1 //===-- SelectionDAGBuild.cpp - Selection-DAG building --------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements routines for translating from LLVM IR into SelectionDAG IR.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "isel"
15 #include "SelectionDAGBuild.h"
16 #include "llvm/ADT/BitVector.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/Constants.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/Function.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/InlineAsm.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/IntrinsicInst.h"
28 #include "llvm/Module.h"
29 #include "llvm/CodeGen/FastISel.h"
30 #include "llvm/CodeGen/GCStrategy.h"
31 #include "llvm/CodeGen/GCMetadata.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineFrameInfo.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/CodeGen/SelectionDAG.h"
40 #include "llvm/CodeGen/DwarfWriter.h"
41 #include "llvm/Analysis/DebugInfo.h"
42 #include "llvm/Target/TargetRegisterInfo.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Target/TargetFrameInfo.h"
45 #include "llvm/Target/TargetInstrInfo.h"
46 #include "llvm/Target/TargetIntrinsicInfo.h"
47 #include "llvm/Target/TargetLowering.h"
48 #include "llvm/Target/TargetMachine.h"
49 #include "llvm/Target/TargetOptions.h"
50 #include "llvm/Support/Compiler.h"
51 #include "llvm/Support/CommandLine.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/MathExtras.h"
54 #include "llvm/Support/raw_ostream.h"
58 /// LimitFloatPrecision - Generate low-precision inline sequences for
59 /// some float libcalls (6, 8 or 12 bits).
60 static unsigned LimitFloatPrecision;
62 static cl::opt<unsigned, true>
63 LimitFPPrecision("limit-float-precision",
64 cl::desc("Generate low-precision inline sequences "
65 "for some float libcalls"),
66 cl::location(LimitFloatPrecision),
69 /// ComputeLinearIndex - Given an LLVM IR aggregate type and a sequence
70 /// of insertvalue or extractvalue indices that identify a member, return
71 /// the linearized index of the start of the member.
73 static unsigned ComputeLinearIndex(const TargetLowering &TLI, const Type *Ty,
74 const unsigned *Indices,
75 const unsigned *IndicesEnd,
76 unsigned CurIndex = 0) {
77 // Base case: We're done.
78 if (Indices && Indices == IndicesEnd)
81 // Given a struct type, recursively traverse the elements.
82 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
83 for (StructType::element_iterator EB = STy->element_begin(),
85 EE = STy->element_end();
87 if (Indices && *Indices == unsigned(EI - EB))
88 return ComputeLinearIndex(TLI, *EI, Indices+1, IndicesEnd, CurIndex);
89 CurIndex = ComputeLinearIndex(TLI, *EI, 0, 0, CurIndex);
93 // Given an array type, recursively traverse the elements.
94 else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
95 const Type *EltTy = ATy->getElementType();
96 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
97 if (Indices && *Indices == i)
98 return ComputeLinearIndex(TLI, EltTy, Indices+1, IndicesEnd, CurIndex);
99 CurIndex = ComputeLinearIndex(TLI, EltTy, 0, 0, CurIndex);
103 // We haven't found the type we're looking for, so keep searching.
107 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
108 /// MVTs that represent all the individual underlying
109 /// non-aggregate types that comprise it.
111 /// If Offsets is non-null, it points to a vector to be filled in
112 /// with the in-memory offsets of each of the individual values.
114 static void ComputeValueVTs(const TargetLowering &TLI, const Type *Ty,
115 SmallVectorImpl<MVT> &ValueVTs,
116 SmallVectorImpl<uint64_t> *Offsets = 0,
117 uint64_t StartingOffset = 0) {
118 // Given a struct type, recursively traverse the elements.
119 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
120 const StructLayout *SL = TLI.getTargetData()->getStructLayout(STy);
121 for (StructType::element_iterator EB = STy->element_begin(),
123 EE = STy->element_end();
125 ComputeValueVTs(TLI, *EI, ValueVTs, Offsets,
126 StartingOffset + SL->getElementOffset(EI - EB));
129 // Given an array type, recursively traverse the elements.
130 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
131 const Type *EltTy = ATy->getElementType();
132 uint64_t EltSize = TLI.getTargetData()->getTypePaddedSize(EltTy);
133 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
134 ComputeValueVTs(TLI, EltTy, ValueVTs, Offsets,
135 StartingOffset + i * EltSize);
138 // Base case: we can get an MVT for this LLVM IR type.
139 ValueVTs.push_back(TLI.getValueType(Ty));
141 Offsets->push_back(StartingOffset);
145 /// RegsForValue - This struct represents the registers (physical or virtual)
146 /// that a particular set of values is assigned, and the type information about
147 /// the value. The most common situation is to represent one value at a time,
148 /// but struct or array values are handled element-wise as multiple values.
149 /// The splitting of aggregates is performed recursively, so that we never
150 /// have aggregate-typed registers. The values at this point do not necessarily
151 /// have legal types, so each value may require one or more registers of some
154 struct VISIBILITY_HIDDEN RegsForValue {
155 /// TLI - The TargetLowering object.
157 const TargetLowering *TLI;
159 /// ValueVTs - The value types of the values, which may not be legal, and
160 /// may need be promoted or synthesized from one or more registers.
162 SmallVector<MVT, 4> ValueVTs;
164 /// RegVTs - The value types of the registers. This is the same size as
165 /// ValueVTs and it records, for each value, what the type of the assigned
166 /// register or registers are. (Individual values are never synthesized
167 /// from more than one type of register.)
169 /// With virtual registers, the contents of RegVTs is redundant with TLI's
170 /// getRegisterType member function, however when with physical registers
171 /// it is necessary to have a separate record of the types.
173 SmallVector<MVT, 4> RegVTs;
175 /// Regs - This list holds the registers assigned to the values.
176 /// Each legal or promoted value requires one register, and each
177 /// expanded value requires multiple registers.
179 SmallVector<unsigned, 4> Regs;
181 RegsForValue() : TLI(0) {}
183 RegsForValue(const TargetLowering &tli,
184 const SmallVector<unsigned, 4> ®s,
185 MVT regvt, MVT valuevt)
186 : TLI(&tli), ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
187 RegsForValue(const TargetLowering &tli,
188 const SmallVector<unsigned, 4> ®s,
189 const SmallVector<MVT, 4> ®vts,
190 const SmallVector<MVT, 4> &valuevts)
191 : TLI(&tli), ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {}
192 RegsForValue(const TargetLowering &tli,
193 unsigned Reg, const Type *Ty) : TLI(&tli) {
194 ComputeValueVTs(tli, Ty, ValueVTs);
196 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
197 MVT ValueVT = ValueVTs[Value];
198 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
199 MVT RegisterVT = TLI->getRegisterType(ValueVT);
200 for (unsigned i = 0; i != NumRegs; ++i)
201 Regs.push_back(Reg + i);
202 RegVTs.push_back(RegisterVT);
207 /// append - Add the specified values to this one.
208 void append(const RegsForValue &RHS) {
210 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
211 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
212 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
216 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
217 /// this value and returns the result as a ValueVTs value. This uses
218 /// Chain/Flag as the input and updates them for the output Chain/Flag.
219 /// If the Flag pointer is NULL, no flag is used.
220 SDValue getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl,
221 SDValue &Chain, SDValue *Flag) const;
223 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
224 /// specified value into the registers specified by this object. This uses
225 /// Chain/Flag as the input and updates them for the output Chain/Flag.
226 /// If the Flag pointer is NULL, no flag is used.
227 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
228 SDValue &Chain, SDValue *Flag) const;
230 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
231 /// operand list. This adds the code marker and includes the number of
232 /// values added into it.
233 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
234 std::vector<SDValue> &Ops) const;
238 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
239 /// PHI nodes or outside of the basic block that defines it, or used by a
240 /// switch or atomic instruction, which may expand to multiple basic blocks.
241 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
242 if (isa<PHINode>(I)) return true;
243 BasicBlock *BB = I->getParent();
244 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
245 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
246 // FIXME: Remove switchinst special case.
247 isa<SwitchInst>(*UI))
252 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
253 /// entry block, return true. This includes arguments used by switches, since
254 /// the switch may expand into multiple basic blocks.
255 static bool isOnlyUsedInEntryBlock(Argument *A, bool EnableFastISel) {
256 // With FastISel active, we may be splitting blocks, so force creation
257 // of virtual registers for all non-dead arguments.
258 // Don't force virtual registers for byval arguments though, because
259 // fast-isel can't handle those in all cases.
260 if (EnableFastISel && !A->hasByValAttr())
261 return A->use_empty();
263 BasicBlock *Entry = A->getParent()->begin();
264 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
265 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
266 return false; // Use not in entry block.
270 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli)
274 void FunctionLoweringInfo::set(Function &fn, MachineFunction &mf,
276 bool EnableFastISel) {
279 RegInfo = &MF->getRegInfo();
281 // Create a vreg for each argument register that is not dead and is used
282 // outside of the entry block for the function.
283 for (Function::arg_iterator AI = Fn->arg_begin(), E = Fn->arg_end();
285 if (!isOnlyUsedInEntryBlock(AI, EnableFastISel))
286 InitializeRegForValue(AI);
288 // Initialize the mapping of values to registers. This is only set up for
289 // instruction values that are used outside of the block that defines
291 Function::iterator BB = Fn->begin(), EB = Fn->end();
292 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
293 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
294 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
295 const Type *Ty = AI->getAllocatedType();
296 uint64_t TySize = TLI.getTargetData()->getTypePaddedSize(Ty);
298 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
301 TySize *= CUI->getZExtValue(); // Get total allocated size.
302 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
303 StaticAllocaMap[AI] =
304 MF->getFrameInfo()->CreateStackObject(TySize, Align);
307 for (; BB != EB; ++BB)
308 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
309 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
310 if (!isa<AllocaInst>(I) ||
311 !StaticAllocaMap.count(cast<AllocaInst>(I)))
312 InitializeRegForValue(I);
314 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
315 // also creates the initial PHI MachineInstrs, though none of the input
316 // operands are populated.
317 for (BB = Fn->begin(), EB = Fn->end(); BB != EB; ++BB) {
318 MachineBasicBlock *MBB = mf.CreateMachineBasicBlock(BB);
322 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
326 for (BasicBlock::iterator
327 I = BB->begin(), E = BB->end(); I != E; ++I) {
328 if (CallInst *CI = dyn_cast<CallInst>(I)) {
329 if (Function *F = CI->getCalledFunction()) {
330 switch (F->getIntrinsicID()) {
332 case Intrinsic::dbg_stoppoint: {
333 DwarfWriter *DW = DAG.getDwarfWriter();
334 DbgStopPointInst *SPI = cast<DbgStopPointInst>(I);
336 if (DW && DW->ValidDebugInfo(SPI->getContext())) {
337 DICompileUnit CU(cast<GlobalVariable>(SPI->getContext()));
339 unsigned SrcFile = DW->getOrCreateSourceID(CU.getDirectory(Dir),
341 unsigned idx = MF->getOrCreateDebugLocID(SrcFile,
344 DL = DebugLoc::get(idx);
349 case Intrinsic::dbg_func_start: {
350 DwarfWriter *DW = DAG.getDwarfWriter();
352 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
353 Value *SP = FSI->getSubprogram();
355 if (DW->ValidDebugInfo(SP)) {
356 DISubprogram Subprogram(cast<GlobalVariable>(SP));
357 DICompileUnit CU(Subprogram.getCompileUnit());
359 unsigned SrcFile = DW->getOrCreateSourceID(CU.getDirectory(Dir),
361 unsigned Line = Subprogram.getLineNumber();
362 DL = DebugLoc::get(MF->getOrCreateDebugLocID(SrcFile, Line, 0));
372 PN = dyn_cast<PHINode>(I);
373 if (!PN || PN->use_empty()) continue;
375 unsigned PHIReg = ValueMap[PN];
376 assert(PHIReg && "PHI node does not have an assigned virtual register!");
378 SmallVector<MVT, 4> ValueVTs;
379 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
380 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
381 MVT VT = ValueVTs[vti];
382 unsigned NumRegisters = TLI.getNumRegisters(VT);
383 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
384 for (unsigned i = 0; i != NumRegisters; ++i)
385 BuildMI(MBB, DL, TII->get(TargetInstrInfo::PHI), PHIReg + i);
386 PHIReg += NumRegisters;
392 unsigned FunctionLoweringInfo::MakeReg(MVT VT) {
393 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
396 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
397 /// the correctly promoted or expanded types. Assign these registers
398 /// consecutive vreg numbers and return the first assigned number.
400 /// In the case that the given value has struct or array type, this function
401 /// will assign registers for each member or element.
403 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
404 SmallVector<MVT, 4> ValueVTs;
405 ComputeValueVTs(TLI, V->getType(), ValueVTs);
407 unsigned FirstReg = 0;
408 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
409 MVT ValueVT = ValueVTs[Value];
410 MVT RegisterVT = TLI.getRegisterType(ValueVT);
412 unsigned NumRegs = TLI.getNumRegisters(ValueVT);
413 for (unsigned i = 0; i != NumRegs; ++i) {
414 unsigned R = MakeReg(RegisterVT);
415 if (!FirstReg) FirstReg = R;
421 /// getCopyFromParts - Create a value that contains the specified legal parts
422 /// combined into the value they represent. If the parts combine to a type
423 /// larger then ValueVT then AssertOp can be used to specify whether the extra
424 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
425 /// (ISD::AssertSext).
426 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl,
427 const SDValue *Parts,
428 unsigned NumParts, MVT PartVT, MVT ValueVT,
429 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
430 assert(NumParts > 0 && "No parts to assemble!");
431 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
432 SDValue Val = Parts[0];
435 // Assemble the value from multiple parts.
436 if (!ValueVT.isVector()) {
437 unsigned PartBits = PartVT.getSizeInBits();
438 unsigned ValueBits = ValueVT.getSizeInBits();
440 // Assemble the power of 2 part.
441 unsigned RoundParts = NumParts & (NumParts - 1) ?
442 1 << Log2_32(NumParts) : NumParts;
443 unsigned RoundBits = PartBits * RoundParts;
444 MVT RoundVT = RoundBits == ValueBits ?
445 ValueVT : MVT::getIntegerVT(RoundBits);
448 MVT HalfVT = ValueVT.isInteger() ?
449 MVT::getIntegerVT(RoundBits/2) :
450 MVT::getFloatingPointVT(RoundBits/2);
452 if (RoundParts > 2) {
453 Lo = getCopyFromParts(DAG, dl, Parts, RoundParts/2, PartVT, HalfVT);
454 Hi = getCopyFromParts(DAG, dl, Parts+RoundParts/2, RoundParts/2,
457 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]);
458 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]);
460 if (TLI.isBigEndian())
462 Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi);
464 if (RoundParts < NumParts) {
465 // Assemble the trailing non-power-of-2 part.
466 unsigned OddParts = NumParts - RoundParts;
467 MVT OddVT = MVT::getIntegerVT(OddParts * PartBits);
468 Hi = getCopyFromParts(DAG, dl,
469 Parts+RoundParts, OddParts, PartVT, OddVT);
471 // Combine the round and odd parts.
473 if (TLI.isBigEndian())
475 MVT TotalVT = MVT::getIntegerVT(NumParts * PartBits);
476 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi);
477 Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi,
478 DAG.getConstant(Lo.getValueType().getSizeInBits(),
479 TLI.getPointerTy()));
480 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo);
481 Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi);
484 // Handle a multi-element vector.
485 MVT IntermediateVT, RegisterVT;
486 unsigned NumIntermediates;
488 TLI.getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
490 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
491 NumParts = NumRegs; // Silence a compiler warning.
492 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
493 assert(RegisterVT == Parts[0].getValueType() &&
494 "Part type doesn't match part!");
496 // Assemble the parts into intermediate operands.
497 SmallVector<SDValue, 8> Ops(NumIntermediates);
498 if (NumIntermediates == NumParts) {
499 // If the register was not expanded, truncate or copy the value,
501 for (unsigned i = 0; i != NumParts; ++i)
502 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1,
503 PartVT, IntermediateVT);
504 } else if (NumParts > 0) {
505 // If the intermediate type was expanded, build the intermediate operands
507 assert(NumParts % NumIntermediates == 0 &&
508 "Must expand into a divisible number of parts!");
509 unsigned Factor = NumParts / NumIntermediates;
510 for (unsigned i = 0; i != NumIntermediates; ++i)
511 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor,
512 PartVT, IntermediateVT);
515 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the intermediate
517 Val = DAG.getNode(IntermediateVT.isVector() ?
518 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl,
519 ValueVT, &Ops[0], NumIntermediates);
523 // There is now one part, held in Val. Correct it to match ValueVT.
524 PartVT = Val.getValueType();
526 if (PartVT == ValueVT)
529 if (PartVT.isVector()) {
530 assert(ValueVT.isVector() && "Unknown vector conversion!");
531 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
534 if (ValueVT.isVector()) {
535 assert(ValueVT.getVectorElementType() == PartVT &&
536 ValueVT.getVectorNumElements() == 1 &&
537 "Only trivial scalar-to-vector conversions should get here!");
538 return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val);
541 if (PartVT.isInteger() &&
542 ValueVT.isInteger()) {
543 if (ValueVT.bitsLT(PartVT)) {
544 // For a truncate, see if we have any information to
545 // indicate whether the truncated bits will always be
546 // zero or sign-extension.
547 if (AssertOp != ISD::DELETED_NODE)
548 Val = DAG.getNode(AssertOp, dl, PartVT, Val,
549 DAG.getValueType(ValueVT));
550 return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
552 return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val);
556 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
557 if (ValueVT.bitsLT(Val.getValueType()))
558 // FP_ROUND's are always exact here.
559 return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val,
560 DAG.getIntPtrConstant(1));
561 return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val);
564 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
565 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
567 assert(0 && "Unknown mismatch!");
571 /// getCopyToParts - Create a series of nodes that contain the specified value
572 /// split into legal parts. If the parts contain more bits than Val, then, for
573 /// integers, ExtendKind can be used to specify how to generate the extra bits.
574 static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, SDValue Val,
575 SDValue *Parts, unsigned NumParts, MVT PartVT,
576 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
577 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
578 MVT PtrVT = TLI.getPointerTy();
579 MVT ValueVT = Val.getValueType();
580 unsigned PartBits = PartVT.getSizeInBits();
581 unsigned OrigNumParts = NumParts;
582 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
587 if (!ValueVT.isVector()) {
588 if (PartVT == ValueVT) {
589 assert(NumParts == 1 && "No-op copy with multiple parts!");
594 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
595 // If the parts cover more bits than the value has, promote the value.
596 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
597 assert(NumParts == 1 && "Do not know what to promote to!");
598 Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val);
599 } else if (PartVT.isInteger() && ValueVT.isInteger()) {
600 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
601 Val = DAG.getNode(ExtendKind, dl, ValueVT, Val);
603 assert(0 && "Unknown mismatch!");
605 } else if (PartBits == ValueVT.getSizeInBits()) {
606 // Different types of the same size.
