPtr:
643 return SelectGetElementPtr(I);
645 case Instruction::Br: {
646 BranchInst *BI = cast<BranchInst>(I);
648 if (BI->isUnconditional()) {
649 BasicBlock *LLVMSucc = BI->getSuccessor(0);
650 MachineBasicBlock *MSucc = MBBMap[LLVMSucc];
651 FastEmitBranch(MSucc);
655 // Conditional branches are not handed yet.
656 // Halt "fast" selection and bail.
660 case Instruction::Unreachable:
664 case Instruction::PHI:
665 // PHI nodes are already emitted.
668 case Instruction::Alloca:
669 // FunctionLowering has the static-sized case covered.
670 if (StaticAllocaMap.count(cast<AllocaInst>(I)))
673 // Dynamic-sized alloca is not handled yet.
676 case Instruction::Call:
677 return SelectCall(I);
679 case Instruction::BitCast:
680 return SelectBitCast(I);
682 case Instruction::FPToSI:
683 return SelectCast(I, ISD::FP_TO_SINT);
684 case Instruction::ZExt:
685 return SelectCast(I, ISD::ZERO_EXTEND);
686 case Instruction::SExt:
687 return SelectCast(I, ISD::SIGN_EXTEND);
688 case Instruction::Trunc:
689 return SelectCast(I, ISD::TRUNCATE);
690 case Instruction::SIToFP:
691 return SelectCast(I, ISD::SINT_TO_FP);
693 case Instruction::IntToPtr: // Deliberate fall-through.
694 case Instruction::PtrToInt: {
695 EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
696 EVT DstVT = TLI.getValueType(I->getType());
697 if (DstVT.bitsGT(SrcVT))
698 return SelectCast(I, ISD::ZERO_EXTEND);
699 if (DstVT.bitsLT(SrcVT))
700 return SelectCast(I, ISD::TRUNCATE);
701 unsigned Reg = getRegForValue(I->getOperand(0));
702 if (Reg == 0) return false;
703 UpdateValueMap(I, Reg);
708 // Unhandled instruction. Halt "fast" selection and bail.
713 FastISel::FastISel(MachineFunction &mf,
714 MachineModuleInfo *mmi,
716 DenseMap<const Value *, unsigned> &vm,
717 DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
718 DenseMap<const AllocaInst *, int> &am
720 , SmallSet<Instruction*, 8> &cil
733 MRI(MF.getRegInfo()),
734 MFI(*MF.getFrameInfo()),
735 MCP(*MF.getConstantPool()),
737 TD(*TM.getTargetData()),
738 TII(*TM.getInstrInfo()),
739 TLI(*TM.getTargetLowering()) {
742 FastISel::~FastISel() {}
744 unsigned FastISel::FastEmit_(MVT, MVT,
749 unsigned FastISel::FastEmit_r(MVT, MVT,
750 ISD::NodeType, unsigned /*Op0*/) {
754 unsigned FastISel::FastEmit_rr(MVT, MVT,
755 ISD::NodeType, unsigned /*Op0*/,
760 unsigned FastISel::FastEmit_i(MVT, MVT, ISD::NodeType, uint64_t /*Imm*/) {
764 unsigned FastISel::FastEmit_f(MVT, MVT,
765 ISD::NodeType, ConstantFP * /*FPImm*/) {
769 unsigned FastISel::FastEmit_ri(MVT, MVT,
770 ISD::NodeType, unsigned /*Op0*/,
775 unsigned FastISel::FastEmit_rf(MVT, MVT,
776 ISD::NodeType, unsigned /*Op0*/,
777 ConstantFP * /*FPImm*/) {
781 unsigned FastISel::FastEmit_rri(MVT, MVT,
783 unsigned /*Op0*/, unsigned /*Op1*/,
788 /// FastEmit_ri_ - This method is a wrapper of FastEmit_ri. It first tries
789 /// to emit an instruction with an immediate operand using FastEmit_ri.
790 /// If that fails, it materializes the immediate into a register and try
791 /// FastEmit_rr instead.
792 unsigned FastISel::FastEmit_ri_(MVT VT, ISD::NodeType Opcode,
793 unsigned Op0, uint64_t Imm,
795 // First check if immediate type is legal. If not, we can't use the ri form.
796 unsigned ResultReg = FastEmit_ri(VT, VT, Opcode, Op0, Imm);
799 unsigned MaterialReg = FastEmit_i(ImmType, ImmType, ISD::Constant, Imm);
800 if (MaterialReg == 0)
802 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
805 /// FastEmit_rf_ - This method is a wrapper of FastEmit_ri. It first tries
806 /// to emit an instruction with a floating-point immediate operand using
807 /// FastEmit_rf. If that fails, it materializes the immediate into a register
808 /// and try FastEmit_rr instead.
809 unsigned FastISel::FastEmit_rf_(MVT VT, ISD::NodeType Opcode,
810 unsigned Op0, ConstantFP *FPImm,
812 // First check if immediate type is legal. If not, we can't use the rf form.
813 unsigned ResultReg = FastEmit_rf(VT, VT, Opcode, Op0, FPImm);
817 // Materialize the constant in a register.
