1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "llvm/Analysis/AliasAnalysis.h"
16 #include "llvm/CodeGen/SelectionDAGISel.h"
17 #include "llvm/CodeGen/ScheduleDAG.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/InlineAsm.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineDebugInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineJumpTableInfo.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/SchedulerRegistry.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/SSARegMap.h"
36 #include "llvm/Target/MRegisterInfo.h"
37 #include "llvm/Target/TargetData.h"
38 #include "llvm/Target/TargetFrameInfo.h"
39 #include "llvm/Target/TargetInstrInfo.h"
40 #include "llvm/Target/TargetLowering.h"
41 #include "llvm/Target/TargetMachine.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/Compiler.h"
55 ViewISelDAGs("view-isel-dags", cl::Hidden,
56 cl::desc("Pop up a window to show isel dags as they are selected"));
58 ViewSchedDAGs("view-sched-dags", cl::Hidden,
59 cl::desc("Pop up a window to show sched dags as they are processed"));
61 static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0;
65 //===---------------------------------------------------------------------===//
67 /// RegisterScheduler class - Track the registration of instruction schedulers.
69 //===---------------------------------------------------------------------===//
70 MachinePassRegistry RegisterScheduler::Registry;
72 //===---------------------------------------------------------------------===//
74 /// ISHeuristic command line option for instruction schedulers.
76 //===---------------------------------------------------------------------===//
78 cl::opt<RegisterScheduler::FunctionPassCtor, false,
79 RegisterPassParser<RegisterScheduler> >
81 cl::init(&createDefaultScheduler),
82 cl::desc("Instruction schedulers available:"));
84 static RegisterScheduler
85 defaultListDAGScheduler("default", " Best scheduler for the target",
86 createDefaultScheduler);
90 /// RegsForValue - This struct represents the physical registers that a
91 /// particular value is assigned and the type information about the value.
92 /// This is needed because values can be promoted into larger registers and
93 /// expanded into multiple smaller registers than the value.
94 struct VISIBILITY_HIDDEN RegsForValue {
95 /// Regs - This list hold the register (for legal and promoted values)
96 /// or register set (for expanded values) that the value should be assigned
98 std::vector<unsigned> Regs;
100 /// RegVT - The value type of each register.
102 MVT::ValueType RegVT;
104 /// ValueVT - The value type of the LLVM value, which may be promoted from
105 /// RegVT or made from merging the two expanded parts.
106 MVT::ValueType ValueVT;
108 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {}
110 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
111 : RegVT(regvt), ValueVT(valuevt) {
114 RegsForValue(const std::vector<unsigned> ®s,
115 MVT::ValueType regvt, MVT::ValueType valuevt)
116 : Regs(regs), RegVT(regvt), ValueVT(valuevt) {
119 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
120 /// this value and returns the result as a ValueVT value. This uses
121 /// Chain/Flag as the input and updates them for the output Chain/Flag.
122 SDOperand getCopyFromRegs(SelectionDAG &DAG,
123 SDOperand &Chain, SDOperand &Flag) const;
125 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
126 /// specified value into the registers specified by this object. This uses
127 /// Chain/Flag as the input and updates them for the output Chain/Flag.
128 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
129 SDOperand &Chain, SDOperand &Flag,
130 MVT::ValueType PtrVT) const;
132 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
133 /// operand list. This adds the code marker and includes the number of
134 /// values added into it.
135 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
136 std::vector<SDOperand> &Ops) const;
141 //===--------------------------------------------------------------------===//
142 /// createDefaultScheduler - This creates an instruction scheduler appropriate
144 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
146 MachineBasicBlock *BB) {
147 TargetLowering &TLI = IS->getTargetLowering();
149 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
150 return createTDListDAGScheduler(IS, DAG, BB);
152 assert(TLI.getSchedulingPreference() ==
153 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
154 return createBURRListDAGScheduler(IS, DAG, BB);
159 //===--------------------------------------------------------------------===//
160 /// FunctionLoweringInfo - This contains information that is global to a
161 /// function that is used when lowering a region of the function.
162 class FunctionLoweringInfo {
169 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF);
171 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry.
172 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
174 /// ValueMap - Since we emit code for the function a basic block at a time,
175 /// we must remember which virtual registers hold the values for
176 /// cross-basic-block values.
177 std::map<const Value*, unsigned> ValueMap;
179 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in
180 /// the entry block. This allows the allocas to be efficiently referenced
181 /// anywhere in the function.
182 std::map<const AllocaInst*, int> StaticAllocaMap;
184 unsigned MakeReg(MVT::ValueType VT) {
185 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
188 /// isExportedInst - Return true if the specified value is an instruction
189 /// exported from its block.
190 bool isExportedInst(const Value *V) {
191 return ValueMap.count(V);
194 unsigned CreateRegForValue(const Value *V);
196 unsigned InitializeRegForValue(const Value *V) {
197 unsigned &R = ValueMap[V];
198 assert(R == 0 && "Already initialized this value register!");
199 return R = CreateRegForValue(V);
204 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
205 /// PHI nodes or outside of the basic block that defines it, or used by a
206 /// switch instruction, which may expand to multiple basic blocks.
207 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
208 if (isa<PHINode>(I)) return true;
209 BasicBlock *BB = I->getParent();
210 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
211 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
212 // FIXME: Remove switchinst special case.
213 isa<SwitchInst>(*UI))
218 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
219 /// entry block, return true. This includes arguments used by switches, since
220 /// the switch may expand into multiple basic blocks.
221 static bool isOnlyUsedInEntryBlock(Argument *A) {
222 BasicBlock *Entry = A->getParent()->begin();
223 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
224 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
225 return false; // Use not in entry block.
229 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli,
230 Function &fn, MachineFunction &mf)
231 : TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) {
233 // Create a vreg for each argument register that is not dead and is used
234 // outside of the entry block for the function.
235 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end();
237 if (!isOnlyUsedInEntryBlock(AI))
238 InitializeRegForValue(AI);
240 // Initialize the mapping of values to registers. This is only set up for
241 // instruction values that are used outside of the block that defines
243 Function::iterator BB = Fn.begin(), EB = Fn.end();
244 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
245 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
246 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
247 const Type *Ty = AI->getAllocatedType();
248 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
250 std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
253 // If the alignment of the value is smaller than the size of the
254 // value, and if the size of the value is particularly small
255 // (<= 8 bytes), round up to the size of the value for potentially
256 // better performance.
258 // FIXME: This could be made better with a preferred alignment hook in
259 // TargetData. It serves primarily to 8-byte align doubles for X86.
260 if (Align < TySize && TySize <= 8) Align = TySize;
261 TySize *= CUI->getZExtValue(); // Get total allocated size.
262 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
263 StaticAllocaMap[AI] =
264 MF.getFrameInfo()->CreateStackObject((unsigned)TySize, Align);
267 for (; BB != EB; ++BB)
268 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
269 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
270 if (!isa<AllocaInst>(I) ||
271 !StaticAllocaMap.count(cast<AllocaInst>(I)))
272 InitializeRegForValue(I);
274 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
275 // also creates the initial PHI MachineInstrs, though none of the input
276 // operands are populated.
277 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) {
278 MachineBasicBlock *MBB = new MachineBasicBlock(BB);
280 MF.getBasicBlockList().push_back(MBB);
282 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
285 for (BasicBlock::iterator I = BB->begin();
286 (PN = dyn_cast<PHINode>(I)); ++I)
287 if (!PN->use_empty()) {
288 MVT::ValueType VT = TLI.getValueType(PN->getType());
289 unsigned NumElements;
290 if (VT != MVT::Vector)
291 NumElements = TLI.getNumElements(VT);
293 MVT::ValueType VT1,VT2;
295 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
298 unsigned PHIReg = ValueMap[PN];
299 assert(PHIReg &&"PHI node does not have an assigned virtual register!");
300 for (unsigned i = 0; i != NumElements; ++i)
301 BuildMI(MBB, TargetInstrInfo::PHI, PN->getNumOperands(), PHIReg+i);
306 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
307 /// the correctly promoted or expanded types. Assign these registers
308 /// consecutive vreg numbers and return the first assigned number.
309 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
310 MVT::ValueType VT = TLI.getValueType(V->getType());
312 // The number of multiples of registers that we need, to, e.g., split up
313 // a <2 x int64> -> 4 x i32 registers.
314 unsigned NumVectorRegs = 1;
316 // If this is a packed type, figure out what type it will decompose into
317 // and how many of the elements it will use.
318 if (VT == MVT::Vector) {
319 const PackedType *PTy = cast<PackedType>(V->getType());
320 unsigned NumElts = PTy->getNumElements();
321 MVT::ValueType EltTy = TLI.getValueType(PTy->getElementType());
323 // Divide the input until we get to a supported size. This will always
324 // end with a scalar if the target doesn't support vectors.
325 while (NumElts > 1 && !TLI.isTypeLegal(getVectorType(EltTy, NumElts))) {
332 VT = getVectorType(EltTy, NumElts);
335 // The common case is that we will only create one register for this
336 // value. If we have that case, create and return the virtual register.
337 unsigned NV = TLI.getNumElements(VT);
339 // If we are promoting this value, pick the next largest supported type.
340 MVT::ValueType PromotedType = TLI.getTypeToTransformTo(VT);
341 unsigned Reg = MakeReg(PromotedType);
342 // If this is a vector of supported or promoted types (e.g. 4 x i16),
343 // create all of the registers.
344 for (unsigned i = 1; i != NumVectorRegs; ++i)
345 MakeReg(PromotedType);
349 // If this value is represented with multiple target registers, make sure
350 // to create enough consecutive registers of the right (smaller) type.
351 unsigned NT = VT-1; // Find the type to use.
352 while (TLI.getNumElements((MVT::ValueType)NT) != 1)
355 unsigned R = MakeReg((MVT::ValueType)NT);
356 for (unsigned i = 1; i != NV*NumVectorRegs; ++i)
357 MakeReg((MVT::ValueType)NT);
361 //===----------------------------------------------------------------------===//
362 /// SelectionDAGLowering - This is the common target-independent lowering
363 /// implementation that is parameterized by a TargetLowering object.
364 /// Also, targets can overload any lowering method.
367 class SelectionDAGLowering {
368 MachineBasicBlock *CurMBB;
370 std::map<const Value*, SDOperand> NodeMap;
372 /// PendingLoads - Loads are not emitted to the program immediately. We bunch
373 /// them up and then emit token factor nodes when possible. This allows us to
374 /// get simple disambiguation between loads without worrying about alias
376 std::vector<SDOperand> PendingLoads;
378 /// Case - A pair of values to record the Value for a switch case, and the
379 /// case's target basic block.
380 typedef std::pair<Constant*, MachineBasicBlock*> Case;
381 typedef std::vector<Case>::iterator CaseItr;
382 typedef std::pair<CaseItr, CaseItr> CaseRange;
384 /// CaseRec - A struct with ctor used in lowering switches to a binary tree
385 /// of conditional branches.
387 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) :
388 CaseBB(bb), LT(lt), GE(ge), Range(r) {}
390 /// CaseBB - The MBB in which to emit the compare and branch
391 MachineBasicBlock *CaseBB;
392 /// LT, GE - If nonzero, we know the current case value must be less-than or
393 /// greater-than-or-equal-to these Constants.
396 /// Range - A pair of iterators representing the range of case values to be
397 /// processed at this point in the binary search tree.
401 /// The comparison function for sorting Case values.
403 bool operator () (const Case& C1, const Case& C2) {
404 if (const ConstantInt* I1 = dyn_cast<const ConstantInt>(C1.first))
405 if (I1->getType()->isUnsigned())
406 return I1->getZExtValue() <
407 cast<const ConstantInt>(C2.first)->getZExtValue();
409 return cast<const ConstantInt>(C1.first)->getSExtValue() <
410 cast<const ConstantInt>(C2.first)->getSExtValue();
415 // TLI - This is information that describes the available target features we
416 // need for lowering. This indicates when operations are unavailable,
417 // implemented with a libcall, etc.
420 const TargetData *TD;
422 /// SwitchCases - Vector of CaseBlock structures used to communicate
423 /// SwitchInst code generation information.
424 std::vector<SelectionDAGISel::CaseBlock> SwitchCases;
425 SelectionDAGISel::JumpTable JT;
427 /// FuncInfo - Information about the function as a whole.
429 FunctionLoweringInfo &FuncInfo;
431 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli,
432 FunctionLoweringInfo &funcinfo)
433 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()),
434 JT(0,0,0,0), FuncInfo(funcinfo) {
437 /// getRoot - Return the current virtual root of the Selection DAG.
439 SDOperand getRoot() {
440 if (PendingLoads.empty())
441 return DAG.getRoot();
443 if (PendingLoads.size() == 1) {
444 SDOperand Root = PendingLoads[0];
446 PendingLoads.clear();
450 // Otherwise, we have to make a token factor node.
451 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
452 &PendingLoads[0], PendingLoads.size());
453 PendingLoads.clear();
458 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg);
460 void visit(Instruction &I) { visit(I.getOpcode(), I); }
462 void visit(unsigned Opcode, User &I) {
464 default: assert(0 && "Unknown instruction type encountered!");
466 // Build the switch statement using the Instruction.def file.
467 #define HANDLE_INST(NUM, OPCODE, CLASS) \
468 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
469 #include "llvm/Instruction.def"
473 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; }
475 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr,
476 const Value *SV, SDOperand Root,
479 SDOperand getIntPtrConstant(uint64_t Val) {
480 return DAG.getConstant(Val, TLI.getPointerTy());
483 SDOperand getValue(const Value *V);
485 const SDOperand &setValue(const Value *V, SDOperand NewN) {
486 SDOperand &N = NodeMap[V];
487 assert(N.Val == 0 && "Already set a value for this node!");
491 RegsForValue GetRegistersForValue(const std::string &ConstrCode,
493 bool OutReg, bool InReg,
494 std::set<unsigned> &OutputRegs,
495 std::set<unsigned> &InputRegs);
497 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB,
498 MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
500 void ExportFromCurrentBlock(Value *V);
502 // Terminator instructions.
503 void visitRet(ReturnInst &I);
504 void visitBr(BranchInst &I);
505 void visitSwitch(SwitchInst &I);
506 void visitUnreachable(UnreachableInst &I) { /* noop */ }
508 // Helper for visitSwitch
509 void visitSwitchCase(SelectionDAGISel::CaseBlock &CB);
510 void visitJumpTable(SelectionDAGISel::JumpTable &JT);
512 // These all get lowered before this pass.
513 void visitInvoke(InvokeInst &I) { assert(0 && "TODO"); }
514 void visitUnwind(UnwindInst &I) { assert(0 && "TODO"); }
516 void visitIntBinary(User &I, unsigned IntOp, unsigned VecOp);
517 void visitFPBinary(User &I, unsigned FPOp, unsigned VecOp);
518 void visitShift(User &I, unsigned Opcode);
519 void visitAdd(User &I) {
520 if (I.getType()->isFloatingPoint())
521 visitFPBinary(I, ISD::FADD, ISD::VADD);
523 visitIntBinary(I, ISD::ADD, ISD::VADD);
525 void visitSub(User &I);
526 void visitMul(User &I) {
527 if (I.getType()->isFloatingPoint())
528 visitFPBinary(I, ISD::FMUL, ISD::VMUL);
530 visitIntBinary(I, ISD::MUL, ISD::VMUL);
532 void visitUDiv(User &I) { visitIntBinary(I, ISD::UDIV, ISD::VUDIV); }
533 void visitSDiv(User &I) { visitIntBinary(I, ISD::SDIV, ISD::VSDIV); }
534 void visitFDiv(User &I) { visitFPBinary(I, ISD::FDIV, ISD::VSDIV); }
535 void visitRem(User &I) {
536 const Type *Ty = I.getType();
537 if (Ty->isFloatingPoint())
538 visitFPBinary(I, ISD::FREM, 0);
540 visitIntBinary(I, Ty->isSigned() ? ISD::SREM : ISD::UREM, 0);
542 void visitAnd(User &I) { visitIntBinary(I, ISD::AND, ISD::VAND); }
543 void visitOr (User &I) { visitIntBinary(I, ISD::OR, ISD::VOR); }
544 void visitXor(User &I) { visitIntBinary(I, ISD::XOR, ISD::VXOR); }
545 void visitShl(User &I) { visitShift(I, ISD::SHL); }
546 void visitShr(User &I) {
547 visitShift(I, I.getType()->isUnsigned() ? ISD::SRL : ISD::SRA);
550 void visitSetCC(User &I, ISD::CondCode SignedOpc, ISD::CondCode UnsignedOpc,
551 ISD::CondCode FPOpc);
552 void visitSetEQ(User &I) { visitSetCC(I, ISD::SETEQ, ISD::SETEQ,
554 void visitSetNE(User &I) { visitSetCC(I, ISD::SETNE, ISD::SETNE,
556 void visitSetLE(User &I) { visitSetCC(I, ISD::SETLE, ISD::SETULE,
558 void visitSetGE(User &I) { visitSetCC(I, ISD::SETGE, ISD::SETUGE,
560 void visitSetLT(User &I) { visitSetCC(I, ISD::SETLT, ISD::SETULT,
562 void visitSetGT(User &I) { visitSetCC(I, ISD::SETGT, ISD::SETUGT,
565 void visitExtractElement(User &I);
566 void visitInsertElement(User &I);
567 void visitShuffleVector(User &I);
569 void visitGetElementPtr(User &I);
570 void visitCast(User &I);
571 void visitSelect(User &I);
573 void visitMalloc(MallocInst &I);
574 void visitFree(FreeInst &I);
575 void visitAlloca(AllocaInst &I);
576 void visitLoad(LoadInst &I);
577 void visitStore(StoreInst &I);
578 void visitPHI(PHINode &I) { } // PHI nodes are handled specially.