607 assert(NumParts == 1 && PartVT != ValueVT);
608 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
609 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
610 // If the parts cover less bits than value has, truncate the value.
611 if (PartVT.isInteger() && ValueVT.isInteger()) {
612 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
613 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
615 assert(0 && "Unknown mismatch!");
619 // The value may have changed - recompute ValueVT.
620 ValueVT = Val.getValueType();
621 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
622 "Failed to tile the value with PartVT!");
625 assert(PartVT == ValueVT && "Type conversion failed!");
630 // Expand the value into multiple parts.
631 if (NumParts & (NumParts - 1)) {
632 // The number of parts is not a power of 2. Split off and copy the tail.
633 assert(PartVT.isInteger() && ValueVT.isInteger() &&
634 "Do not know what to expand to!");
635 unsigned RoundParts = 1 << Log2_32(NumParts);
636 unsigned RoundBits = RoundParts * PartBits;
637 unsigned OddParts = NumParts - RoundParts;
638 SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val,
639 DAG.getConstant(RoundBits,
640 TLI.getPointerTy()));
641 getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, OddParts, PartVT);
642 if (TLI.isBigEndian())
643 // The odd parts were reversed by getCopyToParts - unreverse them.
644 std::reverse(Parts + RoundParts, Parts + NumParts);
645 NumParts = RoundParts;
646 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
647 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
650 // The number of parts is a power of 2. Repeatedly bisect the value using
652 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl,
653 MVT::getIntegerVT(ValueVT.getSizeInBits()),
655 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
656 for (unsigned i = 0; i < NumParts; i += StepSize) {
657 unsigned ThisBits = StepSize * PartBits / 2;
658 MVT ThisVT = MVT::getIntegerVT (ThisBits);
659 SDValue &Part0 = Parts[i];
660 SDValue &Part1 = Parts[i+StepSize/2];
662 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
664 DAG.getConstant(1, PtrVT));
665 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
667 DAG.getConstant(0, PtrVT));
669 if (ThisBits == PartBits && ThisVT != PartVT) {
670 Part0 = DAG.getNode(ISD::BIT_CONVERT, dl,
672 Part1 = DAG.getNode(ISD::BIT_CONVERT, dl,
678 if (TLI.isBigEndian())
679 std::reverse(Parts, Parts + OrigNumParts);
686 if (PartVT != ValueVT) {
687 if (PartVT.isVector()) {
688 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
690 assert(ValueVT.getVectorElementType() == PartVT &&
691 ValueVT.getVectorNumElements() == 1 &&
692 "Only trivial vector-to-scalar conversions should get here!");
693 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
695 DAG.getConstant(0, PtrVT));
703 // Handle a multi-element vector.
704 MVT IntermediateVT, RegisterVT;
705 unsigned NumIntermediates;
706 unsigned NumRegs = TLI
707 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
709 unsigned NumElements = ValueVT.getVectorNumElements();
711 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
712 NumParts = NumRegs; // Silence a compiler warning.
713 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
715 // Split the vector into intermediate operands.
716 SmallVector<SDValue, 8> Ops(NumIntermediates);
717 for (unsigned i = 0; i != NumIntermediates; ++i)
718 if (IntermediateVT.isVector())
719 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,
721 DAG.getConstant(i * (NumElements / NumIntermediates),
724 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
726 DAG.getConstant(i, PtrVT));
728 // Split the intermediate operands into legal parts.
729 if (NumParts == NumIntermediates) {
730 // If the register was not expanded, promote or copy the value,
732 for (unsigned i = 0; i != NumParts; ++i)
733 getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT);
734 } else if (NumParts > 0) {
735 // If the intermediate type was expanded, split each the value into
737 assert(NumParts % NumIntermediates == 0 &&
738 "Must expand into a divisible number of parts!");
739 unsigned Factor = NumParts / NumIntermediates;
740 for (unsigned i = 0; i != NumIntermediates; ++i)
741 getCopyToParts(DAG, dl, Ops[i], &Parts[i * Factor], Factor, PartVT);
746 void SelectionDAGLowering::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
749 TD = DAG.getTarget().getTargetData();
752 /// clear - Clear out the curret SelectionDAG and the associated
753 /// state and prepare this SelectionDAGLowering object to be used
754 /// for a new block. This doesn't clear out information about
755 /// additional blocks that are needed to complete switch lowering
756 /// or PHI node updating; that information is cleared out as it is
758 void SelectionDAGLowering::clear() {
760 PendingLoads.clear();
761 PendingExports.clear();
763 CurDebugLoc = DebugLoc::getUnknownLoc();
766 /// getRoot - Return the current virtual root of the Selection DAG,
767 /// flushing any PendingLoad items. This must be done before emitting
768 /// a store or any other node that may need to be ordered after any
769 /// prior load instructions.
771 SDValue SelectionDAGLowering::getRoot() {
772 if (PendingLoads.empty())
773 return DAG.getRoot();
775 if (PendingLoads.size() == 1) {
776 SDValue Root = PendingLoads[0];
778 PendingLoads.clear();
782 // Otherwise, we have to make a token factor node.
783 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
784 &PendingLoads[0], PendingLoads.size());
785 PendingLoads.clear();
790 /// getControlRoot - Similar to getRoot, but instead of flushing all the
791 /// PendingLoad items, flush all the PendingExports items. It is necessary
792 /// to do this before emitting a terminator instruction.
794 SDValue SelectionDAGLowering::getControlRoot() {
795 SDValue Root = DAG.getRoot();
797 if (PendingExports.empty())
800 // Turn all of the CopyToReg chains into one factored node.
801 if (Root.getOpcode() != ISD::EntryToken) {
802 unsigned i = 0, e = PendingExports.size();
803 for (; i != e; ++i) {
804 assert(PendingExports[i].getNode()->getNumOperands() > 1);
805 if (PendingExports[i].getNode()->getOperand(0) == Root)
806 break; // Don't add the root if we already indirectly depend on it.
810 PendingExports.push_back(Root);
813 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
815 PendingExports.size());
816 PendingExports.clear();
821 void SelectionDAGLowering::visit(Instruction &I) {
822 visit(I.getOpcode(), I);
825 void SelectionDAGLowering::visit(unsigned Opcode, User &I) {
826 // Note: this doesn't use InstVisitor, because it has to work with
827 // ConstantExpr's in addition to instructions.
829 default: assert(0 && "Unknown instruction type encountered!");
831 // Build the switch statement using the Instruction.def file.
832 #define HANDLE_INST(NUM, OPCODE, CLASS) \
833 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
834 #include "llvm/Instruction.def"
838 void SelectionDAGLowering::visitAdd(User &I) {
839 if (I.getType()->isFPOrFPVector())
840 visitBinary(I, ISD::FADD);
842 visitBinary(I, ISD::ADD);
845 void SelectionDAGLowering::visitMul(User &I) {
846 if (I.getType()->isFPOrFPVector())
847 visitBinary(I, ISD::FMUL);
849 visitBinary(I, ISD::MUL);
852 SDValue SelectionDAGLowering::getValue(const Value *V) {
853 SDValue &N = NodeMap[V];
854 if (N.getNode()) return N;
856 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
857 MVT VT = TLI.getValueType(V->getType(), true);
859 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
860 return N = DAG.getConstant(*CI, VT);
862 if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
863 return N = DAG.getGlobalAddress(GV, VT);
865 if (isa<ConstantPointerNull>(C))
866 return N = DAG.getConstant(0, TLI.getPointerTy());
868 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C))
869 return N = DAG.getConstantFP(*CFP, VT);
871 if (isa<UndefValue>(C) && !isa<VectorType>(V->getType()) &&
872 !V->getType()->isAggregateType())
873 return N = DAG.getUNDEF(VT);
875 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
876 visit(CE->getOpcode(), *CE);
877 SDValue N1 = NodeMap[V];
878 assert(N1.getNode() && "visit didn't populate the ValueMap!");
882 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
883 SmallVector<SDValue, 4> Constants;
884 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
886 SDNode *Val = getValue(*OI).getNode();
887 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
888 Constants.push_back(SDValue(Val, i));
890 return DAG.getMergeValues(&Constants[0], Constants.size(),
894 if (isa<StructType>(C->getType()) || isa<ArrayType>(C->getType())) {
895 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
896 "Unknown struct or array constant!");
898 SmallVector<MVT, 4> ValueVTs;
899 ComputeValueVTs(TLI, C->getType(), ValueVTs);
900 unsigned NumElts = ValueVTs.size();
902 return SDValue(); // empty struct
903 SmallVector<SDValue, 4> Constants(NumElts);
904 for (unsigned i = 0; i != NumElts; ++i) {
905 MVT EltVT = ValueVTs[i];
906 if (isa<UndefValue>(C))
907 Constants[i] = DAG.getUNDEF(EltVT);
908 else if (EltVT.isFloatingPoint())
909 Constants[i] = DAG.getConstantFP(0, EltVT);
911 Constants[i] = DAG.getConstant(0, EltVT);
913 return DAG.getMergeValues(&Constants[0], NumElts, getCurDebugLoc());
916 const VectorType *VecTy = cast<VectorType>(V->getType());
917 unsigned NumElements = VecTy->getNumElements();
919 // Now that we know the number and type of the elements, get that number of
920 // elements into the Ops array based on what kind of constant it is.
921 SmallVector<SDValue, 16> Ops;
922 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
923 for (unsigned i = 0; i != NumElements; ++i)
924 Ops.push_back(getValue(CP->getOperand(i)));
926 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
927 "Unknown vector constant!");
928 MVT EltVT = TLI.getValueType(VecTy->getElementType());
931 if (isa<UndefValue>(C))
932 Op = DAG.getUNDEF(EltVT);
933 else if (EltVT.isFloatingPoint())
934 Op = DAG.getConstantFP(0, EltVT);
936 Op = DAG.getConstant(0, EltVT);
937 Ops.assign(NumElements, Op);
940 // Create a BUILD_VECTOR node.
941 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
942 VT, &Ops[0], Ops.size());
945 // If this is a static alloca, generate it as the frameindex instead of
947 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
948 DenseMap<const AllocaInst*, int>::iterator SI =
949 FuncInfo.StaticAllocaMap.find(AI);
950 if (SI != FuncInfo.StaticAllocaMap.end())
951 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
954 unsigned InReg = FuncInfo.ValueMap[V];
955 assert(InReg && "Value not in map!");
957 RegsForValue RFV(TLI, InReg, V->getType());
958 SDValue Chain = DAG.getEntryNode();
959 return RFV.getCopyFromRegs(DAG, getCurDebugLoc(), Chain, NULL);
963 void SelectionDAGLowering::visitRet(ReturnInst &I) {
964 if (I.getNumOperands() == 0) {
965 DAG.setRoot(DAG.getNode(ISD::RET, getCurDebugLoc(),
966 MVT::Other, getControlRoot()));
970 SmallVector<SDValue, 8> NewValues;
971 NewValues.push_back(getControlRoot());
972 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
973 SmallVector<MVT, 4> ValueVTs;
974 ComputeValueVTs(TLI, I.getOperand(i)->getType(), ValueVTs);
975 unsigned NumValues = ValueVTs.size();
976 if (NumValues == 0) continue;
978 SDValue RetOp = getValue(I.getOperand(i));
979 for (unsigned j = 0, f = NumValues; j != f; ++j) {
980 MVT VT = ValueVTs[j];
982 // FIXME: C calling convention requires the return type to be promoted to
983 // at least 32-bit. But this is not necessary for non-C calling
985 if (VT.isInteger()) {
986 MVT MinVT = TLI.getRegisterType(MVT::i32);
987 if (VT.bitsLT(MinVT))
991 unsigned NumParts = TLI.getNumRegisters(VT);
992 MVT PartVT = TLI.getRegisterType(VT);
993 SmallVector<SDValue, 4> Parts(NumParts);
994 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
996 const Function *F = I.getParent()->getParent();
997 if (F->paramHasAttr(0, Attribute::SExt))
998 ExtendKind = ISD::SIGN_EXTEND;
999 else if (F->paramHasAttr(0, Attribute::ZExt))
1000 ExtendKind = ISD::ZERO_EXTEND;
1002 getCopyToParts(DAG, getCurDebugLoc(),
1003 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
1004 &Parts[0], NumParts, PartVT, ExtendKind);
1006 // 'inreg' on function refers to return value
1007 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
1008 if (F->paramHasAttr(0, Attribute::InReg))
1010 for (unsigned i = 0; i < NumParts; ++i) {
1011 NewValues.push_back(Parts[i]);
1012 NewValues.push_back(DAG.getArgFlags(Flags));
1016 DAG.setRoot(DAG.getNode(ISD::RET, getCurDebugLoc(), MVT::Other,
1017 &NewValues[0], NewValues.size()));
1020 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1021 /// the current basic block, add it to ValueMap now so that we'll get a
1023 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
1024 // No need to export constants.
1025 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1027 // Already exported?
1028 if (FuncInfo.isExportedInst(V)) return;
1030 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1031 CopyValueToVirtualRegister(V, Reg);
1034 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
1035 const BasicBlock *FromBB) {
1036 // The operands of the setcc have to be in this block. We don't know
1037 // how to export them from some other block.
1038 if (Instruction *VI = dyn_cast<Instruction>(V)) {
1039 // Can export from current BB.
1040 if (VI->getParent() == FromBB)
1043 // Is already exported, noop.
1044 return FuncInfo.isExportedInst(V);
1047 // If this is an argument, we can export it if the BB is the entry block or
1048 // if it is already exported.
1049 if (isa<Argument>(V)) {
1050 if (FromBB == &FromBB->getParent()->getEntryBlock())
1053 // Otherwise, can only export this if it is already exported.
1054 return FuncInfo.isExportedInst(V);
1057 // Otherwise, constants can always be exported.
1061 static bool InBlock(const Value *V, const BasicBlock *BB) {
1062 if (const Instruction *I = dyn_cast<Instruction>(V))
1063 return I->getParent() == BB;
1067 /// getFCmpCondCode - Return the ISD condition code corresponding to
1068 /// the given LLVM IR floating-point condition code. This includes
1069 /// consideration of global floating-point math flags.
1071 static ISD::CondCode getFCmpCondCode(FCmpInst::Predicate Pred) {
1072 ISD::CondCode FPC, FOC;
1074 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1075 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1076 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1077 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1078 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1079 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1080 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1081 case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
1082 case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
1083 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1084 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1085 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1086 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1087 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1088 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1089 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1091 assert(0 && "Invalid FCmp predicate opcode!");
1092 FOC = FPC = ISD::SETFALSE;
1095 if (FiniteOnlyFPMath())
1101 /// getICmpCondCode - Return the ISD condition code corresponding to
1102 /// the given LLVM IR integer condition code.
1104 static ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred) {
1106 case ICmpInst::ICMP_EQ: return ISD::SETEQ;
1107 case ICmpInst::ICMP_NE: return ISD::SETNE;
1108 case ICmpInst::ICMP_SLE: return ISD::SETLE;
1109 case ICmpInst::ICMP_ULE: return ISD::SETULE;
1110 case ICmpInst::ICMP_SGE: return ISD::SETGE;
1111 case ICmpInst::ICMP_UGE: return ISD::SETUGE;
1112 case ICmpInst::ICMP_SLT: return ISD::SETLT;
1113 case ICmpInst::ICMP_ULT: return ISD::SETULT;
1114 case ICmpInst::ICMP_SGT: return ISD::SETGT;
1115 case ICmpInst::ICMP_UGT: return ISD::SETUGT;
1117 assert(0 && "Invalid ICmp predicate opcode!");
1122 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1123 /// This function emits a branch and is used at the leaves of an OR or an
1124 /// AND operator tree.
1127 SelectionDAGLowering::EmitBranchForMergedCondition(Value *Cond,
1128 MachineBasicBlock *TBB,
1129 MachineBasicBlock *FBB,
1130 MachineBasicBlock *CurBB) {
1131 const BasicBlock *BB = CurBB->getBasicBlock();
1133 // If the leaf of the tree is a comparison, merge the condition into
1135 if (CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1136 // The operands of the cmp have to be in this block. We don't know
1137 // how to export them from some other block. If this is the first block
1138 // of the sequence, no exporting is needed.
1139 if (CurBB == CurMBB ||
1140 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1141 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1142 ISD::CondCode Condition;
1143 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1144 Condition = getICmpCondCode(IC->getPredicate());
1145 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1146 Condition = getFCmpCondCode(FC->getPredicate());
1148 Condition = ISD::SETEQ; // silence warning.
1149 assert(0 && "Unknown compare instruction");
1152 CaseBlock CB(Condition, BOp->getOperand(0),
1153 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1154 SwitchCases.push_back(CB);
1159 // Create a CaseBlock record representing this branch.
1160 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(),
1161 NULL, TBB, FBB, CurBB);
1162 SwitchCases.push_back(CB);
1165 /// FindMergedConditions - If Cond is an expression like
1166 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
1167 MachineBasicBlock *TBB,
1168 MachineBasicBlock *FBB,
1169 MachineBasicBlock *CurBB,
1171 // If this node is not part of the or/and tree, emit it as a branch.
1172 Instruction *BOp = dyn_cast<Instruction>(Cond);
1173 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1174 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1175 BOp->getParent() != CurBB->getBasicBlock() ||
1176 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1177 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1178 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB);
1182 // Create TmpBB after CurBB.
1183 MachineFunction::iterator BBI = CurBB;
1184 MachineFunction &MF = DAG.getMachineFunction();
1185 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1186 CurBB->getParent()->insert(++BBI, TmpBB);
1188 if (Opc == Instruction::Or) {
1189 // Codegen X | Y as:
1197 // Emit the LHS condition.
1198 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1200 // Emit the RHS condition into TmpBB.
1201 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1203 assert(Opc == Instruction::And && "Unknown merge op!");
1204 // Codegen X & Y as:
1211 // This requires creation of TmpBB after CurBB.
1213 // Emit the LHS condition.
1214 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1216 // Emit the RHS condition into TmpBB.
1217 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1221 /// If the set of cases should be emitted as a series of branches, return true.
1222 /// If we should emit this as a bunch of and/or'd together conditions, return
1225 SelectionDAGLowering::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1226 if (Cases.size() != 2) return true;
1228 // If this is two comparisons of the same values or'd or and'd together, they
1229 // will get folded into a single comparison, so don't emit two blocks.
1230 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1231 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1232 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1233 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1240 void SelectionDAGLowering::visitBr(BranchInst &I) {
1241 // Update machine-CFG edges.
1242 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1244 // Figure out which block is immediately after the current one.
1245 MachineBasicBlock *NextBlock = 0;
1246 MachineFunction::iterator BBI = CurMBB;
1247 if (++BBI != CurMBB->getParent()->end())
1250 if (I.isUnconditional()) {
1251 // Update machine-CFG edges.
