1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "llvm/ADT/BitVector.h"
16 #include "llvm/Analysis/AliasAnalysis.h"
17 #include "llvm/CodeGen/SelectionDAGISel.h"
18 #include "llvm/CodeGen/ScheduleDAG.h"
19 #include "llvm/Constants.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/InlineAsm.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/ParameterAttributes.h"
29 #include "llvm/CodeGen/Collector.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/SchedulerRegistry.h"
37 #include "llvm/CodeGen/SelectionDAG.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetData.h"
40 #include "llvm/Target/TargetFrameInfo.h"
41 #include "llvm/Target/TargetInstrInfo.h"
42 #include "llvm/Target/TargetLowering.h"
43 #include "llvm/Target/TargetMachine.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/Support/MathExtras.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/Compiler.h"
53 ViewISelDAGs("view-isel-dags", cl::Hidden,
54 cl::desc("Pop up a window to show isel dags as they are selected"));
56 ViewSchedDAGs("view-sched-dags", cl::Hidden,
57 cl::desc("Pop up a window to show sched dags as they are processed"));
59 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
60 cl::desc("Pop up a window to show SUnit dags after they are processed"));
62 static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0, ViewSUnitDAGs = 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> >
80 ISHeuristic("pre-RA-sched",
81 cl::init(&createDefaultScheduler),
82 cl::desc("Instruction schedulers available (before register"
85 static RegisterScheduler
86 defaultListDAGScheduler("default", " Best scheduler for the target",
87 createDefaultScheduler);
90 namespace { struct SDISelAsmOperandInfo; }
93 /// RegsForValue - This struct represents the physical registers that a
94 /// particular value is assigned and the type information about the value.
95 /// This is needed because values can be promoted into larger registers and
96 /// expanded into multiple smaller registers than the value.
97 struct VISIBILITY_HIDDEN RegsForValue {
98 /// Regs - This list holds the register (for legal and promoted values)
99 /// or register set (for expanded values) that the value should be assigned
101 std::vector<unsigned> Regs;
103 /// RegVT - The value type of each register.
105 MVT::ValueType RegVT;
107 /// ValueVT - The value type of the LLVM value, which may be promoted from
108 /// RegVT or made from merging the two expanded parts.
109 MVT::ValueType ValueVT;
111 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {}
113 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
114 : RegVT(regvt), ValueVT(valuevt) {
117 RegsForValue(const std::vector<unsigned> ®s,
118 MVT::ValueType regvt, MVT::ValueType valuevt)
119 : Regs(regs), RegVT(regvt), ValueVT(valuevt) {
122 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
123 /// this value and returns the result as a ValueVT value. This uses
124 /// Chain/Flag as the input and updates them for the output Chain/Flag.
125 /// If the Flag pointer is NULL, no flag is used.
126 SDOperand getCopyFromRegs(SelectionDAG &DAG,
127 SDOperand &Chain, SDOperand *Flag) const;
129 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
130 /// specified value into the registers specified by this object. This uses
131 /// Chain/Flag as the input and updates them for the output Chain/Flag.
132 /// If the Flag pointer is NULL, no flag is used.
133 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
134 SDOperand &Chain, SDOperand *Flag) const;
136 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
137 /// operand list. This adds the code marker and includes the number of
138 /// values added into it.
139 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
140 std::vector<SDOperand> &Ops) const;
145 //===--------------------------------------------------------------------===//
146 /// createDefaultScheduler - This creates an instruction scheduler appropriate
148 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
150 MachineBasicBlock *BB) {
151 TargetLowering &TLI = IS->getTargetLowering();
153 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
154 return createTDListDAGScheduler(IS, DAG, BB);
156 assert(TLI.getSchedulingPreference() ==
157 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
158 return createBURRListDAGScheduler(IS, DAG, BB);
163 //===--------------------------------------------------------------------===//
164 /// FunctionLoweringInfo - This contains information that is global to a
165 /// function that is used when lowering a region of the function.
166 class FunctionLoweringInfo {
171 MachineRegisterInfo &RegInfo;
173 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF);
175 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry.
176 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
178 /// ValueMap - Since we emit code for the function a basic block at a time,
179 /// we must remember which virtual registers hold the values for
180 /// cross-basic-block values.
181 DenseMap<const Value*, unsigned> ValueMap;
183 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in
184 /// the entry block. This allows the allocas to be efficiently referenced
185 /// anywhere in the function.
186 std::map<const AllocaInst*, int> StaticAllocaMap;
189 SmallSet<Instruction*, 8> CatchInfoLost;
190 SmallSet<Instruction*, 8> CatchInfoFound;
193 unsigned MakeReg(MVT::ValueType VT) {
194 return RegInfo.createVirtualRegister(TLI.getRegClassFor(VT));
197 /// isExportedInst - Return true if the specified value is an instruction
198 /// exported from its block.
199 bool isExportedInst(const Value *V) {
200 return ValueMap.count(V);
203 unsigned CreateRegForValue(const Value *V);
205 unsigned InitializeRegForValue(const Value *V) {
206 unsigned &R = ValueMap[V];
207 assert(R == 0 && "Already initialized this value register!");
208 return R = CreateRegForValue(V);
213 /// isSelector - Return true if this instruction is a call to the
214 /// eh.selector intrinsic.
215 static bool isSelector(Instruction *I) {
216 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
217 return (II->getIntrinsicID() == Intrinsic::eh_selector_i32 ||
218 II->getIntrinsicID() == Intrinsic::eh_selector_i64);
222 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
223 /// PHI nodes or outside of the basic block that defines it, or used by a
224 /// switch or atomic instruction, which may expand to multiple basic blocks.
225 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
226 if (isa<PHINode>(I)) return true;
227 BasicBlock *BB = I->getParent();
228 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
229 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
230 // FIXME: Remove switchinst special case.
231 isa<SwitchInst>(*UI))
236 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
237 /// entry block, return true. This includes arguments used by switches, since
238 /// the switch may expand into multiple basic blocks.
239 static bool isOnlyUsedInEntryBlock(Argument *A) {
240 BasicBlock *Entry = A->getParent()->begin();
241 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
242 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
243 return false; // Use not in entry block.
247 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli,
248 Function &fn, MachineFunction &mf)
249 : TLI(tli), Fn(fn), MF(mf), RegInfo(MF.getRegInfo()) {
251 // Create a vreg for each argument register that is not dead and is used
252 // outside of the entry block for the function.
253 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end();
255 if (!isOnlyUsedInEntryBlock(AI))
256 InitializeRegForValue(AI);
258 // Initialize the mapping of values to registers. This is only set up for
259 // instruction values that are used outside of the block that defines
261 Function::iterator BB = Fn.begin(), EB = Fn.end();
262 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
263 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
264 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
265 const Type *Ty = AI->getAllocatedType();
266 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
268 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
271 TySize *= CUI->getZExtValue(); // Get total allocated size.
272 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
273 StaticAllocaMap[AI] =
274 MF.getFrameInfo()->CreateStackObject(TySize, Align);
277 for (; BB != EB; ++BB)
278 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
279 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
280 if (!isa<AllocaInst>(I) ||
281 !StaticAllocaMap.count(cast<AllocaInst>(I)))
282 InitializeRegForValue(I);
284 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
285 // also creates the initial PHI MachineInstrs, though none of the input
286 // operands are populated.
287 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) {
288 MachineBasicBlock *MBB = new MachineBasicBlock(BB);
290 MF.getBasicBlockList().push_back(MBB);
292 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
295 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){
296 if (PN->use_empty()) continue;
298 MVT::ValueType VT = TLI.getValueType(PN->getType());
299 unsigned NumRegisters = TLI.getNumRegisters(VT);
300 unsigned PHIReg = ValueMap[PN];
301 assert(PHIReg && "PHI node does not have an assigned virtual register!");
302 const TargetInstrInfo *TII = TLI.getTargetMachine().getInstrInfo();
303 for (unsigned i = 0; i != NumRegisters; ++i)
304 BuildMI(MBB, TII->get(TargetInstrInfo::PHI), PHIReg+i);
309 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
310 /// the correctly promoted or expanded types. Assign these registers
311 /// consecutive vreg numbers and return the first assigned number.
312 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
313 MVT::ValueType VT = TLI.getValueType(V->getType());
315 unsigned NumRegisters = TLI.getNumRegisters(VT);
316 MVT::ValueType RegisterVT = TLI.getRegisterType(VT);
318 unsigned R = MakeReg(RegisterVT);
319 for (unsigned i = 1; i != NumRegisters; ++i)
325 //===----------------------------------------------------------------------===//
326 /// SelectionDAGLowering - This is the common target-independent lowering
327 /// implementation that is parameterized by a TargetLowering object.
328 /// Also, targets can overload any lowering method.
331 class SelectionDAGLowering {
332 MachineBasicBlock *CurMBB;
334 DenseMap<const Value*, SDOperand> NodeMap;
336 /// PendingLoads - Loads are not emitted to the program immediately. We bunch
337 /// them up and then emit token factor nodes when possible. This allows us to
338 /// get simple disambiguation between loads without worrying about alias
340 std::vector<SDOperand> PendingLoads;
342 /// Case - A struct to record the Value for a switch case, and the
343 /// case's target basic block.
347 MachineBasicBlock* BB;
349 Case() : Low(0), High(0), BB(0) { }
350 Case(Constant* low, Constant* high, MachineBasicBlock* bb) :
351 Low(low), High(high), BB(bb) { }
352 uint64_t size() const {
353 uint64_t rHigh = cast<ConstantInt>(High)->getSExtValue();
354 uint64_t rLow = cast<ConstantInt>(Low)->getSExtValue();
355 return (rHigh - rLow + 1ULL);
361 MachineBasicBlock* BB;
364 CaseBits(uint64_t mask, MachineBasicBlock* bb, unsigned bits):
365 Mask(mask), BB(bb), Bits(bits) { }
368 typedef std::vector<Case> CaseVector;
369 typedef std::vector<CaseBits> CaseBitsVector;
370 typedef CaseVector::iterator CaseItr;
371 typedef std::pair<CaseItr, CaseItr> CaseRange;
373 /// CaseRec - A struct with ctor used in lowering switches to a binary tree
374 /// of conditional branches.
376 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) :
377 CaseBB(bb), LT(lt), GE(ge), Range(r) {}
379 /// CaseBB - The MBB in which to emit the compare and branch
380 MachineBasicBlock *CaseBB;
381 /// LT, GE - If nonzero, we know the current case value must be less-than or
382 /// greater-than-or-equal-to these Constants.
385 /// Range - A pair of iterators representing the range of case values to be
386 /// processed at this point in the binary search tree.
390 typedef std::vector<CaseRec> CaseRecVector;
392 /// The comparison function for sorting the switch case values in the vector.
393 /// WARNING: Case ranges should be disjoint!
395 bool operator () (const Case& C1, const Case& C2) {
396 assert(isa<ConstantInt>(C1.Low) && isa<ConstantInt>(C2.High));
397 const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low);
398 const ConstantInt* CI2 = cast<const ConstantInt>(C2.High);
399 return CI1->getValue().slt(CI2->getValue());
404 bool operator () (const CaseBits& C1, const CaseBits& C2) {
405 return C1.Bits > C2.Bits;
409 unsigned Clusterify(CaseVector& Cases, const SwitchInst &SI);
412 // TLI - This is information that describes the available target features we
413 // need for lowering. This indicates when operations are unavailable,
414 // implemented with a libcall, etc.
417 const TargetData *TD;
420 /// SwitchCases - Vector of CaseBlock structures used to communicate
421 /// SwitchInst code generation information.
422 std::vector<SelectionDAGISel::CaseBlock> SwitchCases;
423 /// JTCases - Vector of JumpTable structures used to communicate
424 /// SwitchInst code generation information.
425 std::vector<SelectionDAGISel::JumpTableBlock> JTCases;
426 std::vector<SelectionDAGISel::BitTestBlock> BitTestCases;
428 /// FuncInfo - Information about the function as a whole.
430 FunctionLoweringInfo &FuncInfo;
432 /// GCI - Garbage collection metadata for the function.
433 CollectorMetadata *GCI;
435 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli,
437 FunctionLoweringInfo &funcinfo,
438 CollectorMetadata *gci)
439 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()), AA(aa),
440 FuncInfo(funcinfo), GCI(gci) {
443 /// getRoot - Return the current virtual root of the Selection DAG.
445 SDOperand getRoot() {
446 if (PendingLoads.empty())
447 return DAG.getRoot();
449 if (PendingLoads.size() == 1) {
450 SDOperand Root = PendingLoads[0];
452 PendingLoads.clear();
456 // Otherwise, we have to make a token factor node.
457 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
458 &PendingLoads[0], PendingLoads.size());
459 PendingLoads.clear();
464 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg);
466 void visit(Instruction &I) { visit(I.getOpcode(), I); }
468 void visit(unsigned Opcode, User &I) {
469 // Note: this doesn't use InstVisitor, because it has to work with
470 // ConstantExpr's in addition to instructions.
472 default: assert(0 && "Unknown instruction type encountered!");
474 // Build the switch statement using the Instruction.def file.
475 #define HANDLE_INST(NUM, OPCODE, CLASS) \
476 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
477 #include "llvm/Instruction.def"
481 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; }
483 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr,
484 const Value *SV, SDOperand Root,
485 bool isVolatile, unsigned Alignment);
487 SDOperand getValue(const Value *V);
489 void setValue(const Value *V, SDOperand NewN) {
490 SDOperand &N = NodeMap[V];
491 assert(N.Val == 0 && "Already set a value for this node!");
495 void GetRegistersForValue(SDISelAsmOperandInfo &OpInfo, bool HasEarlyClobber,
496 std::set<unsigned> &OutputRegs,
497 std::set<unsigned> &InputRegs);
499 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB,
500 MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
502 bool isExportableFromCurrentBlock(Value *V, const BasicBlock *FromBB);
503 void ExportFromCurrentBlock(Value *V);
504 void LowerCallTo(CallSite CS, SDOperand Callee, bool IsTailCall,
505 MachineBasicBlock *LandingPad = NULL);
507 // Terminator instructions.
508 void visitRet(ReturnInst &I);
509 void visitBr(BranchInst &I);
510 void visitSwitch(SwitchInst &I);
511 void visitUnreachable(UnreachableInst &I) { /* noop */ }
513 // Helpers for visitSwitch
514 bool handleSmallSwitchRange(CaseRec& CR,
515 CaseRecVector& WorkList,
517 MachineBasicBlock* Default);
518 bool handleJTSwitchCase(CaseRec& CR,
519 CaseRecVector& WorkList,
521 MachineBasicBlock* Default);
522 bool handleBTSplitSwitchCase(CaseRec& CR,
523 CaseRecVector& WorkList,
525 MachineBasicBlock* Default);
526 bool handleBitTestsSwitchCase(CaseRec& CR,
527 CaseRecVector& WorkList,
529 MachineBasicBlock* Default);
530 void visitSwitchCase(SelectionDAGISel::CaseBlock &CB);
531 void visitBitTestHeader(SelectionDAGISel::BitTestBlock &B);
532 void visitBitTestCase(MachineBasicBlock* NextMBB,
534 SelectionDAGISel::BitTestCase &B);
535 void visitJumpTable(SelectionDAGISel::JumpTable &JT);
536 void visitJumpTableHeader(SelectionDAGISel::JumpTable &JT,
537 SelectionDAGISel::JumpTableHeader &JTH);
539 // These all get lowered before this pass.
540 void visitInvoke(InvokeInst &I);
541 void visitUnwind(UnwindInst &I);
543 void visitBinary(User &I, unsigned OpCode);
544 void visitShift(User &I, unsigned Opcode);
545 void visitAdd(User &I) {
546 if (I.getType()->isFPOrFPVector())
547 visitBinary(I, ISD::FADD);
549 visitBinary(I, ISD::ADD);
551 void visitSub(User &I);
552 void visitMul(User &I) {
553 if (I.getType()->isFPOrFPVector())
554 visitBinary(I, ISD::FMUL);
556 visitBinary(I, ISD::MUL);
558 void visitURem(User &I) { visitBinary(I, ISD::UREM); }
559 void visitSRem(User &I) { visitBinary(I, ISD::SREM); }
560 void visitFRem(User &I) { visitBinary(I, ISD::FREM); }
561 void visitUDiv(User &I) { visitBinary(I, ISD::UDIV); }
562 void visitSDiv(User &I) { visitBinary(I, ISD::SDIV); }
563 void visitFDiv(User &I) { visitBinary(I, ISD::FDIV); }
564 void visitAnd (User &I) { visitBinary(I, ISD::AND); }
565 void visitOr (User &I) { visitBinary(I, ISD::OR); }
566 void visitXor (User &I) { visitBinary(I, ISD::XOR); }
567 void visitShl (User &I) { visitShift(I, ISD::SHL); }
568 void visitLShr(User &I) { visitShift(I, ISD::SRL); }
569 void visitAShr(User &I) { visitShift(I, ISD::SRA); }
570 void visitICmp(User &I);
571 void visitFCmp(User &I);
572 // Visit the conversion instructions
573 void visitTrunc(User &I);
574 void visitZExt(User &I);
575 void visitSExt(User &I);
576 void visitFPTrunc(User &I);
577 void visitFPExt(User &I);
578 void visitFPToUI(User &I);
579 void visitFPToSI(User &I);
580 void visitUIToFP(User &I);
581 void visitSIToFP(User &I);
582 void visitPtrToInt(User &I);
583 void visitIntToPtr(User &I);
584 void visitBitCast(User &I);
586 void visitExtractElement(User &I);
587 void visitInsertElement(User &I);
588 void visitShuffleVector(User &I);
590 void visitGetElementPtr(User &I);
591 void visitSelect(User &I);
593 void visitMalloc(MallocInst &I);
594 void visitFree(FreeInst &I);
595 void visitAlloca(AllocaInst &I);
596 void visitLoad(LoadInst &I);
597 void visitStore(StoreInst &I);
598 void visitPHI(PHINode &I) { } // PHI nodes are handled specially.
599 void visitCall(CallInst &I);
600 void visitInlineAsm(CallSite CS);
601 const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic);
602 void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic);
604 void visitVAStart(CallInst &I);
605 void visitVAArg(VAArgInst &I);
606 void visitVAEnd(CallInst &I);
607 void visitVACopy(CallInst &I);
609 void visitMemIntrinsic(CallInst &I, unsigned Op);
611 void visitGetResult(GetResultInst &I) {
612 assert (0 && "getresult unimplemented");
615 void visitUserOp1(Instruction &I) {
616 assert(0 && "UserOp1 should not exist at instruction selection time!");
619 void visitUserOp2(Instruction &I) {
620 assert(0 && "UserOp2 should not exist at instruction selection time!");
624 } // end namespace llvm
627 /// getCopyFromParts - Create a value that contains the specified legal parts
628 /// combined into the value they represent. If the parts combine to a type
629 /// larger then ValueVT then AssertOp can be used to specify whether the extra
630 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
631 /// (ISD::AssertSext). Likewise TruncExact is used for floating point types to
632 /// indicate that the extra bits can be discarded without losing precision.
633 static SDOperand getCopyFromParts(SelectionDAG &DAG,
634 const SDOperand *Parts,
636 MVT::ValueType PartVT,
637 MVT::ValueType ValueVT,
638 ISD::NodeType AssertOp = ISD::DELETED_NODE,
639 bool TruncExact = false) {
640 assert(NumParts > 0 && "No parts to assemble!");
641 TargetLowering &TLI = DAG.getTargetLoweringInfo();
642 SDOperand Val = Parts[0];
645 // Assemble the value from multiple parts.
646 if (!MVT::isVector(ValueVT)) {
647 unsigned PartBits = MVT::getSizeInBits(PartVT);
648 unsigned ValueBits = MVT::getSizeInBits(ValueVT);
650 // Assemble the power of 2 part.
651 unsigned RoundParts = NumParts & (NumParts - 1) ?
652 1 << Log2_32(NumParts) : NumParts;
653 unsigned RoundBits = PartBits * RoundParts;
654 MVT::ValueType RoundVT = RoundBits == ValueBits ?
655 ValueVT : MVT::getIntegerType(RoundBits);
658 if (RoundParts > 2) {
659 MVT::ValueType HalfVT = MVT::getIntegerType(RoundBits/2);
660 Lo = getCopyFromParts(DAG, Parts, RoundParts/2, PartVT, HalfVT);
661 Hi = getCopyFromParts(DAG, Parts+RoundParts/2, RoundParts/2,
667 if (TLI.isBigEndian())
669 Val = DAG.getNode(ISD::BUILD_PAIR, RoundVT, Lo, Hi);
671 if (RoundParts < NumParts) {
672 // Assemble the trailing non-power-of-2 part.
673 unsigned OddParts = NumParts - RoundParts;
674 MVT::ValueType OddVT = MVT::getIntegerType(OddParts * PartBits);
675 Hi = getCopyFromParts(DAG, Parts+RoundParts, OddParts, PartVT, OddVT);
677 // Combine the round and odd parts.