579 void visitCall(CallInst &I);
580 void visitInlineAsm(CallInst &I);
581 const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic);
582 void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic);
584 void visitVAStart(CallInst &I);
585 void visitVAArg(VAArgInst &I);
586 void visitVAEnd(CallInst &I);
587 void visitVACopy(CallInst &I);
588 void visitFrameReturnAddress(CallInst &I, bool isFrameAddress);
590 void visitMemIntrinsic(CallInst &I, unsigned Op);
592 void visitUserOp1(Instruction &I) {
593 assert(0 && "UserOp1 should not exist at instruction selection time!");
596 void visitUserOp2(Instruction &I) {
597 assert(0 && "UserOp2 should not exist at instruction selection time!");
601 } // end namespace llvm
603 SDOperand SelectionDAGLowering::getValue(const Value *V) {
604 SDOperand &N = NodeMap[V];
607 const Type *VTy = V->getType();
608 MVT::ValueType VT = TLI.getValueType(VTy);
609 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
611 visit(CE->getOpcode(), *CE);
612 assert(N.Val && "visit didn't populate the ValueMap!");
614 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
615 return N = DAG.getGlobalAddress(GV, VT);
616 } else if (isa<ConstantPointerNull>(C)) {
617 return N = DAG.getConstant(0, TLI.getPointerTy());
618 } else if (isa<UndefValue>(C)) {
619 if (!isa<PackedType>(VTy))
620 return N = DAG.getNode(ISD::UNDEF, VT);
622 // Create a VBUILD_VECTOR of undef nodes.
623 const PackedType *PTy = cast<PackedType>(VTy);
624 unsigned NumElements = PTy->getNumElements();
625 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
627 SmallVector<SDOperand, 8> Ops;
628 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT));
630 // Create a VConstant node with generic Vector type.
631 Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
632 Ops.push_back(DAG.getValueType(PVT));
633 return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
634 &Ops[0], Ops.size());
635 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
636 return N = DAG.getConstantFP(CFP->getValue(), VT);
637 } else if (const PackedType *PTy = dyn_cast<PackedType>(VTy)) {
638 unsigned NumElements = PTy->getNumElements();
639 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
641 // Now that we know the number and type of the elements, push a
642 // Constant or ConstantFP node onto the ops list for each element of
643 // the packed constant.
644 SmallVector<SDOperand, 8> Ops;
645 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(C)) {
646 for (unsigned i = 0; i != NumElements; ++i)
647 Ops.push_back(getValue(CP->getOperand(i)));
649 assert(isa<ConstantAggregateZero>(C) && "Unknown packed constant!");
651 if (MVT::isFloatingPoint(PVT))
652 Op = DAG.getConstantFP(0, PVT);
654 Op = DAG.getConstant(0, PVT);
655 Ops.assign(NumElements, Op);
658 // Create a VBUILD_VECTOR node with generic Vector type.
659 Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
660 Ops.push_back(DAG.getValueType(PVT));
661 return N = DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector,&Ops[0],Ops.size());
663 // Canonicalize all constant ints to be unsigned.
664 return N = DAG.getConstant(cast<ConstantIntegral>(C)->getZExtValue(),VT);
668 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
669 std::map<const AllocaInst*, int>::iterator SI =
670 FuncInfo.StaticAllocaMap.find(AI);
671 if (SI != FuncInfo.StaticAllocaMap.end())
672 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
675 std::map<const Value*, unsigned>::const_iterator VMI =
676 FuncInfo.ValueMap.find(V);
677 assert(VMI != FuncInfo.ValueMap.end() && "Value not in map!");
679 unsigned InReg = VMI->second;
681 // If this type is not legal, make it so now.
682 if (VT != MVT::Vector) {
683 MVT::ValueType DestVT = TLI.getTypeToTransformTo(VT);
685 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT);
687 // Source must be expanded. This input value is actually coming from the
688 // register pair VMI->second and VMI->second+1.
689 N = DAG.getNode(ISD::BUILD_PAIR, VT, N,
690 DAG.getCopyFromReg(DAG.getEntryNode(), InReg+1, DestVT));
691 } else if (DestVT > VT) { // Promotion case
692 if (MVT::isFloatingPoint(VT))
693 N = DAG.getNode(ISD::FP_ROUND, VT, N);
695 N = DAG.getNode(ISD::TRUNCATE, VT, N);
698 // Otherwise, if this is a vector, make it available as a generic vector
700 MVT::ValueType PTyElementVT, PTyLegalElementVT;
701 const PackedType *PTy = cast<PackedType>(VTy);
702 unsigned NE = TLI.getPackedTypeBreakdown(PTy, PTyElementVT,
705 // Build a VBUILD_VECTOR with the input registers.
706 SmallVector<SDOperand, 8> Ops;
707 if (PTyElementVT == PTyLegalElementVT) {
708 // If the value types are legal, just VBUILD the CopyFromReg nodes.
709 for (unsigned i = 0; i != NE; ++i)
710 Ops.push_back(DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
712 } else if (PTyElementVT < PTyLegalElementVT) {
713 // If the register was promoted, use TRUNCATE of FP_ROUND as appropriate.
714 for (unsigned i = 0; i != NE; ++i) {
715 SDOperand Op = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
717 if (MVT::isFloatingPoint(PTyElementVT))
718 Op = DAG.getNode(ISD::FP_ROUND, PTyElementVT, Op);
720 Op = DAG.getNode(ISD::TRUNCATE, PTyElementVT, Op);
724 // If the register was expanded, use BUILD_PAIR.
725 assert((NE & 1) == 0 && "Must expand into a multiple of 2 elements!");
726 for (unsigned i = 0; i != NE/2; ++i) {
727 SDOperand Op0 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
729 SDOperand Op1 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
731 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Op0, Op1));
735 Ops.push_back(DAG.getConstant(NE, MVT::i32));
736 Ops.push_back(DAG.getValueType(PTyLegalElementVT));
737 N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
739 // Finally, use a VBIT_CONVERT to make this available as the appropriate
741 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
742 DAG.getConstant(PTy->getNumElements(),
744 DAG.getValueType(TLI.getValueType(PTy->getElementType())));
751 void SelectionDAGLowering::visitRet(ReturnInst &I) {
752 if (I.getNumOperands() == 0) {
753 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot()));
756 SmallVector<SDOperand, 8> NewValues;
757 NewValues.push_back(getRoot());
758 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
759 SDOperand RetOp = getValue(I.getOperand(i));
760 bool isSigned = I.getOperand(i)->getType()->isSigned();
762 // If this is an integer return value, we need to promote it ourselves to
763 // the full width of a register, since LegalizeOp will use ANY_EXTEND rather
765 // FIXME: C calling convention requires the return type to be promoted to
766 // at least 32-bit. But this is not necessary for non-C calling conventions.
767 if (MVT::isInteger(RetOp.getValueType()) &&
768 RetOp.getValueType() < MVT::i64) {
769 MVT::ValueType TmpVT;
770 if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote)
771 TmpVT = TLI.getTypeToTransformTo(MVT::i32);
776 RetOp = DAG.getNode(ISD::SIGN_EXTEND, TmpVT, RetOp);
778 RetOp = DAG.getNode(ISD::ZERO_EXTEND, TmpVT, RetOp);
780 NewValues.push_back(RetOp);
781 NewValues.push_back(DAG.getConstant(isSigned, MVT::i32));
783 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other,
784 &NewValues[0], NewValues.size()));
787 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
788 /// the current basic block, add it to ValueMap now so that we'll get a
790 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
791 // No need to export constants.
792 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
795 if (FuncInfo.isExportedInst(V)) return;
797 unsigned Reg = FuncInfo.InitializeRegForValue(V);
798 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg));
801 /// FindMergedConditions - If Cond is an expression like
802 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
803 MachineBasicBlock *TBB,
804 MachineBasicBlock *FBB,
805 MachineBasicBlock *CurBB,
807 // FIXME: HANDLE AND.
810 // If this node is not part of the or/and tree, emit it as a branch.
811 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Cond);
813 if (!BOp || (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
814 BOp->getParent() != CurBB->getBasicBlock()) {
815 const BasicBlock *BB = CurBB->getBasicBlock();
817 // If the leaf of the tree is a setcond inst, merge the condition into the
819 if (BOp && isa<SetCondInst>(BOp) &&
820 // The operands of the setcc have to be in this block. We don't know
821 // how to export them from some other block.
822 (!isa<Instruction>(BOp->getOperand(0)) ||
823 cast<Instruction>(BOp->getOperand(0))->getParent() == BB ||
824 FuncInfo.isExportedInst(BOp->getOperand(0))) &&
825 (!isa<Instruction>(BOp->getOperand(1)) ||
826 cast<Instruction>(BOp->getOperand(1))->getParent() == BB ||
827 FuncInfo.isExportedInst(BOp->getOperand(1)))) {
828 ExportFromCurrentBlock(BOp->getOperand(0));
829 ExportFromCurrentBlock(BOp->getOperand(1));
831 ISD::CondCode SignCond, UnsCond, FPCond, Condition;
832 switch (BOp->getOpcode()) {
833 default: assert(0 && "Unknown setcc opcode!");
834 case Instruction::SetEQ:
835 SignCond = ISD::SETEQ;
836 UnsCond = ISD::SETEQ;
837 FPCond = ISD::SETOEQ;
839 case Instruction::SetNE:
840 SignCond = ISD::SETNE;
841 UnsCond = ISD::SETNE;
842 FPCond = ISD::SETUNE;
844 case Instruction::SetLE:
845 SignCond = ISD::SETLE;
846 UnsCond = ISD::SETULE;
847 FPCond = ISD::SETOLE;
849 case Instruction::SetGE:
850 SignCond = ISD::SETGE;
851 UnsCond = ISD::SETUGE;
852 FPCond = ISD::SETOGE;
854 case Instruction::SetLT:
855 SignCond = ISD::SETLT;
856 UnsCond = ISD::SETULT;
857 FPCond = ISD::SETOLT;
859 case Instruction::SetGT:
860 SignCond = ISD::SETGT;
861 UnsCond = ISD::SETUGT;
862 FPCond = ISD::SETOGT;
866 const Type *OpType = BOp->getOperand(0)->getType();
867 if (const PackedType *PTy = dyn_cast<PackedType>(OpType))
868 OpType = PTy->getElementType();
870 if (!FiniteOnlyFPMath() && OpType->isFloatingPoint())
872 else if (OpType->isUnsigned())
875 Condition = SignCond;
877 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0),
878 BOp->getOperand(1), TBB, FBB, CurBB);
879 SwitchCases.push_back(CB);
883 // Create a CaseBlock record representing this branch.
884 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantBool::getTrue(),
886 SwitchCases.push_back(CB);
887 ExportFromCurrentBlock(Cond);
892 // Create TmpBB after CurBB.
893 MachineFunction::iterator BBI = CurBB;
894 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock());
895 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB);
897 if (Opc == Instruction::Or) {
906 // Emit the LHS condition.
907 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
909 // Emit the RHS condition into TmpBB.
910 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
912 assert(Opc == Instruction::And && "Unknown merge op!");
920 // This requires creation of TmpBB after CurBB.
922 // Emit the LHS condition.
923 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
925 // Emit the RHS condition into TmpBB.
926 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
930 void SelectionDAGLowering::visitBr(BranchInst &I) {
931 // Update machine-CFG edges.
932 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
934 // Figure out which block is immediately after the current one.
935 MachineBasicBlock *NextBlock = 0;
936 MachineFunction::iterator BBI = CurMBB;
937 if (++BBI != CurMBB->getParent()->end())
940 if (I.isUnconditional()) {
941 // If this is not a fall-through branch, emit the branch.
942 if (Succ0MBB != NextBlock)
943 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
944 DAG.getBasicBlock(Succ0MBB)));
946 // Update machine-CFG edges.
947 CurMBB->addSuccessor(Succ0MBB);
952 // If this condition is one of the special cases we handle, do special stuff
954 Value *CondVal = I.getCondition();
955 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
957 // If this is a series of conditions that are or'd or and'd together, emit
958 // this as a sequence of branches instead of setcc's with and/or operations.
959 // For example, instead of something like:
972 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
973 if (BOp->hasOneUse() &&
974 (BOp->getOpcode() == Instruction::And ||
975 BOp->getOpcode() == Instruction::Or)) {
976 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
977 visitSwitchCase(SwitchCases[0]);
978 SwitchCases.erase(SwitchCases.begin());
983 // Create a CaseBlock record representing this branch.
984 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantBool::getTrue(),
985 Succ0MBB, Succ1MBB, CurMBB);
986 // Use visitSwitchCase to actually insert the fast branch sequence for this
991 /// visitSwitchCase - Emits the necessary code to represent a single node in
992 /// the binary search tree resulting from lowering a switch instruction.
993 void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) {
995 SDOperand CondLHS = getValue(CB.CmpLHS);
997 // Build the setcc now, fold "(X == true)" to X and "(X == false)" to !X to
998 // handle common cases produced by branch lowering.
999 if (CB.CmpRHS == ConstantBool::getTrue() && CB.CC == ISD::SETEQ)
1001 else if (CB.CmpRHS == ConstantBool::getFalse() && CB.CC == ISD::SETEQ) {
1002 SDOperand True = DAG.getConstant(1, CondLHS.getValueType());
1003 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True);
1005 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1007 // Set NextBlock to be the MBB immediately after the current one, if any.
1008 // This is used to avoid emitting unnecessary branches to the next block.
1009 MachineBasicBlock *NextBlock = 0;
1010 MachineFunction::iterator BBI = CurMBB;
1011 if (++BBI != CurMBB->getParent()->end())
1014 // If the lhs block is the next block, invert the condition so that we can
1015 // fall through to the lhs instead of the rhs block.
1016 if (CB.TrueBB == NextBlock) {
1017 std::swap(CB.TrueBB, CB.FalseBB);
1018 SDOperand True = DAG.getConstant(1, Cond.getValueType());
1019 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
1021 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
1022 DAG.getBasicBlock(CB.TrueBB));
1023 if (CB.FalseBB == NextBlock)
1024 DAG.setRoot(BrCond);
1026 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1027 DAG.getBasicBlock(CB.FalseBB)));
1028 // Update successor info
1029 CurMBB->addSuccessor(CB.TrueBB);
1030 CurMBB->addSuccessor(CB.FalseBB);
1033 void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) {
1034 // Emit the code for the jump table
1035 MVT::ValueType PTy = TLI.getPointerTy();
1036 assert((PTy == MVT::i32 || PTy == MVT::i64) &&
1037 "Jump table entries are 32-bit values");
1038 bool isPIC = TLI.getTargetMachine().getRelocationModel() == Reloc::PIC_;
1039 // PIC jump table entries are 32-bit values.
1040 unsigned EntrySize = isPIC ? 4 : MVT::getSizeInBits(PTy)/8;
1041 SDOperand Copy = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy);
1042 SDOperand IDX = DAG.getNode(ISD::MUL, PTy, Copy,
1043 DAG.getConstant(EntrySize, PTy));
1044 SDOperand TAB = DAG.getJumpTable(JT.JTI,PTy);
1045 SDOperand ADD = DAG.getNode(ISD::ADD, PTy, IDX, TAB);
1046 SDOperand LD = DAG.getLoad(isPIC ? MVT::i32 : PTy, Copy.getValue(1), ADD,
1049 // For Pic, the sequence is:
1050 // BRIND(load(Jumptable + index) + RelocBase)
1051 // RelocBase is the JumpTable on PPC and X86, GOT on Alpha
1053 if (TLI.usesGlobalOffsetTable())
1054 Reloc = DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, PTy);
1057 ADD = (PTy != MVT::i32) ? DAG.getNode(ISD::SIGN_EXTEND, PTy, LD) : LD;
1058 ADD = DAG.getNode(ISD::ADD, PTy, ADD, Reloc);
1059 DAG.setRoot(DAG.getNode(ISD::BRIND, MVT::Other, LD.getValue(1), ADD));
1061 DAG.setRoot(DAG.getNode(ISD::BRIND, MVT::Other, LD.getValue(1), LD));
1065 void SelectionDAGLowering::visitSwitch(SwitchInst &I) {
1066 // Figure out which block is immediately after the current one.
