1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "llvm/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/MachineModuleInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/SchedulerRegistry.h"
35 #include "llvm/CodeGen/SelectionDAG.h"
36 #include "llvm/CodeGen/SSARegMap.h"
37 #include "llvm/Target/MRegisterInfo.h"
38 #include "llvm/Target/TargetData.h"
39 #include "llvm/Target/TargetFrameInfo.h"
40 #include "llvm/Target/TargetInstrInfo.h"
41 #include "llvm/Target/TargetLowering.h"
42 #include "llvm/Target/TargetMachine.h"
43 #include "llvm/Target/TargetOptions.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/Compiler.h"
52 ViewISelDAGs("view-isel-dags", cl::Hidden,
53 cl::desc("Pop up a window to show isel dags as they are selected"));
55 ViewSchedDAGs("view-sched-dags", cl::Hidden,
56 cl::desc("Pop up a window to show sched dags as they are processed"));
58 ViewSUnitDAGs("view-sunit-dags", cl::Hidden,
59 cl::desc("Pop up a window to show SUnit dags after they are processed"));
61 static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0, ViewSUnitDAGs = 0;
64 //===---------------------------------------------------------------------===//
66 /// RegisterScheduler class - Track the registration of instruction schedulers.
68 //===---------------------------------------------------------------------===//
69 MachinePassRegistry RegisterScheduler::Registry;
71 //===---------------------------------------------------------------------===//
73 /// ISHeuristic command line option for instruction schedulers.
75 //===---------------------------------------------------------------------===//
77 cl::opt<RegisterScheduler::FunctionPassCtor, false,
78 RegisterPassParser<RegisterScheduler> >
79 ISHeuristic("pre-RA-sched",
80 cl::init(&createDefaultScheduler),
81 cl::desc("Instruction schedulers available (before register allocation):"));
83 static RegisterScheduler
84 defaultListDAGScheduler("default", " Best scheduler for the target",
85 createDefaultScheduler);
88 namespace { struct AsmOperandInfo; }
91 /// RegsForValue - This struct represents the physical registers that a
92 /// particular value is assigned and the type information about the value.
93 /// This is needed because values can be promoted into larger registers and
94 /// expanded into multiple smaller registers than the value.
95 struct VISIBILITY_HIDDEN RegsForValue {
96 /// Regs - This list holds the register (for legal and promoted values)
97 /// or register set (for expanded values) that the value should be assigned
99 std::vector<unsigned> Regs;
101 /// RegVT - The value type of each register.
103 MVT::ValueType RegVT;
105 /// ValueVT - The value type of the LLVM value, which may be promoted from
106 /// RegVT or made from merging the two expanded parts.
107 MVT::ValueType ValueVT;
109 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {}
111 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
112 : RegVT(regvt), ValueVT(valuevt) {
115 RegsForValue(const std::vector<unsigned> ®s,
116 MVT::ValueType regvt, MVT::ValueType valuevt)
117 : Regs(regs), RegVT(regvt), ValueVT(valuevt) {
120 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
121 /// this value and returns the result as a ValueVT value. This uses
122 /// Chain/Flag as the input and updates them for the output Chain/Flag.
123 /// If the Flag pointer is NULL, no flag is used.
124 SDOperand getCopyFromRegs(SelectionDAG &DAG,
125 SDOperand &Chain, SDOperand *Flag) const;
127 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
128 /// specified value into the registers specified by this object. This uses
129 /// Chain/Flag as the input and updates them for the output Chain/Flag.
130 /// If the Flag pointer is NULL, no flag is used.
131 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
132 SDOperand &Chain, SDOperand *Flag) const;
134 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
135 /// operand list. This adds the code marker and includes the number of
136 /// values added into it.
137 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
138 std::vector<SDOperand> &Ops) const;
143 //===--------------------------------------------------------------------===//
144 /// createDefaultScheduler - This creates an instruction scheduler appropriate
146 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
148 MachineBasicBlock *BB) {
149 TargetLowering &TLI = IS->getTargetLowering();
151 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
152 return createTDListDAGScheduler(IS, DAG, BB);
154 assert(TLI.getSchedulingPreference() ==
155 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
156 return createBURRListDAGScheduler(IS, DAG, BB);
161 //===--------------------------------------------------------------------===//
162 /// FunctionLoweringInfo - This contains information that is global to a
163 /// function that is used when lowering a region of the function.
164 class FunctionLoweringInfo {
171 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF);
173 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry.
174 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
176 /// ValueMap - Since we emit code for the function a basic block at a time,
177 /// we must remember which virtual registers hold the values for
178 /// cross-basic-block values.
179 DenseMap<const Value*, unsigned> ValueMap;
181 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in
182 /// the entry block. This allows the allocas to be efficiently referenced
183 /// anywhere in the function.
184 std::map<const AllocaInst*, int> StaticAllocaMap;
187 SmallSet<Instruction*, 8> CatchInfoLost;
188 SmallSet<Instruction*, 8> CatchInfoFound;
191 unsigned MakeReg(MVT::ValueType VT) {
192 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
195 /// isExportedInst - Return true if the specified value is an instruction
196 /// exported from its block.
197 bool isExportedInst(const Value *V) {
198 return ValueMap.count(V);
201 unsigned CreateRegForValue(const Value *V);
203 unsigned InitializeRegForValue(const Value *V) {
204 unsigned &R = ValueMap[V];
205 assert(R == 0 && "Already initialized this value register!");
206 return R = CreateRegForValue(V);
211 /// isSelector - Return true if this instruction is a call to the
212 /// eh.selector intrinsic.
213 static bool isSelector(Instruction *I) {
214 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
215 return (II->getIntrinsicID() == Intrinsic::eh_selector_i32 ||
216 II->getIntrinsicID() == Intrinsic::eh_selector_i64);
220 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
221 /// PHI nodes or outside of the basic block that defines it, or used by a
222 /// switch instruction, which may expand to multiple basic blocks.
223 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
224 if (isa<PHINode>(I)) return true;
225 BasicBlock *BB = I->getParent();
226 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
227 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
228 // FIXME: Remove switchinst special case.
229 isa<SwitchInst>(*UI))
234 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
235 /// entry block, return true. This includes arguments used by switches, since
236 /// the switch may expand into multiple basic blocks.
237 static bool isOnlyUsedInEntryBlock(Argument *A) {
238 BasicBlock *Entry = A->getParent()->begin();
239 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
240 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
241 return false; // Use not in entry block.
245 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli,
246 Function &fn, MachineFunction &mf)
247 : TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) {
249 // Create a vreg for each argument register that is not dead and is used
250 // outside of the entry block for the function.
251 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end();
253 if (!isOnlyUsedInEntryBlock(AI))
254 InitializeRegForValue(AI);
256 // Initialize the mapping of values to registers. This is only set up for
257 // instruction values that are used outside of the block that defines
259 Function::iterator BB = Fn.begin(), EB = Fn.end();
260 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
261 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
262 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
263 const Type *Ty = AI->getAllocatedType();
264 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
266 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
269 TySize *= CUI->getZExtValue(); // Get total allocated size.
270 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
271 StaticAllocaMap[AI] =
272 MF.getFrameInfo()->CreateStackObject(TySize, Align);
275 for (; BB != EB; ++BB)
276 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
277 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
278 if (!isa<AllocaInst>(I) ||
279 !StaticAllocaMap.count(cast<AllocaInst>(I)))
280 InitializeRegForValue(I);
282 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
283 // also creates the initial PHI MachineInstrs, though none of the input
284 // operands are populated.
285 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) {
286 MachineBasicBlock *MBB = new MachineBasicBlock(BB);
288 MF.getBasicBlockList().push_back(MBB);
290 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
293 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){
294 if (PN->use_empty()) continue;
296 MVT::ValueType VT = TLI.getValueType(PN->getType());
297 unsigned NumRegisters = TLI.getNumRegisters(VT);
298 unsigned PHIReg = ValueMap[PN];
299 assert(PHIReg && "PHI node does not have an assigned virtual register!");
300 const TargetInstrInfo *TII = TLI.getTargetMachine().getInstrInfo();
301 for (unsigned i = 0; i != NumRegisters; ++i)
302 BuildMI(MBB, TII->get(TargetInstrInfo::PHI), PHIReg+i);
307 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
308 /// the correctly promoted or expanded types. Assign these registers
309 /// consecutive vreg numbers and return the first assigned number.
310 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
311 MVT::ValueType VT = TLI.getValueType(V->getType());
313 unsigned NumRegisters = TLI.getNumRegisters(VT);
314 MVT::ValueType RegisterVT = TLI.getRegisterType(VT);
316 unsigned R = MakeReg(RegisterVT);
317 for (unsigned i = 1; i != NumRegisters; ++i)
323 //===----------------------------------------------------------------------===//
324 /// SelectionDAGLowering - This is the common target-independent lowering
325 /// implementation that is parameterized by a TargetLowering object.
326 /// Also, targets can overload any lowering method.
329 class SelectionDAGLowering {
330 MachineBasicBlock *CurMBB;
332 DenseMap<const Value*, SDOperand> NodeMap;
334 /// PendingLoads - Loads are not emitted to the program immediately. We bunch
335 /// them up and then emit token factor nodes when possible. This allows us to
336 /// get simple disambiguation between loads without worrying about alias
338 std::vector<SDOperand> PendingLoads;
340 /// Case - A struct to record the Value for a switch case, and the
341 /// case's target basic block.
345 MachineBasicBlock* BB;
347 Case() : Low(0), High(0), BB(0) { }
348 Case(Constant* low, Constant* high, MachineBasicBlock* bb) :
349 Low(low), High(high), BB(bb) { }
350 uint64_t size() const {
351 uint64_t rHigh = cast<ConstantInt>(High)->getSExtValue();
352 uint64_t rLow = cast<ConstantInt>(Low)->getSExtValue();
353 return (rHigh - rLow + 1ULL);
359 MachineBasicBlock* BB;
362 CaseBits(uint64_t mask, MachineBasicBlock* bb, unsigned bits):
363 Mask(mask), BB(bb), Bits(bits) { }
366 typedef std::vector<Case> CaseVector;
367 typedef std::vector<CaseBits> CaseBitsVector;
368 typedef CaseVector::iterator CaseItr;
369 typedef std::pair<CaseItr, CaseItr> CaseRange;
371 /// CaseRec - A struct with ctor used in lowering switches to a binary tree
372 /// of conditional branches.
374 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) :
375 CaseBB(bb), LT(lt), GE(ge), Range(r) {}
377 /// CaseBB - The MBB in which to emit the compare and branch
378 MachineBasicBlock *CaseBB;
379 /// LT, GE - If nonzero, we know the current case value must be less-than or
380 /// greater-than-or-equal-to these Constants.
383 /// Range - A pair of iterators representing the range of case values to be
384 /// processed at this point in the binary search tree.
388 typedef std::vector<CaseRec> CaseRecVector;
390 /// The comparison function for sorting the switch case values in the vector.
391 /// WARNING: Case ranges should be disjoint!
393 bool operator () (const Case& C1, const Case& C2) {
394 assert(isa<ConstantInt>(C1.Low) && isa<ConstantInt>(C2.High));
395 const ConstantInt* CI1 = cast<const ConstantInt>(C1.Low);
396 const ConstantInt* CI2 = cast<const ConstantInt>(C2.High);
397 return CI1->getValue().slt(CI2->getValue());
402 bool operator () (const CaseBits& C1, const CaseBits& C2) {
403 return C1.Bits > C2.Bits;
407 unsigned Clusterify(CaseVector& Cases, const SwitchInst &SI);
410 // TLI - This is information that describes the available target features we
411 // need for lowering. This indicates when operations are unavailable,
412 // implemented with a libcall, etc.
415 const TargetData *TD;
418 /// SwitchCases - Vector of CaseBlock structures used to communicate
419 /// SwitchInst code generation information.
420 std::vector<SelectionDAGISel::CaseBlock> SwitchCases;
421 /// JTCases - Vector of JumpTable structures used to communicate
422 /// SwitchInst code generation information.
423 std::vector<SelectionDAGISel::JumpTableBlock> JTCases;
424 std::vector<SelectionDAGISel::BitTestBlock> BitTestCases;
426 /// FuncInfo - Information about the function as a whole.
428 FunctionLoweringInfo &FuncInfo;
430 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli,
432 FunctionLoweringInfo &funcinfo)
433 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()), AA(aa),
437 /// getRoot - Return the current virtual root of the Selection DAG.
439 SDOperand getRoot() {
440 if (PendingLoads.empty())
441 return DAG.getRoot();
443 if (PendingLoads.size() == 1) {
444 SDOperand Root = PendingLoads[0];
446 PendingLoads.clear();
450 // Otherwise, we have to make a token factor node.
451 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
452 &PendingLoads[0], PendingLoads.size());
453 PendingLoads.clear();
458 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg);
460 void visit(Instruction &I) { visit(I.getOpcode(), I); }
462 void visit(unsigned Opcode, User &I) {
463 // Note: this doesn't use InstVisitor, because it has to work with
464 // ConstantExpr's in addition to instructions.
466 default: assert(0 && "Unknown instruction type encountered!");
468 // Build the switch statement using the Instruction.def file.
469 #define HANDLE_INST(NUM, OPCODE, CLASS) \
470 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
471 #include "llvm/Instruction.def"
475 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; }
477 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr,
478 const Value *SV, SDOperand Root,
479 bool isVolatile, unsigned Alignment);
481 SDOperand getIntPtrConstant(uint64_t Val) {
482 return DAG.getConstant(Val, TLI.getPointerTy());
485 SDOperand getValue(const Value *V);
487 void setValue(const Value *V, SDOperand NewN) {
488 SDOperand &N = NodeMap[V];
489 assert(N.Val == 0 && "Already set a value for this node!");
493 void GetRegistersForValue(AsmOperandInfo &OpInfo, bool HasEarlyClobber,
494 std::set<unsigned> &OutputRegs,
495 std::set<unsigned> &InputRegs);
497 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB,
498 MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
500 bool isExportableFromCurrentBlock(Value *V, const BasicBlock *FromBB);
501 void ExportFromCurrentBlock(Value *V);
502 void LowerCallTo(Instruction &I,
503 const Type *CalledValueTy, unsigned CallingConv,
504 bool IsTailCall, SDOperand Callee, unsigned OpIdx,
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(CallInst &I);
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 visitUserOp1(Instruction &I) {
612 assert(0 && "UserOp1 should not exist at instruction selection time!");
615 void visitUserOp2(Instruction &I) {
616 assert(0 && "UserOp2 should not exist at instruction selection time!");
620 } // end namespace llvm
623 /// getCopyFromParts - Create a value that contains the
624 /// specified legal parts combined into the value they represent.
625 static SDOperand getCopyFromParts(SelectionDAG &DAG,
626 const SDOperand *Parts,
628 MVT::ValueType PartVT,
629 MVT::ValueType ValueVT,
630 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
631 if (!MVT::isVector(ValueVT) || NumParts == 1) {
632 SDOperand Val = Parts[0];
634 // If the value was expanded, copy from the top part.
636 assert(NumParts == 2 &&
637 "Cannot expand to more than 2 elts yet!");
638 SDOperand Hi = Parts[1];
639 if (!DAG.getTargetLoweringInfo().isLittleEndian())
641 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi);
644 // Otherwise, if the value was promoted or extended, truncate it to the
646 if (PartVT == ValueVT)
649 if (MVT::isVector(PartVT)) {
650 assert(MVT::isVector(ValueVT) && "Unknown vector conversion!");
651 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
654 if (MVT::isVector(ValueVT)) {
655 assert(NumParts == 1 &&
656 MVT::getVectorElementType(ValueVT) == PartVT &&
657 MVT::getVectorNumElements(ValueVT) == 1 &&
658 "Only trivial scalar-to-vector conversions should get here!");
659 return DAG.getNode(ISD::BUILD_VECTOR, ValueVT, Val);
662 if (MVT::isInteger(PartVT) &&
663 MVT::isInteger(ValueVT)) {
664 if (ValueVT < PartVT) {
665 // For a truncate, see if we have any information to
666 // indicate whether the truncated bits will always be
667 // zero or sign-extension.
668 if (AssertOp != ISD::DELETED_NODE)
669 Val = DAG.getNode(AssertOp, PartVT, Val,
670 DAG.getValueType(ValueVT));
671 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
673 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val);
677 if (MVT::isFloatingPoint(PartVT) &&
678 MVT::isFloatingPoint(ValueVT))
679 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val);
681 if (MVT::getSizeInBits(PartVT) ==
682 MVT::getSizeInBits(ValueVT))
683 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
685 assert(0 && "Unknown mismatch!");
688 // Handle a multi-element vector.
689 MVT::ValueType IntermediateVT, RegisterVT;
690 unsigned NumIntermediates;
692 DAG.getTargetLoweringInfo()
693 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
696 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
697 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
698 assert(RegisterVT == Parts[0].getValueType() &&
699 "Part type doesn't match part!");
701 // Assemble the parts into intermediate operands.
702 SmallVector<SDOperand, 8> Ops(NumIntermediates);
703 if (NumIntermediates == NumParts) {
704 // If the register was not expanded, truncate or copy the value,
706 for (unsigned i = 0; i != NumParts; ++i)
707 Ops[i] = getCopyFromParts(DAG, &Parts[i], 1,
708 PartVT, IntermediateVT);
709 } else if (NumParts > 0) {
710 // If the intermediate type was expanded, build the intermediate operands
712 assert(NumParts % NumIntermediates == 0 &&
713 "Must expand into a divisible number of parts!");
714 unsigned Factor = NumParts / NumIntermediates;
715 for (unsigned i = 0; i != NumIntermediates; ++i)
716 Ops[i] = getCopyFromParts(DAG, &Parts[i * Factor], Factor,
717 PartVT, IntermediateVT);
720 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the intermediate
722 return DAG.getNode(MVT::isVector(IntermediateVT) ?
723 ISD::CONCAT_VECTORS :
725 ValueVT, &Ops[0], NumIntermediates);
728 /// getCopyToParts - Create a series of nodes that contain the
729 /// specified value split into legal parts.
730 static void getCopyToParts(SelectionDAG &DAG,
734 MVT::ValueType PartVT) {
735 TargetLowering &TLI = DAG.getTargetLoweringInfo();
736 MVT::ValueType PtrVT = TLI.getPointerTy();
737 MVT::ValueType ValueVT = Val.getValueType();
739 if (!MVT::isVector(ValueVT) || NumParts == 1) {
740 // If the value was expanded, copy from the parts.
742 for (unsigned i = 0; i != NumParts; ++i)
743 Parts[i] = DAG.getNode(ISD::EXTRACT_ELEMENT, PartVT, Val,
744 DAG.getConstant(i, PtrVT));
745 if (!DAG.getTargetLoweringInfo().isLittleEndian())
746 std::reverse(Parts, Parts + NumParts);
750 // If there is a single part and the types differ, this must be
752 if (PartVT != ValueVT) {
753 if (MVT::isVector(PartVT)) {
754 assert(MVT::isVector(ValueVT) &&
755 "Not a vector-vector cast?");
756 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
757 } else if (MVT::isVector(ValueVT)) {
758 assert(NumParts == 1 &&
759 MVT::getVectorElementType(ValueVT) == PartVT &&
760 MVT::getVectorNumElements(ValueVT) == 1 &&
761 "Only trivial vector-to-scalar conversions should get here!");
762 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, PartVT, Val,
763 DAG.getConstant(0, PtrVT));
764 } else if (MVT::isInteger(PartVT) && MVT::isInteger(ValueVT)) {
765 if (PartVT < ValueVT)
766 Val = DAG.getNode(ISD::TRUNCATE, PartVT, Val);
768 Val = DAG.getNode(ISD::ANY_EXTEND, PartVT, Val);
769 } else if (MVT::isFloatingPoint(PartVT) &&
770 MVT::isFloatingPoint(ValueVT)) {
771 Val = DAG.getNode(ISD::FP_EXTEND, PartVT, Val);
772 } else if (MVT::getSizeInBits(PartVT) ==
773 MVT::getSizeInBits(ValueVT)) {
774 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
776 assert(0 && "Unknown mismatch!");
783 // Handle a multi-element vector.