1252 CurMBB->addSuccessor(Succ0MBB);
1254 // If this is not a fall-through branch, emit the branch.
1255 if (Succ0MBB != NextBlock)
1256 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1257 MVT::Other, getControlRoot(),
1258 DAG.getBasicBlock(Succ0MBB)));
1262 // If this condition is one of the special cases we handle, do special stuff
1264 Value *CondVal = I.getCondition();
1265 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1267 // If this is a series of conditions that are or'd or and'd together, emit
1268 // this as a sequence of branches instead of setcc's with and/or operations.
1269 // For example, instead of something like:
1282 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1283 if (BOp->hasOneUse() &&
1284 (BOp->getOpcode() == Instruction::And ||
1285 BOp->getOpcode() == Instruction::Or)) {
1286 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1287 // If the compares in later blocks need to use values not currently
1288 // exported from this block, export them now. This block should always
1289 // be the first entry.
1290 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1292 // Allow some cases to be rejected.
1293 if (ShouldEmitAsBranches(SwitchCases)) {
1294 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1295 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1296 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1299 // Emit the branch for this block.
1300 visitSwitchCase(SwitchCases[0]);
1301 SwitchCases.erase(SwitchCases.begin());
1305 // Okay, we decided not to do this, remove any inserted MBB's and clear
1307 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1308 CurMBB->getParent()->erase(SwitchCases[i].ThisBB);
1310 SwitchCases.clear();
1314 // Create a CaseBlock record representing this branch.
1315 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(),
1316 NULL, Succ0MBB, Succ1MBB, CurMBB);
1317 // Use visitSwitchCase to actually insert the fast branch sequence for this
1319 visitSwitchCase(CB);
1322 /// visitSwitchCase - Emits the necessary code to represent a single node in
1323 /// the binary search tree resulting from lowering a switch instruction.
1324 void SelectionDAGLowering::visitSwitchCase(CaseBlock &CB) {
1326 SDValue CondLHS = getValue(CB.CmpLHS);
1327 DebugLoc dl = getCurDebugLoc();
1329 // Build the setcc now.
1330 if (CB.CmpMHS == NULL) {
1331 // Fold "(X == true)" to X and "(X == false)" to !X to
1332 // handle common cases produced by branch lowering.
1333 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ)
1335 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) {
1336 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1337 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1339 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1341 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1343 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1344 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1346 SDValue CmpOp = getValue(CB.CmpMHS);
1347 MVT VT = CmpOp.getValueType();
1349 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1350 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1353 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1354 VT, CmpOp, DAG.getConstant(Low, VT));
1355 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1356 DAG.getConstant(High-Low, VT), ISD::SETULE);
1360 // Update successor info
1361 CurMBB->addSuccessor(CB.TrueBB);
1362 CurMBB->addSuccessor(CB.FalseBB);
1364 // Set NextBlock to be the MBB immediately after the current one, if any.
1365 // This is used to avoid emitting unnecessary branches to the next block.
1366 MachineBasicBlock *NextBlock = 0;
1367 MachineFunction::iterator BBI = CurMBB;
1368 if (++BBI != CurMBB->getParent()->end())
1371 // If the lhs block is the next block, invert the condition so that we can
1372 // fall through to the lhs instead of the rhs block.
1373 if (CB.TrueBB == NextBlock) {
1374 std::swap(CB.TrueBB, CB.FalseBB);
1375 SDValue True = DAG.getConstant(1, Cond.getValueType());
1376 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1378 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1379 MVT::Other, getControlRoot(), Cond,
1380 DAG.getBasicBlock(CB.TrueBB));
1382 // If the branch was constant folded, fix up the CFG.
1383 if (BrCond.getOpcode() == ISD::BR) {
1384 CurMBB->removeSuccessor(CB.FalseBB);
1385 DAG.setRoot(BrCond);
1387 // Otherwise, go ahead and insert the false branch.
1388 if (BrCond == getControlRoot())
1389 CurMBB->removeSuccessor(CB.TrueBB);
1391 if (CB.FalseBB == NextBlock)
1392 DAG.setRoot(BrCond);
1394 DAG.setRoot(DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1395 DAG.getBasicBlock(CB.FalseBB)));
1399 /// visitJumpTable - Emit JumpTable node in the current MBB
1400 void SelectionDAGLowering::visitJumpTable(JumpTable &JT) {
1401 // Emit the code for the jump table
1402 assert(JT.Reg != -1U && "Should lower JT Header first!");
1403 MVT PTy = TLI.getPointerTy();
1404 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1406 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1407 DAG.setRoot(DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1408 MVT::Other, Index.getValue(1),
1412 /// visitJumpTableHeader - This function emits necessary code to produce index
1413 /// in the JumpTable from switch case.
1414 void SelectionDAGLowering::visitJumpTableHeader(JumpTable &JT,
1415 JumpTableHeader &JTH) {
1416 // Subtract the lowest switch case value from the value being switched on and
1417 // conditional branch to default mbb if the result is greater than the
1418 // difference between smallest and largest cases.
1419 SDValue SwitchOp = getValue(JTH.SValue);
1420 MVT VT = SwitchOp.getValueType();
1421 SDValue SUB = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1422 DAG.getConstant(JTH.First, VT));
1424 // The SDNode we just created, which holds the value being switched on minus
1425 // the the smallest case value, needs to be copied to a virtual register so it
1426 // can be used as an index into the jump table in a subsequent basic block.
1427 // This value may be smaller or larger than the target's pointer type, and
1428 // therefore require extension or truncating.
1429 if (VT.bitsGT(TLI.getPointerTy()))
1430 SwitchOp = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
1431 TLI.getPointerTy(), SUB);
1433 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
1434 TLI.getPointerTy(), SUB);
1436 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1437 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1438 JumpTableReg, SwitchOp);
1439 JT.Reg = JumpTableReg;
1441 // Emit the range check for the jump table, and branch to the default block
1442 // for the switch statement if the value being switched on exceeds the largest
1443 // case in the switch.
1444 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1445 TLI.getSetCCResultType(SUB.getValueType()), SUB,
1446 DAG.getConstant(JTH.Last-JTH.First,VT),
1449 // Set NextBlock to be the MBB immediately after the current one, if any.
1450 // This is used to avoid emitting unnecessary branches to the next block.
1451 MachineBasicBlock *NextBlock = 0;
1452 MachineFunction::iterator BBI = CurMBB;
1453 if (++BBI != CurMBB->getParent()->end())
1456 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1457 MVT::Other, CopyTo, CMP,
1458 DAG.getBasicBlock(JT.Default));
1460 if (JT.MBB == NextBlock)
1461 DAG.setRoot(BrCond);
1463 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1464 DAG.getBasicBlock(JT.MBB)));
1467 /// visitBitTestHeader - This function emits necessary code to produce value
1468 /// suitable for "bit tests"
1469 void SelectionDAGLowering::visitBitTestHeader(BitTestBlock &B) {
1470 // Subtract the minimum value
1471 SDValue SwitchOp = getValue(B.SValue);
1472 MVT VT = SwitchOp.getValueType();
1473 SDValue SUB = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1474 DAG.getConstant(B.First, VT));
1477 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1478 TLI.getSetCCResultType(SUB.getValueType()),
1479 SUB, DAG.getConstant(B.Range, VT),
1483 if (VT.bitsGT(TLI.getPointerTy()))
1484 ShiftOp = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
1485 TLI.getPointerTy(), SUB);
1487 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
1488 TLI.getPointerTy(), SUB);
1490 B.Reg = FuncInfo.MakeReg(TLI.getPointerTy());
1491 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1494 // Set NextBlock to be the MBB immediately after the current one, if any.
1495 // This is used to avoid emitting unnecessary branches to the next block.
1496 MachineBasicBlock *NextBlock = 0;
1497 MachineFunction::iterator BBI = CurMBB;
1498 if (++BBI != CurMBB->getParent()->end())
1501 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1503 CurMBB->addSuccessor(B.Default);
1504 CurMBB->addSuccessor(MBB);
1506 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1507 MVT::Other, CopyTo, RangeCmp,
1508 DAG.getBasicBlock(B.Default));
1510 if (MBB == NextBlock)
1511 DAG.setRoot(BrRange);
1513 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1514 DAG.getBasicBlock(MBB)));
1517 /// visitBitTestCase - this function produces one "bit test"
1518 void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB,
1521 // Make desired shift
1522 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1523 TLI.getPointerTy());
1524 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1526 DAG.getConstant(1, TLI.getPointerTy()),
1529 // Emit bit tests and jumps
1530 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1531 TLI.getPointerTy(), SwitchVal,
1532 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1533 SDValue AndCmp = DAG.getSetCC(getCurDebugLoc(),
1534 TLI.getSetCCResultType(AndOp.getValueType()),
1535 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1538 CurMBB->addSuccessor(B.TargetBB);
1539 CurMBB->addSuccessor(NextMBB);
1541 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1542 MVT::Other, getControlRoot(),
1543 AndCmp, DAG.getBasicBlock(B.TargetBB));
1545 // Set NextBlock to be the MBB immediately after the current one, if any.
1546 // This is used to avoid emitting unnecessary branches to the next block.
1547 MachineBasicBlock *NextBlock = 0;
1548 MachineFunction::iterator BBI = CurMBB;
1549 if (++BBI != CurMBB->getParent()->end())
1552 if (NextMBB == NextBlock)
1555 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1556 DAG.getBasicBlock(NextMBB)));
1559 void SelectionDAGLowering::visitInvoke(InvokeInst &I) {
1560 // Retrieve successors.
1561 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1562 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1564 const Value *Callee(I.getCalledValue());
1565 if (isa<InlineAsm>(Callee))
1568 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1570 // If the value of the invoke is used outside of its defining block, make it
1571 // available as a virtual register.
1572 if (!I.use_empty()) {
1573 DenseMap<const Value*, unsigned>::iterator VMI = FuncInfo.ValueMap.find(&I);
1574 if (VMI != FuncInfo.ValueMap.end())
1575 CopyValueToVirtualRegister(&I, VMI->second);
1578 // Update successor info
1579 CurMBB->addSuccessor(Return);
1580 CurMBB->addSuccessor(LandingPad);
1582 // Drop into normal successor.
1583 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1584 MVT::Other, getControlRoot(),
1585 DAG.getBasicBlock(Return)));
1588 void SelectionDAGLowering::visitUnwind(UnwindInst &I) {
1591 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1592 /// small case ranges).
1593 bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR,
1594 CaseRecVector& WorkList,
1596 MachineBasicBlock* Default) {
1597 Case& BackCase = *(CR.Range.second-1);
1599 // Size is the number of Cases represented by this range.
1600 size_t Size = CR.Range.second - CR.Range.first;
1604 // Get the MachineFunction which holds the current MBB. This is used when
1605 // inserting any additional MBBs necessary to represent the switch.
1606 MachineFunction *CurMF = CurMBB->getParent();
1608 // Figure out which block is immediately after the current one.
1609 MachineBasicBlock *NextBlock = 0;
1610 MachineFunction::iterator BBI = CR.CaseBB;
1612 if (++BBI != CurMBB->getParent()->end())
1615 // TODO: If any two of the cases has the same destination, and if one value
1616 // is the same as the other, but has one bit unset that the other has set,
1617 // use bit manipulation to do two compares at once. For example:
1618 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1620 // Rearrange the case blocks so that the last one falls through if possible.
1621 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1622 // The last case block won't fall through into 'NextBlock' if we emit the
1623 // branches in this order. See if rearranging a case value would help.
1624 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1625 if (I->BB == NextBlock) {
1626 std::swap(*I, BackCase);
1632 // Create a CaseBlock record representing a conditional branch to
1633 // the Case's target mbb if the value being switched on SV is equal
1635 MachineBasicBlock *CurBlock = CR.CaseBB;
1636 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1637 MachineBasicBlock *FallThrough;
1639 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1640 CurMF->insert(BBI, FallThrough);
1642 // If the last case doesn't match, go to the default block.
1643 FallThrough = Default;
1646 Value *RHS, *LHS, *MHS;
1648 if (I->High == I->Low) {
1649 // This is just small small case range :) containing exactly 1 case
1651 LHS = SV; RHS = I->High; MHS = NULL;
1654 LHS = I->Low; MHS = SV; RHS = I->High;
1656 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1658 // If emitting the first comparison, just call visitSwitchCase to emit the
1659 // code into the current block. Otherwise, push the CaseBlock onto the
1660 // vector to be later processed by SDISel, and insert the node's MBB
1661 // before the next MBB.
1662 if (CurBlock == CurMBB)
1663 visitSwitchCase(CB);
1665 SwitchCases.push_back(CB);
1667 CurBlock = FallThrough;
1673 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1674 return !DisableJumpTables &&
1675 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1676 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1679 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1680 APInt LastExt(Last), FirstExt(First);
1681 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1682 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1683 return (LastExt - FirstExt + 1ULL);
1686 /// handleJTSwitchCase - Emit jumptable for current switch case range
1687 bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR,
1688 CaseRecVector& WorkList,
1690 MachineBasicBlock* Default) {
1691 Case& FrontCase = *CR.Range.first;
1692 Case& BackCase = *(CR.Range.second-1);
1694 const APInt& First = cast<ConstantInt>(FrontCase.Low)->getValue();
1695 const APInt& Last = cast<ConstantInt>(BackCase.High)->getValue();
1698 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1702 if (!areJTsAllowed(TLI) || TSize <= 3)
1705 APInt Range = ComputeRange(First, Last);
1706 double Density = (double)TSize / Range.roundToDouble();
1710 DEBUG(errs() << "Lowering jump table\n"
1711 << "First entry: " << First << ". Last entry: " << Last << '\n'
1712 << "Range: " << Range
1713 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1715 // Get the MachineFunction which holds the current MBB. This is used when
1716 // inserting any additional MBBs necessary to represent the switch.
1717 MachineFunction *CurMF = CurMBB->getParent();
1719 // Figure out which block is immediately after the current one.
1720 MachineBasicBlock *NextBlock = 0;
1721 MachineFunction::iterator BBI = CR.CaseBB;
1723 if (++BBI != CurMBB->getParent()->end())
1726 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1728 // Create a new basic block to hold the code for loading the address
1729 // of the jump table, and jumping to it. Update successor information;
1730 // we will either branch to the default case for the switch, or the jump
1732 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1733 CurMF->insert(BBI, JumpTableBB);
1734 CR.CaseBB->addSuccessor(Default);
1735 CR.CaseBB->addSuccessor(JumpTableBB);
1737 // Build a vector of destination BBs, corresponding to each target
1738 // of the jump table. If the value of the jump table slot corresponds to
1739 // a case statement, push the case's BB onto the vector, otherwise, push
1741 std::vector<MachineBasicBlock*> DestBBs;
1743 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1744 const APInt& Low = cast<ConstantInt>(I->Low)->getValue();
1745 const APInt& High = cast<ConstantInt>(I->High)->getValue();
1747 if (Low.sle(TEI) && TEI.sle(High)) {
1748 DestBBs.push_back(I->BB);
1752 DestBBs.push_back(Default);
1756 // Update successor info. Add one edge to each unique successor.
1757 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1758 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1759 E = DestBBs.end(); I != E; ++I) {
1760 if (!SuccsHandled[(*I)->getNumber()]) {
1761 SuccsHandled[(*I)->getNumber()] = true;
1762 JumpTableBB->addSuccessor(*I);
1766 // Create a jump table index for this jump table, or return an existing
1768 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1770 // Set the jump table information so that we can codegen it as a second
1771 // MachineBasicBlock
1772 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1773 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == CurMBB));
1774 if (CR.CaseBB == CurMBB)
1775 visitJumpTableHeader(JT, JTH);
1777 JTCases.push_back(JumpTableBlock(JTH, JT));
1782 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1784 bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR,
1785 CaseRecVector& WorkList,
1787 MachineBasicBlock* Default) {
1788 // Get the MachineFunction which holds the current MBB. This is used when
1789 // inserting any additional MBBs necessary to represent the switch.
1790 MachineFunction *CurMF = CurMBB->getParent();
1792 // Figure out which block is immediately after the current one.
1793 MachineBasicBlock *NextBlock = 0;
1794 MachineFunction::iterator BBI = CR.CaseBB;
1796 if (++BBI != CurMBB->getParent()->end())
1799 Case& FrontCase = *CR.Range.first;
1800 Case& BackCase = *(CR.Range.second-1);
1801 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1803 // Size is the number of Cases represented by this range.
1804 unsigned Size = CR.Range.second - CR.Range.first;
1806 const APInt& First = cast<ConstantInt>(FrontCase.Low)->getValue();
1807 const APInt& Last = cast<ConstantInt>(BackCase.High)->getValue();
1809 CaseItr Pivot = CR.Range.first + Size/2;
1811 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1812 // (heuristically) allow us to emit JumpTable's later.
1814 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1818 size_t LSize = FrontCase.size();
1819 size_t RSize = TSize-LSize;
1820 DEBUG(errs() << "Selecting best pivot: \n"
1821 << "First: " << First << ", Last: " << Last <<'\n'
1822 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1823 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1825 const APInt& LEnd = cast<ConstantInt>(I->High)->getValue();
1826 const APInt& RBegin = cast<ConstantInt>(J->Low)->getValue();
1827 APInt Range = ComputeRange(LEnd, RBegin);
1828 assert((Range - 2ULL).isNonNegative() &&
1829 "Invalid case distance");
1830 double LDensity = (double)LSize / (LEnd - First + 1ULL).roundToDouble();
1831 double RDensity = (double)RSize / (Last - RBegin + 1ULL).roundToDouble();
1832 double Metric = Range.logBase2()*(LDensity+RDensity);
1833 // Should always split in some non-trivial place
1834 DEBUG(errs() <<"=>Step\n"
1835 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1836 << "LDensity: " << LDensity
1837 << ", RDensity: " << RDensity << '\n'
1838 << "Metric: " << Metric << '\n');
1839 if (FMetric < Metric) {
1842 DEBUG(errs() << "Current metric set to: " << FMetric << '\n');
1848 if (areJTsAllowed(TLI)) {
1849 // If our case is dense we *really* should handle it earlier!
1850 assert((FMetric > 0) && "Should handle dense range earlier!");
1852 Pivot = CR.Range.first + Size/2;
1855 CaseRange LHSR(CR.Range.first, Pivot);
1856 CaseRange RHSR(Pivot, CR.Range.second);
1857 Constant *C = Pivot->Low;
1858 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1860 // We know that we branch to the LHS if the Value being switched on is
1861 // less than the Pivot value, C. We use this to optimize our binary
1862 // tree a bit, by recognizing that if SV is greater than or equal to the
1863 // LHS's Case Value, and that Case Value is exactly one less than the
1864 // Pivot's Value, then we can branch directly to the LHS's Target,
1865 // rather than creating a leaf node for it.
