1 //===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements the SelectionDAGISel class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "llvm/Analysis/AliasAnalysis.h"
16 #include "llvm/CodeGen/SelectionDAGISel.h"
17 #include "llvm/CodeGen/ScheduleDAG.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/Constants.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Function.h"
22 #include "llvm/GlobalVariable.h"
23 #include "llvm/InlineAsm.h"
24 #include "llvm/Instructions.h"
25 #include "llvm/Intrinsics.h"
26 #include "llvm/IntrinsicInst.h"
27 #include "llvm/CodeGen/IntrinsicLowering.h"
28 #include "llvm/CodeGen/MachineDebugInfo.h"
29 #include "llvm/CodeGen/MachineFunction.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineJumpTableInfo.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/SchedulerRegistry.h"
34 #include "llvm/CodeGen/SelectionDAG.h"
35 #include "llvm/CodeGen/SSARegMap.h"
36 #include "llvm/Target/MRegisterInfo.h"
37 #include "llvm/Target/TargetData.h"
38 #include "llvm/Target/TargetFrameInfo.h"
39 #include "llvm/Target/TargetInstrInfo.h"
40 #include "llvm/Target/TargetLowering.h"
41 #include "llvm/Target/TargetMachine.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/Compiler.h"
53 ViewISelDAGs("view-isel-dags", cl::Hidden,
54 cl::desc("Pop up a window to show isel dags as they are selected"));
56 ViewSchedDAGs("view-sched-dags", cl::Hidden,
57 cl::desc("Pop up a window to show sched dags as they are processed"));
59 static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0;
63 //===---------------------------------------------------------------------===//
65 /// RegisterScheduler class - Track the registration of instruction schedulers.
67 //===---------------------------------------------------------------------===//
68 MachinePassRegistry RegisterScheduler::Registry;
70 //===---------------------------------------------------------------------===//
72 /// ISHeuristic command line option for instruction schedulers.
74 //===---------------------------------------------------------------------===//
76 cl::opt<RegisterScheduler::FunctionPassCtor, false,
77 RegisterPassParser<RegisterScheduler> >
79 cl::init(&createDefaultScheduler),
80 cl::desc("Instruction schedulers available:"));
82 static RegisterScheduler
83 defaultListDAGScheduler("default", " Best scheduler for the target",
84 createDefaultScheduler);
88 /// RegsForValue - This struct represents the physical registers that a
89 /// particular value is assigned and the type information about the value.
90 /// This is needed because values can be promoted into larger registers and
91 /// expanded into multiple smaller registers than the value.
92 struct VISIBILITY_HIDDEN RegsForValue {
93 /// Regs - This list hold the register (for legal and promoted values)
94 /// or register set (for expanded values) that the value should be assigned
96 std::vector<unsigned> Regs;
98 /// RegVT - The value type of each register.
100 MVT::ValueType RegVT;
102 /// ValueVT - The value type of the LLVM value, which may be promoted from
103 /// RegVT or made from merging the two expanded parts.
104 MVT::ValueType ValueVT;
106 RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {}
108 RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
109 : RegVT(regvt), ValueVT(valuevt) {
112 RegsForValue(const std::vector<unsigned> ®s,
113 MVT::ValueType regvt, MVT::ValueType valuevt)
114 : Regs(regs), RegVT(regvt), ValueVT(valuevt) {
117 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
118 /// this value and returns the result as a ValueVT value. This uses
119 /// Chain/Flag as the input and updates them for the output Chain/Flag.
120 SDOperand getCopyFromRegs(SelectionDAG &DAG,
121 SDOperand &Chain, SDOperand &Flag) const;
123 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
124 /// specified value into the registers specified by this object. This uses
125 /// Chain/Flag as the input and updates them for the output Chain/Flag.
126 void getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
127 SDOperand &Chain, SDOperand &Flag,
128 MVT::ValueType PtrVT) const;
130 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
131 /// operand list. This adds the code marker and includes the number of
132 /// values added into it.
133 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
134 std::vector<SDOperand> &Ops) const;
139 //===--------------------------------------------------------------------===//
140 /// createDefaultScheduler - This creates an instruction scheduler appropriate
142 ScheduleDAG* createDefaultScheduler(SelectionDAGISel *IS,
144 MachineBasicBlock *BB) {
145 TargetLowering &TLI = IS->getTargetLowering();
147 if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency) {
148 return createTDListDAGScheduler(IS, DAG, BB);
150 assert(TLI.getSchedulingPreference() ==
151 TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
152 return createBURRListDAGScheduler(IS, DAG, BB);
157 //===--------------------------------------------------------------------===//
158 /// FunctionLoweringInfo - This contains information that is global to a
159 /// function that is used when lowering a region of the function.
160 class FunctionLoweringInfo {
167 FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF);
169 /// MBBMap - A mapping from LLVM basic blocks to their machine code entry.
170 std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
172 /// ValueMap - Since we emit code for the function a basic block at a time,
173 /// we must remember which virtual registers hold the values for
174 /// cross-basic-block values.
175 std::map<const Value*, unsigned> ValueMap;
177 /// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in
178 /// the entry block. This allows the allocas to be efficiently referenced
179 /// anywhere in the function.
180 std::map<const AllocaInst*, int> StaticAllocaMap;
182 unsigned MakeReg(MVT::ValueType VT) {
183 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
186 /// isExportedInst - Return true if the specified value is an instruction
187 /// exported from its block.
188 bool isExportedInst(const Value *V) {
189 return ValueMap.count(V);
192 unsigned CreateRegForValue(const Value *V);
194 unsigned InitializeRegForValue(const Value *V) {
195 unsigned &R = ValueMap[V];
196 assert(R == 0 && "Already initialized this value register!");
197 return R = CreateRegForValue(V);
202 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
203 /// PHI nodes or outside of the basic block that defines it, or used by a
204 /// switch instruction, which may expand to multiple basic blocks.
205 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
206 if (isa<PHINode>(I)) return true;
207 BasicBlock *BB = I->getParent();
208 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
209 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
210 // FIXME: Remove switchinst special case.
211 isa<SwitchInst>(*UI))
216 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
217 /// entry block, return true. This includes arguments used by switches, since
218 /// the switch may expand into multiple basic blocks.
219 static bool isOnlyUsedInEntryBlock(Argument *A) {
220 BasicBlock *Entry = A->getParent()->begin();
221 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
222 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
223 return false; // Use not in entry block.
227 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli,
228 Function &fn, MachineFunction &mf)
229 : TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) {
231 // Create a vreg for each argument register that is not dead and is used
232 // outside of the entry block for the function.
233 for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end();
235 if (!isOnlyUsedInEntryBlock(AI))
236 InitializeRegForValue(AI);
238 // Initialize the mapping of values to registers. This is only set up for
239 // instruction values that are used outside of the block that defines
241 Function::iterator BB = Fn.begin(), EB = Fn.end();
242 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
243 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
244 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
245 const Type *Ty = AI->getAllocatedType();
246 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
248 std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
251 // If the alignment of the value is smaller than the size of the
252 // value, and if the size of the value is particularly small
253 // (<= 8 bytes), round up to the size of the value for potentially
254 // better performance.
256 // FIXME: This could be made better with a preferred alignment hook in
257 // TargetData. It serves primarily to 8-byte align doubles for X86.
258 if (Align < TySize && TySize <= 8) Align = TySize;
259 TySize *= CUI->getZExtValue(); // Get total allocated size.
260 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
261 StaticAllocaMap[AI] =
262 MF.getFrameInfo()->CreateStackObject((unsigned)TySize, Align);
265 for (; BB != EB; ++BB)
266 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
267 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
268 if (!isa<AllocaInst>(I) ||
269 !StaticAllocaMap.count(cast<AllocaInst>(I)))
270 InitializeRegForValue(I);
272 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
273 // also creates the initial PHI MachineInstrs, though none of the input
274 // operands are populated.
275 for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) {
276 MachineBasicBlock *MBB = new MachineBasicBlock(BB);
278 MF.getBasicBlockList().push_back(MBB);
280 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
283 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){
284 if (PN->use_empty()) continue;
286 MVT::ValueType VT = TLI.getValueType(PN->getType());
287 unsigned NumElements;
288 if (VT != MVT::Vector)
289 NumElements = TLI.getNumElements(VT);
291 MVT::ValueType VT1,VT2;
293 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
296 unsigned PHIReg = ValueMap[PN];
297 assert(PHIReg && "PHI node does not have an assigned virtual register!");
298 for (unsigned i = 0; i != NumElements; ++i)
299 BuildMI(MBB, TargetInstrInfo::PHI, PN->getNumOperands(), PHIReg+i);
304 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
305 /// the correctly promoted or expanded types. Assign these registers
306 /// consecutive vreg numbers and return the first assigned number.
307 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
308 MVT::ValueType VT = TLI.getValueType(V->getType());
310 // The number of multiples of registers that we need, to, e.g., split up
311 // a <2 x int64> -> 4 x i32 registers.
312 unsigned NumVectorRegs = 1;
314 // If this is a packed type, figure out what type it will decompose into
315 // and how many of the elements it will use.
316 if (VT == MVT::Vector) {
317 const PackedType *PTy = cast<PackedType>(V->getType());
318 unsigned NumElts = PTy->getNumElements();
319 MVT::ValueType EltTy = TLI.getValueType(PTy->getElementType());
321 // Divide the input until we get to a supported size. This will always
322 // end with a scalar if the target doesn't support vectors.
323 while (NumElts > 1 && !TLI.isTypeLegal(getVectorType(EltTy, NumElts))) {
330 VT = getVectorType(EltTy, NumElts);
333 // The common case is that we will only create one register for this
334 // value. If we have that case, create and return the virtual register.
335 unsigned NV = TLI.getNumElements(VT);
337 // If we are promoting this value, pick the next largest supported type.
338 MVT::ValueType PromotedType = TLI.getTypeToTransformTo(VT);
339 unsigned Reg = MakeReg(PromotedType);
340 // If this is a vector of supported or promoted types (e.g. 4 x i16),
341 // create all of the registers.
342 for (unsigned i = 1; i != NumVectorRegs; ++i)
343 MakeReg(PromotedType);
347 // If this value is represented with multiple target registers, make sure
348 // to create enough consecutive registers of the right (smaller) type.
349 unsigned NT = VT-1; // Find the type to use.
350 while (TLI.getNumElements((MVT::ValueType)NT) != 1)
353 unsigned R = MakeReg((MVT::ValueType)NT);
354 for (unsigned i = 1; i != NV*NumVectorRegs; ++i)
355 MakeReg((MVT::ValueType)NT);
359 //===----------------------------------------------------------------------===//
360 /// SelectionDAGLowering - This is the common target-independent lowering
361 /// implementation that is parameterized by a TargetLowering object.
362 /// Also, targets can overload any lowering method.
365 class SelectionDAGLowering {
366 MachineBasicBlock *CurMBB;
368 std::map<const Value*, SDOperand> NodeMap;
370 /// PendingLoads - Loads are not emitted to the program immediately. We bunch
371 /// them up and then emit token factor nodes when possible. This allows us to
372 /// get simple disambiguation between loads without worrying about alias
374 std::vector<SDOperand> PendingLoads;
376 /// Case - A pair of values to record the Value for a switch case, and the
377 /// case's target basic block.
378 typedef std::pair<Constant*, MachineBasicBlock*> Case;
379 typedef std::vector<Case>::iterator CaseItr;
380 typedef std::pair<CaseItr, CaseItr> CaseRange;
382 /// CaseRec - A struct with ctor used in lowering switches to a binary tree
383 /// of conditional branches.
385 CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) :
386 CaseBB(bb), LT(lt), GE(ge), Range(r) {}
388 /// CaseBB - The MBB in which to emit the compare and branch
389 MachineBasicBlock *CaseBB;
390 /// LT, GE - If nonzero, we know the current case value must be less-than or
391 /// greater-than-or-equal-to these Constants.
394 /// Range - A pair of iterators representing the range of case values to be
395 /// processed at this point in the binary search tree.
399 /// The comparison function for sorting Case values.
401 bool operator () (const Case& C1, const Case& C2) {
402 if (const ConstantInt* I1 = dyn_cast<const ConstantInt>(C1.first))
403 if (I1->getType()->isUnsigned())
404 return I1->getZExtValue() <
405 cast<const ConstantInt>(C2.first)->getZExtValue();
407 return cast<const ConstantInt>(C1.first)->getSExtValue() <
408 cast<const ConstantInt>(C2.first)->getSExtValue();
413 // TLI - This is information that describes the available target features we
414 // need for lowering. This indicates when operations are unavailable,
415 // implemented with a libcall, etc.
418 const TargetData *TD;
420 /// SwitchCases - Vector of CaseBlock structures used to communicate
421 /// SwitchInst code generation information.
422 std::vector<SelectionDAGISel::CaseBlock> SwitchCases;
423 SelectionDAGISel::JumpTable JT;
425 /// FuncInfo - Information about the function as a whole.
427 FunctionLoweringInfo &FuncInfo;
429 SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli,
430 FunctionLoweringInfo &funcinfo)
431 : TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()),
432 JT(0,0,0,0), FuncInfo(funcinfo) {
435 /// getRoot - Return the current virtual root of the Selection DAG.
437 SDOperand getRoot() {
438 if (PendingLoads.empty())
439 return DAG.getRoot();
441 if (PendingLoads.size() == 1) {
442 SDOperand Root = PendingLoads[0];
444 PendingLoads.clear();
448 // Otherwise, we have to make a token factor node.
449 SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
450 &PendingLoads[0], PendingLoads.size());
451 PendingLoads.clear();
456 SDOperand CopyValueToVirtualRegister(Value *V, unsigned Reg);
458 void visit(Instruction &I) { visit(I.getOpcode(), I); }
460 void visit(unsigned Opcode, User &I) {
462 default: assert(0 && "Unknown instruction type encountered!");
464 // Build the switch statement using the Instruction.def file.
465 #define HANDLE_INST(NUM, OPCODE, CLASS) \
466 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
467 #include "llvm/Instruction.def"
471 void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; }
473 SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr,
474 const Value *SV, SDOperand Root,
477 SDOperand getIntPtrConstant(uint64_t Val) {
478 return DAG.getConstant(Val, TLI.getPointerTy());
481 SDOperand getValue(const Value *V);
483 const SDOperand &setValue(const Value *V, SDOperand NewN) {
484 SDOperand &N = NodeMap[V];
485 assert(N.Val == 0 && "Already set a value for this node!");
489 RegsForValue GetRegistersForValue(const std::string &ConstrCode,
491 bool OutReg, bool InReg,
492 std::set<unsigned> &OutputRegs,
493 std::set<unsigned> &InputRegs);
495 void FindMergedConditions(Value *Cond, MachineBasicBlock *TBB,
496 MachineBasicBlock *FBB, MachineBasicBlock *CurBB,
498 bool isExportableFromCurrentBlock(Value *V, const BasicBlock *FromBB);
499 void ExportFromCurrentBlock(Value *V);
501 // Terminator instructions.
502 void visitRet(ReturnInst &I);
503 void visitBr(BranchInst &I);
504 void visitSwitch(SwitchInst &I);
505 void visitUnreachable(UnreachableInst &I) { /* noop */ }
507 // Helper for visitSwitch
508 void visitSwitchCase(SelectionDAGISel::CaseBlock &CB);
509 void visitJumpTable(SelectionDAGISel::JumpTable &JT);
511 // These all get lowered before this pass.
512 void visitInvoke(InvokeInst &I) { assert(0 && "TODO"); }
513 void visitUnwind(UnwindInst &I) { assert(0 && "TODO"); }
515 void visitIntBinary(User &I, unsigned IntOp, unsigned VecOp);
516 void visitFPBinary(User &I, unsigned FPOp, unsigned VecOp);
517 void visitShift(User &I, unsigned Opcode);
518 void visitAdd(User &I) {
519 if (I.getType()->isFloatingPoint())
520 visitFPBinary(I, ISD::FADD, ISD::VADD);
522 visitIntBinary(I, ISD::ADD, ISD::VADD);
524 void visitSub(User &I);
525 void visitMul(User &I) {
526 if (I.getType()->isFloatingPoint())
527 visitFPBinary(I, ISD::FMUL, ISD::VMUL);
529 visitIntBinary(I, ISD::MUL, ISD::VMUL);
531 void visitUDiv(User &I) { visitIntBinary(I, ISD::UDIV, ISD::VUDIV); }
532 void visitSDiv(User &I) { visitIntBinary(I, ISD::SDIV, ISD::VSDIV); }
533 void visitFDiv(User &I) { visitFPBinary(I, ISD::FDIV, ISD::VSDIV); }
534 void visitRem(User &I) {
535 const Type *Ty = I.getType();
536 if (Ty->isFloatingPoint())
537 visitFPBinary(I, ISD::FREM, 0);
539 visitIntBinary(I, Ty->isSigned() ? ISD::SREM : ISD::UREM, 0);
541 void visitAnd(User &I) { visitIntBinary(I, ISD::AND, ISD::VAND); }
542 void visitOr (User &I) { visitIntBinary(I, ISD::OR, ISD::VOR); }
543 void visitXor(User &I) { visitIntBinary(I, ISD::XOR, ISD::VXOR); }
544 void visitShl(User &I) { visitShift(I, ISD::SHL); }
545 void visitShr(User &I) {
546 visitShift(I, I.getType()->isUnsigned() ? ISD::SRL : ISD::SRA);
549 void visitSetCC(User &I, ISD::CondCode SignedOpc, ISD::CondCode UnsignedOpc,
550 ISD::CondCode FPOpc);
551 void visitSetEQ(User &I) { visitSetCC(I, ISD::SETEQ, ISD::SETEQ,
553 void visitSetNE(User &I) { visitSetCC(I, ISD::SETNE, ISD::SETNE,
555 void visitSetLE(User &I) { visitSetCC(I, ISD::SETLE, ISD::SETULE,
557 void visitSetGE(User &I) { visitSetCC(I, ISD::SETGE, ISD::SETUGE,
559 void visitSetLT(User &I) { visitSetCC(I, ISD::SETLT, ISD::SETULT,
561 void visitSetGT(User &I) { visitSetCC(I, ISD::SETGT, ISD::SETUGT,
564 void visitExtractElement(User &I);
565 void visitInsertElement(User &I);
566 void visitShuffleVector(User &I);
568 void visitGetElementPtr(User &I);
569 void visitCast(User &I);
570 void visitSelect(User &I);
572 void visitMalloc(MallocInst &I);
573 void visitFree(FreeInst &I);
574 void visitAlloca(AllocaInst &I);
575 void visitLoad(LoadInst &I);
576 void visitStore(StoreInst &I);
577 void visitPHI(PHINode &I) { } // PHI nodes are handled specially.