784 MVT::ValueType IntermediateVT, RegisterVT;
785 unsigned NumIntermediates;
787 DAG.getTargetLoweringInfo()
788 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
790 unsigned NumElements = MVT::getVectorNumElements(ValueVT);
792 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
793 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
795 // Split the vector into intermediate operands.
796 SmallVector<SDOperand, 8> Ops(NumIntermediates);
797 for (unsigned i = 0; i != NumIntermediates; ++i)
798 if (MVT::isVector(IntermediateVT))
799 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR,
801 DAG.getConstant(i * (NumElements / NumIntermediates),
804 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
806 DAG.getConstant(i, PtrVT));
808 // Split the intermediate operands into legal parts.
809 if (NumParts == NumIntermediates) {
810 // If the register was not expanded, promote or copy the value,
812 for (unsigned i = 0; i != NumParts; ++i)
813 getCopyToParts(DAG, Ops[i], &Parts[i], 1, PartVT);
814 } else if (NumParts > 0) {
815 // If the intermediate type was expanded, split each the value into
817 assert(NumParts % NumIntermediates == 0 &&
818 "Must expand into a divisible number of parts!");
819 unsigned Factor = NumParts / NumIntermediates;
820 for (unsigned i = 0; i != NumIntermediates; ++i)
821 getCopyToParts(DAG, Ops[i], &Parts[i * Factor], Factor, PartVT);
826 SDOperand SelectionDAGLowering::getValue(const Value *V) {
827 SDOperand &N = NodeMap[V];
830 const Type *VTy = V->getType();
831 MVT::ValueType VT = TLI.getValueType(VTy);
832 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
833 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
834 visit(CE->getOpcode(), *CE);
835 SDOperand N1 = NodeMap[V];
836 assert(N1.Val && "visit didn't populate the ValueMap!");
838 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
839 return N = DAG.getGlobalAddress(GV, VT);
840 } else if (isa<ConstantPointerNull>(C)) {
841 return N = DAG.getConstant(0, TLI.getPointerTy());
842 } else if (isa<UndefValue>(C)) {
843 if (!isa<VectorType>(VTy))
844 return N = DAG.getNode(ISD::UNDEF, VT);
846 // Create a BUILD_VECTOR of undef nodes.
847 const VectorType *PTy = cast<VectorType>(VTy);
848 unsigned NumElements = PTy->getNumElements();
849 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
851 SmallVector<SDOperand, 8> Ops;
852 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT));
854 // Create a VConstant node with generic Vector type.
855 MVT::ValueType VT = MVT::getVectorType(PVT, NumElements);
856 return N = DAG.getNode(ISD::BUILD_VECTOR, VT,
857 &Ops[0], Ops.size());
858 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
859 return N = DAG.getConstantFP(CFP->getValueAPF(), VT);
860 } else if (const VectorType *PTy = dyn_cast<VectorType>(VTy)) {
861 unsigned NumElements = PTy->getNumElements();
862 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
864 // Now that we know the number and type of the elements, push a
865 // Constant or ConstantFP node onto the ops list for each element of
866 // the vector constant.
867 SmallVector<SDOperand, 8> Ops;
868 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
869 for (unsigned i = 0; i != NumElements; ++i)
870 Ops.push_back(getValue(CP->getOperand(i)));
872 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
874 if (MVT::isFloatingPoint(PVT))
875 Op = DAG.getConstantFP(0, PVT);
877 Op = DAG.getConstant(0, PVT);
878 Ops.assign(NumElements, Op);
881 // Create a BUILD_VECTOR node.
882 MVT::ValueType VT = MVT::getVectorType(PVT, NumElements);
883 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0],
886 // Canonicalize all constant ints to be unsigned.
887 return N = DAG.getConstant(cast<ConstantInt>(C)->getZExtValue(),VT);
891 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
892 std::map<const AllocaInst*, int>::iterator SI =
893 FuncInfo.StaticAllocaMap.find(AI);
894 if (SI != FuncInfo.StaticAllocaMap.end())
895 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
898 unsigned InReg = FuncInfo.ValueMap[V];
899 assert(InReg && "Value not in map!");
901 MVT::ValueType RegisterVT = TLI.getRegisterType(VT);
902 unsigned NumRegs = TLI.getNumRegisters(VT);
904 std::vector<unsigned> Regs(NumRegs);
905 for (unsigned i = 0; i != NumRegs; ++i)
908 RegsForValue RFV(Regs, RegisterVT, VT);
909 SDOperand Chain = DAG.getEntryNode();
911 return RFV.getCopyFromRegs(DAG, Chain, NULL);
915 void SelectionDAGLowering::visitRet(ReturnInst &I) {
916 if (I.getNumOperands() == 0) {
917 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot()));
920 SmallVector<SDOperand, 8> NewValues;
921 NewValues.push_back(getRoot());
922 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
923 SDOperand RetOp = getValue(I.getOperand(i));
925 // If this is an integer return value, we need to promote it ourselves to
926 // the full width of a register, since getCopyToParts and Legalize will use
927 // ANY_EXTEND rather than sign/zero.
928 // FIXME: C calling convention requires the return type to be promoted to
929 // at least 32-bit. But this is not necessary for non-C calling conventions.
930 if (MVT::isInteger(RetOp.getValueType()) &&
931 RetOp.getValueType() < MVT::i64) {
932 MVT::ValueType TmpVT;
933 if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote)
934 TmpVT = TLI.getTypeToTransformTo(MVT::i32);
937 const FunctionType *FTy = I.getParent()->getParent()->getFunctionType();
938 const ParamAttrsList *Attrs = FTy->getParamAttrs();
939 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
940 if (Attrs && Attrs->paramHasAttr(0, ParamAttr::SExt))
941 ExtendKind = ISD::SIGN_EXTEND;
942 if (Attrs && Attrs->paramHasAttr(0, ParamAttr::ZExt))
943 ExtendKind = ISD::ZERO_EXTEND;
944 RetOp = DAG.getNode(ExtendKind, TmpVT, RetOp);
945 NewValues.push_back(RetOp);
946 NewValues.push_back(DAG.getConstant(false, MVT::i32));
948 MVT::ValueType VT = RetOp.getValueType();
949 unsigned NumParts = TLI.getNumRegisters(VT);
950 MVT::ValueType PartVT = TLI.getRegisterType(VT);
951 SmallVector<SDOperand, 4> Parts(NumParts);
952 getCopyToParts(DAG, RetOp, &Parts[0], NumParts, PartVT);
953 for (unsigned i = 0; i < NumParts; ++i) {
954 NewValues.push_back(Parts[i]);
955 NewValues.push_back(DAG.getConstant(false, MVT::i32));
959 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other,
960 &NewValues[0], NewValues.size()));
963 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
964 /// the current basic block, add it to ValueMap now so that we'll get a
966 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
967 // No need to export constants.
968 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
971 if (FuncInfo.isExportedInst(V)) return;
973 unsigned Reg = FuncInfo.InitializeRegForValue(V);
974 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg));
977 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
978 const BasicBlock *FromBB) {
979 // The operands of the setcc have to be in this block. We don't know
980 // how to export them from some other block.
981 if (Instruction *VI = dyn_cast<Instruction>(V)) {
982 // Can export from current BB.
983 if (VI->getParent() == FromBB)
986 // Is already exported, noop.
987 return FuncInfo.isExportedInst(V);
990 // If this is an argument, we can export it if the BB is the entry block or
991 // if it is already exported.
992 if (isa<Argument>(V)) {
993 if (FromBB == &FromBB->getParent()->getEntryBlock())
996 // Otherwise, can only export this if it is already exported.
997 return FuncInfo.isExportedInst(V);
1000 // Otherwise, constants can always be exported.
1004 static bool InBlock(const Value *V, const BasicBlock *BB) {
1005 if (const Instruction *I = dyn_cast<Instruction>(V))
1006 return I->getParent() == BB;
1010 /// FindMergedConditions - If Cond is an expression like
1011 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
1012 MachineBasicBlock *TBB,
1013 MachineBasicBlock *FBB,
1014 MachineBasicBlock *CurBB,
1016 // If this node is not part of the or/and tree, emit it as a branch.
1017 Instruction *BOp = dyn_cast<Instruction>(Cond);
1019 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1020 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1021 BOp->getParent() != CurBB->getBasicBlock() ||
1022 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1023 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1024 const BasicBlock *BB = CurBB->getBasicBlock();
1026 // If the leaf of the tree is a comparison, merge the condition into
1028 if ((isa<ICmpInst>(Cond) || isa<FCmpInst>(Cond)) &&
1029 // The operands of the cmp have to be in this block. We don't know
1030 // how to export them from some other block. If this is the first block
1031 // of the sequence, no exporting is needed.
1033 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1034 isExportableFromCurrentBlock(BOp->getOperand(1), BB)))) {
1035 BOp = cast<Instruction>(Cond);
1036 ISD::CondCode Condition;
1037 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1038 switch (IC->getPredicate()) {
1039 default: assert(0 && "Unknown icmp predicate opcode!");
1040 case ICmpInst::ICMP_EQ: Condition = ISD::SETEQ; break;
1041 case ICmpInst::ICMP_NE: Condition = ISD::SETNE; break;
1042 case ICmpInst::ICMP_SLE: Condition = ISD::SETLE; break;
1043 case ICmpInst::ICMP_ULE: Condition = ISD::SETULE; break;
1044 case ICmpInst::ICMP_SGE: Condition = ISD::SETGE; break;
1045 case ICmpInst::ICMP_UGE: Condition = ISD::SETUGE; break;
1046 case ICmpInst::ICMP_SLT: Condition = ISD::SETLT; break;
1047 case ICmpInst::ICMP_ULT: Condition = ISD::SETULT; break;
1048 case ICmpInst::ICMP_SGT: Condition = ISD::SETGT; break;
1049 case ICmpInst::ICMP_UGT: Condition = ISD::SETUGT; break;
1051 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1052 ISD::CondCode FPC, FOC;
1053 switch (FC->getPredicate()) {
1054 default: assert(0 && "Unknown fcmp predicate opcode!");
1055 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1056 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1057 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1058 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1059 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1060 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1061 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1062 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
1063 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
1064 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1065 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1066 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1067 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1068 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1069 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1070 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1072 if (FiniteOnlyFPMath())
1077 Condition = ISD::SETEQ; // silence warning.
1078 assert(0 && "Unknown compare instruction");
1081 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0),
1082 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1083 SwitchCases.push_back(CB);
1087 // Create a CaseBlock record representing this branch.
1088 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(),
1089 NULL, TBB, FBB, CurBB);
1090 SwitchCases.push_back(CB);
1095 // Create TmpBB after CurBB.
1096 MachineFunction::iterator BBI = CurBB;
1097 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock());
1098 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB);
1100 if (Opc == Instruction::Or) {
1101 // Codegen X | Y as:
1109 // Emit the LHS condition.
1110 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1112 // Emit the RHS condition into TmpBB.
1113 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1115 assert(Opc == Instruction::And && "Unknown merge op!");
1116 // Codegen X & Y as:
1123 // This requires creation of TmpBB after CurBB.
1125 // Emit the LHS condition.
1126 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1128 // Emit the RHS condition into TmpBB.
1129 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1133 /// If the set of cases should be emitted as a series of branches, return true.
1134 /// If we should emit this as a bunch of and/or'd together conditions, return
1137 ShouldEmitAsBranches(const std::vector<SelectionDAGISel::CaseBlock> &Cases) {
1138 if (Cases.size() != 2) return true;
1140 // If this is two comparisons of the same values or'd or and'd together, they
1141 // will get folded into a single comparison, so don't emit two blocks.
1142 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1143 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1144 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1145 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1152 void SelectionDAGLowering::visitBr(BranchInst &I) {
1153 // Update machine-CFG edges.
1154 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1156 // Figure out which block is immediately after the current one.
1157 MachineBasicBlock *NextBlock = 0;
1158 MachineFunction::iterator BBI = CurMBB;
1159 if (++BBI != CurMBB->getParent()->end())
1162 if (I.isUnconditional()) {
1163 // If this is not a fall-through branch, emit the branch.
1164 if (Succ0MBB != NextBlock)
1165 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1166 DAG.getBasicBlock(Succ0MBB)));
1168 // Update machine-CFG edges.
1169 CurMBB->addSuccessor(Succ0MBB);
1174 // If this condition is one of the special cases we handle, do special stuff
1176 Value *CondVal = I.getCondition();
1177 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1179 // If this is a series of conditions that are or'd or and'd together, emit
1180 // this as a sequence of branches instead of setcc's with and/or operations.
1181 // For example, instead of something like:
1194 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1195 if (BOp->hasOneUse() &&
1196 (BOp->getOpcode() == Instruction::And ||
1197 BOp->getOpcode() == Instruction::Or)) {
1198 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1199 // If the compares in later blocks need to use values not currently
1200 // exported from this block, export them now. This block should always
1201 // be the first entry.
1202 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1204 // Allow some cases to be rejected.
1205 if (ShouldEmitAsBranches(SwitchCases)) {
1206 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1207 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1208 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1211 // Emit the branch for this block.
1212 visitSwitchCase(SwitchCases[0]);
1213 SwitchCases.erase(SwitchCases.begin());
1217 // Okay, we decided not to do this, remove any inserted MBB's and clear
1219 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1220 CurMBB->getParent()->getBasicBlockList().erase(SwitchCases[i].ThisBB);
1222 SwitchCases.clear();
1226 // Create a CaseBlock record representing this branch.
1227 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(),
1228 NULL, Succ0MBB, Succ1MBB, CurMBB);
1229 // Use visitSwitchCase to actually insert the fast branch sequence for this
1231 visitSwitchCase(CB);
1234 /// visitSwitchCase - Emits the necessary code to represent a single node in
1235 /// the binary search tree resulting from lowering a switch instruction.
1236 void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) {
1238 SDOperand CondLHS = getValue(CB.CmpLHS);
1240 // Build the setcc now.
1241 if (CB.CmpMHS == NULL) {
1242 // Fold "(X == true)" to X and "(X == false)" to !X to
1243 // handle common cases produced by branch lowering.
1244 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ)
1246 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) {
1247 SDOperand True = DAG.getConstant(1, CondLHS.getValueType());
1248 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True);
1250 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1252 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1254 uint64_t Low = cast<ConstantInt>(CB.CmpLHS)->getSExtValue();
1255 uint64_t High = cast<ConstantInt>(CB.CmpRHS)->getSExtValue();
1257 SDOperand CmpOp = getValue(CB.CmpMHS);
1258 MVT::ValueType VT = CmpOp.getValueType();
1260 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1261 Cond = DAG.getSetCC(MVT::i1, CmpOp, DAG.getConstant(High, VT), ISD::SETLE);
1263 SDOperand SUB = DAG.getNode(ISD::SUB, VT, CmpOp, DAG.getConstant(Low, VT));
1264 Cond = DAG.getSetCC(MVT::i1, SUB,
1265 DAG.getConstant(High-Low, VT), ISD::SETULE);
1270 // Set NextBlock to be the MBB immediately after the current one, if any.
1271 // This is used to avoid emitting unnecessary branches to the next block.
1272 MachineBasicBlock *NextBlock = 0;
1273 MachineFunction::iterator BBI = CurMBB;
1274 if (++BBI != CurMBB->getParent()->end())
1277 // If the lhs block is the next block, invert the condition so that we can
1278 // fall through to the lhs instead of the rhs block.
1279 if (CB.TrueBB == NextBlock) {
1280 std::swap(CB.TrueBB, CB.FalseBB);
1281 SDOperand True = DAG.getConstant(1, Cond.getValueType());
1282 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
1284 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
1285 DAG.getBasicBlock(CB.TrueBB));
1286 if (CB.FalseBB == NextBlock)
1287 DAG.setRoot(BrCond);
1289 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1290 DAG.getBasicBlock(CB.FalseBB)));
1291 // Update successor info
1292 CurMBB->addSuccessor(CB.TrueBB);
1293 CurMBB->addSuccessor(CB.FalseBB);
1296 /// visitJumpTable - Emit JumpTable node in the current MBB
1297 void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) {
1298 // Emit the code for the jump table
1299 assert(JT.Reg != -1U && "Should lower JT Header first!");
1300 MVT::ValueType PTy = TLI.getPointerTy();
1301 SDOperand Index = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy);
1302 SDOperand Table = DAG.getJumpTable(JT.JTI, PTy);
1303 DAG.setRoot(DAG.getNode(ISD::BR_JT, MVT::Other, Index.getValue(1),
1308 /// visitJumpTableHeader - This function emits necessary code to produce index
1309 /// in the JumpTable from switch case.
1310 void SelectionDAGLowering::visitJumpTableHeader(SelectionDAGISel::JumpTable &JT,
1311 SelectionDAGISel::JumpTableHeader &JTH) {
1312 // Subtract the lowest switch case value from the value being switched on
1313 // and conditional branch to default mbb if the result is greater than the
1314 // difference between smallest and largest cases.
1315 SDOperand SwitchOp = getValue(JTH.SValue);
1316 MVT::ValueType VT = SwitchOp.getValueType();
1317 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1318 DAG.getConstant(JTH.First, VT));
1320 // The SDNode we just created, which holds the value being switched on
1321 // minus the the smallest case value, needs to be copied to a virtual
1322 // register so it can be used as an index into the jump table in a
1323 // subsequent basic block. This value may be smaller or larger than the
1324 // target's pointer type, and therefore require extension or truncating.
1325 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getPointerTy()))
1326 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
1328 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
1330 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1331 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp);
1332 JT.Reg = JumpTableReg;
1334 // Emit the range check for the jump table, and branch to the default
1335 // block for the switch statement if the value being switched on exceeds
1336 // the largest case in the switch.
1337 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1338 DAG.getConstant(JTH.Last-JTH.First,VT),
1341 // Set NextBlock to be the MBB immediately after the current one, if any.
1342 // This is used to avoid emitting unnecessary branches to the next block.
1343 MachineBasicBlock *NextBlock = 0;
1344 MachineFunction::iterator BBI = CurMBB;
1345 if (++BBI != CurMBB->getParent()->end())
1348 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
1349 DAG.getBasicBlock(JT.Default));
1351 if (JT.MBB == NextBlock)
1352 DAG.setRoot(BrCond);
1354 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1355 DAG.getBasicBlock(JT.MBB)));
1360 /// visitBitTestHeader - This function emits necessary code to produce value
1361 /// suitable for "bit tests"
1362 void SelectionDAGLowering::visitBitTestHeader(SelectionDAGISel::BitTestBlock &B) {
1363 // Subtract the minimum value
1364 SDOperand SwitchOp = getValue(B.SValue);
1365 MVT::ValueType VT = SwitchOp.getValueType();
1366 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1367 DAG.getConstant(B.First, VT));
1370 SDOperand RangeCmp = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1371 DAG.getConstant(B.Range, VT),
1375 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getShiftAmountTy()))
1376 ShiftOp = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), SUB);
1378 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getShiftAmountTy(), SUB);
1380 // Make desired shift
1381 SDOperand SwitchVal = DAG.getNode(ISD::SHL, TLI.getPointerTy(),
1382 DAG.getConstant(1, TLI.getPointerTy()),
1385 unsigned SwitchReg = FuncInfo.MakeReg(TLI.getPointerTy());
1386 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), SwitchReg, SwitchVal);
1389 SDOperand BrRange = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, RangeCmp,
1390 DAG.getBasicBlock(B.Default));
1392 // Set NextBlock to be the MBB immediately after the current one, if any.
1393 // This is used to avoid emitting unnecessary branches to the next block.