1866 if ((LHSR.second - LHSR.first) == 1 &&
1867 LHSR.first->High == CR.GE &&
1868 cast<ConstantInt>(C)->getValue() ==
1869 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
1870 TrueBB = LHSR.first->BB;
1872 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1873 CurMF->insert(BBI, TrueBB);
1874 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1877 // Similar to the optimization above, if the Value being switched on is
1878 // known to be less than the Constant CR.LT, and the current Case Value
1879 // is CR.LT - 1, then we can branch directly to the target block for
1880 // the current Case Value, rather than emitting a RHS leaf node for it.
1881 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1882 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
1883 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
1884 FalseBB = RHSR.first->BB;
1886 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1887 CurMF->insert(BBI, FalseBB);
1888 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1891 // Create a CaseBlock record representing a conditional branch to
1892 // the LHS node if the value being switched on SV is less than C.
1893 // Otherwise, branch to LHS.
1894 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
1896 if (CR.CaseBB == CurMBB)
1897 visitSwitchCase(CB);
1899 SwitchCases.push_back(CB);
1904 /// handleBitTestsSwitchCase - if current case range has few destination and
1905 /// range span less, than machine word bitwidth, encode case range into series
1906 /// of masks and emit bit tests with these masks.
1907 bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR,
1908 CaseRecVector& WorkList,
1910 MachineBasicBlock* Default){
1911 unsigned IntPtrBits = TLI.getPointerTy().getSizeInBits();
1913 Case& FrontCase = *CR.Range.first;
1914 Case& BackCase = *(CR.Range.second-1);
1916 // Get the MachineFunction which holds the current MBB. This is used when
1917 // inserting any additional MBBs necessary to represent the switch.
1918 MachineFunction *CurMF = CurMBB->getParent();
1921 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1923 // Single case counts one, case range - two.
1924 numCmps += (I->Low == I->High ? 1 : 2);
1927 // Count unique destinations
1928 SmallSet<MachineBasicBlock*, 4> Dests;
1929 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1930 Dests.insert(I->BB);
1931 if (Dests.size() > 3)
1932 // Don't bother the code below, if there are too much unique destinations
1935 DEBUG(errs() << "Total number of unique destinations: " << Dests.size() << '\n'
1936 << "Total number of comparisons: " << numCmps << '\n');
1938 // Compute span of values.
1939 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
1940 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
1941 APInt cmpRange = maxValue - minValue;
1943 DEBUG(errs() << "Compare range: " << cmpRange << '\n'
1944 << "Low bound: " << minValue << '\n'
1945 << "High bound: " << maxValue << '\n');
1947 if (cmpRange.uge(APInt(cmpRange.getBitWidth(), IntPtrBits)) ||
1948 (!(Dests.size() == 1 && numCmps >= 3) &&
1949 !(Dests.size() == 2 && numCmps >= 5) &&
1950 !(Dests.size() >= 3 && numCmps >= 6)))
1953 DEBUG(errs() << "Emitting bit tests\n");
1954 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
1956 // Optimize the case where all the case values fit in a
1957 // word without having to subtract minValue. In this case,
1958 // we can optimize away the subtraction.
1959 if (minValue.isNonNegative() &&
1960 maxValue.slt(APInt(maxValue.getBitWidth(), IntPtrBits))) {
1961 cmpRange = maxValue;
1963 lowBound = minValue;
1966 CaseBitsVector CasesBits;
1967 unsigned i, count = 0;
1969 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1970 MachineBasicBlock* Dest = I->BB;
1971 for (i = 0; i < count; ++i)
1972 if (Dest == CasesBits[i].BB)
1976 assert((count < 3) && "Too much destinations to test!");
1977 CasesBits.push_back(CaseBits(0, Dest, 0));
1981 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
1982 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
1984 uint64_t lo = (lowValue - lowBound).getZExtValue();
1985 uint64_t hi = (highValue - lowBound).getZExtValue();
1987 for (uint64_t j = lo; j <= hi; j++) {
1988 CasesBits[i].Mask |= 1ULL << j;
1989 CasesBits[i].Bits++;
1993 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
1997 // Figure out which block is immediately after the current one.
1998 MachineFunction::iterator BBI = CR.CaseBB;
2001 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
2003 DEBUG(errs() << "Cases:\n");
2004 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
2005 DEBUG(errs() << "Mask: " << CasesBits[i].Mask
2006 << ", Bits: " << CasesBits[i].Bits
2007 << ", BB: " << CasesBits[i].BB << '\n');
2009 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2010 CurMF->insert(BBI, CaseBB);
2011 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2016 BitTestBlock BTB(lowBound, cmpRange, SV,
2017 -1U, (CR.CaseBB == CurMBB),
2018 CR.CaseBB, Default, BTC);
2020 if (CR.CaseBB == CurMBB)
2021 visitBitTestHeader(BTB);
2023 BitTestCases.push_back(BTB);
2029 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2030 size_t SelectionDAGLowering::Clusterify(CaseVector& Cases,
2031 const SwitchInst& SI) {
2034 // Start with "simple" cases
2035 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2036 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2037 Cases.push_back(Case(SI.getSuccessorValue(i),
2038 SI.getSuccessorValue(i),
2041 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2043 // Merge case into clusters
2044 if (Cases.size() >= 2)
2045 // Must recompute end() each iteration because it may be
2046 // invalidated by erase if we hold on to it
2047 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2048 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2049 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2050 MachineBasicBlock* nextBB = J->BB;
2051 MachineBasicBlock* currentBB = I->BB;
2053 // If the two neighboring cases go to the same destination, merge them
2054 // into a single case.
2055 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2063 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2064 if (I->Low != I->High)
2065 // A range counts double, since it requires two compares.
2072 void SelectionDAGLowering::visitSwitch(SwitchInst &SI) {
2073 // Figure out which block is immediately after the current one.
2074 MachineBasicBlock *NextBlock = 0;
2075 MachineFunction::iterator BBI = CurMBB;
2077 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2079 // If there is only the default destination, branch to it if it is not the
2080 // next basic block. Otherwise, just fall through.
2081 if (SI.getNumOperands() == 2) {
2082 // Update machine-CFG edges.
2084 // If this is not a fall-through branch, emit the branch.
2085 CurMBB->addSuccessor(Default);
2086 if (Default != NextBlock)
2087 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2088 MVT::Other, getControlRoot(),
2089 DAG.getBasicBlock(Default)));
2093 // If there are any non-default case statements, create a vector of Cases
2094 // representing each one, and sort the vector so that we can efficiently
2095 // create a binary search tree from them.
2097 size_t numCmps = Clusterify(Cases, SI);
2098 DEBUG(errs() << "Clusterify finished. Total clusters: " << Cases.size()
2099 << ". Total compares: " << numCmps << '\n');
2102 // Get the Value to be switched on and default basic blocks, which will be
2103 // inserted into CaseBlock records, representing basic blocks in the binary
2105 Value *SV = SI.getOperand(0);
2107 // Push the initial CaseRec onto the worklist
2108 CaseRecVector WorkList;
2109 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
2111 while (!WorkList.empty()) {
2112 // Grab a record representing a case range to process off the worklist
2113 CaseRec CR = WorkList.back();
2114 WorkList.pop_back();
2116 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
2119 // If the range has few cases (two or less) emit a series of specific
2121 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2124 // If the switch has more than 5 blocks, and at least 40% dense, and the
2125 // target supports indirect branches, then emit a jump table rather than
2126 // lowering the switch to a binary tree of conditional branches.
2127 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2130 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2131 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2132 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2137 void SelectionDAGLowering::visitSub(User &I) {
2138 // -0.0 - X --> fneg
2139 const Type *Ty = I.getType();
2140 if (isa<VectorType>(Ty)) {
2141 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2142 const VectorType *DestTy = cast<VectorType>(I.getType());
2143 const Type *ElTy = DestTy->getElementType();
2144 if (ElTy->isFloatingPoint()) {
2145 unsigned VL = DestTy->getNumElements();
2146 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2147 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2149 SDValue Op2 = getValue(I.getOperand(1));
2150 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2151 Op2.getValueType(), Op2));
2157 if (Ty->isFloatingPoint()) {
2158 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2159 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2160 SDValue Op2 = getValue(I.getOperand(1));
2161 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2162 Op2.getValueType(), Op2));
2167 visitBinary(I, Ty->isFPOrFPVector() ? ISD::FSUB : ISD::SUB);
2170 void SelectionDAGLowering::visitBinary(User &I, unsigned OpCode) {
2171 SDValue Op1 = getValue(I.getOperand(0));
2172 SDValue Op2 = getValue(I.getOperand(1));
2174 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2175 Op1.getValueType(), Op1, Op2));
2178 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
2179 SDValue Op1 = getValue(I.getOperand(0));
2180 SDValue Op2 = getValue(I.getOperand(1));
2181 if (!isa<VectorType>(I.getType())) {
2182 if (TLI.getPointerTy().bitsLT(Op2.getValueType()))
2183 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2184 TLI.getPointerTy(), Op2);
2185 else if (TLI.getPointerTy().bitsGT(Op2.getValueType()))
2186 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2187 TLI.getPointerTy(), Op2);
2190 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2191 Op1.getValueType(), Op1, Op2));
2194 void SelectionDAGLowering::visitICmp(User &I) {
2195 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2196 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2197 predicate = IC->getPredicate();
2198 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2199 predicate = ICmpInst::Predicate(IC->getPredicate());
2200 SDValue Op1 = getValue(I.getOperand(0));
2201 SDValue Op2 = getValue(I.getOperand(1));
2202 ISD::CondCode Opcode = getICmpCondCode(predicate);
2203 setValue(&I, DAG.getSetCC(getCurDebugLoc(),MVT::i1, Op1, Op2, Opcode));
2206 void SelectionDAGLowering::visitFCmp(User &I) {
2207 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2208 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2209 predicate = FC->getPredicate();
2210 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2211 predicate = FCmpInst::Predicate(FC->getPredicate());
2212 SDValue Op1 = getValue(I.getOperand(0));
2213 SDValue Op2 = getValue(I.getOperand(1));
2214 ISD::CondCode Condition = getFCmpCondCode(predicate);
2215 setValue(&I, DAG.getSetCC(getCurDebugLoc(), MVT::i1, Op1, Op2, Condition));
2218 void SelectionDAGLowering::visitVICmp(User &I) {
2219 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2220 if (VICmpInst *IC = dyn_cast<VICmpInst>(&I))
2221 predicate = IC->getPredicate();
2222 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2223 predicate = ICmpInst::Predicate(IC->getPredicate());
2224 SDValue Op1 = getValue(I.getOperand(0));
2225 SDValue Op2 = getValue(I.getOperand(1));
2226 ISD::CondCode Opcode = getICmpCondCode(predicate);
2227 setValue(&I, DAG.getVSetCC(getCurDebugLoc(), Op1.getValueType(),
2231 void SelectionDAGLowering::visitVFCmp(User &I) {
2232 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2233 if (VFCmpInst *FC = dyn_cast<VFCmpInst>(&I))
2234 predicate = FC->getPredicate();
2235 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2236 predicate = FCmpInst::Predicate(FC->getPredicate());
2237 SDValue Op1 = getValue(I.getOperand(0));
2238 SDValue Op2 = getValue(I.getOperand(1));
2239 ISD::CondCode Condition = getFCmpCondCode(predicate);
2240 MVT DestVT = TLI.getValueType(I.getType());
2242 setValue(&I, DAG.getVSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2245 void SelectionDAGLowering::visitSelect(User &I) {
2246 SmallVector<MVT, 4> ValueVTs;
2247 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2248 unsigned NumValues = ValueVTs.size();
2249 if (NumValues != 0) {
2250 SmallVector<SDValue, 4> Values(NumValues);
2251 SDValue Cond = getValue(I.getOperand(0));
2252 SDValue TrueVal = getValue(I.getOperand(1));
2253 SDValue FalseVal = getValue(I.getOperand(2));
2255 for (unsigned i = 0; i != NumValues; ++i)
2256 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2257 TrueVal.getValueType(), Cond,
2258 SDValue(TrueVal.getNode(), TrueVal.getResNo() + i),
2259 SDValue(FalseVal.getNode(), FalseVal.getResNo() + i));
2261 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2262 DAG.getVTList(&ValueVTs[0], NumValues),
2263 &Values[0], NumValues));
2268 void SelectionDAGLowering::visitTrunc(User &I) {
2269 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2270 SDValue N = getValue(I.getOperand(0));
2271 MVT DestVT = TLI.getValueType(I.getType());
2272 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2275 void SelectionDAGLowering::visitZExt(User &I) {
2276 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2277 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2278 SDValue N = getValue(I.getOperand(0));
2279 MVT DestVT = TLI.getValueType(I.getType());
2280 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2283 void SelectionDAGLowering::visitSExt(User &I) {
2284 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2285 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2286 SDValue N = getValue(I.getOperand(0));
2287 MVT DestVT = TLI.getValueType(I.getType());
2288 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2291 void SelectionDAGLowering::visitFPTrunc(User &I) {
2292 // FPTrunc is never a no-op cast, no need to check
2293 SDValue N = getValue(I.getOperand(0));
2294 MVT DestVT = TLI.getValueType(I.getType());
2295 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2296 DestVT, N, DAG.getIntPtrConstant(0)));
2299 void SelectionDAGLowering::visitFPExt(User &I){
2300 // FPTrunc is never a no-op cast, no need to check
2301 SDValue N = getValue(I.getOperand(0));
2302 MVT DestVT = TLI.getValueType(I.getType());
2303 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2306 void SelectionDAGLowering::visitFPToUI(User &I) {
2307 // FPToUI is never a no-op cast, no need to check
2308 SDValue N = getValue(I.getOperand(0));
2309 MVT DestVT = TLI.getValueType(I.getType());
2310 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2313 void SelectionDAGLowering::visitFPToSI(User &I) {
2314 // FPToSI is never a no-op cast, no need to check
2315 SDValue N = getValue(I.getOperand(0));
2316 MVT DestVT = TLI.getValueType(I.getType());
2317 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, getCurDebugLoc(), DestVT, N));
2320 void SelectionDAGLowering::visitUIToFP(User &I) {
2321 // UIToFP is never a no-op cast, no need to check
2322 SDValue N = getValue(I.getOperand(0));
2323 MVT DestVT = TLI.getValueType(I.getType());
2324 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2327 void SelectionDAGLowering::visitSIToFP(User &I){
2328 // SIToFP is never a no-op cast, no need to check
2329 SDValue N = getValue(I.getOperand(0));
2330 MVT DestVT = TLI.getValueType(I.getType());
2331 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2334 void SelectionDAGLowering::visitPtrToInt(User &I) {
2335 // What to do depends on the size of the integer and the size of the pointer.
2336 // We can either truncate, zero extend, or no-op, accordingly.
2337 SDValue N = getValue(I.getOperand(0));
2338 MVT SrcVT = N.getValueType();
2339 MVT DestVT = TLI.getValueType(I.getType());
2341 if (DestVT.bitsLT(SrcVT))
2342 Result = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N);
2344 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2345 Result = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N);
2346 setValue(&I, Result);
2349 void SelectionDAGLowering::visitIntToPtr(User &I) {
2350 // What to do depends on the size of the integer and the size of the pointer.
2351 // We can either truncate, zero extend, or no-op, accordingly.
2352 SDValue N = getValue(I.getOperand(0));
2353 MVT SrcVT = N.getValueType();
2354 MVT DestVT = TLI.getValueType(I.getType());
2355 if (DestVT.bitsLT(SrcVT))
2356 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2358 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2359 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2363 void SelectionDAGLowering::visitBitCast(User &I) {
2364 SDValue N = getValue(I.getOperand(0));
2365 MVT DestVT = TLI.getValueType(I.getType());
2367 // BitCast assures us that source and destination are the same size so this
2368 // is either a BIT_CONVERT or a no-op.
2369 if (DestVT != N.getValueType())
2370 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2371 DestVT, N)); // convert types
2373 setValue(&I, N); // noop cast.
2376 void SelectionDAGLowering::visitInsertElement(User &I) {
2377 SDValue InVec = getValue(I.getOperand(0));
2378 SDValue InVal = getValue(I.getOperand(1));
2379 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2381 getValue(I.getOperand(2)));
2383 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2384 TLI.getValueType(I.getType()),
2385 InVec, InVal, InIdx));
2388 void SelectionDAGLowering::visitExtractElement(User &I) {
2389 SDValue InVec = getValue(I.getOperand(0));
2390 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2392 getValue(I.getOperand(1)));
2393 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2394 TLI.getValueType(I.getType()), InVec, InIdx));
2398 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2399 // from SIndx and increasing to the element length (undefs are allowed).
2400 static bool SequentialMask(SDValue Mask, unsigned SIndx) {
2401 unsigned MaskNumElts = Mask.getNumOperands();
2402 for (unsigned i = 0; i != MaskNumElts; ++i) {
2403 if (Mask.getOperand(i).getOpcode() != ISD::UNDEF) {
2404 unsigned Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
2405 if (Idx != i + SIndx)
2412 void SelectionDAGLowering::visitShuffleVector(User &I) {
2413 SDValue Src1 = getValue(I.getOperand(0));
2414 SDValue Src2 = getValue(I.getOperand(1));
2415 SDValue Mask = getValue(I.getOperand(2));
2417 MVT VT = TLI.getValueType(I.getType());
2418 MVT SrcVT = Src1.getValueType();
2419 int MaskNumElts = Mask.getNumOperands();
2420 int SrcNumElts = SrcVT.getVectorNumElements();
2422 if (SrcNumElts == MaskNumElts) {
2423 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, getCurDebugLoc(),
2424 VT, Src1, Src2, Mask));
2428 // Normalize the shuffle vector since mask and vector length don't match.
2429 MVT MaskEltVT = Mask.getValueType().getVectorElementType();
2431 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2432 // Mask is longer than the source vectors and is a multiple of the source
2433 // vectors. We can use concatenate vector to make the mask and vectors
2435 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2436 // The shuffle is concatenating two vectors together.
2437 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2442 // Pad both vectors with undefs to make them the same length as the mask.
2443 unsigned NumConcat = MaskNumElts / SrcNumElts;
2444 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2446 SDValue* MOps1 = new SDValue[NumConcat];
2447 SDValue* MOps2 = new SDValue[NumConcat];
2450 for (unsigned i = 1; i != NumConcat; ++i) {
2451 MOps1[i] = UndefVal;
2452 MOps2[i] = UndefVal;
2454 Src1 = DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2455 VT, MOps1, NumConcat);
2456 Src2 = DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2457 VT, MOps2, NumConcat);
2462 // Readjust mask for new input vector length.
2463 SmallVector<SDValue, 8> MappedOps;
2464 for (int i = 0; i != MaskNumElts; ++i) {
2465 if (Mask.getOperand(i).getOpcode() == ISD::UNDEF) {
2466 MappedOps.push_back(Mask.getOperand(i));
2468 int Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
2469 if (Idx < SrcNumElts)
2470 MappedOps.push_back(DAG.getConstant(Idx, MaskEltVT));
2472 MappedOps.push_back(DAG.getConstant(Idx + MaskNumElts - SrcNumElts,
2476 Mask = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2477 Mask.getValueType(),
2478 &MappedOps[0], MappedOps.size());
2480 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, getCurDebugLoc(),
2481 VT, Src1, Src2, Mask));
2485 if (SrcNumElts > MaskNumElts) {
2486 // Resulting vector is shorter than the incoming vector.