679 if (TLI.isBigEndian())
681 MVT::ValueType TotalVT = MVT::getIntegerType(NumParts * PartBits);
682 Hi = DAG.getNode(ISD::ANY_EXTEND, TotalVT, Hi);
683 Hi = DAG.getNode(ISD::SHL, TotalVT, Hi,
684 DAG.getConstant(MVT::getSizeInBits(Lo.getValueType()),
685 TLI.getShiftAmountTy()));
686 Lo = DAG.getNode(ISD::ZERO_EXTEND, TotalVT, Lo);
687 Val = DAG.getNode(ISD::OR, TotalVT, Lo, Hi);
690 // Handle a multi-element vector.
691 MVT::ValueType IntermediateVT, RegisterVT;
692 unsigned NumIntermediates;
694 TLI.getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
697 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
698 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
699 assert(RegisterVT == Parts[0].getValueType() &&
700 "Part type doesn't match part!");
702 // Assemble the parts into intermediate operands.
703 SmallVector<SDOperand, 8> Ops(NumIntermediates);
704 if (NumIntermediates == NumParts) {
705 // If the register was not expanded, truncate or copy the value,
707 for (unsigned i = 0; i != NumParts; ++i)
708 Ops[i] = getCopyFromParts(DAG, &Parts[i], 1,
709 PartVT, IntermediateVT);
710 } else if (NumParts > 0) {
711 // If the intermediate type was expanded, build the intermediate operands
713 assert(NumParts % NumIntermediates == 0 &&
714 "Must expand into a divisible number of parts!");
715 unsigned Factor = NumParts / NumIntermediates;
716 for (unsigned i = 0; i != NumIntermediates; ++i)
717 Ops[i] = getCopyFromParts(DAG, &Parts[i * Factor], Factor,
718 PartVT, IntermediateVT);
721 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the intermediate
723 Val = DAG.getNode(MVT::isVector(IntermediateVT) ?
724 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR,
725 ValueVT, &Ops[0], NumIntermediates);
729 // There is now one part, held in Val. Correct it to match ValueVT.
730 PartVT = Val.getValueType();
732 if (PartVT == ValueVT)
735 if (MVT::isVector(PartVT)) {
736 assert(MVT::isVector(ValueVT) && "Unknown vector conversion!");
737 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
740 if (MVT::isVector(ValueVT)) {
741 assert(MVT::getVectorElementType(ValueVT) == PartVT &&
742 MVT::getVectorNumElements(ValueVT) == 1 &&
743 "Only trivial scalar-to-vector conversions should get here!");
744 return DAG.getNode(ISD::BUILD_VECTOR, ValueVT, Val);
747 if (MVT::isInteger(PartVT) &&
748 MVT::isInteger(ValueVT)) {
749 if (MVT::getSizeInBits(ValueVT) < MVT::getSizeInBits(PartVT)) {
750 // For a truncate, see if we have any information to
751 // indicate whether the truncated bits will always be
752 // zero or sign-extension.
753 if (AssertOp != ISD::DELETED_NODE)
754 Val = DAG.getNode(AssertOp, PartVT, Val,
755 DAG.getValueType(ValueVT));
756 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
758 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val);
762 if (MVT::isFloatingPoint(PartVT) && MVT::isFloatingPoint(ValueVT)) {
763 if (ValueVT < Val.getValueType())
764 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val,
765 DAG.getIntPtrConstant(TruncExact));
766 return DAG.getNode(ISD::FP_EXTEND, ValueVT, Val);
769 if (MVT::getSizeInBits(PartVT) == MVT::getSizeInBits(ValueVT))
770 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
772 assert(0 && "Unknown mismatch!");
775 /// getCopyToParts - Create a series of nodes that contain the specified value
776 /// split into legal parts. If the parts contain more bits than Val, then, for
777 /// integers, ExtendKind can be used to specify how to generate the extra bits.
778 static void getCopyToParts(SelectionDAG &DAG,
782 MVT::ValueType PartVT,
783 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
784 TargetLowering &TLI = DAG.getTargetLoweringInfo();
785 MVT::ValueType PtrVT = TLI.getPointerTy();
786 MVT::ValueType ValueVT = Val.getValueType();
787 unsigned PartBits = MVT::getSizeInBits(PartVT);
788 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
793 if (!MVT::isVector(ValueVT)) {
794 if (PartVT == ValueVT) {
795 assert(NumParts == 1 && "No-op copy with multiple parts!");
800 if (NumParts * PartBits > MVT::getSizeInBits(ValueVT)) {
801 // If the parts cover more bits than the value has, promote the value.
802 if (MVT::isFloatingPoint(PartVT) && MVT::isFloatingPoint(ValueVT)) {
803 assert(NumParts == 1 && "Do not know what to promote to!");
804 Val = DAG.getNode(ISD::FP_EXTEND, PartVT, Val);
805 } else if (MVT::isInteger(PartVT) && MVT::isInteger(ValueVT)) {
806 ValueVT = MVT::getIntegerType(NumParts * PartBits);
807 Val = DAG.getNode(ExtendKind, ValueVT, Val);
809 assert(0 && "Unknown mismatch!");
811 } else if (PartBits == MVT::getSizeInBits(ValueVT)) {
812 // Different types of the same size.
813 assert(NumParts == 1 && PartVT != ValueVT);
814 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
815 } else if (NumParts * PartBits < MVT::getSizeInBits(ValueVT)) {
816 // If the parts cover less bits than value has, truncate the value.
817 if (MVT::isInteger(PartVT) && MVT::isInteger(ValueVT)) {
818 ValueVT = MVT::getIntegerType(NumParts * PartBits);
819 Val = DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
821 assert(0 && "Unknown mismatch!");
825 // The value may have changed - recompute ValueVT.
826 ValueVT = Val.getValueType();
827 assert(NumParts * PartBits == MVT::getSizeInBits(ValueVT) &&
828 "Failed to tile the value with PartVT!");
831 assert(PartVT == ValueVT && "Type conversion failed!");
836 // Expand the value into multiple parts.
837 if (NumParts & (NumParts - 1)) {
838 // The number of parts is not a power of 2. Split off and copy the tail.
839 assert(MVT::isInteger(PartVT) && MVT::isInteger(ValueVT) &&
840 "Do not know what to expand to!");
841 unsigned RoundParts = 1 << Log2_32(NumParts);
842 unsigned RoundBits = RoundParts * PartBits;
843 unsigned OddParts = NumParts - RoundParts;
844 SDOperand OddVal = DAG.getNode(ISD::SRL, ValueVT, Val,
845 DAG.getConstant(RoundBits,
846 TLI.getShiftAmountTy()));
847 getCopyToParts(DAG, OddVal, Parts + RoundParts, OddParts, PartVT);
848 if (TLI.isBigEndian())
849 // The odd parts were reversed by getCopyToParts - unreverse them.
850 std::reverse(Parts + RoundParts, Parts + NumParts);
851 NumParts = RoundParts;
852 ValueVT = MVT::getIntegerType(NumParts * PartBits);
853 Val = DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
856 // The number of parts is a power of 2. Repeatedly bisect the value using
859 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
860 for (unsigned i = 0; i < NumParts; i += StepSize) {
861 unsigned ThisBits = StepSize * PartBits / 2;
862 MVT::ValueType ThisVT =
863 ThisBits == PartBits ? PartVT : MVT::getIntegerType (ThisBits);
865 Parts[i+StepSize/2] =
866 DAG.getNode(ISD::EXTRACT_ELEMENT, ThisVT, Parts[i],
867 DAG.getConstant(1, PtrVT));
869 DAG.getNode(ISD::EXTRACT_ELEMENT, ThisVT, Parts[i],
870 DAG.getConstant(0, PtrVT));
874 if (TLI.isBigEndian())
875 std::reverse(Parts, Parts + NumParts);
882 if (PartVT != ValueVT) {
883 if (MVT::isVector(PartVT)) {
884 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
886 assert(MVT::getVectorElementType(ValueVT) == PartVT &&
887 MVT::getVectorNumElements(ValueVT) == 1 &&
888 "Only trivial vector-to-scalar conversions should get here!");
889 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, PartVT, Val,
890 DAG.getConstant(0, PtrVT));
898 // Handle a multi-element vector.
899 MVT::ValueType IntermediateVT, RegisterVT;
900 unsigned NumIntermediates;
902 DAG.getTargetLoweringInfo()
903 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
905 unsigned NumElements = MVT::getVectorNumElements(ValueVT);
907 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
908 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
910 // Split the vector into intermediate operands.
911 SmallVector<SDOperand, 8> Ops(NumIntermediates);
912 for (unsigned i = 0; i != NumIntermediates; ++i)
913 if (MVT::isVector(IntermediateVT))
914 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR,
916 DAG.getConstant(i * (NumElements / NumIntermediates),
919 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
921 DAG.getConstant(i, PtrVT));
923 // Split the intermediate operands into legal parts.
924 if (NumParts == NumIntermediates) {
925 // If the register was not expanded, promote or copy the value,
927 for (unsigned i = 0; i != NumParts; ++i)
928 getCopyToParts(DAG, Ops[i], &Parts[i], 1, PartVT);
929 } else if (NumParts > 0) {
930 // If the intermediate type was expanded, split each the value into
932 assert(NumParts % NumIntermediates == 0 &&
933 "Must expand into a divisible number of parts!");
934 unsigned Factor = NumParts / NumIntermediates;
935 for (unsigned i = 0; i != NumIntermediates; ++i)
936 getCopyToParts(DAG, Ops[i], &Parts[i * Factor], Factor, PartVT);
941 SDOperand SelectionDAGLowering::getValue(const Value *V) {
942 SDOperand &N = NodeMap[V];
945 const Type *VTy = V->getType();
946 MVT::ValueType VT = TLI.getValueType(VTy);
947 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
948 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
949 visit(CE->getOpcode(), *CE);
950 SDOperand N1 = NodeMap[V];
951 assert(N1.Val && "visit didn't populate the ValueMap!");
953 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
954 return N = DAG.getGlobalAddress(GV, VT);
955 } else if (isa<ConstantPointerNull>(C)) {
956 return N = DAG.getConstant(0, TLI.getPointerTy());
957 } else if (isa<UndefValue>(C)) {
958 if (!isa<VectorType>(VTy))
959 return N = DAG.getNode(ISD::UNDEF, VT);
961 // Create a BUILD_VECTOR of undef nodes.
962 const VectorType *PTy = cast<VectorType>(VTy);
963 unsigned NumElements = PTy->getNumElements();
964 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
966 SmallVector<SDOperand, 8> Ops;
967 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT));
969 // Create a VConstant node with generic Vector type.
970 MVT::ValueType VT = MVT::getVectorType(PVT, NumElements);
971 return N = DAG.getNode(ISD::BUILD_VECTOR, VT,
972 &Ops[0], Ops.size());
973 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
974 return N = DAG.getConstantFP(CFP->getValueAPF(), VT);
975 } else if (const VectorType *PTy = dyn_cast<VectorType>(VTy)) {
976 unsigned NumElements = PTy->getNumElements();
977 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
979 // Now that we know the number and type of the elements, push a
980 // Constant or ConstantFP node onto the ops list for each element of
981 // the vector constant.
982 SmallVector<SDOperand, 8> Ops;
983 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
984 for (unsigned i = 0; i != NumElements; ++i)
985 Ops.push_back(getValue(CP->getOperand(i)));
987 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
989 if (MVT::isFloatingPoint(PVT))
990 Op = DAG.getConstantFP(0, PVT);
992 Op = DAG.getConstant(0, PVT);
993 Ops.assign(NumElements, Op);
996 // Create a BUILD_VECTOR node.
997 MVT::ValueType VT = MVT::getVectorType(PVT, NumElements);
998 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0],
1001 // Canonicalize all constant ints to be unsigned.
1002 return N = DAG.getConstant(cast<ConstantInt>(C)->getValue(),VT);
1006 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
1007 std::map<const AllocaInst*, int>::iterator SI =
1008 FuncInfo.StaticAllocaMap.find(AI);
1009 if (SI != FuncInfo.StaticAllocaMap.end())
1010 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
1013 unsigned InReg = FuncInfo.ValueMap[V];
1014 assert(InReg && "Value not in map!");
1016 MVT::ValueType RegisterVT = TLI.getRegisterType(VT);
1017 unsigned NumRegs = TLI.getNumRegisters(VT);
1019 std::vector<unsigned> Regs(NumRegs);
1020 for (unsigned i = 0; i != NumRegs; ++i)
1021 Regs[i] = InReg + i;
1023 RegsForValue RFV(Regs, RegisterVT, VT);
1024 SDOperand Chain = DAG.getEntryNode();
1026 return RFV.getCopyFromRegs(DAG, Chain, NULL);
1030 void SelectionDAGLowering::visitRet(ReturnInst &I) {
1031 if (I.getNumOperands() == 0) {
1032 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot()));
1035 SmallVector<SDOperand, 8> NewValues;
1036 NewValues.push_back(getRoot());
1037 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
1038 SDOperand RetOp = getValue(I.getOperand(i));
1039 MVT::ValueType VT = RetOp.getValueType();
1041 // FIXME: C calling convention requires the return type to be promoted to
1042 // at least 32-bit. But this is not necessary for non-C calling conventions.
1043 if (MVT::isInteger(VT)) {
1044 MVT::ValueType MinVT = TLI.getRegisterType(MVT::i32);
1045 if (MVT::getSizeInBits(VT) < MVT::getSizeInBits(MinVT))
1049 unsigned NumParts = TLI.getNumRegisters(VT);
1050 MVT::ValueType PartVT = TLI.getRegisterType(VT);
1051 SmallVector<SDOperand, 4> Parts(NumParts);
1052 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
1054 const Function *F = I.getParent()->getParent();
1055 if (F->paramHasAttr(0, ParamAttr::SExt))
1056 ExtendKind = ISD::SIGN_EXTEND;
1057 else if (F->paramHasAttr(0, ParamAttr::ZExt))
1058 ExtendKind = ISD::ZERO_EXTEND;
1060 getCopyToParts(DAG, RetOp, &Parts[0], NumParts, PartVT, ExtendKind);
1062 for (unsigned i = 0; i < NumParts; ++i) {
1063 NewValues.push_back(Parts[i]);
1064 NewValues.push_back(DAG.getConstant(false, MVT::i32));
1067 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other,
1068 &NewValues[0], NewValues.size()));
1071 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1072 /// the current basic block, add it to ValueMap now so that we'll get a
1074 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
1075 // No need to export constants.
1076 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1078 // Already exported?
1079 if (FuncInfo.isExportedInst(V)) return;
1081 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1082 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg));
1085 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
1086 const BasicBlock *FromBB) {
1087 // The operands of the setcc have to be in this block. We don't know
1088 // how to export them from some other block.
1089 if (Instruction *VI = dyn_cast<Instruction>(V)) {
1090 // Can export from current BB.
1091 if (VI->getParent() == FromBB)
1094 // Is already exported, noop.
1095 return FuncInfo.isExportedInst(V);
1098 // If this is an argument, we can export it if the BB is the entry block or
1099 // if it is already exported.
1100 if (isa<Argument>(V)) {
1101 if (FromBB == &FromBB->getParent()->getEntryBlock())
1104 // Otherwise, can only export this if it is already exported.
1105 return FuncInfo.isExportedInst(V);
1108 // Otherwise, constants can always be exported.
1112 static bool InBlock(const Value *V, const BasicBlock *BB) {
1113 if (const Instruction *I = dyn_cast<Instruction>(V))
1114 return I->getParent() == BB;
1118 /// FindMergedConditions - If Cond is an expression like
1119 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
1120 MachineBasicBlock *TBB,
1121 MachineBasicBlock *FBB,
1122 MachineBasicBlock *CurBB,
1124 // If this node is not part of the or/and tree, emit it as a branch.
1125 Instruction *BOp = dyn_cast<Instruction>(Cond);
1127 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1128 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1129 BOp->getParent() != CurBB->getBasicBlock() ||
1130 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1131 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1132 const BasicBlock *BB = CurBB->getBasicBlock();
1134 // If the leaf of the tree is a comparison, merge the condition into
1136 if ((isa<ICmpInst>(Cond) || isa<FCmpInst>(Cond)) &&
1137 // The operands of the cmp have to be in this block. We don't know
1138 // how to export them from some other block. If this is the first block
1139 // of the sequence, no exporting is needed.
1141 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1142 isExportableFromCurrentBlock(BOp->getOperand(1), BB)))) {
1143 BOp = cast<Instruction>(Cond);
1144 ISD::CondCode Condition;
1145 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1146 switch (IC->getPredicate()) {
1147 default: assert(0 && "Unknown icmp predicate opcode!");
1148 case ICmpInst::ICMP_EQ: Condition = ISD::SETEQ; break;
1149 case ICmpInst::ICMP_NE: Condition = ISD::SETNE; break;
1150 case ICmpInst::ICMP_SLE: Condition = ISD::SETLE; break;
1151 case ICmpInst::ICMP_ULE: Condition = ISD::SETULE; break;
1152 case ICmpInst::ICMP_SGE: Condition = ISD::SETGE; break;
1153 case ICmpInst::ICMP_UGE: Condition = ISD::SETUGE; break;
1154 case ICmpInst::ICMP_SLT: Condition = ISD::SETLT; break;
1155 case ICmpInst::ICMP_ULT: Condition = ISD::SETULT; break;
1156 case ICmpInst::ICMP_SGT: Condition = ISD::SETGT; break;
1157 case ICmpInst::ICMP_UGT: Condition = ISD::SETUGT; break;
1159 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1160 ISD::CondCode FPC, FOC;
1161 switch (FC->getPredicate()) {
1162 default: assert(0 && "Unknown fcmp predicate opcode!");
1163 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1164 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1165 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1166 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1167 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1168 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1169 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1170 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
1171 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
1172 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1173 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1174 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1175 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1176 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1177 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1178 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1180 if (FiniteOnlyFPMath())
1185 Condition = ISD::SETEQ; // silence warning.
1186 assert(0 && "Unknown compare instruction");
1189 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0),
1190 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1191 SwitchCases.push_back(CB);
1195 // Create a CaseBlock record representing this branch.
1196 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(),
1197 NULL, TBB, FBB, CurBB);
1198 SwitchCases.push_back(CB);
1203 // Create TmpBB after CurBB.
1204 MachineFunction::iterator BBI = CurBB;
1205 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock());
1206 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB);
1208 if (Opc == Instruction::Or) {
1209 // Codegen X | Y as:
1217 // Emit the LHS condition.
1218 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1220 // Emit the RHS condition into TmpBB.
1221 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1223 assert(Opc == Instruction::And && "Unknown merge op!");
1224 // Codegen X & Y as:
1231 // This requires creation of TmpBB after CurBB.
1233 // Emit the LHS condition.
1234 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1236 // Emit the RHS condition into TmpBB.
1237 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1241 /// If the set of cases should be emitted as a series of branches, return true.
1242 /// If we should emit this as a bunch of and/or'd together conditions, return
1245 ShouldEmitAsBranches(const std::vector<SelectionDAGISel::CaseBlock> &Cases) {
1246 if (Cases.size() != 2) return true;
1248 // If this is two comparisons of the same values or'd or and'd together, they
1249 // will get folded into a single comparison, so don't emit two blocks.
1250 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1251 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1252 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1253 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1260 void SelectionDAGLowering::visitBr(BranchInst &I) {
1261 // Update machine-CFG edges.
1262 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1264 // Figure out which block is immediately after the current one.
1265 MachineBasicBlock *NextBlock = 0;
1266 MachineFunction::iterator BBI = CurMBB;
1267 if (++BBI != CurMBB->getParent()->end())
1270 if (I.isUnconditional()) {
1271 // If this is not a fall-through branch, emit the branch.
1272 if (Succ0MBB != NextBlock)
1273 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1274 DAG.getBasicBlock(Succ0MBB)));
1276 // Update machine-CFG edges.
1277 CurMBB->addSuccessor(Succ0MBB);
1281 // If this condition is one of the special cases we handle, do special stuff
1283 Value *CondVal = I.getCondition();
1284 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1286 // If this is a series of conditions that are or'd or and'd together, emit
1287 // this as a sequence of branches instead of setcc's with and/or operations.
1288 // For example, instead of something like:
1301 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1302 if (BOp->hasOneUse() &&
1303 (BOp->getOpcode() == Instruction::And ||
1304 BOp->getOpcode() == Instruction::Or)) {
1305 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1306 // If the compares in later blocks need to use values not currently
1307 // exported from this block, export them now. This block should always
1308 // be the first entry.
1309 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1311 // Allow some cases to be rejected.
1312 if (ShouldEmitAsBranches(SwitchCases)) {
1313 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1314 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1315 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1318 // Emit the branch for this block.
1319 visitSwitchCase(SwitchCases[0]);
1320 SwitchCases.erase(SwitchCases.begin());
1324 // Okay, we decided not to do this, remove any inserted MBB's and clear
1326 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1327 CurMBB->getParent()->getBasicBlockList().erase(SwitchCases[i].ThisBB);
1329 SwitchCases.clear();
1333 // Create a CaseBlock record representing this branch.