1394 MachineBasicBlock *NextBlock = 0;
1395 MachineFunction::iterator BBI = CurMBB;
1396 if (++BBI != CurMBB->getParent()->end())
1399 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1400 if (MBB == NextBlock)
1401 DAG.setRoot(BrRange);
1403 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, CopyTo,
1404 DAG.getBasicBlock(MBB)));
1406 CurMBB->addSuccessor(B.Default);
1407 CurMBB->addSuccessor(MBB);
1412 /// visitBitTestCase - this function produces one "bit test"
1413 void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB,
1415 SelectionDAGISel::BitTestCase &B) {
1416 // Emit bit tests and jumps
1417 SDOperand SwitchVal = DAG.getCopyFromReg(getRoot(), Reg, TLI.getPointerTy());
1419 SDOperand AndOp = DAG.getNode(ISD::AND, TLI.getPointerTy(),
1421 DAG.getConstant(B.Mask,
1422 TLI.getPointerTy()));
1423 SDOperand AndCmp = DAG.getSetCC(TLI.getSetCCResultTy(), AndOp,
1424 DAG.getConstant(0, TLI.getPointerTy()),
1426 SDOperand BrAnd = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(),
1427 AndCmp, DAG.getBasicBlock(B.TargetBB));
1429 // Set NextBlock to be the MBB immediately after the current one, if any.
1430 // This is used to avoid emitting unnecessary branches to the next block.
1431 MachineBasicBlock *NextBlock = 0;
1432 MachineFunction::iterator BBI = CurMBB;
1433 if (++BBI != CurMBB->getParent()->end())
1436 if (NextMBB == NextBlock)
1439 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrAnd,
1440 DAG.getBasicBlock(NextMBB)));
1442 CurMBB->addSuccessor(B.TargetBB);
1443 CurMBB->addSuccessor(NextMBB);
1448 void SelectionDAGLowering::visitInvoke(InvokeInst &I) {
1449 // Retrieve successors.
1450 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1451 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1453 LowerCallTo(I, I.getCalledValue()->getType(),
1456 getValue(I.getOperand(0)),
1459 // If the value of the invoke is used outside of its defining block, make it
1460 // available as a virtual register.
1461 if (!I.use_empty()) {
1462 DenseMap<const Value*, unsigned>::iterator VMI = FuncInfo.ValueMap.find(&I);
1463 if (VMI != FuncInfo.ValueMap.end())
1464 DAG.setRoot(CopyValueToVirtualRegister(&I, VMI->second));
1467 // Drop into normal successor.
1468 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1469 DAG.getBasicBlock(Return)));
1471 // Update successor info
1472 CurMBB->addSuccessor(Return);
1473 CurMBB->addSuccessor(LandingPad);
1476 void SelectionDAGLowering::visitUnwind(UnwindInst &I) {
1479 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1480 /// small case ranges).
1481 bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR,
1482 CaseRecVector& WorkList,
1484 MachineBasicBlock* Default) {
1485 Case& BackCase = *(CR.Range.second-1);
1487 // Size is the number of Cases represented by this range.
1488 unsigned Size = CR.Range.second - CR.Range.first;
1492 // Get the MachineFunction which holds the current MBB. This is used when
1493 // inserting any additional MBBs necessary to represent the switch.
1494 MachineFunction *CurMF = CurMBB->getParent();
1496 // Figure out which block is immediately after the current one.
1497 MachineBasicBlock *NextBlock = 0;
1498 MachineFunction::iterator BBI = CR.CaseBB;
1500 if (++BBI != CurMBB->getParent()->end())
1503 // TODO: If any two of the cases has the same destination, and if one value
1504 // is the same as the other, but has one bit unset that the other has set,
1505 // use bit manipulation to do two compares at once. For example:
1506 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1508 // Rearrange the case blocks so that the last one falls through if possible.
1509 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1510 // The last case block won't fall through into 'NextBlock' if we emit the
1511 // branches in this order. See if rearranging a case value would help.
1512 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1513 if (I->BB == NextBlock) {
1514 std::swap(*I, BackCase);
1520 // Create a CaseBlock record representing a conditional branch to
1521 // the Case's target mbb if the value being switched on SV is equal
1523 MachineBasicBlock *CurBlock = CR.CaseBB;
1524 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1525 MachineBasicBlock *FallThrough;
1527 FallThrough = new MachineBasicBlock(CurBlock->getBasicBlock());
1528 CurMF->getBasicBlockList().insert(BBI, FallThrough);
1530 // If the last case doesn't match, go to the default block.
1531 FallThrough = Default;
1534 Value *RHS, *LHS, *MHS;
1536 if (I->High == I->Low) {
1537 // This is just small small case range :) containing exactly 1 case
1539 LHS = SV; RHS = I->High; MHS = NULL;
1542 LHS = I->Low; MHS = SV; RHS = I->High;
1544 SelectionDAGISel::CaseBlock CB(CC, LHS, RHS, MHS,
1545 I->BB, FallThrough, CurBlock);
1547 // If emitting the first comparison, just call visitSwitchCase to emit the
1548 // code into the current block. Otherwise, push the CaseBlock onto the
1549 // vector to be later processed by SDISel, and insert the node's MBB
1550 // before the next MBB.
1551 if (CurBlock == CurMBB)
1552 visitSwitchCase(CB);
1554 SwitchCases.push_back(CB);
1556 CurBlock = FallThrough;
1562 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1563 return (TLI.isOperationLegal(ISD::BR_JT, MVT::Other) ||
1564 TLI.isOperationLegal(ISD::BRIND, MVT::Other));
1567 /// handleJTSwitchCase - Emit jumptable for current switch case range
1568 bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR,
1569 CaseRecVector& WorkList,
1571 MachineBasicBlock* Default) {
1572 Case& FrontCase = *CR.Range.first;
1573 Case& BackCase = *(CR.Range.second-1);
1575 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1576 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1579 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1583 if (!areJTsAllowed(TLI) || TSize <= 3)
1586 double Density = (double)TSize / (double)((Last - First) + 1ULL);
1590 DOUT << "Lowering jump table\n"
1591 << "First entry: " << First << ". Last entry: " << Last << "\n"
1592 << "Size: " << TSize << ". Density: " << Density << "\n\n";
1594 // Get the MachineFunction which holds the current MBB. This is used when
1595 // inserting any additional MBBs necessary to represent the switch.
1596 MachineFunction *CurMF = CurMBB->getParent();
1598 // Figure out which block is immediately after the current one.
1599 MachineBasicBlock *NextBlock = 0;
1600 MachineFunction::iterator BBI = CR.CaseBB;
1602 if (++BBI != CurMBB->getParent()->end())
1605 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1607 // Create a new basic block to hold the code for loading the address
1608 // of the jump table, and jumping to it. Update successor information;
1609 // we will either branch to the default case for the switch, or the jump
1611 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB);
1612 CurMF->getBasicBlockList().insert(BBI, JumpTableBB);
1613 CR.CaseBB->addSuccessor(Default);
1614 CR.CaseBB->addSuccessor(JumpTableBB);
1616 // Build a vector of destination BBs, corresponding to each target
1617 // of the jump table. If the value of the jump table slot corresponds to
1618 // a case statement, push the case's BB onto the vector, otherwise, push
1620 std::vector<MachineBasicBlock*> DestBBs;
1621 int64_t TEI = First;
1622 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1623 int64_t Low = cast<ConstantInt>(I->Low)->getSExtValue();
1624 int64_t High = cast<ConstantInt>(I->High)->getSExtValue();
1626 if ((Low <= TEI) && (TEI <= High)) {
1627 DestBBs.push_back(I->BB);
1631 DestBBs.push_back(Default);
1635 // Update successor info. Add one edge to each unique successor.
1636 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1637 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1638 E = DestBBs.end(); I != E; ++I) {
1639 if (!SuccsHandled[(*I)->getNumber()]) {
1640 SuccsHandled[(*I)->getNumber()] = true;
1641 JumpTableBB->addSuccessor(*I);
1645 // Create a jump table index for this jump table, or return an existing
1647 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1649 // Set the jump table information so that we can codegen it as a second
1650 // MachineBasicBlock
1651 SelectionDAGISel::JumpTable JT(-1U, JTI, JumpTableBB, Default);
1652 SelectionDAGISel::JumpTableHeader JTH(First, Last, SV, CR.CaseBB,
1653 (CR.CaseBB == CurMBB));
1654 if (CR.CaseBB == CurMBB)
1655 visitJumpTableHeader(JT, JTH);
1657 JTCases.push_back(SelectionDAGISel::JumpTableBlock(JTH, JT));
1662 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1664 bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR,
1665 CaseRecVector& WorkList,
1667 MachineBasicBlock* Default) {
1668 // Get the MachineFunction which holds the current MBB. This is used when
1669 // inserting any additional MBBs necessary to represent the switch.
1670 MachineFunction *CurMF = CurMBB->getParent();
1672 // Figure out which block is immediately after the current one.
1673 MachineBasicBlock *NextBlock = 0;
1674 MachineFunction::iterator BBI = CR.CaseBB;
1676 if (++BBI != CurMBB->getParent()->end())
1679 Case& FrontCase = *CR.Range.first;
1680 Case& BackCase = *(CR.Range.second-1);
1681 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1683 // Size is the number of Cases represented by this range.
1684 unsigned Size = CR.Range.second - CR.Range.first;
1686 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1687 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1689 CaseItr Pivot = CR.Range.first + Size/2;
1691 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1692 // (heuristically) allow us to emit JumpTable's later.
1694 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1698 uint64_t LSize = FrontCase.size();
1699 uint64_t RSize = TSize-LSize;
1700 DOUT << "Selecting best pivot: \n"
1701 << "First: " << First << ", Last: " << Last <<"\n"
1702 << "LSize: " << LSize << ", RSize: " << RSize << "\n";
1703 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1705 int64_t LEnd = cast<ConstantInt>(I->High)->getSExtValue();
1706 int64_t RBegin = cast<ConstantInt>(J->Low)->getSExtValue();
1707 assert((RBegin-LEnd>=1) && "Invalid case distance");
1708 double LDensity = (double)LSize / (double)((LEnd - First) + 1ULL);
1709 double RDensity = (double)RSize / (double)((Last - RBegin) + 1ULL);
1710 double Metric = Log2_64(RBegin-LEnd)*(LDensity+RDensity);
1711 // Should always split in some non-trivial place
1713 << "LEnd: " << LEnd << ", RBegin: " << RBegin << "\n"
1714 << "LDensity: " << LDensity << ", RDensity: " << RDensity << "\n"
1715 << "Metric: " << Metric << "\n";
1716 if (FMetric < Metric) {
1719 DOUT << "Current metric set to: " << FMetric << "\n";
1725 if (areJTsAllowed(TLI)) {
1726 // If our case is dense we *really* should handle it earlier!
1727 assert((FMetric > 0) && "Should handle dense range earlier!");
1729 Pivot = CR.Range.first + Size/2;
1732 CaseRange LHSR(CR.Range.first, Pivot);
1733 CaseRange RHSR(Pivot, CR.Range.second);
1734 Constant *C = Pivot->Low;
1735 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1737 // We know that we branch to the LHS if the Value being switched on is
1738 // less than the Pivot value, C. We use this to optimize our binary
1739 // tree a bit, by recognizing that if SV is greater than or equal to the
1740 // LHS's Case Value, and that Case Value is exactly one less than the
1741 // Pivot's Value, then we can branch directly to the LHS's Target,
1742 // rather than creating a leaf node for it.
1743 if ((LHSR.second - LHSR.first) == 1 &&
1744 LHSR.first->High == CR.GE &&
1745 cast<ConstantInt>(C)->getSExtValue() ==
1746 (cast<ConstantInt>(CR.GE)->getSExtValue() + 1LL)) {
1747 TrueBB = LHSR.first->BB;
1749 TrueBB = new MachineBasicBlock(LLVMBB);
1750 CurMF->getBasicBlockList().insert(BBI, TrueBB);
1751 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1754 // Similar to the optimization above, if the Value being switched on is
1755 // known to be less than the Constant CR.LT, and the current Case Value
1756 // is CR.LT - 1, then we can branch directly to the target block for
1757 // the current Case Value, rather than emitting a RHS leaf node for it.
1758 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1759 cast<ConstantInt>(RHSR.first->Low)->getSExtValue() ==
1760 (cast<ConstantInt>(CR.LT)->getSExtValue() - 1LL)) {
1761 FalseBB = RHSR.first->BB;
1763 FalseBB = new MachineBasicBlock(LLVMBB);
1764 CurMF->getBasicBlockList().insert(BBI, FalseBB);
1765 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1768 // Create a CaseBlock record representing a conditional branch to
1769 // the LHS node if the value being switched on SV is less than C.
1770 // Otherwise, branch to LHS.
1771 SelectionDAGISel::CaseBlock CB(ISD::SETLT, SV, C, NULL,
1772 TrueBB, FalseBB, CR.CaseBB);
1774 if (CR.CaseBB == CurMBB)
1775 visitSwitchCase(CB);
1777 SwitchCases.push_back(CB);
1782 /// handleBitTestsSwitchCase - if current case range has few destination and
1783 /// range span less, than machine word bitwidth, encode case range into series
1784 /// of masks and emit bit tests with these masks.
1785 bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR,
1786 CaseRecVector& WorkList,
1788 MachineBasicBlock* Default){
1789 unsigned IntPtrBits = MVT::getSizeInBits(TLI.getPointerTy());
1791 Case& FrontCase = *CR.Range.first;
1792 Case& BackCase = *(CR.Range.second-1);
1794 // Get the MachineFunction which holds the current MBB. This is used when
1795 // inserting any additional MBBs necessary to represent the switch.
1796 MachineFunction *CurMF = CurMBB->getParent();
1798 unsigned numCmps = 0;
1799 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1801 // Single case counts one, case range - two.
1802 if (I->Low == I->High)
1808 // Count unique destinations
1809 SmallSet<MachineBasicBlock*, 4> Dests;
1810 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1811 Dests.insert(I->BB);
1812 if (Dests.size() > 3)
1813 // Don't bother the code below, if there are too much unique destinations
1816 DOUT << "Total number of unique destinations: " << Dests.size() << "\n"
1817 << "Total number of comparisons: " << numCmps << "\n";
1819 // Compute span of values.
1820 Constant* minValue = FrontCase.Low;
1821 Constant* maxValue = BackCase.High;
1822 uint64_t range = cast<ConstantInt>(maxValue)->getSExtValue() -
1823 cast<ConstantInt>(minValue)->getSExtValue();
1824 DOUT << "Compare range: " << range << "\n"
1825 << "Low bound: " << cast<ConstantInt>(minValue)->getSExtValue() << "\n"
1826 << "High bound: " << cast<ConstantInt>(maxValue)->getSExtValue() << "\n";
1828 if (range>=IntPtrBits ||
1829 (!(Dests.size() == 1 && numCmps >= 3) &&
1830 !(Dests.size() == 2 && numCmps >= 5) &&
1831 !(Dests.size() >= 3 && numCmps >= 6)))
1834 DOUT << "Emitting bit tests\n";
1835 int64_t lowBound = 0;
1837 // Optimize the case where all the case values fit in a
1838 // word without having to subtract minValue. In this case,
1839 // we can optimize away the subtraction.
1840 if (cast<ConstantInt>(minValue)->getSExtValue() >= 0 &&
1841 cast<ConstantInt>(maxValue)->getSExtValue() < IntPtrBits) {
1842 range = cast<ConstantInt>(maxValue)->getSExtValue();
1844 lowBound = cast<ConstantInt>(minValue)->getSExtValue();
1847 CaseBitsVector CasesBits;
1848 unsigned i, count = 0;
1850 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1851 MachineBasicBlock* Dest = I->BB;
1852 for (i = 0; i < count; ++i)
1853 if (Dest == CasesBits[i].BB)
1857 assert((count < 3) && "Too much destinations to test!");
1858 CasesBits.push_back(CaseBits(0, Dest, 0));
1862 uint64_t lo = cast<ConstantInt>(I->Low)->getSExtValue() - lowBound;
1863 uint64_t hi = cast<ConstantInt>(I->High)->getSExtValue() - lowBound;
1865 for (uint64_t j = lo; j <= hi; j++) {
1866 CasesBits[i].Mask |= 1ULL << j;
1867 CasesBits[i].Bits++;
1871 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
1873 SelectionDAGISel::BitTestInfo BTC;
1875 // Figure out which block is immediately after the current one.
1876 MachineFunction::iterator BBI = CR.CaseBB;
1879 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1882 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
1883 DOUT << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits
1884 << ", BB: " << CasesBits[i].BB << "\n";
1886 MachineBasicBlock *CaseBB = new MachineBasicBlock(LLVMBB);
1887 CurMF->getBasicBlockList().insert(BBI, CaseBB);
1888 BTC.push_back(SelectionDAGISel::BitTestCase(CasesBits[i].Mask,
1893 SelectionDAGISel::BitTestBlock BTB(lowBound, range, SV,
1894 -1U, (CR.CaseBB == CurMBB),
1895 CR.CaseBB, Default, BTC);
1897 if (CR.CaseBB == CurMBB)
1898 visitBitTestHeader(BTB);
1900 BitTestCases.push_back(BTB);
1906 // Clusterify - Transform simple list of Cases into list of CaseRange's
1907 unsigned SelectionDAGLowering::Clusterify(CaseVector& Cases,
1908 const SwitchInst& SI) {
1909 unsigned numCmps = 0;
1911 // Start with "simple" cases
1912 for (unsigned i = 1; i < SI.getNumSuccessors(); ++i) {
1913 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
1914 Cases.push_back(Case(SI.getSuccessorValue(i),
1915 SI.getSuccessorValue(i),
1918 sort(Cases.begin(), Cases.end(), CaseCmp());
1920 // Merge case into clusters
1921 if (Cases.size()>=2)
1922 // Must recompute end() each iteration because it may be
1923 // invalidated by erase if we hold on to it
1924 for (CaseItr I=Cases.begin(), J=++(Cases.begin()); J!=Cases.end(); ) {
1925 int64_t nextValue = cast<ConstantInt>(J->Low)->getSExtValue();
1926 int64_t currentValue = cast<ConstantInt>(I->High)->getSExtValue();
1927 MachineBasicBlock* nextBB = J->BB;
1928 MachineBasicBlock* currentBB = I->BB;
1930 // If the two neighboring cases go to the same destination, merge them
1931 // into a single case.
1932 if ((nextValue-currentValue==1) && (currentBB == nextBB)) {
1940 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
1941 if (I->Low != I->High)
1942 // A range counts double, since it requires two compares.
1949 void SelectionDAGLowering::visitSwitch(SwitchInst &SI) {
1950 // Figure out which block is immediately after the current one.
1951 MachineBasicBlock *NextBlock = 0;
1952 MachineFunction::iterator BBI = CurMBB;
1954 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
1956 // If there is only the default destination, branch to it if it is not the
1957 // next basic block. Otherwise, just fall through.
1958 if (SI.getNumOperands() == 2) {
1959 // Update machine-CFG edges.
1961 // If this is not a fall-through branch, emit the branch.
1962 if (Default != NextBlock)
1963 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1964 DAG.getBasicBlock(Default)));
1966 CurMBB->addSuccessor(Default);
1970 // If there are any non-default case statements, create a vector of Cases
1971 // representing each one, and sort the vector so that we can efficiently
1972 // create a binary search tree from them.
1974 unsigned numCmps = Clusterify(Cases, SI);
1975 DOUT << "Clusterify finished. Total clusters: " << Cases.size()
1976 << ". Total compares: " << numCmps << "\n";
1978 // Get the Value to be switched on and default basic blocks, which will be
1979 // inserted into CaseBlock records, representing basic blocks in the binary
1981 Value *SV = SI.getOperand(0);
1983 // Push the initial CaseRec onto the worklist
1984 CaseRecVector WorkList;
1985 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
1987 while (!WorkList.empty()) {
1988 // Grab a record representing a case range to process off the worklist
1989 CaseRec CR = WorkList.back();
1990 WorkList.pop_back();
1992 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
1995 // If the range has few cases (two or less) emit a series of specific
1997 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2000 // If the switch has more than 5 blocks, and at least 40% dense, and the
2001 // target supports indirect branches, then emit a jump table rather than
2002 // lowering the switch to a binary tree of conditional branches.