1067 MachineBasicBlock *NextBlock = 0;
1068 MachineFunction::iterator BBI = CurMBB;
1070 if (++BBI != CurMBB->getParent()->end())
1073 MachineBasicBlock *Default = FuncInfo.MBBMap[I.getDefaultDest()];
1075 // If there is only the default destination, branch to it if it is not the
1076 // next basic block. Otherwise, just fall through.
1077 if (I.getNumOperands() == 2) {
1078 // Update machine-CFG edges.
1080 // If this is not a fall-through branch, emit the branch.
1081 if (Default != NextBlock)
1082 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1083 DAG.getBasicBlock(Default)));
1085 CurMBB->addSuccessor(Default);
1089 // If there are any non-default case statements, create a vector of Cases
1090 // representing each one, and sort the vector so that we can efficiently
1091 // create a binary search tree from them.
1092 std::vector<Case> Cases;
1094 for (unsigned i = 1; i < I.getNumSuccessors(); ++i) {
1095 MachineBasicBlock *SMBB = FuncInfo.MBBMap[I.getSuccessor(i)];
1096 Cases.push_back(Case(I.getSuccessorValue(i), SMBB));
1099 std::sort(Cases.begin(), Cases.end(), CaseCmp());
1101 // Get the Value to be switched on and default basic blocks, which will be
1102 // inserted into CaseBlock records, representing basic blocks in the binary
1104 Value *SV = I.getOperand(0);
1106 // Get the MachineFunction which holds the current MBB. This is used during
1107 // emission of jump tables, and when inserting any additional MBBs necessary
1108 // to represent the switch.
1109 MachineFunction *CurMF = CurMBB->getParent();
1110 const BasicBlock *LLVMBB = CurMBB->getBasicBlock();
1112 // If the switch has few cases (two or less) emit a series of specific
1114 if (Cases.size() < 3) {
1115 // TODO: If any two of the cases has the same destination, and if one value
1116 // is the same as the other, but has one bit unset that the other has set,
1117 // use bit manipulation to do two compares at once. For example:
1118 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1120 // Rearrange the case blocks so that the last one falls through if possible.
1121 if (NextBlock && Default != NextBlock && Cases.back().second != NextBlock) {
1122 // The last case block won't fall through into 'NextBlock' if we emit the
1123 // branches in this order. See if rearranging a case value would help.
1124 for (unsigned i = 0, e = Cases.size()-1; i != e; ++i) {
1125 if (Cases[i].second == NextBlock) {
1126 std::swap(Cases[i], Cases.back());
1132 // Create a CaseBlock record representing a conditional branch to
1133 // the Case's target mbb if the value being switched on SV is equal
1135 MachineBasicBlock *CurBlock = CurMBB;
1136 for (unsigned i = 0, e = Cases.size(); i != e; ++i) {
1137 MachineBasicBlock *FallThrough;
1139 FallThrough = new MachineBasicBlock(CurMBB->getBasicBlock());
1140 CurMF->getBasicBlockList().insert(BBI, FallThrough);
1142 // If the last case doesn't match, go to the default block.
1143 FallThrough = Default;
1146 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, Cases[i].first,
1147 Cases[i].second, FallThrough, CurBlock);
1149 // If emitting the first comparison, just call visitSwitchCase to emit the
1150 // code into the current block. Otherwise, push the CaseBlock onto the
1151 // vector to be later processed by SDISel, and insert the node's MBB
1152 // before the next MBB.
1153 if (CurBlock == CurMBB)
1154 visitSwitchCase(CB);
1156 SwitchCases.push_back(CB);
1158 CurBlock = FallThrough;
1163 // If the switch has more than 5 blocks, and at least 31.25% dense, and the
1164 // target supports indirect branches, then emit a jump table rather than
1165 // lowering the switch to a binary tree of conditional branches.
1166 if (TLI.isOperationLegal(ISD::BRIND, TLI.getPointerTy()) &&
1168 uint64_t First =cast<ConstantIntegral>(Cases.front().first)->getZExtValue();
1169 uint64_t Last = cast<ConstantIntegral>(Cases.back().first)->getZExtValue();
1170 double Density = (double)Cases.size() / (double)((Last - First) + 1ULL);
1172 if (Density >= 0.3125) {
1173 // Create a new basic block to hold the code for loading the address
1174 // of the jump table, and jumping to it. Update successor information;
1175 // we will either branch to the default case for the switch, or the jump
1177 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB);
1178 CurMF->getBasicBlockList().insert(BBI, JumpTableBB);
1179 CurMBB->addSuccessor(Default);
1180 CurMBB->addSuccessor(JumpTableBB);
1182 // Subtract the lowest switch case value from the value being switched on
1183 // and conditional branch to default mbb if the result is greater than the
1184 // difference between smallest and largest cases.
1185 SDOperand SwitchOp = getValue(SV);
1186 MVT::ValueType VT = SwitchOp.getValueType();
1187 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1188 DAG.getConstant(First, VT));
1190 // The SDNode we just created, which holds the value being switched on
1191 // minus the the smallest case value, needs to be copied to a virtual
1192 // register so it can be used as an index into the jump table in a
1193 // subsequent basic block. This value may be smaller or larger than the
1194 // target's pointer type, and therefore require extension or truncating.
1195 if (VT > TLI.getPointerTy())
1196 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
1198 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
1200 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1201 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp);
1203 // Emit the range check for the jump table, and branch to the default
1204 // block for the switch statement if the value being switched on exceeds
1205 // the largest case in the switch.
1206 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1207 DAG.getConstant(Last-First,VT), ISD::SETUGT);
1208 DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
1209 DAG.getBasicBlock(Default)));
1211 // Build a vector of destination BBs, corresponding to each target
1212 // of the jump table. If the value of the jump table slot corresponds to
1213 // a case statement, push the case's BB onto the vector, otherwise, push
1215 std::vector<MachineBasicBlock*> DestBBs;
1216 uint64_t TEI = First;
1217 for (CaseItr ii = Cases.begin(), ee = Cases.end(); ii != ee; ++TEI)
1218 if (cast<ConstantIntegral>(ii->first)->getZExtValue() == TEI) {
1219 DestBBs.push_back(ii->second);
1222 DestBBs.push_back(Default);
1225 // Update successor info
1226 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1227 E = DestBBs.end(); I != E; ++I)
1228 JumpTableBB->addSuccessor(*I);
1230 // Create a jump table index for this jump table, or return an existing
1232 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1234 // Set the jump table information so that we can codegen it as a second
1235 // MachineBasicBlock
1236 JT.Reg = JumpTableReg;
1238 JT.MBB = JumpTableBB;
1239 JT.Default = Default;
1244 // Push the initial CaseRec onto the worklist
1245 std::vector<CaseRec> CaseVec;
1246 CaseVec.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
1248 while (!CaseVec.empty()) {
1249 // Grab a record representing a case range to process off the worklist
1250 CaseRec CR = CaseVec.back();
1253 // Size is the number of Cases represented by this range. If Size is 1,
1254 // then we are processing a leaf of the binary search tree. Otherwise,
1255 // we need to pick a pivot, and push left and right ranges onto the
1257 unsigned Size = CR.Range.second - CR.Range.first;
1260 // Create a CaseBlock record representing a conditional branch to
1261 // the Case's target mbb if the value being switched on SV is equal
1262 // to C. Otherwise, branch to default.
1263 Constant *C = CR.Range.first->first;
1264 MachineBasicBlock *Target = CR.Range.first->second;
1265 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, C, Target, Default,
1268 // If the MBB representing the leaf node is the current MBB, then just
1269 // call visitSwitchCase to emit the code into the current block.
1270 // Otherwise, push the CaseBlock onto the vector to be later processed
1271 // by SDISel, and insert the node's MBB before the next MBB.
1272 if (CR.CaseBB == CurMBB)
1273 visitSwitchCase(CB);
1275 SwitchCases.push_back(CB);
1277 // split case range at pivot
1278 CaseItr Pivot = CR.Range.first + (Size / 2);
1279 CaseRange LHSR(CR.Range.first, Pivot);
1280 CaseRange RHSR(Pivot, CR.Range.second);
1281 Constant *C = Pivot->first;
1282 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1284 // We know that we branch to the LHS if the Value being switched on is
1285 // less than the Pivot value, C. We use this to optimize our binary
1286 // tree a bit, by recognizing that if SV is greater than or equal to the
1287 // LHS's Case Value, and that Case Value is exactly one less than the
1288 // Pivot's Value, then we can branch directly to the LHS's Target,
1289 // rather than creating a leaf node for it.
1290 if ((LHSR.second - LHSR.first) == 1 &&
1291 LHSR.first->first == CR.GE &&
1292 cast<ConstantIntegral>(C)->getZExtValue() ==
1293 (cast<ConstantIntegral>(CR.GE)->getZExtValue() + 1ULL)) {
1294 TrueBB = LHSR.first->second;
1296 TrueBB = new MachineBasicBlock(LLVMBB);
1297 CurMF->getBasicBlockList().insert(BBI, TrueBB);
1298 CaseVec.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1301 // Similar to the optimization above, if the Value being switched on is
1302 // known to be less than the Constant CR.LT, and the current Case Value
1303 // is CR.LT - 1, then we can branch directly to the target block for
1304 // the current Case Value, rather than emitting a RHS leaf node for it.
1305 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1306 cast<ConstantIntegral>(RHSR.first->first)->getZExtValue() ==
1307 (cast<ConstantIntegral>(CR.LT)->getZExtValue() - 1ULL)) {
1308 FalseBB = RHSR.first->second;
1310 FalseBB = new MachineBasicBlock(LLVMBB);
1311 CurMF->getBasicBlockList().insert(BBI, FalseBB);
1312 CaseVec.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1315 // Create a CaseBlock record representing a conditional branch to
1316 // the LHS node if the value being switched on SV is less than C.
1317 // Otherwise, branch to LHS.
1318 ISD::CondCode CC = C->getType()->isSigned() ? ISD::SETLT : ISD::SETULT;
1319 SelectionDAGISel::CaseBlock CB(CC, SV, C, TrueBB, FalseBB, CR.CaseBB);
1321 if (CR.CaseBB == CurMBB)
1322 visitSwitchCase(CB);
1324 SwitchCases.push_back(CB);
1329 void SelectionDAGLowering::visitSub(User &I) {
1330 // -0.0 - X --> fneg
1331 if (I.getType()->isFloatingPoint()) {
1332 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
1333 if (CFP->isExactlyValue(-0.0)) {
1334 SDOperand Op2 = getValue(I.getOperand(1));
1335 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
1338 visitFPBinary(I, ISD::FSUB, ISD::VSUB);
1340 visitIntBinary(I, ISD::SUB, ISD::VSUB);
1344 SelectionDAGLowering::visitIntBinary(User &I, unsigned IntOp, unsigned VecOp) {
1345 const Type *Ty = I.getType();
1346 SDOperand Op1 = getValue(I.getOperand(0));
1347 SDOperand Op2 = getValue(I.getOperand(1));
1349 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1350 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32);
1351 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType()));
1352 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ));
1354 setValue(&I, DAG.getNode(IntOp, Op1.getValueType(), Op1, Op2));
1359 SelectionDAGLowering::visitFPBinary(User &I, unsigned FPOp, unsigned VecOp) {
1360 const Type *Ty = I.getType();
1361 SDOperand Op1 = getValue(I.getOperand(0));
1362 SDOperand Op2 = getValue(I.getOperand(1));
1364 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1365 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32);
1366 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType()));
1367 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ));
1369 setValue(&I, DAG.getNode(FPOp, Op1.getValueType(), Op1, Op2));
1373 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
1374 SDOperand Op1 = getValue(I.getOperand(0));
1375 SDOperand Op2 = getValue(I.getOperand(1));
1377 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
1379 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
1382 void SelectionDAGLowering::visitSetCC(User &I,ISD::CondCode SignedOpcode,
1383 ISD::CondCode UnsignedOpcode,
1384 ISD::CondCode FPOpcode) {
1385 SDOperand Op1 = getValue(I.getOperand(0));
1386 SDOperand Op2 = getValue(I.getOperand(1));
1387 ISD::CondCode Opcode = SignedOpcode;
1388 if (!FiniteOnlyFPMath() && I.getOperand(0)->getType()->isFloatingPoint())
1390 else if (I.getOperand(0)->getType()->isUnsigned())
1391 Opcode = UnsignedOpcode;
1392 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
1395 void SelectionDAGLowering::visitSelect(User &I) {
1396 SDOperand Cond = getValue(I.getOperand(0));
1397 SDOperand TrueVal = getValue(I.getOperand(1));
1398 SDOperand FalseVal = getValue(I.getOperand(2));
1399 if (!isa<PackedType>(I.getType())) {
1400 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
1401 TrueVal, FalseVal));
1403 setValue(&I, DAG.getNode(ISD::VSELECT, MVT::Vector, Cond, TrueVal, FalseVal,
1404 *(TrueVal.Val->op_end()-2),
1405 *(TrueVal.Val->op_end()-1)));
1409 void SelectionDAGLowering::visitCast(User &I) {
1410 SDOperand N = getValue(I.getOperand(0));
1411 MVT::ValueType SrcVT = N.getValueType();
1412 MVT::ValueType DestVT = TLI.getValueType(I.getType());
1414 if (DestVT == MVT::Vector) {
1415 // This is a cast to a vector from something else. This is always a bit
1416 // convert. Get information about the input vector.
1417 const PackedType *DestTy = cast<PackedType>(I.getType());
1418 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
1419 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N,
1420 DAG.getConstant(DestTy->getNumElements(),MVT::i32),
1421 DAG.getValueType(EltVT)));
1422 } else if (SrcVT == DestVT) {
1423 setValue(&I, N); // noop cast.
1424 } else if (DestVT == MVT::i1) {
1425 // Cast to bool is a comparison against zero, not truncation to zero.
1426 SDOperand Zero = isInteger(SrcVT) ? DAG.getConstant(0, N.getValueType()) :
1427 DAG.getConstantFP(0.0, N.getValueType());
1428 setValue(&I, DAG.getSetCC(MVT::i1, N, Zero, ISD::SETNE));
1429 } else if (isInteger(SrcVT)) {
1430 if (isInteger(DestVT)) { // Int -> Int cast
1431 if (DestVT < SrcVT) // Truncating cast?
1432 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
1433 else if (I.getOperand(0)->getType()->isSigned())
1434 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
1436 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
1437 } else if (isFloatingPoint(DestVT)) { // Int -> FP cast
1438 if (I.getOperand(0)->getType()->isSigned())
1439 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
1441 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
1443 assert(0 && "Unknown cast!");
1445 } else if (isFloatingPoint(SrcVT)) {
1446 if (isFloatingPoint(DestVT)) { // FP -> FP cast
1447 if (DestVT < SrcVT) // Rounding cast?
1448 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N));
1450 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
1451 } else if (isInteger(DestVT)) { // FP -> Int cast.
1452 if (I.getType()->isSigned())
1453 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
1455 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
1457 assert(0 && "Unknown cast!");
1460 assert(SrcVT == MVT::Vector && "Unknown cast!");
1461 assert(DestVT != MVT::Vector && "Casts to vector already handled!");
1462 // This is a cast from a vector to something else. This is always a bit
1463 // convert. Get information about the input vector.
1464 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N));
1468 void SelectionDAGLowering::visitInsertElement(User &I) {
1469 SDOperand InVec = getValue(I.getOperand(0));
1470 SDOperand InVal = getValue(I.getOperand(1));
1471 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
1472 getValue(I.getOperand(2)));
1474 SDOperand Num = *(InVec.Val->op_end()-2);
1475 SDOperand Typ = *(InVec.Val->op_end()-1);
1476 setValue(&I, DAG.getNode(ISD::VINSERT_VECTOR_ELT, MVT::Vector,
1477 InVec, InVal, InIdx, Num, Typ));
1480 void SelectionDAGLowering::visitExtractElement(User &I) {
1481 SDOperand InVec = getValue(I.getOperand(0));
1482 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
1483 getValue(I.getOperand(1)));
1484 SDOperand Typ = *(InVec.Val->op_end()-1);
1485 setValue(&I, DAG.getNode(ISD::VEXTRACT_VECTOR_ELT,
1486 TLI.getValueType(I.getType()), InVec, InIdx));
1489 void SelectionDAGLowering::visitShuffleVector(User &I) {
1490 SDOperand V1 = getValue(I.getOperand(0));
1491 SDOperand V2 = getValue(I.getOperand(1));
1492 SDOperand Mask = getValue(I.getOperand(2));
1494 SDOperand Num = *(V1.Val->op_end()-2);
1495 SDOperand Typ = *(V2.Val->op_end()-1);
1496 setValue(&I, DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
1497 V1, V2, Mask, Num, Typ));
1501 void SelectionDAGLowering::visitGetElementPtr(User &I) {
1502 SDOperand N = getValue(I.getOperand(0));
1503 const Type *Ty = I.getOperand(0)->getType();
1505 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
1508 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
1509 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
1512 uint64_t Offset = TD->getStructLayout(StTy)->MemberOffsets[Field];
1513 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
1514 getIntPtrConstant(Offset));
1516 Ty = StTy->getElementType(Field);
1518 Ty = cast<SequentialType>(Ty)->getElementType();
1520 // If this is a constant subscript, handle it quickly.