1334 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(),
1335 NULL, Succ0MBB, Succ1MBB, CurMBB);
1336 // Use visitSwitchCase to actually insert the fast branch sequence for this
1338 visitSwitchCase(CB);
1341 /// visitSwitchCase - Emits the necessary code to represent a single node in
1342 /// the binary search tree resulting from lowering a switch instruction.
1343 void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) {
1345 SDOperand CondLHS = getValue(CB.CmpLHS);
1347 // Build the setcc now.
1348 if (CB.CmpMHS == NULL) {
1349 // Fold "(X == true)" to X and "(X == false)" to !X to
1350 // handle common cases produced by branch lowering.
1351 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ)
1353 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) {
1354 SDOperand True = DAG.getConstant(1, CondLHS.getValueType());
1355 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True);
1357 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1359 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1361 uint64_t Low = cast<ConstantInt>(CB.CmpLHS)->getSExtValue();
1362 uint64_t High = cast<ConstantInt>(CB.CmpRHS)->getSExtValue();
1364 SDOperand CmpOp = getValue(CB.CmpMHS);
1365 MVT::ValueType VT = CmpOp.getValueType();
1367 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1368 Cond = DAG.getSetCC(MVT::i1, CmpOp, DAG.getConstant(High, VT), ISD::SETLE);
1370 SDOperand SUB = DAG.getNode(ISD::SUB, VT, CmpOp, DAG.getConstant(Low, VT));
1371 Cond = DAG.getSetCC(MVT::i1, SUB,
1372 DAG.getConstant(High-Low, VT), ISD::SETULE);
1377 // Set NextBlock to be the MBB immediately after the current one, if any.
1378 // This is used to avoid emitting unnecessary branches to the next block.
1379 MachineBasicBlock *NextBlock = 0;
1380 MachineFunction::iterator BBI = CurMBB;
1381 if (++BBI != CurMBB->getParent()->end())
1384 // If the lhs block is the next block, invert the condition so that we can
1385 // fall through to the lhs instead of the rhs block.
1386 if (CB.TrueBB == NextBlock) {
1387 std::swap(CB.TrueBB, CB.FalseBB);
1388 SDOperand True = DAG.getConstant(1, Cond.getValueType());
1389 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
1391 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
1392 DAG.getBasicBlock(CB.TrueBB));
1393 if (CB.FalseBB == NextBlock)
1394 DAG.setRoot(BrCond);
1396 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1397 DAG.getBasicBlock(CB.FalseBB)));
1398 // Update successor info
1399 CurMBB->addSuccessor(CB.TrueBB);
1400 CurMBB->addSuccessor(CB.FalseBB);
1403 /// visitJumpTable - Emit JumpTable node in the current MBB
1404 void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) {
1405 // Emit the code for the jump table
1406 assert(JT.Reg != -1U && "Should lower JT Header first!");
1407 MVT::ValueType PTy = TLI.getPointerTy();
1408 SDOperand Index = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy);
1409 SDOperand Table = DAG.getJumpTable(JT.JTI, PTy);
1410 DAG.setRoot(DAG.getNode(ISD::BR_JT, MVT::Other, Index.getValue(1),
1415 /// visitJumpTableHeader - This function emits necessary code to produce index
1416 /// in the JumpTable from switch case.
1417 void SelectionDAGLowering::visitJumpTableHeader(SelectionDAGISel::JumpTable &JT,
1418 SelectionDAGISel::JumpTableHeader &JTH) {
1419 // Subtract the lowest switch case value from the value being switched on
1420 // and conditional branch to default mbb if the result is greater than the
1421 // difference between smallest and largest cases.
1422 SDOperand SwitchOp = getValue(JTH.SValue);
1423 MVT::ValueType VT = SwitchOp.getValueType();
1424 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1425 DAG.getConstant(JTH.First, VT));
1427 // The SDNode we just created, which holds the value being switched on
1428 // minus the the smallest case value, needs to be copied to a virtual
1429 // register so it can be used as an index into the jump table in a
1430 // subsequent basic block. This value may be smaller or larger than the
1431 // target's pointer type, and therefore require extension or truncating.
1432 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getPointerTy()))
1433 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
1435 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
1437 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1438 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp);
1439 JT.Reg = JumpTableReg;
1441 // Emit the range check for the jump table, and branch to the default
1442 // block for the switch statement if the value being switched on exceeds
1443 // the largest case in the switch.
1444 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1445 DAG.getConstant(JTH.Last-JTH.First,VT),
1448 // Set NextBlock to be the MBB immediately after the current one, if any.
1449 // This is used to avoid emitting unnecessary branches to the next block.
1450 MachineBasicBlock *NextBlock = 0;
1451 MachineFunction::iterator BBI = CurMBB;
1452 if (++BBI != CurMBB->getParent()->end())
1455 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
1456 DAG.getBasicBlock(JT.Default));
1458 if (JT.MBB == NextBlock)
1459 DAG.setRoot(BrCond);
1461 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1462 DAG.getBasicBlock(JT.MBB)));
1467 /// visitBitTestHeader - This function emits necessary code to produce value
1468 /// suitable for "bit tests"
1469 void SelectionDAGLowering::visitBitTestHeader(SelectionDAGISel::BitTestBlock &B) {
1470 // Subtract the minimum value
1471 SDOperand SwitchOp = getValue(B.SValue);
1472 MVT::ValueType VT = SwitchOp.getValueType();
1473 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1474 DAG.getConstant(B.First, VT));
1477 SDOperand RangeCmp = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1478 DAG.getConstant(B.Range, VT),
1482 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getShiftAmountTy()))
1483 ShiftOp = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), SUB);
1485 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getShiftAmountTy(), SUB);
1487 // Make desired shift
1488 SDOperand SwitchVal = DAG.getNode(ISD::SHL, TLI.getPointerTy(),
1489 DAG.getConstant(1, TLI.getPointerTy()),
1492 unsigned SwitchReg = FuncInfo.MakeReg(TLI.getPointerTy());
1493 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), SwitchReg, SwitchVal);
1496 SDOperand BrRange = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, RangeCmp,
1497 DAG.getBasicBlock(B.Default));
1499 // Set NextBlock to be the MBB immediately after the current one, if any.
1500 // This is used to avoid emitting unnecessary branches to the next block.
1501 MachineBasicBlock *NextBlock = 0;
1502 MachineFunction::iterator BBI = CurMBB;
1503 if (++BBI != CurMBB->getParent()->end())
1506 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1507 if (MBB == NextBlock)
1508 DAG.setRoot(BrRange);
1510 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, CopyTo,
1511 DAG.getBasicBlock(MBB)));
1513 CurMBB->addSuccessor(B.Default);
1514 CurMBB->addSuccessor(MBB);
1519 /// visitBitTestCase - this function produces one "bit test"
1520 void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB,
1522 SelectionDAGISel::BitTestCase &B) {
1523 // Emit bit tests and jumps
1524 SDOperand SwitchVal = DAG.getCopyFromReg(getRoot(), Reg, TLI.getPointerTy());
1526 SDOperand AndOp = DAG.getNode(ISD::AND, TLI.getPointerTy(),
1528 DAG.getConstant(B.Mask,
1529 TLI.getPointerTy()));
1530 SDOperand AndCmp = DAG.getSetCC(TLI.getSetCCResultTy(), AndOp,
1531 DAG.getConstant(0, TLI.getPointerTy()),
1533 SDOperand BrAnd = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(),
1534 AndCmp, DAG.getBasicBlock(B.TargetBB));
1536 // Set NextBlock to be the MBB immediately after the current one, if any.
1537 // This is used to avoid emitting unnecessary branches to the next block.
1538 MachineBasicBlock *NextBlock = 0;
1539 MachineFunction::iterator BBI = CurMBB;
1540 if (++BBI != CurMBB->getParent()->end())
1543 if (NextMBB == NextBlock)
1546 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrAnd,
1547 DAG.getBasicBlock(NextMBB)));
1549 CurMBB->addSuccessor(B.TargetBB);
1550 CurMBB->addSuccessor(NextMBB);
1555 void SelectionDAGLowering::visitInvoke(InvokeInst &I) {
1556 // Retrieve successors.
1557 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1558 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1560 if (isa<InlineAsm>(I.getCalledValue()))
1563 LowerCallTo(&I, getValue(I.getOperand(0)), false, LandingPad);
1565 // If the value of the invoke is used outside of its defining block, make it
1566 // available as a virtual register.
1567 if (!I.use_empty()) {
1568 DenseMap<const Value*, unsigned>::iterator VMI = FuncInfo.ValueMap.find(&I);
1569 if (VMI != FuncInfo.ValueMap.end())
1570 DAG.setRoot(CopyValueToVirtualRegister(&I, VMI->second));
1573 // Drop into normal successor.
1574 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1575 DAG.getBasicBlock(Return)));
1577 // Update successor info
1578 CurMBB->addSuccessor(Return);
1579 CurMBB->addSuccessor(LandingPad);
1582 void SelectionDAGLowering::visitUnwind(UnwindInst &I) {
1585 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1586 /// small case ranges).
1587 bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR,
1588 CaseRecVector& WorkList,
1590 MachineBasicBlock* Default) {
1591 Case& BackCase = *(CR.Range.second-1);
1593 // Size is the number of Cases represented by this range.
1594 unsigned Size = CR.Range.second - CR.Range.first;
1598 // Get the MachineFunction which holds the current MBB. This is used when
1599 // inserting any additional MBBs necessary to represent the switch.
1600 MachineFunction *CurMF = CurMBB->getParent();
1602 // Figure out which block is immediately after the current one.
1603 MachineBasicBlock *NextBlock = 0;
1604 MachineFunction::iterator BBI = CR.CaseBB;
1606 if (++BBI != CurMBB->getParent()->end())
1609 // TODO: If any two of the cases has the same destination, and if one value
1610 // is the same as the other, but has one bit unset that the other has set,
1611 // use bit manipulation to do two compares at once. For example:
1612 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1614 // Rearrange the case blocks so that the last one falls through if possible.
1615 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1616 // The last case block won't fall through into 'NextBlock' if we emit the
1617 // branches in this order. See if rearranging a case value would help.
1618 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1619 if (I->BB == NextBlock) {
1620 std::swap(*I, BackCase);
1626 // Create a CaseBlock record representing a conditional branch to
1627 // the Case's target mbb if the value being switched on SV is equal
1629 MachineBasicBlock *CurBlock = CR.CaseBB;
1630 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1631 MachineBasicBlock *FallThrough;
1633 FallThrough = new MachineBasicBlock(CurBlock->getBasicBlock());
1634 CurMF->getBasicBlockList().insert(BBI, FallThrough);
1636 // If the last case doesn't match, go to the default block.
1637 FallThrough = Default;
1640 Value *RHS, *LHS, *MHS;
1642 if (I->High == I->Low) {
1643 // This is just small small case range :) containing exactly 1 case
1645 LHS = SV; RHS = I->High; MHS = NULL;
1648 LHS = I->Low; MHS = SV; RHS = I->High;
1650 SelectionDAGISel::CaseBlock CB(CC, LHS, RHS, MHS,
1651 I->BB, FallThrough, CurBlock);
1653 // If emitting the first comparison, just call visitSwitchCase to emit the
1654 // code into the current block. Otherwise, push the CaseBlock onto the
1655 // vector to be later processed by SDISel, and insert the node's MBB
1656 // before the next MBB.
1657 if (CurBlock == CurMBB)
1658 visitSwitchCase(CB);
1660 SwitchCases.push_back(CB);
1662 CurBlock = FallThrough;
1668 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1669 return (TLI.isOperationLegal(ISD::BR_JT, MVT::Other) ||
1670 TLI.isOperationLegal(ISD::BRIND, MVT::Other));
1673 /// handleJTSwitchCase - Emit jumptable for current switch case range
1674 bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR,
1675 CaseRecVector& WorkList,
1677 MachineBasicBlock* Default) {
1678 Case& FrontCase = *CR.Range.first;
1679 Case& BackCase = *(CR.Range.second-1);
1681 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1682 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1685 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1689 if (!areJTsAllowed(TLI) || TSize <= 3)
1692 double Density = (double)TSize / (double)((Last - First) + 1ULL);
1696 DOUT << "Lowering jump table\n"
1697 << "First entry: " << First << ". Last entry: " << Last << "\n"
1698 << "Size: " << TSize << ". Density: " << Density << "\n\n";
1700 // Get the MachineFunction which holds the current MBB. This is used when
1701 // inserting any additional MBBs necessary to represent the switch.
1702 MachineFunction *CurMF = CurMBB->getParent();
1704 // Figure out which block is immediately after the current one.
1705 MachineBasicBlock *NextBlock = 0;
1706 MachineFunction::iterator BBI = CR.CaseBB;
1708 if (++BBI != CurMBB->getParent()->end())
1711 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1713 // Create a new basic block to hold the code for loading the address
1714 // of the jump table, and jumping to it. Update successor information;
1715 // we will either branch to the default case for the switch, or the jump
1717 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB);
1718 CurMF->getBasicBlockList().insert(BBI, JumpTableBB);
1719 CR.CaseBB->addSuccessor(Default);
1720 CR.CaseBB->addSuccessor(JumpTableBB);
1722 // Build a vector of destination BBs, corresponding to each target
1723 // of the jump table. If the value of the jump table slot corresponds to
1724 // a case statement, push the case's BB onto the vector, otherwise, push
1726 std::vector<MachineBasicBlock*> DestBBs;
1727 int64_t TEI = First;
1728 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1729 int64_t Low = cast<ConstantInt>(I->Low)->getSExtValue();
1730 int64_t High = cast<ConstantInt>(I->High)->getSExtValue();
1732 if ((Low <= TEI) && (TEI <= High)) {
1733 DestBBs.push_back(I->BB);
1737 DestBBs.push_back(Default);
1741 // Update successor info. Add one edge to each unique successor.
1742 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1743 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1744 E = DestBBs.end(); I != E; ++I) {
1745 if (!SuccsHandled[(*I)->getNumber()]) {
1746 SuccsHandled[(*I)->getNumber()] = true;
1747 JumpTableBB->addSuccessor(*I);
1751 // Create a jump table index for this jump table, or return an existing
1753 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1755 // Set the jump table information so that we can codegen it as a second
1756 // MachineBasicBlock
1757 SelectionDAGISel::JumpTable JT(-1U, JTI, JumpTableBB, Default);
1758 SelectionDAGISel::JumpTableHeader JTH(First, Last, SV, CR.CaseBB,
1759 (CR.CaseBB == CurMBB));
1760 if (CR.CaseBB == CurMBB)
1761 visitJumpTableHeader(JT, JTH);
1763 JTCases.push_back(SelectionDAGISel::JumpTableBlock(JTH, JT));
1768 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1770 bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR,
1771 CaseRecVector& WorkList,
1773 MachineBasicBlock* Default) {
1774 // Get the MachineFunction which holds the current MBB. This is used when
1775 // inserting any additional MBBs necessary to represent the switch.
1776 MachineFunction *CurMF = CurMBB->getParent();
1778 // Figure out which block is immediately after the current one.
1779 MachineBasicBlock *NextBlock = 0;
1780 MachineFunction::iterator BBI = CR.CaseBB;
1782 if (++BBI != CurMBB->getParent()->end())
1785 Case& FrontCase = *CR.Range.first;
1786 Case& BackCase = *(CR.Range.second-1);
1787 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1789 // Size is the number of Cases represented by this range.
1790 unsigned Size = CR.Range.second - CR.Range.first;
1792 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1793 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1795 CaseItr Pivot = CR.Range.first + Size/2;
1797 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1798 // (heuristically) allow us to emit JumpTable's later.
1800 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1804 uint64_t LSize = FrontCase.size();
1805 uint64_t RSize = TSize-LSize;
1806 DOUT << "Selecting best pivot: \n"
1807 << "First: " << First << ", Last: " << Last <<"\n"
1808 << "LSize: " << LSize << ", RSize: " << RSize << "\n";
1809 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1811 int64_t LEnd = cast<ConstantInt>(I->High)->getSExtValue();
1812 int64_t RBegin = cast<ConstantInt>(J->Low)->getSExtValue();
1813 assert((RBegin-LEnd>=1) && "Invalid case distance");
1814 double LDensity = (double)LSize / (double)((LEnd - First) + 1ULL);
1815 double RDensity = (double)RSize / (double)((Last - RBegin) + 1ULL);
1816 double Metric = Log2_64(RBegin-LEnd)*(LDensity+RDensity);
1817 // Should always split in some non-trivial place
1819 << "LEnd: " << LEnd << ", RBegin: " << RBegin << "\n"
1820 << "LDensity: " << LDensity << ", RDensity: " << RDensity << "\n"
1821 << "Metric: " << Metric << "\n";
1822 if (FMetric < Metric) {
1825 DOUT << "Current metric set to: " << FMetric << "\n";
1831 if (areJTsAllowed(TLI)) {
1832 // If our case is dense we *really* should handle it earlier!
1833 assert((FMetric > 0) && "Should handle dense range earlier!");
1835 Pivot = CR.Range.first + Size/2;
1838 CaseRange LHSR(CR.Range.first, Pivot);
1839 CaseRange RHSR(Pivot, CR.Range.second);
1840 Constant *C = Pivot->Low;
1841 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1843 // We know that we branch to the LHS if the Value being switched on is
1844 // less than the Pivot value, C. We use this to optimize our binary
1845 // tree a bit, by recognizing that if SV is greater than or equal to the
1846 // LHS's Case Value, and that Case Value is exactly one less than the
1847 // Pivot's Value, then we can branch directly to the LHS's Target,
1848 // rather than creating a leaf node for it.
1849 if ((LHSR.second - LHSR.first) == 1 &&
1850 LHSR.first->High == CR.GE &&
1851 cast<ConstantInt>(C)->getSExtValue() ==
1852 (cast<ConstantInt>(CR.GE)->getSExtValue() + 1LL)) {
1853 TrueBB = LHSR.first->BB;
1855 TrueBB = new MachineBasicBlock(LLVMBB);
1856 CurMF->getBasicBlockList().insert(BBI, TrueBB);
1857 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1860 // Similar to the optimization above, if the Value being switched on is
1861 // known to be less than the Constant CR.LT, and the current Case Value
1862 // is CR.LT - 1, then we can branch directly to the target block for
1863 // the current Case Value, rather than emitting a RHS leaf node for it.
1864 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1865 cast<ConstantInt>(RHSR.first->Low)->getSExtValue() ==
1866 (cast<ConstantInt>(CR.LT)->getSExtValue() - 1LL)) {
1867 FalseBB = RHSR.first->BB;
1869 FalseBB = new MachineBasicBlock(LLVMBB);
1870 CurMF->getBasicBlockList().insert(BBI, FalseBB);
1871 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1874 // Create a CaseBlock record representing a conditional branch to
1875 // the LHS node if the value being switched on SV is less than C.
1876 // Otherwise, branch to LHS.
1877 SelectionDAGISel::CaseBlock CB(ISD::SETLT, SV, C, NULL,
1878 TrueBB, FalseBB, CR.CaseBB);
1880 if (CR.CaseBB == CurMBB)
1881 visitSwitchCase(CB);
1883 SwitchCases.push_back(CB);
1888 /// handleBitTestsSwitchCase - if current case range has few destination and
1889 /// range span less, than machine word bitwidth, encode case range into series
1890 /// of masks and emit bit tests with these masks.
1891 bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR,
1892 CaseRecVector& WorkList,
1894 MachineBasicBlock* Default){
1895 unsigned IntPtrBits = MVT::getSizeInBits(TLI.getPointerTy());
1897 Case& FrontCase = *CR.Range.first;
1898 Case& BackCase = *(CR.Range.second-1);
1900 // Get the MachineFunction which holds the current MBB. This is used when
1901 // inserting any additional MBBs necessary to represent the switch.
1902 MachineFunction *CurMF = CurMBB->getParent();
1904 unsigned numCmps = 0;
1905 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1907 // Single case counts one, case range - two.
1908 if (I->Low == I->High)
1914 // Count unique destinations
1915 SmallSet<MachineBasicBlock*, 4> Dests;
1916 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1917 Dests.insert(I->BB);
1918 if (Dests.size() > 3)
1919 // Don't bother the code below, if there are too much unique destinations
1922 DOUT << "Total number of unique destinations: " << Dests.size() << "\n"
1923 << "Total number of comparisons: " << numCmps << "\n";
1925 // Compute span of values.
1926 Constant* minValue = FrontCase.Low;
1927 Constant* maxValue = BackCase.High;
1928 uint64_t range = cast<ConstantInt>(maxValue)->getSExtValue() -
1929 cast<ConstantInt>(minValue)->getSExtValue();
1930 DOUT << "Compare range: " << range << "\n"
1931 << "Low bound: " << cast<ConstantInt>(minValue)->getSExtValue() << "\n"
1932 << "High bound: " << cast<ConstantInt>(maxValue)->getSExtValue() << "\n";
1934 if (range>=IntPtrBits ||
1935 (!(Dests.size() == 1 && numCmps >= 3) &&
1936 !(Dests.size() == 2 && numCmps >= 5) &&
1937 !(Dests.size() >= 3 && numCmps >= 6)))
1940 DOUT << "Emitting bit tests\n";
1941 int64_t lowBound = 0;
1943 // Optimize the case where all the case values fit in a
1944 // word without having to subtract minValue. In this case,
1945 // we can optimize away the subtraction.