578 void visitCall(CallInst &I);
579 void visitInlineAsm(CallInst &I);
580 const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic);
581 void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic);
583 void visitVAStart(CallInst &I);
584 void visitVAArg(VAArgInst &I);
585 void visitVAEnd(CallInst &I);
586 void visitVACopy(CallInst &I);
587 void visitFrameReturnAddress(CallInst &I, bool isFrameAddress);
589 void visitMemIntrinsic(CallInst &I, unsigned Op);
591 void visitUserOp1(Instruction &I) {
592 assert(0 && "UserOp1 should not exist at instruction selection time!");
595 void visitUserOp2(Instruction &I) {
596 assert(0 && "UserOp2 should not exist at instruction selection time!");
600 } // end namespace llvm
602 SDOperand SelectionDAGLowering::getValue(const Value *V) {
603 SDOperand &N = NodeMap[V];
606 const Type *VTy = V->getType();
607 MVT::ValueType VT = TLI.getValueType(VTy);
608 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
609 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
610 visit(CE->getOpcode(), *CE);
611 assert(N.Val && "visit didn't populate the ValueMap!");
613 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
614 return N = DAG.getGlobalAddress(GV, VT);
615 } else if (isa<ConstantPointerNull>(C)) {
616 return N = DAG.getConstant(0, TLI.getPointerTy());
617 } else if (isa<UndefValue>(C)) {
618 if (!isa<PackedType>(VTy))
619 return N = DAG.getNode(ISD::UNDEF, VT);
621 // Create a VBUILD_VECTOR of undef nodes.
622 const PackedType *PTy = cast<PackedType>(VTy);
623 unsigned NumElements = PTy->getNumElements();
624 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
626 SmallVector<SDOperand, 8> Ops;
627 Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT));
629 // Create a VConstant node with generic Vector type.
630 Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
631 Ops.push_back(DAG.getValueType(PVT));
632 return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector,
633 &Ops[0], Ops.size());
634 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
635 return N = DAG.getConstantFP(CFP->getValue(), VT);
636 } else if (const PackedType *PTy = dyn_cast<PackedType>(VTy)) {
637 unsigned NumElements = PTy->getNumElements();
638 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
640 // Now that we know the number and type of the elements, push a
641 // Constant or ConstantFP node onto the ops list for each element of
642 // the packed constant.
643 SmallVector<SDOperand, 8> Ops;
644 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(C)) {
645 for (unsigned i = 0; i != NumElements; ++i)
646 Ops.push_back(getValue(CP->getOperand(i)));
648 assert(isa<ConstantAggregateZero>(C) && "Unknown packed constant!");
650 if (MVT::isFloatingPoint(PVT))
651 Op = DAG.getConstantFP(0, PVT);
653 Op = DAG.getConstant(0, PVT);
654 Ops.assign(NumElements, Op);
657 // Create a VBUILD_VECTOR node with generic Vector type.
658 Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
659 Ops.push_back(DAG.getValueType(PVT));
660 return N = DAG.getNode(ISD::VBUILD_VECTOR,MVT::Vector,&Ops[0],Ops.size());
662 // Canonicalize all constant ints to be unsigned.
663 return N = DAG.getConstant(cast<ConstantIntegral>(C)->getZExtValue(),VT);
667 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
668 std::map<const AllocaInst*, int>::iterator SI =
669 FuncInfo.StaticAllocaMap.find(AI);
670 if (SI != FuncInfo.StaticAllocaMap.end())
671 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
674 std::map<const Value*, unsigned>::const_iterator VMI =
675 FuncInfo.ValueMap.find(V);
676 assert(VMI != FuncInfo.ValueMap.end() && "Value not in map!");
678 unsigned InReg = VMI->second;
680 // If this type is not legal, make it so now.
681 if (VT != MVT::Vector) {
682 MVT::ValueType DestVT = TLI.getTypeToTransformTo(VT);
684 N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT);
686 // Source must be expanded. This input value is actually coming from the
687 // register pair VMI->second and VMI->second+1.
688 N = DAG.getNode(ISD::BUILD_PAIR, VT, N,
689 DAG.getCopyFromReg(DAG.getEntryNode(), InReg+1, DestVT));
690 } else if (DestVT > VT) { // Promotion case
691 if (MVT::isFloatingPoint(VT))
692 N = DAG.getNode(ISD::FP_ROUND, VT, N);
694 N = DAG.getNode(ISD::TRUNCATE, VT, N);
697 // Otherwise, if this is a vector, make it available as a generic vector
699 MVT::ValueType PTyElementVT, PTyLegalElementVT;
700 const PackedType *PTy = cast<PackedType>(VTy);
701 unsigned NE = TLI.getPackedTypeBreakdown(PTy, PTyElementVT,
704 // Build a VBUILD_VECTOR with the input registers.
705 SmallVector<SDOperand, 8> Ops;
706 if (PTyElementVT == PTyLegalElementVT) {
707 // If the value types are legal, just VBUILD the CopyFromReg nodes.
708 for (unsigned i = 0; i != NE; ++i)
709 Ops.push_back(DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
711 } else if (PTyElementVT < PTyLegalElementVT) {
712 // If the register was promoted, use TRUNCATE of FP_ROUND as appropriate.
713 for (unsigned i = 0; i != NE; ++i) {
714 SDOperand Op = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
716 if (MVT::isFloatingPoint(PTyElementVT))
717 Op = DAG.getNode(ISD::FP_ROUND, PTyElementVT, Op);
719 Op = DAG.getNode(ISD::TRUNCATE, PTyElementVT, Op);
723 // If the register was expanded, use BUILD_PAIR.
724 assert((NE & 1) == 0 && "Must expand into a multiple of 2 elements!");
725 for (unsigned i = 0; i != NE/2; ++i) {
726 SDOperand Op0 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
728 SDOperand Op1 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
730 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Op0, Op1));
734 Ops.push_back(DAG.getConstant(NE, MVT::i32));
735 Ops.push_back(DAG.getValueType(PTyLegalElementVT));
736 N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, &Ops[0], Ops.size());
738 // Finally, use a VBIT_CONVERT to make this available as the appropriate
740 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
741 DAG.getConstant(PTy->getNumElements(),
743 DAG.getValueType(TLI.getValueType(PTy->getElementType())));
750 void SelectionDAGLowering::visitRet(ReturnInst &I) {
751 if (I.getNumOperands() == 0) {
752 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot()));
755 SmallVector<SDOperand, 8> NewValues;
756 NewValues.push_back(getRoot());
757 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
758 SDOperand RetOp = getValue(I.getOperand(i));
759 bool isSigned = I.getOperand(i)->getType()->isSigned();
761 // If this is an integer return value, we need to promote it ourselves to
762 // the full width of a register, since LegalizeOp will use ANY_EXTEND rather
764 // FIXME: C calling convention requires the return type to be promoted to
765 // at least 32-bit. But this is not necessary for non-C calling conventions.
766 if (MVT::isInteger(RetOp.getValueType()) &&
767 RetOp.getValueType() < MVT::i64) {
768 MVT::ValueType TmpVT;
769 if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote)
770 TmpVT = TLI.getTypeToTransformTo(MVT::i32);
775 RetOp = DAG.getNode(ISD::SIGN_EXTEND, TmpVT, RetOp);
777 RetOp = DAG.getNode(ISD::ZERO_EXTEND, TmpVT, RetOp);
779 NewValues.push_back(RetOp);
780 NewValues.push_back(DAG.getConstant(isSigned, MVT::i32));
782 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other,
783 &NewValues[0], NewValues.size()));
786 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
787 /// the current basic block, add it to ValueMap now so that we'll get a
789 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
790 // No need to export constants.
791 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
794 if (FuncInfo.isExportedInst(V)) return;
796 unsigned Reg = FuncInfo.InitializeRegForValue(V);
797 PendingLoads.push_back(CopyValueToVirtualRegister(V, Reg));
800 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
801 const BasicBlock *FromBB) {
802 // The operands of the setcc have to be in this block. We don't know
803 // how to export them from some other block.
804 if (Instruction *VI = dyn_cast<Instruction>(V)) {
805 // Can export from current BB.
806 if (VI->getParent() == FromBB)
809 // Is already exported, noop.
810 return FuncInfo.isExportedInst(V);
813 // If this is an argument, we can export it if the BB is the entry block or
814 // if it is already exported.
815 if (isa<Argument>(V)) {
816 if (FromBB == &FromBB->getParent()->getEntryBlock())
819 // Otherwise, can only export this if it is already exported.
820 return FuncInfo.isExportedInst(V);
823 // Otherwise, constants can always be exported.
827 /// FindMergedConditions - If Cond is an expression like
828 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
829 MachineBasicBlock *TBB,
830 MachineBasicBlock *FBB,
831 MachineBasicBlock *CurBB,
833 // If this node is not part of the or/and tree, emit it as a branch.
834 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Cond);
836 if (!BOp || (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
837 BOp->getParent() != CurBB->getBasicBlock()) {
838 const BasicBlock *BB = CurBB->getBasicBlock();
840 // If the leaf of the tree is a setcond inst, merge the condition into the
842 if (BOp && isa<SetCondInst>(BOp) &&
843 // The operands of the setcc have to be in this block. We don't know
844 // how to export them from some other block. If this is the first block
845 // of the sequence, no exporting is needed.
847 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
848 isExportableFromCurrentBlock(BOp->getOperand(1), BB)))) {
849 ISD::CondCode SignCond, UnsCond, FPCond, Condition;
850 switch (BOp->getOpcode()) {
851 default: assert(0 && "Unknown setcc opcode!");
852 case Instruction::SetEQ:
853 SignCond = ISD::SETEQ;
854 UnsCond = ISD::SETEQ;
855 FPCond = ISD::SETOEQ;
857 case Instruction::SetNE:
858 SignCond = ISD::SETNE;
859 UnsCond = ISD::SETNE;
860 FPCond = ISD::SETUNE;
862 case Instruction::SetLE:
863 SignCond = ISD::SETLE;
864 UnsCond = ISD::SETULE;
865 FPCond = ISD::SETOLE;
867 case Instruction::SetGE:
868 SignCond = ISD::SETGE;
869 UnsCond = ISD::SETUGE;
870 FPCond = ISD::SETOGE;
872 case Instruction::SetLT:
873 SignCond = ISD::SETLT;
874 UnsCond = ISD::SETULT;
875 FPCond = ISD::SETOLT;
877 case Instruction::SetGT:
878 SignCond = ISD::SETGT;
879 UnsCond = ISD::SETUGT;
880 FPCond = ISD::SETOGT;
884 const Type *OpType = BOp->getOperand(0)->getType();
885 if (const PackedType *PTy = dyn_cast<PackedType>(OpType))
886 OpType = PTy->getElementType();
888 if (!FiniteOnlyFPMath() && OpType->isFloatingPoint())
890 else if (OpType->isUnsigned())
893 Condition = SignCond;
895 SelectionDAGISel::CaseBlock CB(Condition, BOp->getOperand(0),
896 BOp->getOperand(1), TBB, FBB, CurBB);
897 SwitchCases.push_back(CB);
901 // Create a CaseBlock record representing this branch.
902 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, Cond, ConstantBool::getTrue(),
904 SwitchCases.push_back(CB);
909 // Create TmpBB after CurBB.
910 MachineFunction::iterator BBI = CurBB;
911 MachineBasicBlock *TmpBB = new MachineBasicBlock(CurBB->getBasicBlock());
912 CurBB->getParent()->getBasicBlockList().insert(++BBI, TmpBB);
914 if (Opc == Instruction::Or) {
923 // Emit the LHS condition.
924 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
926 // Emit the RHS condition into TmpBB.
927 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
929 assert(Opc == Instruction::And && "Unknown merge op!");
937 // This requires creation of TmpBB after CurBB.
939 // Emit the LHS condition.
940 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
942 // Emit the RHS condition into TmpBB.
943 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
947 void SelectionDAGLowering::visitBr(BranchInst &I) {
948 // Update machine-CFG edges.
949 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
951 // Figure out which block is immediately after the current one.
952 MachineBasicBlock *NextBlock = 0;
953 MachineFunction::iterator BBI = CurMBB;
954 if (++BBI != CurMBB->getParent()->end())
957 if (I.isUnconditional()) {
958 // If this is not a fall-through branch, emit the branch.
959 if (Succ0MBB != NextBlock)
960 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
961 DAG.getBasicBlock(Succ0MBB)));
963 // Update machine-CFG edges.
964 CurMBB->addSuccessor(Succ0MBB);
969 // If this condition is one of the special cases we handle, do special stuff
971 Value *CondVal = I.getCondition();
972 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
974 // If this is a series of conditions that are or'd or and'd together, emit
975 // this as a sequence of branches instead of setcc's with and/or operations.
976 // For example, instead of something like:
989 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
990 if (BOp->hasOneUse() &&
991 (BOp->getOpcode() == Instruction::And ||
992 BOp->getOpcode() == Instruction::Or)) {
993 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
995 // If the compares in later blocks need to use values not currently
996 // exported from this block, export them now. This block should always be
998 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1000 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1001 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1002 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1005 // Emit the branch for this block.
1006 visitSwitchCase(SwitchCases[0]);
1007 SwitchCases.erase(SwitchCases.begin());
1012 // Create a CaseBlock record representing this branch.
1013 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, CondVal, ConstantBool::getTrue(),
1014 Succ0MBB, Succ1MBB, CurMBB);
1015 // Use visitSwitchCase to actually insert the fast branch sequence for this
1017 visitSwitchCase(CB);
1020 /// visitSwitchCase - Emits the necessary code to represent a single node in
1021 /// the binary search tree resulting from lowering a switch instruction.
1022 void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) {
1024 SDOperand CondLHS = getValue(CB.CmpLHS);
1026 // Build the setcc now, fold "(X == true)" to X and "(X == false)" to !X to
1027 // handle common cases produced by branch lowering.
1028 if (CB.CmpRHS == ConstantBool::getTrue() && CB.CC == ISD::SETEQ)
1030 else if (CB.CmpRHS == ConstantBool::getFalse() && CB.CC == ISD::SETEQ) {
1031 SDOperand True = DAG.getConstant(1, CondLHS.getValueType());
1032 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True);
1034 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1036 // Set NextBlock to be the MBB immediately after the current one, if any.
1037 // This is used to avoid emitting unnecessary branches to the next block.
1038 MachineBasicBlock *NextBlock = 0;
1039 MachineFunction::iterator BBI = CurMBB;
1040 if (++BBI != CurMBB->getParent()->end())
1043 // If the lhs block is the next block, invert the condition so that we can
1044 // fall through to the lhs instead of the rhs block.
1045 if (CB.TrueBB == NextBlock) {
1046 std::swap(CB.TrueBB, CB.FalseBB);
1047 SDOperand True = DAG.getConstant(1, Cond.getValueType());
1048 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
1050 SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
1051 DAG.getBasicBlock(CB.TrueBB));
1052 if (CB.FalseBB == NextBlock)
1053 DAG.setRoot(BrCond);
1055 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1056 DAG.getBasicBlock(CB.FalseBB)));
1057 // Update successor info
1058 CurMBB->addSuccessor(CB.TrueBB);
1059 CurMBB->addSuccessor(CB.FalseBB);
1062 void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) {
1063 // Emit the code for the jump table
1064 MVT::ValueType PTy = TLI.getPointerTy();
1065 assert((PTy == MVT::i32 || PTy == MVT::i64) &&
1066 "Jump table entries are 32-bit values");
1067 bool isPIC = TLI.getTargetMachine().getRelocationModel() == Reloc::PIC_;
1068 // PIC jump table entries are 32-bit values.
1069 unsigned EntrySize = isPIC ? 4 : MVT::getSizeInBits(PTy)/8;
1070 SDOperand Copy = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy);
1071 SDOperand IDX = DAG.getNode(ISD::MUL, PTy, Copy,
1072 DAG.getConstant(EntrySize, PTy));
1073 SDOperand TAB = DAG.getJumpTable(JT.JTI,PTy);
1074 SDOperand ADD = DAG.getNode(ISD::ADD, PTy, IDX, TAB);
1075 SDOperand LD = DAG.getLoad(isPIC ? MVT::i32 : PTy, Copy.getValue(1), ADD,
1078 // For Pic, the sequence is:
1079 // BRIND(load(Jumptable + index) + RelocBase)
1080 // RelocBase is the JumpTable on PPC and X86, GOT on Alpha
1082 if (TLI.usesGlobalOffsetTable())
1083 Reloc = DAG.getNode(ISD::GLOBAL_OFFSET_TABLE, PTy);
1086 ADD = (PTy != MVT::i32) ? DAG.getNode(ISD::SIGN_EXTEND, PTy, LD) : LD;
1087 ADD = DAG.getNode(ISD::ADD, PTy, ADD, Reloc);
1088 DAG.setRoot(DAG.getNode(ISD::BRIND, MVT::Other, LD.getValue(1), ADD));
1090 DAG.setRoot(DAG.getNode(ISD::BRIND, MVT::Other, LD.getValue(1), LD));
1094 void SelectionDAGLowering::visitSwitch(SwitchInst &I) {
1095 // Figure out which block is immediately after the current one.