1521 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
1522 if (CI->getZExtValue() == 0) continue;
1524 if (CI->getType()->isSigned())
1526 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
1529 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getZExtValue();
1530 N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs));
1534 // N = N + Idx * ElementSize;
1535 uint64_t ElementSize = TD->getTypeSize(Ty);
1536 SDOperand IdxN = getValue(Idx);
1538 // If the index is smaller or larger than intptr_t, truncate or extend
1540 if (IdxN.getValueType() < N.getValueType()) {
1541 if (Idx->getType()->isSigned())
1542 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
1544 IdxN = DAG.getNode(ISD::ZERO_EXTEND, N.getValueType(), IdxN);
1545 } else if (IdxN.getValueType() > N.getValueType())
1546 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
1548 // If this is a multiply by a power of two, turn it into a shl
1549 // immediately. This is a very common case.
1550 if (isPowerOf2_64(ElementSize)) {
1551 unsigned Amt = Log2_64(ElementSize);
1552 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
1553 DAG.getConstant(Amt, TLI.getShiftAmountTy()));
1554 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
1558 SDOperand Scale = getIntPtrConstant(ElementSize);
1559 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
1560 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
1566 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
1567 // If this is a fixed sized alloca in the entry block of the function,
1568 // allocate it statically on the stack.
1569 if (FuncInfo.StaticAllocaMap.count(&I))
1570 return; // getValue will auto-populate this.
1572 const Type *Ty = I.getAllocatedType();
1573 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
1574 unsigned Align = std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
1577 SDOperand AllocSize = getValue(I.getArraySize());
1578 MVT::ValueType IntPtr = TLI.getPointerTy();
1579 if (IntPtr < AllocSize.getValueType())
1580 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
1581 else if (IntPtr > AllocSize.getValueType())
1582 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
1584 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
1585 getIntPtrConstant(TySize));
1587 // Handle alignment. If the requested alignment is less than or equal to the
1588 // stack alignment, ignore it and round the size of the allocation up to the
1589 // stack alignment size. If the size is greater than the stack alignment, we
1590 // note this in the DYNAMIC_STACKALLOC node.
1591 unsigned StackAlign =
1592 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
1593 if (Align <= StackAlign) {
1595 // Add SA-1 to the size.
1596 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
1597 getIntPtrConstant(StackAlign-1));
1598 // Mask out the low bits for alignment purposes.
1599 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
1600 getIntPtrConstant(~(uint64_t)(StackAlign-1)));
1603 SDOperand Ops[] = { getRoot(), AllocSize, getIntPtrConstant(Align) };
1604 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
1606 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3);
1607 DAG.setRoot(setValue(&I, DSA).getValue(1));
1609 // Inform the Frame Information that we have just allocated a variable-sized
1611 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
1614 void SelectionDAGLowering::visitLoad(LoadInst &I) {
1615 SDOperand Ptr = getValue(I.getOperand(0));
1621 // Do not serialize non-volatile loads against each other.
1622 Root = DAG.getRoot();
1625 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0),
1626 Root, I.isVolatile()));
1629 SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr,
1630 const Value *SV, SDOperand Root,
1633 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1634 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
1635 L = DAG.getVecLoad(PTy->getNumElements(), PVT, Root, Ptr,
1636 DAG.getSrcValue(SV));
1638 L = DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, isVolatile);
1642 DAG.setRoot(L.getValue(1));
1644 PendingLoads.push_back(L.getValue(1));
1650 void SelectionDAGLowering::visitStore(StoreInst &I) {
1651 Value *SrcV = I.getOperand(0);
1652 SDOperand Src = getValue(SrcV);
1653 SDOperand Ptr = getValue(I.getOperand(1));
1654 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1),
1658 /// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot
1659 /// access memory and has no other side effects at all.
1660 static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) {
1661 #define GET_NO_MEMORY_INTRINSICS
1662 #include "llvm/Intrinsics.gen"
1663 #undef GET_NO_MEMORY_INTRINSICS
1667 // IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't
1668 // have any side-effects or if it only reads memory.
1669 static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) {
1670 #define GET_SIDE_EFFECT_INFO
1671 #include "llvm/Intrinsics.gen"
1672 #undef GET_SIDE_EFFECT_INFO
1676 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
1678 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
1679 unsigned Intrinsic) {
1680 bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic);
1681 bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic);
1683 // Build the operand list.
1684 SmallVector<SDOperand, 8> Ops;
1685 if (HasChain) { // If this intrinsic has side-effects, chainify it.
1687 // We don't need to serialize loads against other loads.
1688 Ops.push_back(DAG.getRoot());
1690 Ops.push_back(getRoot());
1694 // Add the intrinsic ID as an integer operand.
1695 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
1697 // Add all operands of the call to the operand list.
1698 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
1699 SDOperand Op = getValue(I.getOperand(i));
1701 // If this is a vector type, force it to the right packed type.
1702 if (Op.getValueType() == MVT::Vector) {
1703 const PackedType *OpTy = cast<PackedType>(I.getOperand(i)->getType());
1704 MVT::ValueType EltVT = TLI.getValueType(OpTy->getElementType());
1706 MVT::ValueType VVT = MVT::getVectorType(EltVT, OpTy->getNumElements());
1707 assert(VVT != MVT::Other && "Intrinsic uses a non-legal type?");
1708 Op = DAG.getNode(ISD::VBIT_CONVERT, VVT, Op);
1711 assert(TLI.isTypeLegal(Op.getValueType()) &&
1712 "Intrinsic uses a non-legal type?");
1716 std::vector<MVT::ValueType> VTs;
1717 if (I.getType() != Type::VoidTy) {
1718 MVT::ValueType VT = TLI.getValueType(I.getType());
1719 if (VT == MVT::Vector) {
1720 const PackedType *DestTy = cast<PackedType>(I.getType());
1721 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
1723 VT = MVT::getVectorType(EltVT, DestTy->getNumElements());
1724 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
1727 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
1731 VTs.push_back(MVT::Other);
1733 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs);
1738 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(),
1739 &Ops[0], Ops.size());
1740 else if (I.getType() != Type::VoidTy)
1741 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(),
1742 &Ops[0], Ops.size());
1744 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(),
1745 &Ops[0], Ops.size());
1748 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1);
1750 PendingLoads.push_back(Chain);
1754 if (I.getType() != Type::VoidTy) {
1755 if (const PackedType *PTy = dyn_cast<PackedType>(I.getType())) {
1756 MVT::ValueType EVT = TLI.getValueType(PTy->getElementType());
1757 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
1758 DAG.getConstant(PTy->getNumElements(), MVT::i32),
1759 DAG.getValueType(EVT));
1761 setValue(&I, Result);
1765 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
1766 /// we want to emit this as a call to a named external function, return the name
1767 /// otherwise lower it and return null.
1769 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
1770 switch (Intrinsic) {
1772 // By default, turn this into a target intrinsic node.
1773 visitTargetIntrinsic(I, Intrinsic);
1775 case Intrinsic::vastart: visitVAStart(I); return 0;
1776 case Intrinsic::vaend: visitVAEnd(I); return 0;
1777 case Intrinsic::vacopy: visitVACopy(I); return 0;
1778 case Intrinsic::returnaddress: visitFrameReturnAddress(I, false); return 0;
1779 case Intrinsic::frameaddress: visitFrameReturnAddress(I, true); return 0;
1780 case Intrinsic::setjmp:
1781 return "_setjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
1783 case Intrinsic::longjmp:
1784 return "_longjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
1786 case Intrinsic::memcpy_i32:
1787 case Intrinsic::memcpy_i64:
1788 visitMemIntrinsic(I, ISD::MEMCPY);
1790 case Intrinsic::memset_i32:
1791 case Intrinsic::memset_i64:
1792 visitMemIntrinsic(I, ISD::MEMSET);
1794 case Intrinsic::memmove_i32:
1795 case Intrinsic::memmove_i64:
1796 visitMemIntrinsic(I, ISD::MEMMOVE);
1799 case Intrinsic::dbg_stoppoint: {
1800 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1801 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
1802 if (DebugInfo && SPI.getContext() && DebugInfo->Verify(SPI.getContext())) {
1806 Ops[1] = getValue(SPI.getLineValue());
1807 Ops[2] = getValue(SPI.getColumnValue());
1809 DebugInfoDesc *DD = DebugInfo->getDescFor(SPI.getContext());
1810 assert(DD && "Not a debug information descriptor");
1811 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD);
1813 Ops[3] = DAG.getString(CompileUnit->getFileName());
1814 Ops[4] = DAG.getString(CompileUnit->getDirectory());
1816 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5));
1821 case Intrinsic::dbg_region_start: {
1822 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1823 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
1824 if (DebugInfo && RSI.getContext() && DebugInfo->Verify(RSI.getContext())) {
1825 unsigned LabelID = DebugInfo->RecordRegionStart(RSI.getContext());
1826 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, getRoot(),
1827 DAG.getConstant(LabelID, MVT::i32)));
1832 case Intrinsic::dbg_region_end: {
1833 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1834 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
1835 if (DebugInfo && REI.getContext() && DebugInfo->Verify(REI.getContext())) {
1836 unsigned LabelID = DebugInfo->RecordRegionEnd(REI.getContext());
1837 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other,
1838 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
1843 case Intrinsic::dbg_func_start: {
1844 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1845 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
1846 if (DebugInfo && FSI.getSubprogram() &&
1847 DebugInfo->Verify(FSI.getSubprogram())) {
1848 unsigned LabelID = DebugInfo->RecordRegionStart(FSI.getSubprogram());
1849 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other,
1850 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
1855 case Intrinsic::dbg_declare: {
1856 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1857 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
1858 if (DebugInfo && DI.getVariable() && DebugInfo->Verify(DI.getVariable())) {
1859 SDOperand AddressOp = getValue(DI.getAddress());
1860 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp))
1861 DebugInfo->RecordVariable(DI.getVariable(), FI->getIndex());
1867 case Intrinsic::isunordered_f32:
1868 case Intrinsic::isunordered_f64:
1869 setValue(&I, DAG.getSetCC(MVT::i1,getValue(I.getOperand(1)),
1870 getValue(I.getOperand(2)), ISD::SETUO));
1873 case Intrinsic::sqrt_f32:
1874 case Intrinsic::sqrt_f64:
1875 setValue(&I, DAG.getNode(ISD::FSQRT,
1876 getValue(I.getOperand(1)).getValueType(),
1877 getValue(I.getOperand(1))));
1879 case Intrinsic::powi_f32:
1880 case Intrinsic::powi_f64:
1881 setValue(&I, DAG.getNode(ISD::FPOWI,
1882 getValue(I.getOperand(1)).getValueType(),
1883 getValue(I.getOperand(1)),
1884 getValue(I.getOperand(2))));
1886 case Intrinsic::pcmarker: {
1887 SDOperand Tmp = getValue(I.getOperand(1));
1888 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
1891 case Intrinsic::readcyclecounter: {
1892 SDOperand Op = getRoot();
1893 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER,
1894 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
1897 DAG.setRoot(Tmp.getValue(1));
1900 case Intrinsic::bswap_i16:
1901 case Intrinsic::bswap_i32:
1902 case Intrinsic::bswap_i64:
1903 setValue(&I, DAG.getNode(ISD::BSWAP,
1904 getValue(I.getOperand(1)).getValueType(),
1905 getValue(I.getOperand(1))));
1907 case Intrinsic::cttz_i8:
1908 case Intrinsic::cttz_i16:
1909 case Intrinsic::cttz_i32:
1910 case Intrinsic::cttz_i64:
1911 setValue(&I, DAG.getNode(ISD::CTTZ,
1912 getValue(I.getOperand(1)).getValueType(),
1913 getValue(I.getOperand(1))));
1915 case Intrinsic::ctlz_i8:
1916 case Intrinsic::ctlz_i16:
1917 case Intrinsic::ctlz_i32:
1918 case Intrinsic::ctlz_i64:
1919 setValue(&I, DAG.getNode(ISD::CTLZ,
1920 getValue(I.getOperand(1)).getValueType(),
1921 getValue(I.getOperand(1))));
1923 case Intrinsic::ctpop_i8:
1924 case Intrinsic::ctpop_i16:
1925 case Intrinsic::ctpop_i32:
1926 case Intrinsic::ctpop_i64:
1927 setValue(&I, DAG.getNode(ISD::CTPOP,
1928 getValue(I.getOperand(1)).getValueType(),
1929 getValue(I.getOperand(1))));
1931 case Intrinsic::stacksave: {
1932 SDOperand Op = getRoot();
1933 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE,
1934 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
1936 DAG.setRoot(Tmp.getValue(1));
1939 case Intrinsic::stackrestore: {
1940 SDOperand Tmp = getValue(I.getOperand(1));
1941 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
1944 case Intrinsic::prefetch:
1945 // FIXME: Currently discarding prefetches.
1951 void SelectionDAGLowering::visitCall(CallInst &I) {
1952 const char *RenameFn = 0;
1953 if (Function *F = I.getCalledFunction()) {
1954 if (F->isExternal())
1955 if (unsigned IID = F->getIntrinsicID()) {
1956 RenameFn = visitIntrinsicCall(I, IID);
1959 } else { // Not an LLVM intrinsic.
1960 const std::string &Name = F->getName();
1961 if (Name[0] == 'c' && (Name == "copysign" || Name == "copysignf")) {
1962 if (I.getNumOperands() == 3 && // Basic sanity checks.
1963 I.getOperand(1)->getType()->isFloatingPoint() &&
1964 I.getType() == I.getOperand(1)->getType() &&
1965 I.getType() == I.getOperand(2)->getType()) {
1966 SDOperand LHS = getValue(I.getOperand(1));
1967 SDOperand RHS = getValue(I.getOperand(2));
1968 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
1972 } else if (Name[0] == 'f' && (Name == "fabs" || Name == "fabsf")) {
1973 if (I.getNumOperands() == 2 && // Basic sanity checks.
1974 I.getOperand(1)->getType()->isFloatingPoint() &&
1975 I.getType() == I.getOperand(1)->getType()) {
1976 SDOperand Tmp = getValue(I.getOperand(1));
1977 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
1980 } else if (Name[0] == 's' && (Name == "sin" || Name == "sinf")) {
1981 if (I.getNumOperands() == 2 && // Basic sanity checks.
1982 I.getOperand(1)->getType()->isFloatingPoint() &&
1983 I.getType() == I.getOperand(1)->getType()) {
1984 SDOperand Tmp = getValue(I.getOperand(1));
1985 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
1988 } else if (Name[0] == 'c' && (Name == "cos" || Name == "cosf")) {
1989 if (I.getNumOperands() == 2 && // Basic sanity checks.
1990 I.getOperand(1)->getType()->isFloatingPoint() &&
1991 I.getType() == I.getOperand(1)->getType()) {
1992 SDOperand Tmp = getValue(I.getOperand(1));
1993 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
1998 } else if (isa<InlineAsm>(I.getOperand(0))) {
2005 Callee = getValue(I.getOperand(0));
2007 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
2008 std::vector<std::pair<SDOperand, const Type*> > Args;
2009 Args.reserve(I.getNumOperands());
2010 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2011 Value *Arg = I.getOperand(i);
2012 SDOperand ArgNode = getValue(Arg);
2013 Args.push_back(std::make_pair(ArgNode, Arg->getType()));
2016 const PointerType *PT = cast<PointerType>(I.getCalledValue()->getType());
2017 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2019 std::pair<SDOperand,SDOperand> Result =
2020 TLI.LowerCallTo(getRoot(), I.getType(), FTy->isVarArg(), I.getCallingConv(),
2021 I.isTailCall(), Callee, Args, DAG);
2022 if (I.getType() != Type::VoidTy)
2023 setValue(&I, Result.first);
2024 DAG.setRoot(Result.second);
2027 SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
2028 SDOperand &Chain, SDOperand &Flag)const{
2029 SDOperand Val = DAG.getCopyFromReg(Chain, Regs[0], RegVT, Flag);
2030 Chain = Val.getValue(1);
2031 Flag = Val.getValue(2);
2033 // If the result was expanded, copy from the top part.