1946 if (cast<ConstantInt>(minValue)->getSExtValue() >= 0 &&
1947 cast<ConstantInt>(maxValue)->getSExtValue() < IntPtrBits) {
1948 range = cast<ConstantInt>(maxValue)->getSExtValue();
1950 lowBound = cast<ConstantInt>(minValue)->getSExtValue();
1953 CaseBitsVector CasesBits;
1954 unsigned i, count = 0;
1956 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1957 MachineBasicBlock* Dest = I->BB;
1958 for (i = 0; i < count; ++i)
1959 if (Dest == CasesBits[i].BB)
1963 assert((count < 3) && "Too much destinations to test!");
1964 CasesBits.push_back(CaseBits(0, Dest, 0));
1968 uint64_t lo = cast<ConstantInt>(I->Low)->getSExtValue() - lowBound;
1969 uint64_t hi = cast<ConstantInt>(I->High)->getSExtValue() - lowBound;
1971 for (uint64_t j = lo; j <= hi; j++) {
1972 CasesBits[i].Mask |= 1ULL << j;
1973 CasesBits[i].Bits++;
1977 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
1979 SelectionDAGISel::BitTestInfo BTC;
1981 // Figure out which block is immediately after the current one.
1982 MachineFunction::iterator BBI = CR.CaseBB;
1985 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1988 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
1989 DOUT << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits
1990 << ", BB: " << CasesBits[i].BB << "\n";
1992 MachineBasicBlock *CaseBB = new MachineBasicBlock(LLVMBB);
1993 CurMF->getBasicBlockList().insert(BBI, CaseBB);
1994 BTC.push_back(SelectionDAGISel::BitTestCase(CasesBits[i].Mask,
1999 SelectionDAGISel::BitTestBlock BTB(lowBound, range, SV,
2000 -1U, (CR.CaseBB == CurMBB),
2001 CR.CaseBB, Default, BTC);
2003 if (CR.CaseBB == CurMBB)
2004 visitBitTestHeader(BTB);
2006 BitTestCases.push_back(BTB);
2012 // Clusterify - Transform simple list of Cases into list of CaseRange's
2013 unsigned SelectionDAGLowering::Clusterify(CaseVector& Cases,
2014 const SwitchInst& SI) {
2015 unsigned numCmps = 0;
2017 // Start with "simple" cases
2018 for (unsigned i = 1; i < SI.getNumSuccessors(); ++i) {
2019 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2020 Cases.push_back(Case(SI.getSuccessorValue(i),
2021 SI.getSuccessorValue(i),
2024 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2026 // Merge case into clusters
2027 if (Cases.size()>=2)
2028 // Must recompute end() each iteration because it may be
2029 // invalidated by erase if we hold on to it
2030 for (CaseItr I=Cases.begin(), J=++(Cases.begin()); J!=Cases.end(); ) {
2031 int64_t nextValue = cast<ConstantInt>(J->Low)->getSExtValue();
2032 int64_t currentValue = cast<ConstantInt>(I->High)->getSExtValue();
2033 MachineBasicBlock* nextBB = J->BB;
2034 MachineBasicBlock* currentBB = I->BB;
2036 // If the two neighboring cases go to the same destination, merge them
2037 // into a single case.
2038 if ((nextValue-currentValue==1) && (currentBB == nextBB)) {
2046 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2047 if (I->Low != I->High)
2048 // A range counts double, since it requires two compares.
2055 void SelectionDAGLowering::visitSwitch(SwitchInst &SI) {
2056 // Figure out which block is immediately after the current one.
2057 MachineBasicBlock *NextBlock = 0;
2058 MachineFunction::iterator BBI = CurMBB;
2060 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2062 // If there is only the default destination, branch to it if it is not the
2063 // next basic block. Otherwise, just fall through.
2064 if (SI.getNumOperands() == 2) {
2065 // Update machine-CFG edges.
2067 // If this is not a fall-through branch, emit the branch.
2068 if (Default != NextBlock)
2069 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
2070 DAG.getBasicBlock(Default)));
2072 CurMBB->addSuccessor(Default);
2076 // If there are any non-default case statements, create a vector of Cases
2077 // representing each one, and sort the vector so that we can efficiently
2078 // create a binary search tree from them.
2080 unsigned numCmps = Clusterify(Cases, SI);
2081 DOUT << "Clusterify finished. Total clusters: " << Cases.size()
2082 << ". Total compares: " << numCmps << "\n";
2084 // Get the Value to be switched on and default basic blocks, which will be
2085 // inserted into CaseBlock records, representing basic blocks in the binary
2087 Value *SV = SI.getOperand(0);
2089 // Push the initial CaseRec onto the worklist
2090 CaseRecVector WorkList;
2091 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
2093 while (!WorkList.empty()) {
2094 // Grab a record representing a case range to process off the worklist
2095 CaseRec CR = WorkList.back();
2096 WorkList.pop_back();
2098 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
2101 // If the range has few cases (two or less) emit a series of specific
2103 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2106 // If the switch has more than 5 blocks, and at least 40% dense, and the
2107 // target supports indirect branches, then emit a jump table rather than
2108 // lowering the switch to a binary tree of conditional branches.
2109 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2112 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2113 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2114 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2119 void SelectionDAGLowering::visitSub(User &I) {
2120 // -0.0 - X --> fneg
2121 const Type *Ty = I.getType();
2122 if (isa<VectorType>(Ty)) {
2123 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2124 const VectorType *DestTy = cast<VectorType>(I.getType());
2125 const Type *ElTy = DestTy->getElementType();
2126 if (ElTy->isFloatingPoint()) {
2127 unsigned VL = DestTy->getNumElements();
2128 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2129 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2131 SDOperand Op2 = getValue(I.getOperand(1));
2132 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2138 if (Ty->isFloatingPoint()) {
2139 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2140 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2141 SDOperand Op2 = getValue(I.getOperand(1));
2142 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2147 visitBinary(I, Ty->isFPOrFPVector() ? ISD::FSUB : ISD::SUB);
2150 void SelectionDAGLowering::visitBinary(User &I, unsigned OpCode) {
2151 SDOperand Op1 = getValue(I.getOperand(0));
2152 SDOperand Op2 = getValue(I.getOperand(1));
2154 setValue(&I, DAG.getNode(OpCode, Op1.getValueType(), Op1, Op2));
2157 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
2158 SDOperand Op1 = getValue(I.getOperand(0));
2159 SDOperand Op2 = getValue(I.getOperand(1));
2161 if (MVT::getSizeInBits(TLI.getShiftAmountTy()) <
2162 MVT::getSizeInBits(Op2.getValueType()))
2163 Op2 = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), Op2);
2164 else if (TLI.getShiftAmountTy() > Op2.getValueType())
2165 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
2167 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
2170 void SelectionDAGLowering::visitICmp(User &I) {
2171 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2172 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2173 predicate = IC->getPredicate();
2174 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2175 predicate = ICmpInst::Predicate(IC->getPredicate());
2176 SDOperand Op1 = getValue(I.getOperand(0));
2177 SDOperand Op2 = getValue(I.getOperand(1));
2178 ISD::CondCode Opcode;
2179 switch (predicate) {
2180 case ICmpInst::ICMP_EQ : Opcode = ISD::SETEQ; break;
2181 case ICmpInst::ICMP_NE : Opcode = ISD::SETNE; break;
2182 case ICmpInst::ICMP_UGT : Opcode = ISD::SETUGT; break;
2183 case ICmpInst::ICMP_UGE : Opcode = ISD::SETUGE; break;
2184 case ICmpInst::ICMP_ULT : Opcode = ISD::SETULT; break;
2185 case ICmpInst::ICMP_ULE : Opcode = ISD::SETULE; break;
2186 case ICmpInst::ICMP_SGT : Opcode = ISD::SETGT; break;
2187 case ICmpInst::ICMP_SGE : Opcode = ISD::SETGE; break;
2188 case ICmpInst::ICMP_SLT : Opcode = ISD::SETLT; break;
2189 case ICmpInst::ICMP_SLE : Opcode = ISD::SETLE; break;
2191 assert(!"Invalid ICmp predicate value");
2192 Opcode = ISD::SETEQ;
2195 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
2198 void SelectionDAGLowering::visitFCmp(User &I) {
2199 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2200 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2201 predicate = FC->getPredicate();
2202 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2203 predicate = FCmpInst::Predicate(FC->getPredicate());
2204 SDOperand Op1 = getValue(I.getOperand(0));
2205 SDOperand Op2 = getValue(I.getOperand(1));
2206 ISD::CondCode Condition, FOC, FPC;
2207 switch (predicate) {
2208 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
2209 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
2210 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
2211 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
2212 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
2213 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
2214 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
2215 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
2216 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
2217 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
2218 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
2219 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
2220 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
2221 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
2222 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
2223 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
2225 assert(!"Invalid FCmp predicate value");
2226 FOC = FPC = ISD::SETFALSE;
2229 if (FiniteOnlyFPMath())
2233 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Condition));
2236 void SelectionDAGLowering::visitSelect(User &I) {
2237 SDOperand Cond = getValue(I.getOperand(0));
2238 SDOperand TrueVal = getValue(I.getOperand(1));
2239 SDOperand FalseVal = getValue(I.getOperand(2));
2240 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
2241 TrueVal, FalseVal));
2245 void SelectionDAGLowering::visitTrunc(User &I) {
2246 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2247 SDOperand N = getValue(I.getOperand(0));
2248 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2249 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2252 void SelectionDAGLowering::visitZExt(User &I) {
2253 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2254 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2255 SDOperand N = getValue(I.getOperand(0));
2256 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2257 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2260 void SelectionDAGLowering::visitSExt(User &I) {
2261 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2262 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2263 SDOperand N = getValue(I.getOperand(0));
2264 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2265 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
2268 void SelectionDAGLowering::visitFPTrunc(User &I) {
2269 // FPTrunc is never a no-op cast, no need to check
2270 SDOperand N = getValue(I.getOperand(0));
2271 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2272 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N, DAG.getIntPtrConstant(0)));
2275 void SelectionDAGLowering::visitFPExt(User &I){
2276 // FPTrunc is never a no-op cast, no need to check
2277 SDOperand N = getValue(I.getOperand(0));
2278 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2279 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
2282 void SelectionDAGLowering::visitFPToUI(User &I) {
2283 // FPToUI is never a no-op cast, no need to check
2284 SDOperand N = getValue(I.getOperand(0));
2285 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2286 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
2289 void SelectionDAGLowering::visitFPToSI(User &I) {
2290 // FPToSI is never a no-op cast, no need to check
2291 SDOperand N = getValue(I.getOperand(0));
2292 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2293 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
2296 void SelectionDAGLowering::visitUIToFP(User &I) {
2297 // UIToFP is never a no-op cast, no need to check
2298 SDOperand N = getValue(I.getOperand(0));
2299 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2300 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
2303 void SelectionDAGLowering::visitSIToFP(User &I){
2304 // UIToFP is never a no-op cast, no need to check
2305 SDOperand N = getValue(I.getOperand(0));
2306 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2307 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
2310 void SelectionDAGLowering::visitPtrToInt(User &I) {
2311 // What to do depends on the size of the integer and the size of the pointer.
2312 // We can either truncate, zero extend, or no-op, accordingly.
2313 SDOperand N = getValue(I.getOperand(0));
2314 MVT::ValueType SrcVT = N.getValueType();
2315 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2317 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT))
2318 Result = DAG.getNode(ISD::TRUNCATE, DestVT, N);
2320 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2321 Result = DAG.getNode(ISD::ZERO_EXTEND, DestVT, N);
2322 setValue(&I, Result);
2325 void SelectionDAGLowering::visitIntToPtr(User &I) {
2326 // What to do depends on the size of the integer and the size of the pointer.
2327 // We can either truncate, zero extend, or no-op, accordingly.
2328 SDOperand N = getValue(I.getOperand(0));
2329 MVT::ValueType SrcVT = N.getValueType();
2330 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2331 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT))
2332 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2334 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2335 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2338 void SelectionDAGLowering::visitBitCast(User &I) {
2339 SDOperand N = getValue(I.getOperand(0));
2340 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2342 // BitCast assures us that source and destination are the same size so this
2343 // is either a BIT_CONVERT or a no-op.
2344 if (DestVT != N.getValueType())
2345 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, DestVT, N)); // convert types
2347 setValue(&I, N); // noop cast.
2350 void SelectionDAGLowering::visitInsertElement(User &I) {
2351 SDOperand InVec = getValue(I.getOperand(0));
2352 SDOperand InVal = getValue(I.getOperand(1));
2353 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2354 getValue(I.getOperand(2)));
2356 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT,
2357 TLI.getValueType(I.getType()),
2358 InVec, InVal, InIdx));
2361 void SelectionDAGLowering::visitExtractElement(User &I) {
2362 SDOperand InVec = getValue(I.getOperand(0));
2363 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2364 getValue(I.getOperand(1)));
2365 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
2366 TLI.getValueType(I.getType()), InVec, InIdx));
2369 void SelectionDAGLowering::visitShuffleVector(User &I) {
2370 SDOperand V1 = getValue(I.getOperand(0));
2371 SDOperand V2 = getValue(I.getOperand(1));
2372 SDOperand Mask = getValue(I.getOperand(2));
2374 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE,
2375 TLI.getValueType(I.getType()),
2380 void SelectionDAGLowering::visitGetElementPtr(User &I) {
2381 SDOperand N = getValue(I.getOperand(0));
2382 const Type *Ty = I.getOperand(0)->getType();
2384 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2387 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2388 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2391 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2392 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
2393 DAG.getIntPtrConstant(Offset));
2395 Ty = StTy->getElementType(Field);
2397 Ty = cast<SequentialType>(Ty)->getElementType();
2399 // If this is a constant subscript, handle it quickly.
2400 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2401 if (CI->getZExtValue() == 0) continue;
2403 TD->getABITypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2404 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
2405 DAG.getIntPtrConstant(Offs));
2409 // N = N + Idx * ElementSize;
2410 uint64_t ElementSize = TD->getABITypeSize(Ty);
2411 SDOperand IdxN = getValue(Idx);
2413 // If the index is smaller or larger than intptr_t, truncate or extend
2415 if (IdxN.getValueType() < N.getValueType()) {
2416 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
2417 } else if (IdxN.getValueType() > N.getValueType())
2418 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
2420 // If this is a multiply by a power of two, turn it into a shl
2421 // immediately. This is a very common case.
2422 if (isPowerOf2_64(ElementSize)) {
2423 unsigned Amt = Log2_64(ElementSize);
2424 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
2425 DAG.getConstant(Amt, TLI.getShiftAmountTy()));
2426 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
2430 SDOperand Scale = DAG.getIntPtrConstant(ElementSize);
2431 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
2432 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
2438 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
2439 // If this is a fixed sized alloca in the entry block of the function,
2440 // allocate it statically on the stack.
2441 if (FuncInfo.StaticAllocaMap.count(&I))
2442 return; // getValue will auto-populate this.
2444 const Type *Ty = I.getAllocatedType();
2445 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
2447 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2450 SDOperand AllocSize = getValue(I.getArraySize());
2451 MVT::ValueType IntPtr = TLI.getPointerTy();
2452 if (IntPtr < AllocSize.getValueType())
2453 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
2454 else if (IntPtr > AllocSize.getValueType())
2455 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
2457 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
2458 DAG.getIntPtrConstant(TySize));
2460 // Handle alignment. If the requested alignment is less than or equal to
2461 // the stack alignment, ignore it. If the size is greater than or equal to
2462 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2463 unsigned StackAlign =
2464 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2465 if (Align <= StackAlign)
2468 // Round the size of the allocation up to the stack alignment size
2469 // by add SA-1 to the size.
2470 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
2471 DAG.getIntPtrConstant(StackAlign-1));
2472 // Mask out the low bits for alignment purposes.
2473 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
2474 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2476 SDOperand Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2477 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
2479 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3);
2481 DAG.setRoot(DSA.getValue(1));
2483 // Inform the Frame Information that we have just allocated a variable-sized
2485 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
2488 void SelectionDAGLowering::visitLoad(LoadInst &I) {
2489 SDOperand Ptr = getValue(I.getOperand(0));
2495 // Do not serialize non-volatile loads against each other.
2496 Root = DAG.getRoot();
2499 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0),
2500 Root, I.isVolatile(), I.getAlignment()));
2503 SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr,
2504 const Value *SV, SDOperand Root,
2506 unsigned Alignment) {
2508 DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, 0,
2509 isVolatile, Alignment);
2512 DAG.setRoot(L.getValue(1));
2514 PendingLoads.push_back(L.getValue(1));
2520 void SelectionDAGLowering::visitStore(StoreInst &I) {
2521 Value *SrcV = I.getOperand(0);
2522 SDOperand Src = getValue(SrcV);
2523 SDOperand Ptr = getValue(I.getOperand(1));
2524 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1), 0,
2525 I.isVolatile(), I.getAlignment()));
2528 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2530 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
2531 unsigned Intrinsic) {
2532 bool HasChain = !I.doesNotAccessMemory();
2533 bool OnlyLoad = HasChain && I.onlyReadsMemory();
2535 // Build the operand list.
2536 SmallVector<SDOperand, 8> Ops;
2537 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2539 // We don't need to serialize loads against other loads.
2540 Ops.push_back(DAG.getRoot());
2542 Ops.push_back(getRoot());
2546 // Add the intrinsic ID as an integer operand.
2547 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2549 // Add all operands of the call to the operand list.
2550 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2551 SDOperand Op = getValue(I.getOperand(i));
2552 assert(TLI.isTypeLegal(Op.getValueType()) &&
2553 "Intrinsic uses a non-legal type?");
2557 std::vector<MVT::ValueType> VTs;
2558 if (I.getType() != Type::VoidTy) {
2559 MVT::ValueType VT = TLI.getValueType(I.getType());
2560 if (MVT::isVector(VT)) {
2561 const VectorType *DestTy = cast<VectorType>(I.getType());
2562 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
2564 VT = MVT::getVectorType(EltVT, DestTy->getNumElements());
2565 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
2568 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
2572 VTs.push_back(MVT::Other);
2574 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs);
2579 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(),
2580 &Ops[0], Ops.size());
2581 else if (I.getType() != Type::VoidTy)
2582 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(),
2583 &Ops[0], Ops.size());
2585 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(),
2586 &Ops[0], Ops.size());
2589 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1);
2591 PendingLoads.push_back(Chain);
2595 if (I.getType() != Type::VoidTy) {
2596 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
2597 MVT::ValueType VT = TLI.getValueType(PTy);
2598 Result = DAG.getNode(ISD::BIT_CONVERT, VT, Result);
2600 setValue(&I, Result);
2604 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
2605 static GlobalVariable *ExtractTypeInfo (Value *V) {
2606 V = IntrinsicInst::StripPointerCasts(V);
2607 GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2608 assert ((GV || isa<ConstantPointerNull>(V)) &&
2609 "TypeInfo must be a global variable or NULL");
2613 /// addCatchInfo - Extract the personality and type infos from an eh.selector
2614 /// call, and add them to the specified machine basic block.
2615 static void addCatchInfo(CallInst &I, MachineModuleInfo *MMI,
2616 MachineBasicBlock *MBB) {
2617 // Inform the MachineModuleInfo of the personality for this landing pad.
2618 ConstantExpr *CE = cast<ConstantExpr>(I.getOperand(2));
2619 assert(CE->getOpcode() == Instruction::BitCast &&
2620 isa<Function>(CE->getOperand(0)) &&
2621 "Personality should be a function");
2622 MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
2624 // Gather all the type infos for this landing pad and pass them along to
2625 // MachineModuleInfo.
2626 std::vector<GlobalVariable *> TyInfo;
2627 unsigned N = I.getNumOperands();
2629 for (unsigned i = N - 1; i > 2; --i) {
2630 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(i))) {
2631 unsigned FilterLength = CI->getZExtValue();
2632 unsigned FirstCatch = i + FilterLength + !FilterLength;
2633 assert (FirstCatch <= N && "Invalid filter length");
2635 if (FirstCatch < N) {
2636 TyInfo.reserve(N - FirstCatch);
2637 for (unsigned j = FirstCatch; j < N; ++j)
2638 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2639 MMI->addCatchTypeInfo(MBB, TyInfo);
2643 if (!FilterLength) {
2645 MMI->addCleanup(MBB);
2648 TyInfo.reserve(FilterLength - 1);
2649 for (unsigned j = i + 1; j < FirstCatch; ++j)
2650 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2651 MMI->addFilterTypeInfo(MBB, TyInfo);
2660 TyInfo.reserve(N - 3);
2661 for (unsigned j = 3; j < N; ++j)
2662 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2663 MMI->addCatchTypeInfo(MBB, TyInfo);
2667 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
2668 /// we want to emit this as a call to a named external function, return the name
2669 /// otherwise lower it and return null.