2003 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2006 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2007 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2008 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2013 void SelectionDAGLowering::visitSub(User &I) {
2014 // -0.0 - X --> fneg
2015 const Type *Ty = I.getType();
2016 if (isa<VectorType>(Ty)) {
2017 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2018 const VectorType *DestTy = cast<VectorType>(I.getType());
2019 const Type *ElTy = DestTy->getElementType();
2020 if (ElTy->isFloatingPoint()) {
2021 unsigned VL = DestTy->getNumElements();
2022 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2023 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2025 SDOperand Op2 = getValue(I.getOperand(1));
2026 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2032 if (Ty->isFloatingPoint()) {
2033 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2034 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2035 SDOperand Op2 = getValue(I.getOperand(1));
2036 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2041 visitBinary(I, Ty->isFPOrFPVector() ? ISD::FSUB : ISD::SUB);
2044 void SelectionDAGLowering::visitBinary(User &I, unsigned OpCode) {
2045 SDOperand Op1 = getValue(I.getOperand(0));
2046 SDOperand Op2 = getValue(I.getOperand(1));
2048 setValue(&I, DAG.getNode(OpCode, Op1.getValueType(), Op1, Op2));
2051 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
2052 SDOperand Op1 = getValue(I.getOperand(0));
2053 SDOperand Op2 = getValue(I.getOperand(1));
2055 if (MVT::getSizeInBits(TLI.getShiftAmountTy()) <
2056 MVT::getSizeInBits(Op2.getValueType()))
2057 Op2 = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), Op2);
2058 else if (TLI.getShiftAmountTy() > Op2.getValueType())
2059 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
2061 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
2064 void SelectionDAGLowering::visitICmp(User &I) {
2065 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2066 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2067 predicate = IC->getPredicate();
2068 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2069 predicate = ICmpInst::Predicate(IC->getPredicate());
2070 SDOperand Op1 = getValue(I.getOperand(0));
2071 SDOperand Op2 = getValue(I.getOperand(1));
2072 ISD::CondCode Opcode;
2073 switch (predicate) {
2074 case ICmpInst::ICMP_EQ : Opcode = ISD::SETEQ; break;
2075 case ICmpInst::ICMP_NE : Opcode = ISD::SETNE; break;
2076 case ICmpInst::ICMP_UGT : Opcode = ISD::SETUGT; break;
2077 case ICmpInst::ICMP_UGE : Opcode = ISD::SETUGE; break;
2078 case ICmpInst::ICMP_ULT : Opcode = ISD::SETULT; break;
2079 case ICmpInst::ICMP_ULE : Opcode = ISD::SETULE; break;
2080 case ICmpInst::ICMP_SGT : Opcode = ISD::SETGT; break;
2081 case ICmpInst::ICMP_SGE : Opcode = ISD::SETGE; break;
2082 case ICmpInst::ICMP_SLT : Opcode = ISD::SETLT; break;
2083 case ICmpInst::ICMP_SLE : Opcode = ISD::SETLE; break;
2085 assert(!"Invalid ICmp predicate value");
2086 Opcode = ISD::SETEQ;
2089 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
2092 void SelectionDAGLowering::visitFCmp(User &I) {
2093 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2094 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2095 predicate = FC->getPredicate();
2096 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2097 predicate = FCmpInst::Predicate(FC->getPredicate());
2098 SDOperand Op1 = getValue(I.getOperand(0));
2099 SDOperand Op2 = getValue(I.getOperand(1));
2100 ISD::CondCode Condition, FOC, FPC;
2101 switch (predicate) {
2102 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
2103 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
2104 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
2105 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
2106 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
2107 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
2108 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
2109 case FCmpInst::FCMP_ORD: FOC = ISD::SETEQ; FPC = ISD::SETO; break;
2110 case FCmpInst::FCMP_UNO: FOC = ISD::SETNE; FPC = ISD::SETUO; break;
2111 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
2112 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
2113 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
2114 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
2115 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
2116 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
2117 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
2119 assert(!"Invalid FCmp predicate value");
2120 FOC = FPC = ISD::SETFALSE;
2123 if (FiniteOnlyFPMath())
2127 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Condition));
2130 void SelectionDAGLowering::visitSelect(User &I) {
2131 SDOperand Cond = getValue(I.getOperand(0));
2132 SDOperand TrueVal = getValue(I.getOperand(1));
2133 SDOperand FalseVal = getValue(I.getOperand(2));
2134 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
2135 TrueVal, FalseVal));
2139 void SelectionDAGLowering::visitTrunc(User &I) {
2140 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2141 SDOperand N = getValue(I.getOperand(0));
2142 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2143 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2146 void SelectionDAGLowering::visitZExt(User &I) {
2147 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2148 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2149 SDOperand N = getValue(I.getOperand(0));
2150 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2151 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2154 void SelectionDAGLowering::visitSExt(User &I) {
2155 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2156 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2157 SDOperand N = getValue(I.getOperand(0));
2158 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2159 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
2162 void SelectionDAGLowering::visitFPTrunc(User &I) {
2163 // FPTrunc is never a no-op cast, no need to check
2164 SDOperand N = getValue(I.getOperand(0));
2165 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2166 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N));
2169 void SelectionDAGLowering::visitFPExt(User &I){
2170 // FPTrunc is never a no-op cast, no need to check
2171 SDOperand N = getValue(I.getOperand(0));
2172 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2173 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
2176 void SelectionDAGLowering::visitFPToUI(User &I) {
2177 // FPToUI is never a no-op cast, no need to check
2178 SDOperand N = getValue(I.getOperand(0));
2179 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2180 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
2183 void SelectionDAGLowering::visitFPToSI(User &I) {
2184 // FPToSI is never a no-op cast, no need to check
2185 SDOperand N = getValue(I.getOperand(0));
2186 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2187 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
2190 void SelectionDAGLowering::visitUIToFP(User &I) {
2191 // UIToFP is never a no-op cast, no need to check
2192 SDOperand N = getValue(I.getOperand(0));
2193 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2194 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
2197 void SelectionDAGLowering::visitSIToFP(User &I){
2198 // UIToFP is never a no-op cast, no need to check
2199 SDOperand N = getValue(I.getOperand(0));
2200 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2201 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
2204 void SelectionDAGLowering::visitPtrToInt(User &I) {
2205 // What to do depends on the size of the integer and the size of the pointer.
2206 // We can either truncate, zero extend, or no-op, accordingly.
2207 SDOperand N = getValue(I.getOperand(0));
2208 MVT::ValueType SrcVT = N.getValueType();
2209 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2211 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT))
2212 Result = DAG.getNode(ISD::TRUNCATE, DestVT, N);
2214 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2215 Result = DAG.getNode(ISD::ZERO_EXTEND, DestVT, N);
2216 setValue(&I, Result);
2219 void SelectionDAGLowering::visitIntToPtr(User &I) {
2220 // What to do depends on the size of the integer and the size of the pointer.
2221 // We can either truncate, zero extend, or no-op, accordingly.
2222 SDOperand N = getValue(I.getOperand(0));
2223 MVT::ValueType SrcVT = N.getValueType();
2224 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2225 if (MVT::getSizeInBits(DestVT) < MVT::getSizeInBits(SrcVT))
2226 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2228 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2229 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2232 void SelectionDAGLowering::visitBitCast(User &I) {
2233 SDOperand N = getValue(I.getOperand(0));
2234 MVT::ValueType DestVT = TLI.getValueType(I.getType());
2236 // BitCast assures us that source and destination are the same size so this
2237 // is either a BIT_CONVERT or a no-op.
2238 if (DestVT != N.getValueType())
2239 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, DestVT, N)); // convert types
2241 setValue(&I, N); // noop cast.
2244 void SelectionDAGLowering::visitInsertElement(User &I) {
2245 SDOperand InVec = getValue(I.getOperand(0));
2246 SDOperand InVal = getValue(I.getOperand(1));
2247 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2248 getValue(I.getOperand(2)));
2250 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT,
2251 TLI.getValueType(I.getType()),
2252 InVec, InVal, InIdx));
2255 void SelectionDAGLowering::visitExtractElement(User &I) {
2256 SDOperand InVec = getValue(I.getOperand(0));
2257 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2258 getValue(I.getOperand(1)));
2259 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
2260 TLI.getValueType(I.getType()), InVec, InIdx));
2263 void SelectionDAGLowering::visitShuffleVector(User &I) {
2264 SDOperand V1 = getValue(I.getOperand(0));
2265 SDOperand V2 = getValue(I.getOperand(1));
2266 SDOperand Mask = getValue(I.getOperand(2));
2268 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE,
2269 TLI.getValueType(I.getType()),
2274 void SelectionDAGLowering::visitGetElementPtr(User &I) {
2275 SDOperand N = getValue(I.getOperand(0));
2276 const Type *Ty = I.getOperand(0)->getType();
2278 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2281 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2282 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2285 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2286 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
2287 getIntPtrConstant(Offset));
2289 Ty = StTy->getElementType(Field);
2291 Ty = cast<SequentialType>(Ty)->getElementType();
2293 // If this is a constant subscript, handle it quickly.
2294 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2295 if (CI->getZExtValue() == 0) continue;
2297 TD->getABITypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2298 N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs));
2302 // N = N + Idx * ElementSize;
2303 uint64_t ElementSize = TD->getABITypeSize(Ty);
2304 SDOperand IdxN = getValue(Idx);
2306 // If the index is smaller or larger than intptr_t, truncate or extend
2308 if (IdxN.getValueType() < N.getValueType()) {
2309 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
2310 } else if (IdxN.getValueType() > N.getValueType())
2311 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
2313 // If this is a multiply by a power of two, turn it into a shl
2314 // immediately. This is a very common case.
2315 if (isPowerOf2_64(ElementSize)) {
2316 unsigned Amt = Log2_64(ElementSize);
2317 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
2318 DAG.getConstant(Amt, TLI.getShiftAmountTy()));
2319 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
2323 SDOperand Scale = getIntPtrConstant(ElementSize);
2324 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
2325 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
2331 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
2332 // If this is a fixed sized alloca in the entry block of the function,
2333 // allocate it statically on the stack.
2334 if (FuncInfo.StaticAllocaMap.count(&I))
2335 return; // getValue will auto-populate this.
2337 const Type *Ty = I.getAllocatedType();
2338 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
2340 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2343 SDOperand AllocSize = getValue(I.getArraySize());
2344 MVT::ValueType IntPtr = TLI.getPointerTy();
2345 if (IntPtr < AllocSize.getValueType())
2346 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
2347 else if (IntPtr > AllocSize.getValueType())
2348 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
2350 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
2351 getIntPtrConstant(TySize));
2353 // Handle alignment. If the requested alignment is less than or equal to
2354 // the stack alignment, ignore it. If the size is greater than or equal to
2355 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2356 unsigned StackAlign =
2357 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2358 if (Align <= StackAlign)
2361 // Round the size of the allocation up to the stack alignment size
2362 // by add SA-1 to the size.
2363 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
2364 getIntPtrConstant(StackAlign-1));
2365 // Mask out the low bits for alignment purposes.
2366 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
2367 getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2369 SDOperand Ops[] = { getRoot(), AllocSize, getIntPtrConstant(Align) };
2370 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
2372 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3);
2374 DAG.setRoot(DSA.getValue(1));
2376 // Inform the Frame Information that we have just allocated a variable-sized
2378 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
2381 void SelectionDAGLowering::visitLoad(LoadInst &I) {
2382 SDOperand Ptr = getValue(I.getOperand(0));
2388 // Do not serialize non-volatile loads against each other.
2389 Root = DAG.getRoot();
2392 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0),
2393 Root, I.isVolatile(), I.getAlignment()));
2396 SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr,
2397 const Value *SV, SDOperand Root,
2399 unsigned Alignment) {
2401 DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, 0,
2402 isVolatile, Alignment);
2405 DAG.setRoot(L.getValue(1));
2407 PendingLoads.push_back(L.getValue(1));
2413 void SelectionDAGLowering::visitStore(StoreInst &I) {
2414 Value *SrcV = I.getOperand(0);
2415 SDOperand Src = getValue(SrcV);
2416 SDOperand Ptr = getValue(I.getOperand(1));
2417 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1), 0,
2418 I.isVolatile(), I.getAlignment()));
2421 /// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot
2422 /// access memory and has no other side effects at all.
2423 static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) {
2424 #define GET_NO_MEMORY_INTRINSICS
2425 #include "llvm/Intrinsics.gen"
2426 #undef GET_NO_MEMORY_INTRINSICS
2430 // IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't
2431 // have any side-effects or if it only reads memory.
2432 static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) {
2433 #define GET_SIDE_EFFECT_INFO
2434 #include "llvm/Intrinsics.gen"
2435 #undef GET_SIDE_EFFECT_INFO
2439 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2441 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
2442 unsigned Intrinsic) {
2443 bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic);
2444 bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic);
2446 // Build the operand list.
2447 SmallVector<SDOperand, 8> Ops;
2448 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2450 // We don't need to serialize loads against other loads.
2451 Ops.push_back(DAG.getRoot());
2453 Ops.push_back(getRoot());
2457 // Add the intrinsic ID as an integer operand.
2458 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2460 // Add all operands of the call to the operand list.
2461 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2462 SDOperand Op = getValue(I.getOperand(i));
2463 assert(TLI.isTypeLegal(Op.getValueType()) &&
2464 "Intrinsic uses a non-legal type?");
2468 std::vector<MVT::ValueType> VTs;
2469 if (I.getType() != Type::VoidTy) {
2470 MVT::ValueType VT = TLI.getValueType(I.getType());
2471 if (MVT::isVector(VT)) {
2472 const VectorType *DestTy = cast<VectorType>(I.getType());
2473 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
2475 VT = MVT::getVectorType(EltVT, DestTy->getNumElements());
2476 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
2479 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
2483 VTs.push_back(MVT::Other);
2485 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs);
2490 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(),
2491 &Ops[0], Ops.size());
2492 else if (I.getType() != Type::VoidTy)
2493 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(),
2494 &Ops[0], Ops.size());
2496 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(),
2497 &Ops[0], Ops.size());
2500 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1);
2502 PendingLoads.push_back(Chain);
2506 if (I.getType() != Type::VoidTy) {
2507 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
2508 MVT::ValueType VT = TLI.getValueType(PTy);
2509 Result = DAG.getNode(ISD::BIT_CONVERT, VT, Result);
2511 setValue(&I, Result);
2515 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
2516 static GlobalVariable *ExtractTypeInfo (Value *V) {
2517 V = IntrinsicInst::StripPointerCasts(V);
2518 GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2519 assert (GV || isa<ConstantPointerNull>(V) &&
2520 "TypeInfo must be a global variable or NULL");
2524 /// addCatchInfo - Extract the personality and type infos from an eh.selector
2525 /// call, and add them to the specified machine basic block.
2526 static void addCatchInfo(CallInst &I, MachineModuleInfo *MMI,
2527 MachineBasicBlock *MBB) {
2528 // Inform the MachineModuleInfo of the personality for this landing pad.
2529 ConstantExpr *CE = cast<ConstantExpr>(I.getOperand(2));
2530 assert(CE->getOpcode() == Instruction::BitCast &&
2531 isa<Function>(CE->getOperand(0)) &&
2532 "Personality should be a function");
2533 MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
2535 // Gather all the type infos for this landing pad and pass them along to
2536 // MachineModuleInfo.
2537 std::vector<GlobalVariable *> TyInfo;
2538 unsigned N = I.getNumOperands();
2540 for (unsigned i = N - 1; i > 2; --i) {
2541 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(i))) {
2542 unsigned FilterLength = CI->getZExtValue();
2543 unsigned FirstCatch = i + FilterLength + !FilterLength;
2544 assert (FirstCatch <= N && "Invalid filter length");
2546 if (FirstCatch < N) {
2547 TyInfo.reserve(N - FirstCatch);
2548 for (unsigned j = FirstCatch; j < N; ++j)
2549 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2550 MMI->addCatchTypeInfo(MBB, TyInfo);
2554 if (!FilterLength) {
2556 MMI->addCleanup(MBB);
2559 TyInfo.reserve(FilterLength - 1);
2560 for (unsigned j = i + 1; j < FirstCatch; ++j)
2561 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2562 MMI->addFilterTypeInfo(MBB, TyInfo);
2571 TyInfo.reserve(N - 3);
2572 for (unsigned j = 3; j < N; ++j)
2573 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2574 MMI->addCatchTypeInfo(MBB, TyInfo);
2578 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
2579 /// we want to emit this as a call to a named external function, return the name
2580 /// otherwise lower it and return null.
2582 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
2583 switch (Intrinsic) {
2585 // By default, turn this into a target intrinsic node.
2586 visitTargetIntrinsic(I, Intrinsic);
2588 case Intrinsic::vastart: visitVAStart(I); return 0;
2589 case Intrinsic::vaend: visitVAEnd(I); return 0;
2590 case Intrinsic::vacopy: visitVACopy(I); return 0;
2591 case Intrinsic::returnaddress:
2592 setValue(&I, DAG.getNode(ISD::RETURNADDR, TLI.getPointerTy(),
2593 getValue(I.getOperand(1))));
2595 case Intrinsic::frameaddress:
2596 setValue(&I, DAG.getNode(ISD::FRAMEADDR, TLI.getPointerTy(),
2597 getValue(I.getOperand(1))));
2599 case Intrinsic::setjmp:
2600 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
2602 case Intrinsic::longjmp:
2603 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
2605 case Intrinsic::memcpy_i32:
2606 case Intrinsic::memcpy_i64:
2607 visitMemIntrinsic(I, ISD::MEMCPY);
2609 case Intrinsic::memset_i32:
2610 case Intrinsic::memset_i64:
2611 visitMemIntrinsic(I, ISD::MEMSET);
2613 case Intrinsic::memmove_i32:
2614 case Intrinsic::memmove_i64:
2615 visitMemIntrinsic(I, ISD::MEMMOVE);
2618 case Intrinsic::dbg_stoppoint: {
2619 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2620 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
2621 if (MMI && SPI.getContext() && MMI->Verify(SPI.getContext())) {
2625 Ops[1] = getValue(SPI.getLineValue());
2626 Ops[2] = getValue(SPI.getColumnValue());
2628 DebugInfoDesc *DD = MMI->getDescFor(SPI.getContext());
2629 assert(DD && "Not a debug information descriptor");
2630 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD);
2632 Ops[3] = DAG.getString(CompileUnit->getFileName());
2633 Ops[4] = DAG.getString(CompileUnit->getDirectory());
2635 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5));
2640 case Intrinsic::dbg_region_start: {
2641 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2642 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
2643 if (MMI && RSI.getContext() && MMI->Verify(RSI.getContext())) {
2644 unsigned LabelID = MMI->RecordRegionStart(RSI.getContext());
2645 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
2646 DAG.getConstant(LabelID, MVT::i32)));
2651 case Intrinsic::dbg_region_end: {
2652 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2653 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
2654 if (MMI && REI.getContext() && MMI->Verify(REI.getContext())) {
2655 unsigned LabelID = MMI->RecordRegionEnd(REI.getContext());
2656 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other,
2657 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
2662 case Intrinsic::dbg_func_start: {
2663 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2664 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
2665 if (MMI && FSI.getSubprogram() &&
2666 MMI->Verify(FSI.getSubprogram())) {
2667 unsigned LabelID = MMI->RecordRegionStart(FSI.getSubprogram());
2668 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other,
2669 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
2674 case Intrinsic::dbg_declare: {
2675 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2676 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
2677 if (MMI && DI.getVariable() && MMI->Verify(DI.getVariable())) {
2678 SDOperand AddressOp = getValue(DI.getAddress());
2679 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp))
2680 MMI->RecordVariable(DI.getVariable(), FI->getIndex());
2686 case Intrinsic::eh_exception: {
2687 if (ExceptionHandling) {
2688 if (!CurMBB->isLandingPad()) {
2689 // FIXME: Mark exception register as live in. Hack for PR1508.