2034 if (Regs.size() > 1) {
2035 assert(Regs.size() == 2 &&
2036 "Cannot expand to more than 2 elts yet!");
2037 SDOperand Hi = DAG.getCopyFromReg(Chain, Regs[1], RegVT, Flag);
2038 Chain = Hi.getValue(1);
2039 Flag = Hi.getValue(2);
2040 if (DAG.getTargetLoweringInfo().isLittleEndian())
2041 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi);
2043 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Hi, Val);
2046 // Otherwise, if the return value was promoted or extended, truncate it to the
2047 // appropriate type.
2048 if (RegVT == ValueVT)
2051 if (MVT::isInteger(RegVT)) {
2052 if (ValueVT < RegVT)
2053 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
2055 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val);
2057 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val);
2061 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
2062 /// specified value into the registers specified by this object. This uses
2063 /// Chain/Flag as the input and updates them for the output Chain/Flag.
2064 void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
2065 SDOperand &Chain, SDOperand &Flag,
2066 MVT::ValueType PtrVT) const {
2067 if (Regs.size() == 1) {
2068 // If there is a single register and the types differ, this must be
2070 if (RegVT != ValueVT) {
2071 if (MVT::isInteger(RegVT)) {
2072 if (RegVT < ValueVT)
2073 Val = DAG.getNode(ISD::TRUNCATE, RegVT, Val);
2075 Val = DAG.getNode(ISD::ANY_EXTEND, RegVT, Val);
2077 Val = DAG.getNode(ISD::FP_EXTEND, RegVT, Val);
2079 Chain = DAG.getCopyToReg(Chain, Regs[0], Val, Flag);
2080 Flag = Chain.getValue(1);
2082 std::vector<unsigned> R(Regs);
2083 if (!DAG.getTargetLoweringInfo().isLittleEndian())
2084 std::reverse(R.begin(), R.end());
2086 for (unsigned i = 0, e = R.size(); i != e; ++i) {
2087 SDOperand Part = DAG.getNode(ISD::EXTRACT_ELEMENT, RegVT, Val,
2088 DAG.getConstant(i, PtrVT));
2089 Chain = DAG.getCopyToReg(Chain, R[i], Part, Flag);
2090 Flag = Chain.getValue(1);
2095 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
2096 /// operand list. This adds the code marker and includes the number of
2097 /// values added into it.
2098 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
2099 std::vector<SDOperand> &Ops) const {
2100 Ops.push_back(DAG.getConstant(Code | (Regs.size() << 3), MVT::i32));
2101 for (unsigned i = 0, e = Regs.size(); i != e; ++i)
2102 Ops.push_back(DAG.getRegister(Regs[i], RegVT));
2105 /// isAllocatableRegister - If the specified register is safe to allocate,
2106 /// i.e. it isn't a stack pointer or some other special register, return the
2107 /// register class for the register. Otherwise, return null.
2108 static const TargetRegisterClass *
2109 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
2110 const TargetLowering &TLI, const MRegisterInfo *MRI) {
2111 MVT::ValueType FoundVT = MVT::Other;
2112 const TargetRegisterClass *FoundRC = 0;
2113 for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(),
2114 E = MRI->regclass_end(); RCI != E; ++RCI) {
2115 MVT::ValueType ThisVT = MVT::Other;
2117 const TargetRegisterClass *RC = *RCI;
2118 // If none of the the value types for this register class are valid, we
2119 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2120 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2122 if (TLI.isTypeLegal(*I)) {
2123 // If we have already found this register in a different register class,
2124 // choose the one with the largest VT specified. For example, on
2125 // PowerPC, we favor f64 register classes over f32.
2126 if (FoundVT == MVT::Other ||
2127 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) {
2134 if (ThisVT == MVT::Other) continue;
2136 // NOTE: This isn't ideal. In particular, this might allocate the
2137 // frame pointer in functions that need it (due to them not being taken
2138 // out of allocation, because a variable sized allocation hasn't been seen
2139 // yet). This is a slight code pessimization, but should still work.
2140 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
2141 E = RC->allocation_order_end(MF); I != E; ++I)
2143 // We found a matching register class. Keep looking at others in case
2144 // we find one with larger registers that this physreg is also in.
2153 RegsForValue SelectionDAGLowering::
2154 GetRegistersForValue(const std::string &ConstrCode,
2155 MVT::ValueType VT, bool isOutReg, bool isInReg,
2156 std::set<unsigned> &OutputRegs,
2157 std::set<unsigned> &InputRegs) {
2158 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
2159 TLI.getRegForInlineAsmConstraint(ConstrCode, VT);
2160 std::vector<unsigned> Regs;
2162 unsigned NumRegs = VT != MVT::Other ? TLI.getNumElements(VT) : 1;
2163 MVT::ValueType RegVT;
2164 MVT::ValueType ValueVT = VT;
2166 if (PhysReg.first) {
2167 if (VT == MVT::Other)
2168 ValueVT = *PhysReg.second->vt_begin();
2170 // Get the actual register value type. This is important, because the user
2171 // may have asked for (e.g.) the AX register in i32 type. We need to
2172 // remember that AX is actually i16 to get the right extension.
2173 RegVT = *PhysReg.second->vt_begin();
2175 // This is a explicit reference to a physical register.
2176 Regs.push_back(PhysReg.first);
2178 // If this is an expanded reference, add the rest of the regs to Regs.
2180 TargetRegisterClass::iterator I = PhysReg.second->begin();
2181 TargetRegisterClass::iterator E = PhysReg.second->end();
2182 for (; *I != PhysReg.first; ++I)
2183 assert(I != E && "Didn't find reg!");
2185 // Already added the first reg.
2187 for (; NumRegs; --NumRegs, ++I) {
2188 assert(I != E && "Ran out of registers to allocate!");
2192 return RegsForValue(Regs, RegVT, ValueVT);
2195 // This is a reference to a register class. Allocate NumRegs consecutive,
2196 // available, registers from the class.
2197 std::vector<unsigned> RegClassRegs =
2198 TLI.getRegClassForInlineAsmConstraint(ConstrCode, VT);
2200 const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo();
2201 MachineFunction &MF = *CurMBB->getParent();
2202 unsigned NumAllocated = 0;
2203 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
2204 unsigned Reg = RegClassRegs[i];
2205 // See if this register is available.
2206 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
2207 (isInReg && InputRegs.count(Reg))) { // Already used.
2208 // Make sure we find consecutive registers.
2213 // Check to see if this register is allocatable (i.e. don't give out the
2215 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, MRI);
2217 // Make sure we find consecutive registers.
2222 // Okay, this register is good, we can use it.
2225 // If we allocated enough consecutive
2226 if (NumAllocated == NumRegs) {
2227 unsigned RegStart = (i-NumAllocated)+1;
2228 unsigned RegEnd = i+1;
2229 // Mark all of the allocated registers used.
2230 for (unsigned i = RegStart; i != RegEnd; ++i) {
2231 unsigned Reg = RegClassRegs[i];
2232 Regs.push_back(Reg);
2233 if (isOutReg) OutputRegs.insert(Reg); // Mark reg used.
2234 if (isInReg) InputRegs.insert(Reg); // Mark reg used.
2237 return RegsForValue(Regs, *RC->vt_begin(), VT);
2241 // Otherwise, we couldn't allocate enough registers for this.
2242 return RegsForValue();
2246 /// visitInlineAsm - Handle a call to an InlineAsm object.
2248 void SelectionDAGLowering::visitInlineAsm(CallInst &I) {
2249 InlineAsm *IA = cast<InlineAsm>(I.getOperand(0));
2251 SDOperand AsmStr = DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
2254 // Note, we treat inline asms both with and without side-effects as the same.
2255 // If an inline asm doesn't have side effects and doesn't access memory, we
2256 // could not choose to not chain it.
2257 bool hasSideEffects = IA->hasSideEffects();
2259 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
2260 std::vector<MVT::ValueType> ConstraintVTs;
2262 /// AsmNodeOperands - A list of pairs. The first element is a register, the
2263 /// second is a bitfield where bit #0 is set if it is a use and bit #1 is set
2264 /// if it is a def of that register.
2265 std::vector<SDOperand> AsmNodeOperands;
2266 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain
2267 AsmNodeOperands.push_back(AsmStr);
2269 SDOperand Chain = getRoot();
2272 // We fully assign registers here at isel time. This is not optimal, but
2273 // should work. For register classes that correspond to LLVM classes, we
2274 // could let the LLVM RA do its thing, but we currently don't. Do a prepass
2275 // over the constraints, collecting fixed registers that we know we can't use.
2276 std::set<unsigned> OutputRegs, InputRegs;
2278 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2279 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
2280 std::string &ConstraintCode = Constraints[i].Codes[0];
2282 MVT::ValueType OpVT;
2284 // Compute the value type for each operand and add it to ConstraintVTs.
2285 switch (Constraints[i].Type) {
2286 case InlineAsm::isOutput:
2287 if (!Constraints[i].isIndirectOutput) {
2288 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
2289 OpVT = TLI.getValueType(I.getType());
2291 const Type *OpTy = I.getOperand(OpNum)->getType();
2292 OpVT = TLI.getValueType(cast<PointerType>(OpTy)->getElementType());
2293 OpNum++; // Consumes a call operand.
2296 case InlineAsm::isInput:
2297 OpVT = TLI.getValueType(I.getOperand(OpNum)->getType());
2298 OpNum++; // Consumes a call operand.
2300 case InlineAsm::isClobber:
2305 ConstraintVTs.push_back(OpVT);
2307 if (TLI.getRegForInlineAsmConstraint(ConstraintCode, OpVT).first == 0)
2308 continue; // Not assigned a fixed reg.
2310 // Build a list of regs that this operand uses. This always has a single
2311 // element for promoted/expanded operands.
2312 RegsForValue Regs = GetRegistersForValue(ConstraintCode, OpVT,
2314 OutputRegs, InputRegs);
2316 switch (Constraints[i].Type) {
2317 case InlineAsm::isOutput:
2318 // We can't assign any other output to this register.
2319 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2320 // If this is an early-clobber output, it cannot be assigned to the same
2321 // value as the input reg.
2322 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
2323 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2325 case InlineAsm::isInput:
2326 // We can't assign any other input to this register.
2327 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2329 case InlineAsm::isClobber:
2330 // Clobbered regs cannot be used as inputs or outputs.
2331 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2332 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2337 // Loop over all of the inputs, copying the operand values into the
2338 // appropriate registers and processing the output regs.
2339 RegsForValue RetValRegs;
2340 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
2343 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2344 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
2345 std::string &ConstraintCode = Constraints[i].Codes[0];
2347 switch (Constraints[i].Type) {
2348 case InlineAsm::isOutput: {
2349 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
2350 if (ConstraintCode.size() == 1) // not a physreg name.
2351 CTy = TLI.getConstraintType(ConstraintCode[0]);
2353 if (CTy == TargetLowering::C_Memory) {
2355 SDOperand InOperandVal = getValue(I.getOperand(OpNum));
2357 // Check that the operand (the address to store to) isn't a float.
2358 if (!MVT::isInteger(InOperandVal.getValueType()))
2359 assert(0 && "MATCH FAIL!");
2361 if (!Constraints[i].isIndirectOutput)
2362 assert(0 && "MATCH FAIL!");
2364 OpNum++; // Consumes a call operand.
2366 // Extend/truncate to the right pointer type if needed.
2367 MVT::ValueType PtrType = TLI.getPointerTy();
2368 if (InOperandVal.getValueType() < PtrType)
2369 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
2370 else if (InOperandVal.getValueType() > PtrType)
2371 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
2373 // Add information to the INLINEASM node to know about this output.
2374 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
2375 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2376 AsmNodeOperands.push_back(InOperandVal);
2380 // Otherwise, this is a register output.
2381 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
2383 // If this is an early-clobber output, or if there is an input
2384 // constraint that matches this, we need to reserve the input register
2385 // so no other inputs allocate to it.
2386 bool UsesInputRegister = false;
2387 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
2388 UsesInputRegister = true;
2390 // Copy the output from the appropriate register. Find a register that
2393 GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
2394 true, UsesInputRegister,
2395 OutputRegs, InputRegs);
2396 assert(!Regs.Regs.empty() && "Couldn't allocate output reg!");
2398 if (!Constraints[i].isIndirectOutput) {
2399 assert(RetValRegs.Regs.empty() &&
2400 "Cannot have multiple output constraints yet!");
2401 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
2404 IndirectStoresToEmit.push_back(std::make_pair(Regs,
2405 I.getOperand(OpNum)));
2406 OpNum++; // Consumes a call operand.
2409 // Add information to the INLINEASM node to know that this register is
2411 Regs.AddInlineAsmOperands(2 /*REGDEF*/, DAG, AsmNodeOperands);
2414 case InlineAsm::isInput: {
2415 SDOperand InOperandVal = getValue(I.getOperand(OpNum));
2416 OpNum++; // Consumes a call operand.
2418 if (isdigit(ConstraintCode[0])) { // Matching constraint?
2419 // If this is required to match an output register we have already set,
2420 // just use its register.
2421 unsigned OperandNo = atoi(ConstraintCode.c_str());
2423 // Scan until we find the definition we already emitted of this operand.
2424 // When we find it, create a RegsForValue operand.
2425 unsigned CurOp = 2; // The first operand.
2426 for (; OperandNo; --OperandNo) {
2427 // Advance to the next operand.
2429 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
2430 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
2431 (NumOps & 7) == 4 /*MEM*/) &&
2432 "Skipped past definitions?");
2433 CurOp += (NumOps>>3)+1;
2437 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
2438 assert((NumOps & 7) == 2 /*REGDEF*/ &&
2439 "Skipped past definitions?");
2441 // Add NumOps>>3 registers to MatchedRegs.
2442 RegsForValue MatchedRegs;
2443 MatchedRegs.ValueVT = InOperandVal.getValueType();
2444 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType();
2445 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
2446 unsigned Reg=cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
2447 MatchedRegs.Regs.push_back(Reg);
2450 // Use the produced MatchedRegs object to
2451 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag,
2452 TLI.getPointerTy());
2453 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
2457 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
2458 if (ConstraintCode.size() == 1) // not a physreg name.
2459 CTy = TLI.getConstraintType(ConstraintCode[0]);
2461 if (CTy == TargetLowering::C_Other) {
2462 if (!TLI.isOperandValidForConstraint(InOperandVal, ConstraintCode[0]))
2463 assert(0 && "MATCH FAIL!");
2465 // Add information to the INLINEASM node to know about this input.
2466 unsigned ResOpType = 3 /*IMM*/ | (1 << 3);
2467 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2468 AsmNodeOperands.push_back(InOperandVal);
2470 } else if (CTy == TargetLowering::C_Memory) {
2473 // Check that the operand isn't a float.
2474 if (!MVT::isInteger(InOperandVal.getValueType()))
2475 assert(0 && "MATCH FAIL!");
2477 // Extend/truncate to the right pointer type if needed.
2478 MVT::ValueType PtrType = TLI.getPointerTy();
2479 if (InOperandVal.getValueType() < PtrType)
2480 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
2481 else if (InOperandVal.getValueType() > PtrType)
2482 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
2484 // Add information to the INLINEASM node to know about this input.
2485 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
2486 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2487 AsmNodeOperands.push_back(InOperandVal);
2491 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
2493 // Copy the input into the appropriate registers.
2494 RegsForValue InRegs =
2495 GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
2496 false, true, OutputRegs, InputRegs);
2497 // FIXME: should be match fail.
2498 assert(!InRegs.Regs.empty() && "Couldn't allocate input reg!");
2500 InRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, TLI.getPointerTy());
2502 InRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG, AsmNodeOperands);
2505 case InlineAsm::isClobber: {
2506 RegsForValue ClobberedRegs =
2507 GetRegistersForValue(ConstraintCode, MVT::Other, false, false,
2508 OutputRegs, InputRegs);
2509 // Add the clobbered value to the operand list, so that the register
2510 // allocator is aware that the physreg got clobbered.
2511 if (!ClobberedRegs.Regs.empty())
2512 ClobberedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG, AsmNodeOperands);
2518 // Finish up input operands.
2519 AsmNodeOperands[0] = Chain;
2520 if (Flag.Val) AsmNodeOperands.push_back(Flag);
2522 Chain = DAG.getNode(ISD::INLINEASM,
2523 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
2524 &AsmNodeOperands[0], AsmNodeOperands.size());
2525 Flag = Chain.getValue(1);
2527 // If this asm returns a register value, copy the result from that register
2528 // and set it as the value of the call.
2529 if (!RetValRegs.Regs.empty())
2530 setValue(&I, RetValRegs.getCopyFromRegs(DAG, Chain, Flag));
2532 std::vector<std::pair<SDOperand, Value*> > StoresToEmit;
2534 // Process indirect outputs, first output all of the flagged copies out of
2536 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
2537 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
2538 Value *Ptr = IndirectStoresToEmit[i].second;
2539 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, Flag);
2540 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
2543 // Emit the non-flagged stores from the physregs.