2487 if (SrcNumElts == MaskNumElts && SequentialMask(Mask,0)) {
2488 // Shuffle extracts 1st vector.
2493 if (SrcNumElts == MaskNumElts && SequentialMask(Mask,MaskNumElts)) {
2494 // Shuffle extracts 2nd vector.
2499 // Analyze the access pattern of the vector to see if we can extract
2500 // two subvectors and do the shuffle. The analysis is done by calculating
2501 // the range of elements the mask access on both vectors.
2502 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2503 int MaxRange[2] = {-1, -1};
2505 for (int i = 0; i != MaskNumElts; ++i) {
2506 SDValue Arg = Mask.getOperand(i);
2507 if (Arg.getOpcode() != ISD::UNDEF) {
2508 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2509 int Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2511 if (Idx >= SrcNumElts) {
2515 if (Idx > MaxRange[Input])
2516 MaxRange[Input] = Idx;
2517 if (Idx < MinRange[Input])
2518 MinRange[Input] = Idx;
2522 // Check if the access is smaller than the vector size and can we find
2523 // a reasonable extract index.
2524 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not Extract.
2525 int StartIdx[2]; // StartIdx to extract from
2526 for (int Input=0; Input < 2; ++Input) {
2527 if (MinRange[Input] == SrcNumElts+1 && MaxRange[Input] == -1) {
2528 RangeUse[Input] = 0; // Unused
2529 StartIdx[Input] = 0;
2530 } else if (MaxRange[Input] - MinRange[Input] < MaskNumElts) {
2531 // Fits within range but we should see if we can find a good
2532 // start index that is a multiple of the mask length.
2533 if (MaxRange[Input] < MaskNumElts) {
2534 RangeUse[Input] = 1; // Extract from beginning of the vector
2535 StartIdx[Input] = 0;
2537 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2538 if (MaxRange[Input] - StartIdx[Input] < MaskNumElts &&
2539 StartIdx[Input] + MaskNumElts < SrcNumElts)
2540 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2545 if (RangeUse[0] == 0 && RangeUse[0] == 0) {
2546 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2549 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2550 // Extract appropriate subvector and generate a vector shuffle
2551 for (int Input=0; Input < 2; ++Input) {
2552 SDValue& Src = Input == 0 ? Src1 : Src2;
2553 if (RangeUse[Input] == 0) {
2554 Src = DAG.getUNDEF(VT);
2556 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2557 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2560 // Calculate new mask.
2561 SmallVector<SDValue, 8> MappedOps;
2562 for (int i = 0; i != MaskNumElts; ++i) {
2563 SDValue Arg = Mask.getOperand(i);
2564 if (Arg.getOpcode() == ISD::UNDEF) {
2565 MappedOps.push_back(Arg);
2567 int Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2568 if (Idx < SrcNumElts)
2569 MappedOps.push_back(DAG.getConstant(Idx - StartIdx[0], MaskEltVT));
2571 Idx = Idx - SrcNumElts - StartIdx[1] + MaskNumElts;
2572 MappedOps.push_back(DAG.getConstant(Idx, MaskEltVT));
2576 Mask = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2577 Mask.getValueType(),
2578 &MappedOps[0], MappedOps.size());
2579 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, getCurDebugLoc(),
2580 VT, Src1, Src2, Mask));
2585 // We can't use either concat vectors or extract subvectors so fall back to
2586 // replacing the shuffle with extract and build vector.
2587 // to insert and build vector.
2588 MVT EltVT = VT.getVectorElementType();
2589 MVT PtrVT = TLI.getPointerTy();
2590 SmallVector<SDValue,8> Ops;
2591 for (int i = 0; i != MaskNumElts; ++i) {
2592 SDValue Arg = Mask.getOperand(i);
2593 if (Arg.getOpcode() == ISD::UNDEF) {
2594 Ops.push_back(DAG.getUNDEF(EltVT));
2596 assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
2597 int Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
2598 if (Idx < SrcNumElts)
2599 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2600 EltVT, Src1, DAG.getConstant(Idx, PtrVT)));
2602 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2604 DAG.getConstant(Idx - SrcNumElts, PtrVT)));
2607 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2608 VT, &Ops[0], Ops.size()));
2611 void SelectionDAGLowering::visitInsertValue(InsertValueInst &I) {
2612 const Value *Op0 = I.getOperand(0);
2613 const Value *Op1 = I.getOperand(1);
2614 const Type *AggTy = I.getType();
2615 const Type *ValTy = Op1->getType();
2616 bool IntoUndef = isa<UndefValue>(Op0);
2617 bool FromUndef = isa<UndefValue>(Op1);
2619 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2620 I.idx_begin(), I.idx_end());
2622 SmallVector<MVT, 4> AggValueVTs;
2623 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2624 SmallVector<MVT, 4> ValValueVTs;
2625 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2627 unsigned NumAggValues = AggValueVTs.size();
2628 unsigned NumValValues = ValValueVTs.size();
2629 SmallVector<SDValue, 4> Values(NumAggValues);
2631 SDValue Agg = getValue(Op0);
2632 SDValue Val = getValue(Op1);
2634 // Copy the beginning value(s) from the original aggregate.
2635 for (; i != LinearIndex; ++i)
2636 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2637 SDValue(Agg.getNode(), Agg.getResNo() + i);
2638 // Copy values from the inserted value(s).
2639 for (; i != LinearIndex + NumValValues; ++i)
2640 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2641 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2642 // Copy remaining value(s) from the original aggregate.
2643 for (; i != NumAggValues; ++i)
2644 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2645 SDValue(Agg.getNode(), Agg.getResNo() + i);
2647 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2648 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2649 &Values[0], NumAggValues));
2652 void SelectionDAGLowering::visitExtractValue(ExtractValueInst &I) {
2653 const Value *Op0 = I.getOperand(0);
2654 const Type *AggTy = Op0->getType();
2655 const Type *ValTy = I.getType();
2656 bool OutOfUndef = isa<UndefValue>(Op0);
2658 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2659 I.idx_begin(), I.idx_end());
2661 SmallVector<MVT, 4> ValValueVTs;
2662 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2664 unsigned NumValValues = ValValueVTs.size();
2665 SmallVector<SDValue, 4> Values(NumValValues);
2667 SDValue Agg = getValue(Op0);
2668 // Copy out the selected value(s).
2669 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2670 Values[i - LinearIndex] =
2672 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2673 SDValue(Agg.getNode(), Agg.getResNo() + i);
2675 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2676 DAG.getVTList(&ValValueVTs[0], NumValValues),
2677 &Values[0], NumValValues));
2681 void SelectionDAGLowering::visitGetElementPtr(User &I) {
2682 SDValue N = getValue(I.getOperand(0));
2683 const Type *Ty = I.getOperand(0)->getType();
2685 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2688 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2689 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2692 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2693 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2694 DAG.getIntPtrConstant(Offset));
2696 Ty = StTy->getElementType(Field);
2698 Ty = cast<SequentialType>(Ty)->getElementType();
2700 // If this is a constant subscript, handle it quickly.
2701 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2702 if (CI->getZExtValue() == 0) continue;
2704 TD->getTypePaddedSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2706 unsigned PtrBits = TLI.getPointerTy().getSizeInBits();
2708 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2710 DAG.getConstant(Offs, MVT::i64));
2712 OffsVal = DAG.getIntPtrConstant(Offs);
2713 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2718 // N = N + Idx * ElementSize;
2719 uint64_t ElementSize = TD->getTypePaddedSize(Ty);
2720 SDValue IdxN = getValue(Idx);
2722 // If the index is smaller or larger than intptr_t, truncate or extend
2724 if (IdxN.getValueType().bitsLT(N.getValueType()))
2725 IdxN = DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(),
2726 N.getValueType(), IdxN);
2727 else if (IdxN.getValueType().bitsGT(N.getValueType()))
2728 IdxN = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2729 N.getValueType(), IdxN);
2731 // If this is a multiply by a power of two, turn it into a shl
2732 // immediately. This is a very common case.
2733 if (ElementSize != 1) {
2734 if (isPowerOf2_64(ElementSize)) {
2735 unsigned Amt = Log2_64(ElementSize);
2736 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2737 N.getValueType(), IdxN,
2738 DAG.getConstant(Amt, TLI.getPointerTy()));
2740 SDValue Scale = DAG.getIntPtrConstant(ElementSize);
2741 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2742 N.getValueType(), IdxN, Scale);
2746 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2747 N.getValueType(), N, IdxN);
2753 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
2754 // If this is a fixed sized alloca in the entry block of the function,
2755 // allocate it statically on the stack.
2756 if (FuncInfo.StaticAllocaMap.count(&I))
2757 return; // getValue will auto-populate this.
2759 const Type *Ty = I.getAllocatedType();
2760 uint64_t TySize = TLI.getTargetData()->getTypePaddedSize(Ty);
2762 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2765 SDValue AllocSize = getValue(I.getArraySize());
2767 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), AllocSize.getValueType(),
2769 DAG.getConstant(TySize, AllocSize.getValueType()));
2773 MVT IntPtr = TLI.getPointerTy();
2774 if (IntPtr.bitsLT(AllocSize.getValueType()))
2775 AllocSize = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2777 else if (IntPtr.bitsGT(AllocSize.getValueType()))
2778 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2781 // Handle alignment. If the requested alignment is less than or equal to
2782 // the stack alignment, ignore it. If the size is greater than or equal to
2783 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2784 unsigned StackAlign =
2785 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2786 if (Align <= StackAlign)
2789 // Round the size of the allocation up to the stack alignment size
2790 // by add SA-1 to the size.
2791 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2792 AllocSize.getValueType(), AllocSize,
2793 DAG.getIntPtrConstant(StackAlign-1));
2794 // Mask out the low bits for alignment purposes.
2795 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2796 AllocSize.getValueType(), AllocSize,
2797 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2799 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2800 const MVT *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
2802 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2805 DAG.setRoot(DSA.getValue(1));
2807 // Inform the Frame Information that we have just allocated a variable-sized
2809 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
2812 void SelectionDAGLowering::visitLoad(LoadInst &I) {
2813 const Value *SV = I.getOperand(0);
2814 SDValue Ptr = getValue(SV);
2816 const Type *Ty = I.getType();
2817 bool isVolatile = I.isVolatile();
2818 unsigned Alignment = I.getAlignment();
2820 SmallVector<MVT, 4> ValueVTs;
2821 SmallVector<uint64_t, 4> Offsets;
2822 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2823 unsigned NumValues = ValueVTs.size();
2828 bool ConstantMemory = false;
2830 // Serialize volatile loads with other side effects.
2832 else if (AA->pointsToConstantMemory(SV)) {
2833 // Do not serialize (non-volatile) loads of constant memory with anything.
2834 Root = DAG.getEntryNode();
2835 ConstantMemory = true;
2837 // Do not serialize non-volatile loads against each other.
2838 Root = DAG.getRoot();
2841 SmallVector<SDValue, 4> Values(NumValues);
2842 SmallVector<SDValue, 4> Chains(NumValues);
2843 MVT PtrVT = Ptr.getValueType();
2844 for (unsigned i = 0; i != NumValues; ++i) {
2845 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2846 DAG.getNode(ISD::ADD, getCurDebugLoc(),
2848 DAG.getConstant(Offsets[i], PtrVT)),
2850 isVolatile, Alignment);
2852 Chains[i] = L.getValue(1);
2855 if (!ConstantMemory) {
2856 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2858 &Chains[0], NumValues);
2862 PendingLoads.push_back(Chain);
2865 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2866 DAG.getVTList(&ValueVTs[0], NumValues),
2867 &Values[0], NumValues));
2871 void SelectionDAGLowering::visitStore(StoreInst &I) {
2872 Value *SrcV = I.getOperand(0);
2873 Value *PtrV = I.getOperand(1);
2875 SmallVector<MVT, 4> ValueVTs;
2876 SmallVector<uint64_t, 4> Offsets;
2877 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2878 unsigned NumValues = ValueVTs.size();
2882 // Get the lowered operands. Note that we do this after
2883 // checking if NumResults is zero, because with zero results
2884 // the operands won't have values in the map.
2885 SDValue Src = getValue(SrcV);
2886 SDValue Ptr = getValue(PtrV);
2888 SDValue Root = getRoot();
2889 SmallVector<SDValue, 4> Chains(NumValues);
2890 MVT PtrVT = Ptr.getValueType();
2891 bool isVolatile = I.isVolatile();
2892 unsigned Alignment = I.getAlignment();
2893 for (unsigned i = 0; i != NumValues; ++i)
2894 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
2895 SDValue(Src.getNode(), Src.getResNo() + i),
2896 DAG.getNode(ISD::ADD, getCurDebugLoc(),
2898 DAG.getConstant(Offsets[i], PtrVT)),
2900 isVolatile, Alignment);
2902 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2903 MVT::Other, &Chains[0], NumValues));
2906 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2908 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
2909 unsigned Intrinsic) {
2910 bool HasChain = !I.doesNotAccessMemory();
2911 bool OnlyLoad = HasChain && I.onlyReadsMemory();
2913 // Build the operand list.
2914 SmallVector<SDValue, 8> Ops;
2915 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2917 // We don't need to serialize loads against other loads.
2918 Ops.push_back(DAG.getRoot());
2920 Ops.push_back(getRoot());
2924 // Info is set by getTgtMemInstrinsic
2925 TargetLowering::IntrinsicInfo Info;
2926 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
2928 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
2929 if (!IsTgtIntrinsic)
2930 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2932 // Add all operands of the call to the operand list.
2933 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2934 SDValue Op = getValue(I.getOperand(i));
2935 assert(TLI.isTypeLegal(Op.getValueType()) &&
2936 "Intrinsic uses a non-legal type?");
2940 std::vector<MVT> VTs;
2941 if (I.getType() != Type::VoidTy) {
2942 MVT VT = TLI.getValueType(I.getType());
2943 if (VT.isVector()) {
2944 const VectorType *DestTy = cast<VectorType>(I.getType());
2945 MVT EltVT = TLI.getValueType(DestTy->getElementType());
2947 VT = MVT::getVectorVT(EltVT, DestTy->getNumElements());
2948 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
2951 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
2955 VTs.push_back(MVT::Other);
2957 const MVT *VTList = DAG.getNodeValueTypes(VTs);
2961 if (IsTgtIntrinsic) {
2962 // This is target intrinsic that touches memory
2963 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
2965 &Ops[0], Ops.size(),
2966 Info.memVT, Info.ptrVal, Info.offset,
2967 Info.align, Info.vol,
2968 Info.readMem, Info.writeMem);
2971 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
2973 &Ops[0], Ops.size());
2974 else if (I.getType() != Type::VoidTy)
2975 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
2977 &Ops[0], Ops.size());
2979 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
2981 &Ops[0], Ops.size());
2984 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
2986 PendingLoads.push_back(Chain);
2990 if (I.getType() != Type::VoidTy) {
2991 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
2992 MVT VT = TLI.getValueType(PTy);
2993 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
2995 setValue(&I, Result);
2999 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
3000 static GlobalVariable *ExtractTypeInfo(Value *V) {
3001 V = V->stripPointerCasts();
3002 GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
3003 assert ((GV || isa<ConstantPointerNull>(V)) &&
3004 "TypeInfo must be a global variable or NULL");
3010 /// AddCatchInfo - Extract the personality and type infos from an eh.selector
3011 /// call, and add them to the specified machine basic block.
3012 void AddCatchInfo(CallInst &I, MachineModuleInfo *MMI,
3013 MachineBasicBlock *MBB) {
3014 // Inform the MachineModuleInfo of the personality for this landing pad.
3015 ConstantExpr *CE = cast<ConstantExpr>(I.getOperand(2));
3016 assert(CE->getOpcode() == Instruction::BitCast &&
3017 isa<Function>(CE->getOperand(0)) &&
3018 "Personality should be a function");
3019 MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
3021 // Gather all the type infos for this landing pad and pass them along to
3022 // MachineModuleInfo.
3023 std::vector<GlobalVariable *> TyInfo;
3024 unsigned N = I.getNumOperands();
3026 for (unsigned i = N - 1; i > 2; --i) {
3027 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(i))) {
3028 unsigned FilterLength = CI->getZExtValue();
3029 unsigned FirstCatch = i + FilterLength + !FilterLength;
3030 assert (FirstCatch <= N && "Invalid filter length");
3032 if (FirstCatch < N) {
3033 TyInfo.reserve(N - FirstCatch);
3034 for (unsigned j = FirstCatch; j < N; ++j)
3035 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
3036 MMI->addCatchTypeInfo(MBB, TyInfo);
3040 if (!FilterLength) {
3042 MMI->addCleanup(MBB);
3045 TyInfo.reserve(FilterLength - 1);
3046 for (unsigned j = i + 1; j < FirstCatch; ++j)
3047 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
3048 MMI->addFilterTypeInfo(MBB, TyInfo);
3057 TyInfo.reserve(N - 3);
3058 for (unsigned j = 3; j < N; ++j)
3059 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
3060 MMI->addCatchTypeInfo(MBB, TyInfo);
3066 /// GetSignificand - Get the significand and build it into a floating-point
3067 /// number with exponent of 1:
3069 /// Op = (Op & 0x007fffff) | 0x3f800000;
3071 /// where Op is the hexidecimal representation of floating point value.
3073 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3074 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3075 DAG.getConstant(0x007fffff, MVT::i32));
3076 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3077 DAG.getConstant(0x3f800000, MVT::i32));
3078 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3081 /// GetExponent - Get the exponent:
3083 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3085 /// where Op is the hexidecimal representation of floating point value.
3087 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3089 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3090 DAG.getConstant(0x7f800000, MVT::i32));
3091 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3092 DAG.getConstant(23, TLI.getPointerTy()));
3093 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3094 DAG.getConstant(127, MVT::i32));
3095 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3098 /// getF32Constant - Get 32-bit floating point constant.
3100 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3101 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3104 /// Inlined utility function to implement binary input atomic intrinsics for
3105 /// visitIntrinsicCall: I is a call instruction
3106 /// Op is the associated NodeType for I
3108 SelectionDAGLowering::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) {
3109 SDValue Root = getRoot();
3111 DAG.getAtomic(Op, getCurDebugLoc(),
3112 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
3114 getValue(I.getOperand(1)),
3115 getValue(I.getOperand(2)),
3118 DAG.setRoot(L.getValue(1));
3122 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3124 SelectionDAGLowering::implVisitAluOverflow(CallInst &I, ISD::NodeType Op) {
3125 SDValue Op1 = getValue(I.getOperand(1));
3126 SDValue Op2 = getValue(I.getOperand(2));
3128 MVT ValueVTs[] = { Op1.getValueType(), MVT::i1 };
3129 SDValue Ops[] = { Op1, Op2 };
3131 SDValue Result = DAG.getNode(Op, getCurDebugLoc(),
3132 DAG.getVTList(&ValueVTs[0], 2), &Ops[0], 2);
3134 setValue(&I, Result);
3138 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3139 /// limited-precision mode.