1096 MachineBasicBlock *NextBlock = 0;
1097 MachineFunction::iterator BBI = CurMBB;
1099 if (++BBI != CurMBB->getParent()->end())
1102 MachineBasicBlock *Default = FuncInfo.MBBMap[I.getDefaultDest()];
1104 // If there is only the default destination, branch to it if it is not the
1105 // next basic block. Otherwise, just fall through.
1106 if (I.getNumOperands() == 2) {
1107 // Update machine-CFG edges.
1109 // If this is not a fall-through branch, emit the branch.
1110 if (Default != NextBlock)
1111 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
1112 DAG.getBasicBlock(Default)));
1114 CurMBB->addSuccessor(Default);
1118 // If there are any non-default case statements, create a vector of Cases
1119 // representing each one, and sort the vector so that we can efficiently
1120 // create a binary search tree from them.
1121 std::vector<Case> Cases;
1123 for (unsigned i = 1; i < I.getNumSuccessors(); ++i) {
1124 MachineBasicBlock *SMBB = FuncInfo.MBBMap[I.getSuccessor(i)];
1125 Cases.push_back(Case(I.getSuccessorValue(i), SMBB));
1128 std::sort(Cases.begin(), Cases.end(), CaseCmp());
1130 // Get the Value to be switched on and default basic blocks, which will be
1131 // inserted into CaseBlock records, representing basic blocks in the binary
1133 Value *SV = I.getOperand(0);
1135 // Get the MachineFunction which holds the current MBB. This is used during
1136 // emission of jump tables, and when inserting any additional MBBs necessary
1137 // to represent the switch.
1138 MachineFunction *CurMF = CurMBB->getParent();
1139 const BasicBlock *LLVMBB = CurMBB->getBasicBlock();
1141 // If the switch has few cases (two or less) emit a series of specific
1143 if (Cases.size() < 3) {
1144 // TODO: If any two of the cases has the same destination, and if one value
1145 // is the same as the other, but has one bit unset that the other has set,
1146 // use bit manipulation to do two compares at once. For example:
1147 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1149 // Rearrange the case blocks so that the last one falls through if possible.
1150 if (NextBlock && Default != NextBlock && Cases.back().second != NextBlock) {
1151 // The last case block won't fall through into 'NextBlock' if we emit the
1152 // branches in this order. See if rearranging a case value would help.
1153 for (unsigned i = 0, e = Cases.size()-1; i != e; ++i) {
1154 if (Cases[i].second == NextBlock) {
1155 std::swap(Cases[i], Cases.back());
1161 // Create a CaseBlock record representing a conditional branch to
1162 // the Case's target mbb if the value being switched on SV is equal
1164 MachineBasicBlock *CurBlock = CurMBB;
1165 for (unsigned i = 0, e = Cases.size(); i != e; ++i) {
1166 MachineBasicBlock *FallThrough;
1168 FallThrough = new MachineBasicBlock(CurMBB->getBasicBlock());
1169 CurMF->getBasicBlockList().insert(BBI, FallThrough);
1171 // If the last case doesn't match, go to the default block.
1172 FallThrough = Default;
1175 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, Cases[i].first,
1176 Cases[i].second, FallThrough, CurBlock);
1178 // If emitting the first comparison, just call visitSwitchCase to emit the
1179 // code into the current block. Otherwise, push the CaseBlock onto the
1180 // vector to be later processed by SDISel, and insert the node's MBB
1181 // before the next MBB.
1182 if (CurBlock == CurMBB)
1183 visitSwitchCase(CB);
1185 SwitchCases.push_back(CB);
1187 CurBlock = FallThrough;
1192 // If the switch has more than 5 blocks, and at least 31.25% dense, and the
1193 // target supports indirect branches, then emit a jump table rather than
1194 // lowering the switch to a binary tree of conditional branches.
1195 if (TLI.isOperationLegal(ISD::BRIND, TLI.getPointerTy()) &&
1197 uint64_t First =cast<ConstantIntegral>(Cases.front().first)->getZExtValue();
1198 uint64_t Last = cast<ConstantIntegral>(Cases.back().first)->getZExtValue();
1199 double Density = (double)Cases.size() / (double)((Last - First) + 1ULL);
1201 if (Density >= 0.3125) {
1202 // Create a new basic block to hold the code for loading the address
1203 // of the jump table, and jumping to it. Update successor information;
1204 // we will either branch to the default case for the switch, or the jump
1206 MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB);
1207 CurMF->getBasicBlockList().insert(BBI, JumpTableBB);
1208 CurMBB->addSuccessor(Default);
1209 CurMBB->addSuccessor(JumpTableBB);
1211 // Subtract the lowest switch case value from the value being switched on
1212 // and conditional branch to default mbb if the result is greater than the
1213 // difference between smallest and largest cases.
1214 SDOperand SwitchOp = getValue(SV);
1215 MVT::ValueType VT = SwitchOp.getValueType();
1216 SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1217 DAG.getConstant(First, VT));
1219 // The SDNode we just created, which holds the value being switched on
1220 // minus the the smallest case value, needs to be copied to a virtual
1221 // register so it can be used as an index into the jump table in a
1222 // subsequent basic block. This value may be smaller or larger than the
1223 // target's pointer type, and therefore require extension or truncating.
1224 if (VT > TLI.getPointerTy())
1225 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
1227 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
1229 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1230 SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp);
1232 // Emit the range check for the jump table, and branch to the default
1233 // block for the switch statement if the value being switched on exceeds
1234 // the largest case in the switch.
1235 SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
1236 DAG.getConstant(Last-First,VT), ISD::SETUGT);
1237 DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
1238 DAG.getBasicBlock(Default)));
1240 // Build a vector of destination BBs, corresponding to each target
1241 // of the jump table. If the value of the jump table slot corresponds to
1242 // a case statement, push the case's BB onto the vector, otherwise, push
1244 std::vector<MachineBasicBlock*> DestBBs;
1245 uint64_t TEI = First;
1246 for (CaseItr ii = Cases.begin(), ee = Cases.end(); ii != ee; ++TEI)
1247 if (cast<ConstantIntegral>(ii->first)->getZExtValue() == TEI) {
1248 DestBBs.push_back(ii->second);
1251 DestBBs.push_back(Default);
1254 // Update successor info. Add one edge to each unique successor.
1255 // Vector bool would be better, but vector<bool> is really slow.
1256 std::vector<unsigned char> SuccsHandled;
1257 SuccsHandled.resize(CurMBB->getParent()->getNumBlockIDs());
1259 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1260 E = DestBBs.end(); I != E; ++I) {
1261 if (!SuccsHandled[(*I)->getNumber()]) {
1262 SuccsHandled[(*I)->getNumber()] = true;
1263 JumpTableBB->addSuccessor(*I);
1267 // Create a jump table index for this jump table, or return an existing
1269 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1271 // Set the jump table information so that we can codegen it as a second
1272 // MachineBasicBlock
1273 JT.Reg = JumpTableReg;
1275 JT.MBB = JumpTableBB;
1276 JT.Default = Default;
1281 // Push the initial CaseRec onto the worklist
1282 std::vector<CaseRec> CaseVec;
1283 CaseVec.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
1285 while (!CaseVec.empty()) {
1286 // Grab a record representing a case range to process off the worklist
1287 CaseRec CR = CaseVec.back();
1290 // Size is the number of Cases represented by this range. If Size is 1,
1291 // then we are processing a leaf of the binary search tree. Otherwise,
1292 // we need to pick a pivot, and push left and right ranges onto the
1294 unsigned Size = CR.Range.second - CR.Range.first;
1297 // Create a CaseBlock record representing a conditional branch to
1298 // the Case's target mbb if the value being switched on SV is equal
1299 // to C. Otherwise, branch to default.
1300 Constant *C = CR.Range.first->first;
1301 MachineBasicBlock *Target = CR.Range.first->second;
1302 SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, C, Target, Default,
1305 // If the MBB representing the leaf node is the current MBB, then just
1306 // call visitSwitchCase to emit the code into the current block.
1307 // Otherwise, push the CaseBlock onto the vector to be later processed
1308 // by SDISel, and insert the node's MBB before the next MBB.
1309 if (CR.CaseBB == CurMBB)
1310 visitSwitchCase(CB);
1312 SwitchCases.push_back(CB);
1314 // split case range at pivot
1315 CaseItr Pivot = CR.Range.first + (Size / 2);
1316 CaseRange LHSR(CR.Range.first, Pivot);
1317 CaseRange RHSR(Pivot, CR.Range.second);
1318 Constant *C = Pivot->first;
1319 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1321 // We know that we branch to the LHS if the Value being switched on is
1322 // less than the Pivot value, C. We use this to optimize our binary
1323 // tree a bit, by recognizing that if SV is greater than or equal to the
1324 // LHS's Case Value, and that Case Value is exactly one less than the
1325 // Pivot's Value, then we can branch directly to the LHS's Target,
1326 // rather than creating a leaf node for it.
1327 if ((LHSR.second - LHSR.first) == 1 &&
1328 LHSR.first->first == CR.GE &&
1329 cast<ConstantIntegral>(C)->getZExtValue() ==
1330 (cast<ConstantIntegral>(CR.GE)->getZExtValue() + 1ULL)) {
1331 TrueBB = LHSR.first->second;
1333 TrueBB = new MachineBasicBlock(LLVMBB);
1334 CurMF->getBasicBlockList().insert(BBI, TrueBB);
1335 CaseVec.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1338 // Similar to the optimization above, if the Value being switched on is
1339 // known to be less than the Constant CR.LT, and the current Case Value
1340 // is CR.LT - 1, then we can branch directly to the target block for
1341 // the current Case Value, rather than emitting a RHS leaf node for it.
1342 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1343 cast<ConstantIntegral>(RHSR.first->first)->getZExtValue() ==
1344 (cast<ConstantIntegral>(CR.LT)->getZExtValue() - 1ULL)) {
1345 FalseBB = RHSR.first->second;
1347 FalseBB = new MachineBasicBlock(LLVMBB);
1348 CurMF->getBasicBlockList().insert(BBI, FalseBB);
1349 CaseVec.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1352 // Create a CaseBlock record representing a conditional branch to
1353 // the LHS node if the value being switched on SV is less than C.
1354 // Otherwise, branch to LHS.
1355 ISD::CondCode CC = C->getType()->isSigned() ? ISD::SETLT : ISD::SETULT;
1356 SelectionDAGISel::CaseBlock CB(CC, SV, C, TrueBB, FalseBB, CR.CaseBB);
1358 if (CR.CaseBB == CurMBB)
1359 visitSwitchCase(CB);
1361 SwitchCases.push_back(CB);
1366 void SelectionDAGLowering::visitSub(User &I) {
1367 // -0.0 - X --> fneg
1368 if (I.getType()->isFloatingPoint()) {
1369 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
1370 if (CFP->isExactlyValue(-0.0)) {
1371 SDOperand Op2 = getValue(I.getOperand(1));
1372 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
1375 visitFPBinary(I, ISD::FSUB, ISD::VSUB);
1377 visitIntBinary(I, ISD::SUB, ISD::VSUB);
1381 SelectionDAGLowering::visitIntBinary(User &I, unsigned IntOp, unsigned VecOp) {
1382 const Type *Ty = I.getType();
1383 SDOperand Op1 = getValue(I.getOperand(0));
1384 SDOperand Op2 = getValue(I.getOperand(1));
1386 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1387 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32);
1388 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType()));
1389 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ));
1391 setValue(&I, DAG.getNode(IntOp, Op1.getValueType(), Op1, Op2));
1396 SelectionDAGLowering::visitFPBinary(User &I, unsigned FPOp, unsigned VecOp) {
1397 const Type *Ty = I.getType();
1398 SDOperand Op1 = getValue(I.getOperand(0));
1399 SDOperand Op2 = getValue(I.getOperand(1));
1401 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1402 SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32);
1403 SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType()));
1404 setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ));
1406 setValue(&I, DAG.getNode(FPOp, Op1.getValueType(), Op1, Op2));
1410 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
1411 SDOperand Op1 = getValue(I.getOperand(0));
1412 SDOperand Op2 = getValue(I.getOperand(1));
1414 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
1416 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
1419 void SelectionDAGLowering::visitSetCC(User &I,ISD::CondCode SignedOpcode,
1420 ISD::CondCode UnsignedOpcode,
1421 ISD::CondCode FPOpcode) {
1422 SDOperand Op1 = getValue(I.getOperand(0));
1423 SDOperand Op2 = getValue(I.getOperand(1));
1424 ISD::CondCode Opcode = SignedOpcode;
1425 if (!FiniteOnlyFPMath() && I.getOperand(0)->getType()->isFloatingPoint())
1427 else if (I.getOperand(0)->getType()->isUnsigned())
1428 Opcode = UnsignedOpcode;
1429 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
1432 void SelectionDAGLowering::visitSelect(User &I) {
1433 SDOperand Cond = getValue(I.getOperand(0));
1434 SDOperand TrueVal = getValue(I.getOperand(1));
1435 SDOperand FalseVal = getValue(I.getOperand(2));
1436 if (!isa<PackedType>(I.getType())) {
1437 setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
1438 TrueVal, FalseVal));
1440 setValue(&I, DAG.getNode(ISD::VSELECT, MVT::Vector, Cond, TrueVal, FalseVal,
1441 *(TrueVal.Val->op_end()-2),
1442 *(TrueVal.Val->op_end()-1)));
1446 void SelectionDAGLowering::visitCast(User &I) {
1447 SDOperand N = getValue(I.getOperand(0));
1448 MVT::ValueType SrcVT = N.getValueType();
1449 MVT::ValueType DestVT = TLI.getValueType(I.getType());
1451 if (DestVT == MVT::Vector) {
1452 // This is a cast to a vector from something else. This is always a bit
1453 // convert. Get information about the input vector.
1454 const PackedType *DestTy = cast<PackedType>(I.getType());
1455 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
1456 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N,
1457 DAG.getConstant(DestTy->getNumElements(),MVT::i32),
1458 DAG.getValueType(EltVT)));
1459 } else if (SrcVT == DestVT) {
1460 setValue(&I, N); // noop cast.
1461 } else if (DestVT == MVT::i1) {
1462 // Cast to bool is a comparison against zero, not truncation to zero.
1463 SDOperand Zero = isInteger(SrcVT) ? DAG.getConstant(0, N.getValueType()) :
1464 DAG.getConstantFP(0.0, N.getValueType());
1465 setValue(&I, DAG.getSetCC(MVT::i1, N, Zero, ISD::SETNE));
1466 } else if (isInteger(SrcVT)) {
1467 if (isInteger(DestVT)) { // Int -> Int cast
1468 if (DestVT < SrcVT) // Truncating cast?
1469 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
1470 else if (I.getOperand(0)->getType()->isSigned())
1471 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
1473 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
1474 } else if (isFloatingPoint(DestVT)) { // Int -> FP cast
1475 if (I.getOperand(0)->getType()->isSigned())
1476 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
1478 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
1480 assert(0 && "Unknown cast!");
1482 } else if (isFloatingPoint(SrcVT)) {
1483 if (isFloatingPoint(DestVT)) { // FP -> FP cast
1484 if (DestVT < SrcVT) // Rounding cast?
1485 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N));
1487 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
1488 } else if (isInteger(DestVT)) { // FP -> Int cast.
1489 if (I.getType()->isSigned())
1490 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
1492 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
1494 assert(0 && "Unknown cast!");
1497 assert(SrcVT == MVT::Vector && "Unknown cast!");
1498 assert(DestVT != MVT::Vector && "Casts to vector already handled!");
1499 // This is a cast from a vector to something else. This is always a bit
1500 // convert. Get information about the input vector.
1501 setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N));
1505 void SelectionDAGLowering::visitInsertElement(User &I) {
1506 SDOperand InVec = getValue(I.getOperand(0));
1507 SDOperand InVal = getValue(I.getOperand(1));
1508 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
1509 getValue(I.getOperand(2)));
1511 SDOperand Num = *(InVec.Val->op_end()-2);
1512 SDOperand Typ = *(InVec.Val->op_end()-1);
1513 setValue(&I, DAG.getNode(ISD::VINSERT_VECTOR_ELT, MVT::Vector,
1514 InVec, InVal, InIdx, Num, Typ));
1517 void SelectionDAGLowering::visitExtractElement(User &I) {
1518 SDOperand InVec = getValue(I.getOperand(0));
1519 SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
1520 getValue(I.getOperand(1)));
1521 SDOperand Typ = *(InVec.Val->op_end()-1);
1522 setValue(&I, DAG.getNode(ISD::VEXTRACT_VECTOR_ELT,
1523 TLI.getValueType(I.getType()), InVec, InIdx));
1526 void SelectionDAGLowering::visitShuffleVector(User &I) {
1527 SDOperand V1 = getValue(I.getOperand(0));
1528 SDOperand V2 = getValue(I.getOperand(1));
1529 SDOperand Mask = getValue(I.getOperand(2));
1531 SDOperand Num = *(V1.Val->op_end()-2);
1532 SDOperand Typ = *(V2.Val->op_end()-1);
1533 setValue(&I, DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
1534 V1, V2, Mask, Num, Typ));
1538 void SelectionDAGLowering::visitGetElementPtr(User &I) {
1539 SDOperand N = getValue(I.getOperand(0));
1540 const Type *Ty = I.getOperand(0)->getType();
1542 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
1545 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
1546 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
1549 uint64_t Offset = TD->getStructLayout(StTy)->MemberOffsets[Field];
1550 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
1551 getIntPtrConstant(Offset));
1553 Ty = StTy->getElementType(Field);
1555 Ty = cast<SequentialType>(Ty)->getElementType();
1557 // If this is a constant subscript, handle it quickly.