2544 SmallVector<SDOperand, 8> OutChains;
2545 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
2546 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first,
2547 getValue(StoresToEmit[i].second),
2548 StoresToEmit[i].second, 0));
2549 if (!OutChains.empty())
2550 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2551 &OutChains[0], OutChains.size());
2556 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
2557 SDOperand Src = getValue(I.getOperand(0));
2559 MVT::ValueType IntPtr = TLI.getPointerTy();
2561 if (IntPtr < Src.getValueType())
2562 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
2563 else if (IntPtr > Src.getValueType())
2564 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
2566 // Scale the source by the type size.
2567 uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType());
2568 Src = DAG.getNode(ISD::MUL, Src.getValueType(),
2569 Src, getIntPtrConstant(ElementSize));
2571 std::vector<std::pair<SDOperand, const Type*> > Args;
2572 Args.push_back(std::make_pair(Src, TLI.getTargetData()->getIntPtrType()));
2574 std::pair<SDOperand,SDOperand> Result =
2575 TLI.LowerCallTo(getRoot(), I.getType(), false, CallingConv::C, true,
2576 DAG.getExternalSymbol("malloc", IntPtr),
2578 setValue(&I, Result.first); // Pointers always fit in registers
2579 DAG.setRoot(Result.second);
2582 void SelectionDAGLowering::visitFree(FreeInst &I) {
2583 std::vector<std::pair<SDOperand, const Type*> > Args;
2584 Args.push_back(std::make_pair(getValue(I.getOperand(0)),
2585 TLI.getTargetData()->getIntPtrType()));
2586 MVT::ValueType IntPtr = TLI.getPointerTy();
2587 std::pair<SDOperand,SDOperand> Result =
2588 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, CallingConv::C, true,
2589 DAG.getExternalSymbol("free", IntPtr), Args, DAG);
2590 DAG.setRoot(Result.second);
2593 // InsertAtEndOfBasicBlock - This method should be implemented by targets that
2594 // mark instructions with the 'usesCustomDAGSchedInserter' flag. These
2595 // instructions are special in various ways, which require special support to
2596 // insert. The specified MachineInstr is created but not inserted into any
2597 // basic blocks, and the scheduler passes ownership of it to this method.
2598 MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
2599 MachineBasicBlock *MBB) {
2600 std::cerr << "If a target marks an instruction with "
2601 "'usesCustomDAGSchedInserter', it must implement "
2602 "TargetLowering::InsertAtEndOfBasicBlock!\n";
2607 void SelectionDAGLowering::visitVAStart(CallInst &I) {
2608 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
2609 getValue(I.getOperand(1)),
2610 DAG.getSrcValue(I.getOperand(1))));
2613 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
2614 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
2615 getValue(I.getOperand(0)),
2616 DAG.getSrcValue(I.getOperand(0)));
2618 DAG.setRoot(V.getValue(1));
2621 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
2622 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
2623 getValue(I.getOperand(1)),
2624 DAG.getSrcValue(I.getOperand(1))));
2627 void SelectionDAGLowering::visitVACopy(CallInst &I) {
2628 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
2629 getValue(I.getOperand(1)),
2630 getValue(I.getOperand(2)),
2631 DAG.getSrcValue(I.getOperand(1)),
2632 DAG.getSrcValue(I.getOperand(2))));
2635 /// TargetLowering::LowerArguments - This is the default LowerArguments
2636 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
2637 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
2638 /// integrated into SDISel.
2639 std::vector<SDOperand>
2640 TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
2641 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
2642 std::vector<SDOperand> Ops;
2643 Ops.push_back(DAG.getRoot());
2644 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
2645 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
2647 // Add one result value for each formal argument.
2648 std::vector<MVT::ValueType> RetVals;
2649 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
2650 MVT::ValueType VT = getValueType(I->getType());
2652 switch (getTypeAction(VT)) {
2653 default: assert(0 && "Unknown type action!");
2655 RetVals.push_back(VT);
2658 RetVals.push_back(getTypeToTransformTo(VT));
2661 if (VT != MVT::Vector) {
2662 // If this is a large integer, it needs to be broken up into small
2663 // integers. Figure out what the destination type is and how many small
2664 // integers it turns into.
2665 MVT::ValueType NVT = getTypeToTransformTo(VT);
2666 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2667 for (unsigned i = 0; i != NumVals; ++i)
2668 RetVals.push_back(NVT);
2670 // Otherwise, this is a vector type. We only support legal vectors
2672 unsigned NumElems = cast<PackedType>(I->getType())->getNumElements();
2673 const Type *EltTy = cast<PackedType>(I->getType())->getElementType();
2675 // Figure out if there is a Packed type corresponding to this Vector
2676 // type. If so, convert to the packed type.
2677 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2678 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2679 RetVals.push_back(TVT);
2681 assert(0 && "Don't support illegal by-val vector arguments yet!");
2688 RetVals.push_back(MVT::Other);
2691 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS,
2692 DAG.getNodeValueTypes(RetVals), RetVals.size(),
2693 &Ops[0], Ops.size()).Val;
2695 DAG.setRoot(SDOperand(Result, Result->getNumValues()-1));
2697 // Set up the return result vector.
2700 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
2701 MVT::ValueType VT = getValueType(I->getType());
2703 switch (getTypeAction(VT)) {
2704 default: assert(0 && "Unknown type action!");
2706 Ops.push_back(SDOperand(Result, i++));
2709 SDOperand Op(Result, i++);
2710 if (MVT::isInteger(VT)) {
2711 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
2713 Op = DAG.getNode(AssertOp, Op.getValueType(), Op, DAG.getValueType(VT));
2714 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2716 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
2717 Op = DAG.getNode(ISD::FP_ROUND, VT, Op);
2723 if (VT != MVT::Vector) {
2724 // If this is a large integer, it needs to be reassembled from small
2725 // integers. Figure out what the source elt type is and how many small
2727 MVT::ValueType NVT = getTypeToTransformTo(VT);
2728 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2730 SDOperand Lo = SDOperand(Result, i++);
2731 SDOperand Hi = SDOperand(Result, i++);
2733 if (!isLittleEndian())
2736 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Lo, Hi));
2738 // Value scalarized into many values. Unimp for now.
2739 assert(0 && "Cannot expand i64 -> i16 yet!");
2742 // Otherwise, this is a vector type. We only support legal vectors
2744 const PackedType *PTy = cast<PackedType>(I->getType());
2745 unsigned NumElems = PTy->getNumElements();
2746 const Type *EltTy = PTy->getElementType();
2748 // Figure out if there is a Packed type corresponding to this Vector
2749 // type. If so, convert to the packed type.
2750 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2751 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2752 SDOperand N = SDOperand(Result, i++);
2753 // Handle copies from generic vectors to registers.
2754 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
2755 DAG.getConstant(NumElems, MVT::i32),
2756 DAG.getValueType(getValueType(EltTy)));
2759 assert(0 && "Don't support illegal by-val vector arguments yet!");
2770 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
2771 /// implementation, which just inserts an ISD::CALL node, which is later custom
2772 /// lowered by the target to something concrete. FIXME: When all targets are
2773 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
2774 std::pair<SDOperand, SDOperand>
2775 TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
2776 unsigned CallingConv, bool isTailCall,
2778 ArgListTy &Args, SelectionDAG &DAG) {
2779 SmallVector<SDOperand, 32> Ops;
2780 Ops.push_back(Chain); // Op#0 - Chain
2781 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC
2782 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg
2783 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail
2784 Ops.push_back(Callee);
2786 // Handle all of the outgoing arguments.
2787 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
2788 MVT::ValueType VT = getValueType(Args[i].second);
2789 SDOperand Op = Args[i].first;
2790 bool isSigned = Args[i].second->isSigned();
2791 switch (getTypeAction(VT)) {
2792 default: assert(0 && "Unknown type action!");
2795 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2798 if (MVT::isInteger(VT)) {
2799 unsigned ExtOp = isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2800 Op = DAG.getNode(ExtOp, getTypeToTransformTo(VT), Op);
2802 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
2803 Op = DAG.getNode(ISD::FP_EXTEND, getTypeToTransformTo(VT), Op);
2806 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2809 if (VT != MVT::Vector) {
2810 // If this is a large integer, it needs to be broken down into small
2811 // integers. Figure out what the source elt type is and how many small
2813 MVT::ValueType NVT = getTypeToTransformTo(VT);
2814 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2816 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, NVT, Op,
2817 DAG.getConstant(0, getPointerTy()));
2818 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, NVT, Op,
2819 DAG.getConstant(1, getPointerTy()));
2820 if (!isLittleEndian())
2824 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2826 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2828 // Value scalarized into many values. Unimp for now.
2829 assert(0 && "Cannot expand i64 -> i16 yet!");
2832 // Otherwise, this is a vector type. We only support legal vectors
2834 const PackedType *PTy = cast<PackedType>(Args[i].second);
2835 unsigned NumElems = PTy->getNumElements();
2836 const Type *EltTy = PTy->getElementType();
2838 // Figure out if there is a Packed type corresponding to this Vector
2839 // type. If so, convert to the packed type.
2840 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2841 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2842 // Insert a VBIT_CONVERT of the MVT::Vector type to the packed type.
2843 Op = DAG.getNode(ISD::VBIT_CONVERT, TVT, Op);
2845 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2847 assert(0 && "Don't support illegal by-val vector call args yet!");
2855 // Figure out the result value types.
2856 SmallVector<MVT::ValueType, 4> RetTys;
2858 if (RetTy != Type::VoidTy) {
2859 MVT::ValueType VT = getValueType(RetTy);
2860 switch (getTypeAction(VT)) {
2861 default: assert(0 && "Unknown type action!");
2863 RetTys.push_back(VT);
2866 RetTys.push_back(getTypeToTransformTo(VT));
2869 if (VT != MVT::Vector) {
2870 // If this is a large integer, it needs to be reassembled from small
2871 // integers. Figure out what the source elt type is and how many small
2873 MVT::ValueType NVT = getTypeToTransformTo(VT);
2874 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2875 for (unsigned i = 0; i != NumVals; ++i)
2876 RetTys.push_back(NVT);
2878 // Otherwise, this is a vector type. We only support legal vectors
2880 const PackedType *PTy = cast<PackedType>(RetTy);
2881 unsigned NumElems = PTy->getNumElements();
2882 const Type *EltTy = PTy->getElementType();
2884 // Figure out if there is a Packed type corresponding to this Vector
2885 // type. If so, convert to the packed type.
2886 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2887 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2888 RetTys.push_back(TVT);
2890 assert(0 && "Don't support illegal by-val vector call results yet!");
2897 RetTys.push_back(MVT::Other); // Always has a chain.
2899 // Finally, create the CALL node.
2900 SDOperand Res = DAG.getNode(ISD::CALL,
2901 DAG.getVTList(&RetTys[0], RetTys.size()),
2902 &Ops[0], Ops.size());
2904 // This returns a pair of operands. The first element is the
2905 // return value for the function (if RetTy is not VoidTy). The second
2906 // element is the outgoing token chain.
2908 if (RetTys.size() != 1) {
2909 MVT::ValueType VT = getValueType(RetTy);
2910 if (RetTys.size() == 2) {
2913 // If this value was promoted, truncate it down.
2914 if (ResVal.getValueType() != VT) {
2915 if (VT == MVT::Vector) {
2916 // Insert a VBITCONVERT to convert from the packed result type to the
2917 // MVT::Vector type.
2918 unsigned NumElems = cast<PackedType>(RetTy)->getNumElements();
2919 const Type *EltTy = cast<PackedType>(RetTy)->getElementType();
2921 // Figure out if there is a Packed type corresponding to this Vector
2922 // type. If so, convert to the packed type.
2923 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2924 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2925 // Insert a VBIT_CONVERT of the FORMAL_ARGUMENTS to a
2926 // "N x PTyElementVT" MVT::Vector type.
2927 ResVal = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, ResVal,
2928 DAG.getConstant(NumElems, MVT::i32),
2929 DAG.getValueType(getValueType(EltTy)));
2933 } else if (MVT::isInteger(VT)) {
2934 unsigned AssertOp = RetTy->isSigned() ?
2935 ISD::AssertSext : ISD::AssertZext;
2936 ResVal = DAG.getNode(AssertOp, ResVal.getValueType(), ResVal,
2937 DAG.getValueType(VT));
2938 ResVal = DAG.getNode(ISD::TRUNCATE, VT, ResVal);
2940 assert(MVT::isFloatingPoint(VT));
2941 ResVal = DAG.getNode(ISD::FP_ROUND, VT, ResVal);
2944 } else if (RetTys.size() == 3) {
2945 ResVal = DAG.getNode(ISD::BUILD_PAIR, VT,
2946 Res.getValue(0), Res.getValue(1));
2949 assert(0 && "Case not handled yet!");
2953 return std::make_pair(ResVal, Res.getValue(Res.Val->getNumValues()-1));
2958 // It is always conservatively correct for llvm.returnaddress and
2959 // llvm.frameaddress to return 0.
2961 // FIXME: Change this to insert a FRAMEADDR/RETURNADDR node, and have that be
2962 // expanded to 0 if the target wants.
2963 std::pair<SDOperand, SDOperand>
2964 TargetLowering::LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain,
2965 unsigned Depth, SelectionDAG &DAG) {
2966 return std::make_pair(DAG.getConstant(0, getPointerTy()), Chain);
2969 SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
2970 assert(0 && "LowerOperation not implemented for this target!");
2975 SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op,
2976 SelectionDAG &DAG) {
2977 assert(0 && "CustomPromoteOperation not implemented for this target!");
2982 void SelectionDAGLowering::visitFrameReturnAddress(CallInst &I, bool isFrame) {
2983 unsigned Depth = (unsigned)cast<ConstantInt>(I.getOperand(1))->getZExtValue();
2984 std::pair<SDOperand,SDOperand> Result =
2985 TLI.LowerFrameReturnAddress(isFrame, getRoot(), Depth, DAG);
2986 setValue(&I, Result.first);
2987 DAG.setRoot(Result.second);
2990 /// getMemsetValue - Vectorized representation of the memset value
2992 static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT,
2993 SelectionDAG &DAG) {
2994 MVT::ValueType CurVT = VT;
2995 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
2996 uint64_t Val = C->getValue() & 255;
2998 while (CurVT != MVT::i8) {
2999 Val = (Val << Shift) | Val;
3001 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
3003 return DAG.getConstant(Val, VT);
3005 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
3007 while (CurVT != MVT::i8) {
3009 DAG.getNode(ISD::OR, VT,
3010 DAG.getNode(ISD::SHL, VT, Value,
3011 DAG.getConstant(Shift, MVT::i8)), Value);
3013 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
3020 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3021 /// used when a memcpy is turned into a memset when the source is a constant
3023 static SDOperand getMemsetStringVal(MVT::ValueType VT,
3024 SelectionDAG &DAG, TargetLowering &TLI,
3025 std::string &Str, unsigned Offset) {
3026 MVT::ValueType CurVT = VT;
3028 unsigned MSB = getSizeInBits(VT) / 8;
3029 if (TLI.isLittleEndian())
3030 Offset = Offset + MSB - 1;
3031 for (unsigned i = 0; i != MSB; ++i) {
3032 Val = (Val << 8) | Str[Offset];
3033 Offset += TLI.isLittleEndian() ? -1 : 1;
3035 return DAG.getConstant(Val, VT);
3038 /// getMemBasePlusOffset - Returns base and offset node for the
3039 static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset,
3040 SelectionDAG &DAG, TargetLowering &TLI) {
3041 MVT::ValueType VT = Base.getValueType();
3042 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
3045 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3046 /// to replace the memset / memcpy is below the threshold. It also returns the
3047 /// types of the sequence of memory ops to perform memset / memcpy.
3048 static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps,
3049 unsigned Limit, uint64_t Size,
3050 unsigned Align, TargetLowering &TLI) {
3053 if (TLI.allowsUnalignedMemoryAccesses()) {
3056 switch (Align & 7) {
3072 MVT::ValueType LVT = MVT::i64;
3073 while (!TLI.isTypeLegal(LVT))
3074 LVT = (MVT::ValueType)((unsigned)LVT - 1);
3075 assert(MVT::isInteger(LVT));
3080 unsigned NumMemOps = 0;
3082 unsigned VTSize = getSizeInBits(VT) / 8;
3083 while (VTSize > Size) {
3084 VT = (MVT::ValueType)((unsigned)VT - 1);
3087 assert(MVT::isInteger(VT));
3089 if (++NumMemOps > Limit)
3091 MemOps.push_back(VT);
3098 void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) {
3099 SDOperand Op1 = getValue(I.getOperand(1));
3100 SDOperand Op2 = getValue(I.getOperand(2));
3101 SDOperand Op3 = getValue(I.getOperand(3));
3102 SDOperand Op4 = getValue(I.getOperand(4));
3103 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue();
3104 if (Align == 0) Align = 1;
3106 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) {
3107 std::vector<MVT::ValueType> MemOps;
3109 // Expand memset / memcpy to a series of load / store ops
3110 // if the size operand falls below a certain threshold.