3141 SelectionDAGLowering::visitExp(CallInst &I) {
3143 DebugLoc dl = getCurDebugLoc();
3145 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3146 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3147 SDValue Op = getValue(I.getOperand(1));
3149 // Put the exponent in the right bit position for later addition to the
3152 // #define LOG2OFe 1.4426950f
3153 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3154 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3155 getF32Constant(DAG, 0x3fb8aa3b));
3156 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3158 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3159 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3160 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3162 // IntegerPartOfX <<= 23;
3163 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3164 DAG.getConstant(23, TLI.getPointerTy()));
3166 if (LimitFloatPrecision <= 6) {
3167 // For floating-point precision of 6:
3169 // TwoToFractionalPartOfX =
3171 // (0.735607626f + 0.252464424f * x) * x;
3173 // error 0.0144103317, which is 6 bits
3174 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3175 getF32Constant(DAG, 0x3e814304));
3176 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3177 getF32Constant(DAG, 0x3f3c50c8));
3178 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3179 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3180 getF32Constant(DAG, 0x3f7f5e7e));
3181 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3183 // Add the exponent into the result in integer domain.
3184 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3185 TwoToFracPartOfX, IntegerPartOfX);
3187 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3188 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3189 // For floating-point precision of 12:
3191 // TwoToFractionalPartOfX =
3194 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3196 // 0.000107046256 error, which is 13 to 14 bits
3197 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3198 getF32Constant(DAG, 0x3da235e3));
3199 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3200 getF32Constant(DAG, 0x3e65b8f3));
3201 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3202 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3203 getF32Constant(DAG, 0x3f324b07));
3204 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3205 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3206 getF32Constant(DAG, 0x3f7ff8fd));
3207 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3209 // Add the exponent into the result in integer domain.
3210 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3211 TwoToFracPartOfX, IntegerPartOfX);
3213 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3214 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3215 // For floating-point precision of 18:
3217 // TwoToFractionalPartOfX =
3221 // (0.554906021e-1f +
3222 // (0.961591928e-2f +
3223 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3225 // error 2.47208000*10^(-7), which is better than 18 bits
3226 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3227 getF32Constant(DAG, 0x3924b03e));
3228 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3229 getF32Constant(DAG, 0x3ab24b87));
3230 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3231 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3232 getF32Constant(DAG, 0x3c1d8c17));
3233 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3234 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3235 getF32Constant(DAG, 0x3d634a1d));
3236 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3237 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3238 getF32Constant(DAG, 0x3e75fe14));
3239 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3240 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3241 getF32Constant(DAG, 0x3f317234));
3242 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3243 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3244 getF32Constant(DAG, 0x3f800000));
3245 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3248 // Add the exponent into the result in integer domain.
3249 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3250 TwoToFracPartOfX, IntegerPartOfX);
3252 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3255 // No special expansion.
3256 result = DAG.getNode(ISD::FEXP, dl,
3257 getValue(I.getOperand(1)).getValueType(),
3258 getValue(I.getOperand(1)));
3261 setValue(&I, result);
3264 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3265 /// limited-precision mode.
3267 SelectionDAGLowering::visitLog(CallInst &I) {
3269 DebugLoc dl = getCurDebugLoc();
3271 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3272 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3273 SDValue Op = getValue(I.getOperand(1));
3274 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3276 // Scale the exponent by log(2) [0.69314718f].
3277 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3278 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3279 getF32Constant(DAG, 0x3f317218));
3281 // Get the significand and build it into a floating-point number with
3283 SDValue X = GetSignificand(DAG, Op1, dl);
3285 if (LimitFloatPrecision <= 6) {
3286 // For floating-point precision of 6:
3290 // (1.4034025f - 0.23903021f * x) * x;
3292 // error 0.0034276066, which is better than 8 bits
3293 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3294 getF32Constant(DAG, 0xbe74c456));
3295 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3296 getF32Constant(DAG, 0x3fb3a2b1));
3297 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3298 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3299 getF32Constant(DAG, 0x3f949a29));
3301 result = DAG.getNode(ISD::FADD, dl,
3302 MVT::f32, LogOfExponent, LogOfMantissa);
3303 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3304 // For floating-point precision of 12:
3310 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3312 // error 0.000061011436, which is 14 bits
3313 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3314 getF32Constant(DAG, 0xbd67b6d6));
3315 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3316 getF32Constant(DAG, 0x3ee4f4b8));
3317 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3318 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3319 getF32Constant(DAG, 0x3fbc278b));
3320 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3321 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3322 getF32Constant(DAG, 0x40348e95));
3323 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3324 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3325 getF32Constant(DAG, 0x3fdef31a));
3327 result = DAG.getNode(ISD::FADD, dl,
3328 MVT::f32, LogOfExponent, LogOfMantissa);
3329 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3330 // For floating-point precision of 18:
3338 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3340 // error 0.0000023660568, which is better than 18 bits
3341 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3342 getF32Constant(DAG, 0xbc91e5ac));
3343 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3344 getF32Constant(DAG, 0x3e4350aa));
3345 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3346 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3347 getF32Constant(DAG, 0x3f60d3e3));
3348 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3349 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3350 getF32Constant(DAG, 0x4011cdf0));
3351 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3352 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3353 getF32Constant(DAG, 0x406cfd1c));
3354 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3355 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3356 getF32Constant(DAG, 0x408797cb));
3357 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3358 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3359 getF32Constant(DAG, 0x4006dcab));
3361 result = DAG.getNode(ISD::FADD, dl,
3362 MVT::f32, LogOfExponent, LogOfMantissa);
3365 // No special expansion.
3366 result = DAG.getNode(ISD::FLOG, dl,
3367 getValue(I.getOperand(1)).getValueType(),
3368 getValue(I.getOperand(1)));
3371 setValue(&I, result);
3374 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3375 /// limited-precision mode.
3377 SelectionDAGLowering::visitLog2(CallInst &I) {
3379 DebugLoc dl = getCurDebugLoc();
3381 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3382 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3383 SDValue Op = getValue(I.getOperand(1));
3384 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3386 // Get the exponent.
3387 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3389 // Get the significand and build it into a floating-point number with
3391 SDValue X = GetSignificand(DAG, Op1, dl);
3393 // Different possible minimax approximations of significand in
3394 // floating-point for various degrees of accuracy over [1,2].
3395 if (LimitFloatPrecision <= 6) {
3396 // For floating-point precision of 6:
3398 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3400 // error 0.0049451742, which is more than 7 bits
3401 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3402 getF32Constant(DAG, 0xbeb08fe0));
3403 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3404 getF32Constant(DAG, 0x40019463));
3405 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3406 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3407 getF32Constant(DAG, 0x3fd6633d));
3409 result = DAG.getNode(ISD::FADD, dl,
3410 MVT::f32, LogOfExponent, Log2ofMantissa);
3411 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3412 // For floating-point precision of 12:
3418 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3420 // error 0.0000876136000, which is better than 13 bits
3421 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3422 getF32Constant(DAG, 0xbda7262e));
3423 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3424 getF32Constant(DAG, 0x3f25280b));
3425 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3426 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3427 getF32Constant(DAG, 0x4007b923));
3428 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3429 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3430 getF32Constant(DAG, 0x40823e2f));
3431 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3432 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3433 getF32Constant(DAG, 0x4020d29c));
3435 result = DAG.getNode(ISD::FADD, dl,
3436 MVT::f32, LogOfExponent, Log2ofMantissa);
3437 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3438 // For floating-point precision of 18:
3447 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3449 // error 0.0000018516, which is better than 18 bits
3450 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3451 getF32Constant(DAG, 0xbcd2769e));
3452 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3453 getF32Constant(DAG, 0x3e8ce0b9));
3454 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3455 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3456 getF32Constant(DAG, 0x3fa22ae7));
3457 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3458 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3459 getF32Constant(DAG, 0x40525723));
3460 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3461 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3462 getF32Constant(DAG, 0x40aaf200));
3463 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3464 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3465 getF32Constant(DAG, 0x40c39dad));
3466 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3467 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3468 getF32Constant(DAG, 0x4042902c));
3470 result = DAG.getNode(ISD::FADD, dl,
3471 MVT::f32, LogOfExponent, Log2ofMantissa);
3474 // No special expansion.
3475 result = DAG.getNode(ISD::FLOG2, dl,
3476 getValue(I.getOperand(1)).getValueType(),
3477 getValue(I.getOperand(1)));
3480 setValue(&I, result);
3483 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3484 /// limited-precision mode.
3486 SelectionDAGLowering::visitLog10(CallInst &I) {
3488 DebugLoc dl = getCurDebugLoc();
3490 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3491 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3492 SDValue Op = getValue(I.getOperand(1));
3493 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3495 // Scale the exponent by log10(2) [0.30102999f].
3496 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3497 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3498 getF32Constant(DAG, 0x3e9a209a));
3500 // Get the significand and build it into a floating-point number with
3502 SDValue X = GetSignificand(DAG, Op1, dl);
3504 if (LimitFloatPrecision <= 6) {
3505 // For floating-point precision of 6:
3507 // Log10ofMantissa =
3509 // (0.60948995f - 0.10380950f * x) * x;
3511 // error 0.0014886165, which is 6 bits
3512 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3513 getF32Constant(DAG, 0xbdd49a13));
3514 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3515 getF32Constant(DAG, 0x3f1c0789));
3516 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3517 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3518 getF32Constant(DAG, 0x3f011300));
3520 result = DAG.getNode(ISD::FADD, dl,
3521 MVT::f32, LogOfExponent, Log10ofMantissa);
3522 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3523 // For floating-point precision of 12:
3525 // Log10ofMantissa =
3528 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3530 // error 0.00019228036, which is better than 12 bits
3531 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3532 getF32Constant(DAG, 0x3d431f31));
3533 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3534 getF32Constant(DAG, 0x3ea21fb2));
3535 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3536 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3537 getF32Constant(DAG, 0x3f6ae232));
3538 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3539 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3540 getF32Constant(DAG, 0x3f25f7c3));
3542 result = DAG.getNode(ISD::FADD, dl,
3543 MVT::f32, LogOfExponent, Log10ofMantissa);
3544 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3545 // For floating-point precision of 18:
3547 // Log10ofMantissa =
3552 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3554 // error 0.0000037995730, which is better than 18 bits
3555 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3556 getF32Constant(DAG, 0x3c5d51ce));
3557 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3558 getF32Constant(DAG, 0x3e00685a));
3559 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3560 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3561 getF32Constant(DAG, 0x3efb6798));
3562 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3563 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3564 getF32Constant(DAG, 0x3f88d192));
3565 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3566 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3567 getF32Constant(DAG, 0x3fc4316c));
3568 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3569 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3570 getF32Constant(DAG, 0x3f57ce70));
3572 result = DAG.getNode(ISD::FADD, dl,
3573 MVT::f32, LogOfExponent, Log10ofMantissa);
3576 // No special expansion.
3577 result = DAG.getNode(ISD::FLOG10, dl,
3578 getValue(I.getOperand(1)).getValueType(),
3579 getValue(I.getOperand(1)));
3582 setValue(&I, result);
3585 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3586 /// limited-precision mode.
3588 SelectionDAGLowering::visitExp2(CallInst &I) {
3590 DebugLoc dl = getCurDebugLoc();
3592 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3593 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3594 SDValue Op = getValue(I.getOperand(1));
3596 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3598 // FractionalPartOfX = x - (float)IntegerPartOfX;
3599 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3600 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3602 // IntegerPartOfX <<= 23;
3603 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3604 DAG.getConstant(23, TLI.getPointerTy()));
3606 if (LimitFloatPrecision <= 6) {
3607 // For floating-point precision of 6:
3609 // TwoToFractionalPartOfX =
3611 // (0.735607626f + 0.252464424f * x) * x;
3613 // error 0.0144103317, which is 6 bits
3614 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3615 getF32Constant(DAG, 0x3e814304));
3616 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3617 getF32Constant(DAG, 0x3f3c50c8));
3618 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3619 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3620 getF32Constant(DAG, 0x3f7f5e7e));
3621 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3622 SDValue TwoToFractionalPartOfX =
3623 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3625 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3626 MVT::f32, TwoToFractionalPartOfX);
3627 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3628 // For floating-point precision of 12:
3630 // TwoToFractionalPartOfX =
3633 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3635 // error 0.000107046256, which is 13 to 14 bits
3636 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3637 getF32Constant(DAG, 0x3da235e3));
3638 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3639 getF32Constant(DAG, 0x3e65b8f3));
3640 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3641 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3642 getF32Constant(DAG, 0x3f324b07));
3643 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3644 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3645 getF32Constant(DAG, 0x3f7ff8fd));
3646 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3647 SDValue TwoToFractionalPartOfX =
3648 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3650 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3651 MVT::f32, TwoToFractionalPartOfX);
3652 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3653 // For floating-point precision of 18:
3655 // TwoToFractionalPartOfX =
3659 // (0.554906021e-1f +
3660 // (0.961591928e-2f +
3661 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3662 // error 2.47208000*10^(-7), which is better than 18 bits
3663 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3664 getF32Constant(DAG, 0x3924b03e));
3665 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3666 getF32Constant(DAG, 0x3ab24b87));
3667 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3668 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3669 getF32Constant(DAG, 0x3c1d8c17));
3670 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3671 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3672 getF32Constant(DAG, 0x3d634a1d));
3673 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3674 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3675 getF32Constant(DAG, 0x3e75fe14));
3676 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3677 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3678 getF32Constant(DAG, 0x3f317234));
3679 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3680 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3681 getF32Constant(DAG, 0x3f800000));
3682 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3683 SDValue TwoToFractionalPartOfX =
3684 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3686 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3687 MVT::f32, TwoToFractionalPartOfX);
3690 // No special expansion.
3691 result = DAG.getNode(ISD::FEXP2, dl,
3692 getValue(I.getOperand(1)).getValueType(),
3693 getValue(I.getOperand(1)));
3696 setValue(&I, result);
3699 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3700 /// limited-precision mode with x == 10.0f.
3702 SelectionDAGLowering::visitPow(CallInst &I) {
3704 Value *Val = I.getOperand(1);
3705 DebugLoc dl = getCurDebugLoc();
3706 bool IsExp10 = false;
3708 if (getValue(Val).getValueType() == MVT::f32 &&
3709 getValue(I.getOperand(2)).getValueType() == MVT::f32 &&
3710 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3711 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3712 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3714 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3719 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3720 SDValue Op = getValue(I.getOperand(2));
3722 // Put the exponent in the right bit position for later addition to the
3725 // #define LOG2OF10 3.3219281f
3726 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3727 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3728 getF32Constant(DAG, 0x40549a78));
3729 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3731 // FractionalPartOfX = x - (float)IntegerPartOfX;
3732 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3733 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3735 // IntegerPartOfX <<= 23;
3736 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3737 DAG.getConstant(23, TLI.getPointerTy()));
3739 if (LimitFloatPrecision <= 6) {
3740 // For floating-point precision of 6:
3742 // twoToFractionalPartOfX =
3744 // (0.735607626f + 0.252464424f * x) * x;
3746 // error 0.0144103317, which is 6 bits
3747 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3748 getF32Constant(DAG, 0x3e814304));
3749 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3750 getF32Constant(DAG, 0x3f3c50c8));
3751 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3752 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3753 getF32Constant(DAG, 0x3f7f5e7e));
3754 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3755 SDValue TwoToFractionalPartOfX =
3756 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3758 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3759 MVT::f32, TwoToFractionalPartOfX);
3760 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3761 // For floating-point precision of 12:
3763 // TwoToFractionalPartOfX =
3766 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3768 // error 0.000107046256, which is 13 to 14 bits
3769 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3770 getF32Constant(DAG, 0x3da235e3));
3771 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3772 getF32Constant(DAG, 0x3e65b8f3));
3773 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3774 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3775 getF32Constant(DAG, 0x3f324b07));
3776 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3777 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3778 getF32Constant(DAG, 0x3f7ff8fd));
3779 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3780 SDValue TwoToFractionalPartOfX =
3781 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3783 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3784 MVT::f32, TwoToFractionalPartOfX);
3785 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3786 // For floating-point precision of 18:
3788 // TwoToFractionalPartOfX =
3792 // (0.554906021e-1f +
3793 // (0.961591928e-2f +
3794 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3795 // error 2.47208000*10^(-7), which is better than 18 bits
3796 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3797 getF32Constant(DAG, 0x3924b03e));
3798 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3799 getF32Constant(DAG, 0x3ab24b87));
3800 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3801 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3802 getF32Constant(DAG, 0x3c1d8c17));
3803 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3804 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3805 getF32Constant(DAG, 0x3d634a1d));
3806 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3807 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3808 getF32Constant(DAG, 0x3e75fe14));
3809 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3810 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3811 getF32Constant(DAG, 0x3f317234));
3812 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3813 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3814 getF32Constant(DAG, 0x3f800000));
3815 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3816 SDValue TwoToFractionalPartOfX =
3817 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3819 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3820 MVT::f32, TwoToFractionalPartOfX);
3823 // No special expansion.
3824 result = DAG.getNode(ISD::FPOW, dl,
3825 getValue(I.getOperand(1)).getValueType(),
3826 getValue(I.getOperand(1)),
3827 getValue(I.getOperand(2)));
3830 setValue(&I, result);
3833 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3834 /// we want to emit this as a call to a named external function, return the name
3835 /// otherwise lower it and return null.
3837 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
3838 DebugLoc dl = getCurDebugLoc();
3839 switch (Intrinsic) {
3841 // By default, turn this into a target intrinsic node.
3842 visitTargetIntrinsic(I, Intrinsic);
3844 case Intrinsic::vastart: visitVAStart(I); return 0;
3845 case Intrinsic::vaend: visitVAEnd(I); return 0;
3846 case Intrinsic::vacopy: visitVACopy(I); return 0;
3847 case Intrinsic::returnaddress:
3848 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
3849 getValue(I.getOperand(1))));
3851 case Intrinsic::frameaddress:
3852 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
3853 getValue(I.getOperand(1))));
3855 case Intrinsic::setjmp:
3856 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
3858 case Intrinsic::longjmp:
3859 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
3861 case Intrinsic::memcpy: {
3862 SDValue Op1 = getValue(I.getOperand(1));
3863 SDValue Op2 = getValue(I.getOperand(2));
3864 SDValue Op3 = getValue(I.getOperand(3));
3865 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3866 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
3867 I.getOperand(1), 0, I.getOperand(2), 0));
3870 case Intrinsic::memset: {
3871 SDValue Op1 = getValue(I.getOperand(1));
3872 SDValue Op2 = getValue(I.getOperand(2));
3873 SDValue Op3 = getValue(I.getOperand(3));
3874 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3875 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align,
3876 I.getOperand(1), 0));
3879 case Intrinsic::memmove: {
3880 SDValue Op1 = getValue(I.getOperand(1));
3881 SDValue Op2 = getValue(I.getOperand(2));
3882 SDValue Op3 = getValue(I.getOperand(3));
3883 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3885 // If the source and destination are known to not be aliases, we can
3886 // lower memmove as memcpy.