1558 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
1559 if (CI->getZExtValue() == 0) continue;
1561 if (CI->getType()->isSigned())
1563 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
1566 TD->getTypeSize(Ty)*cast<ConstantInt>(CI)->getZExtValue();
1567 N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs));
1571 // N = N + Idx * ElementSize;
1572 uint64_t ElementSize = TD->getTypeSize(Ty);
1573 SDOperand IdxN = getValue(Idx);
1575 // If the index is smaller or larger than intptr_t, truncate or extend
1577 if (IdxN.getValueType() < N.getValueType()) {
1578 if (Idx->getType()->isSigned())
1579 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
1581 IdxN = DAG.getNode(ISD::ZERO_EXTEND, N.getValueType(), IdxN);
1582 } else if (IdxN.getValueType() > N.getValueType())
1583 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
1585 // If this is a multiply by a power of two, turn it into a shl
1586 // immediately. This is a very common case.
1587 if (isPowerOf2_64(ElementSize)) {
1588 unsigned Amt = Log2_64(ElementSize);
1589 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
1590 DAG.getConstant(Amt, TLI.getShiftAmountTy()));
1591 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
1595 SDOperand Scale = getIntPtrConstant(ElementSize);
1596 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
1597 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
1603 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
1604 // If this is a fixed sized alloca in the entry block of the function,
1605 // allocate it statically on the stack.
1606 if (FuncInfo.StaticAllocaMap.count(&I))
1607 return; // getValue will auto-populate this.
1609 const Type *Ty = I.getAllocatedType();
1610 uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
1611 unsigned Align = std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
1614 SDOperand AllocSize = getValue(I.getArraySize());
1615 MVT::ValueType IntPtr = TLI.getPointerTy();
1616 if (IntPtr < AllocSize.getValueType())
1617 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
1618 else if (IntPtr > AllocSize.getValueType())
1619 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
1621 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
1622 getIntPtrConstant(TySize));
1624 // Handle alignment. If the requested alignment is less than or equal to the
1625 // stack alignment, ignore it and round the size of the allocation up to the
1626 // stack alignment size. If the size is greater than the stack alignment, we
1627 // note this in the DYNAMIC_STACKALLOC node.
1628 unsigned StackAlign =
1629 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
1630 if (Align <= StackAlign) {
1632 // Add SA-1 to the size.
1633 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
1634 getIntPtrConstant(StackAlign-1));
1635 // Mask out the low bits for alignment purposes.
1636 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
1637 getIntPtrConstant(~(uint64_t)(StackAlign-1)));
1640 SDOperand Ops[] = { getRoot(), AllocSize, getIntPtrConstant(Align) };
1641 const MVT::ValueType *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
1643 SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3);
1644 DAG.setRoot(setValue(&I, DSA).getValue(1));
1646 // Inform the Frame Information that we have just allocated a variable-sized
1648 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
1651 void SelectionDAGLowering::visitLoad(LoadInst &I) {
1652 SDOperand Ptr = getValue(I.getOperand(0));
1658 // Do not serialize non-volatile loads against each other.
1659 Root = DAG.getRoot();
1662 setValue(&I, getLoadFrom(I.getType(), Ptr, I.getOperand(0),
1663 Root, I.isVolatile()));
1666 SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr,
1667 const Value *SV, SDOperand Root,
1670 if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
1671 MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
1672 L = DAG.getVecLoad(PTy->getNumElements(), PVT, Root, Ptr,
1673 DAG.getSrcValue(SV));
1675 L = DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SV, isVolatile);
1679 DAG.setRoot(L.getValue(1));
1681 PendingLoads.push_back(L.getValue(1));
1687 void SelectionDAGLowering::visitStore(StoreInst &I) {
1688 Value *SrcV = I.getOperand(0);
1689 SDOperand Src = getValue(SrcV);
1690 SDOperand Ptr = getValue(I.getOperand(1));
1691 DAG.setRoot(DAG.getStore(getRoot(), Src, Ptr, I.getOperand(1),
1695 /// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot
1696 /// access memory and has no other side effects at all.
1697 static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) {
1698 #define GET_NO_MEMORY_INTRINSICS
1699 #include "llvm/Intrinsics.gen"
1700 #undef GET_NO_MEMORY_INTRINSICS
1704 // IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't
1705 // have any side-effects or if it only reads memory.
1706 static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) {
1707 #define GET_SIDE_EFFECT_INFO
1708 #include "llvm/Intrinsics.gen"
1709 #undef GET_SIDE_EFFECT_INFO
1713 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
1715 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
1716 unsigned Intrinsic) {
1717 bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic);
1718 bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic);
1720 // Build the operand list.
1721 SmallVector<SDOperand, 8> Ops;
1722 if (HasChain) { // If this intrinsic has side-effects, chainify it.
1724 // We don't need to serialize loads against other loads.
1725 Ops.push_back(DAG.getRoot());
1727 Ops.push_back(getRoot());
1731 // Add the intrinsic ID as an integer operand.
1732 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
1734 // Add all operands of the call to the operand list.
1735 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
1736 SDOperand Op = getValue(I.getOperand(i));
1738 // If this is a vector type, force it to the right packed type.
1739 if (Op.getValueType() == MVT::Vector) {
1740 const PackedType *OpTy = cast<PackedType>(I.getOperand(i)->getType());
1741 MVT::ValueType EltVT = TLI.getValueType(OpTy->getElementType());
1743 MVT::ValueType VVT = MVT::getVectorType(EltVT, OpTy->getNumElements());
1744 assert(VVT != MVT::Other && "Intrinsic uses a non-legal type?");
1745 Op = DAG.getNode(ISD::VBIT_CONVERT, VVT, Op);
1748 assert(TLI.isTypeLegal(Op.getValueType()) &&
1749 "Intrinsic uses a non-legal type?");
1753 std::vector<MVT::ValueType> VTs;
1754 if (I.getType() != Type::VoidTy) {
1755 MVT::ValueType VT = TLI.getValueType(I.getType());
1756 if (VT == MVT::Vector) {
1757 const PackedType *DestTy = cast<PackedType>(I.getType());
1758 MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
1760 VT = MVT::getVectorType(EltVT, DestTy->getNumElements());
1761 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
1764 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
1768 VTs.push_back(MVT::Other);
1770 const MVT::ValueType *VTList = DAG.getNodeValueTypes(VTs);
1775 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(),
1776 &Ops[0], Ops.size());
1777 else if (I.getType() != Type::VoidTy)
1778 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(),
1779 &Ops[0], Ops.size());
1781 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(),
1782 &Ops[0], Ops.size());
1785 SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1);
1787 PendingLoads.push_back(Chain);
1791 if (I.getType() != Type::VoidTy) {
1792 if (const PackedType *PTy = dyn_cast<PackedType>(I.getType())) {
1793 MVT::ValueType EVT = TLI.getValueType(PTy->getElementType());
1794 Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
1795 DAG.getConstant(PTy->getNumElements(), MVT::i32),
1796 DAG.getValueType(EVT));
1798 setValue(&I, Result);
1802 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
1803 /// we want to emit this as a call to a named external function, return the name
1804 /// otherwise lower it and return null.
1806 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
1807 switch (Intrinsic) {
1809 // By default, turn this into a target intrinsic node.
1810 visitTargetIntrinsic(I, Intrinsic);
1812 case Intrinsic::vastart: visitVAStart(I); return 0;
1813 case Intrinsic::vaend: visitVAEnd(I); return 0;
1814 case Intrinsic::vacopy: visitVACopy(I); return 0;
1815 case Intrinsic::returnaddress: visitFrameReturnAddress(I, false); return 0;
1816 case Intrinsic::frameaddress: visitFrameReturnAddress(I, true); return 0;
1817 case Intrinsic::setjmp:
1818 return "_setjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
1820 case Intrinsic::longjmp:
1821 return "_longjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
1823 case Intrinsic::memcpy_i32:
1824 case Intrinsic::memcpy_i64:
1825 visitMemIntrinsic(I, ISD::MEMCPY);
1827 case Intrinsic::memset_i32:
1828 case Intrinsic::memset_i64:
1829 visitMemIntrinsic(I, ISD::MEMSET);
1831 case Intrinsic::memmove_i32:
1832 case Intrinsic::memmove_i64:
1833 visitMemIntrinsic(I, ISD::MEMMOVE);
1836 case Intrinsic::dbg_stoppoint: {
1837 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1838 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
1839 if (DebugInfo && SPI.getContext() && DebugInfo->Verify(SPI.getContext())) {
1843 Ops[1] = getValue(SPI.getLineValue());
1844 Ops[2] = getValue(SPI.getColumnValue());
1846 DebugInfoDesc *DD = DebugInfo->getDescFor(SPI.getContext());
1847 assert(DD && "Not a debug information descriptor");
1848 CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD);
1850 Ops[3] = DAG.getString(CompileUnit->getFileName());
1851 Ops[4] = DAG.getString(CompileUnit->getDirectory());
1853 DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops, 5));
1858 case Intrinsic::dbg_region_start: {
1859 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1860 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
1861 if (DebugInfo && RSI.getContext() && DebugInfo->Verify(RSI.getContext())) {
1862 unsigned LabelID = DebugInfo->RecordRegionStart(RSI.getContext());
1863 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, getRoot(),
1864 DAG.getConstant(LabelID, MVT::i32)));
1869 case Intrinsic::dbg_region_end: {
1870 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1871 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
1872 if (DebugInfo && REI.getContext() && DebugInfo->Verify(REI.getContext())) {
1873 unsigned LabelID = DebugInfo->RecordRegionEnd(REI.getContext());
1874 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other,
1875 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
1880 case Intrinsic::dbg_func_start: {
1881 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1882 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
1883 if (DebugInfo && FSI.getSubprogram() &&
1884 DebugInfo->Verify(FSI.getSubprogram())) {
1885 unsigned LabelID = DebugInfo->RecordRegionStart(FSI.getSubprogram());
1886 DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other,
1887 getRoot(), DAG.getConstant(LabelID, MVT::i32)));
1892 case Intrinsic::dbg_declare: {
1893 MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
1894 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
1895 if (DebugInfo && DI.getVariable() && DebugInfo->Verify(DI.getVariable())) {
1896 SDOperand AddressOp = getValue(DI.getAddress());
1897 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp))
1898 DebugInfo->RecordVariable(DI.getVariable(), FI->getIndex());
1904 case Intrinsic::isunordered_f32:
1905 case Intrinsic::isunordered_f64:
1906 setValue(&I, DAG.getSetCC(MVT::i1,getValue(I.getOperand(1)),
1907 getValue(I.getOperand(2)), ISD::SETUO));
1910 case Intrinsic::sqrt_f32:
1911 case Intrinsic::sqrt_f64:
1912 setValue(&I, DAG.getNode(ISD::FSQRT,
1913 getValue(I.getOperand(1)).getValueType(),
1914 getValue(I.getOperand(1))));
1916 case Intrinsic::powi_f32:
1917 case Intrinsic::powi_f64:
1918 setValue(&I, DAG.getNode(ISD::FPOWI,
1919 getValue(I.getOperand(1)).getValueType(),
1920 getValue(I.getOperand(1)),
1921 getValue(I.getOperand(2))));
1923 case Intrinsic::pcmarker: {
1924 SDOperand Tmp = getValue(I.getOperand(1));
1925 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
1928 case Intrinsic::readcyclecounter: {
1929 SDOperand Op = getRoot();
1930 SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER,
1931 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
1934 DAG.setRoot(Tmp.getValue(1));
1937 case Intrinsic::bswap_i16:
1938 case Intrinsic::bswap_i32:
1939 case Intrinsic::bswap_i64:
1940 setValue(&I, DAG.getNode(ISD::BSWAP,
1941 getValue(I.getOperand(1)).getValueType(),
1942 getValue(I.getOperand(1))));
1944 case Intrinsic::cttz_i8:
1945 case Intrinsic::cttz_i16:
1946 case Intrinsic::cttz_i32:
1947 case Intrinsic::cttz_i64:
1948 setValue(&I, DAG.getNode(ISD::CTTZ,
1949 getValue(I.getOperand(1)).getValueType(),
1950 getValue(I.getOperand(1))));
1952 case Intrinsic::ctlz_i8:
1953 case Intrinsic::ctlz_i16:
1954 case Intrinsic::ctlz_i32:
1955 case Intrinsic::ctlz_i64:
1956 setValue(&I, DAG.getNode(ISD::CTLZ,
1957 getValue(I.getOperand(1)).getValueType(),
1958 getValue(I.getOperand(1))));
1960 case Intrinsic::ctpop_i8:
1961 case Intrinsic::ctpop_i16:
1962 case Intrinsic::ctpop_i32:
1963 case Intrinsic::ctpop_i64:
1964 setValue(&I, DAG.getNode(ISD::CTPOP,
1965 getValue(I.getOperand(1)).getValueType(),
1966 getValue(I.getOperand(1))));
1968 case Intrinsic::stacksave: {
1969 SDOperand Op = getRoot();
1970 SDOperand Tmp = DAG.getNode(ISD::STACKSAVE,
1971 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
1973 DAG.setRoot(Tmp.getValue(1));
1976 case Intrinsic::stackrestore: {
1977 SDOperand Tmp = getValue(I.getOperand(1));
1978 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
1981 case Intrinsic::prefetch:
1982 // FIXME: Currently discarding prefetches.
1988 void SelectionDAGLowering::visitCall(CallInst &I) {
1989 const char *RenameFn = 0;
1990 if (Function *F = I.getCalledFunction()) {
1991 if (F->isExternal())
1992 if (unsigned IID = F->getIntrinsicID()) {
1993 RenameFn = visitIntrinsicCall(I, IID);
1996 } else { // Not an LLVM intrinsic.
1997 const std::string &Name = F->getName();
1998 if (Name[0] == 'c' && (Name == "copysign" || Name == "copysignf")) {
1999 if (I.getNumOperands() == 3 && // Basic sanity checks.
2000 I.getOperand(1)->getType()->isFloatingPoint() &&
2001 I.getType() == I.getOperand(1)->getType() &&
2002 I.getType() == I.getOperand(2)->getType()) {
2003 SDOperand LHS = getValue(I.getOperand(1));
2004 SDOperand RHS = getValue(I.getOperand(2));
2005 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
2009 } else if (Name[0] == 'f' && (Name == "fabs" || Name == "fabsf")) {
2010 if (I.getNumOperands() == 2 && // Basic sanity checks.
2011 I.getOperand(1)->getType()->isFloatingPoint() &&
2012 I.getType() == I.getOperand(1)->getType()) {
2013 SDOperand Tmp = getValue(I.getOperand(1));
2014 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
2017 } else if (Name[0] == 's' && (Name == "sin" || Name == "sinf")) {
2018 if (I.getNumOperands() == 2 && // Basic sanity checks.
2019 I.getOperand(1)->getType()->isFloatingPoint() &&
2020 I.getType() == I.getOperand(1)->getType()) {
2021 SDOperand Tmp = getValue(I.getOperand(1));
2022 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
2025 } else if (Name[0] == 'c' && (Name == "cos" || Name == "cosf")) {
2026 if (I.getNumOperands() == 2 && // Basic sanity checks.
2027 I.getOperand(1)->getType()->isFloatingPoint() &&
2028 I.getType() == I.getOperand(1)->getType()) {
2029 SDOperand Tmp = getValue(I.getOperand(1));
2030 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
2035 } else if (isa<InlineAsm>(I.getOperand(0))) {
2042 Callee = getValue(I.getOperand(0));
2044 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
2045 std::vector<std::pair<SDOperand, const Type*> > Args;
2046 Args.reserve(I.getNumOperands());
2047 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2048 Value *Arg = I.getOperand(i);
2049 SDOperand ArgNode = getValue(Arg);
2050 Args.push_back(std::make_pair(ArgNode, Arg->getType()));
2053 const PointerType *PT = cast<PointerType>(I.getCalledValue()->getType());
2054 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
2056 std::pair<SDOperand,SDOperand> Result =
2057 TLI.LowerCallTo(getRoot(), I.getType(), FTy->isVarArg(), I.getCallingConv(),
2058 I.isTailCall(), Callee, Args, DAG);
2059 if (I.getType() != Type::VoidTy)
2060 setValue(&I, Result.first);
2061 DAG.setRoot(Result.second);
2064 SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
2065 SDOperand &Chain, SDOperand &Flag)const{
2066 SDOperand Val = DAG.getCopyFromReg(Chain, Regs[0], RegVT, Flag);
2067 Chain = Val.getValue(1);
2068 Flag = Val.getValue(2);
2070 // If the result was expanded, copy from the top part.
2071 if (Regs.size() > 1) {
2072 assert(Regs.size() == 2 &&
2073 "Cannot expand to more than 2 elts yet!");
2074 SDOperand Hi = DAG.getCopyFromReg(Chain, Regs[1], RegVT, Flag);
2075 Chain = Hi.getValue(1);
2076 Flag = Hi.getValue(2);
2077 if (DAG.getTargetLoweringInfo().isLittleEndian())
2078 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi);
2080 return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Hi, Val);
2083 // Otherwise, if the return value was promoted or extended, truncate it to the
2084 // appropriate type.