2690 unsigned Reg = TLI.getExceptionAddressRegister();
2691 if (Reg) CurMBB->addLiveIn(Reg);
2693 // Insert the EXCEPTIONADDR instruction.
2694 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
2696 Ops[0] = DAG.getRoot();
2697 SDOperand Op = DAG.getNode(ISD::EXCEPTIONADDR, VTs, Ops, 1);
2699 DAG.setRoot(Op.getValue(1));
2701 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2706 case Intrinsic::eh_selector_i32:
2707 case Intrinsic::eh_selector_i64: {
2708 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2709 MVT::ValueType VT = (Intrinsic == Intrinsic::eh_selector_i32 ?
2710 MVT::i32 : MVT::i64);
2712 if (ExceptionHandling && MMI) {
2713 if (CurMBB->isLandingPad())
2714 addCatchInfo(I, MMI, CurMBB);
2717 FuncInfo.CatchInfoLost.insert(&I);
2719 // FIXME: Mark exception selector register as live in. Hack for PR1508.
2720 unsigned Reg = TLI.getExceptionSelectorRegister();
2721 if (Reg) CurMBB->addLiveIn(Reg);
2724 // Insert the EHSELECTION instruction.
2725 SDVTList VTs = DAG.getVTList(VT, MVT::Other);
2727 Ops[0] = getValue(I.getOperand(1));
2729 SDOperand Op = DAG.getNode(ISD::EHSELECTION, VTs, Ops, 2);
2731 DAG.setRoot(Op.getValue(1));
2733 setValue(&I, DAG.getConstant(0, VT));
2739 case Intrinsic::eh_typeid_for_i32:
2740 case Intrinsic::eh_typeid_for_i64: {
2741 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2742 MVT::ValueType VT = (Intrinsic == Intrinsic::eh_typeid_for_i32 ?
2743 MVT::i32 : MVT::i64);
2746 // Find the type id for the given typeinfo.
2747 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
2749 unsigned TypeID = MMI->getTypeIDFor(GV);
2750 setValue(&I, DAG.getConstant(TypeID, VT));
2752 // Return something different to eh_selector.
2753 setValue(&I, DAG.getConstant(1, VT));
2759 case Intrinsic::eh_return: {
2760 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2762 if (MMI && ExceptionHandling) {
2763 MMI->setCallsEHReturn(true);
2764 DAG.setRoot(DAG.getNode(ISD::EH_RETURN,
2767 getValue(I.getOperand(1)),
2768 getValue(I.getOperand(2))));
2770 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2776 case Intrinsic::eh_unwind_init: {
2777 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
2778 MMI->setCallsUnwindInit(true);
2784 case Intrinsic::eh_dwarf_cfa: {
2785 if (ExceptionHandling) {
2786 MVT::ValueType VT = getValue(I.getOperand(1)).getValueType();
2788 if (MVT::getSizeInBits(VT) > MVT::getSizeInBits(TLI.getPointerTy()))
2789 CfaArg = DAG.getNode(ISD::TRUNCATE,
2790 TLI.getPointerTy(), getValue(I.getOperand(1)));
2792 CfaArg = DAG.getNode(ISD::SIGN_EXTEND,
2793 TLI.getPointerTy(), getValue(I.getOperand(1)));
2795 SDOperand Offset = DAG.getNode(ISD::ADD,
2797 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET,
2798 TLI.getPointerTy()),
2800 setValue(&I, DAG.getNode(ISD::ADD,
2802 DAG.getNode(ISD::FRAMEADDR,
2805 TLI.getPointerTy())),
2808 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
2814 case Intrinsic::sqrt:
2815 setValue(&I, DAG.getNode(ISD::FSQRT,
2816 getValue(I.getOperand(1)).getValueType(),
2817 getValue(I.getOperand(1))));
2819 case Intrinsic::powi:
2820 setValue(&I, DAG.getNode(ISD::FPOWI,
2821 getValue(I.getOperand(1)).getValueType(),
2822 getValue(I.getOperand(1)),
2823 getValue(I.getOperand(2))));
2825 case Intrinsic::sin:
2826 setValue(&I, DAG.getNode(ISD::FSIN,
2827 getValue(I.getOperand(1)).getValueType(),
2828 getValue(I.getOperand(1))));
2830 case Intrinsic::cos:
2831 setValue(&I, DAG.getNode(ISD::FCOS,
2832 getValue(I.getOperand(1)).getValueType(),
2833 getValue(I.getOperand(1))));
2835 case Intrinsic::pow:
2836 setValue(&I, DAG.getNode(ISD::FPOW,
2837 getValue(I.getOperand(1)).getValueType(),
2838 getValue(I.getOperand(1)),
2839 getValue(I.getOperand(2))));
2841 case Intrinsic::pcmarker: {
2842 SDOperand Tmp = getValue(I.getOperand(1));
2843 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
2846 case Intrinsic::readcyclecounter: {
2847 SDOperand Op = getRoot();
2848 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER,
2849 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
2852 DAG.setRoot(Tmp.getValue(1));
2855 case Intrinsic::part_select: {
2856 // Currently not implemented: just abort
2857 assert(0 && "part_select intrinsic not implemented");
2860 case Intrinsic::part_set: {
2861 // Currently not implemented: just abort
2862 assert(0 && "part_set intrinsic not implemented");
2865 case Intrinsic::bswap:
2866 setValue(&I, DAG.getNode(ISD::BSWAP,
2867 getValue(I.getOperand(1)).getValueType(),
2868 getValue(I.getOperand(1))));
2870 case Intrinsic::cttz: {
2871 SDOperand Arg = getValue(I.getOperand(1));
2872 MVT::ValueType Ty = Arg.getValueType();
2873 SDOperand result = DAG.getNode(ISD::CTTZ, Ty, Arg);
2874 setValue(&I, result);
2877 case Intrinsic::ctlz: {
2878 SDOperand Arg = getValue(I.getOperand(1));
2879 MVT::ValueType Ty = Arg.getValueType();
2880 SDOperand result = DAG.getNode(ISD::CTLZ, Ty, Arg);
2881 setValue(&I, result);
2884 case Intrinsic::ctpop: {
2885 SDOperand Arg = getValue(I.getOperand(1));
2886 MVT::ValueType Ty = Arg.getValueType();
2887 SDOperand result = DAG.getNode(ISD::CTPOP, Ty, Arg);
2888 setValue(&I, result);
2891 case Intrinsic::stacksave: {
2892 SDOperand Op = getRoot();
2893 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE,
2894 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
2896 DAG.setRoot(Tmp.getValue(1));
2899 case Intrinsic::stackrestore: {
2900 SDOperand Tmp = getValue(I.getOperand(1));
2901 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
2904 case Intrinsic::prefetch:
2905 // FIXME: Currently discarding prefetches.
2908 case Intrinsic::var_annotation:
2909 // Discard annotate attributes
2912 case Intrinsic::init_trampoline: {
2914 cast<Function>(IntrinsicInst::StripPointerCasts(I.getOperand(2)));
2918 Ops[1] = getValue(I.getOperand(1));
2919 Ops[2] = getValue(I.getOperand(2));
2920 Ops[3] = getValue(I.getOperand(3));
2921 Ops[4] = DAG.getSrcValue(I.getOperand(1));
2922 Ops[5] = DAG.getSrcValue(F);
2924 SDOperand Tmp = DAG.getNode(ISD::TRAMPOLINE,
2925 DAG.getNodeValueTypes(TLI.getPointerTy(),
2930 DAG.setRoot(Tmp.getValue(1));
2937 void SelectionDAGLowering::LowerCallTo(Instruction &I,
2938 const Type *CalledValueTy,
2939 unsigned CallingConv,
2941 SDOperand Callee, unsigned OpIdx,
2942 MachineBasicBlock *LandingPad) {
2943 const PointerType *PT = cast<PointerType>(CalledValueTy);
2944 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2945 const ParamAttrsList *Attrs = FTy->getParamAttrs();
2946 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
2947 unsigned BeginLabel = 0, EndLabel = 0;
2949 TargetLowering::ArgListTy Args;
2950 TargetLowering::ArgListEntry Entry;
2951 Args.reserve(I.getNumOperands());
2952 for (unsigned i = OpIdx, e = I.getNumOperands(); i != e; ++i) {
2953 Value *Arg = I.getOperand(i);
2954 SDOperand ArgNode = getValue(Arg);
2955 Entry.Node = ArgNode; Entry.Ty = Arg->getType();
2957 unsigned attrInd = i - OpIdx + 1;
2958 Entry.isSExt = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::SExt);
2959 Entry.isZExt = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::ZExt);
2960 Entry.isInReg = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::InReg);
2961 Entry.isSRet = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::StructRet);
2962 Entry.isNest = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::Nest);
2963 Entry.isByVal = Attrs && Attrs->paramHasAttr(attrInd, ParamAttr::ByVal);
2964 Args.push_back(Entry);
2967 if (ExceptionHandling && MMI && LandingPad) {
2968 // Insert a label before the invoke call to mark the try range. This can be
2969 // used to detect deletion of the invoke via the MachineModuleInfo.
2970 BeginLabel = MMI->NextLabelID();
2971 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
2972 DAG.getConstant(BeginLabel, MVT::i32)));
2975 std::pair<SDOperand,SDOperand> Result =
2976 TLI.LowerCallTo(getRoot(), I.getType(),
2977 Attrs && Attrs->paramHasAttr(0, ParamAttr::SExt),
2978 FTy->isVarArg(), CallingConv, IsTailCall,
2980 if (I.getType() != Type::VoidTy)
2981 setValue(&I, Result.first);
2982 DAG.setRoot(Result.second);
2984 if (ExceptionHandling && MMI && LandingPad) {
2985 // Insert a label at the end of the invoke call to mark the try range. This
2986 // can be used to detect deletion of the invoke via the MachineModuleInfo.
2987 EndLabel = MMI->NextLabelID();
2988 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, getRoot(),
2989 DAG.getConstant(EndLabel, MVT::i32)));
2991 // Inform MachineModuleInfo of range.
2992 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
2997 void SelectionDAGLowering::visitCall(CallInst &I) {
2998 const char *RenameFn = 0;
2999 if (Function *F = I.getCalledFunction()) {
3000 if (F->isDeclaration()) {
3001 if (unsigned IID = F->getIntrinsicID()) {
3002 RenameFn = visitIntrinsicCall(I, IID);
3008 // Check for well-known libc/libm calls. If the function is internal, it
3009 // can't be a library call.
3010 unsigned NameLen = F->getNameLen();
3011 if (!F->hasInternalLinkage() && NameLen) {
3012 const char *NameStr = F->getNameStart();
3013 if (NameStr[0] == 'c' &&
3014 ((NameLen == 8 && !strcmp(NameStr, "copysign")) ||
3015 (NameLen == 9 && !strcmp(NameStr, "copysignf")))) {
3016 if (I.getNumOperands() == 3 && // Basic sanity checks.
3017 I.getOperand(1)->getType()->isFloatingPoint() &&
3018 I.getType() == I.getOperand(1)->getType() &&
3019 I.getType() == I.getOperand(2)->getType()) {
3020 SDOperand LHS = getValue(I.getOperand(1));
3021 SDOperand RHS = getValue(I.getOperand(2));
3022 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
3026 } else if (NameStr[0] == 'f' &&
3027 ((NameLen == 4 && !strcmp(NameStr, "fabs")) ||
3028 (NameLen == 5 && !strcmp(NameStr, "fabsf")) ||
3029 (NameLen == 5 && !strcmp(NameStr, "fabsl")))) {
3030 if (I.getNumOperands() == 2 && // Basic sanity checks.
3031 I.getOperand(1)->getType()->isFloatingPoint() &&
3032 I.getType() == I.getOperand(1)->getType()) {
3033 SDOperand Tmp = getValue(I.getOperand(1));
3034 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
3037 } else if (NameStr[0] == 's' &&
3038 ((NameLen == 3 && !strcmp(NameStr, "sin")) ||
3039 (NameLen == 4 && !strcmp(NameStr, "sinf")) ||
3040 (NameLen == 4 && !strcmp(NameStr, "sinl")))) {
3041 if (I.getNumOperands() == 2 && // Basic sanity checks.
3042 I.getOperand(1)->getType()->isFloatingPoint() &&
3043 I.getType() == I.getOperand(1)->getType()) {
3044 SDOperand Tmp = getValue(I.getOperand(1));
3045 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
3048 } else if (NameStr[0] == 'c' &&
3049 ((NameLen == 3 && !strcmp(NameStr, "cos")) ||
3050 (NameLen == 4 && !strcmp(NameStr, "cosf")) ||
3051 (NameLen == 4 && !strcmp(NameStr, "cosl")))) {
3052 if (I.getNumOperands() == 2 && // Basic sanity checks.
3053 I.getOperand(1)->getType()->isFloatingPoint() &&
3054 I.getType() == I.getOperand(1)->getType()) {
3055 SDOperand Tmp = getValue(I.getOperand(1));
3056 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
3061 } else if (isa<InlineAsm>(I.getOperand(0))) {
3068 Callee = getValue(I.getOperand(0));
3070 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
3072 LowerCallTo(I, I.getCalledValue()->getType(),
3080 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
3081 /// this value and returns the result as a ValueVT value. This uses
3082 /// Chain/Flag as the input and updates them for the output Chain/Flag.
3083 /// If the Flag pointer is NULL, no flag is used.
3084 SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
3085 SDOperand &Chain, SDOperand *Flag)const{
3086 // Copy the legal parts from the registers.
3087 unsigned NumParts = Regs.size();
3088 SmallVector<SDOperand, 8> Parts(NumParts);
3089 for (unsigned i = 0; i != NumParts; ++i) {
3090 SDOperand Part = Flag ?
3091 DAG.getCopyFromReg(Chain, Regs[i], RegVT, *Flag) :
3092 DAG.getCopyFromReg(Chain, Regs[i], RegVT);
3093 Chain = Part.getValue(1);
3095 *Flag = Part.getValue(2);
3099 // Assemble the legal parts into the final value.
3100 return getCopyFromParts(DAG, &Parts[0], NumParts, RegVT, ValueVT);
3103 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
3104 /// specified value into the registers specified by this object. This uses
3105 /// Chain/Flag as the input and updates them for the output Chain/Flag.
3106 /// If the Flag pointer is NULL, no flag is used.
3107 void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
3108 SDOperand &Chain, SDOperand *Flag) const {
3109 // Get the list of the values's legal parts.
3110 unsigned NumParts = Regs.size();
3111 SmallVector<SDOperand, 8> Parts(NumParts);
3112 getCopyToParts(DAG, Val, &Parts[0], NumParts, RegVT);
3114 // Copy the parts into the registers.
3115 for (unsigned i = 0; i != NumParts; ++i) {
3116 SDOperand Part = Flag ?
3117 DAG.getCopyToReg(Chain, Regs[i], Parts[i], *Flag) :
3118 DAG.getCopyToReg(Chain, Regs[i], Parts[i]);
3119 Chain = Part.getValue(0);
3121 *Flag = Part.getValue(1);
3125 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
3126 /// operand list. This adds the code marker and includes the number of
3127 /// values added into it.
3128 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
3129 std::vector<SDOperand> &Ops) const {
3130 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
3131 Ops.push_back(DAG.getTargetConstant(Code | (Regs.size() << 3), IntPtrTy));
3132 for (unsigned i = 0, e = Regs.size(); i != e; ++i)
3133 Ops.push_back(DAG.getRegister(Regs[i], RegVT));
3136 /// isAllocatableRegister - If the specified register is safe to allocate,
3137 /// i.e. it isn't a stack pointer or some other special register, return the
3138 /// register class for the register. Otherwise, return null.
3139 static const TargetRegisterClass *
3140 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
3141 const TargetLowering &TLI, const MRegisterInfo *MRI) {
3142 MVT::ValueType FoundVT = MVT::Other;
3143 const TargetRegisterClass *FoundRC = 0;
3144 for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(),
3145 E = MRI->regclass_end(); RCI != E; ++RCI) {
3146 MVT::ValueType ThisVT = MVT::Other;
3148 const TargetRegisterClass *RC = *RCI;
3149 // If none of the the value types for this register class are valid, we
3150 // can't use it. For example, 64-bit reg classes on 32-bit targets.
3151 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
3153 if (TLI.isTypeLegal(*I)) {
3154 // If we have already found this register in a different register class,
3155 // choose the one with the largest VT specified. For example, on
3156 // PowerPC, we favor f64 register classes over f32.
3157 if (FoundVT == MVT::Other ||
3158 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) {
3165 if (ThisVT == MVT::Other) continue;
3167 // NOTE: This isn't ideal. In particular, this might allocate the
3168 // frame pointer in functions that need it (due to them not being taken
3169 // out of allocation, because a variable sized allocation hasn't been seen
3170 // yet). This is a slight code pessimization, but should still work.
3171 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
3172 E = RC->allocation_order_end(MF); I != E; ++I)
3174 // We found a matching register class. Keep looking at others in case
3175 // we find one with larger registers that this physreg is also in.
3186 /// AsmOperandInfo - This contains information for each constraint that we are
3188 struct AsmOperandInfo : public InlineAsm::ConstraintInfo {
3189 /// ConstraintCode - This contains the actual string for the code, like "m".
3190 std::string ConstraintCode;
3192 /// ConstraintType - Information about the constraint code, e.g. Register,
3193 /// RegisterClass, Memory, Other, Unknown.
3194 TargetLowering::ConstraintType ConstraintType;
3196 /// CallOperand/CallOperandval - If this is the result output operand or a
3197 /// clobber, this is null, otherwise it is the incoming operand to the
3198 /// CallInst. This gets modified as the asm is processed.
3199 SDOperand CallOperand;
3200 Value *CallOperandVal;
3202 /// ConstraintVT - The ValueType for the operand value.
3203 MVT::ValueType ConstraintVT;
3205 /// AssignedRegs - If this is a register or register class operand, this
3206 /// contains the set of register corresponding to the operand.
3207 RegsForValue AssignedRegs;
3209 AsmOperandInfo(const InlineAsm::ConstraintInfo &info)
3210 : InlineAsm::ConstraintInfo(info),
3211 ConstraintType(TargetLowering::C_Unknown),
3212 CallOperand(0,0), CallOperandVal(0), ConstraintVT(MVT::Other) {
3215 void ComputeConstraintToUse(const TargetLowering &TLI);
3217 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
3218 /// busy in OutputRegs/InputRegs.
3219 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
3220 std::set<unsigned> &OutputRegs,
3221 std::set<unsigned> &InputRegs) const {
3223 OutputRegs.insert(AssignedRegs.Regs.begin(), AssignedRegs.Regs.end());
3225 InputRegs.insert(AssignedRegs.Regs.begin(), AssignedRegs.Regs.end());
3228 } // end anon namespace.
3230 /// getConstraintGenerality - Return an integer indicating how general CT is.
3231 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) {
3233 default: assert(0 && "Unknown constraint type!");
3234 case TargetLowering::C_Other:
3235 case TargetLowering::C_Unknown:
3237 case TargetLowering::C_Register:
3239 case TargetLowering::C_RegisterClass:
3241 case TargetLowering::C_Memory:
3246 void AsmOperandInfo::ComputeConstraintToUse(const TargetLowering &TLI) {
3247 assert(!Codes.empty() && "Must have at least one constraint");
3249 std::string *Current = &Codes[0];
3250 TargetLowering::ConstraintType CurType = TLI.getConstraintType(*Current);
3251 if (Codes.size() == 1) { // Single-letter constraints ('r') are very common.
3252 ConstraintCode = *Current;
3253 ConstraintType = CurType;
3257 unsigned CurGenerality = getConstraintGenerality(CurType);
3259 // If we have multiple constraints, try to pick the most general one ahead
3260 // of time. This isn't a wonderful solution, but handles common cases.