2671 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
2672 switch (Intrinsic) {
2674 // By default, turn this into a target intrinsic node.
2675 visitTargetIntrinsic(I, Intrinsic);
2677 case Intrinsic::vastart: visitVAStart(I); return 0;
2678 case Intrinsic::vaend: visitVAEnd(I); return 0;
2679 case Intrinsic::vacopy: visitVACopy(I); return 0;
2680 case Intrinsic::returnaddress:
2681 setValue(&I, DAG.getNode(ISD::RETURNADDR, TLI.getPointerTy(),
2682 getValue(I.getOperand(1))));
2684 case Intrinsic::frameaddress:
2685 setValue(&I, DAG.getNode(ISD::FRAMEADDR, TLI.getPointerTy(),
2686 getValue(I.getOperand(1))));
2688 case Intrinsic::setjmp:
2689 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
2691 case Intrinsic::longjmp:
2692 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
2694 case Intrinsic::memcpy_i32:
2695 case Intrinsic::memcpy_i64:
2696 visitMemIntrinsic(I, ISD::MEMCPY);
2698 case Intrinsic::memset_i32:
2699 case Intrinsic::memset_i64:
2700 visitMemIntrinsic(I, ISD::MEMSET);
2702 case Intrinsic::memmove_i32:
2703 case Intrinsic::memmove_i64:
2704 visitMemIntrinsic(I, ISD::MEMMOVE);
2707 case Intrinsic::dbg_stoppoint: {
2708 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2709 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
2710 if (MMI && SPI.getContext() && MMI->Verify(SPI.getContext())) {
2714 Ops[1] = getValue(SPI.getLineValue());
2715 Ops[2] = getValue(SPI.getColumnValue());
2717 DebugInfoDesc *DD = MMI->getDescFor(SPI.getContext());
2718 assert(DD && "Not a debug information descriptor");
2719 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD);
2721 Ops[3] = DAG.getString(CompileUnit->getFileName());
2722 Ops[4] = DAG.getString(CompileUnit->getDirectory());
2724 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5));
2729 case Intrinsic::dbg_region_start: {
2730 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2731 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
2732 if (MMI && RSI.getContext() && MMI->Verify(RSI.getContext())) {
2733 unsigned LabelID = MMI->RecordRegionStart(RSI.getContext());
2734 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
2735 DAG.getConstant(LabelID, MVT::i32),
2736 DAG.getConstant(0, MVT::i32)));
2741 case Intrinsic::dbg_region_end: {
2742 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2743 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
2744 if (MMI && REI.getContext() && MMI->Verify(REI.getContext())) {
2745 unsigned LabelID = MMI->RecordRegionEnd(REI.getContext());
2746 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
2747 DAG.getConstant(LabelID, MVT::i32),
2748 DAG.getConstant(0, MVT::i32)));
2753 case Intrinsic::dbg_func_start: {
2754 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2756 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
2757 Value *SP = FSI.getSubprogram();
2758 if (SP && MMI->Verify(SP)) {
2759 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is
2760 // what (most?) gdb expects.
2761 DebugInfoDesc *DD = MMI->getDescFor(SP);
2762 assert(DD && "Not a debug information descriptor");
2763 SubprogramDesc *Subprogram = cast<SubprogramDesc>(DD);
2764 const CompileUnitDesc *CompileUnit = Subprogram->getFile();
2765 unsigned SrcFile = MMI->RecordSource(CompileUnit->getDirectory(),
2766 CompileUnit->getFileName());
2767 // Record the source line but does create a label. It will be emitted
2768 // at asm emission time.
2769 MMI->RecordSourceLine(Subprogram->getLine(), 0, SrcFile);
2774 case Intrinsic::dbg_declare: {
2775 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2776 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
2777 Value *Variable = DI.getVariable();
2778 if (MMI && Variable && MMI->Verify(Variable))
2779 DAG.setRoot(DAG.getNode(ISD::DECLARE, MVT::Other, getRoot(),
2780 getValue(DI.getAddress()), getValue(Variable)));
2784 case Intrinsic::eh_exception: {
2785 if (ExceptionHandling) {
2786 if (!CurMBB->isLandingPad()) {
2787 // FIXME: Mark exception register as live in. Hack for PR1508.
2788 unsigned Reg = TLI.getExceptionAddressRegister();
2789 if (Reg) CurMBB->addLiveIn(Reg);
2791 // Insert the EXCEPTIONADDR instruction.
2792 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
2794 Ops[0] = DAG.getRoot();
2795 SDOperand Op = DAG.getNode(ISD::EXCEPTIONADDR, VTs, Ops, 1);
2797 DAG.setRoot(Op.getValue(1));
2799 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2804 case Intrinsic::eh_selector_i32:
2805 case Intrinsic::eh_selector_i64: {
2806 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2807 MVT::ValueType VT = (Intrinsic == Intrinsic::eh_selector_i32 ?
2808 MVT::i32 : MVT::i64);
2810 if (ExceptionHandling && MMI) {
2811 if (CurMBB->isLandingPad())
2812 addCatchInfo(I, MMI, CurMBB);
2815 FuncInfo.CatchInfoLost.insert(&I);
2817 // FIXME: Mark exception selector register as live in. Hack for PR1508.
2818 unsigned Reg = TLI.getExceptionSelectorRegister();
2819 if (Reg) CurMBB->addLiveIn(Reg);
2822 // Insert the EHSELECTION instruction.
2823 SDVTList VTs = DAG.getVTList(VT, MVT::Other);
2825 Ops[0] = getValue(I.getOperand(1));
2827 SDOperand Op = DAG.getNode(ISD::EHSELECTION, VTs, Ops, 2);
2829 DAG.setRoot(Op.getValue(1));
2831 setValue(&I, DAG.getConstant(0, VT));
2837 case Intrinsic::eh_typeid_for_i32:
2838 case Intrinsic::eh_typeid_for_i64: {
2839 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2840 MVT::ValueType VT = (Intrinsic == Intrinsic::eh_typeid_for_i32 ?
2841 MVT::i32 : MVT::i64);
2844 // Find the type id for the given typeinfo.
2845 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
2847 unsigned TypeID = MMI->getTypeIDFor(GV);
2848 setValue(&I, DAG.getConstant(TypeID, VT));
2850 // Return something different to eh_selector.
2851 setValue(&I, DAG.getConstant(1, VT));
2857 case Intrinsic::eh_return: {
2858 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2860 if (MMI && ExceptionHandling) {
2861 MMI->setCallsEHReturn(true);
2862 DAG.setRoot(DAG.getNode(ISD::EH_RETURN,
2865 getValue(I.getOperand(1)),
2866 getValue(I.getOperand(2))));
2868 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2874 case Intrinsic::eh_unwind_init: {
2875 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
2876 MMI->setCallsUnwindInit(true);
2882 case Intrinsic::eh_dwarf_cfa: {
2883 if (ExceptionHandling) {
2884 MVT::ValueType VT = getValue(I.getOperand(1)).getValueType();
2886 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getPointerTy()))
2887 CfaArg = DAG.getNode(ISD::TRUNCATE,
2888 TLI.getPointerTy(), getValue(I.getOperand(1)));
2890 CfaArg = DAG.getNode(ISD::SIGN_EXTEND,
2891 TLI.getPointerTy(), getValue(I.getOperand(1)));
2893 SDOperand Offset = DAG.getNode(ISD::ADD,
2895 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET,
2896 TLI.getPointerTy()),
2898 setValue(&I, DAG.getNode(ISD::ADD,
2900 DAG.getNode(ISD::FRAMEADDR,
2903 TLI.getPointerTy())),
2906 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2912 case Intrinsic::sqrt:
2913 setValue(&I, DAG.getNode(ISD::FSQRT,
2914 getValue(I.getOperand(1)).getValueType(),
2915 getValue(I.getOperand(1))));
2917 case Intrinsic::powi:
2918 setValue(&I, DAG.getNode(ISD::FPOWI,
2919 getValue(I.getOperand(1)).getValueType(),
2920 getValue(I.getOperand(1)),
2921 getValue(I.getOperand(2))));
2923 case Intrinsic::sin:
2924 setValue(&I, DAG.getNode(ISD::FSIN,
2925 getValue(I.getOperand(1)).getValueType(),
2926 getValue(I.getOperand(1))));
2928 case Intrinsic::cos:
2929 setValue(&I, DAG.getNode(ISD::FCOS,
2930 getValue(I.getOperand(1)).getValueType(),
2931 getValue(I.getOperand(1))));
2933 case Intrinsic::pow:
2934 setValue(&I, DAG.getNode(ISD::FPOW,
2935 getValue(I.getOperand(1)).getValueType(),
2936 getValue(I.getOperand(1)),
2937 getValue(I.getOperand(2))));
2939 case Intrinsic::pcmarker: {
2940 SDOperand Tmp = getValue(I.getOperand(1));
2941 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
2944 case Intrinsic::readcyclecounter: {
2945 SDOperand Op = getRoot();
2946 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER,
2947 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
2950 DAG.setRoot(Tmp.getValue(1));
2953 case Intrinsic::part_select: {
2954 // Currently not implemented: just abort
2955 assert(0 && "part_select intrinsic not implemented");
2958 case Intrinsic::part_set: {
2959 // Currently not implemented: just abort
2960 assert(0 && "part_set intrinsic not implemented");
2963 case Intrinsic::bswap:
2964 setValue(&I, DAG.getNode(ISD::BSWAP,
2965 getValue(I.getOperand(1)).getValueType(),
2966 getValue(I.getOperand(1))));
2968 case Intrinsic::cttz: {
2969 SDOperand Arg = getValue(I.getOperand(1));
2970 MVT::ValueType Ty = Arg.getValueType();
2971 SDOperand result = DAG.getNode(ISD::CTTZ, Ty, Arg);
2972 setValue(&I, result);
2975 case Intrinsic::ctlz: {
2976 SDOperand Arg = getValue(I.getOperand(1));
2977 MVT::ValueType Ty = Arg.getValueType();
2978 SDOperand result = DAG.getNode(ISD::CTLZ, Ty, Arg);
2979 setValue(&I, result);
2982 case Intrinsic::ctpop: {
2983 SDOperand Arg = getValue(I.getOperand(1));
2984 MVT::ValueType Ty = Arg.getValueType();
2985 SDOperand result = DAG.getNode(ISD::CTPOP, Ty, Arg);
2986 setValue(&I, result);
2989 case Intrinsic::stacksave: {
2990 SDOperand Op = getRoot();
2991 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE,
2992 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
2994 DAG.setRoot(Tmp.getValue(1));
2997 case Intrinsic::stackrestore: {
2998 SDOperand Tmp = getValue(I.getOperand(1));
2999 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
3002 case Intrinsic::var_annotation:
3003 // Discard annotate attributes
3006 case Intrinsic::init_trampoline: {
3008 cast<Function>(IntrinsicInst::StripPointerCasts(I.getOperand(2)));
3012 Ops[1] = getValue(I.getOperand(1));
3013 Ops[2] = getValue(I.getOperand(2));
3014 Ops[3] = getValue(I.getOperand(3));
3015 Ops[4] = DAG.getSrcValue(I.getOperand(1));
3016 Ops[5] = DAG.getSrcValue(F);
3018 SDOperand Tmp = DAG.getNode(ISD::TRAMPOLINE,
3019 DAG.getNodeValueTypes(TLI.getPointerTy(),
3024 DAG.setRoot(Tmp.getValue(1));
3028 case Intrinsic::gcroot:
3030 Value *Alloca = I.getOperand(1);
3031 Constant *TypeMap = cast<Constant>(I.getOperand(2));
3033 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).Val);
3034 GCI->addStackRoot(FI->getIndex(), TypeMap);
3038 case Intrinsic::gcread:
3039 case Intrinsic::gcwrite:
3040 assert(0 && "Collector failed to lower gcread/gcwrite intrinsics!");
3043 case Intrinsic::flt_rounds: {
3044 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, MVT::i32));
3048 case Intrinsic::trap: {
3049 DAG.setRoot(DAG.getNode(ISD::TRAP, MVT::Other, getRoot()));
3052 case Intrinsic::prefetch: {
3055 Ops[1] = getValue(I.getOperand(1));
3056 Ops[2] = getValue(I.getOperand(2));
3057 Ops[3] = getValue(I.getOperand(3));
3058 DAG.setRoot(DAG.getNode(ISD::PREFETCH, MVT::Other, &Ops[0], 4));
3062 case Intrinsic::memory_barrier: {
3065 for (int x = 1; x < 6; ++x)
3066 Ops[x] = getValue(I.getOperand(x));
3068 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, MVT::Other, &Ops[0], 6));
3071 case Intrinsic::atomic_lcs: {
3072 SDOperand Root = getRoot();
3073 SDOperand O3 = getValue(I.getOperand(3));
3074 SDOperand L = DAG.getAtomic(ISD::ATOMIC_LCS, Root,
3075 getValue(I.getOperand(1)),
3076 getValue(I.getOperand(2)),
3077 O3, O3.getValueType());
3079 DAG.setRoot(L.getValue(1));
3082 case Intrinsic::atomic_las: {
3083 SDOperand Root = getRoot();
3084 SDOperand O2 = getValue(I.getOperand(2));
3085 SDOperand L = DAG.getAtomic(ISD::ATOMIC_LAS, Root,
3086 getValue(I.getOperand(1)),
3087 O2, O2.getValueType());
3089 DAG.setRoot(L.getValue(1));
3092 case Intrinsic::atomic_swap: {
3093 SDOperand Root = getRoot();
3094 SDOperand O2 = getValue(I.getOperand(2));
3095 SDOperand L = DAG.getAtomic(ISD::ATOMIC_SWAP, Root,
3096 getValue(I.getOperand(1)),
3097 O2, O2.getValueType());
3099 DAG.setRoot(L.getValue(1));
3107 void SelectionDAGLowering::LowerCallTo(CallSite CS, SDOperand Callee,
3109 MachineBasicBlock *LandingPad) {
3110 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
3111 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
3112 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3113 unsigned BeginLabel = 0, EndLabel = 0;
3115 TargetLowering::ArgListTy Args;
3116 TargetLowering::ArgListEntry Entry;
3117 Args.reserve(CS.arg_size());
3118 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
3120 SDOperand ArgNode = getValue(*i);
3121 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
3123 unsigned attrInd = i - CS.arg_begin() + 1;
3124 Entry.isSExt = CS.paramHasAttr(attrInd, ParamAttr::SExt);
3125 Entry.isZExt = CS.paramHasAttr(attrInd, ParamAttr::ZExt);
3126 Entry.isInReg = CS.paramHasAttr(attrInd, ParamAttr::InReg);
3127 Entry.isSRet = CS.paramHasAttr(attrInd, ParamAttr::StructRet);
3128 Entry.isNest = CS.paramHasAttr(attrInd, ParamAttr::Nest);
3129 Entry.isByVal = CS.paramHasAttr(attrInd, ParamAttr::ByVal);
3130 Entry.Alignment = CS.getParamAlignment(attrInd);
3131 Args.push_back(Entry);
3134 bool MarkTryRange = LandingPad ||
3135 // C++ requires special handling of 'nounwind' calls.
3136 (CS.doesNotThrow());
3138 if (MarkTryRange && ExceptionHandling && MMI) {
3139 // Insert a label before the invoke call to mark the try range. This can be
3140 // used to detect deletion of the invoke via the MachineModuleInfo.
3141 BeginLabel = MMI->NextLabelID();
3142 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
3143 DAG.getConstant(BeginLabel, MVT::i32),
3144 DAG.getConstant(1, MVT::i32)));
3147 std::pair<SDOperand,SDOperand> Result =
3148 TLI.LowerCallTo(getRoot(), CS.getType(),
3149 CS.paramHasAttr(0, ParamAttr::SExt),
3150 CS.paramHasAttr(0, ParamAttr::ZExt),
3151 FTy->isVarArg(), CS.getCallingConv(), IsTailCall,
3153 if (CS.getType() != Type::VoidTy)
3154 setValue(CS.getInstruction(), Result.first);
3155 DAG.setRoot(Result.second);
3157 if (MarkTryRange && ExceptionHandling && MMI) {
3158 // Insert a label at the end of the invoke call to mark the try range. This
3159 // can be used to detect deletion of the invoke via the MachineModuleInfo.
3160 EndLabel = MMI->NextLabelID();
3161 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
3162 DAG.getConstant(EndLabel, MVT::i32),
3163 DAG.getConstant(1, MVT::i32)));
3165 // Inform MachineModuleInfo of range.
3166 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
3171 void SelectionDAGLowering::visitCall(CallInst &I) {
3172 const char *RenameFn = 0;
3173 if (Function *F = I.getCalledFunction()) {
3174 if (F->isDeclaration()) {
3175 if (unsigned IID = F->getIntrinsicID()) {
3176 RenameFn = visitIntrinsicCall(I, IID);
3182 // Check for well-known libc/libm calls. If the function is internal, it
3183 // can't be a library call.
3184 unsigned NameLen = F->getNameLen();
3185 if (!F->hasInternalLinkage() && NameLen) {
3186 const char *NameStr = F->getNameStart();
3187 if (NameStr[0] == 'c' &&
3188 ((NameLen == 8 && !strcmp(NameStr, "copysign")) ||
3189 (NameLen == 9 && !strcmp(NameStr, "copysignf")))) {
3190 if (I.getNumOperands() == 3 && // Basic sanity checks.
3191 I.getOperand(1)->getType()->isFloatingPoint() &&
3192 I.getType() == I.getOperand(1)->getType() &&
3193 I.getType() == I.getOperand(2)->getType()) {
3194 SDOperand LHS = getValue(I.getOperand(1));
3195 SDOperand RHS = getValue(I.getOperand(2));
3196 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
3200 } else if (NameStr[0] == 'f' &&
3201 ((NameLen == 4 && !strcmp(NameStr, "fabs")) ||
3202 (NameLen == 5 && !strcmp(NameStr, "fabsf")) ||
3203 (NameLen == 5 && !strcmp(NameStr, "fabsl")))) {
3204 if (I.getNumOperands() == 2 && // Basic sanity checks.
3205 I.getOperand(1)->getType()->isFloatingPoint() &&
3206 I.getType() == I.getOperand(1)->getType()) {
3207 SDOperand Tmp = getValue(I.getOperand(1));
3208 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
3211 } else if (NameStr[0] == 's' &&
3212 ((NameLen == 3 && !strcmp(NameStr, "sin")) ||
3213 (NameLen == 4 && !strcmp(NameStr, "sinf")) ||
3214 (NameLen == 4 && !strcmp(NameStr, "sinl")))) {
3215 if (I.getNumOperands() == 2 && // Basic sanity checks.
3216 I.getOperand(1)->getType()->isFloatingPoint() &&
3217 I.getType() == I.getOperand(1)->getType()) {
3218 SDOperand Tmp = getValue(I.getOperand(1));
3219 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
3222 } else if (NameStr[0] == 'c' &&
3223 ((NameLen == 3 && !strcmp(NameStr, "cos")) ||
3224 (NameLen == 4 && !strcmp(NameStr, "cosf")) ||
3225 (NameLen == 4 && !strcmp(NameStr, "cosl")))) {
3226 if (I.getNumOperands() == 2 && // Basic sanity checks.
3227 I.getOperand(1)->getType()->isFloatingPoint() &&
3228 I.getType() == I.getOperand(1)->getType()) {
3229 SDOperand Tmp = getValue(I.getOperand(1));
3230 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
3235 } else if (isa<InlineAsm>(I.getOperand(0))) {
3242 Callee = getValue(I.getOperand(0));
3244 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
3246 LowerCallTo(&I, Callee, I.isTailCall());
3250 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
3251 /// this value and returns the result as a ValueVT value. This uses
3252 /// Chain/Flag as the input and updates them for the output Chain/Flag.
3253 /// If the Flag pointer is NULL, no flag is used.
3254 SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
3255 SDOperand &Chain, SDOperand *Flag)const{
3256 // Copy the legal parts from the registers.
3257 unsigned NumParts = Regs.size();
3258 SmallVector<SDOperand, 8> Parts(NumParts);
3259 for (unsigned i = 0; i != NumParts; ++i) {
3260 SDOperand Part = Flag ?
3261 DAG.getCopyFromReg(Chain, Regs[i], RegVT, *Flag) :
3262 DAG.getCopyFromReg(Chain, Regs[i], RegVT);
3263 Chain = Part.getValue(1);
3265 *Flag = Part.getValue(2);
3269 // Assemble the legal parts into the final value.
3270 return getCopyFromParts(DAG, &Parts[0], NumParts, RegVT, ValueVT);
3273 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
3274 /// specified value into the registers specified by this object. This uses
3275 /// Chain/Flag as the input and updates them for the output Chain/Flag.
3276 /// If the Flag pointer is NULL, no flag is used.
3277 void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
3278 SDOperand &Chain, SDOperand *Flag) const {
3279 // Get the list of the values's legal parts.