2085 if (RegVT == ValueVT)
2088 if (MVT::isInteger(RegVT)) {
2089 if (ValueVT < RegVT)
2090 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
2092 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val);
2094 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val);
2098 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
2099 /// specified value into the registers specified by this object. This uses
2100 /// Chain/Flag as the input and updates them for the output Chain/Flag.
2101 void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
2102 SDOperand &Chain, SDOperand &Flag,
2103 MVT::ValueType PtrVT) const {
2104 if (Regs.size() == 1) {
2105 // If there is a single register and the types differ, this must be
2107 if (RegVT != ValueVT) {
2108 if (MVT::isInteger(RegVT)) {
2109 if (RegVT < ValueVT)
2110 Val = DAG.getNode(ISD::TRUNCATE, RegVT, Val);
2112 Val = DAG.getNode(ISD::ANY_EXTEND, RegVT, Val);
2114 Val = DAG.getNode(ISD::FP_EXTEND, RegVT, Val);
2116 Chain = DAG.getCopyToReg(Chain, Regs[0], Val, Flag);
2117 Flag = Chain.getValue(1);
2119 std::vector<unsigned> R(Regs);
2120 if (!DAG.getTargetLoweringInfo().isLittleEndian())
2121 std::reverse(R.begin(), R.end());
2123 for (unsigned i = 0, e = R.size(); i != e; ++i) {
2124 SDOperand Part = DAG.getNode(ISD::EXTRACT_ELEMENT, RegVT, Val,
2125 DAG.getConstant(i, PtrVT));
2126 Chain = DAG.getCopyToReg(Chain, R[i], Part, Flag);
2127 Flag = Chain.getValue(1);
2132 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
2133 /// operand list. This adds the code marker and includes the number of
2134 /// values added into it.
2135 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
2136 std::vector<SDOperand> &Ops) const {
2137 Ops.push_back(DAG.getConstant(Code | (Regs.size() << 3), MVT::i32));
2138 for (unsigned i = 0, e = Regs.size(); i != e; ++i)
2139 Ops.push_back(DAG.getRegister(Regs[i], RegVT));
2142 /// isAllocatableRegister - If the specified register is safe to allocate,
2143 /// i.e. it isn't a stack pointer or some other special register, return the
2144 /// register class for the register. Otherwise, return null.
2145 static const TargetRegisterClass *
2146 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
2147 const TargetLowering &TLI, const MRegisterInfo *MRI) {
2148 MVT::ValueType FoundVT = MVT::Other;
2149 const TargetRegisterClass *FoundRC = 0;
2150 for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(),
2151 E = MRI->regclass_end(); RCI != E; ++RCI) {
2152 MVT::ValueType ThisVT = MVT::Other;
2154 const TargetRegisterClass *RC = *RCI;
2155 // If none of the the value types for this register class are valid, we
2156 // can't use it. For example, 64-bit reg classes on 32-bit targets.
2157 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
2159 if (TLI.isTypeLegal(*I)) {
2160 // If we have already found this register in a different register class,
2161 // choose the one with the largest VT specified. For example, on
2162 // PowerPC, we favor f64 register classes over f32.
2163 if (FoundVT == MVT::Other ||
2164 MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) {
2171 if (ThisVT == MVT::Other) continue;
2173 // NOTE: This isn't ideal. In particular, this might allocate the
2174 // frame pointer in functions that need it (due to them not being taken
2175 // out of allocation, because a variable sized allocation hasn't been seen
2176 // yet). This is a slight code pessimization, but should still work.
2177 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
2178 E = RC->allocation_order_end(MF); I != E; ++I)
2180 // We found a matching register class. Keep looking at others in case
2181 // we find one with larger registers that this physreg is also in.
2190 RegsForValue SelectionDAGLowering::
2191 GetRegistersForValue(const std::string &ConstrCode,
2192 MVT::ValueType VT, bool isOutReg, bool isInReg,
2193 std::set<unsigned> &OutputRegs,
2194 std::set<unsigned> &InputRegs) {
2195 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
2196 TLI.getRegForInlineAsmConstraint(ConstrCode, VT);
2197 std::vector<unsigned> Regs;
2199 unsigned NumRegs = VT != MVT::Other ? TLI.getNumElements(VT) : 1;
2200 MVT::ValueType RegVT;
2201 MVT::ValueType ValueVT = VT;
2203 if (PhysReg.first) {
2204 if (VT == MVT::Other)
2205 ValueVT = *PhysReg.second->vt_begin();
2207 // Get the actual register value type. This is important, because the user
2208 // may have asked for (e.g.) the AX register in i32 type. We need to
2209 // remember that AX is actually i16 to get the right extension.
2210 RegVT = *PhysReg.second->vt_begin();
2212 // This is a explicit reference to a physical register.
2213 Regs.push_back(PhysReg.first);
2215 // If this is an expanded reference, add the rest of the regs to Regs.
2217 TargetRegisterClass::iterator I = PhysReg.second->begin();
2218 TargetRegisterClass::iterator E = PhysReg.second->end();
2219 for (; *I != PhysReg.first; ++I)
2220 assert(I != E && "Didn't find reg!");
2222 // Already added the first reg.
2224 for (; NumRegs; --NumRegs, ++I) {
2225 assert(I != E && "Ran out of registers to allocate!");
2229 return RegsForValue(Regs, RegVT, ValueVT);
2232 // This is a reference to a register class. Allocate NumRegs consecutive,
2233 // available, registers from the class.
2234 std::vector<unsigned> RegClassRegs =
2235 TLI.getRegClassForInlineAsmConstraint(ConstrCode, VT);
2237 const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo();
2238 MachineFunction &MF = *CurMBB->getParent();
2239 unsigned NumAllocated = 0;
2240 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
2241 unsigned Reg = RegClassRegs[i];
2242 // See if this register is available.
2243 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
2244 (isInReg && InputRegs.count(Reg))) { // Already used.
2245 // Make sure we find consecutive registers.
2250 // Check to see if this register is allocatable (i.e. don't give out the
2252 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, MRI);
2254 // Make sure we find consecutive registers.
2259 // Okay, this register is good, we can use it.
2262 // If we allocated enough consecutive
2263 if (NumAllocated == NumRegs) {
2264 unsigned RegStart = (i-NumAllocated)+1;
2265 unsigned RegEnd = i+1;
2266 // Mark all of the allocated registers used.
2267 for (unsigned i = RegStart; i != RegEnd; ++i) {
2268 unsigned Reg = RegClassRegs[i];
2269 Regs.push_back(Reg);
2270 if (isOutReg) OutputRegs.insert(Reg); // Mark reg used.
2271 if (isInReg) InputRegs.insert(Reg); // Mark reg used.
2274 return RegsForValue(Regs, *RC->vt_begin(), VT);
2278 // Otherwise, we couldn't allocate enough registers for this.
2279 return RegsForValue();
2283 /// visitInlineAsm - Handle a call to an InlineAsm object.
2285 void SelectionDAGLowering::visitInlineAsm(CallInst &I) {
2286 InlineAsm *IA = cast<InlineAsm>(I.getOperand(0));
2288 SDOperand AsmStr = DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
2291 // Note, we treat inline asms both with and without side-effects as the same.
2292 // If an inline asm doesn't have side effects and doesn't access memory, we
2293 // could not choose to not chain it.
2294 bool hasSideEffects = IA->hasSideEffects();
2296 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
2297 std::vector<MVT::ValueType> ConstraintVTs;
2299 /// AsmNodeOperands - A list of pairs. The first element is a register, the
2300 /// second is a bitfield where bit #0 is set if it is a use and bit #1 is set
2301 /// if it is a def of that register.
2302 std::vector<SDOperand> AsmNodeOperands;
2303 AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain
2304 AsmNodeOperands.push_back(AsmStr);
2306 SDOperand Chain = getRoot();
2309 // We fully assign registers here at isel time. This is not optimal, but
2310 // should work. For register classes that correspond to LLVM classes, we
2311 // could let the LLVM RA do its thing, but we currently don't. Do a prepass
2312 // over the constraints, collecting fixed registers that we know we can't use.
2313 std::set<unsigned> OutputRegs, InputRegs;
2315 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2316 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
2317 std::string &ConstraintCode = Constraints[i].Codes[0];
2319 MVT::ValueType OpVT;
2321 // Compute the value type for each operand and add it to ConstraintVTs.
2322 switch (Constraints[i].Type) {
2323 case InlineAsm::isOutput:
2324 if (!Constraints[i].isIndirectOutput) {
2325 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
2326 OpVT = TLI.getValueType(I.getType());
2328 const Type *OpTy = I.getOperand(OpNum)->getType();
2329 OpVT = TLI.getValueType(cast<PointerType>(OpTy)->getElementType());
2330 OpNum++; // Consumes a call operand.
2333 case InlineAsm::isInput:
2334 OpVT = TLI.getValueType(I.getOperand(OpNum)->getType());
2335 OpNum++; // Consumes a call operand.
2337 case InlineAsm::isClobber:
2342 ConstraintVTs.push_back(OpVT);
2344 if (TLI.getRegForInlineAsmConstraint(ConstraintCode, OpVT).first == 0)
2345 continue; // Not assigned a fixed reg.
2347 // Build a list of regs that this operand uses. This always has a single
2348 // element for promoted/expanded operands.
2349 RegsForValue Regs = GetRegistersForValue(ConstraintCode, OpVT,
2351 OutputRegs, InputRegs);
2353 switch (Constraints[i].Type) {
2354 case InlineAsm::isOutput:
2355 // We can't assign any other output to this register.
2356 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2357 // If this is an early-clobber output, it cannot be assigned to the same
2358 // value as the input reg.
2359 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
2360 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2362 case InlineAsm::isInput:
2363 // We can't assign any other input to this register.
2364 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2366 case InlineAsm::isClobber:
2367 // Clobbered regs cannot be used as inputs or outputs.
2368 InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2369 OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
2374 // Loop over all of the inputs, copying the operand values into the
2375 // appropriate registers and processing the output regs.
2376 RegsForValue RetValRegs;
2377 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
2380 for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
2381 assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
2382 std::string &ConstraintCode = Constraints[i].Codes[0];
2384 switch (Constraints[i].Type) {
2385 case InlineAsm::isOutput: {
2386 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
2387 if (ConstraintCode.size() == 1) // not a physreg name.
2388 CTy = TLI.getConstraintType(ConstraintCode[0]);
2390 if (CTy == TargetLowering::C_Memory) {
2392 SDOperand InOperandVal = getValue(I.getOperand(OpNum));
2394 // Check that the operand (the address to store to) isn't a float.
2395 if (!MVT::isInteger(InOperandVal.getValueType()))
2396 assert(0 && "MATCH FAIL!");
2398 if (!Constraints[i].isIndirectOutput)
2399 assert(0 && "MATCH FAIL!");
2401 OpNum++; // Consumes a call operand.
2403 // Extend/truncate to the right pointer type if needed.
2404 MVT::ValueType PtrType = TLI.getPointerTy();
2405 if (InOperandVal.getValueType() < PtrType)
2406 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
2407 else if (InOperandVal.getValueType() > PtrType)
2408 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
2410 // Add information to the INLINEASM node to know about this output.
2411 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
2412 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2413 AsmNodeOperands.push_back(InOperandVal);
2417 // Otherwise, this is a register output.
2418 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
2420 // If this is an early-clobber output, or if there is an input
2421 // constraint that matches this, we need to reserve the input register
2422 // so no other inputs allocate to it.
2423 bool UsesInputRegister = false;
2424 if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
2425 UsesInputRegister = true;
2427 // Copy the output from the appropriate register. Find a register that
2430 GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
2431 true, UsesInputRegister,
2432 OutputRegs, InputRegs);
2433 assert(!Regs.Regs.empty() && "Couldn't allocate output reg!");
2435 if (!Constraints[i].isIndirectOutput) {
2436 assert(RetValRegs.Regs.empty() &&
2437 "Cannot have multiple output constraints yet!");
2438 assert(I.getType() != Type::VoidTy && "Bad inline asm!");
2441 IndirectStoresToEmit.push_back(std::make_pair(Regs,
2442 I.getOperand(OpNum)));
2443 OpNum++; // Consumes a call operand.
2446 // Add information to the INLINEASM node to know that this register is
2448 Regs.AddInlineAsmOperands(2 /*REGDEF*/, DAG, AsmNodeOperands);
2451 case InlineAsm::isInput: {
2452 SDOperand InOperandVal = getValue(I.getOperand(OpNum));
2453 OpNum++; // Consumes a call operand.
2455 if (isdigit(ConstraintCode[0])) { // Matching constraint?
2456 // If this is required to match an output register we have already set,
2457 // just use its register.
2458 unsigned OperandNo = atoi(ConstraintCode.c_str());
2460 // Scan until we find the definition we already emitted of this operand.
2461 // When we find it, create a RegsForValue operand.
2462 unsigned CurOp = 2; // The first operand.
2463 for (; OperandNo; --OperandNo) {
2464 // Advance to the next operand.
2466 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
2467 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
2468 (NumOps & 7) == 4 /*MEM*/) &&
2469 "Skipped past definitions?");
2470 CurOp += (NumOps>>3)+1;
2474 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
2475 assert((NumOps & 7) == 2 /*REGDEF*/ &&
2476 "Skipped past definitions?");
2478 // Add NumOps>>3 registers to MatchedRegs.
2479 RegsForValue MatchedRegs;
2480 MatchedRegs.ValueVT = InOperandVal.getValueType();
2481 MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType();
2482 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
2483 unsigned Reg=cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
2484 MatchedRegs.Regs.push_back(Reg);
2487 // Use the produced MatchedRegs object to
2488 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag,
2489 TLI.getPointerTy());
2490 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
2494 TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
2495 if (ConstraintCode.size() == 1) // not a physreg name.
2496 CTy = TLI.getConstraintType(ConstraintCode[0]);
2498 if (CTy == TargetLowering::C_Other) {
2499 if (!TLI.isOperandValidForConstraint(InOperandVal, ConstraintCode[0]))
2500 assert(0 && "MATCH FAIL!");
2502 // Add information to the INLINEASM node to know about this input.
2503 unsigned ResOpType = 3 /*IMM*/ | (1 << 3);
2504 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2505 AsmNodeOperands.push_back(InOperandVal);
2507 } else if (CTy == TargetLowering::C_Memory) {
2510 // Check that the operand isn't a float.
2511 if (!MVT::isInteger(InOperandVal.getValueType()))
2512 assert(0 && "MATCH FAIL!");
2514 // Extend/truncate to the right pointer type if needed.
2515 MVT::ValueType PtrType = TLI.getPointerTy();
2516 if (InOperandVal.getValueType() < PtrType)
2517 InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
2518 else if (InOperandVal.getValueType() > PtrType)
2519 InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
2521 // Add information to the INLINEASM node to know about this input.
2522 unsigned ResOpType = 4/*MEM*/ | (1 << 3);
2523 AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
2524 AsmNodeOperands.push_back(InOperandVal);
2528 assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
2530 // Copy the input into the appropriate registers.
2531 RegsForValue InRegs =
2532 GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
2533 false, true, OutputRegs, InputRegs);
2534 // FIXME: should be match fail.
2535 assert(!InRegs.Regs.empty() && "Couldn't allocate input reg!");
2537 InRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag, TLI.getPointerTy());
2539 InRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG, AsmNodeOperands);
2542 case InlineAsm::isClobber: {
2543 RegsForValue ClobberedRegs =
2544 GetRegistersForValue(ConstraintCode, MVT::Other, false, false,
2545 OutputRegs, InputRegs);
2546 // Add the clobbered value to the operand list, so that the register
2547 // allocator is aware that the physreg got clobbered.
2548 if (!ClobberedRegs.Regs.empty())
2549 ClobberedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG, AsmNodeOperands);
2555 // Finish up input operands.
2556 AsmNodeOperands[0] = Chain;
2557 if (Flag.Val) AsmNodeOperands.push_back(Flag);
2559 Chain = DAG.getNode(ISD::INLINEASM,
2560 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
2561 &AsmNodeOperands[0], AsmNodeOperands.size());
2562 Flag = Chain.getValue(1);
2564 // If this asm returns a register value, copy the result from that register
2565 // and set it as the value of the call.
2566 if (!RetValRegs.Regs.empty())
2567 setValue(&I, RetValRegs.getCopyFromRegs(DAG, Chain, Flag));
2569 std::vector<std::pair<SDOperand, Value*> > StoresToEmit;
2571 // Process indirect outputs, first output all of the flagged copies out of
2573 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
2574 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
2575 Value *Ptr = IndirectStoresToEmit[i].second;
2576 SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, Flag);
2577 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
2580 // Emit the non-flagged stores from the physregs.
2581 SmallVector<SDOperand, 8> OutChains;
2582 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
2583 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first,
2584 getValue(StoresToEmit[i].second),
2585 StoresToEmit[i].second, 0));
2586 if (!OutChains.empty())
2587 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2588 &OutChains[0], OutChains.size());
2593 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
2594 SDOperand Src = getValue(I.getOperand(0));
2596 MVT::ValueType IntPtr = TLI.getPointerTy();
2598 if (IntPtr < Src.getValueType())
2599 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
2600 else if (IntPtr > Src.getValueType())
2601 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
2603 // Scale the source by the type size.