3111 SmallVector<SDOperand, 8> OutChains;
3113 default: break; // Do nothing for now.
3115 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(),
3116 Size->getValue(), Align, TLI)) {
3117 unsigned NumMemOps = MemOps.size();
3118 unsigned Offset = 0;
3119 for (unsigned i = 0; i < NumMemOps; i++) {
3120 MVT::ValueType VT = MemOps[i];
3121 unsigned VTSize = getSizeInBits(VT) / 8;
3122 SDOperand Value = getMemsetValue(Op2, VT, DAG);
3123 SDOperand Store = DAG.getStore(getRoot(), Value,
3124 getMemBasePlusOffset(Op1, Offset, DAG, TLI),
3125 I.getOperand(1), Offset);
3126 OutChains.push_back(Store);
3133 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(),
3134 Size->getValue(), Align, TLI)) {
3135 unsigned NumMemOps = MemOps.size();
3136 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0;
3137 GlobalAddressSDNode *G = NULL;
3139 bool CopyFromStr = false;
3141 if (Op2.getOpcode() == ISD::GlobalAddress)
3142 G = cast<GlobalAddressSDNode>(Op2);
3143 else if (Op2.getOpcode() == ISD::ADD &&
3144 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3145 Op2.getOperand(1).getOpcode() == ISD::Constant) {
3146 G = cast<GlobalAddressSDNode>(Op2.getOperand(0));
3147 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue();
3150 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3152 Str = GV->getStringValue(false);
3160 for (unsigned i = 0; i < NumMemOps; i++) {
3161 MVT::ValueType VT = MemOps[i];
3162 unsigned VTSize = getSizeInBits(VT) / 8;
3163 SDOperand Value, Chain, Store;
3166 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
3169 DAG.getStore(Chain, Value,
3170 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
3171 I.getOperand(1), DstOff);
3173 Value = DAG.getLoad(VT, getRoot(),
3174 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI),
3175 I.getOperand(2), SrcOff);
3176 Chain = Value.getValue(1);
3178 DAG.getStore(Chain, Value,
3179 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
3180 I.getOperand(1), DstOff);
3182 OutChains.push_back(Store);
3191 if (!OutChains.empty()) {
3192 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
3193 &OutChains[0], OutChains.size()));
3198 DAG.setRoot(DAG.getNode(Op, MVT::Other, getRoot(), Op1, Op2, Op3, Op4));
3201 //===----------------------------------------------------------------------===//
3202 // SelectionDAGISel code
3203 //===----------------------------------------------------------------------===//
3205 unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) {
3206 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
3209 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
3210 // FIXME: we only modify the CFG to split critical edges. This
3211 // updates dom and loop info.
3212 AU.addRequired<AliasAnalysis>();
3216 /// OptimizeNoopCopyExpression - We have determined that the specified cast
3217 /// instruction is a noop copy (e.g. it's casting from one pointer type to
3218 /// another, int->uint, or int->sbyte on PPC.
3220 /// Return true if any changes are made.
3221 static bool OptimizeNoopCopyExpression(CastInst *CI) {
3222 BasicBlock *DefBB = CI->getParent();
3224 /// InsertedCasts - Only insert a cast in each block once.
3225 std::map<BasicBlock*, CastInst*> InsertedCasts;
3227 bool MadeChange = false;
3228 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
3230 Use &TheUse = UI.getUse();
3231 Instruction *User = cast<Instruction>(*UI);
3233 // Figure out which BB this cast is used in. For PHI's this is the
3234 // appropriate predecessor block.
3235 BasicBlock *UserBB = User->getParent();
3236 if (PHINode *PN = dyn_cast<PHINode>(User)) {
3237 unsigned OpVal = UI.getOperandNo()/2;
3238 UserBB = PN->getIncomingBlock(OpVal);
3241 // Preincrement use iterator so we don't invalidate it.
3244 // If this user is in the same block as the cast, don't change the cast.
3245 if (UserBB == DefBB) continue;
3247 // If we have already inserted a cast into this block, use it.
3248 CastInst *&InsertedCast = InsertedCasts[UserBB];
3250 if (!InsertedCast) {
3251 BasicBlock::iterator InsertPt = UserBB->begin();
3252 while (isa<PHINode>(InsertPt)) ++InsertPt;
3255 new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
3259 // Replace a use of the cast with a use of the new casat.
3260 TheUse = InsertedCast;
3263 // If we removed all uses, nuke the cast.
3264 if (CI->use_empty())
3265 CI->eraseFromParent();
3270 /// InsertGEPComputeCode - Insert code into BB to compute Ptr+PtrOffset,
3271 /// casting to the type of GEPI.
3272 static Instruction *InsertGEPComputeCode(Instruction *&V, BasicBlock *BB,
3273 Instruction *GEPI, Value *Ptr,
3275 if (V) return V; // Already computed.
3277 BasicBlock::iterator InsertPt;
3278 if (BB == GEPI->getParent()) {
3279 // If insert into the GEP's block, insert right after the GEP.
3283 // Otherwise, insert at the top of BB, after any PHI nodes
3284 InsertPt = BB->begin();
3285 while (isa<PHINode>(InsertPt)) ++InsertPt;
3288 // If Ptr is itself a cast, but in some other BB, emit a copy of the cast into
3289 // BB so that there is only one value live across basic blocks (the cast
3291 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
3292 if (CI->getParent() != BB && isa<PointerType>(CI->getOperand(0)->getType()))
3293 Ptr = new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
3295 // Add the offset, cast it to the right type.
3296 Ptr = BinaryOperator::createAdd(Ptr, PtrOffset, "", InsertPt);
3297 return V = new CastInst(Ptr, GEPI->getType(), "", InsertPt);
3300 /// ReplaceUsesOfGEPInst - Replace all uses of RepPtr with inserted code to
3301 /// compute its value. The RepPtr value can be computed with Ptr+PtrOffset. One
3302 /// trivial way of doing this would be to evaluate Ptr+PtrOffset in RepPtr's
3303 /// block, then ReplaceAllUsesWith'ing everything. However, we would prefer to
3304 /// sink PtrOffset into user blocks where doing so will likely allow us to fold
3305 /// the constant add into a load or store instruction. Additionally, if a user
3306 /// is a pointer-pointer cast, we look through it to find its users.
3307 static void ReplaceUsesOfGEPInst(Instruction *RepPtr, Value *Ptr,
3308 Constant *PtrOffset, BasicBlock *DefBB,
3309 GetElementPtrInst *GEPI,
3310 std::map<BasicBlock*,Instruction*> &InsertedExprs) {
3311 while (!RepPtr->use_empty()) {
3312 Instruction *User = cast<Instruction>(RepPtr->use_back());
3314 // If the user is a Pointer-Pointer cast, recurse.
3315 if (isa<CastInst>(User) && isa<PointerType>(User->getType())) {
3316 ReplaceUsesOfGEPInst(User, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs);
3318 // Drop the use of RepPtr. The cast is dead. Don't delete it now, else we
3319 // could invalidate an iterator.
3320 User->setOperand(0, UndefValue::get(RepPtr->getType()));
3324 // If this is a load of the pointer, or a store through the pointer, emit
3325 // the increment into the load/store block.
3326 Instruction *NewVal;
3327 if (isa<LoadInst>(User) ||
3328 (isa<StoreInst>(User) && User->getOperand(0) != RepPtr)) {
3329 NewVal = InsertGEPComputeCode(InsertedExprs[User->getParent()],
3330 User->getParent(), GEPI,
3333 // If this use is not foldable into the addressing mode, use a version
3334 // emitted in the GEP block.
3335 NewVal = InsertGEPComputeCode(InsertedExprs[DefBB], DefBB, GEPI,
3339 if (GEPI->getType() != RepPtr->getType()) {
3340 BasicBlock::iterator IP = NewVal;
3342 NewVal = new CastInst(NewVal, RepPtr->getType(), "", IP);
3344 User->replaceUsesOfWith(RepPtr, NewVal);
3349 /// OptimizeGEPExpression - Since we are doing basic-block-at-a-time instruction
3350 /// selection, we want to be a bit careful about some things. In particular, if
3351 /// we have a GEP instruction that is used in a different block than it is
3352 /// defined, the addressing expression of the GEP cannot be folded into loads or
3353 /// stores that use it. In this case, decompose the GEP and move constant
3354 /// indices into blocks that use it.
3355 static bool OptimizeGEPExpression(GetElementPtrInst *GEPI,
3356 const TargetData *TD) {
3357 // If this GEP is only used inside the block it is defined in, there is no
3358 // need to rewrite it.
3359 bool isUsedOutsideDefBB = false;
3360 BasicBlock *DefBB = GEPI->getParent();
3361 for (Value::use_iterator UI = GEPI->use_begin(), E = GEPI->use_end();
3363 if (cast<Instruction>(*UI)->getParent() != DefBB) {
3364 isUsedOutsideDefBB = true;
3368 if (!isUsedOutsideDefBB) return false;
3370 // If this GEP has no non-zero constant indices, there is nothing we can do,
3372 bool hasConstantIndex = false;
3373 bool hasVariableIndex = false;
3374 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
3375 E = GEPI->op_end(); OI != E; ++OI) {
3376 if (ConstantInt *CI = dyn_cast<ConstantInt>(*OI)) {
3377 if (CI->getZExtValue()) {
3378 hasConstantIndex = true;
3382 hasVariableIndex = true;
3386 // If this is a "GEP X, 0, 0, 0", turn this into a cast.
3387 if (!hasConstantIndex && !hasVariableIndex) {
3388 Value *NC = new CastInst(GEPI->getOperand(0), GEPI->getType(),
3389 GEPI->getName(), GEPI);
3390 GEPI->replaceAllUsesWith(NC);
3391 GEPI->eraseFromParent();
3395 // If this is a GEP &Alloca, 0, 0, forward subst the frame index into uses.
3396 if (!hasConstantIndex && !isa<AllocaInst>(GEPI->getOperand(0)))
3399 // Otherwise, decompose the GEP instruction into multiplies and adds. Sum the
3400 // constant offset (which we now know is non-zero) and deal with it later.
3401 uint64_t ConstantOffset = 0;
3402 const Type *UIntPtrTy = TD->getIntPtrType();
3403 Value *Ptr = new CastInst(GEPI->getOperand(0), UIntPtrTy, "", GEPI);
3404 const Type *Ty = GEPI->getOperand(0)->getType();
3406 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
3407 E = GEPI->op_end(); OI != E; ++OI) {
3409 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
3410 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3412 ConstantOffset += TD->getStructLayout(StTy)->MemberOffsets[Field];
3413 Ty = StTy->getElementType(Field);
3415 Ty = cast<SequentialType>(Ty)->getElementType();
3417 // Handle constant subscripts.
3418 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3419 if (CI->getZExtValue() == 0) continue;
3420 if (CI->getType()->isSigned())
3421 ConstantOffset += (int64_t)TD->getTypeSize(Ty)*CI->getSExtValue();
3423 ConstantOffset += TD->getTypeSize(Ty)*CI->getZExtValue();
3427 // Ptr = Ptr + Idx * ElementSize;
3429 // Cast Idx to UIntPtrTy if needed.
3430 Idx = new CastInst(Idx, UIntPtrTy, "", GEPI);
3432 uint64_t ElementSize = TD->getTypeSize(Ty);
3433 // Mask off bits that should not be set.
3434 ElementSize &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
3435 Constant *SizeCst = ConstantInt::get(UIntPtrTy, ElementSize);
3437 // Multiply by the element size and add to the base.
3438 Idx = BinaryOperator::createMul(Idx, SizeCst, "", GEPI);
3439 Ptr = BinaryOperator::createAdd(Ptr, Idx, "", GEPI);
3443 // Make sure that the offset fits in uintptr_t.
3444 ConstantOffset &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
3445 Constant *PtrOffset = ConstantInt::get(UIntPtrTy, ConstantOffset);
3447 // Okay, we have now emitted all of the variable index parts to the BB that
3448 // the GEP is defined in. Loop over all of the using instructions, inserting
3449 // an "add Ptr, ConstantOffset" into each block that uses it and update the
3450 // instruction to use the newly computed value, making GEPI dead. When the
3451 // user is a load or store instruction address, we emit the add into the user
3452 // block, otherwise we use a canonical version right next to the gep (these
3453 // won't be foldable as addresses, so we might as well share the computation).
3455 std::map<BasicBlock*,Instruction*> InsertedExprs;
3456 ReplaceUsesOfGEPInst(GEPI, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs);
3458 // Finally, the GEP is dead, remove it.
3459 GEPI->eraseFromParent();
3464 /// SplitCritEdgesForPHIConstants - If this block has any PHI nodes with
3465 /// constant operands, and if any of the edges feeding the PHI node are
3466 /// critical, split them so that the assignments of a constant to a register
3467 /// will not be executed on a path that isn't relevant.
3468 void SelectionDAGISel::SplitCritEdgesForPHIConstants(BasicBlock *BB) {
3469 // The most common case is that this is a PHI node with two incoming
3470 // successors handle this case efficiently, because it is simple.
3471 PHINode *PN = cast<PHINode>(BB->begin());
3472 if (PN->getNumIncomingValues() == 2) {
3473 // If neither edge is critical, we never need to split.
3474 if (PN->getIncomingBlock(0)->getTerminator()->getNumSuccessors() == 1 &&
3475 PN->getIncomingBlock(1)->getTerminator()->getNumSuccessors() == 1)
3478 BasicBlock::iterator BBI = BB->begin();
3479 while ((PN = dyn_cast<PHINode>(BBI++))) {
3480 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3481 if (isa<Constant>(PN->getIncomingValue(i)))
3482 SplitCriticalEdge(PN->getIncomingBlock(i), BB);
3487 // Otherwise, things are a bit trickier.
3491 BasicBlock::iterator BBI = BB->begin();
3492 while ((PN = dyn_cast<PHINode>(BBI++))) {
3493 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
3494 if (isa<Constant>(PN->getIncomingValue(i)))
3495 SplitCriticalEdge(PN->getIncomingBlock(i), BB);
3500 bool SelectionDAGISel::runOnFunction(Function &Fn) {
3501 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
3502 RegMap = MF.getSSARegMap();
3503 DEBUG(std::cerr << "\n\n\n=== " << Fn.getName() << "\n");
3505 // First, split all critical edges for PHI nodes with incoming values that are
3506 // constants, this way the load of the constant into a vreg will not be placed
3507 // into MBBs that are used some other way.
3509 // In this pass we also look for GEP and cast instructions that are used
3510 // across basic blocks and rewrite them to improve basic-block-at-a-time
3514 bool MadeChange = true;
3515 while (MadeChange) {
3517 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
3518 // If this block has any PHI nodes with constant operands, and if any of the
3519 // edges feeding the PHI node are critical, split them.
3520 if (isa<PHINode>(BB->begin()))
3521 SplitCritEdgesForPHIConstants(BB);
3523 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3524 Instruction *I = BBI++;
3525 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3526 MadeChange |= OptimizeGEPExpression(GEPI, TLI.getTargetData());
3527 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3528 // If the source of the cast is a constant, then this should have
3529 // already been constant folded. The only reason NOT to constant fold
3530 // it is if something (e.g. LSR) was careful to place the constant
3531 // evaluation in a block other than then one that uses it (e.g. to hoist
3532 // the address of globals out of a loop). If this is the case, we don't
3533 // want to forward-subst the cast.
3534 if (isa<Constant>(CI->getOperand(0)))
3537 // If this is a noop copy, sink it into user blocks to reduce the number
3538 // of virtual registers that must be created and coallesced.
3539 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
3540 MVT::ValueType DstVT = TLI.getValueType(CI->getType());
3542 // This is an fp<->int conversion?
3543 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT))
3546 // If this is an extension, it will be a zero or sign extension, which
3548 if (SrcVT < DstVT) continue;
3550 // If these values will be promoted, find out what they will be promoted
3551 // to. This helps us consider truncates on PPC as noop copies when they
3553 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
3554 SrcVT = TLI.getTypeToTransformTo(SrcVT);
3555 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
3556 DstVT = TLI.getTypeToTransformTo(DstVT);
3558 // If, after promotion, these are the same types, this is a noop copy.