3261 for (unsigned j = 1, e = Codes.size(); j != e; ++j) {
3262 TargetLowering::ConstraintType ThisType = TLI.getConstraintType(Codes[j]);
3263 unsigned ThisGenerality = getConstraintGenerality(ThisType);
3264 if (ThisGenerality > CurGenerality) {
3265 // This constraint letter is more general than the previous one,
3268 Current = &Codes[j];
3269 CurGenerality = ThisGenerality;
3273 ConstraintCode = *Current;
3274 ConstraintType = CurType;
3278 void SelectionDAGLowering::
3279 GetRegistersForValue(AsmOperandInfo &OpInfo, bool HasEarlyClobber,
3280 std::set<unsigned> &OutputRegs,
3281 std::set<unsigned> &InputRegs) {
3282 // Compute whether this value requires an input register, an output register,
3284 bool isOutReg = false;
3285 bool isInReg = false;
3286 switch (OpInfo.Type) {
3287 case InlineAsm::isOutput:
3290 // If this is an early-clobber output, or if there is an input
3291 // constraint that matches this, we need to reserve the input register
3292 // so no other inputs allocate to it.
3293 isInReg = OpInfo.isEarlyClobber || OpInfo.hasMatchingInput;
3295 case InlineAsm::isInput:
3299 case InlineAsm::isClobber:
3306 MachineFunction &MF = DAG.getMachineFunction();
3307 std::vector<unsigned> Regs;
3309 // If this is a constraint for a single physreg, or a constraint for a
3310 // register class, find it.
3311 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
3312 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
3313 OpInfo.ConstraintVT);
3315 unsigned NumRegs = 1;
3316 if (OpInfo.ConstraintVT != MVT::Other)
3317 NumRegs = TLI.getNumRegisters(OpInfo.ConstraintVT);
3318 MVT::ValueType RegVT;
3319 MVT::ValueType ValueVT = OpInfo.ConstraintVT;
3322 // If this is a constraint for a specific physical register, like {r17},
3324 if (PhysReg.first) {
3325 if (OpInfo.ConstraintVT == MVT::Other)
3326 ValueVT = *PhysReg.second->vt_begin();
3328 // Get the actual register value type. This is important, because the user
3329 // may have asked for (e.g.) the AX register in i32 type. We need to
3330 // remember that AX is actually i16 to get the right extension.
3331 RegVT = *PhysReg.second->vt_begin();
3333 // This is a explicit reference to a physical register.
3334 Regs.push_back(PhysReg.first);
3336 // If this is an expanded reference, add the rest of the regs to Regs.
3338 TargetRegisterClass::iterator I = PhysReg.second->begin();
3339 TargetRegisterClass::iterator E = PhysReg.second->end();
3340 for (; *I != PhysReg.first; ++I)
3341 assert(I != E && "Didn't find reg!");
3343 // Already added the first reg.
3345 for (; NumRegs; --NumRegs, ++I) {
3346 assert(I != E && "Ran out of registers to allocate!");
3350 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
3351 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs);
3355 // Otherwise, if this was a reference to an LLVM register class, create vregs
3356 // for this reference.
3357 std::vector<unsigned> RegClassRegs;
3358 const TargetRegisterClass *RC = PhysReg.second;
3360 // If this is an early clobber or tied register, our regalloc doesn't know
3361 // how to maintain the constraint. If it isn't, go ahead and create vreg
3362 // and let the regalloc do the right thing.
3363 if (!OpInfo.hasMatchingInput && !OpInfo.isEarlyClobber &&
3364 // If there is some other early clobber and this is an input register,
3365 // then we are forced to pre-allocate the input reg so it doesn't
3366 // conflict with the earlyclobber.
3367 !(OpInfo.Type == InlineAsm::isInput && HasEarlyClobber)) {
3368 RegVT = *PhysReg.second->vt_begin();
3370 if (OpInfo.ConstraintVT == MVT::Other)
3373 // Create the appropriate number of virtual registers.
3374 SSARegMap *RegMap = MF.getSSARegMap();
3375 for (; NumRegs; --NumRegs)
3376 Regs.push_back(RegMap->createVirtualRegister(PhysReg.second));
3378 OpInfo.AssignedRegs = RegsForValue(Regs, RegVT, ValueVT);
3379 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs);
3383 // Otherwise, we can't allocate it. Let the code below figure out how to
3384 // maintain these constraints.
3385 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end());
3388 // This is a reference to a register class that doesn't directly correspond
3389 // to an LLVM register class. Allocate NumRegs consecutive, available,
3390 // registers from the class.
3391 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
3392 OpInfo.ConstraintVT);
3395 const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo();
3396 unsigned NumAllocated = 0;
3397 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
3398 unsigned Reg = RegClassRegs[i];
3399 // See if this register is available.
3400 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
3401 (isInReg && InputRegs.count(Reg))) { // Already used.
3402 // Make sure we find consecutive registers.
3407 // Check to see if this register is allocatable (i.e. don't give out the
3410 RC = isAllocatableRegister(Reg, MF, TLI, MRI);
3411 if (!RC) { // Couldn't allocate this register.
3412 // Reset NumAllocated to make sure we return consecutive registers.
3418 // Okay, this register is good, we can use it.
3421 // If we allocated enough consecutive registers, succeed.
3422 if (NumAllocated == NumRegs) {
3423 unsigned RegStart = (i-NumAllocated)+1;
3424 unsigned RegEnd = i+1;
3425 // Mark all of the allocated registers used.
3426 for (unsigned i = RegStart; i != RegEnd; ++i)
3427 Regs.push_back(RegClassRegs[i]);
3429 OpInfo.AssignedRegs = RegsForValue(Regs, *RC->vt_begin(),
3430 OpInfo.ConstraintVT);
3431 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs);
3436 // Otherwise, we couldn't allocate enough registers for this.
3441 /// visitInlineAsm - Handle a call to an InlineAsm object.
3443 void SelectionDAGLowering::visitInlineAsm(CallInst &I) {
3444 InlineAsm *IA = cast<InlineAsm>(I.getOperand(0));
3446 /// ConstraintOperands - Information about all of the constraints.
3447 std::vector<AsmOperandInfo> ConstraintOperands;
3449 SDOperand Chain = getRoot();
3452 std::set<unsigned> OutputRegs, InputRegs;
3454 // Do a prepass over the constraints, canonicalizing them, and building up the
3455 // ConstraintOperands list.
3456 std::vector<InlineAsm::ConstraintInfo>
3457 ConstraintInfos = IA->ParseConstraints();
3459 // SawEarlyClobber - Keep track of whether we saw an earlyclobber output
3460 // constraint. If so, we can't let the register allocator allocate any input
3461 // registers, because it will not know to avoid the earlyclobbered output reg.
3462 bool SawEarlyClobber = false;
3464 unsigned OpNo = 1; // OpNo - The operand of the CallInst.
3465 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
3466 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i]));
3467 AsmOperandInfo &OpInfo = ConstraintOperands.back();
3469 MVT::ValueType OpVT = MVT::Other;
3471 // Compute the value type for each operand.
3472 switch (OpInfo.Type) {
3473 case InlineAsm::isOutput:
3474 if (!OpInfo.isIndirect) {
3475 // The return value of the call is this value. As such, there is no
3476 // corresponding argument.
3477 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
3478 OpVT = TLI.getValueType(I.getType());
3480 OpInfo.CallOperandVal = I.getOperand(OpNo++);
3483 case InlineAsm::isInput:
3484 OpInfo.CallOperandVal = I.getOperand(OpNo++);
3486 case InlineAsm::isClobber:
3491 // If this is an input or an indirect output, process the call argument.
3492 if (OpInfo.CallOperandVal) {
3493 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
3494 const Type *OpTy = OpInfo.CallOperandVal->getType();
3495 // If this is an indirect operand, the operand is a pointer to the
3497 if (OpInfo.isIndirect)
3498 OpTy = cast<PointerType>(OpTy)->getElementType();
3500 // If OpTy is not a first-class value, it may be a struct/union that we
3501 // can tile with integers.
3502 if (!OpTy->isFirstClassType() && OpTy->isSized()) {
3503 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
3511 OpTy = IntegerType::get(BitSize);
3516 OpVT = TLI.getValueType(OpTy, true);
3519 OpInfo.ConstraintVT = OpVT;
3521 // Compute the constraint code and ConstraintType to use.
3522 OpInfo.ComputeConstraintToUse(TLI);
3524 // Keep track of whether we see an earlyclobber.
3525 SawEarlyClobber |= OpInfo.isEarlyClobber;
3527 // If this is a memory input, and if the operand is not indirect, do what we
3528 // need to to provide an address for the memory input.
3529 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
3530 !OpInfo.isIndirect) {
3531 assert(OpInfo.Type == InlineAsm::isInput &&
3532 "Can only indirectify direct input operands!");
3534 // Memory operands really want the address of the value. If we don't have
3535 // an indirect input, put it in the constpool if we can, otherwise spill
3536 // it to a stack slot.
3538 // If the operand is a float, integer, or vector constant, spill to a
3539 // constant pool entry to get its address.
3540 Value *OpVal = OpInfo.CallOperandVal;
3541 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
3542 isa<ConstantVector>(OpVal)) {
3543 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
3544 TLI.getPointerTy());
3546 // Otherwise, create a stack slot and emit a store to it before the
3548 const Type *Ty = OpVal->getType();
3549 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
3550 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
3551 MachineFunction &MF = DAG.getMachineFunction();
3552 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align);
3553 SDOperand StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
3554 Chain = DAG.getStore(Chain, OpInfo.CallOperand, StackSlot, NULL, 0);
3555 OpInfo.CallOperand = StackSlot;
3558 // There is no longer a Value* corresponding to this operand.
3559 OpInfo.CallOperandVal = 0;
3560 // It is now an indirect operand.
3561 OpInfo.isIndirect = true;
3564 // If this constraint is for a specific register, allocate it before
3566 if (OpInfo.ConstraintType == TargetLowering::C_Register)
3567 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs);
3569 ConstraintInfos.clear();
3572 // Second pass - Loop over all of the operands, assigning virtual or physregs
3573 // to registerclass operands.
3574 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
3575 AsmOperandInfo &OpInfo = ConstraintOperands[i];
3577 // C_Register operands have already been allocated, Other/Memory don't need
3579 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
3580 GetRegistersForValue(OpInfo, SawEarlyClobber, OutputRegs, InputRegs);
3583 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
3584 std::vector<SDOperand> AsmNodeOperands;
3585 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain
3586 AsmNodeOperands.push_back(
3587 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other));
3590 // Loop over all of the inputs, copying the operand values into the
3591 // appropriate registers and processing the output regs.
3592 RegsForValue RetValRegs;
3594 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
3595 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
3597 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
3598 AsmOperandInfo &OpInfo = ConstraintOperands[i];
3600 switch (OpInfo.Type) {
3601 case InlineAsm::isOutput: {
3602 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
3603 OpInfo.ConstraintType != TargetLowering::C_Register) {
3604 // Memory output, or 'other' output (e.g. 'X' constraint).
3605 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
3607 // Add information to the INLINEASM node to know about this output.
3608 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
3609 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3610 TLI.getPointerTy()));
3611 AsmNodeOperands.push_back(OpInfo.CallOperand);
3615 // Otherwise, this is a register or register class output.
3617 // Copy the output from the appropriate register. Find a register that
3619 if (OpInfo.AssignedRegs.Regs.empty()) {
3620 cerr << "Couldn't allocate output reg for contraint '"
3621 << OpInfo.ConstraintCode << "'!\n";
3625 if (!OpInfo.isIndirect) {
3626 // This is the result value of the call.
3627 assert(RetValRegs.Regs.empty() &&
3628 "Cannot have multiple output constraints yet!");
3629 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
3630 RetValRegs = OpInfo.AssignedRegs;
3632 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
3633 OpInfo.CallOperandVal));
3636 // Add information to the INLINEASM node to know that this register is
3638 OpInfo.AssignedRegs.AddInlineAsmOperands(2 /*REGDEF*/, DAG,
3642 case InlineAsm::isInput: {
3643 SDOperand InOperandVal = OpInfo.CallOperand;
3645 if (isdigit(OpInfo.ConstraintCode[0])) { // Matching constraint?
3646 // If this is required to match an output register we have already set,
3647 // just use its register.
3648 unsigned OperandNo = atoi(OpInfo.ConstraintCode.c_str());
3650 // Scan until we find the definition we already emitted of this operand.
3651 // When we find it, create a RegsForValue operand.
3652 unsigned CurOp = 2; // The first operand.
3653 for (; OperandNo; --OperandNo) {
3654 // Advance to the next operand.
3656 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
3657 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
3658 (NumOps & 7) == 4 /*MEM*/) &&
3659 "Skipped past definitions?");
3660 CurOp += (NumOps>>3)+1;
3664 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
3665 if ((NumOps & 7) == 2 /*REGDEF*/) {
3666 // Add NumOps>>3 registers to MatchedRegs.
3667 RegsForValue MatchedRegs;
3668 MatchedRegs.ValueVT = InOperandVal.getValueType();
3669 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType();
3670 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
3672 cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
3673 MatchedRegs.Regs.push_back(Reg);
3676 // Use the produced MatchedRegs object to
3677 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
3678 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
3681 assert((NumOps & 7) == 4/*MEM*/ && "Unknown matching constraint!");
3682 assert(0 && "matching constraints for memory operands unimp");
3686 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
3687 assert(!OpInfo.isIndirect &&
3688 "Don't know how to handle indirect other inputs yet!");
3690 std::vector<SDOperand> Ops;
3691 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
3694 cerr << "Invalid operand for inline asm constraint '"
3695 << OpInfo.ConstraintCode << "'!\n";
3699 // Add information to the INLINEASM node to know about this input.
3700 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
3701 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3702 TLI.getPointerTy()));
3703 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
3705 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
3706 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
3707 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
3708 "Memory operands expect pointer values");
3710 // Add information to the INLINEASM node to know about this input.
3711 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
3712 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
3713 TLI.getPointerTy()));
3714 AsmNodeOperands.push_back(InOperandVal);
3718 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
3719 OpInfo.ConstraintType == TargetLowering::C_Register) &&
3720 "Unknown constraint type!");
3721 assert(!OpInfo.isIndirect &&
3722 "Don't know how to handle indirect register inputs yet!");
3724 // Copy the input into the appropriate registers.
3725 assert(!OpInfo.AssignedRegs.Regs.empty() &&
3726 "Couldn't allocate input reg!");
3728 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
3730 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG,
3734 case InlineAsm::isClobber: {
3735 // Add the clobbered value to the operand list, so that the register
3736 // allocator is aware that the physreg got clobbered.
3737 if (!OpInfo.AssignedRegs.Regs.empty())
3738 OpInfo.AssignedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG,
3745 // Finish up input operands.
3746 AsmNodeOperands[0] = Chain;
3747 if (Flag.Val) AsmNodeOperands.push_back(Flag);
3749 Chain = DAG.getNode(ISD::INLINEASM,
3750 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
3751 &AsmNodeOperands[0], AsmNodeOperands.size());
3752 Flag = Chain.getValue(1);
3754 // If this asm returns a register value, copy the result from that register
3755 // and set it as the value of the call.
3756 if (!RetValRegs.Regs.empty()) {
3757 SDOperand Val = RetValRegs.getCopyFromRegs(DAG, Chain, &Flag);
3759 // If the result of the inline asm is a vector, it may have the wrong
3760 // width/num elts. Make sure to convert it to the right type with
3762 if (MVT::isVector(Val.getValueType())) {
3763 const VectorType *VTy = cast<VectorType>(I.getType());
3764 MVT::ValueType DesiredVT = TLI.getValueType(VTy);
3766 Val = DAG.getNode(ISD::BIT_CONVERT, DesiredVT, Val);
3772 std::vector<std::pair<SDOperand, Value*> > StoresToEmit;
3774 // Process indirect outputs, first output all of the flagged copies out of
3776 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
3777 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
3778 Value *Ptr = IndirectStoresToEmit[i].second;
3779 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, &Flag);
3780 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
3783 // Emit the non-flagged stores from the physregs.
3784 SmallVector<SDOperand, 8> OutChains;
3785 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
3786 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first,
3787 getValue(StoresToEmit[i].second),
3788 StoresToEmit[i].second, 0));
3789 if (!OutChains.empty())
3790 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
3791 &OutChains[0], OutChains.size());
3796 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
3797 SDOperand Src = getValue(I.getOperand(0));
3799 MVT::ValueType IntPtr = TLI.getPointerTy();
3801 if (IntPtr < Src.getValueType())
3802 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
3803 else if (IntPtr > Src.getValueType())
3804 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
3806 // Scale the source by the type size.
3807 uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType());
3808 Src = DAG.getNode(ISD::MUL, Src.getValueType(),
3809 Src, getIntPtrConstant(ElementSize));
3811 TargetLowering::ArgListTy Args;
3812 TargetLowering::ArgListEntry Entry;
3814 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3815 Args.push_back(Entry);
3817 std::pair<SDOperand,SDOperand> Result =
3818 TLI.LowerCallTo(getRoot(), I.getType(), false, false, CallingConv::C, true,
3819 DAG.getExternalSymbol("malloc", IntPtr),
3821 setValue(&I, Result.first); // Pointers always fit in registers
3822 DAG.setRoot(Result.second);
3825 void SelectionDAGLowering::visitFree(FreeInst &I) {
3826 TargetLowering::ArgListTy Args;
3827 TargetLowering::ArgListEntry Entry;
3828 Entry.Node = getValue(I.getOperand(0));
3829 Entry.Ty = TLI.getTargetData()->getIntPtrType();
3830 Args.push_back(Entry);
3831 MVT::ValueType IntPtr = TLI.getPointerTy();
3832 std::pair<SDOperand,SDOperand> Result =
3833 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, CallingConv::C, true,
3834 DAG.getExternalSymbol("free", IntPtr), Args, DAG);
3835 DAG.setRoot(Result.second);
3838 // InsertAtEndOfBasicBlock - This method should be implemented by targets that
3839 // mark instructions with the 'usesCustomDAGSchedInserter' flag. These
3840 // instructions are special in various ways, which require special support to
3841 // insert. The specified MachineInstr is created but not inserted into any
3842 // basic blocks, and the scheduler passes ownership of it to this method.
3843 MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
3844 MachineBasicBlock *MBB) {
3845 cerr << "If a target marks an instruction with "
3846 << "'usesCustomDAGSchedInserter', it must implement "
3847 << "TargetLowering::InsertAtEndOfBasicBlock!\n";
3852 void SelectionDAGLowering::visitVAStart(CallInst &I) {
3853 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
3854 getValue(I.getOperand(1)),
3855 DAG.getSrcValue(I.getOperand(1))));
3858 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
3859 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
3860 getValue(I.getOperand(0)),
3861 DAG.getSrcValue(I.getOperand(0)));
3863 DAG.setRoot(V.getValue(1));
3866 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
3867 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
3868 getValue(I.getOperand(1)),
3869 DAG.getSrcValue(I.getOperand(1))));
3872 void SelectionDAGLowering::visitVACopy(CallInst &I) {
3873 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
3874 getValue(I.getOperand(1)),
3875 getValue(I.getOperand(2)),
3876 DAG.getSrcValue(I.getOperand(1)),
3877 DAG.getSrcValue(I.getOperand(2))));
3880 /// TargetLowering::LowerArguments - This is the default LowerArguments
3881 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
3882 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
3883 /// integrated into SDISel.
3884 std::vector<SDOperand>
3885 TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
3886 const FunctionType *FTy = F.getFunctionType();
3887 const ParamAttrsList *Attrs = FTy->getParamAttrs();
3888 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
3889 std::vector<SDOperand> Ops;
3890 Ops.push_back(DAG.getRoot());
3891 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
3892 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
3894 // Add one result value for each formal argument.