2604 uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType());
2605 Src = DAG.getNode(ISD::MUL, Src.getValueType(),
2606 Src, getIntPtrConstant(ElementSize));
2608 std::vector<std::pair<SDOperand, const Type*> > Args;
2609 Args.push_back(std::make_pair(Src, TLI.getTargetData()->getIntPtrType()));
2611 std::pair<SDOperand,SDOperand> Result =
2612 TLI.LowerCallTo(getRoot(), I.getType(), false, CallingConv::C, true,
2613 DAG.getExternalSymbol("malloc", IntPtr),
2615 setValue(&I, Result.first); // Pointers always fit in registers
2616 DAG.setRoot(Result.second);
2619 void SelectionDAGLowering::visitFree(FreeInst &I) {
2620 std::vector<std::pair<SDOperand, const Type*> > Args;
2621 Args.push_back(std::make_pair(getValue(I.getOperand(0)),
2622 TLI.getTargetData()->getIntPtrType()));
2623 MVT::ValueType IntPtr = TLI.getPointerTy();
2624 std::pair<SDOperand,SDOperand> Result =
2625 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, CallingConv::C, true,
2626 DAG.getExternalSymbol("free", IntPtr), Args, DAG);
2627 DAG.setRoot(Result.second);
2630 // InsertAtEndOfBasicBlock - This method should be implemented by targets that
2631 // mark instructions with the 'usesCustomDAGSchedInserter' flag. These
2632 // instructions are special in various ways, which require special support to
2633 // insert. The specified MachineInstr is created but not inserted into any
2634 // basic blocks, and the scheduler passes ownership of it to this method.
2635 MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
2636 MachineBasicBlock *MBB) {
2637 std::cerr << "If a target marks an instruction with "
2638 "'usesCustomDAGSchedInserter', it must implement "
2639 "TargetLowering::InsertAtEndOfBasicBlock!\n";
2644 void SelectionDAGLowering::visitVAStart(CallInst &I) {
2645 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
2646 getValue(I.getOperand(1)),
2647 DAG.getSrcValue(I.getOperand(1))));
2650 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
2651 SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
2652 getValue(I.getOperand(0)),
2653 DAG.getSrcValue(I.getOperand(0)));
2655 DAG.setRoot(V.getValue(1));
2658 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
2659 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
2660 getValue(I.getOperand(1)),
2661 DAG.getSrcValue(I.getOperand(1))));
2664 void SelectionDAGLowering::visitVACopy(CallInst &I) {
2665 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
2666 getValue(I.getOperand(1)),
2667 getValue(I.getOperand(2)),
2668 DAG.getSrcValue(I.getOperand(1)),
2669 DAG.getSrcValue(I.getOperand(2))));
2672 /// TargetLowering::LowerArguments - This is the default LowerArguments
2673 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
2674 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
2675 /// integrated into SDISel.
2676 std::vector<SDOperand>
2677 TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
2678 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
2679 std::vector<SDOperand> Ops;
2680 Ops.push_back(DAG.getRoot());
2681 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
2682 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
2684 // Add one result value for each formal argument.
2685 std::vector<MVT::ValueType> RetVals;
2686 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
2687 MVT::ValueType VT = getValueType(I->getType());
2689 switch (getTypeAction(VT)) {
2690 default: assert(0 && "Unknown type action!");
2692 RetVals.push_back(VT);
2695 RetVals.push_back(getTypeToTransformTo(VT));
2698 if (VT != MVT::Vector) {
2699 // If this is a large integer, it needs to be broken up into small
2700 // integers. Figure out what the destination type is and how many small
2701 // integers it turns into.
2702 MVT::ValueType NVT = getTypeToTransformTo(VT);
2703 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2704 for (unsigned i = 0; i != NumVals; ++i)
2705 RetVals.push_back(NVT);
2707 // Otherwise, this is a vector type. We only support legal vectors
2709 unsigned NumElems = cast<PackedType>(I->getType())->getNumElements();
2710 const Type *EltTy = cast<PackedType>(I->getType())->getElementType();
2712 // Figure out if there is a Packed type corresponding to this Vector
2713 // type. If so, convert to the packed type.
2714 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2715 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2716 RetVals.push_back(TVT);
2718 assert(0 && "Don't support illegal by-val vector arguments yet!");
2725 RetVals.push_back(MVT::Other);
2728 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS,
2729 DAG.getNodeValueTypes(RetVals), RetVals.size(),
2730 &Ops[0], Ops.size()).Val;
2732 DAG.setRoot(SDOperand(Result, Result->getNumValues()-1));
2734 // Set up the return result vector.
2737 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
2738 MVT::ValueType VT = getValueType(I->getType());
2740 switch (getTypeAction(VT)) {
2741 default: assert(0 && "Unknown type action!");
2743 Ops.push_back(SDOperand(Result, i++));
2746 SDOperand Op(Result, i++);
2747 if (MVT::isInteger(VT)) {
2748 unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
2750 Op = DAG.getNode(AssertOp, Op.getValueType(), Op, DAG.getValueType(VT));
2751 Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
2753 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
2754 Op = DAG.getNode(ISD::FP_ROUND, VT, Op);
2760 if (VT != MVT::Vector) {
2761 // If this is a large integer, it needs to be reassembled from small
2762 // integers. Figure out what the source elt type is and how many small
2764 MVT::ValueType NVT = getTypeToTransformTo(VT);
2765 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2767 SDOperand Lo = SDOperand(Result, i++);
2768 SDOperand Hi = SDOperand(Result, i++);
2770 if (!isLittleEndian())
2773 Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Lo, Hi));
2775 // Value scalarized into many values. Unimp for now.
2776 assert(0 && "Cannot expand i64 -> i16 yet!");
2779 // Otherwise, this is a vector type. We only support legal vectors
2781 const PackedType *PTy = cast<PackedType>(I->getType());
2782 unsigned NumElems = PTy->getNumElements();
2783 const Type *EltTy = PTy->getElementType();
2785 // Figure out if there is a Packed type corresponding to this Vector
2786 // type. If so, convert to the packed type.
2787 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2788 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2789 SDOperand N = SDOperand(Result, i++);
2790 // Handle copies from generic vectors to registers.
2791 N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
2792 DAG.getConstant(NumElems, MVT::i32),
2793 DAG.getValueType(getValueType(EltTy)));
2796 assert(0 && "Don't support illegal by-val vector arguments yet!");
2807 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
2808 /// implementation, which just inserts an ISD::CALL node, which is later custom
2809 /// lowered by the target to something concrete. FIXME: When all targets are
2810 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
2811 std::pair<SDOperand, SDOperand>
2812 TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg,
2813 unsigned CallingConv, bool isTailCall,
2815 ArgListTy &Args, SelectionDAG &DAG) {
2816 SmallVector<SDOperand, 32> Ops;
2817 Ops.push_back(Chain); // Op#0 - Chain
2818 Ops.push_back(DAG.getConstant(CallingConv, getPointerTy())); // Op#1 - CC
2819 Ops.push_back(DAG.getConstant(isVarArg, getPointerTy())); // Op#2 - VarArg
2820 Ops.push_back(DAG.getConstant(isTailCall, getPointerTy())); // Op#3 - Tail
2821 Ops.push_back(Callee);
2823 // Handle all of the outgoing arguments.
2824 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
2825 MVT::ValueType VT = getValueType(Args[i].second);
2826 SDOperand Op = Args[i].first;
2827 bool isSigned = Args[i].second->isSigned();
2828 switch (getTypeAction(VT)) {
2829 default: assert(0 && "Unknown type action!");
2832 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2835 if (MVT::isInteger(VT)) {
2836 unsigned ExtOp = isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2837 Op = DAG.getNode(ExtOp, getTypeToTransformTo(VT), Op);
2839 assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
2840 Op = DAG.getNode(ISD::FP_EXTEND, getTypeToTransformTo(VT), Op);
2843 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2846 if (VT != MVT::Vector) {
2847 // If this is a large integer, it needs to be broken down into small
2848 // integers. Figure out what the source elt type is and how many small
2850 MVT::ValueType NVT = getTypeToTransformTo(VT);
2851 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2853 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, NVT, Op,
2854 DAG.getConstant(0, getPointerTy()));
2855 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, NVT, Op,
2856 DAG.getConstant(1, getPointerTy()));
2857 if (!isLittleEndian())
2861 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2863 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2865 // Value scalarized into many values. Unimp for now.
2866 assert(0 && "Cannot expand i64 -> i16 yet!");
2869 // Otherwise, this is a vector type. We only support legal vectors
2871 const PackedType *PTy = cast<PackedType>(Args[i].second);
2872 unsigned NumElems = PTy->getNumElements();
2873 const Type *EltTy = PTy->getElementType();
2875 // Figure out if there is a Packed type corresponding to this Vector
2876 // type. If so, convert to the packed type.
2877 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2878 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2879 // Insert a VBIT_CONVERT of the MVT::Vector type to the packed type.
2880 Op = DAG.getNode(ISD::VBIT_CONVERT, TVT, Op);
2882 Ops.push_back(DAG.getConstant(isSigned, MVT::i32));
2884 assert(0 && "Don't support illegal by-val vector call args yet!");
2892 // Figure out the result value types.
2893 SmallVector<MVT::ValueType, 4> RetTys;
2895 if (RetTy != Type::VoidTy) {
2896 MVT::ValueType VT = getValueType(RetTy);
2897 switch (getTypeAction(VT)) {
2898 default: assert(0 && "Unknown type action!");
2900 RetTys.push_back(VT);
2903 RetTys.push_back(getTypeToTransformTo(VT));
2906 if (VT != MVT::Vector) {
2907 // If this is a large integer, it needs to be reassembled from small
2908 // integers. Figure out what the source elt type is and how many small
2910 MVT::ValueType NVT = getTypeToTransformTo(VT);
2911 unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
2912 for (unsigned i = 0; i != NumVals; ++i)
2913 RetTys.push_back(NVT);
2915 // Otherwise, this is a vector type. We only support legal vectors
2917 const PackedType *PTy = cast<PackedType>(RetTy);
2918 unsigned NumElems = PTy->getNumElements();
2919 const Type *EltTy = PTy->getElementType();
2921 // Figure out if there is a Packed type corresponding to this Vector
2922 // type. If so, convert to the packed type.
2923 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2924 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2925 RetTys.push_back(TVT);
2927 assert(0 && "Don't support illegal by-val vector call results yet!");
2934 RetTys.push_back(MVT::Other); // Always has a chain.
2936 // Finally, create the CALL node.
2937 SDOperand Res = DAG.getNode(ISD::CALL,
2938 DAG.getVTList(&RetTys[0], RetTys.size()),
2939 &Ops[0], Ops.size());
2941 // This returns a pair of operands. The first element is the
2942 // return value for the function (if RetTy is not VoidTy). The second
2943 // element is the outgoing token chain.
2945 if (RetTys.size() != 1) {
2946 MVT::ValueType VT = getValueType(RetTy);
2947 if (RetTys.size() == 2) {
2950 // If this value was promoted, truncate it down.
2951 if (ResVal.getValueType() != VT) {
2952 if (VT == MVT::Vector) {
2953 // Insert a VBITCONVERT to convert from the packed result type to the
2954 // MVT::Vector type.
2955 unsigned NumElems = cast<PackedType>(RetTy)->getNumElements();
2956 const Type *EltTy = cast<PackedType>(RetTy)->getElementType();
2958 // Figure out if there is a Packed type corresponding to this Vector
2959 // type. If so, convert to the packed type.
2960 MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
2961 if (TVT != MVT::Other && isTypeLegal(TVT)) {
2962 // Insert a VBIT_CONVERT of the FORMAL_ARGUMENTS to a
2963 // "N x PTyElementVT" MVT::Vector type.
2964 ResVal = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, ResVal,
2965 DAG.getConstant(NumElems, MVT::i32),
2966 DAG.getValueType(getValueType(EltTy)));
2970 } else if (MVT::isInteger(VT)) {
2971 unsigned AssertOp = RetTy->isSigned() ?
2972 ISD::AssertSext : ISD::AssertZext;
2973 ResVal = DAG.getNode(AssertOp, ResVal.getValueType(), ResVal,
2974 DAG.getValueType(VT));
2975 ResVal = DAG.getNode(ISD::TRUNCATE, VT, ResVal);
2977 assert(MVT::isFloatingPoint(VT));
2978 ResVal = DAG.getNode(ISD::FP_ROUND, VT, ResVal);
2981 } else if (RetTys.size() == 3) {
2982 ResVal = DAG.getNode(ISD::BUILD_PAIR, VT,
2983 Res.getValue(0), Res.getValue(1));
2986 assert(0 && "Case not handled yet!");
2990 return std::make_pair(ResVal, Res.getValue(Res.Val->getNumValues()-1));
2995 // It is always conservatively correct for llvm.returnaddress and
2996 // llvm.frameaddress to return 0.
2998 // FIXME: Change this to insert a FRAMEADDR/RETURNADDR node, and have that be
2999 // expanded to 0 if the target wants.
3000 std::pair<SDOperand, SDOperand>
3001 TargetLowering::LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain,
3002 unsigned Depth, SelectionDAG &DAG) {
3003 return std::make_pair(DAG.getConstant(0, getPointerTy()), Chain);
3006 SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
3007 assert(0 && "LowerOperation not implemented for this target!");
3012 SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op,
3013 SelectionDAG &DAG) {
3014 assert(0 && "CustomPromoteOperation not implemented for this target!");
3019 void SelectionDAGLowering::visitFrameReturnAddress(CallInst &I, bool isFrame) {
3020 unsigned Depth = (unsigned)cast<ConstantInt>(I.getOperand(1))->getZExtValue();
3021 std::pair<SDOperand,SDOperand> Result =
3022 TLI.LowerFrameReturnAddress(isFrame, getRoot(), Depth, DAG);
3023 setValue(&I, Result.first);
3024 DAG.setRoot(Result.second);
3027 /// getMemsetValue - Vectorized representation of the memset value
3029 static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT,
3030 SelectionDAG &DAG) {
3031 MVT::ValueType CurVT = VT;
3032 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
3033 uint64_t Val = C->getValue() & 255;
3035 while (CurVT != MVT::i8) {
3036 Val = (Val << Shift) | Val;
3038 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
3040 return DAG.getConstant(Val, VT);
3042 Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
3044 while (CurVT != MVT::i8) {
3046 DAG.getNode(ISD::OR, VT,
3047 DAG.getNode(ISD::SHL, VT, Value,
3048 DAG.getConstant(Shift, MVT::i8)), Value);
3050 CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
3057 /// getMemsetStringVal - Similar to getMemsetValue. Except this is only
3058 /// used when a memcpy is turned into a memset when the source is a constant
3060 static SDOperand getMemsetStringVal(MVT::ValueType VT,
3061 SelectionDAG &DAG, TargetLowering &TLI,
3062 std::string &Str, unsigned Offset) {
3063 MVT::ValueType CurVT = VT;
3065 unsigned MSB = getSizeInBits(VT) / 8;
3066 if (TLI.isLittleEndian())
3067 Offset = Offset + MSB - 1;
3068 for (unsigned i = 0; i != MSB; ++i) {
3069 Val = (Val << 8) | Str[Offset];
3070 Offset += TLI.isLittleEndian() ? -1 : 1;
3072 return DAG.getConstant(Val, VT);
3075 /// getMemBasePlusOffset - Returns base and offset node for the
3076 static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset,
3077 SelectionDAG &DAG, TargetLowering &TLI) {
3078 MVT::ValueType VT = Base.getValueType();
3079 return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
3082 /// MeetsMaxMemopRequirement - Determines if the number of memory ops required
3083 /// to replace the memset / memcpy is below the threshold. It also returns the
3084 /// types of the sequence of memory ops to perform memset / memcpy.
3085 static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps,
3086 unsigned Limit, uint64_t Size,
3087 unsigned Align, TargetLowering &TLI) {
3090 if (TLI.allowsUnalignedMemoryAccesses()) {
3093 switch (Align & 7) {
3109 MVT::ValueType LVT = MVT::i64;
3110 while (!TLI.isTypeLegal(LVT))
3111 LVT = (MVT::ValueType)((unsigned)LVT - 1);
3112 assert(MVT::isInteger(LVT));
3117 unsigned NumMemOps = 0;
3119 unsigned VTSize = getSizeInBits(VT) / 8;
3120 while (VTSize > Size) {
3121 VT = (MVT::ValueType)((unsigned)VT - 1);
3124 assert(MVT::isInteger(VT));
3126 if (++NumMemOps > Limit)
3128 MemOps.push_back(VT);
3135 void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) {
3136 SDOperand Op1 = getValue(I.getOperand(1));
3137 SDOperand Op2 = getValue(I.getOperand(2));
3138 SDOperand Op3 = getValue(I.getOperand(3));
3139 SDOperand Op4 = getValue(I.getOperand(4));
3140 unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue();
3141 if (Align == 0) Align = 1;
3143 if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) {
3144 std::vector<MVT::ValueType> MemOps;
3146 // Expand memset / memcpy to a series of load / store ops
3147 // if the size operand falls below a certain threshold.
3148 SmallVector<SDOperand, 8> OutChains;
3150 default: break; // Do nothing for now.