3280 unsigned NumParts = Regs.size();
3281 SmallVector<SDOperand, 8> Parts(NumParts);
3282 getCopyToParts(DAG, Val, &Parts[0], NumParts, RegVT);
3284 // Copy the parts into the registers.
3285 for (unsigned i = 0; i != NumParts; ++i) {
3286 SDOperand Part = Flag ?
3287 DAG.getCopyToReg(Chain, Regs[i], Parts[i], *Flag) :
3288 DAG.getCopyToReg(Chain, Regs[i], Parts[i]);
3289 Chain = Part.getValue(0);
3291 *Flag = Part.getValue(1);
3295 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
3296 /// operand list. This adds the code marker and includes the number of
3297 /// values added into it.
3298 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
3299 std::vector<SDOperand> &Ops) const {
3300 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
3301 Ops.push_back(DAG.getTargetConstant(Code | (Regs.size() << 3), IntPtrTy));
3302 for (unsigned i = 0, e = Regs.size(); i != e; ++i)
3303 Ops.push_back(DAG.getRegister(Regs[i], RegVT));
3306 /// isAllocatableRegister - If the specified register is safe to allocate,
3307 /// i.e. it isn't a stack pointer or some other special register, return the
3308 /// register class for the register. Otherwise, return null.
3309 static const TargetRegisterClass *
3310 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
3311 const TargetLowering &TLI,
3312 const TargetRegisterInfo *TRI) {
3313 MVT::ValueType FoundVT = MVT::Other;
3314 const TargetRegisterClass *FoundRC = 0;
3315 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
3316 E = TRI->regclass_end(); RCI != E; ++RCI) {
3317 MVT::ValueType ThisVT = MVT::Other;
3319 const TargetRegisterClass *RC = *RCI;
3320 // If none of the the value types for this register class are valid, we
3321 // can't use it. For example, 64-bit reg classes on 32-bit targets.
3322 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
3324 if (TLI.isTypeLegal(*I)) {
3325 // If we have already found this register in a different register class,
3326 // choose the one with the largest VT specified. For example, on
3327 // PowerPC, we favor f64 register classes over f32.
3328 if (FoundVT == MVT::Other ||
3329 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) {
3336 if (ThisVT == MVT::Other) continue;
3338 // NOTE: This isn't ideal. In particular, this might allocate the
3339 // frame pointer in functions that need it (due to them not being taken
3340 // out of allocation, because a variable sized allocation hasn't been seen
3341 // yet). This is a slight code pessimization, but should still work.
3342 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
3343 E = RC->allocation_order_end(MF); I != E; ++I)
3345 // We found a matching register class. Keep looking at others in case
3346 // we find one with larger registers that this physreg is also in.
3357 /// AsmOperandInfo - This contains information for each constraint that we are
3359 struct SDISelAsmOperandInfo : public TargetLowering::AsmOperandInfo {
3360 /// CallOperand - If this is the result output operand or a clobber
3361 /// this is null, otherwise it is the incoming operand to the CallInst.
3362 /// This gets modified as the asm is processed.
3363 SDOperand CallOperand;
3365 /// AssignedRegs - If this is a register or register class operand, this
3366 /// contains the set of register corresponding to the operand.
3367 RegsForValue AssignedRegs;
3369 SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
3370 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
3373 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
3374 /// busy in OutputRegs/InputRegs.
3375 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
3376 std::set<unsigned> &OutputRegs,
3377 std::set<unsigned> &InputRegs,
3378 const TargetRegisterInfo &TRI) const {
3380 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
3381 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
3384 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
3385 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
3390 /// MarkRegAndAliases - Mark the specified register and all aliases in the
3392 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
3393 const TargetRegisterInfo &TRI) {
3394 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
3396 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
3397 for (; *Aliases; ++Aliases)
3398 Regs.insert(*Aliases);
3401 } // end anon namespace.
3404 /// GetRegistersForValue - Assign registers (virtual or physical) for the
3405 /// specified operand. We prefer to assign virtual registers, to allow the
3406 /// register allocator handle the assignment process. However, if the asm uses
3407 /// features that we can't model on machineinstrs, we have SDISel do the
3408 /// allocation. This produces generally horrible, but correct, code.
3410 /// OpInfo describes the operand.
3411 /// HasEarlyClobber is true if there are any early clobber constraints (=&r)
3412 /// or any explicitly clobbered registers.
3413 /// Input and OutputRegs are the set of already allocated physical registers.
3415 void SelectionDAGLowering::
3416 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo, bool HasEarlyClobber,
3417 std::set<unsigned> &OutputRegs,
3418 std::set<unsigned> &InputRegs) {
3419 // Compute whether this value requires an input register, an output register,
3421 bool isOutReg = false;
3422 bool isInReg = false;
3423 switch (OpInfo.Type) {
3424 case InlineAsm::isOutput:
3427 // If this is an early-clobber output, or if there is an input
3428 // constraint that matches this, we need to reserve the input register
3429 // so no other inputs allocate to it.
3430 isInReg = OpInfo.isEarlyClobber || OpInfo.hasMatchingInput;
3432 case InlineAsm::isInput:
3436 case InlineAsm::isClobber:
3443 MachineFunction &MF = DAG.getMachineFunction();
3444 std::vector<unsigned> Regs;
3446 // If this is a constraint for a single physreg, or a constraint for a
3447 // register class, find it.
3448 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
3449 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
3450 OpInfo.ConstraintVT);
3452 unsigned NumRegs = 1;
3453 if (OpInfo.ConstraintVT != MVT::Other)
3454 NumRegs = TLI.getNumRegisters(OpInfo.ConstraintVT);
3455 MVT::ValueType RegVT;
3456 MVT::ValueType ValueVT = OpInfo.ConstraintVT;
3459 // If this is a constraint for a specific physical register, like {r17},
3461 if (PhysReg.first) {
3462 if (OpInfo.ConstraintVT == MVT::Other)
3463 ValueVT = *PhysReg.second->vt_begin();
3465 // Get the actual register value type. This is important, because the user
3466 // may have asked for (e.g.) the AX register in i32 type. We need to
3467 // remember that AX is actually i16 to get the right extension.
3468 RegVT = *PhysReg.second->vt_begin();
3470 // This is a explicit reference to a physical register.
3471 Regs.push_back(PhysReg.first);
3473 // If this is an expanded reference, add the rest of the regs to Regs.
3475 TargetRegisterClass::iterator I = PhysReg.second->begin();
3476 TargetRegisterClass::iterator E = PhysReg.second->end();
3477 for (; *I != PhysReg.first; ++I)
3478 assert(I != E && "Didn't find reg!");
3480 // Already added the first reg.
3482 for (; NumRegs; --NumRegs, ++I) {
3483 assert(I != E && "Ran out of registers to allocate!");
3487 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
3488 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
3489 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
3493 // Otherwise, if this was a reference to an LLVM register class, create vregs
3494 // for this reference.
3495 std::vector<unsigned> RegClassRegs;
3496 const TargetRegisterClass *RC = PhysReg.second;
3498 // If this is an early clobber or tied register, our regalloc doesn't know
3499 // how to maintain the constraint. If it isn't, go ahead and create vreg
3500 // and let the regalloc do the right thing.
3501 if (!OpInfo.hasMatchingInput && !OpInfo.isEarlyClobber &&
3502 // If there is some other early clobber and this is an input register,
3503 // then we are forced to pre-allocate the input reg so it doesn't
3504 // conflict with the earlyclobber.
3505 !(OpInfo.Type == InlineAsm::isInput && HasEarlyClobber)) {
3506 RegVT = *PhysReg.second->vt_begin();
3508 if (OpInfo.ConstraintVT == MVT::Other)
3511 // Create the appropriate number of virtual registers.
3512 MachineRegisterInfo &RegInfo = MF.getRegInfo();
3513 for (; NumRegs; --NumRegs)
3514 Regs.push_back(RegInfo.createVirtualRegister(PhysReg.second));
3516 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
3520 // Otherwise, we can't allocate it. Let the code below figure out how to
3521 // maintain these constraints.
3522 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end());
3525 // This is a reference to a register class that doesn't directly correspond
3526 // to an LLVM register class. Allocate NumRegs consecutive, available,
3527 // registers from the class.
3528 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
3529 OpInfo.ConstraintVT);
3532 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
3533 unsigned NumAllocated = 0;
3534 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
3535 unsigned Reg = RegClassRegs[i];
3536 // See if this register is available.
3537 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
3538 (isInReg && InputRegs.count(Reg))) { // Already used.
3539 // Make sure we find consecutive registers.
3544 // Check to see if this register is allocatable (i.e. don't give out the
3547 RC = isAllocatableRegister(Reg, MF, TLI, TRI);
3548 if (!RC) { // Couldn't allocate this register.
3549 // Reset NumAllocated to make sure we return consecutive registers.
3555 // Okay, this register is good, we can use it.
3558 // If we allocated enough consecutive registers, succeed.
3559 if (NumAllocated == NumRegs) {
3560 unsigned RegStart = (i-NumAllocated)+1;
3561 unsigned RegEnd = i+1;
3562 // Mark all of the allocated registers used.
3563 for (unsigned i = RegStart; i != RegEnd; ++i)
3564 Regs.push_back(RegClassRegs[i]);
3566 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
3567 OpInfo.ConstraintVT);
3568 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
3573 // Otherwise, we couldn't allocate enough registers for this.
3578 /// visitInlineAsm - Handle a call to an InlineAsm object.
3580 void SelectionDAGLowering::visitInlineAsm(CallSite CS) {
3581 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
3583 /// ConstraintOperands - Information about all of the constraints.
3584 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
3586 SDOperand Chain = getRoot();
3589 std::set<unsigned> OutputRegs, InputRegs;
3591 // Do a prepass over the constraints, canonicalizing them, and building up the
3592 // ConstraintOperands list.
3593 std::vector<InlineAsm::ConstraintInfo>
3594 ConstraintInfos = IA->ParseConstraints();
3596 // SawEarlyClobber - Keep track of whether we saw an earlyclobber output
3597 // constraint. If so, we can't let the register allocator allocate any input
3598 // registers, because it will not know to avoid the earlyclobbered output reg.
3599 bool SawEarlyClobber = false;
3601 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
3602 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
3603 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
3604 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
3606 MVT::ValueType OpVT = MVT::Other;
3608 // Compute the value type for each operand.
3609 switch (OpInfo.Type) {
3610 case InlineAsm::isOutput:
3611 if (!OpInfo.isIndirect) {
3612 // The return value of the call is this value. As such, there is no
3613 // corresponding argument.
3614 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
3615 OpVT = TLI.getValueType(CS.getType());
3617 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
3620 case InlineAsm::isInput:
3621 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
3623 case InlineAsm::isClobber:
3628 // If this is an input or an indirect output, process the call argument.
3629 // BasicBlocks are labels, currently appearing only in asm's.
3630 if (OpInfo.CallOperandVal) {
3631 if (isa<BasicBlock>(OpInfo.CallOperandVal))
3632 OpInfo.CallOperand =
3633 DAG.getBasicBlock(FuncInfo.MBBMap[cast<BasicBlock>(
3634 OpInfo.CallOperandVal)]);
3636 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
3637 const Type *OpTy = OpInfo.CallOperandVal->getType();
3638 // If this is an indirect operand, the operand is a pointer to the
3640 if (OpInfo.isIndirect)
3641 OpTy = cast<PointerType>(OpTy)->getElementType();
3643 // If OpTy is not a first-class value, it may be a struct/union that we
3644 // can tile with integers.
3645 if (!OpTy->isFirstClassType() && OpTy->isSized()) {
3646 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
3654 OpTy = IntegerType::get(BitSize);
3659 OpVT = TLI.getValueType(OpTy, true);
3663 OpInfo.ConstraintVT = OpVT;
3665 // Compute the constraint code and ConstraintType to use.
3666 OpInfo.ComputeConstraintToUse(TLI);
3668 // Keep track of whether we see an earlyclobber.
3669 SawEarlyClobber |= OpInfo.isEarlyClobber;
3671 // If we see a clobber of a register, it is an early clobber.
3672 if (!SawEarlyClobber &&
3673 OpInfo.Type == InlineAsm::isClobber &&
3674 OpInfo.ConstraintType == TargetLowering::C_Register) {
3675 // Note that we want to ignore things that we don't trick here, like
3676 // dirflag, fpsr, flags, etc.
3677 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
3678 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
3679 OpInfo.ConstraintVT);
3680 if (PhysReg.first || PhysReg.second) {
3681 // This is a register we know of.
3682 SawEarlyClobber = true;
3686 // If this is a memory input, and if the operand is not indirect, do what we
3687 // need to to provide an address for the memory input.
3688 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3689 !OpInfo.isIndirect) {
3690 assert(OpInfo.Type == InlineAsm::isInput &&
3691 "Can only indirectify direct input operands!");
3693 // Memory operands really want the address of the value. If we don't have
3694 // an indirect input, put it in the constpool if we can, otherwise spill
3695 // it to a stack slot.
3697 // If the operand is a float, integer, or vector constant, spill to a
3698 // constant pool entry to get its address.
3699 Value *OpVal = OpInfo.CallOperandVal;
3700 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
3701 isa<ConstantVector>(OpVal)) {
3702 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
3703 TLI.getPointerTy());
3705 // Otherwise, create a stack slot and emit a store to it before the
3707 const Type *Ty = OpVal->getType();
3708 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
3709 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
3710 MachineFunction &MF = DAG.getMachineFunction();
3711 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align);
3712 SDOperand StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
3713 Chain = DAG.getStore(Chain, OpInfo.CallOperand, StackSlot, NULL, 0);
3714 OpInfo.CallOperand = StackSlot;
3717 // There is no longer a Value* corresponding to this operand.
3718 OpInfo.CallOperandVal = 0;
3719 // It is now an indirect operand.
3720 OpInfo.isIndirect = true;
3723 // If this constraint is for a specific register, allocate it before
3725 if (OpInfo.ConstraintType == TargetLowering::C_Register)
3726 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs);
3728 ConstraintInfos.clear();
3731 // Second pass - Loop over all of the operands, assigning virtual or physregs
3732 // to registerclass operands.
3733 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
3734 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
3736 // C_Register operands have already been allocated, Other/Memory don't need
3738 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
3739 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs);
3742 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
3743 std::vector<SDOperand> AsmNodeOperands;
3744 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain
3745 AsmNodeOperands.push_back(
3746 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other));
3749 // Loop over all of the inputs, copying the operand values into the
3750 // appropriate registers and processing the output regs.
3751 RegsForValue RetValRegs;
3753 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
3754 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
3756 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
3757 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
3759 switch (OpInfo.Type) {
3760 case InlineAsm::isOutput: {
3761 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
3762 OpInfo.ConstraintType != TargetLowering::C_Register) {
3763 // Memory output, or 'other' output (e.g. 'X' constraint).
3764 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
3766 // Add information to the INLINEASM node to know about this output.
3767 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
3768 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3769 TLI.getPointerTy()));
3770 AsmNodeOperands.push_back(OpInfo.CallOperand);
3774 // Otherwise, this is a register or register class output.
3776 // Copy the output from the appropriate register. Find a register that
3778 if (OpInfo.AssignedRegs.Regs.empty()) {
3779 cerr << "Couldn't allocate output reg for contraint '"
3780 << OpInfo.ConstraintCode << "'!\n";
3784 if (!OpInfo.isIndirect) {
3785 // This is the result value of the call.
3786 assert(RetValRegs.Regs.empty() &&
3787 "Cannot have multiple output constraints yet!");
3788 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
3789 RetValRegs = OpInfo.AssignedRegs;
3791 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
3792 OpInfo.CallOperandVal));
3795 // Add information to the INLINEASM node to know that this register is
3797 OpInfo.AssignedRegs.AddInlineAsmOperands(2 /*REGDEF*/, DAG,
3801 case InlineAsm::isInput: {
3802 SDOperand InOperandVal = OpInfo.CallOperand;
3804 if (isdigit(OpInfo.ConstraintCode[0])) { // Matching constraint?
3805 // If this is required to match an output register we have already set,
3806 // just use its register.
3807 unsigned OperandNo = atoi(OpInfo.ConstraintCode.c_str());
3809 // Scan until we find the definition we already emitted of this operand.
3810 // When we find it, create a RegsForValue operand.
3811 unsigned CurOp = 2; // The first operand.
3812 for (; OperandNo; --OperandNo) {
3813 // Advance to the next operand.
3815 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
3816 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
3817 (NumOps & 7) == 4 /*MEM*/) &&
3818 "Skipped past definitions?");
3819 CurOp += (NumOps>>3)+1;
3823 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
3824 if ((NumOps & 7) == 2 /*REGDEF*/) {
3825 // Add NumOps>>3 registers to MatchedRegs.
3826 RegsForValue MatchedRegs;
3827 MatchedRegs.ValueVT = InOperandVal.getValueType();
3828 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType();
3829 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
3831 cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
3832 MatchedRegs.Regs.push_back(Reg);
3835 // Use the produced MatchedRegs object to
3836 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
3837 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
3840 assert((NumOps & 7) == 4/*MEM*/ && "Unknown matching constraint!");
3841 assert((NumOps >> 3) == 1 && "Unexpected number of operands");
3842 // Add information to the INLINEASM node to know about this input.
3843 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
3844 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3845 TLI.getPointerTy()));
3846 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
3851 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
3852 assert(!OpInfo.isIndirect &&
3853 "Don't know how to handle indirect other inputs yet!");
3855 std::vector<SDOperand> Ops;
3856 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
3859 cerr << "Invalid operand for inline asm constraint '"
3860 << OpInfo.ConstraintCode << "'!\n";
3864 // Add information to the INLINEASM node to know about this input.
3865 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
3866 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3867 TLI.getPointerTy()));
3868 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
3870 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
3871 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
3872 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
3873 "Memory operands expect pointer values");
3875 // Add information to the INLINEASM node to know about this input.
3876 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
3877 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3878 TLI.getPointerTy()));
3879 AsmNodeOperands.push_back(InOperandVal);
3883 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
3884 OpInfo.ConstraintType == TargetLowering::C_Register) &&
3885 "Unknown constraint type!");
3886 assert(!OpInfo.isIndirect &&
3887 "Don't know how to handle indirect register inputs yet!");
3889 // Copy the input into the appropriate registers.
3890 assert(!OpInfo.AssignedRegs.Regs.empty() &&
3891 "Couldn't allocate input reg!");
3893 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
3895 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG,
3899 case InlineAsm::isClobber: {
3900 // Add the clobbered value to the operand list, so that the register
3901 // allocator is aware that the physreg got clobbered.
3902 if (!OpInfo.AssignedRegs.Regs.empty())
3903 OpInfo.AssignedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG,
3910 // Finish up input operands.
3911 AsmNodeOperands[0] = Chain;
3912 if (Flag.Val) AsmNodeOperands.push_back(Flag);
3914 Chain = DAG.getNode(ISD::INLINEASM,
3915 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
3916 &AsmNodeOperands[0], AsmNodeOperands.size());
3917 Flag = Chain.getValue(1);
3919 // If this asm returns a register value, copy the result from that register
3920 // and set it as the value of the call.
3921 if (!RetValRegs.Regs.empty()) {
3922 SDOperand Val = RetValRegs.getCopyFromRegs(DAG, Chain, &Flag);
3924 // If the result of the inline asm is a vector, it may have the wrong
3925 // width/num elts. Make sure to convert it to the right type with
3927 if (MVT::isVector(Val.getValueType())) {
3928 const VectorType *VTy = cast<VectorType>(CS.getType());
3929 MVT::ValueType DesiredVT = TLI.getValueType(VTy);
3931 Val = DAG.getNode(ISD::BIT_CONVERT, DesiredVT, Val);
3934 setValue(CS.getInstruction(), Val);
3937 std::vector<std::pair<SDOperand, Value*> > StoresToEmit;
3939 // Process indirect outputs, first output all of the flagged copies out of
3941 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
3942 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
3943 Value *Ptr = IndirectStoresToEmit[i].second;
3944 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, &Flag);
3945 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
3948 // Emit the non-flagged stores from the physregs.
3949 SmallVector<SDOperand, 8> OutChains;
3950 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
3951 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first,
3952 getValue(StoresToEmit[i].second),
3953 StoresToEmit[i].second, 0));
3954 if (!OutChains.empty())
3955 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
3956 &OutChains[0], OutChains.size());
3961 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
3962 SDOperand Src = getValue(I.getOperand(0));
3964 MVT::ValueType IntPtr = TLI.getPointerTy();
3966 if (IntPtr < Src.getValueType())
3967 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
3968 else if (IntPtr > Src.getValueType())
3969 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
3971 // Scale the source by the type size.