3560 MadeChange |= OptimizeNoopCopyExpression(CI);
3566 FunctionLoweringInfo FuncInfo(TLI, Fn, MF);
3568 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
3569 SelectBasicBlock(I, MF, FuncInfo);
3574 SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V,
3576 SDOperand Op = getValue(V);
3577 assert((Op.getOpcode() != ISD::CopyFromReg ||
3578 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
3579 "Copy from a reg to the same reg!");
3581 // If this type is not legal, we must make sure to not create an invalid
3583 MVT::ValueType SrcVT = Op.getValueType();
3584 MVT::ValueType DestVT = TLI.getTypeToTransformTo(SrcVT);
3585 if (SrcVT == DestVT) {
3586 return DAG.getCopyToReg(getRoot(), Reg, Op);
3587 } else if (SrcVT == MVT::Vector) {
3588 // Handle copies from generic vectors to registers.
3589 MVT::ValueType PTyElementVT, PTyLegalElementVT;
3590 unsigned NE = TLI.getPackedTypeBreakdown(cast<PackedType>(V->getType()),
3591 PTyElementVT, PTyLegalElementVT);
3593 // Insert a VBIT_CONVERT of the input vector to a "N x PTyElementVT"
3594 // MVT::Vector type.
3595 Op = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Op,
3596 DAG.getConstant(NE, MVT::i32),
3597 DAG.getValueType(PTyElementVT));
3599 // Loop over all of the elements of the resultant vector,
3600 // VEXTRACT_VECTOR_ELT'ing them, converting them to PTyLegalElementVT, then
3601 // copying them into output registers.
3602 SmallVector<SDOperand, 8> OutChains;
3603 SDOperand Root = getRoot();
3604 for (unsigned i = 0; i != NE; ++i) {
3605 SDOperand Elt = DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, PTyElementVT,
3606 Op, DAG.getConstant(i, TLI.getPointerTy()));
3607 if (PTyElementVT == PTyLegalElementVT) {
3608 // Elements are legal.
3609 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
3610 } else if (PTyLegalElementVT > PTyElementVT) {
3611 // Elements are promoted.
3612 if (MVT::isFloatingPoint(PTyLegalElementVT))
3613 Elt = DAG.getNode(ISD::FP_EXTEND, PTyLegalElementVT, Elt);
3615 Elt = DAG.getNode(ISD::ANY_EXTEND, PTyLegalElementVT, Elt);
3616 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
3618 // Elements are expanded.
3619 // The src value is expanded into multiple registers.
3620 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
3621 Elt, DAG.getConstant(0, TLI.getPointerTy()));
3622 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
3623 Elt, DAG.getConstant(1, TLI.getPointerTy()));
3624 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Lo));
3625 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Hi));
3628 return DAG.getNode(ISD::TokenFactor, MVT::Other,
3629 &OutChains[0], OutChains.size());
3630 } else if (SrcVT < DestVT) {
3631 // The src value is promoted to the register.
3632 if (MVT::isFloatingPoint(SrcVT))
3633 Op = DAG.getNode(ISD::FP_EXTEND, DestVT, Op);
3635 Op = DAG.getNode(ISD::ANY_EXTEND, DestVT, Op);
3636 return DAG.getCopyToReg(getRoot(), Reg, Op);
3638 // The src value is expanded into multiple registers.
3639 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
3640 Op, DAG.getConstant(0, TLI.getPointerTy()));
3641 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
3642 Op, DAG.getConstant(1, TLI.getPointerTy()));
3643 Op = DAG.getCopyToReg(getRoot(), Reg, Lo);
3644 return DAG.getCopyToReg(Op, Reg+1, Hi);
3648 void SelectionDAGISel::
3649 LowerArguments(BasicBlock *BB, SelectionDAGLowering &SDL,
3650 std::vector<SDOperand> &UnorderedChains) {
3651 // If this is the entry block, emit arguments.
3652 Function &F = *BB->getParent();
3653 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo;
3654 SDOperand OldRoot = SDL.DAG.getRoot();
3655 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG);
3658 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
3660 if (!AI->use_empty()) {
3661 SDL.setValue(AI, Args[a]);
3663 // If this argument is live outside of the entry block, insert a copy from
3664 // whereever we got it to the vreg that other BB's will reference it as.
3665 if (FuncInfo.ValueMap.count(AI)) {
3667 SDL.CopyValueToVirtualRegister(AI, FuncInfo.ValueMap[AI]);
3668 UnorderedChains.push_back(Copy);
3672 // Finally, if the target has anything special to do, allow it to do so.
3673 // FIXME: this should insert code into the DAG!
3674 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction());
3677 void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB,
3678 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate,
3679 FunctionLoweringInfo &FuncInfo) {
3680 SelectionDAGLowering SDL(DAG, TLI, FuncInfo);
3682 std::vector<SDOperand> UnorderedChains;
3684 // Lower any arguments needed in this block if this is the entry block.
3685 if (LLVMBB == &LLVMBB->getParent()->front())
3686 LowerArguments(LLVMBB, SDL, UnorderedChains);
3688 BB = FuncInfo.MBBMap[LLVMBB];
3689 SDL.setCurrentBasicBlock(BB);
3691 // Lower all of the non-terminator instructions.
3692 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end();
3696 // Ensure that all instructions which are used outside of their defining
3697 // blocks are available as virtual registers.
3698 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I)
3699 if (!I->use_empty() && !isa<PHINode>(I)) {
3700 std::map<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I);
3701 if (VMI != FuncInfo.ValueMap.end())
3702 UnorderedChains.push_back(
3703 SDL.CopyValueToVirtualRegister(I, VMI->second));
3706 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
3707 // ensure constants are generated when needed. Remember the virtual registers
3708 // that need to be added to the Machine PHI nodes as input. We cannot just
3709 // directly add them, because expansion might result in multiple MBB's for one
3710 // BB. As such, the start of the BB might correspond to a different MBB than
3714 // Emit constants only once even if used by multiple PHI nodes.
3715 std::map<Constant*, unsigned> ConstantsOut;
3717 // Check successor nodes PHI nodes that expect a constant to be available from
3719 TerminatorInst *TI = LLVMBB->getTerminator();
3720 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
3721 BasicBlock *SuccBB = TI->getSuccessor(succ);
3722 if (!isa<PHINode>(SuccBB->begin())) continue;
3724 MachineBasicBlock::iterator MBBI = FuncInfo.MBBMap[SuccBB]->begin();
3727 // At this point we know that there is a 1-1 correspondence between LLVM PHI
3728 // nodes and Machine PHI nodes, but the incoming operands have not been
3730 for (BasicBlock::iterator I = SuccBB->begin();
3731 (PN = dyn_cast<PHINode>(I)); ++I)
3732 if (!PN->use_empty()) {
3734 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
3735 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
3736 unsigned &RegOut = ConstantsOut[C];
3738 RegOut = FuncInfo.CreateRegForValue(C);
3739 UnorderedChains.push_back(
3740 SDL.CopyValueToVirtualRegister(C, RegOut));
3744 Reg = FuncInfo.ValueMap[PHIOp];
3746 assert(isa<AllocaInst>(PHIOp) &&
3747 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
3748 "Didn't codegen value into a register!??");
3749 Reg = FuncInfo.CreateRegForValue(PHIOp);
3750 UnorderedChains.push_back(
3751 SDL.CopyValueToVirtualRegister(PHIOp, Reg));
3755 // Remember that this register needs to added to the machine PHI node as
3756 // the input for this MBB.
3757 MVT::ValueType VT = TLI.getValueType(PN->getType());
3758 unsigned NumElements;
3759 if (VT != MVT::Vector)
3760 NumElements = TLI.getNumElements(VT);
3762 MVT::ValueType VT1,VT2;
3764 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
3767 for (unsigned i = 0, e = NumElements; i != e; ++i)
3768 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
3771 ConstantsOut.clear();
3773 // Turn all of the unordered chains into one factored node.
3774 if (!UnorderedChains.empty()) {
3775 SDOperand Root = SDL.getRoot();
3776 if (Root.getOpcode() != ISD::EntryToken) {
3777 unsigned i = 0, e = UnorderedChains.size();
3778 for (; i != e; ++i) {
3779 assert(UnorderedChains[i].Val->getNumOperands() > 1);
3780 if (UnorderedChains[i].Val->getOperand(0) == Root)
3781 break; // Don't add the root if we already indirectly depend on it.
3785 UnorderedChains.push_back(Root);
3787 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
3788 &UnorderedChains[0], UnorderedChains.size()));
3791 // Lower the terminator after the copies are emitted.
3792 SDL.visit(*LLVMBB->getTerminator());
3794 // Copy over any CaseBlock records that may now exist due to SwitchInst
3795 // lowering, as well as any jump table information.
3796 SwitchCases.clear();
3797 SwitchCases = SDL.SwitchCases;
3800 // Make sure the root of the DAG is up-to-date.
3801 DAG.setRoot(SDL.getRoot());
3804 void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) {
3805 // Get alias analysis for load/store combining.
3806 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
3808 // Run the DAG combiner in pre-legalize mode.
3809 DAG.Combine(false, AA);
3811 DEBUG(std::cerr << "Lowered selection DAG:\n");
3814 // Second step, hack on the DAG until it only uses operations and types that
3815 // the target supports.
3818 DEBUG(std::cerr << "Legalized selection DAG:\n");
3821 // Run the DAG combiner in post-legalize mode.
3822 DAG.Combine(true, AA);
3824 if (ViewISelDAGs) DAG.viewGraph();
3826 // Third, instruction select all of the operations to machine code, adding the
3827 // code to the MachineBasicBlock.
3828 InstructionSelectBasicBlock(DAG);
3830 DEBUG(std::cerr << "Selected machine code:\n");
3834 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF,
3835 FunctionLoweringInfo &FuncInfo) {
3836 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate;
3838 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3841 // First step, lower LLVM code to some DAG. This DAG may use operations and
3842 // types that are not supported by the target.
3843 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo);
3845 // Second step, emit the lowered DAG as machine code.
3846 CodeGenAndEmitDAG(DAG);
3849 // Next, now that we know what the last MBB the LLVM BB expanded is, update
3850 // PHI nodes in successors.
3851 if (SwitchCases.empty() && JT.Reg == 0) {
3852 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
3853 MachineInstr *PHI = PHINodesToUpdate[i].first;
3854 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3855 "This is not a machine PHI node that we are updating!");
3856 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
3857 PHI->addMachineBasicBlockOperand(BB);
3862 // If the JumpTable record is filled in, then we need to emit a jump table.
3863 // Updating the PHI nodes is tricky in this case, since we need to determine
3864 // whether the PHI is a successor of the range check MBB or the jump table MBB
3866 assert(SwitchCases.empty() && "Cannot have jump table and lowered switch");
3867 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3869 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
3870 MachineBasicBlock *RangeBB = BB;
3871 // Set the current basic block to the mbb we wish to insert the code into
3873 SDL.setCurrentBasicBlock(BB);
3875 SDL.visitJumpTable(JT);
3876 SDAG.setRoot(SDL.getRoot());
3877 CodeGenAndEmitDAG(SDAG);
3879 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
3880 MachineInstr *PHI = PHINodesToUpdate[pi].first;
3881 MachineBasicBlock *PHIBB = PHI->getParent();
3882 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3883 "This is not a machine PHI node that we are updating!");
3884 if (PHIBB == JT.Default) {
3885 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
3886 PHI->addMachineBasicBlockOperand(RangeBB);
3888 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
3889 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
3890 PHI->addMachineBasicBlockOperand(BB);
3896 // If the switch block involved a branch to one of the actual successors, we
3897 // need to update PHI nodes in that block.
3898 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
3899 MachineInstr *PHI = PHINodesToUpdate[i].first;
3900 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3901 "This is not a machine PHI node that we are updating!");
3902 if (BB->isSuccessor(PHI->getParent())) {
3903 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
3904 PHI->addMachineBasicBlockOperand(BB);
3908 // If we generated any switch lowering information, build and codegen any
3909 // additional DAGs necessary.
3910 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) {
3911 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3913 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
3915 // Set the current basic block to the mbb we wish to insert the code into
3916 BB = SwitchCases[i].ThisBB;
3917 SDL.setCurrentBasicBlock(BB);
3920 SDL.visitSwitchCase(SwitchCases[i]);
3921 SDAG.setRoot(SDL.getRoot());
3922 CodeGenAndEmitDAG(SDAG);
3924 // Handle any PHI nodes in successors of this chunk, as if we were coming
3925 // from the original BB before switch expansion. Note that PHI nodes can
3926 // occur multiple times in PHINodesToUpdate. We have to be very careful to
3927 // handle them the right number of times.
3928 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
3929 for (MachineBasicBlock::iterator Phi = BB->begin();
3930 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
3931 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
3932 for (unsigned pn = 0; ; ++pn) {
3933 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!");
3934 if (PHINodesToUpdate[pn].first == Phi) {
3935 Phi->addRegOperand(PHINodesToUpdate[pn].second, false);
3936 Phi->addMachineBasicBlockOperand(SwitchCases[i].ThisBB);
3942 // Don't process RHS if same block as LHS.
3943 if (BB == SwitchCases[i].FalseBB)
3944 SwitchCases[i].FalseBB = 0;
3946 // If we haven't handled the RHS, do so now. Otherwise, we're done.
3947 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB;
3948 SwitchCases[i].FalseBB = 0;
3950 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0);
3955 //===----------------------------------------------------------------------===//
3956 /// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
3957 /// target node in the graph.
3958 void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) {
3959 if (ViewSchedDAGs) DAG.viewGraph();
3961 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
3965 RegisterScheduler::setDefault(Ctor);
3968 ScheduleDAG *SL = Ctor(this, &DAG, BB);
3974 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
3975 return new HazardRecognizer();
3978 //===----------------------------------------------------------------------===//
3979 // Helper functions used by the generated instruction selector.
3980 //===----------------------------------------------------------------------===//
3981 // Calls to these methods are generated by tblgen.
3983 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
3984 /// the dag combiner simplified the 255, we still want to match. RHS is the
3985 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
3986 /// specified in the .td file (e.g. 255).
3987 bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS,
3988 int64_t DesiredMaskS) {
3989 uint64_t ActualMask = RHS->getValue();
3990 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
3992 // If the actual mask exactly matches, success!
3993 if (ActualMask == DesiredMask)
3996 // If the actual AND mask is allowing unallowed bits, this doesn't match.
3997 if (ActualMask & ~DesiredMask)
4000 // Otherwise, the DAG Combiner may have proven that the value coming in is
4001 // either already zero or is not demanded. Check for known zero input bits.
4002 uint64_t NeededMask = DesiredMask & ~ActualMask;
4003 if (getTargetLowering().MaskedValueIsZero(LHS, NeededMask))
4006 // TODO: check to see if missing bits are just not demanded.
4008 // Otherwise, this pattern doesn't match.
4012 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
4013 /// the dag combiner simplified the 255, we still want to match. RHS is the
4014 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
4015 /// specified in the .td file (e.g. 255).
4016 bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS,
4017 int64_t DesiredMaskS) {
4018 uint64_t ActualMask = RHS->getValue();
4019 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
4021 // If the actual mask exactly matches, success!
4022 if (ActualMask == DesiredMask)
4025 // If the actual AND mask is allowing unallowed bits, this doesn't match.
4026 if (ActualMask & ~DesiredMask)
4029 // Otherwise, the DAG Combiner may have proven that the value coming in is
4030 // either already zero or is not demanded. Check for known zero input bits.
4031 uint64_t NeededMask = DesiredMask & ~ActualMask;
4033 uint64_t KnownZero, KnownOne;
4034 getTargetLowering().ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
4036 // If all the missing bits in the or are already known to be set, match!
4037 if ((NeededMask & KnownOne) == NeededMask)
4040 // TODO: check to see if missing bits are just not demanded.
4042 // Otherwise, this pattern doesn't match.
4047 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
4048 /// by tblgen. Others should not call it.
4049 void SelectionDAGISel::
4050 SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) {
4051 std::vector<SDOperand> InOps;
4052 std::swap(InOps, Ops);
4054 Ops.push_back(InOps[0]); // input chain.
4055 Ops.push_back(InOps[1]); // input asm string.
4057 unsigned i = 2, e = InOps.size();
4058 if (InOps[e-1].getValueType() == MVT::Flag)
4059 --e; // Don't process a flag operand if it is here.
4062 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
4063 if ((Flags & 7) != 4 /*MEM*/) {
4064 // Just skip over this operand, copying the operands verbatim.
4065 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
4066 i += (Flags >> 3) + 1;
4068 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
4069 // Otherwise, this is a memory operand. Ask the target to select it.
4070 std::vector<SDOperand> SelOps;
4071 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) {
4072 std::cerr << "Could not match memory address. Inline asm failure!\n";
4076 // Add this to the output node.
4077 Ops.push_back(DAG.getConstant(4/*MEM*/ | (SelOps.size() << 3), MVT::i32));
4078 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
4083 // Add the flag input back if present.
4084 if (e != InOps.size())
4085 Ops.push_back(InOps.back());