3152 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(),
3153 Size->getValue(), Align, TLI)) {
3154 unsigned NumMemOps = MemOps.size();
3155 unsigned Offset = 0;
3156 for (unsigned i = 0; i < NumMemOps; i++) {
3157 MVT::ValueType VT = MemOps[i];
3158 unsigned VTSize = getSizeInBits(VT) / 8;
3159 SDOperand Value = getMemsetValue(Op2, VT, DAG);
3160 SDOperand Store = DAG.getStore(getRoot(), Value,
3161 getMemBasePlusOffset(Op1, Offset, DAG, TLI),
3162 I.getOperand(1), Offset);
3163 OutChains.push_back(Store);
3170 if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(),
3171 Size->getValue(), Align, TLI)) {
3172 unsigned NumMemOps = MemOps.size();
3173 unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0;
3174 GlobalAddressSDNode *G = NULL;
3176 bool CopyFromStr = false;
3178 if (Op2.getOpcode() == ISD::GlobalAddress)
3179 G = cast<GlobalAddressSDNode>(Op2);
3180 else if (Op2.getOpcode() == ISD::ADD &&
3181 Op2.getOperand(0).getOpcode() == ISD::GlobalAddress &&
3182 Op2.getOperand(1).getOpcode() == ISD::Constant) {
3183 G = cast<GlobalAddressSDNode>(Op2.getOperand(0));
3184 SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue();
3187 GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
3189 Str = GV->getStringValue(false);
3197 for (unsigned i = 0; i < NumMemOps; i++) {
3198 MVT::ValueType VT = MemOps[i];
3199 unsigned VTSize = getSizeInBits(VT) / 8;
3200 SDOperand Value, Chain, Store;
3203 Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
3206 DAG.getStore(Chain, Value,
3207 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
3208 I.getOperand(1), DstOff);
3210 Value = DAG.getLoad(VT, getRoot(),
3211 getMemBasePlusOffset(Op2, SrcOff, DAG, TLI),
3212 I.getOperand(2), SrcOff);
3213 Chain = Value.getValue(1);
3215 DAG.getStore(Chain, Value,
3216 getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
3217 I.getOperand(1), DstOff);
3219 OutChains.push_back(Store);
3228 if (!OutChains.empty()) {
3229 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
3230 &OutChains[0], OutChains.size()));
3235 DAG.setRoot(DAG.getNode(Op, MVT::Other, getRoot(), Op1, Op2, Op3, Op4));
3238 //===----------------------------------------------------------------------===//
3239 // SelectionDAGISel code
3240 //===----------------------------------------------------------------------===//
3242 unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) {
3243 return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
3246 void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
3247 // FIXME: we only modify the CFG to split critical edges. This
3248 // updates dom and loop info.
3249 AU.addRequired<AliasAnalysis>();
3253 /// OptimizeNoopCopyExpression - We have determined that the specified cast
3254 /// instruction is a noop copy (e.g. it's casting from one pointer type to
3255 /// another, int->uint, or int->sbyte on PPC.
3257 /// Return true if any changes are made.
3258 static bool OptimizeNoopCopyExpression(CastInst *CI) {
3259 BasicBlock *DefBB = CI->getParent();
3261 /// InsertedCasts - Only insert a cast in each block once.
3262 std::map<BasicBlock*, CastInst*> InsertedCasts;
3264 bool MadeChange = false;
3265 for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
3267 Use &TheUse = UI.getUse();
3268 Instruction *User = cast<Instruction>(*UI);
3270 // Figure out which BB this cast is used in. For PHI's this is the
3271 // appropriate predecessor block.
3272 BasicBlock *UserBB = User->getParent();
3273 if (PHINode *PN = dyn_cast<PHINode>(User)) {
3274 unsigned OpVal = UI.getOperandNo()/2;
3275 UserBB = PN->getIncomingBlock(OpVal);
3278 // Preincrement use iterator so we don't invalidate it.
3281 // If this user is in the same block as the cast, don't change the cast.
3282 if (UserBB == DefBB) continue;
3284 // If we have already inserted a cast into this block, use it.
3285 CastInst *&InsertedCast = InsertedCasts[UserBB];
3287 if (!InsertedCast) {
3288 BasicBlock::iterator InsertPt = UserBB->begin();
3289 while (isa<PHINode>(InsertPt)) ++InsertPt;
3292 new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
3296 // Replace a use of the cast with a use of the new casat.
3297 TheUse = InsertedCast;
3300 // If we removed all uses, nuke the cast.
3301 if (CI->use_empty())
3302 CI->eraseFromParent();
3307 /// InsertGEPComputeCode - Insert code into BB to compute Ptr+PtrOffset,
3308 /// casting to the type of GEPI.
3309 static Instruction *InsertGEPComputeCode(Instruction *&V, BasicBlock *BB,
3310 Instruction *GEPI, Value *Ptr,
3312 if (V) return V; // Already computed.
3314 BasicBlock::iterator InsertPt;
3315 if (BB == GEPI->getParent()) {
3316 // If insert into the GEP's block, insert right after the GEP.
3320 // Otherwise, insert at the top of BB, after any PHI nodes
3321 InsertPt = BB->begin();
3322 while (isa<PHINode>(InsertPt)) ++InsertPt;
3325 // If Ptr is itself a cast, but in some other BB, emit a copy of the cast into
3326 // BB so that there is only one value live across basic blocks (the cast
3328 if (CastInst *CI = dyn_cast<CastInst>(Ptr))
3329 if (CI->getParent() != BB && isa<PointerType>(CI->getOperand(0)->getType()))
3330 Ptr = new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
3332 // Add the offset, cast it to the right type.
3333 Ptr = BinaryOperator::createAdd(Ptr, PtrOffset, "", InsertPt);
3334 return V = new CastInst(Ptr, GEPI->getType(), "", InsertPt);
3337 /// ReplaceUsesOfGEPInst - Replace all uses of RepPtr with inserted code to
3338 /// compute its value. The RepPtr value can be computed with Ptr+PtrOffset. One
3339 /// trivial way of doing this would be to evaluate Ptr+PtrOffset in RepPtr's
3340 /// block, then ReplaceAllUsesWith'ing everything. However, we would prefer to
3341 /// sink PtrOffset into user blocks where doing so will likely allow us to fold
3342 /// the constant add into a load or store instruction. Additionally, if a user
3343 /// is a pointer-pointer cast, we look through it to find its users.
3344 static void ReplaceUsesOfGEPInst(Instruction *RepPtr, Value *Ptr,
3345 Constant *PtrOffset, BasicBlock *DefBB,
3346 GetElementPtrInst *GEPI,
3347 std::map<BasicBlock*,Instruction*> &InsertedExprs) {
3348 while (!RepPtr->use_empty()) {
3349 Instruction *User = cast<Instruction>(RepPtr->use_back());
3351 // If the user is a Pointer-Pointer cast, recurse.
3352 if (isa<CastInst>(User) && isa<PointerType>(User->getType())) {
3353 ReplaceUsesOfGEPInst(User, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs);
3355 // Drop the use of RepPtr. The cast is dead. Don't delete it now, else we
3356 // could invalidate an iterator.
3357 User->setOperand(0, UndefValue::get(RepPtr->getType()));
3361 // If this is a load of the pointer, or a store through the pointer, emit
3362 // the increment into the load/store block.
3363 Instruction *NewVal;
3364 if (isa<LoadInst>(User) ||
3365 (isa<StoreInst>(User) && User->getOperand(0) != RepPtr)) {
3366 NewVal = InsertGEPComputeCode(InsertedExprs[User->getParent()],
3367 User->getParent(), GEPI,
3370 // If this use is not foldable into the addressing mode, use a version
3371 // emitted in the GEP block.
3372 NewVal = InsertGEPComputeCode(InsertedExprs[DefBB], DefBB, GEPI,
3376 if (GEPI->getType() != RepPtr->getType()) {
3377 BasicBlock::iterator IP = NewVal;
3379 NewVal = new CastInst(NewVal, RepPtr->getType(), "", IP);
3381 User->replaceUsesOfWith(RepPtr, NewVal);
3386 /// OptimizeGEPExpression - Since we are doing basic-block-at-a-time instruction
3387 /// selection, we want to be a bit careful about some things. In particular, if
3388 /// we have a GEP instruction that is used in a different block than it is
3389 /// defined, the addressing expression of the GEP cannot be folded into loads or
3390 /// stores that use it. In this case, decompose the GEP and move constant
3391 /// indices into blocks that use it.
3392 static bool OptimizeGEPExpression(GetElementPtrInst *GEPI,
3393 const TargetData *TD) {
3394 // If this GEP is only used inside the block it is defined in, there is no
3395 // need to rewrite it.
3396 bool isUsedOutsideDefBB = false;
3397 BasicBlock *DefBB = GEPI->getParent();
3398 for (Value::use_iterator UI = GEPI->use_begin(), E = GEPI->use_end();
3400 if (cast<Instruction>(*UI)->getParent() != DefBB) {
3401 isUsedOutsideDefBB = true;
3405 if (!isUsedOutsideDefBB) return false;
3407 // If this GEP has no non-zero constant indices, there is nothing we can do,
3409 bool hasConstantIndex = false;
3410 bool hasVariableIndex = false;
3411 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
3412 E = GEPI->op_end(); OI != E; ++OI) {
3413 if (ConstantInt *CI = dyn_cast<ConstantInt>(*OI)) {
3414 if (CI->getZExtValue()) {
3415 hasConstantIndex = true;
3419 hasVariableIndex = true;
3423 // If this is a "GEP X, 0, 0, 0", turn this into a cast.
3424 if (!hasConstantIndex && !hasVariableIndex) {
3425 Value *NC = new CastInst(GEPI->getOperand(0), GEPI->getType(),
3426 GEPI->getName(), GEPI);
3427 GEPI->replaceAllUsesWith(NC);
3428 GEPI->eraseFromParent();
3432 // If this is a GEP &Alloca, 0, 0, forward subst the frame index into uses.
3433 if (!hasConstantIndex && !isa<AllocaInst>(GEPI->getOperand(0)))
3436 // Otherwise, decompose the GEP instruction into multiplies and adds. Sum the
3437 // constant offset (which we now know is non-zero) and deal with it later.
3438 uint64_t ConstantOffset = 0;
3439 const Type *UIntPtrTy = TD->getIntPtrType();
3440 Value *Ptr = new CastInst(GEPI->getOperand(0), UIntPtrTy, "", GEPI);
3441 const Type *Ty = GEPI->getOperand(0)->getType();
3443 for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
3444 E = GEPI->op_end(); OI != E; ++OI) {
3446 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
3447 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
3449 ConstantOffset += TD->getStructLayout(StTy)->MemberOffsets[Field];
3450 Ty = StTy->getElementType(Field);
3452 Ty = cast<SequentialType>(Ty)->getElementType();
3454 // Handle constant subscripts.
3455 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
3456 if (CI->getZExtValue() == 0) continue;
3457 if (CI->getType()->isSigned())
3458 ConstantOffset += (int64_t)TD->getTypeSize(Ty)*CI->getSExtValue();
3460 ConstantOffset += TD->getTypeSize(Ty)*CI->getZExtValue();
3464 // Ptr = Ptr + Idx * ElementSize;
3466 // Cast Idx to UIntPtrTy if needed.
3467 Idx = new CastInst(Idx, UIntPtrTy, "", GEPI);
3469 uint64_t ElementSize = TD->getTypeSize(Ty);
3470 // Mask off bits that should not be set.
3471 ElementSize &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
3472 Constant *SizeCst = ConstantInt::get(UIntPtrTy, ElementSize);
3474 // Multiply by the element size and add to the base.
3475 Idx = BinaryOperator::createMul(Idx, SizeCst, "", GEPI);
3476 Ptr = BinaryOperator::createAdd(Ptr, Idx, "", GEPI);
3480 // Make sure that the offset fits in uintptr_t.
3481 ConstantOffset &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
3482 Constant *PtrOffset = ConstantInt::get(UIntPtrTy, ConstantOffset);
3484 // Okay, we have now emitted all of the variable index parts to the BB that
3485 // the GEP is defined in. Loop over all of the using instructions, inserting
3486 // an "add Ptr, ConstantOffset" into each block that uses it and update the
3487 // instruction to use the newly computed value, making GEPI dead. When the
3488 // user is a load or store instruction address, we emit the add into the user
3489 // block, otherwise we use a canonical version right next to the gep (these
3490 // won't be foldable as addresses, so we might as well share the computation).
3492 std::map<BasicBlock*,Instruction*> InsertedExprs;
3493 ReplaceUsesOfGEPInst(GEPI, Ptr, PtrOffset, DefBB, GEPI, InsertedExprs);
3495 // Finally, the GEP is dead, remove it.
3496 GEPI->eraseFromParent();
3502 /// SplitEdgeNicely - Split the critical edge from TI to it's specified
3503 /// successor if it will improve codegen. We only do this if the successor has
3504 /// phi nodes (otherwise critical edges are ok). If there is already another
3505 /// predecessor of the succ that is empty (and thus has no phi nodes), use it
3506 /// instead of introducing a new block.
3507 static void SplitEdgeNicely(TerminatorInst *TI, unsigned SuccNum, Pass *P) {
3508 BasicBlock *TIBB = TI->getParent();
3509 BasicBlock *Dest = TI->getSuccessor(SuccNum);
3510 assert(isa<PHINode>(Dest->begin()) &&
3511 "This should only be called if Dest has a PHI!");
3513 /// TIPHIValues - This array is lazily computed to determine the values of
3514 /// PHIs in Dest that TI would provide.
3515 std::vector<Value*> TIPHIValues;
3517 // Check to see if Dest has any blocks that can be used as a split edge for
3519 for (pred_iterator PI = pred_begin(Dest), E = pred_end(Dest); PI != E; ++PI) {
3520 BasicBlock *Pred = *PI;
3521 // To be usable, the pred has to end with an uncond branch to the dest.
3522 BranchInst *PredBr = dyn_cast<BranchInst>(Pred->getTerminator());
3523 if (!PredBr || !PredBr->isUnconditional() ||
3524 // Must be empty other than the branch.
3525 &Pred->front() != PredBr)
3528 // Finally, since we know that Dest has phi nodes in it, we have to make
3529 // sure that jumping to Pred will have the same affect as going to Dest in
3530 // terms of PHI values.
3533 bool FoundMatch = true;
3534 for (BasicBlock::iterator I = Dest->begin();
3535 (PN = dyn_cast<PHINode>(I)); ++I, ++PHINo) {
3536 if (PHINo == TIPHIValues.size())
3537 TIPHIValues.push_back(PN->getIncomingValueForBlock(TIBB));
3539 // If the PHI entry doesn't work, we can't use this pred.
3540 if (TIPHIValues[PHINo] != PN->getIncomingValueForBlock(Pred)) {
3546 // If we found a workable predecessor, change TI to branch to Succ.
3548 Dest->removePredecessor(TIBB);
3549 TI->setSuccessor(SuccNum, Pred);
3554 SplitCriticalEdge(TI, SuccNum, P, true);
3558 bool SelectionDAGISel::runOnFunction(Function &Fn) {
3559 MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
3560 RegMap = MF.getSSARegMap();
3561 DEBUG(std::cerr << "\n\n\n=== " << Fn.getName() << "\n");
3563 // First, split all critical edges.
3565 // In this pass we also look for GEP and cast instructions that are used
3566 // across basic blocks and rewrite them to improve basic-block-at-a-time
3569 bool MadeChange = true;
3570 while (MadeChange) {
3572 for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
3573 // Split all critical edges where the dest block has a PHI.
3574 TerminatorInst *BBTI = BB->getTerminator();
3575 if (BBTI->getNumSuccessors() > 1) {
3576 for (unsigned i = 0, e = BBTI->getNumSuccessors(); i != e; ++i)
3577 if (isa<PHINode>(BBTI->getSuccessor(i)->begin()) &&
3578 isCriticalEdge(BBTI, i, true))
3579 SplitEdgeNicely(BBTI, i, this);
3583 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
3584 Instruction *I = BBI++;
3585 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
3586 MadeChange |= OptimizeGEPExpression(GEPI, TLI.getTargetData());
3587 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
3588 // If the source of the cast is a constant, then this should have
3589 // already been constant folded. The only reason NOT to constant fold
3590 // it is if something (e.g. LSR) was careful to place the constant
3591 // evaluation in a block other than then one that uses it (e.g. to hoist
3592 // the address of globals out of a loop). If this is the case, we don't
3593 // want to forward-subst the cast.
3594 if (isa<Constant>(CI->getOperand(0)))
3597 // If this is a noop copy, sink it into user blocks to reduce the number
3598 // of virtual registers that must be created and coallesced.
3599 MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
3600 MVT::ValueType DstVT = TLI.getValueType(CI->getType());
3602 // This is an fp<->int conversion?
3603 if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT))
3606 // If this is an extension, it will be a zero or sign extension, which
3608 if (SrcVT < DstVT) continue;
3610 // If these values will be promoted, find out what they will be promoted
3611 // to. This helps us consider truncates on PPC as noop copies when they
3613 if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
3614 SrcVT = TLI.getTypeToTransformTo(SrcVT);
3615 if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
3616 DstVT = TLI.getTypeToTransformTo(DstVT);
3618 // If, after promotion, these are the same types, this is a noop copy.
3620 MadeChange |= OptimizeNoopCopyExpression(CI);
3626 FunctionLoweringInfo FuncInfo(TLI, Fn, MF);
3628 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
3629 SelectBasicBlock(I, MF, FuncInfo);
3634 SDOperand SelectionDAGLowering::CopyValueToVirtualRegister(Value *V,
3636 SDOperand Op = getValue(V);
3637 assert((Op.getOpcode() != ISD::CopyFromReg ||
3638 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
3639 "Copy from a reg to the same reg!");
3641 // If this type is not legal, we must make sure to not create an invalid
3643 MVT::ValueType SrcVT = Op.getValueType();
3644 MVT::ValueType DestVT = TLI.getTypeToTransformTo(SrcVT);
3645 if (SrcVT == DestVT) {
3646 return DAG.getCopyToReg(getRoot(), Reg, Op);
3647 } else if (SrcVT == MVT::Vector) {
3648 // Handle copies from generic vectors to registers.
3649 MVT::ValueType PTyElementVT, PTyLegalElementVT;
3650 unsigned NE = TLI.getPackedTypeBreakdown(cast<PackedType>(V->getType()),
3651 PTyElementVT, PTyLegalElementVT);
3653 // Insert a VBIT_CONVERT of the input vector to a "N x PTyElementVT"
3654 // MVT::Vector type.
3655 Op = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Op,
3656 DAG.getConstant(NE, MVT::i32),
3657 DAG.getValueType(PTyElementVT));
3659 // Loop over all of the elements of the resultant vector,
3660 // VEXTRACT_VECTOR_ELT'ing them, converting them to PTyLegalElementVT, then
3661 // copying them into output registers.