818 unsigned MaterialReg = FastEmit_f(ImmType, ImmType, ISD::ConstantFP, FPImm);
819 if (MaterialReg == 0) {
820 // If the target doesn't have a way to directly enter a floating-point
821 // value into a register, use an alternate approach.
822 // TODO: The current approach only supports floating-point constants
823 // that can be constructed by conversion from integer values. This should
824 // be replaced by code that creates a load from a constant-pool entry,
825 // which will require some target-specific work.
826 const APFloat &Flt = FPImm->getValueAPF();
827 EVT IntVT = TLI.getPointerTy();
830 uint32_t IntBitWidth = IntVT.getSizeInBits();
832 (void) Flt.convertToInteger(x, IntBitWidth, /*isSigned=*/true,
833 APFloat::rmTowardZero, &isExact);
836 APInt IntVal(IntBitWidth, 2, x);
838 unsigned IntegerReg = FastEmit_i(IntVT.getSimpleVT(), IntVT.getSimpleVT(),
839 ISD::Constant, IntVal.getZExtValue());
842 MaterialReg = FastEmit_r(IntVT.getSimpleVT(), VT,
843 ISD::SINT_TO_FP, IntegerReg);
844 if (MaterialReg == 0)
847 return FastEmit_rr(VT, VT, Opcode, Op0, MaterialReg);
850 unsigned FastISel::createResultReg(const TargetRegisterClass* RC) {
851 return MRI.createVirtualRegister(RC);
854 unsigned FastISel::FastEmitInst_(unsigned MachineInstOpcode,
855 const TargetRegisterClass* RC) {
856 unsigned ResultReg = createResultReg(RC);
857 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
859 BuildMI(MBB, DL, II, ResultReg);
863 unsigned FastISel::FastEmitInst_r(unsigned MachineInstOpcode,
864 const TargetRegisterClass *RC,
866 unsigned ResultReg = createResultReg(RC);
867 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
869 if (II.getNumDefs() >= 1)
870 BuildMI(MBB, DL, II, ResultReg).addReg(Op0);
872 BuildMI(MBB, DL, II).addReg(Op0);
873 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
874 II.ImplicitDefs[0], RC, RC);
882 unsigned FastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
883 const TargetRegisterClass *RC,
884 unsigned Op0, unsigned Op1) {
885 unsigned ResultReg = createResultReg(RC);
886 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
888 if (II.getNumDefs() >= 1)
889 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1);
891 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1);
892 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
893 II.ImplicitDefs[0], RC, RC);
900 unsigned FastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
901 const TargetRegisterClass *RC,
902 unsigned Op0, uint64_t Imm) {
903 unsigned ResultReg = createResultReg(RC);
904 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
906 if (II.getNumDefs() >= 1)
907 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Imm);
909 BuildMI(MBB, DL, II).addReg(Op0).addImm(Imm);
910 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
911 II.ImplicitDefs[0], RC, RC);
918 unsigned FastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
919 const TargetRegisterClass *RC,
920 unsigned Op0, ConstantFP *FPImm) {
921 unsigned ResultReg = createResultReg(RC);
922 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
924 if (II.getNumDefs() >= 1)
925 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addFPImm(FPImm);
927 BuildMI(MBB, DL, II).addReg(Op0).addFPImm(FPImm);
928 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
929 II.ImplicitDefs[0], RC, RC);
936 unsigned FastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
937 const TargetRegisterClass *RC,
938 unsigned Op0, unsigned Op1, uint64_t Imm) {
939 unsigned ResultReg = createResultReg(RC);
940 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
942 if (II.getNumDefs() >= 1)
943 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addReg(Op1).addImm(Imm);
945 BuildMI(MBB, DL, II).addReg(Op0).addReg(Op1).addImm(Imm);
946 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
947 II.ImplicitDefs[0], RC, RC);
954 unsigned FastISel::FastEmitInst_i(unsigned MachineInstOpcode,
955 const TargetRegisterClass *RC,
957 unsigned ResultReg = createResultReg(RC);
958 const TargetInstrDesc &II = TII.get(MachineInstOpcode);
960 if (II.getNumDefs() >= 1)
961 BuildMI(MBB, DL, II, ResultReg).addImm(Imm);
963 BuildMI(MBB, DL, II).addImm(Imm);
964 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
965 II.ImplicitDefs[0], RC, RC);
972 unsigned FastISel::FastEmitInst_extractsubreg(MVT RetVT,
973 unsigned Op0, uint32_t Idx) {
974 const TargetRegisterClass* RC = MRI.getRegClass(Op0);
976 unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
977 const TargetInstrDesc &II = TII.get(TargetInstrInfo::EXTRACT_SUBREG);
979 if (II.getNumDefs() >= 1)
980 BuildMI(MBB, DL, II, ResultReg).addReg(Op0).addImm(Idx);
982 BuildMI(MBB, DL, II).addReg(Op0).addImm(Idx);
983 bool InsertedCopy = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
984 II.ImplicitDefs[0], RC, RC);
991 /// FastEmitZExtFromI1 - Emit MachineInstrs to compute the value of Op
992 /// with all but the least significant bit set to zero.
993 unsigned FastISel::FastEmitZExtFromI1(MVT VT, unsigned Op) {
994 return FastEmit_ri(VT, VT, ISD::AND, Op, 1);