3887 uint64_t Size = -1ULL;
3888 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
3889 Size = C->getZExtValue();
3890 if (AA->alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
3891 AliasAnalysis::NoAlias) {
3892 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
3893 I.getOperand(1), 0, I.getOperand(2), 0));
3897 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align,
3898 I.getOperand(1), 0, I.getOperand(2), 0));
3901 case Intrinsic::dbg_stoppoint: {
3902 DwarfWriter *DW = DAG.getDwarfWriter();
3903 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
3904 if (DW && DW->ValidDebugInfo(SPI.getContext())) {
3905 MachineFunction &MF = DAG.getMachineFunction();
3906 DAG.setRoot(DAG.getDbgStopPoint(getRoot(),
3910 DICompileUnit CU(cast<GlobalVariable>(SPI.getContext()));
3911 std::string Dir, FN;
3912 unsigned SrcFile = DW->getOrCreateSourceID(CU.getDirectory(Dir),
3913 CU.getFilename(FN));
3914 unsigned idx = MF.getOrCreateDebugLocID(SrcFile,
3915 SPI.getLine(), SPI.getColumn());
3916 setCurDebugLoc(DebugLoc::get(idx));
3920 case Intrinsic::dbg_region_start: {
3921 DwarfWriter *DW = DAG.getDwarfWriter();
3922 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
3923 if (DW && DW->ValidDebugInfo(RSI.getContext())) {
3925 DW->RecordRegionStart(cast<GlobalVariable>(RSI.getContext()));
3927 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3928 getRoot(), LabelID));
3933 case Intrinsic::dbg_region_end: {
3934 DwarfWriter *DW = DAG.getDwarfWriter();
3935 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
3936 if (DW && DW->ValidDebugInfo(REI.getContext())) {
3938 DW->RecordRegionEnd(cast<GlobalVariable>(REI.getContext()));
3940 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3941 getRoot(), LabelID));
3946 case Intrinsic::dbg_func_start: {
3947 DwarfWriter *DW = DAG.getDwarfWriter();
3949 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
3950 Value *SP = FSI.getSubprogram();
3951 if (SP && DW->ValidDebugInfo(SP)) {
3952 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is
3953 // what (most?) gdb expects.
3954 MachineFunction &MF = DAG.getMachineFunction();
3955 DISubprogram Subprogram(cast<GlobalVariable>(SP));
3956 DICompileUnit CompileUnit = Subprogram.getCompileUnit();
3957 std::string Dir, FN;
3958 unsigned SrcFile = DW->getOrCreateSourceID(CompileUnit.getDirectory(Dir),
3959 CompileUnit.getFilename(FN));
3961 // Record the source line but does not create a label for the normal
3962 // function start. It will be emitted at asm emission time. However,
3963 // create a label if this is a beginning of inlined function.
3964 unsigned Line = Subprogram.getLineNumber();
3967 unsigned LabelID = DW->RecordSourceLine(Line, 0, SrcFile);
3968 if (DW->getRecordSourceLineCount() != 1)
3969 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3970 getRoot(), LabelID));
3973 setCurDebugLoc(DebugLoc::get(MF.getOrCreateDebugLocID(SrcFile, Line, 0)));
3978 case Intrinsic::dbg_declare: {
3980 DwarfWriter *DW = DAG.getDwarfWriter();
3981 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
3982 Value *Variable = DI.getVariable();
3983 if (DW && DW->ValidDebugInfo(Variable))
3984 DAG.setRoot(DAG.getNode(ISD::DECLARE, dl, MVT::Other, getRoot(),
3985 getValue(DI.getAddress()), getValue(Variable)));
3987 // FIXME: Do something sensible here when we support debug declare.
3991 case Intrinsic::eh_exception: {
3992 if (!CurMBB->isLandingPad()) {
3993 // FIXME: Mark exception register as live in. Hack for PR1508.
3994 unsigned Reg = TLI.getExceptionAddressRegister();
3995 if (Reg) CurMBB->addLiveIn(Reg);
3997 // Insert the EXCEPTIONADDR instruction.
3998 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
4000 Ops[0] = DAG.getRoot();
4001 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
4003 DAG.setRoot(Op.getValue(1));
4007 case Intrinsic::eh_selector_i32:
4008 case Intrinsic::eh_selector_i64: {
4009 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4010 MVT VT = (Intrinsic == Intrinsic::eh_selector_i32 ?
4011 MVT::i32 : MVT::i64);
4014 if (CurMBB->isLandingPad())
4015 AddCatchInfo(I, MMI, CurMBB);
4018 FuncInfo.CatchInfoLost.insert(&I);
4020 // FIXME: Mark exception selector register as live in. Hack for PR1508.
4021 unsigned Reg = TLI.getExceptionSelectorRegister();
4022 if (Reg) CurMBB->addLiveIn(Reg);
4025 // Insert the EHSELECTION instruction.
4026 SDVTList VTs = DAG.getVTList(VT, MVT::Other);
4028 Ops[0] = getValue(I.getOperand(1));
4030 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4032 DAG.setRoot(Op.getValue(1));
4034 setValue(&I, DAG.getConstant(0, VT));
4040 case Intrinsic::eh_typeid_for_i32:
4041 case Intrinsic::eh_typeid_for_i64: {
4042 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4043 MVT VT = (Intrinsic == Intrinsic::eh_typeid_for_i32 ?
4044 MVT::i32 : MVT::i64);
4047 // Find the type id for the given typeinfo.
4048 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
4050 unsigned TypeID = MMI->getTypeIDFor(GV);
4051 setValue(&I, DAG.getConstant(TypeID, VT));
4053 // Return something different to eh_selector.
4054 setValue(&I, DAG.getConstant(1, VT));
4060 case Intrinsic::eh_return_i32:
4061 case Intrinsic::eh_return_i64:
4062 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4063 MMI->setCallsEHReturn(true);
4064 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4067 getValue(I.getOperand(1)),
4068 getValue(I.getOperand(2))));
4070 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
4074 case Intrinsic::eh_unwind_init:
4075 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4076 MMI->setCallsUnwindInit(true);
4081 case Intrinsic::eh_dwarf_cfa: {
4082 MVT VT = getValue(I.getOperand(1)).getValueType();
4084 if (VT.bitsGT(TLI.getPointerTy()))
4085 CfaArg = DAG.getNode(ISD::TRUNCATE, dl,
4086 TLI.getPointerTy(), getValue(I.getOperand(1)));
4088 CfaArg = DAG.getNode(ISD::SIGN_EXTEND, dl,
4089 TLI.getPointerTy(), getValue(I.getOperand(1)));
4091 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4093 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4094 TLI.getPointerTy()),
4096 setValue(&I, DAG.getNode(ISD::ADD, dl,
4098 DAG.getNode(ISD::FRAMEADDR, dl,
4101 TLI.getPointerTy())),
4106 case Intrinsic::convertff:
4107 case Intrinsic::convertfsi:
4108 case Intrinsic::convertfui:
4109 case Intrinsic::convertsif:
4110 case Intrinsic::convertuif:
4111 case Intrinsic::convertss:
4112 case Intrinsic::convertsu:
4113 case Intrinsic::convertus:
4114 case Intrinsic::convertuu: {
4115 ISD::CvtCode Code = ISD::CVT_INVALID;
4116 switch (Intrinsic) {
4117 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4118 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4119 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4120 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4121 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4122 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4123 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4124 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4125 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4127 MVT DestVT = TLI.getValueType(I.getType());
4128 Value* Op1 = I.getOperand(1);
4129 setValue(&I, DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4130 DAG.getValueType(DestVT),
4131 DAG.getValueType(getValue(Op1).getValueType()),
4132 getValue(I.getOperand(2)),
4133 getValue(I.getOperand(3)),
4138 case Intrinsic::sqrt:
4139 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4140 getValue(I.getOperand(1)).getValueType(),
4141 getValue(I.getOperand(1))));
4143 case Intrinsic::powi:
4144 setValue(&I, DAG.getNode(ISD::FPOWI, dl,
4145 getValue(I.getOperand(1)).getValueType(),
4146 getValue(I.getOperand(1)),
4147 getValue(I.getOperand(2))));
4149 case Intrinsic::sin:
4150 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4151 getValue(I.getOperand(1)).getValueType(),
4152 getValue(I.getOperand(1))));
4154 case Intrinsic::cos:
4155 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4156 getValue(I.getOperand(1)).getValueType(),
4157 getValue(I.getOperand(1))));
4159 case Intrinsic::log:
4162 case Intrinsic::log2:
4165 case Intrinsic::log10:
4168 case Intrinsic::exp:
4171 case Intrinsic::exp2:
4174 case Intrinsic::pow:
4177 case Intrinsic::pcmarker: {
4178 SDValue Tmp = getValue(I.getOperand(1));
4179 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4182 case Intrinsic::readcyclecounter: {
4183 SDValue Op = getRoot();
4184 SDValue Tmp = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4185 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
4188 DAG.setRoot(Tmp.getValue(1));
4191 case Intrinsic::part_select: {
4192 // Currently not implemented: just abort
4193 assert(0 && "part_select intrinsic not implemented");
4196 case Intrinsic::part_set: {
4197 // Currently not implemented: just abort
4198 assert(0 && "part_set intrinsic not implemented");
4201 case Intrinsic::bswap:
4202 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4203 getValue(I.getOperand(1)).getValueType(),
4204 getValue(I.getOperand(1))));
4206 case Intrinsic::cttz: {
4207 SDValue Arg = getValue(I.getOperand(1));
4208 MVT Ty = Arg.getValueType();
4209 SDValue result = DAG.getNode(ISD::CTTZ, dl, Ty, Arg);
4210 setValue(&I, result);
4213 case Intrinsic::ctlz: {
4214 SDValue Arg = getValue(I.getOperand(1));
4215 MVT Ty = Arg.getValueType();
4216 SDValue result = DAG.getNode(ISD::CTLZ, dl, Ty, Arg);
4217 setValue(&I, result);
4220 case Intrinsic::ctpop: {
4221 SDValue Arg = getValue(I.getOperand(1));
4222 MVT Ty = Arg.getValueType();
4223 SDValue result = DAG.getNode(ISD::CTPOP, dl, Ty, Arg);
4224 setValue(&I, result);
4227 case Intrinsic::stacksave: {
4228 SDValue Op = getRoot();
4229 SDValue Tmp = DAG.getNode(ISD::STACKSAVE, dl,
4230 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
4232 DAG.setRoot(Tmp.getValue(1));
4235 case Intrinsic::stackrestore: {
4236 SDValue Tmp = getValue(I.getOperand(1));
4237 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Tmp));
4240 case Intrinsic::stackprotector: {
4241 // Emit code into the DAG to store the stack guard onto the stack.
4242 MachineFunction &MF = DAG.getMachineFunction();
4243 MachineFrameInfo *MFI = MF.getFrameInfo();
4244 MVT PtrTy = TLI.getPointerTy();
4246 SDValue Src = getValue(I.getOperand(1)); // The guard's value.
4247 AllocaInst *Slot = cast<AllocaInst>(I.getOperand(2));
4249 int FI = FuncInfo.StaticAllocaMap[Slot];
4250 MFI->setStackProtectorIndex(FI);
4252 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4254 // Store the stack protector onto the stack.
4255 SDValue Result = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4256 PseudoSourceValue::getFixedStack(FI),
4258 setValue(&I, Result);
4259 DAG.setRoot(Result);
4262 case Intrinsic::var_annotation:
4263 // Discard annotate attributes
4266 case Intrinsic::init_trampoline: {
4267 const Function *F = cast<Function>(I.getOperand(2)->stripPointerCasts());
4271 Ops[1] = getValue(I.getOperand(1));
4272 Ops[2] = getValue(I.getOperand(2));
4273 Ops[3] = getValue(I.getOperand(3));
4274 Ops[4] = DAG.getSrcValue(I.getOperand(1));
4275 Ops[5] = DAG.getSrcValue(F);
4277 SDValue Tmp = DAG.getNode(ISD::TRAMPOLINE, dl,
4278 DAG.getNodeValueTypes(TLI.getPointerTy(),
4283 DAG.setRoot(Tmp.getValue(1));
4287 case Intrinsic::gcroot:
4289 Value *Alloca = I.getOperand(1);
4290 Constant *TypeMap = cast<Constant>(I.getOperand(2));
4292 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4293 GFI->addStackRoot(FI->getIndex(), TypeMap);
4297 case Intrinsic::gcread:
4298 case Intrinsic::gcwrite:
4299 assert(0 && "GC failed to lower gcread/gcwrite intrinsics!");
4302 case Intrinsic::flt_rounds: {
4303 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4307 case Intrinsic::trap: {
4308 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4312 case Intrinsic::uadd_with_overflow:
4313 return implVisitAluOverflow(I, ISD::UADDO);
4314 case Intrinsic::sadd_with_overflow:
4315 return implVisitAluOverflow(I, ISD::SADDO);
4316 case Intrinsic::usub_with_overflow:
4317 return implVisitAluOverflow(I, ISD::USUBO);
4318 case Intrinsic::ssub_with_overflow:
4319 return implVisitAluOverflow(I, ISD::SSUBO);
4320 case Intrinsic::umul_with_overflow:
4321 return implVisitAluOverflow(I, ISD::UMULO);
4322 case Intrinsic::smul_with_overflow:
4323 return implVisitAluOverflow(I, ISD::SMULO);
4325 case Intrinsic::prefetch: {
4328 Ops[1] = getValue(I.getOperand(1));
4329 Ops[2] = getValue(I.getOperand(2));
4330 Ops[3] = getValue(I.getOperand(3));
4331 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4335 case Intrinsic::memory_barrier: {
4338 for (int x = 1; x < 6; ++x)
4339 Ops[x] = getValue(I.getOperand(x));
4341 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4344 case Intrinsic::atomic_cmp_swap: {
4345 SDValue Root = getRoot();
4347 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4348 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
4350 getValue(I.getOperand(1)),
4351 getValue(I.getOperand(2)),
4352 getValue(I.getOperand(3)),
4355 DAG.setRoot(L.getValue(1));
4358 case Intrinsic::atomic_load_add:
4359 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4360 case Intrinsic::atomic_load_sub:
4361 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4362 case Intrinsic::atomic_load_or:
4363 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4364 case Intrinsic::atomic_load_xor:
4365 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4366 case Intrinsic::atomic_load_and:
4367 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4368 case Intrinsic::atomic_load_nand:
4369 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4370 case Intrinsic::atomic_load_max:
4371 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4372 case Intrinsic::atomic_load_min:
4373 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4374 case Intrinsic::atomic_load_umin:
4375 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4376 case Intrinsic::atomic_load_umax:
4377 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4378 case Intrinsic::atomic_swap:
4379 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4384 void SelectionDAGLowering::LowerCallTo(CallSite CS, SDValue Callee,
4386 MachineBasicBlock *LandingPad) {
4387 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4388 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4389 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4390 unsigned BeginLabel = 0, EndLabel = 0;
4392 TargetLowering::ArgListTy Args;
4393 TargetLowering::ArgListEntry Entry;
4394 Args.reserve(CS.arg_size());
4395 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4397 SDValue ArgNode = getValue(*i);
4398 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4400 unsigned attrInd = i - CS.arg_begin() + 1;
4401 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4402 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4403 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4404 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4405 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4406 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4407 Entry.Alignment = CS.getParamAlignment(attrInd);
4408 Args.push_back(Entry);
4411 if (LandingPad && MMI) {
4412 // Insert a label before the invoke call to mark the try range. This can be
4413 // used to detect deletion of the invoke via the MachineModuleInfo.
4414 BeginLabel = MMI->NextLabelID();
4415 // Both PendingLoads and PendingExports must be flushed here;
4416 // this call might not return.
4418 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4419 getControlRoot(), BeginLabel));
4422 std::pair<SDValue,SDValue> Result =
4423 TLI.LowerCallTo(getRoot(), CS.getType(),
4424 CS.paramHasAttr(0, Attribute::SExt),
4425 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4426 CS.paramHasAttr(0, Attribute::InReg),
4427 CS.getCallingConv(),
4428 IsTailCall && PerformTailCallOpt,
4429 Callee, Args, DAG, getCurDebugLoc());
4430 if (CS.getType() != Type::VoidTy)
4431 setValue(CS.getInstruction(), Result.first);
4432 DAG.setRoot(Result.second);
4434 if (LandingPad && MMI) {
4435 // Insert a label at the end of the invoke call to mark the try range. This
4436 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4437 EndLabel = MMI->NextLabelID();
4438 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4439 getRoot(), EndLabel));
4441 // Inform MachineModuleInfo of range.
4442 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
4447 void SelectionDAGLowering::visitCall(CallInst &I) {
4448 const char *RenameFn = 0;
4449 if (Function *F = I.getCalledFunction()) {
4450 if (F->isDeclaration()) {
4451 const TargetIntrinsicInfo *II = TLI.getTargetMachine().getIntrinsicInfo();
4453 if (unsigned IID = II->getIntrinsicID(F)) {
4454 RenameFn = visitIntrinsicCall(I, IID);
4459 if (unsigned IID = F->getIntrinsicID()) {
4460 RenameFn = visitIntrinsicCall(I, IID);
4466 // Check for well-known libc/libm calls. If the function is internal, it
4467 // can't be a library call.
4468 unsigned NameLen = F->getNameLen();
4469 if (!F->hasLocalLinkage() && NameLen) {
4470 const char *NameStr = F->getNameStart();
4471 if (NameStr[0] == 'c' &&
4472 ((NameLen == 8 && !strcmp(NameStr, "copysign")) ||
4473 (NameLen == 9 && !strcmp(NameStr, "copysignf")))) {
4474 if (I.getNumOperands() == 3 && // Basic sanity checks.
4475 I.getOperand(1)->getType()->isFloatingPoint() &&
4476 I.getType() == I.getOperand(1)->getType() &&
4477 I.getType() == I.getOperand(2)->getType()) {
4478 SDValue LHS = getValue(I.getOperand(1));
4479 SDValue RHS = getValue(I.getOperand(2));
4480 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4481 LHS.getValueType(), LHS, RHS));
4484 } else if (NameStr[0] == 'f' &&
4485 ((NameLen == 4 && !strcmp(NameStr, "fabs")) ||
4486 (NameLen == 5 && !strcmp(NameStr, "fabsf")) ||
4487 (NameLen == 5 && !strcmp(NameStr, "fabsl")))) {
4488 if (I.getNumOperands() == 2 && // Basic sanity checks.