3662 SmallVector<SDOperand, 8> OutChains;
3663 SDOperand Root = getRoot();
3664 for (unsigned i = 0; i != NE; ++i) {
3665 SDOperand Elt = DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, PTyElementVT,
3666 Op, DAG.getConstant(i, TLI.getPointerTy()));
3667 if (PTyElementVT == PTyLegalElementVT) {
3668 // Elements are legal.
3669 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
3670 } else if (PTyLegalElementVT > PTyElementVT) {
3671 // Elements are promoted.
3672 if (MVT::isFloatingPoint(PTyLegalElementVT))
3673 Elt = DAG.getNode(ISD::FP_EXTEND, PTyLegalElementVT, Elt);
3675 Elt = DAG.getNode(ISD::ANY_EXTEND, PTyLegalElementVT, Elt);
3676 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
3678 // Elements are expanded.
3679 // The src value is expanded into multiple registers.
3680 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
3681 Elt, DAG.getConstant(0, TLI.getPointerTy()));
3682 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
3683 Elt, DAG.getConstant(1, TLI.getPointerTy()));
3684 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Lo));
3685 OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Hi));
3688 return DAG.getNode(ISD::TokenFactor, MVT::Other,
3689 &OutChains[0], OutChains.size());
3690 } else if (SrcVT < DestVT) {
3691 // The src value is promoted to the register.
3692 if (MVT::isFloatingPoint(SrcVT))
3693 Op = DAG.getNode(ISD::FP_EXTEND, DestVT, Op);
3695 Op = DAG.getNode(ISD::ANY_EXTEND, DestVT, Op);
3696 return DAG.getCopyToReg(getRoot(), Reg, Op);
3698 // The src value is expanded into multiple registers.
3699 SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
3700 Op, DAG.getConstant(0, TLI.getPointerTy()));
3701 SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
3702 Op, DAG.getConstant(1, TLI.getPointerTy()));
3703 Op = DAG.getCopyToReg(getRoot(), Reg, Lo);
3704 return DAG.getCopyToReg(Op, Reg+1, Hi);
3708 void SelectionDAGISel::
3709 LowerArguments(BasicBlock *BB, SelectionDAGLowering &SDL,
3710 std::vector<SDOperand> &UnorderedChains) {
3711 // If this is the entry block, emit arguments.
3712 Function &F = *BB->getParent();
3713 FunctionLoweringInfo &FuncInfo = SDL.FuncInfo;
3714 SDOperand OldRoot = SDL.DAG.getRoot();
3715 std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG);
3718 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
3720 if (!AI->use_empty()) {
3721 SDL.setValue(AI, Args[a]);
3723 // If this argument is live outside of the entry block, insert a copy from
3724 // whereever we got it to the vreg that other BB's will reference it as.
3725 if (FuncInfo.ValueMap.count(AI)) {
3727 SDL.CopyValueToVirtualRegister(AI, FuncInfo.ValueMap[AI]);
3728 UnorderedChains.push_back(Copy);
3732 // Finally, if the target has anything special to do, allow it to do so.
3733 // FIXME: this should insert code into the DAG!
3734 EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction());
3737 void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB,
3738 std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate,
3739 FunctionLoweringInfo &FuncInfo) {
3740 SelectionDAGLowering SDL(DAG, TLI, FuncInfo);
3742 std::vector<SDOperand> UnorderedChains;
3744 // Lower any arguments needed in this block if this is the entry block.
3745 if (LLVMBB == &LLVMBB->getParent()->front())
3746 LowerArguments(LLVMBB, SDL, UnorderedChains);
3748 BB = FuncInfo.MBBMap[LLVMBB];
3749 SDL.setCurrentBasicBlock(BB);
3751 // Lower all of the non-terminator instructions.
3752 for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end();
3756 // Ensure that all instructions which are used outside of their defining
3757 // blocks are available as virtual registers.
3758 for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I)
3759 if (!I->use_empty() && !isa<PHINode>(I)) {
3760 std::map<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I);
3761 if (VMI != FuncInfo.ValueMap.end())
3762 UnorderedChains.push_back(
3763 SDL.CopyValueToVirtualRegister(I, VMI->second));
3766 // Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
3767 // ensure constants are generated when needed. Remember the virtual registers
3768 // that need to be added to the Machine PHI nodes as input. We cannot just
3769 // directly add them, because expansion might result in multiple MBB's for one
3770 // BB. As such, the start of the BB might correspond to a different MBB than
3773 TerminatorInst *TI = LLVMBB->getTerminator();
3775 // Emit constants only once even if used by multiple PHI nodes.
3776 std::map<Constant*, unsigned> ConstantsOut;
3778 // Vector bool would be better, but vector<bool> is really slow.
3779 std::vector<unsigned char> SuccsHandled;
3780 if (TI->getNumSuccessors())
3781 SuccsHandled.resize(BB->getParent()->getNumBlockIDs());
3783 // Check successor nodes PHI nodes that expect a constant to be available from
3785 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
3786 BasicBlock *SuccBB = TI->getSuccessor(succ);
3787 if (!isa<PHINode>(SuccBB->begin())) continue;
3788 MachineBasicBlock *SuccMBB = FuncInfo.MBBMap[SuccBB];
3790 // If this terminator has multiple identical successors (common for
3791 // switches), only handle each succ once.
3792 unsigned SuccMBBNo = SuccMBB->getNumber();
3793 if (SuccsHandled[SuccMBBNo]) continue;
3794 SuccsHandled[SuccMBBNo] = true;
3796 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
3799 // At this point we know that there is a 1-1 correspondence between LLVM PHI
3800 // nodes and Machine PHI nodes, but the incoming operands have not been
3802 for (BasicBlock::iterator I = SuccBB->begin();
3803 (PN = dyn_cast<PHINode>(I)); ++I) {
3804 // Ignore dead phi's.
3805 if (PN->use_empty()) continue;
3808 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
3809 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
3810 unsigned &RegOut = ConstantsOut[C];
3812 RegOut = FuncInfo.CreateRegForValue(C);
3813 UnorderedChains.push_back(
3814 SDL.CopyValueToVirtualRegister(C, RegOut));
3818 Reg = FuncInfo.ValueMap[PHIOp];
3820 assert(isa<AllocaInst>(PHIOp) &&
3821 FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
3822 "Didn't codegen value into a register!??");
3823 Reg = FuncInfo.CreateRegForValue(PHIOp);
3824 UnorderedChains.push_back(
3825 SDL.CopyValueToVirtualRegister(PHIOp, Reg));
3829 // Remember that this register needs to added to the machine PHI node as
3830 // the input for this MBB.
3831 MVT::ValueType VT = TLI.getValueType(PN->getType());
3832 unsigned NumElements;
3833 if (VT != MVT::Vector)
3834 NumElements = TLI.getNumElements(VT);
3836 MVT::ValueType VT1,VT2;
3838 TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
3841 for (unsigned i = 0, e = NumElements; i != e; ++i)
3842 PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
3845 ConstantsOut.clear();
3847 // Turn all of the unordered chains into one factored node.
3848 if (!UnorderedChains.empty()) {
3849 SDOperand Root = SDL.getRoot();
3850 if (Root.getOpcode() != ISD::EntryToken) {
3851 unsigned i = 0, e = UnorderedChains.size();
3852 for (; i != e; ++i) {
3853 assert(UnorderedChains[i].Val->getNumOperands() > 1);
3854 if (UnorderedChains[i].Val->getOperand(0) == Root)
3855 break; // Don't add the root if we already indirectly depend on it.
3859 UnorderedChains.push_back(Root);
3861 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other,
3862 &UnorderedChains[0], UnorderedChains.size()));
3865 // Lower the terminator after the copies are emitted.
3866 SDL.visit(*LLVMBB->getTerminator());
3868 // Copy over any CaseBlock records that may now exist due to SwitchInst
3869 // lowering, as well as any jump table information.
3870 SwitchCases.clear();
3871 SwitchCases = SDL.SwitchCases;
3874 // Make sure the root of the DAG is up-to-date.
3875 DAG.setRoot(SDL.getRoot());
3878 void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) {
3879 // Get alias analysis for load/store combining.
3880 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
3882 // Run the DAG combiner in pre-legalize mode.
3883 DAG.Combine(false, AA);
3885 DEBUG(std::cerr << "Lowered selection DAG:\n");
3888 // Second step, hack on the DAG until it only uses operations and types that
3889 // the target supports.
3892 DEBUG(std::cerr << "Legalized selection DAG:\n");
3895 // Run the DAG combiner in post-legalize mode.
3896 DAG.Combine(true, AA);
3898 if (ViewISelDAGs) DAG.viewGraph();
3900 // Third, instruction select all of the operations to machine code, adding the
3901 // code to the MachineBasicBlock.
3902 InstructionSelectBasicBlock(DAG);
3904 DEBUG(std::cerr << "Selected machine code:\n");
3908 void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF,
3909 FunctionLoweringInfo &FuncInfo) {
3910 std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate;
3912 SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3915 // First step, lower LLVM code to some DAG. This DAG may use operations and
3916 // types that are not supported by the target.
3917 BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo);
3919 // Second step, emit the lowered DAG as machine code.
3920 CodeGenAndEmitDAG(DAG);
3923 // Next, now that we know what the last MBB the LLVM BB expanded is, update
3924 // PHI nodes in successors.
3925 if (SwitchCases.empty() && JT.Reg == 0) {
3926 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
3927 MachineInstr *PHI = PHINodesToUpdate[i].first;
3928 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3929 "This is not a machine PHI node that we are updating!");
3930 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
3931 PHI->addMachineBasicBlockOperand(BB);
3936 // If the JumpTable record is filled in, then we need to emit a jump table.
3937 // Updating the PHI nodes is tricky in this case, since we need to determine
3938 // whether the PHI is a successor of the range check MBB or the jump table MBB
3940 assert(SwitchCases.empty() && "Cannot have jump table and lowered switch");
3941 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3943 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
3944 MachineBasicBlock *RangeBB = BB;
3945 // Set the current basic block to the mbb we wish to insert the code into
3947 SDL.setCurrentBasicBlock(BB);
3949 SDL.visitJumpTable(JT);
3950 SDAG.setRoot(SDL.getRoot());
3951 CodeGenAndEmitDAG(SDAG);
3953 for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
3954 MachineInstr *PHI = PHINodesToUpdate[pi].first;
3955 MachineBasicBlock *PHIBB = PHI->getParent();
3956 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3957 "This is not a machine PHI node that we are updating!");
3958 if (PHIBB == JT.Default) {
3959 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
3960 PHI->addMachineBasicBlockOperand(RangeBB);
3962 if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
3963 PHI->addRegOperand(PHINodesToUpdate[pi].second, false);
3964 PHI->addMachineBasicBlockOperand(BB);
3970 // If the switch block involved a branch to one of the actual successors, we
3971 // need to update PHI nodes in that block.
3972 for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
3973 MachineInstr *PHI = PHINodesToUpdate[i].first;
3974 assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
3975 "This is not a machine PHI node that we are updating!");
3976 if (BB->isSuccessor(PHI->getParent())) {
3977 PHI->addRegOperand(PHINodesToUpdate[i].second, false);
3978 PHI->addMachineBasicBlockOperand(BB);
3982 // If we generated any switch lowering information, build and codegen any
3983 // additional DAGs necessary.
3984 for (unsigned i = 0, e = SwitchCases.size(); i != e; ++i) {
3985 SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
3987 SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
3989 // Set the current basic block to the mbb we wish to insert the code into
3990 BB = SwitchCases[i].ThisBB;
3991 SDL.setCurrentBasicBlock(BB);
3994 SDL.visitSwitchCase(SwitchCases[i]);
3995 SDAG.setRoot(SDL.getRoot());
3996 CodeGenAndEmitDAG(SDAG);
3998 // Handle any PHI nodes in successors of this chunk, as if we were coming
3999 // from the original BB before switch expansion. Note that PHI nodes can
4000 // occur multiple times in PHINodesToUpdate. We have to be very careful to
4001 // handle them the right number of times.
4002 while ((BB = SwitchCases[i].TrueBB)) { // Handle LHS and RHS.
4003 for (MachineBasicBlock::iterator Phi = BB->begin();
4004 Phi != BB->end() && Phi->getOpcode() == TargetInstrInfo::PHI; ++Phi){
4005 // This value for this PHI node is recorded in PHINodesToUpdate, get it.
4006 for (unsigned pn = 0; ; ++pn) {
4007 assert(pn != PHINodesToUpdate.size() && "Didn't find PHI entry!");
4008 if (PHINodesToUpdate[pn].first == Phi) {
4009 Phi->addRegOperand(PHINodesToUpdate[pn].second, false);
4010 Phi->addMachineBasicBlockOperand(SwitchCases[i].ThisBB);
4016 // Don't process RHS if same block as LHS.
4017 if (BB == SwitchCases[i].FalseBB)
4018 SwitchCases[i].FalseBB = 0;
4020 // If we haven't handled the RHS, do so now. Otherwise, we're done.
4021 SwitchCases[i].TrueBB = SwitchCases[i].FalseBB;
4022 SwitchCases[i].FalseBB = 0;
4024 assert(SwitchCases[i].TrueBB == 0 && SwitchCases[i].FalseBB == 0);
4029 //===----------------------------------------------------------------------===//
4030 /// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
4031 /// target node in the graph.
4032 void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) {
4033 if (ViewSchedDAGs) DAG.viewGraph();
4035 RegisterScheduler::FunctionPassCtor Ctor = RegisterScheduler::getDefault();
4039 RegisterScheduler::setDefault(Ctor);
4042 ScheduleDAG *SL = Ctor(this, &DAG, BB);
4048 HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
4049 return new HazardRecognizer();
4052 //===----------------------------------------------------------------------===//
4053 // Helper functions used by the generated instruction selector.
4054 //===----------------------------------------------------------------------===//
4055 // Calls to these methods are generated by tblgen.
4057 /// CheckAndMask - The isel is trying to match something like (and X, 255). If
4058 /// the dag combiner simplified the 255, we still want to match. RHS is the
4059 /// actual value in the DAG on the RHS of an AND, and DesiredMaskS is the value
4060 /// specified in the .td file (e.g. 255).
4061 bool SelectionDAGISel::CheckAndMask(SDOperand LHS, ConstantSDNode *RHS,
4062 int64_t DesiredMaskS) {
4063 uint64_t ActualMask = RHS->getValue();
4064 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
4066 // If the actual mask exactly matches, success!
4067 if (ActualMask == DesiredMask)
4070 // If the actual AND mask is allowing unallowed bits, this doesn't match.
4071 if (ActualMask & ~DesiredMask)
4074 // Otherwise, the DAG Combiner may have proven that the value coming in is
4075 // either already zero or is not demanded. Check for known zero input bits.
4076 uint64_t NeededMask = DesiredMask & ~ActualMask;
4077 if (getTargetLowering().MaskedValueIsZero(LHS, NeededMask))
4080 // TODO: check to see if missing bits are just not demanded.
4082 // Otherwise, this pattern doesn't match.
4086 /// CheckOrMask - The isel is trying to match something like (or X, 255). If
4087 /// the dag combiner simplified the 255, we still want to match. RHS is the
4088 /// actual value in the DAG on the RHS of an OR, and DesiredMaskS is the value
4089 /// specified in the .td file (e.g. 255).
4090 bool SelectionDAGISel::CheckOrMask(SDOperand LHS, ConstantSDNode *RHS,
4091 int64_t DesiredMaskS) {
4092 uint64_t ActualMask = RHS->getValue();
4093 uint64_t DesiredMask =DesiredMaskS & MVT::getIntVTBitMask(LHS.getValueType());
4095 // If the actual mask exactly matches, success!
4096 if (ActualMask == DesiredMask)
4099 // If the actual AND mask is allowing unallowed bits, this doesn't match.
4100 if (ActualMask & ~DesiredMask)
4103 // Otherwise, the DAG Combiner may have proven that the value coming in is
4104 // either already zero or is not demanded. Check for known zero input bits.
4105 uint64_t NeededMask = DesiredMask & ~ActualMask;
4107 uint64_t KnownZero, KnownOne;
4108 getTargetLowering().ComputeMaskedBits(LHS, NeededMask, KnownZero, KnownOne);
4110 // If all the missing bits in the or are already known to be set, match!
4111 if ((NeededMask & KnownOne) == NeededMask)
4114 // TODO: check to see if missing bits are just not demanded.
4116 // Otherwise, this pattern doesn't match.
4121 /// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
4122 /// by tblgen. Others should not call it.
4123 void SelectionDAGISel::
4124 SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) {
4125 std::vector<SDOperand> InOps;
4126 std::swap(InOps, Ops);
4128 Ops.push_back(InOps[0]); // input chain.
4129 Ops.push_back(InOps[1]); // input asm string.
4131 unsigned i = 2, e = InOps.size();
4132 if (InOps[e-1].getValueType() == MVT::Flag)
4133 --e; // Don't process a flag operand if it is here.
4136 unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
4137 if ((Flags & 7) != 4 /*MEM*/) {
4138 // Just skip over this operand, copying the operands verbatim.
4139 Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
4140 i += (Flags >> 3) + 1;
4142 assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
4143 // Otherwise, this is a memory operand. Ask the target to select it.
4144 std::vector<SDOperand> SelOps;
4145 if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) {
4146 std::cerr << "Could not match memory address. Inline asm failure!\n";
4150 // Add this to the output node.
4151 Ops.push_back(DAG.getConstant(4/*MEM*/ | (SelOps.size() << 3), MVT::i32));
4152 Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
4157 // Add the flag input back if present.
4158 if (e != InOps.size())
4159 Ops.push_back(InOps.back());