3895 std::vector<MVT::ValueType> RetVals;
3897 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
3899 MVT::ValueType VT = getValueType(I->getType());
3900 unsigned Flags = ISD::ParamFlags::NoFlagSet;
3901 unsigned OriginalAlignment =
3902 getTargetData()->getABITypeAlignment(I->getType());
3904 // FIXME: Distinguish between a formal with no [sz]ext attribute from one
3905 // that is zero extended!
3906 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::ZExt))
3907 Flags &= ~(ISD::ParamFlags::SExt);
3908 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::SExt))
3909 Flags |= ISD::ParamFlags::SExt;
3910 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::InReg))
3911 Flags |= ISD::ParamFlags::InReg;
3912 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::StructRet))
3913 Flags |= ISD::ParamFlags::StructReturn;
3914 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::ByVal)) {
3915 Flags |= ISD::ParamFlags::ByVal;
3916 const PointerType *Ty = cast<PointerType>(I->getType());
3917 const StructType *STy = cast<StructType>(Ty->getElementType());
3918 unsigned StructAlign =
3919 Log2_32(getTargetData()->getCallFrameTypeAlignment(STy));
3920 unsigned StructSize = getTargetData()->getTypeSize(STy);
3921 Flags |= (StructAlign << ISD::ParamFlags::ByValAlignOffs);
3922 Flags |= (StructSize << ISD::ParamFlags::ByValSizeOffs);
3924 if (Attrs && Attrs->paramHasAttr(j, ParamAttr::Nest))
3925 Flags |= ISD::ParamFlags::Nest;
3926 Flags |= (OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs);
3928 switch (getTypeAction(VT)) {
3929 default: assert(0 && "Unknown type action!");
3931 RetVals.push_back(VT);
3932 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
3935 RetVals.push_back(getTypeToTransformTo(VT));
3936 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
3939 // If this is an illegal type, it needs to be broken up to fit into
3941 MVT::ValueType RegisterVT = getRegisterType(VT);
3942 unsigned NumRegs = getNumRegisters(VT);
3943 for (unsigned i = 0; i != NumRegs; ++i) {
3944 RetVals.push_back(RegisterVT);
3945 // if it isn't first piece, alignment must be 1
3947 Flags = (Flags & (~ISD::ParamFlags::OrigAlignment)) |
3948 (1 << ISD::ParamFlags::OrigAlignmentOffs);
3949 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
3956 RetVals.push_back(MVT::Other);
3959 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS,
3960 DAG.getNodeValueTypes(RetVals), RetVals.size(),
3961 &Ops[0], Ops.size()).Val;
3962 unsigned NumArgRegs = Result->getNumValues() - 1;
3963 DAG.setRoot(SDOperand(Result, NumArgRegs));
3965 // Set up the return result vector.
3969 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
3971 MVT::ValueType VT = getValueType(I->getType());
3973 switch (getTypeAction(VT)) {
3974 default: assert(0 && "Unknown type action!");
3976 Ops.push_back(SDOperand(Result, i++));
3979 SDOperand Op(Result, i++);
3980 if (MVT::isInteger(VT)) {
3981 if (Attrs && Attrs->paramHasAttr(Idx, ParamAttr::SExt))
3982 Op = DAG.getNode(ISD::AssertSext, Op.getValueType(), Op,
3983 DAG.getValueType(VT));
3984 else if (Attrs && Attrs->paramHasAttr(Idx, ParamAttr::ZExt))
3985 Op = DAG.getNode(ISD::AssertZext, Op.getValueType(), Op,
3986 DAG.getValueType(VT));
3987 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
3989 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
3990 Op = DAG.getNode(ISD::FP_ROUND, VT, Op);
3996 MVT::ValueType PartVT = getRegisterType(VT);
3997 unsigned NumParts = getNumRegisters(VT);
3998 SmallVector<SDOperand, 4> Parts(NumParts);
3999 for (unsigned j = 0; j != NumParts; ++j)
4000 Parts[j] = SDOperand(Result, i++);
4001 Ops.push_back(getCopyFromParts(DAG, &Parts[0], NumParts, PartVT, VT));
4006 assert(i == NumArgRegs && "Argument register count mismatch!");
4011 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
4012 /// implementation, which just inserts an ISD::CALL node, which is later custom
4013 /// lowered by the target to something concrete. FIXME: When all targets are
4014 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
4015 std::pair<SDOperand, SDOperand>
4016 TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
4017 bool RetTyIsSigned, bool isVarArg,
4018 unsigned CallingConv, bool isTailCall,
4020 ArgListTy &Args, SelectionDAG &DAG) {
4021 SmallVector<SDOperand, 32> Ops;
4022 Ops.push_back(Chain); // Op#0 - Chain
4023 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC
4024 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg
4025 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail
4026 Ops.push_back(Callee);
4028 // Handle all of the outgoing arguments.
4029 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
4030 MVT::ValueType VT = getValueType(Args[i].Ty);
4031 SDOperand Op = Args[i].Node;
4032 unsigned Flags = ISD::ParamFlags::NoFlagSet;
4033 unsigned OriginalAlignment =
4034 getTargetData()->getABITypeAlignment(Args[i].Ty);
4037 Flags |= ISD::ParamFlags::SExt;
4039 Flags |= ISD::ParamFlags::ZExt;
4040 if (Args[i].isInReg)
4041 Flags |= ISD::ParamFlags::InReg;
4043 Flags |= ISD::ParamFlags::StructReturn;
4044 if (Args[i].isByVal) {
4045 Flags |= ISD::ParamFlags::ByVal;
4046 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
4047 const StructType *STy = cast<StructType>(Ty->getElementType());
4048 unsigned StructAlign =
4049 Log2_32(getTargetData()->getCallFrameTypeAlignment(STy));
4050 unsigned StructSize = getTargetData()->getTypeSize(STy);
4051 Flags |= (StructAlign << ISD::ParamFlags::ByValAlignOffs);
4052 Flags |= (StructSize << ISD::ParamFlags::ByValSizeOffs);
4055 Flags |= ISD::ParamFlags::Nest;
4056 Flags |= OriginalAlignment << ISD::ParamFlags::OrigAlignmentOffs;
4058 switch (getTypeAction(VT)) {
4059 default: assert(0 && "Unknown type action!");
4062 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
4065 if (MVT::isInteger(VT)) {
4068 ExtOp = ISD::SIGN_EXTEND;
4069 else if (Args[i].isZExt)
4070 ExtOp = ISD::ZERO_EXTEND;
4072 ExtOp = ISD::ANY_EXTEND;
4073 Op = DAG.getNode(ExtOp, getTypeToTransformTo(VT), Op);
4075 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
4076 Op = DAG.getNode(ISD::FP_EXTEND, getTypeToTransformTo(VT), Op);
4079 Ops.push_back(DAG.getConstant(Flags, MVT::i32));
4082 MVT::ValueType PartVT = getRegisterType(VT);
4083 unsigned NumParts = getNumRegisters(VT);
4084 SmallVector<SDOperand, 4> Parts(NumParts);
4085 getCopyToParts(DAG, Op, &Parts[0], NumParts, PartVT);
4086 for (unsigned i = 0; i != NumParts; ++i) {
4087 // if it isn't first piece, alignment must be 1
4088 unsigned MyFlags = Flags;
4090 MyFlags = (MyFlags & (~ISD::ParamFlags::OrigAlignment)) |
4091 (1 << ISD::ParamFlags::OrigAlignmentOffs);
4093 Ops.push_back(Parts[i]);
4094 Ops.push_back(DAG.getConstant(MyFlags, MVT::i32));
4101 // Figure out the result value types.
4102 MVT::ValueType VT = getValueType(RetTy);
4103 MVT::ValueType RegisterVT = getRegisterType(VT);
4104 unsigned NumRegs = getNumRegisters(VT);
4105 SmallVector<MVT::ValueType, 4> RetTys(NumRegs);
4106 for (unsigned i = 0; i != NumRegs; ++i)
4107 RetTys[i] = RegisterVT;
4109 RetTys.push_back(MVT::Other); // Always has a chain.
4111 // Create the CALL node.
4112 SDOperand Res = DAG.getNode(ISD::CALL,
4113 DAG.getVTList(&RetTys[0], NumRegs + 1),
4114 &Ops[0], Ops.size());
4115 Chain = Res.getValue(NumRegs);
4117 // Gather up the call result into a single value.
4118 if (RetTy != Type::VoidTy) {
4119 ISD::NodeType AssertOp = ISD::AssertSext;
4121 AssertOp = ISD::AssertZext;
4122 SmallVector<SDOperand, 4> Results(NumRegs);
4123 for (unsigned i = 0; i != NumRegs; ++i)
4124 Results[i] = Res.getValue(i);
4125 Res = getCopyFromParts(DAG, &Results[0], NumRegs, RegisterVT, VT, AssertOp);
4128 return std::make_pair(Res, Chain);
4131 SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
4132 assert(0 && "LowerOperation not implemented for this target!");
4137 SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op,
4138 SelectionDAG &DAG) {
4139 assert(0 && "CustomPromoteOperation not implemented for this target!");
4144 /// getMemsetValue - Vectorized representation of the memset value
4146 static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT,
4147 SelectionDAG &DAG) {
4148 MVT::ValueType CurVT = VT;
4149 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
4150 uint64_t Val = C->getValue() & 255;
4152 while (CurVT != MVT::i8) {
4153 Val = (Val << Shift) | Val;
4155 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
4157 return DAG.getConstant(Val, VT);
4159 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
4161 while (CurVT != MVT::i8) {
4163 DAG.getNode(ISD::OR, VT,
4164 DAG.getNode(ISD::SHL, VT, Value,
4165 DAG.getConstant(Shift, MVT::i8)), Value);
4167 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
4174 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
4175 /// used when a memcpy is turned into a memset when the source is a constant
4177 static SDOperand getMemsetStringVal(MVT::ValueType VT,
4178 SelectionDAG &DAG, TargetLowering &TLI,
4179 std::string &Str, unsigned Offset) {
4181 unsigned MSB = MVT::getSizeInBits(VT) / 8;
4182 if (TLI.isLittleEndian())
4183 Offset = Offset + MSB - 1;
4184 for (unsigned i = 0; i != MSB; ++i) {
4185 Val = (Val << 8) | (unsigned char)Str[Offset];
4186 Offset += TLI.isLittleEndian() ? -1 : 1;
4188 return DAG.getConstant(Val, VT);
4191 /// getMemBasePlusOffset - Returns base and offset node for the
4192 static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset,
4193 SelectionDAG &DAG, TargetLowering &TLI) {
4194 MVT::ValueType VT = Base.getValueType();
4195 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
4198 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
4199 /// to replace the memset / memcpy is below the threshold. It also returns the
4200 /// types of the sequence of memory ops to perform memset / memcpy.
4201 static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps,
4202 unsigned Limit, uint64_t Size,
4203 unsigned Align, TargetLowering &TLI) {
4206 if (TLI.allowsUnalignedMemoryAccesses()) {
4209 switch (Align & 7) {
4225 MVT::ValueType LVT = MVT::i64;
4226 while (!TLI.isTypeLegal(LVT))
4227 LVT = (MVT::ValueType)((unsigned)LVT - 1);
4228 assert(MVT::isInteger(LVT));
4233 unsigned NumMemOps = 0;
4235 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4236 while (VTSize > Size) {
4237 VT = (MVT::ValueType)((unsigned)VT - 1);
4240 assert(MVT::isInteger(VT));
4242 if (++NumMemOps > Limit)
4244 MemOps.push_back(VT);
4251 void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) {
4252 SDOperand Op1 = getValue(I.getOperand(1));
4253 SDOperand Op2 = getValue(I.getOperand(2));
4254 SDOperand Op3 = getValue(I.getOperand(3));
4255 SDOperand Op4 = getValue(I.getOperand(4));
4256 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue();
4257 if (Align == 0) Align = 1;
4259 // If the source and destination are known to not be aliases, we can
4260 // lower memmove as memcpy.
4261 if (Op == ISD::MEMMOVE) {
4262 uint64_t Size = -1ULL;
4263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
4264 Size = C->getValue();
4265 if (AA.alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
4266 AliasAnalysis::NoAlias)
4270 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) {
4271 std::vector<MVT::ValueType> MemOps;
4273 // Expand memset / memcpy to a series of load / store ops
4274 // if the size operand falls below a certain threshold.
4275 SmallVector<SDOperand, 8> OutChains;
4277 default: break; // Do nothing for now.
4279 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(),
4280 Size->getValue(), Align, TLI)) {
4281 unsigned NumMemOps = MemOps.size();
4282 unsigned Offset = 0;
4283 for (unsigned i = 0; i < NumMemOps; i++) {
4284 MVT::ValueType VT = MemOps[i];
4285 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4286 SDOperand Value = getMemsetValue(Op2, VT, DAG);
4287 SDOperand Store = DAG.getStore(getRoot(), Value,
4288 getMemBasePlusOffset(Op1, Offset, DAG, TLI),
4289 I.getOperand(1), Offset);
4290 OutChains.push_back(Store);
4297 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(),
4298 Size->getValue(), Align, TLI)) {
4299 unsigned NumMemOps = MemOps.size();
4300 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0;
4301 GlobalAddressSDNode *G = NULL;
4303 bool CopyFromStr = false;
4305 if (Op2.getOpcode() == ISD::GlobalAddress)
4306 G = cast<GlobalAddressSDNode>(Op2);
4307 else if (Op2.getOpcode() == ISD::ADD &&
4308 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress &&
4309 Op2.getOperand(1).getOpcode() == ISD::Constant) {
4310 G = cast<GlobalAddressSDNode>(Op2.getOperand(0));
4311 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue();
4314 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
4315 if (GV && GV->isConstant()) {
4316 Str = GV->getStringValue(false);
4324 for (unsigned i = 0; i < NumMemOps; i++) {
4325 MVT::ValueType VT = MemOps[i];
4326 unsigned VTSize = MVT::getSizeInBits(VT) / 8;
4327 SDOperand Value, Chain, Store;
4330 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
4333 DAG.getStore(Chain, Value,
4334 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
4335 I.getOperand(1), DstOff);
4337 Value = DAG.getLoad(VT, getRoot(),
4338 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI),
4339 I.getOperand(2), SrcOff);
4340 Chain = Value.getValue(1);
4342 DAG.getStore(Chain, Value,
4343 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
4344 I.getOperand(1), DstOff);
4346 OutChains.push_back(Store);
4355 if (!OutChains.empty()) {
4356 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
4357 &OutChains[0], OutChains.size()));
4362 DAG.setRoot(DAG.getNode(Op, MVT::Other, getRoot(), Op1, Op2, Op3, Op4));
4365 //===----------------------------------------------------------------------===//
4366 // SelectionDAGISel code
4367 //===----------------------------------------------------------------------===//
4369 unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) {
4370 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
4373 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
4374 AU.addRequired<AliasAnalysis>();
4375 AU.setPreservesAll();
4380 bool SelectionDAGISel::runOnFunction(Function &Fn) {
4381 // Get alias analysis for load/store combining.
4382 AA = &getAnalysis<AliasAnalysis>();
4384 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
4385 RegMap = MF.getSSARegMap();
4386 DOUT << "\n\n\n=== " << Fn.getName() << "\n";
4388 FunctionLoweringInfo FuncInfo(TLI, Fn, MF);
4390 if (ExceptionHandling)
4391 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
4392 if (InvokeInst *Invoke = dyn_cast<InvokeInst>(I->getTerminator()))
4393 // Mark landing pad.
4394 FuncInfo.MBBMap[Invoke->getSuccessor(1)]->setIsLandingPad();
4396 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
4397 SelectBasicBlock(I, MF, FuncInfo);
4399 // Add function live-ins to entry block live-in set.
4400 BasicBlock *EntryBB = &Fn.getEntryBlock();
4401 BB = FuncInfo.MBBMap[EntryBB];
4402 if (!MF.livein_empty())
4403 for (MachineFunction::livein_iterator I = MF.livein_begin(),
4404 E = MF.livein_end(); I != E; ++I)
4405 BB->addLiveIn(I->first);
4408 assert(FuncInfo.CatchInfoFound.size() == FuncInfo.CatchInfoLost.size() &&
4409 "Not all catch info was assigned to a landing pad!");
4415 SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V,
4417 SDOperand Op = getValue(V);
4418 assert((Op.getOpcode() != ISD::CopyFromReg ||
4419 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
4420 "Copy from a reg to the same reg!");
4422 MVT::ValueType SrcVT = Op.getValueType();
4423 MVT::ValueType RegisterVT = TLI.getRegisterType(SrcVT);
4424 unsigned NumRegs = TLI.getNumRegisters(SrcVT);
4425 SmallVector<SDOperand, 8> Regs(NumRegs);
4426 SmallVector<SDOperand, 8> Chains(NumRegs);
4428 // Copy the value by legal parts into sequential virtual registers.
4429 getCopyToParts(DAG, Op, &Regs[0], NumRegs, RegisterVT);
4430 for (unsigned i = 0; i != NumRegs; ++i)
4431 Chains[i] = DAG.getCopyToReg(getRoot(), Reg + i, Regs[i]);
4432 return DAG.getNode(ISD::TokenFactor, MVT::Other, &Chains[0], NumRegs);
4435 void SelectionDAGISel::
4436 LowerArguments(BasicBlock *LLVMBB, SelectionDAGLowering &SDL,
4437 std::vector<SDOperand> &UnorderedChains) {
4438 // If this is the entry block, emit arguments.
4439 Function &F = *LLVMBB->getParent();
4440 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo;
4441 SDOperand OldRoot = SDL.DAG.getRoot();
4442 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG);
4445 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
4447 if (!AI->use_empty()) {
4448 SDL.setValue(AI, Args[a]);
4450 // If this argument is live outside of the entry block, insert a copy from
4451 // whereever we got it to the vreg that other BB's will reference it as.
4452 DenseMap<const Value*, unsigned>::iterator VMI=FuncInfo.ValueMap.find(AI);
4453 if (VMI != FuncInfo.ValueMap.end()) {
4454 SDOperand Copy = SDL.CopyValueToVirtualRegister(AI, VMI->second);
4455 UnorderedChains.push_back(Copy);
4459 // Finally, if the target has anything special to do, allow it to do so.
4460 // FIXME: this should insert code into the DAG!
4461 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction());
4464 static void copyCatchInfo(BasicBlock *SrcBB, BasicBlock *DestBB,
4465 MachineModuleInfo *MMI, FunctionLoweringInfo &FLI) {
4466 assert(!FLI.MBBMap[SrcBB]->isLandingPad() &&
4467 "Copying catch info out of a landing pad!");
4468 for (BasicBlock::iterator I = SrcBB->begin(), E = --SrcBB->end(); I != E; ++I)
4469 if (isSelector(I)) {
4470 // Apply the catch info to DestBB.
4471 addCatchInfo(cast<CallInst>(*I), MMI, FLI.MBBMap[DestBB]);
4473 FLI.CatchInfoFound.insert(I);
4478 /// CheckDAGForTailCallsAndFixThem - This Function looks for CALL nodes in the
4479 /// DAG and fixes their tailcall attribute operand
4480 static void CheckDAGForTailCallsAndFixThem(SelectionDAG &DAG,
4481 TargetLowering& TLI) {
4482 SDNode * Ret = NULL;
4483 SDOperand Terminator = DAG.getRoot();
4486 if (Terminator.getOpcode() == ISD::RET) {
4487 Ret = Terminator.Val;
4490 // Fix tail call attribute of CALL nodes.
4491 for (SelectionDAG::allnodes_iterator BE = DAG.allnodes_begin(),
4492 BI = prior(DAG.allnodes_end()); BI != BE; --BI) {
4493 if (BI->getOpcode() == ISD::CALL) {
4494 SDOperand OpRet(Ret, 0);
4495 SDOperand OpCall(static_cast<SDNode*>(BI), 0);
4496 bool isMarkedTailCall =
4497 cast<ConstantSDNode>(OpCall.getOperand(3))->getValue() != 0;
4498 // If CALL node has tail call attribute set to true and the call is not
4499 // eligible (no RET or the target rejects) the attribute is fixed to
4500 // false. The TargetLowering::IsEligibleForTailCallOptimization function
4501 // must correctly identify tail call optimizable calls.