3972 uint64_t ElementSize = TD->getABITypeSize(I.getType()->getElementType());
3973 Src = DAG.getNode(ISD::MUL, Src.getValueType(),
3974 Src, DAG.getIntPtrConstant(ElementSize));
3976 TargetLowering::ArgListTy Args;
3977 TargetLowering::ArgListEntry Entry;
3979 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3980 Args.push_back(Entry);
3982 std::pair<SDOperand,SDOperand> Result =
3983 TLI.LowerCallTo(getRoot(), I.getType(), false, false, false, CallingConv::C,
3984 true, DAG.getExternalSymbol("malloc", IntPtr), Args, DAG);
3985 setValue(&I, Result.first); // Pointers always fit in registers
3986 DAG.setRoot(Result.second);
3989 void SelectionDAGLowering::visitFree(FreeInst &I) {
3990 TargetLowering::ArgListTy Args;
3991 TargetLowering::ArgListEntry Entry;
3992 Entry.Node = getValue(I.getOperand(0));
3993 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3994 Args.push_back(Entry);
3995 MVT::ValueType IntPtr = TLI.getPointerTy();
3996 std::pair<SDOperand,SDOperand> Result =
3997 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, false,
3998 CallingConv::C, true,
3999 DAG.getExternalSymbol("free", IntPtr), Args, DAG);
4000 DAG.setRoot(Result.second);
4003 // EmitInstrWithCustomInserter - This method should be implemented by targets
4004 // that mark instructions with the 'usesCustomDAGSchedInserter' flag. These
4005 // instructions are special in various ways, which require special support to
4006 // insert. The specified MachineInstr is created but not inserted into any
4007 // basic blocks, and the scheduler passes ownership of it to this method.
4008 MachineBasicBlock *TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
4009 MachineBasicBlock *MBB) {
4010 cerr << "If a target marks an instruction with "
4011 << "'usesCustomDAGSchedInserter', it must implement "
4012 << "TargetLowering::EmitInstrWithCustomInserter!\n";
4017 void SelectionDAGLowering::visitVAStart(CallInst &I) {
4018 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
4019 getValue(I.getOperand(1)),
4020 DAG.getSrcValue(I.getOperand(1))));
4023 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
4024 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
4025 getValue(I.getOperand(0)),
4026 DAG.getSrcValue(I.getOperand(0)));
4028 DAG.setRoot(V.getValue(1));
4031 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
4032 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
4033 getValue(I.getOperand(1)),
4034 DAG.getSrcValue(I.getOperand(1))));
4037 void SelectionDAGLowering::visitVACopy(CallInst &I) {
4038 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
4039 getValue(I.getOperand(1)),
4040 getValue(I.getOperand(2)),
4041 DAG.getSrcValue(I.getOperand(1)),
4042 DAG.getSrcValue(I.getOperand(2))));
4045 /// TargetLowering::LowerArguments - This is the default LowerArguments
4046 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
4047 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
4048 /// integrated into SDISel.
4049 std::vector<SDOperand>
4050 TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
4051 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
4052 std::vector<SDOperand> Ops;
4053 Ops.push_back(DAG.getRoot());
4054 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
4055 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
4057 // Add one result value for each formal argument.
4058 std::vector<MVT::ValueType> RetVals;
4060 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
4062 MVT::ValueType VT = getValueType(I->getType());
4063 unsigned Flags = ISD::ParamFlags::NoFlagSet;
4064 unsigned OriginalAlignment =
4065 getTargetData()->getABITypeAlignment(I->getType());
4067 // FIXME: Distinguish between a formal with no [sz]ext attribute from one
4068 // that is zero extended!
4069 if (F.paramHasAttr(j, ParamAttr::ZExt))
4070 Flags &= ~(ISD::ParamFlags::SExt);
4071 if (F.paramHasAttr(j, ParamAttr::SExt))
4072 Flags |= ISD::ParamFlags::SExt;
4073 if (F.paramHasAttr(j, ParamAttr::InReg))
4074 Flags |= ISD::ParamFlags::InReg;
4075 if (F.paramHasAttr(j, ParamAttr::StructRet))
4076 Flags |= ISD::ParamFlags::StructReturn;
4077 if (F.paramHasAttr(j, ParamAttr::ByVal)) {
4078 Flags |= ISD::ParamFlags::ByVal;
4079 const PointerType *Ty = cast<PointerType>(I->getType());
4080 const Type *ElementTy = Ty->getElementType();
4081 unsigned FrameAlign = Log2_32(getByValTypeAlignment(ElementTy));
4082 unsigned FrameSize = getTargetData()->getABITypeSize(ElementTy);
4083 // For ByVal, alignment should be passed from FE. BE will guess if
4084 // this info is not there but there are cases it cannot get right.
4085 if (F.getParamAlignment(j))
4086 FrameAlign = Log2_32(F.getParamAlignment(j));
4087 Flags |= (FrameAlign << ISD::ParamFlags::ByValAlignOffs);
4088 Flags |= (FrameSize << ISD::ParamFlags::ByValSizeOffs);
4090 if (F.paramHasAttr(j, ParamAttr::Nest))
4091 Flags |= ISD::ParamFlags::Nest;
4092 Flags |= (OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs);
4094 MVT::ValueType RegisterVT = getRegisterType(VT);
4095 unsigned NumRegs = getNumRegisters(VT);
4096 for (unsigned i = 0; i != NumRegs; ++i) {
4097 RetVals.push_back(RegisterVT);
4098 // if it isn't first piece, alignment must be 1
4100 Flags = (Flags & (~ISD::ParamFlags::OrigAlignment)) |
4101 (1 << ISD::ParamFlags::OrigAlignmentOffs);
4102 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
4106 RetVals.push_back(MVT::Other);
4109 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS,
4110 DAG.getVTList(&RetVals[0], RetVals.size()),
4111 &Ops[0], Ops.size()).Val;
4113 // Prelower FORMAL_ARGUMENTS. This isn't required for functionality, but
4114 // allows exposing the loads that may be part of the argument access to the
4115 // first DAGCombiner pass.
4116 SDOperand TmpRes = LowerOperation(SDOperand(Result, 0), DAG);
4118 // The number of results should match up, except that the lowered one may have
4119 // an extra flag result.
4120 assert((Result->getNumValues() == TmpRes.Val->getNumValues() ||
4121 (Result->getNumValues()+1 == TmpRes.Val->getNumValues() &&
4122 TmpRes.getValue(Result->getNumValues()).getValueType() == MVT::Flag))
4123 && "Lowering produced unexpected number of results!");
4124 Result = TmpRes.Val;
4126 unsigned NumArgRegs = Result->getNumValues() - 1;
4127 DAG.setRoot(SDOperand(Result, NumArgRegs));
4129 // Set up the return result vector.
4133 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
4135 MVT::ValueType VT = getValueType(I->getType());
4136 MVT::ValueType PartVT = getRegisterType(VT);
4138 unsigned NumParts = getNumRegisters(VT);
4139 SmallVector<SDOperand, 4> Parts(NumParts);
4140 for (unsigned j = 0; j != NumParts; ++j)
4141 Parts[j] = SDOperand(Result, i++);
4143 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4144 if (F.paramHasAttr(Idx, ParamAttr::SExt))
4145 AssertOp = ISD::AssertSext;
4146 else if (F.paramHasAttr(Idx, ParamAttr::ZExt))
4147 AssertOp = ISD::AssertZext;
4149 Ops.push_back(getCopyFromParts(DAG, &Parts[0], NumParts, PartVT, VT,
4152 assert(i == NumArgRegs && "Argument register count mismatch!");
4157 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
4158 /// implementation, which just inserts an ISD::CALL node, which is later custom
4159 /// lowered by the target to something concrete. FIXME: When all targets are
4160 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
4161 std::pair<SDOperand, SDOperand>
4162 TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
4163 bool RetSExt, bool RetZExt, bool isVarArg,
4164 unsigned CallingConv, bool isTailCall,
4166 ArgListTy &Args, SelectionDAG &DAG) {
4167 SmallVector<SDOperand, 32> Ops;
4168 Ops.push_back(Chain); // Op#0 - Chain
4169 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC
4170 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg
4171 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail
4172 Ops.push_back(Callee);
4174 // Handle all of the outgoing arguments.
4175 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
4176 MVT::ValueType VT = getValueType(Args[i].Ty);
4177 SDOperand Op = Args[i].Node;
4178 unsigned Flags = ISD::ParamFlags::NoFlagSet;
4179 unsigned OriginalAlignment =
4180 getTargetData()->getABITypeAlignment(Args[i].Ty);
4183 Flags |= ISD::ParamFlags::SExt;
4185 Flags |= ISD::ParamFlags::ZExt;
4186 if (Args[i].isInReg)
4187 Flags |= ISD::ParamFlags::InReg;
4189 Flags |= ISD::ParamFlags::StructReturn;
4190 if (Args[i].isByVal) {
4191 Flags |= ISD::ParamFlags::ByVal;
4192 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
4193 const Type *ElementTy = Ty->getElementType();
4194 unsigned FrameAlign = Log2_32(getByValTypeAlignment(ElementTy));
4195 unsigned FrameSize = getTargetData()->getABITypeSize(ElementTy);
4196 // For ByVal, alignment should come from FE. BE will guess if this
4197 // info is not there but there are cases it cannot get right.
4198 if (Args[i].Alignment)
4199 FrameAlign = Log2_32(Args[i].Alignment);
4200 Flags |= (FrameAlign << ISD::ParamFlags::ByValAlignOffs);
4201 Flags |= (FrameSize << ISD::ParamFlags::ByValSizeOffs);
4204 Flags |= ISD::ParamFlags::Nest;
4205 Flags |= OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs;
4207 MVT::ValueType PartVT = getRegisterType(VT);
4208 unsigned NumParts = getNumRegisters(VT);
4209 SmallVector<SDOperand, 4> Parts(NumParts);
4210 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
4213 ExtendKind = ISD::SIGN_EXTEND;
4214 else if (Args[i].isZExt)
4215 ExtendKind = ISD::ZERO_EXTEND;
4217 getCopyToParts(DAG, Op, &Parts[0], NumParts, PartVT, ExtendKind);
4219 for (unsigned i = 0; i != NumParts; ++i) {
4220 // if it isn't first piece, alignment must be 1
4221 unsigned MyFlags = Flags;
4223 MyFlags = (MyFlags & (~ISD::ParamFlags::OrigAlignment)) |
4224 (1 << ISD::ParamFlags::OrigAlignmentOffs);
4226 Ops.push_back(Parts[i]);
4227 Ops.push_back(DAG.getConstant(MyFlags, MVT::i32));
4231 // Figure out the result value types.
4232 MVT::ValueType VT = getValueType(RetTy);
4233 MVT::ValueType RegisterVT = getRegisterType(VT);
4234 unsigned NumRegs = getNumRegisters(VT);
4235 SmallVector<MVT::ValueType, 4> RetTys(NumRegs);
4236 for (unsigned i = 0; i != NumRegs; ++i)
4237 RetTys[i] = RegisterVT;
4239 RetTys.push_back(MVT::Other); // Always has a chain.
4241 // Create the CALL node.
4242 SDOperand Res = DAG.getNode(ISD::CALL,
4243 DAG.getVTList(&RetTys[0], NumRegs + 1),
4244 &Ops[0], Ops.size());
4245 Chain = Res.getValue(NumRegs);
4247 // Gather up the call result into a single value.
4248 if (RetTy != Type::VoidTy) {
4249 ISD::NodeType AssertOp = ISD::DELETED_NODE;
4252 AssertOp = ISD::AssertSext;
4254 AssertOp = ISD::AssertZext;
4256 SmallVector<SDOperand, 4> Results(NumRegs);
4257 for (unsigned i = 0; i != NumRegs; ++i)
4258 Results[i] = Res.getValue(i);
4259 Res = getCopyFromParts(DAG, &Results[0], NumRegs, RegisterVT, VT,
4263 return std::make_pair(Res, Chain);
4266 SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4267 assert(0 && "LowerOperation not implemented for this target!");
4272 SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op,
4273 SelectionDAG &DAG) {
4274 assert(0 && "CustomPromoteOperation not implemented for this target!");
4279 /// getMemsetValue - Vectorized representation of the memset value
4281 static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT,
4282 SelectionDAG &DAG) {
4283 MVT::ValueType CurVT = VT;
4284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
4285 uint64_t Val = C->getValue() & 255;
4287 while (CurVT != MVT::i8) {
4288 Val = (Val << Shift) | Val;
4290 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
4292 return DAG.getConstant(Val, VT);
4294 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
4296 while (CurVT != MVT::i8) {
4298 DAG.getNode(ISD::OR, VT,
4299 DAG.getNode(ISD::SHL, VT, Value,
4300 DAG.getConstant(Shift, MVT::i8)), Value);
4302 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
4309 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
4310 /// used when a memcpy is turned into a memset when the source is a constant
4312 static SDOperand getMemsetStringVal(MVT::ValueType VT,
4313 SelectionDAG &DAG, TargetLowering &TLI,
4314 std::string &Str, unsigned Offset) {
4316 unsigned MSB = MVT::getSizeInBits(VT) / 8;
4317 if (TLI.isLittleEndian())
4318 Offset = Offset + MSB - 1;
4319 for (unsigned i = 0; i != MSB; ++i) {
4320 Val = (Val << 8) | (unsigned char)Str[Offset];
4321 Offset += TLI.isLittleEndian() ? -1 : 1;
4323 return DAG.getConstant(Val, VT);
4326 /// getMemBasePlusOffset - Returns base and offset node for the
4327 static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset,
4328 SelectionDAG &DAG, TargetLowering &TLI) {
4329 MVT::ValueType VT = Base.getValueType();
4330 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
4333 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
4334 /// to replace the memset / memcpy is below the threshold. It also returns the
4335 /// types of the sequence of memory ops to perform memset / memcpy.
4336 static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps,
4337 unsigned Limit, uint64_t Size,
4338 unsigned Align, TargetLowering &TLI) {
4341 if (TLI.allowsUnalignedMemoryAccesses()) {
4344 switch (Align & 7) {
4360 MVT::ValueType LVT = MVT::i64;
4361 while (!TLI.isTypeLegal(LVT))
4362 LVT = (MVT::ValueType)((unsigned)LVT - 1);
4363 assert(MVT::isInteger(LVT));
4368 unsigned NumMemOps = 0;
4370 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4371 while (VTSize > Size) {
4372 VT = (MVT::ValueType)((unsigned)VT - 1);
4375 assert(MVT::isInteger(VT));
4377 if (++NumMemOps > Limit)
4379 MemOps.push_back(VT);
4386 void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) {
4387 SDOperand Op1 = getValue(I.getOperand(1));
4388 SDOperand Op2 = getValue(I.getOperand(2));
4389 SDOperand Op3 = getValue(I.getOperand(3));
4390 SDOperand Op4 = getValue(I.getOperand(4));
4391 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue();
4392 if (Align == 0) Align = 1;
4394 // If the source and destination are known to not be aliases, we can
4395 // lower memmove as memcpy.
4396 if (Op == ISD::MEMMOVE) {
4397 uint64_t Size = -1ULL;
4398 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4399 Size = C->getValue();
4400 if (AA.alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
4401 AliasAnalysis::NoAlias)
4405 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) {
4406 std::vector<MVT::ValueType> MemOps;
4408 // Expand memset / memcpy to a series of load / store ops
4409 // if the size operand falls below a certain threshold.
4410 SmallVector<SDOperand, 8> OutChains;
4412 default: break; // Do nothing for now.
4414 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(),
4415 Size->getValue(), Align, TLI)) {
4416 unsigned NumMemOps = MemOps.size();
4417 unsigned Offset = 0;
4418 for (unsigned i = 0; i < NumMemOps; i++) {
4419 MVT::ValueType VT = MemOps[i];
4420 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4421 SDOperand Value = getMemsetValue(Op2, VT, DAG);
4422 SDOperand Store = DAG.getStore(getRoot(), Value,
4423 getMemBasePlusOffset(Op1, Offset, DAG, TLI),
4424 I.getOperand(1), Offset);
4425 OutChains.push_back(Store);
4432 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(),
4433 Size->getValue(), Align, TLI)) {
4434 unsigned NumMemOps = MemOps.size();
4435 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0;
4436 GlobalAddressSDNode *G = NULL;
4438 bool CopyFromStr = false;
4440 if (Op2.getOpcode() == ISD::GlobalAddress)
4441 G = cast<GlobalAddressSDNode>(Op2);
4442 else if (Op2.getOpcode() == ISD::ADD &&
4443 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress &&
4444 Op2.getOperand(1).getOpcode() == ISD::Constant) {
4445 G = cast<GlobalAddressSDNode>(Op2.getOperand(0));
4446 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue();
4449 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
4450 if (GV && GV->isConstant()) {
4451 Str = GV->getStringValue(false);
4459 for (unsigned i = 0; i < NumMemOps; i++) {
4460 MVT::ValueType VT = MemOps[i];
4461 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4462 SDOperand Value, Chain, Store;
4465 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
4468 DAG.getStore(Chain, Value,
4469 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
4470 I.getOperand(1), DstOff);
4472 Value = DAG.getLoad(VT, getRoot(),
4473 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI),
4474 I.getOperand(2), SrcOff, false, Align);
4475 Chain = Value.getValue(1);
4477 DAG.getStore(Chain, Value,
4478 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
4479 I.getOperand(1), DstOff, false, Align);
4481 OutChains.push_back(Store);
4490 if (!OutChains.empty()) {
4491 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
4492 &OutChains[0], OutChains.size()));
4497 SDOperand AlwaysInline = DAG.getConstant(0, MVT::i1);
4501 assert(0 && "Unknown Op");
4503 Node = DAG.getMemcpy(getRoot(), Op1, Op2, Op3, Op4, AlwaysInline);
4506 Node = DAG.getMemmove(getRoot(), Op1, Op2, Op3, Op4, AlwaysInline);
4509 Node = DAG.getMemset(getRoot(), Op1, Op2, Op3, Op4, AlwaysInline);
4515 //===----------------------------------------------------------------------===//
4516 // SelectionDAGISel code
4517 //===----------------------------------------------------------------------===//
4519 unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) {
4520 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
4523 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
4524 AU.addRequired<AliasAnalysis>();
4525 AU.addRequired<CollectorModuleMetadata>();
4526 AU.setPreservesAll();
4531 bool SelectionDAGISel::runOnFunction(Function &Fn) {
4532 // Get alias analysis for load/store combining.
4533 AA = &getAnalysis<AliasAnalysis>();
4535 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
4536 if (MF.getFunction()->hasCollector())
4537 GCI = &getAnalysis<CollectorModuleMetadata>().get(*MF.getFunction());
4540 RegInfo = &MF.getRegInfo();
4541 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
4543 FunctionLoweringInfo FuncInfo(TLI, Fn, MF);
4545 if (ExceptionHandling)
4546 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
4547 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
4548 // Mark landing pad.
4549 FuncInfo.MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
4551 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
4552 SelectBasicBlock(I, MF, FuncInfo);
4554 // Add function live-ins to entry block live-in set.
4555 BasicBlock *EntryBB = &Fn.getEntryBlock();
4556 BB = FuncInfo.MBBMap[EntryBB];
4557 if (!RegInfo->livein_empty())
4558 for (MachineRegisterInfo::livein_iterator I = RegInfo->livein_begin(),
4559 E = RegInfo->livein_end(); I != E; ++I)
4560 BB->addLiveIn(I->first);
4563 assert(FuncInfo.CatchInfoFound.size() == FuncInfo.CatchInfoLost.size() &&
4564 "Not all catch info was assigned to a landing pad!");
4570 SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V,
4572 SDOperand Op = getValue(V);
4573 assert((Op.getOpcode() != ISD::CopyFromReg ||
4574 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
4575 "Copy from a reg to the same reg!");
4577 MVT::ValueType SrcVT = Op.getValueType();
4578 MVT::ValueType RegisterVT = TLI.getRegisterType(SrcVT);
4579 unsigned NumRegs = TLI.getNumRegisters(SrcVT);
4580 SmallVector<SDOperand, 8> Regs(NumRegs);
4581 SmallVector<SDOperand, 8> Chains(NumRegs);
4583 // Copy the value by legal parts into sequential virtual registers.
4584 getCopyToParts(DAG, Op, &Regs[0], NumRegs, RegisterVT);
4585 for (unsigned i = 0; i != NumRegs; ++i)
4586 Chains[i] = DAG.getCopyToReg(getRoot(), Reg + i, Regs[i]);
4587 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Chains[0], NumRegs);
4590 void SelectionDAGISel::
4591 LowerArguments(BasicBlock *LLVMBB, SelectionDAGLowering &SDL,
4592 std::vector<SDOperand> &UnorderedChains) {
4593 // If this is the entry block, emit arguments.
4594 Function &F = *LLVMBB->getParent();
4595 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo;
4596 SDOperand OldRoot = SDL.DAG.getRoot();
4597 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG);
4600 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
4602 if (!AI->use_empty()) {
4603 SDL.setValue(AI, Args[a]);
4605 // If this argument is live outside of the entry block, insert a copy from
4606 // whereever we got it to the vreg that other BB's will reference it as.