4502 if (isMarkedTailCall &&
4504 !TLI.IsEligibleForTailCallOptimization(OpCall, OpRet, DAG))) {
4505 SmallVector<SDOperand, 32> Ops;
4507 for(SDNode::op_iterator I =OpCall.Val->op_begin(),
4508 E=OpCall.Val->op_end(); I!=E; I++, idx++) {
4512 Ops.push_back(DAG.getConstant(false, TLI.getPointerTy()));
4514 DAG.UpdateNodeOperands(OpCall, Ops.begin(), Ops.size());
4520 void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB,
4521 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate,
4522 FunctionLoweringInfo &FuncInfo) {
4523 SelectionDAGLowering SDL(DAG, TLI, *AA, FuncInfo);
4525 std::vector<SDOperand> UnorderedChains;
4527 // Lower any arguments needed in this block if this is the entry block.
4528 if (LLVMBB == &LLVMBB->getParent()->getEntryBlock())
4529 LowerArguments(LLVMBB, SDL, UnorderedChains);
4531 BB = FuncInfo.MBBMap[LLVMBB];
4532 SDL.setCurrentBasicBlock(BB);
4534 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4536 if (ExceptionHandling && MMI && BB->isLandingPad()) {
4537 // Add a label to mark the beginning of the landing pad. Deletion of the
4538 // landing pad can thus be detected via the MachineModuleInfo.
4539 unsigned LabelID = MMI->addLandingPad(BB);
4540 DAG.setRoot(DAG.getNode(ISD::LABEL, MVT::Other, DAG.getEntryNode(),
4541 DAG.getConstant(LabelID, MVT::i32)));
4543 // Mark exception register as live in.
4544 unsigned Reg = TLI.getExceptionAddressRegister();
4545 if (Reg) BB->addLiveIn(Reg);
4547 // Mark exception selector register as live in.
4548 Reg = TLI.getExceptionSelectorRegister();
4549 if (Reg) BB->addLiveIn(Reg);
4551 // FIXME: Hack around an exception handling flaw (PR1508): the personality
4552 // function and list of typeids logically belong to the invoke (or, if you
4553 // like, the basic block containing the invoke), and need to be associated
4554 // with it in the dwarf exception handling tables. Currently however the
4555 // information is provided by an intrinsic (eh.selector) that can be moved
4556 // to unexpected places by the optimizers: if the unwind edge is critical,
4557 // then breaking it can result in the intrinsics being in the successor of
4558 // the landing pad, not the landing pad itself. This results in exceptions
4559 // not being caught because no typeids are associated with the invoke.
4560 // This may not be the only way things can go wrong, but it is the only way
4561 // we try to work around for the moment.
4562 BranchInst *Br = dyn_cast<BranchInst>(LLVMBB->getTerminator());
4564 if (Br && Br->isUnconditional()) { // Critical edge?
4565 BasicBlock::iterator I, E;
4566 for (I = LLVMBB->begin(), E = --LLVMBB->end(); I != E; ++I)
4571 // No catch info found - try to extract some from the successor.
4572 copyCatchInfo(Br->getSuccessor(0), LLVMBB, MMI, FuncInfo);
4576 // Lower all of the non-terminator instructions.
4577 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end();
4581 // Ensure that all instructions which are used outside of their defining
4582 // blocks are available as virtual registers. Invoke is handled elsewhere.
4583 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I)
4584 if (!I->use_empty() && !isa<PHINode>(I) && !isa<InvokeInst>(I)) {
4585 DenseMap<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I);
4586 if (VMI != FuncInfo.ValueMap.end())
4587 UnorderedChains.push_back(
4588 SDL.CopyValueToVirtualRegister(I, VMI->second));
4591 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
4592 // ensure constants are generated when needed. Remember the virtual registers
4593 // that need to be added to the Machine PHI nodes as input. We cannot just
4594 // directly add them, because expansion might result in multiple MBB's for one
4595 // BB. As such, the start of the BB might correspond to a different MBB than
4598 TerminatorInst *TI = LLVMBB->getTerminator();
4600 // Emit constants only once even if used by multiple PHI nodes.
4601 std::map<Constant*, unsigned> ConstantsOut;
4603 // Vector bool would be better, but vector<bool> is really slow.
4604 std::vector<unsigned char> SuccsHandled;
4605 if (TI->getNumSuccessors())
4606 SuccsHandled.resize(BB->getParent()->getNumBlockIDs());
4608 // Check successor nodes' PHI nodes that expect a constant to be available
4610 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
4611 BasicBlock *SuccBB = TI->getSuccessor(succ);
4612 if (!isa<PHINode>(SuccBB->begin())) continue;
4613 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
4615 // If this terminator has multiple identical successors (common for
4616 // switches), only handle each succ once.
4617 unsigned SuccMBBNo = SuccMBB->getNumber();
4618 if (SuccsHandled[SuccMBBNo]) continue;
4619 SuccsHandled[SuccMBBNo] = true;
4621 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
4624 // At this point we know that there is a 1-1 correspondence between LLVM PHI
4625 // nodes and Machine PHI nodes, but the incoming operands have not been
4627 for (BasicBlock::iterator I = SuccBB->begin();
4628 (PN = dyn_cast<PHINode>(I)); ++I) {
4629 // Ignore dead phi's.
4630 if (PN->use_empty()) continue;
4633 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
4635 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
4636 unsigned &RegOut = ConstantsOut[C];
4638 RegOut = FuncInfo.CreateRegForValue(C);
4639 UnorderedChains.push_back(
4640 SDL.CopyValueToVirtualRegister(C, RegOut));
4644 Reg = FuncInfo.ValueMap[PHIOp];
4646 assert(isa<AllocaInst>(PHIOp) &&
4647 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
4648 "Didn't codegen value into a register!??");
4649 Reg = FuncInfo.CreateRegForValue(PHIOp);
4650 UnorderedChains.push_back(
4651 SDL.CopyValueToVirtualRegister(PHIOp, Reg));
4655 // Remember that this register needs to added to the machine PHI node as
4656 // the input for this MBB.
4657 MVT::ValueType VT = TLI.getValueType(PN->getType());
4658 unsigned NumRegisters = TLI.getNumRegisters(VT);
4659 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
4660 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
4663 ConstantsOut.clear();
4665 // Turn all of the unordered chains into one factored node.
4666 if (!UnorderedChains.empty()) {
4667 SDOperand Root = SDL.getRoot();
4668 if (Root.getOpcode() != ISD::EntryToken) {
4669 unsigned i = 0, e = UnorderedChains.size();
4670 for (; i != e; ++i) {
4671 assert(UnorderedChains[i].Val->getNumOperands() > 1);
4672 if (UnorderedChains[i].Val->getOperand(0) == Root)
4673 break; // Don't add the root if we already indirectly depend on it.
4677 UnorderedChains.push_back(Root);
4679 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
4680 &UnorderedChains[0], UnorderedChains.size()));
4683 // Lower the terminator after the copies are emitted.
4684 SDL.visit(*LLVMBB->getTerminator());
4686 // Copy over any CaseBlock records that may now exist due to SwitchInst
4687 // lowering, as well as any jump table information.
4688 SwitchCases.clear();
4689 SwitchCases = SDL.SwitchCases;
4691 JTCases = SDL.JTCases;
4692 BitTestCases.clear();
4693 BitTestCases = SDL.BitTestCases;
4695 // Make sure the root of the DAG is up-to-date.
4696 DAG.setRoot(SDL.getRoot());
4698 // Check whether calls in this block are real tail calls. Fix up CALL nodes
4699 // with correct tailcall attribute so that the target can rely on the tailcall
4700 // attribute indicating whether the call is really eligible for tail call
4702 CheckDAGForTailCallsAndFixThem(DAG, TLI);
4705 void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) {
4706 DOUT << "Lowered selection DAG:\n";
4709 // Run the DAG combiner in pre-legalize mode.
4710 DAG.Combine(false, *AA);
4712 DOUT << "Optimized lowered selection DAG:\n";
4715 // Second step, hack on the DAG until it only uses operations and types that
4716 // the target supports.
4719 DOUT << "Legalized selection DAG:\n";
4722 // Run the DAG combiner in post-legalize mode.
4723 DAG.Combine(true, *AA);
4725 DOUT << "Optimized legalized selection DAG:\n";
4728 if (ViewISelDAGs) DAG.viewGraph();
4730 // Third, instruction select all of the operations to machine code, adding the
4731 // code to the MachineBasicBlock.
4732 InstructionSelectBasicBlock(DAG);
4734 DOUT << "Selected machine code:\n";
4738 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF,
4739 FunctionLoweringInfo &FuncInfo) {
4740 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate;
4742 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4745 // First step, lower LLVM code to some DAG. This DAG may use operations and
4746 // types that are not supported by the target.
4747 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo);
4749 // Second step, emit the lowered DAG as machine code.
4750 CodeGenAndEmitDAG(DAG);
4753 DOUT << "Total amount of phi nodes to update: "
4754 << PHINodesToUpdate.size() << "\n";
4755 DEBUG(for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i)
4756 DOUT << "Node " << i << " : (" << PHINodesToUpdate[i].first
4757 << ", " << PHINodesToUpdate[i].second << ")\n";);
4759 // Next, now that we know what the last MBB the LLVM BB expanded is, update
4760 // PHI nodes in successors.
4761 if (SwitchCases.empty() && JTCases.empty() && BitTestCases.empty()) {
4762 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
4763 MachineInstr *PHI = PHINodesToUpdate[i].first;
4764 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4765 "This is not a machine PHI node that we are updating!");
4766 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
4767 PHI->addMachineBasicBlockOperand(BB);
4772 for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i) {
4773 // Lower header first, if it wasn't already lowered
4774 if (!BitTestCases[i].Emitted) {
4775 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4777 SelectionDAGLowering HSDL(HSDAG, TLI, *AA, FuncInfo);
4778 // Set the current basic block to the mbb we wish to insert the code into
4779 BB = BitTestCases[i].Parent;
4780 HSDL.setCurrentBasicBlock(BB);
4782 HSDL.visitBitTestHeader(BitTestCases[i]);
4783 HSDAG.setRoot(HSDL.getRoot());
4784 CodeGenAndEmitDAG(HSDAG);
4787 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) {
4788 SelectionDAG BSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4790 SelectionDAGLowering BSDL(BSDAG, TLI, *AA, FuncInfo);
4791 // Set the current basic block to the mbb we wish to insert the code into
4792 BB = BitTestCases[i].Cases[j].ThisBB;
4793 BSDL.setCurrentBasicBlock(BB);
4796 BSDL.visitBitTestCase(BitTestCases[i].Cases[j+1].ThisBB,
4797 BitTestCases[i].Reg,
4798 BitTestCases[i].Cases[j]);
4800 BSDL.visitBitTestCase(BitTestCases[i].Default,
4801 BitTestCases[i].Reg,
4802 BitTestCases[i].Cases[j]);
4805 BSDAG.setRoot(BSDL.getRoot());
4806 CodeGenAndEmitDAG(BSDAG);
4810 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
4811 MachineInstr *PHI = PHINodesToUpdate[pi].first;
4812 MachineBasicBlock *PHIBB = PHI->getParent();
4813 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4814 "This is not a machine PHI node that we are updating!");
4815 // This is "default" BB. We have two jumps to it. From "header" BB and
4816 // from last "case" BB.
4817 if (PHIBB == BitTestCases[i].Default) {
4818 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
4819 PHI->addMachineBasicBlockOperand(BitTestCases[i].Parent);
4820 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
4821 PHI->addMachineBasicBlockOperand(BitTestCases[i].Cases.back().ThisBB);
4823 // One of "cases" BB.
4824 for (unsigned j = 0, ej = BitTestCases[i].Cases.size(); j != ej; ++j) {
4825 MachineBasicBlock* cBB = BitTestCases[i].Cases[j].ThisBB;
4826 if (cBB->succ_end() !=
4827 std::find(cBB->succ_begin(),cBB->succ_end(), PHIBB)) {
4828 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
4829 PHI->addMachineBasicBlockOperand(cBB);
4835 // If the JumpTable record is filled in, then we need to emit a jump table.
4836 // Updating the PHI nodes is tricky in this case, since we need to determine
4837 // whether the PHI is a successor of the range check MBB or the jump table MBB
4838 for (unsigned i = 0, e = JTCases.size(); i != e; ++i) {
4839 // Lower header first, if it wasn't already lowered
4840 if (!JTCases[i].first.Emitted) {
4841 SelectionDAG HSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4843 SelectionDAGLowering HSDL(HSDAG, TLI, *AA, FuncInfo);
4844 // Set the current basic block to the mbb we wish to insert the code into
4845 BB = JTCases[i].first.HeaderBB;
4846 HSDL.setCurrentBasicBlock(BB);
4848 HSDL.visitJumpTableHeader(JTCases[i].second, JTCases[i].first);
4849 HSDAG.setRoot(HSDL.getRoot());
4850 CodeGenAndEmitDAG(HSDAG);
4853 SelectionDAG JSDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4855 SelectionDAGLowering JSDL(JSDAG, TLI, *AA, FuncInfo);
4856 // Set the current basic block to the mbb we wish to insert the code into
4857 BB = JTCases[i].second.MBB;
4858 JSDL.setCurrentBasicBlock(BB);
4860 JSDL.visitJumpTable(JTCases[i].second);
4861 JSDAG.setRoot(JSDL.getRoot());
4862 CodeGenAndEmitDAG(JSDAG);
4865 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
4866 MachineInstr *PHI = PHINodesToUpdate[pi].first;
4867 MachineBasicBlock *PHIBB = PHI->getParent();
4868 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4869 "This is not a machine PHI node that we are updating!");
4870 // "default" BB. We can go there only from header BB.
4871 if (PHIBB == JTCases[i].second.Default) {
4872 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
4873 PHI->addMachineBasicBlockOperand(JTCases[i].first.HeaderBB);
4875 // JT BB. Just iterate over successors here
4876 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
4877 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
4878 PHI->addMachineBasicBlockOperand(BB);
4883 // If the switch block involved a branch to one of the actual successors, we
4884 // need to update PHI nodes in that block.
4885 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
4886 MachineInstr *PHI = PHINodesToUpdate[i].first;
4887 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
4888 "This is not a machine PHI node that we are updating!");
4889 if (BB->isSuccessor(PHI->getParent())) {
4890 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
4891 PHI->addMachineBasicBlockOperand(BB);
4895 // If we generated any switch lowering information, build and codegen any
4896 // additional DAGs necessary.
4897 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) {
4898 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineModuleInfo>());
4900 SelectionDAGLowering SDL(SDAG, TLI, *AA, FuncInfo);
4902 // Set the current basic block to the mbb we wish to insert the code into
4903 BB = SwitchCases[i].ThisBB;
4904 SDL.setCurrentBasicBlock(BB);
4907 SDL.visitSwitchCase(SwitchCases[i]);
4908 SDAG.setRoot(SDL.getRoot());
4909 CodeGenAndEmitDAG(SDAG);
4911 // Handle any PHI nodes in successors of this chunk, as if we were coming
4912 // from the original BB before switch expansion. Note that PHI nodes can
4913 // occur multiple times in PHINodesToUpdate. We have to be very careful to
4914 // handle them the right number of times.
4915 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
4916 for (MachineBasicBlock::iterator Phi = BB->begin();
4917 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
4918 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
4919 for (unsigned pn = 0; ; ++pn) {
4920 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!");
4921 if (PHINodesToUpdate[pn].first == Phi) {
4922 Phi->addRegOperand(PHINodesToUpdate[pn].second, false);
4923 Phi->addMachineBasicBlockOperand(SwitchCases[i].ThisBB);
4929 // Don't process RHS if same block as LHS.
4930 if (BB == SwitchCases[i].FalseBB)
4931 SwitchCases[i].FalseBB = 0;
4933 // If we haven't handled the RHS, do so now. Otherwise, we're done.
4934 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB;
4935 SwitchCases[i].FalseBB = 0;
4937 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0);
4942 //===----------------------------------------------------------------------===//
4943 /// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
4944 /// target node in the graph.
4945 void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) {
4946 if (ViewSchedDAGs) DAG.viewGraph();
4948 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
4952 RegisterScheduler::setDefault(Ctor);
4955 ScheduleDAG *SL = Ctor(this, &DAG, BB);
4958 if (ViewSUnitDAGs) SL->viewGraph();
4964 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
4965 return new HazardRecognizer();
4968 //===----------------------------------------------------------------------===//
4969 // Helper functions used by the generated instruction selector.
4970 //===----------------------------------------------------------------------===//
4971 // Calls to these methods are generated by tblgen.
4973 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
4974 /// the dag combiner simplified the 255, we still want to match. RHS is the
4975 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
4976 /// specified in the .td file (e.g. 255).
4977 bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS,
4978 int64_t DesiredMaskS) const {
4979 uint64_t ActualMask = RHS->getValue();
4980 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
4982 // If the actual mask exactly matches, success!
4983 if (ActualMask == DesiredMask)
4986 // If the actual AND mask is allowing unallowed bits, this doesn't match.
4987 if (ActualMask & ~DesiredMask)
4990 // Otherwise, the DAG Combiner may have proven that the value coming in is
4991 // either already zero or is not demanded. Check for known zero input bits.
4992 uint64_t NeededMask = DesiredMask & ~ActualMask;
4993 if (CurDAG->MaskedValueIsZero(LHS, NeededMask))
4996 // TODO: check to see if missing bits are just not demanded.
4998 // Otherwise, this pattern doesn't match.
5002 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
5003 /// the dag combiner simplified the 255, we still want to match. RHS is the
5004 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
5005 /// specified in the .td file (e.g. 255).
5006 bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS,
5007 int64_t DesiredMaskS) const {
5008 uint64_t ActualMask = RHS->getValue();
5009 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
5011 // If the actual mask exactly matches, success!
5012 if (ActualMask == DesiredMask)
5015 // If the actual AND mask is allowing unallowed bits, this doesn't match.
5016 if (ActualMask & ~DesiredMask)
5019 // Otherwise, the DAG Combiner may have proven that the value coming in is
5020 // either already zero or is not demanded. Check for known zero input bits.
5021 uint64_t NeededMask = DesiredMask & ~ActualMask;
5023 uint64_t KnownZero, KnownOne;
5024 CurDAG->ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
5026 // If all the missing bits in the or are already known to be set, match!
5027 if ((NeededMask & KnownOne) == NeededMask)
5030 // TODO: check to see if missing bits are just not demanded.
5032 // Otherwise, this pattern doesn't match.
5037 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
5038 /// by tblgen. Others should not call it.
5039 void SelectionDAGISel::
5040 SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) {
5041 std::vector<SDOperand> InOps;
5042 std::swap(InOps, Ops);
5044 Ops.push_back(InOps[0]); // input chain.
5045 Ops.push_back(InOps[1]); // input asm string.
5047 unsigned i = 2, e = InOps.size();
5048 if (InOps[e-1].getValueType() == MVT::Flag)
5049 --e; // Don't process a flag operand if it is here.
5052 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
5053 if ((Flags & 7) != 4 /*MEM*/) {
5054 // Just skip over this operand, copying the operands verbatim.
5055 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
5056 i += (Flags >> 3) + 1;
5058 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
5059 // Otherwise, this is a memory operand. Ask the target to select it.
5060 std::vector<SDOperand> SelOps;
5061 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) {
5062 cerr << "Could not match memory address. Inline asm failure!\n";
5066 // Add this to the output node.
5067 MVT::ValueType IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
5068 Ops.push_back(DAG.getTargetConstant(4/*MEM*/ | (SelOps.size() << 3),
5070 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
5075 // Add the flag input back if present.
5076 if (e != InOps.size())
5077 Ops.push_back(InOps.back());
5080 char SelectionDAGISel::ID = 0;