4607 DenseMap<const Value*, unsigned>::iterator VMI=FuncInfo.ValueMap.find(AI);
4608 if (VMI != FuncInfo.ValueMap.end()) {
4609 SDOperand Copy = SDL.CopyValueToVirtualRegister(AI, VMI->second);
4610 UnorderedChains.push_back(Copy);
4614 // Finally, if the target has anything special to do, allow it to do so.
4615 // FIXME: this should insert code into the DAG!
4616 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction());
4619 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
4620 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
4621 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
4622 if (isSelector(I)) {
4623 // Apply the catch info to DestBB.
4624 addCatchInfo(cast<CallInst>(*I), MMI, FLI.MBBMap[DestBB]);
4626 if (!FLI.MBBMap[SrcBB]->isLandingPad())
4627 FLI.CatchInfoFound.insert(I);
4632 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
4633 /// DAG and fixes their tailcall attribute operand.
4634 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
4635 TargetLowering& TLI) {
4636 SDNode * Ret = NULL;
4637 SDOperand Terminator = DAG.getRoot();
4640 if (Terminator.getOpcode() == ISD::RET) {
4641 Ret = Terminator.Val;
4644 // Fix tail call attribute of CALL nodes.
4645 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
4646 BI = prior(DAG.allnodes_end()); BI != BE; --BI) {
4647 if (BI->getOpcode() == ISD::CALL) {
4648 SDOperand OpRet(Ret, 0);
4649 SDOperand OpCall(static_cast<SDNode*>(BI), 0);
4650 bool isMarkedTailCall =
4651 cast<ConstantSDNode>(OpCall.getOperand(3))->getValue() != 0;
4652 // If CALL node has tail call attribute set to true and the call is not
4653 // eligible (no RET or the target rejects) the attribute is fixed to
4654 // false. The TargetLowering::IsEligibleForTailCallOptimization function
4655 // must correctly identify tail call optimizable calls.
4656 if (isMarkedTailCall &&
4658 !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG))) {
4659 SmallVector<SDOperand, 32> Ops;
4661 for(SDNode::op_iterator I =OpCall.Val->op_begin(),
4662 E=OpCall.Val->op_end(); I!=E; I++, idx++) {
4666 Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
4668 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
4674 void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB,
4675 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate,
4676 FunctionLoweringInfo &FuncInfo) {
4677 SelectionDAGLowering SDL(DAG, TLI, *AA, FuncInfo, GCI);
4679 std::vector<SDOperand> UnorderedChains;
4681 // Lower any arguments needed in this block if this is the entry block.
4682 if (LLVMBB == &LLVMBB->getParent()->getEntryBlock())
4683 LowerArguments(LLVMBB, SDL, UnorderedChains);
4685 BB = FuncInfo.MBBMap[LLVMBB];
4686 SDL.setCurrentBasicBlock(BB);
4688 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4690 if (ExceptionHandling && MMI && BB->isLandingPad()) {
4691 // Add a label to mark the beginning of the landing pad. Deletion of the
4692 // landing pad can thus be detected via the MachineModuleInfo.
4693 unsigned LabelID = MMI->addLandingPad(BB);
4694 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, DAG.getEntryNode(),
4695 DAG.getConstant(LabelID, MVT::i32),
4696 DAG.getConstant(1, MVT::i32)));
4698 // Mark exception register as live in.
4699 unsigned Reg = TLI.getExceptionAddressRegister();
4700 if (Reg) BB->addLiveIn(Reg);
4702 // Mark exception selector register as live in.
4703 Reg = TLI.getExceptionSelectorRegister();
4704 if (Reg) BB->addLiveIn(Reg);
4706 // FIXME: Hack around an exception handling flaw (PR1508): the personality
4707 // function and list of typeids logically belong to the invoke (or, if you
4708 // like, the basic block containing the invoke), and need to be associated
4709 // with it in the dwarf exception handling tables. Currently however the
4710 // information is provided by an intrinsic (eh.selector) that can be moved
4711 // to unexpected places by the optimizers: if the unwind edge is critical,
4712 // then breaking it can result in the intrinsics being in the successor of
4713 // the landing pad, not the landing pad itself. This results in exceptions
4714 // not being caught because no typeids are associated with the invoke.
4715 // This may not be the only way things can go wrong, but it is the only way
4716 // we try to work around for the moment.
4717 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
4719 if (Br && Br->isUnconditional()) { // Critical edge?
4720 BasicBlock::iterator I, E;
4721 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
4726 // No catch info found - try to extract some from the successor.
4727 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, FuncInfo);
4731 // Lower all of the non-terminator instructions.
4732 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end();
4736 // Ensure that all instructions which are used outside of their defining
4737 // blocks are available as virtual registers. Invoke is handled elsewhere.
4738 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I)
4739 if (!I->use_empty() && !isa<PHINode>(I) && !isa<InvokeInst>(I)) {
4740 DenseMap<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I);
4741 if (VMI != FuncInfo.ValueMap.end())
4742 UnorderedChains.push_back(
4743 SDL.CopyValueToVirtualRegister(I, VMI->second));
4746 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
4747 // ensure constants are generated when needed. Remember the virtual registers
4748 // that need to be added to the Machine PHI nodes as input. We cannot just
4749 // directly add them, because expansion might result in multiple MBB's for one
4750 // BB. As such, the start of the BB might correspond to a different MBB than
4753 TerminatorInst *TI = LLVMBB->getTerminator();
4755 // Emit constants only once even if used by multiple PHI nodes.
4756 std::map<Constant*, unsigned> ConstantsOut;
4758 // Vector bool would be better, but vector<bool> is really slow.
4759 std::vector<unsigned char> SuccsHandled;
4760 if (TI->getNumSuccessors())
4761 SuccsHandled.resize(BB->getParent()->getNumBlockIDs());
4763 // Check successor nodes' PHI nodes that expect a constant to be available
4765 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
4766 BasicBlock *SuccBB = TI->getSuccessor(succ);
4767 if (!isa<PHINode>(SuccBB->begin())) continue;
4768 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
4770 // If this terminator has multiple identical successors (common for
4771 // switches), only handle each succ once.
4772 unsigned SuccMBBNo = SuccMBB->getNumber();
4773 if (SuccsHandled[SuccMBBNo]) continue;
4774 SuccsHandled[SuccMBBNo] = true;
4776 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
4779 // At this point we know that there is a 1-1 correspondence between LLVM PHI
4780 // nodes and Machine PHI nodes, but the incoming operands have not been
4782 for (BasicBlock::iterator I = SuccBB->begin();
4783 (PN = dyn_cast<PHINode>(I)); ++I) {
4784 // Ignore dead phi's.
4785 if (PN->use_empty()) continue;
4788 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
4790 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
4791 unsigned &RegOut = ConstantsOut[C];
4793 RegOut = FuncInfo.CreateRegForValue(C);
4794 UnorderedChains.push_back(
4795 SDL.CopyValueToVirtualRegister(C, RegOut));
4799 Reg = FuncInfo.ValueMap[PHIOp];
4801 assert(isa<AllocaInst>(PHIOp) &&
4802 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
4803 "Didn't codegen value into a register!??");
4804 Reg = FuncInfo.CreateRegForValue(PHIOp);
4805 UnorderedChains.push_back(
4806 SDL.CopyValueToVirtualRegister(PHIOp, Reg));
4810 // Remember that this register needs to added to the machine PHI node as
4811 // the input for this MBB.
4812 MVT::ValueType VT = TLI.getValueType(PN->getType());
4813 unsigned NumRegisters = TLI.getNumRegisters(VT);
4814 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
4815 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
4818 ConstantsOut.clear();
4820 // Turn all of the unordered chains into one factored node.
4821 if (!UnorderedChains.empty()) {
4822 SDOperand Root = SDL.getRoot();
4823 if (Root.getOpcode() != ISD::EntryToken) {
4824 unsigned i = 0, e = UnorderedChains.size();
4825 for (; i != e; ++i) {
4826 assert(UnorderedChains[i].Val->getNumOperands() > 1);
4827 if (UnorderedChains[i].Val->getOperand(0) == Root)
4828 break; // Don't add the root if we already indirectly depend on it.
4832 UnorderedChains.push_back(Root);
4834 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
4835 &UnorderedChains[0], UnorderedChains.size()));
4838 // Lower the terminator after the copies are emitted.
4839 SDL.visit(*LLVMBB->getTerminator());
4841 // Copy over any CaseBlock records that may now exist due to SwitchInst
4842 // lowering, as well as any jump table information.
4843 SwitchCases.clear();
4844 SwitchCases = SDL.SwitchCases;
4846 JTCases = SDL.JTCases;
4847 BitTestCases.clear();
4848 BitTestCases = SDL.BitTestCases;
4850 // Make sure the root of the DAG is up-to-date.
4851 DAG.setRoot(SDL.getRoot());
4853 // Check whether calls in this block are real tail calls. Fix up CALL nodes
4854 // with correct tailcall attribute so that the target can rely on the tailcall
4855 // attribute indicating whether the call is really eligible for tail call
4857 CheckDAGForTailCallsAndFixThem(DAG, TLI);
4860 void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) {
4861 DOUT << "Lowered selection DAG:\n";
4864 // Run the DAG combiner in pre-legalize mode.
4865 DAG.Combine(false, *AA);
4867 DOUT << "Optimized lowered selection DAG:\n";
4870 // Second step, hack on the DAG until it only uses operations and types that
4871 // the target supports.
4872 #if 0 // Enable this some day.
4873 DAG.LegalizeTypes();
4874 // Someday even later, enable a dag combine pass here.
4878 DOUT << "Legalized selection DAG:\n";
4881 // Run the DAG combiner in post-legalize mode.
4882 DAG.Combine(true, *AA);
4884 DOUT << "Optimized legalized selection DAG:\n";
4887 if (ViewISelDAGs) DAG.viewGraph();
4889 // Third, instruction select all of the operations to machine code, adding the
4890 // code to the MachineBasicBlock.
4891 InstructionSelectBasicBlock(DAG);
4893 DOUT << "Selected machine code:\n";
4897 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF,
4898 FunctionLoweringInfo &FuncInfo) {
4899 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate;
4901 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4904 // First step, lower LLVM code to some DAG. This DAG may use operations and
4905 // types that are not supported by the target.
4906 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo);
4908 // Second step, emit the lowered DAG as machine code.
4909 CodeGenAndEmitDAG(DAG);
4912 DOUT << "Total amount of phi nodes to update: "
4913 << PHINodesToUpdate.size() << "\n";
4914 DEBUG(for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i)
4915 DOUT << "Node " << i << " : (" << PHINodesToUpdate[i].first
4916 << ", " << PHINodesToUpdate[i].second << ")\n";);
4918 // Next, now that we know what the last MBB the LLVM BB expanded is, update
4919 // PHI nodes in successors.
4920 if (SwitchCases.empty() && JTCases.empty() && BitTestCases.empty()) {
4921 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
4922 MachineInstr *PHI = PHINodesToUpdate[i].first;
4923 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4924 "This is not a machine PHI node that we are updating!");
4925 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[i].second,
4927 PHI->addOperand(MachineOperand::CreateMBB(BB));
4932 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) {
4933 // Lower header first, if it wasn't already lowered
4934 if (!BitTestCases[i].Emitted) {
4935 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4937 SelectionDAGLowering HSDL(HSDAG, TLI, *AA, FuncInfo, GCI);
4938 // Set the current basic block to the mbb we wish to insert the code into
4939 BB = BitTestCases[i].Parent;
4940 HSDL.setCurrentBasicBlock(BB);
4942 HSDL.visitBitTestHeader(BitTestCases[i]);
4943 HSDAG.setRoot(HSDL.getRoot());
4944 CodeGenAndEmitDAG(HSDAG);
4947 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) {
4948 SelectionDAG BSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4950 SelectionDAGLowering BSDL(BSDAG, TLI, *AA, FuncInfo, GCI);
4951 // Set the current basic block to the mbb we wish to insert the code into
4952 BB = BitTestCases[i].Cases[j].ThisBB;
4953 BSDL.setCurrentBasicBlock(BB);
4956 BSDL.visitBitTestCase(BitTestCases[i].Cases[j+1].ThisBB,
4957 BitTestCases[i].Reg,
4958 BitTestCases[i].Cases[j]);
4960 BSDL.visitBitTestCase(BitTestCases[i].Default,
4961 BitTestCases[i].Reg,
4962 BitTestCases[i].Cases[j]);
4965 BSDAG.setRoot(BSDL.getRoot());
4966 CodeGenAndEmitDAG(BSDAG);
4970 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
4971 MachineInstr *PHI = PHINodesToUpdate[pi].first;
4972 MachineBasicBlock *PHIBB = PHI->getParent();
4973 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4974 "This is not a machine PHI node that we are updating!");
4975 // This is "default" BB. We have two jumps to it. From "header" BB and
4976 // from last "case" BB.
4977 if (PHIBB == BitTestCases[i].Default) {
4978 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pi].second,
4980 PHI->addOperand(MachineOperand::CreateMBB(BitTestCases[i].Parent));
4981 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pi].second,
4983 PHI->addOperand(MachineOperand::CreateMBB(BitTestCases[i].Cases.
4986 // One of "cases" BB.
4987 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) {
4988 MachineBasicBlock* cBB = BitTestCases[i].Cases[j].ThisBB;
4989 if (cBB->succ_end() !=
4990 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
4991 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pi].second,
4993 PHI->addOperand(MachineOperand::CreateMBB(cBB));
4999 // If the JumpTable record is filled in, then we need to emit a jump table.
5000 // Updating the PHI nodes is tricky in this case, since we need to determine
5001 // whether the PHI is a successor of the range check MBB or the jump table MBB
5002 for (unsigned i = 0, e = JTCases.size(); i != e; ++i) {
5003 // Lower header first, if it wasn't already lowered
5004 if (!JTCases[i].first.Emitted) {
5005 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
5007 SelectionDAGLowering HSDL(HSDAG, TLI, *AA, FuncInfo, GCI);
5008 // Set the current basic block to the mbb we wish to insert the code into
5009 BB = JTCases[i].first.HeaderBB;
5010 HSDL.setCurrentBasicBlock(BB);
5012 HSDL.visitJumpTableHeader(JTCases[i].second, JTCases[i].first);
5013 HSDAG.setRoot(HSDL.getRoot());
5014 CodeGenAndEmitDAG(HSDAG);
5017 SelectionDAG JSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
5019 SelectionDAGLowering JSDL(JSDAG, TLI, *AA, FuncInfo, GCI);
5020 // Set the current basic block to the mbb we wish to insert the code into
5021 BB = JTCases[i].second.MBB;
5022 JSDL.setCurrentBasicBlock(BB);
5024 JSDL.visitJumpTable(JTCases[i].second);
5025 JSDAG.setRoot(JSDL.getRoot());
5026 CodeGenAndEmitDAG(JSDAG);
5029 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
5030 MachineInstr *PHI = PHINodesToUpdate[pi].first;
5031 MachineBasicBlock *PHIBB = PHI->getParent();
5032 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
5033 "This is not a machine PHI node that we are updating!");
5034 // "default" BB. We can go there only from header BB.
5035 if (PHIBB == JTCases[i].second.Default) {
5036 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pi].second,
5038 PHI->addOperand(MachineOperand::CreateMBB(JTCases[i].first.HeaderBB));
5040 // JT BB. Just iterate over successors here
5041 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
5042 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pi].second,
5044 PHI->addOperand(MachineOperand::CreateMBB(BB));
5049 // If the switch block involved a branch to one of the actual successors, we
5050 // need to update PHI nodes in that block.
5051 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
5052 MachineInstr *PHI = PHINodesToUpdate[i].first;
5053 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
5054 "This is not a machine PHI node that we are updating!");
5055 if (BB->isSuccessor(PHI->getParent())) {
5056 PHI->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[i].second,
5058 PHI->addOperand(MachineOperand::CreateMBB(BB));
5062 // If we generated any switch lowering information, build and codegen any
5063 // additional DAGs necessary.
5064 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) {
5065 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
5067 SelectionDAGLowering SDL(SDAG, TLI, *AA, FuncInfo, GCI);
5069 // Set the current basic block to the mbb we wish to insert the code into
5070 BB = SwitchCases[i].ThisBB;
5071 SDL.setCurrentBasicBlock(BB);
5074 SDL.visitSwitchCase(SwitchCases[i]);
5075 SDAG.setRoot(SDL.getRoot());
5076 CodeGenAndEmitDAG(SDAG);
5078 // Handle any PHI nodes in successors of this chunk, as if we were coming
5079 // from the original BB before switch expansion. Note that PHI nodes can
5080 // occur multiple times in PHINodesToUpdate. We have to be very careful to
5081 // handle them the right number of times.
5082 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
5083 for (MachineBasicBlock::iterator Phi = BB->begin();
5084 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
5085 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
5086 for (unsigned pn = 0; ; ++pn) {
5087 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!");
5088 if (PHINodesToUpdate[pn].first == Phi) {
5089 Phi->addOperand(MachineOperand::CreateReg(PHINodesToUpdate[pn].
5091 Phi->addOperand(MachineOperand::CreateMBB(SwitchCases[i].ThisBB));
5097 // Don't process RHS if same block as LHS.
5098 if (BB == SwitchCases[i].FalseBB)
5099 SwitchCases[i].FalseBB = 0;
5101 // If we haven't handled the RHS, do so now. Otherwise, we're done.
5102 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB;
5103 SwitchCases[i].FalseBB = 0;
5105 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0);
5110 //===----------------------------------------------------------------------===//
5111 /// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
5112 /// target node in the graph.
5113 void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) {
5114 if (ViewSchedDAGs) DAG.viewGraph();
5116 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
5120 RegisterScheduler::setDefault(Ctor);
5123 ScheduleDAG *SL = Ctor(this, &DAG, BB);
5126 if (ViewSUnitDAGs) SL->viewGraph();
5132 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
5133 return new HazardRecognizer();
5136 //===----------------------------------------------------------------------===//
5137 // Helper functions used by the generated instruction selector.
5138 //===----------------------------------------------------------------------===//
5139 // Calls to these methods are generated by tblgen.
5141 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
5142 /// the dag combiner simplified the 255, we still want to match. RHS is the
5143 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
5144 /// specified in the .td file (e.g. 255).
5145 bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS,
5146 int64_t DesiredMaskS) const {
5147 const APInt &ActualMask = RHS->getAPIntValue();
5148 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
5150 // If the actual mask exactly matches, success!
5151 if (ActualMask == DesiredMask)
5154 // If the actual AND mask is allowing unallowed bits, this doesn't match.
5155 if (ActualMask.intersects(~DesiredMask))
5158 // Otherwise, the DAG Combiner may have proven that the value coming in is
5159 // either already zero or is not demanded. Check for known zero input bits.
5160 APInt NeededMask = DesiredMask & ~ActualMask;
5161 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
5164 // TODO: check to see if missing bits are just not demanded.
5166 // Otherwise, this pattern doesn't match.
5170 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
5171 /// the dag combiner simplified the 255, we still want to match. RHS is the
5172 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
5173 /// specified in the .td file (e.g. 255).
5174 bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS,
5175 int64_t DesiredMaskS) const {
5176 const APInt &ActualMask = RHS->getAPIntValue();
5177 const APInt &DesiredMask = APInt(LHS.getValueSizeInBits(), DesiredMaskS);
5179 // If the actual mask exactly matches, success!
5180 if (ActualMask == DesiredMask)
5183 // If the actual AND mask is allowing unallowed bits, this doesn't match.
5184 if (ActualMask.intersects(~DesiredMask))
5187 // Otherwise, the DAG Combiner may have proven that the value coming in is
5188 // either already zero or is not demanded. Check for known zero input bits.
5189 APInt NeededMask = DesiredMask & ~ActualMask;
5191 APInt KnownZero, KnownOne;
5192 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
5194 // If all the missing bits in the or are already known to be set, match!
5195 if ((NeededMask & KnownOne) == NeededMask)
5198 // TODO: check to see if missing bits are just not demanded.
5200 // Otherwise, this pattern doesn't match.
5205 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
5206 /// by tblgen. Others should not call it.
5207 void SelectionDAGISel::
5208 SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) {
5209 std::vector<SDOperand> InOps;
5210 std::swap(InOps, Ops);
5212 Ops.push_back(InOps[0]); // input chain.
5213 Ops.push_back(InOps[1]); // input asm string.
5215 unsigned i = 2, e = InOps.size();
5216 if (InOps[e-1].getValueType() == MVT::Flag)
5217 --e; // Don't process a flag operand if it is here.
5220 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
5221 if ((Flags & 7) != 4 /*MEM*/) {
5222 // Just skip over this operand, copying the operands verbatim.
5223 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
5224 i += (Flags >> 3) + 1;
5226 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
5227 // Otherwise, this is a memory operand. Ask the target to select it.
5228 std::vector<SDOperand> SelOps;
5229 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) {
5230 cerr << "Could not match memory address. Inline asm failure!\n";
5234 // Add this to the output node.
5235 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
5236 Ops.push_back(DAG.getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
5238 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
5243 // Add the flag input back if present.
5244 if (e != InOps.size())
5245 Ops.push_back(InOps.back());
5248 char SelectionDAGISel::ID = 0;