4489 I.getOperand(1)->getType()->isFloatingPoint() &&
4490 I.getType() == I.getOperand(1)->getType()) {
4491 SDValue Tmp = getValue(I.getOperand(1));
4492 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4493 Tmp.getValueType(), Tmp));
4496 } else if (NameStr[0] == 's' &&
4497 ((NameLen == 3 && !strcmp(NameStr, "sin")) ||
4498 (NameLen == 4 && !strcmp(NameStr, "sinf")) ||
4499 (NameLen == 4 && !strcmp(NameStr, "sinl")))) {
4500 if (I.getNumOperands() == 2 && // Basic sanity checks.
4501 I.getOperand(1)->getType()->isFloatingPoint() &&
4502 I.getType() == I.getOperand(1)->getType()) {
4503 SDValue Tmp = getValue(I.getOperand(1));
4504 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4505 Tmp.getValueType(), Tmp));
4508 } else if (NameStr[0] == 'c' &&
4509 ((NameLen == 3 && !strcmp(NameStr, "cos")) ||
4510 (NameLen == 4 && !strcmp(NameStr, "cosf")) ||
4511 (NameLen == 4 && !strcmp(NameStr, "cosl")))) {
4512 if (I.getNumOperands() == 2 && // Basic sanity checks.
4513 I.getOperand(1)->getType()->isFloatingPoint() &&
4514 I.getType() == I.getOperand(1)->getType()) {
4515 SDValue Tmp = getValue(I.getOperand(1));
4516 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4517 Tmp.getValueType(), Tmp));
4522 } else if (isa<InlineAsm>(I.getOperand(0))) {
4529 Callee = getValue(I.getOperand(0));
4531 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4533 LowerCallTo(&I, Callee, I.isTailCall());
4537 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
4538 /// this value and returns the result as a ValueVT value. This uses
4539 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4540 /// If the Flag pointer is NULL, no flag is used.
4541 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl,
4543 SDValue *Flag) const {
4544 // Assemble the legal parts into the final values.
4545 SmallVector<SDValue, 4> Values(ValueVTs.size());
4546 SmallVector<SDValue, 8> Parts;
4547 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4548 // Copy the legal parts from the registers.
4549 MVT ValueVT = ValueVTs[Value];
4550 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
4551 MVT RegisterVT = RegVTs[Value];
4553 Parts.resize(NumRegs);
4554 for (unsigned i = 0; i != NumRegs; ++i) {
4557 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
4559 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
4560 *Flag = P.getValue(2);
4562 Chain = P.getValue(1);
4564 // If the source register was virtual and if we know something about it,
4565 // add an assert node.
4566 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
4567 RegisterVT.isInteger() && !RegisterVT.isVector()) {
4568 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
4569 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
4570 if (FLI.LiveOutRegInfo.size() > SlotNo) {
4571 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[SlotNo];
4573 unsigned RegSize = RegisterVT.getSizeInBits();
4574 unsigned NumSignBits = LOI.NumSignBits;
4575 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
4577 // FIXME: We capture more information than the dag can represent. For
4578 // now, just use the tightest assertzext/assertsext possible.
4580 MVT FromVT(MVT::Other);
4581 if (NumSignBits == RegSize)
4582 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
4583 else if (NumZeroBits >= RegSize-1)
4584 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
4585 else if (NumSignBits > RegSize-8)
4586 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
4587 else if (NumZeroBits >= RegSize-9)
4588 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
4589 else if (NumSignBits > RegSize-16)
4590 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
4591 else if (NumZeroBits >= RegSize-17)
4592 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
4593 else if (NumSignBits > RegSize-32)
4594 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
4595 else if (NumZeroBits >= RegSize-33)
4596 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
4598 if (FromVT != MVT::Other) {
4599 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
4600 RegisterVT, P, DAG.getValueType(FromVT));
4609 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
4610 NumRegs, RegisterVT, ValueVT);
4615 return DAG.getNode(ISD::MERGE_VALUES, dl,
4616 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
4617 &Values[0], ValueVTs.size());
4620 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
4621 /// specified value into the registers specified by this object. This uses
4622 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4623 /// If the Flag pointer is NULL, no flag is used.
4624 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
4625 SDValue &Chain, SDValue *Flag) const {
4626 // Get the list of the values's legal parts.
4627 unsigned NumRegs = Regs.size();
4628 SmallVector<SDValue, 8> Parts(NumRegs);
4629 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4630 MVT ValueVT = ValueVTs[Value];
4631 unsigned NumParts = TLI->getNumRegisters(ValueVT);
4632 MVT RegisterVT = RegVTs[Value];
4634 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
4635 &Parts[Part], NumParts, RegisterVT);
4639 // Copy the parts into the registers.
4640 SmallVector<SDValue, 8> Chains(NumRegs);
4641 for (unsigned i = 0; i != NumRegs; ++i) {
4644 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
4646 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
4647 *Flag = Part.getValue(1);
4649 Chains[i] = Part.getValue(0);
4652 if (NumRegs == 1 || Flag)
4653 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
4654 // flagged to it. That is the CopyToReg nodes and the user are considered
4655 // a single scheduling unit. If we create a TokenFactor and return it as
4656 // chain, then the TokenFactor is both a predecessor (operand) of the
4657 // user as well as a successor (the TF operands are flagged to the user).
4658 // c1, f1 = CopyToReg
4659 // c2, f2 = CopyToReg
4660 // c3 = TokenFactor c1, c2
4663 Chain = Chains[NumRegs-1];
4665 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
4668 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
4669 /// operand list. This adds the code marker and includes the number of
4670 /// values added into it.
4671 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
4672 std::vector<SDValue> &Ops) const {
4673 MVT IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
4674 Ops.push_back(DAG.getTargetConstant(Code | (Regs.size() << 3), IntPtrTy));
4675 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
4676 unsigned NumRegs = TLI->getNumRegisters(ValueVTs[Value]);
4677 MVT RegisterVT = RegVTs[Value];
4678 for (unsigned i = 0; i != NumRegs; ++i) {
4679 assert(Reg < Regs.size() && "Mismatch in # registers expected");
4680 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
4685 /// isAllocatableRegister - If the specified register is safe to allocate,
4686 /// i.e. it isn't a stack pointer or some other special register, return the
4687 /// register class for the register. Otherwise, return null.
4688 static const TargetRegisterClass *
4689 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
4690 const TargetLowering &TLI,
4691 const TargetRegisterInfo *TRI) {
4692 MVT FoundVT = MVT::Other;
4693 const TargetRegisterClass *FoundRC = 0;
4694 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
4695 E = TRI->regclass_end(); RCI != E; ++RCI) {
4696 MVT ThisVT = MVT::Other;
4698 const TargetRegisterClass *RC = *RCI;
4699 // If none of the the value types for this register class are valid, we
4700 // can't use it. For example, 64-bit reg classes on 32-bit targets.
4701 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
4703 if (TLI.isTypeLegal(*I)) {
4704 // If we have already found this register in a different register class,
4705 // choose the one with the largest VT specified. For example, on
4706 // PowerPC, we favor f64 register classes over f32.
4707 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
4714 if (ThisVT == MVT::Other) continue;
4716 // NOTE: This isn't ideal. In particular, this might allocate the
4717 // frame pointer in functions that need it (due to them not being taken
4718 // out of allocation, because a variable sized allocation hasn't been seen
4719 // yet). This is a slight code pessimization, but should still work.
4720 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
4721 E = RC->allocation_order_end(MF); I != E; ++I)
4723 // We found a matching register class. Keep looking at others in case
4724 // we find one with larger registers that this physreg is also in.
4735 /// AsmOperandInfo - This contains information for each constraint that we are
4737 class VISIBILITY_HIDDEN SDISelAsmOperandInfo :
4738 public TargetLowering::AsmOperandInfo {
4740 /// CallOperand - If this is the result output operand or a clobber
4741 /// this is null, otherwise it is the incoming operand to the CallInst.
4742 /// This gets modified as the asm is processed.
4743 SDValue CallOperand;
4745 /// AssignedRegs - If this is a register or register class operand, this
4746 /// contains the set of register corresponding to the operand.
4747 RegsForValue AssignedRegs;
4749 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
4750 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
4753 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
4754 /// busy in OutputRegs/InputRegs.
4755 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
4756 std::set<unsigned> &OutputRegs,
4757 std::set<unsigned> &InputRegs,
4758 const TargetRegisterInfo &TRI) const {
4760 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4761 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
4764 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4765 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
4769 /// getCallOperandValMVT - Return the MVT of the Value* that this operand
4770 /// corresponds to. If there is no Value* for this operand, it returns
4772 MVT getCallOperandValMVT(const TargetLowering &TLI,
4773 const TargetData *TD) const {
4774 if (CallOperandVal == 0) return MVT::Other;
4776 if (isa<BasicBlock>(CallOperandVal))
4777 return TLI.getPointerTy();
4779 const llvm::Type *OpTy = CallOperandVal->getType();
4781 // If this is an indirect operand, the operand is a pointer to the
4784 OpTy = cast<PointerType>(OpTy)->getElementType();
4786 // If OpTy is not a single value, it may be a struct/union that we
4787 // can tile with integers.
4788 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
4789 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
4798 OpTy = IntegerType::get(BitSize);
4803 return TLI.getValueType(OpTy, true);
4807 /// MarkRegAndAliases - Mark the specified register and all aliases in the
4809 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
4810 const TargetRegisterInfo &TRI) {
4811 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
4813 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
4814 for (; *Aliases; ++Aliases)
4815 Regs.insert(*Aliases);
4818 } // end llvm namespace.
4821 /// GetRegistersForValue - Assign registers (virtual or physical) for the
4822 /// specified operand. We prefer to assign virtual registers, to allow the
4823 /// register allocator handle the assignment process. However, if the asm uses
4824 /// features that we can't model on machineinstrs, we have SDISel do the
4825 /// allocation. This produces generally horrible, but correct, code.
4827 /// OpInfo describes the operand.
4828 /// Input and OutputRegs are the set of already allocated physical registers.
4830 void SelectionDAGLowering::
4831 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
4832 std::set<unsigned> &OutputRegs,
4833 std::set<unsigned> &InputRegs) {
4834 // Compute whether this value requires an input register, an output register,
4836 bool isOutReg = false;
4837 bool isInReg = false;
4838 switch (OpInfo.Type) {
4839 case InlineAsm::isOutput:
4842 // If there is an input constraint that matches this, we need to reserve
4843 // the input register so no other inputs allocate to it.
4844 isInReg = OpInfo.hasMatchingInput();
4846 case InlineAsm::isInput:
4850 case InlineAsm::isClobber:
4857 MachineFunction &MF = DAG.getMachineFunction();
4858 SmallVector<unsigned, 4> Regs;
4860 // If this is a constraint for a single physreg, or a constraint for a
4861 // register class, find it.
4862 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
4863 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
4864 OpInfo.ConstraintVT);
4866 unsigned NumRegs = 1;
4867 if (OpInfo.ConstraintVT != MVT::Other) {
4868 // If this is a FP input in an integer register (or visa versa) insert a bit
4869 // cast of the input value. More generally, handle any case where the input
4870 // value disagrees with the register class we plan to stick this in.
4871 if (OpInfo.Type == InlineAsm::isInput &&
4872 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
4873 // Try to convert to the first MVT that the reg class contains. If the
4874 // types are identical size, use a bitcast to convert (e.g. two differing
4876 MVT RegVT = *PhysReg.second->vt_begin();
4877 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
4878 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
4879 RegVT, OpInfo.CallOperand);
4880 OpInfo.ConstraintVT = RegVT;
4881 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
4882 // If the input is a FP value and we want it in FP registers, do a
4883 // bitcast to the corresponding integer type. This turns an f64 value
4884 // into i64, which can be passed with two i32 values on a 32-bit
4886 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
4887 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
4888 RegVT, OpInfo.CallOperand);
4889 OpInfo.ConstraintVT = RegVT;
4893 NumRegs = TLI.getNumRegisters(OpInfo.ConstraintVT);
4897 MVT ValueVT = OpInfo.ConstraintVT;
4899 // If this is a constraint for a specific physical register, like {r17},
4901 if (PhysReg.first) {
4902 if (OpInfo.ConstraintVT == MVT::Other)
4903 ValueVT = *PhysReg.second->vt_begin();
4905 // Get the actual register value type. This is important, because the user
4906 // may have asked for (e.g.) the AX register in i32 type. We need to
4907 // remember that AX is actually i16 to get the right extension.
4908 RegVT = *PhysReg.second->vt_begin();
4910 // This is a explicit reference to a physical register.
4911 Regs.push_back(PhysReg.first);
4913 // If this is an expanded reference, add the rest of the regs to Regs.
4915 TargetRegisterClass::iterator I = PhysReg.second->begin();
4916 for (; *I != PhysReg.first; ++I)
4917 assert(I != PhysReg.second->end() && "Didn't find reg!");
4919 // Already added the first reg.
4921 for (; NumRegs; --NumRegs, ++I) {
4922 assert(I != PhysReg.second->end() && "Ran out of registers to allocate!");
4926 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4927 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4928 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
4932 // Otherwise, if this was a reference to an LLVM register class, create vregs
4933 // for this reference.
4934 std::vector<unsigned> RegClassRegs;
4935 const TargetRegisterClass *RC = PhysReg.second;
4937 // If this is a tied register, our regalloc doesn't know how to maintain
4938 // the constraint, so we have to pick a register to pin the input/output to.
4939 // If it isn't a matched constraint, go ahead and create vreg and let the
4940 // regalloc do its thing.
4941 if (!OpInfo.hasMatchingInput()) {
4942 RegVT = *PhysReg.second->vt_begin();
4943 if (OpInfo.ConstraintVT == MVT::Other)
4946 // Create the appropriate number of virtual registers.
4947 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4948 for (; NumRegs; --NumRegs)
4949 Regs.push_back(RegInfo.createVirtualRegister(PhysReg.second));
4951 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4955 // Otherwise, we can't allocate it. Let the code below figure out how to
4956 // maintain these constraints.
4957 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end());
4960 // This is a reference to a register class that doesn't directly correspond
4961 // to an LLVM register class. Allocate NumRegs consecutive, available,
4962 // registers from the class.
4963 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
4964 OpInfo.ConstraintVT);
4967 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4968 unsigned NumAllocated = 0;
4969 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
4970 unsigned Reg = RegClassRegs[i];
4971 // See if this register is available.
4972 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
4973 (isInReg && InputRegs.count(Reg))) { // Already used.
4974 // Make sure we find consecutive registers.
4979 // Check to see if this register is allocatable (i.e. don't give out the
4982 RC = isAllocatableRegister(Reg, MF, TLI, TRI);
4983 if (!RC) { // Couldn't allocate this register.
4984 // Reset NumAllocated to make sure we return consecutive registers.
4990 // Okay, this register is good, we can use it.
4993 // If we allocated enough consecutive registers, succeed.
4994 if (NumAllocated == NumRegs) {
4995 unsigned RegStart = (i-NumAllocated)+1;
4996 unsigned RegEnd = i+1;
4997 // Mark all of the allocated registers used.
4998 for (unsigned i = RegStart; i != RegEnd; ++i)
4999 Regs.push_back(RegClassRegs[i]);
5001 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, *RC->vt_begin(),
5002 OpInfo.ConstraintVT);
5003 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
5008 // Otherwise, we couldn't allocate enough registers for this.
5011 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
5012 /// processed uses a memory 'm' constraint.
5014 hasInlineAsmMemConstraint(std::vector<InlineAsm::ConstraintInfo> &CInfos,
5015 const TargetLowering &TLI) {
5016 for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
5017 InlineAsm::ConstraintInfo &CI = CInfos[i];
5018 for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
5019 TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
5020 if (CType == TargetLowering::C_Memory)
5028 /// visitInlineAsm - Handle a call to an InlineAsm object.
5030 void SelectionDAGLowering::visitInlineAsm(CallSite CS) {
5031 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
5033 /// ConstraintOperands - Information about all of the constraints.
5034 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
5036 SDValue Chain = getRoot();
5039 std::set<unsigned> OutputRegs, InputRegs;
5041 // Do a prepass over the constraints, canonicalizing them, and building up the
5042 // ConstraintOperands list.
5043 std::vector<InlineAsm::ConstraintInfo>
5044 ConstraintInfos = IA->ParseConstraints();
5046 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5048 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5049 unsigned ResNo = 0; // ResNo - The result number of the next output.
5050 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5051 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5052 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5054 MVT OpVT = MVT::Other;
5056 // Compute the value type for each operand.
5057 switch (OpInfo.Type) {
5058 case InlineAsm::isOutput:
5059 // Indirect outputs just consume an argument.
5060 if (OpInfo.isIndirect) {
5061 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5065 // The return value of the call is this value. As such, there is no
5066 // corresponding argument.
5067 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
5068 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5069 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5071 assert(ResNo == 0 && "Asm only has one result!");
5072 OpVT = TLI.getValueType(CS.getType());
5076 case InlineAsm::isInput:
5077 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5079 case InlineAsm::isClobber:
5084 // If this is an input or an indirect output, process the call argument.
5085 // BasicBlocks are labels, currently appearing only in asm's.
5086 if (OpInfo.CallOperandVal) {
5087 if (BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5088 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5090 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5093 OpVT = OpInfo.getCallOperandValMVT(TLI, TD);
5096 OpInfo.ConstraintVT = OpVT;
5099 // Second pass over the constraints: compute which constraint option to use
5100 // and assign registers to constraints that want a specific physreg.
5101 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5102 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5104 // If this is an output operand with a matching input operand, look up the
5105 // matching input. If their types mismatch, e.g. one is an integer, the
5106 // other is floating point, or their sizes are different, flag it as an
5108 if (OpInfo.hasMatchingInput()) {
5109 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5110 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5111 if ((OpInfo.ConstraintVT.isInteger() !=
5112 Input.ConstraintVT.isInteger()) ||
5113 (OpInfo.ConstraintVT.getSizeInBits() !=
5114 Input.ConstraintVT.getSizeInBits())) {
5115 cerr << "Unsupported asm: input constraint with a matching output "
5116 << "constraint of incompatible type!\n";
5119 Input.ConstraintVT = OpInfo.ConstraintVT;
5123 // Compute the constraint code and ConstraintType to use.
5124 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, hasMemory, &DAG);
5126 // If this is a memory input, and if the operand is not indirect, do what we
5127 // need to to provide an address for the memory input.
5128 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5129 !OpInfo.isIndirect) {
5130 assert(OpInfo.Type == InlineAsm::isInput &&
5131 "Can only indirectify direct input operands!");
5133 // Memory operands really want the address of the value. If we don't have
5134 // an indirect input, put it in the constpool if we can, otherwise spill
5135 // it to a stack slot.
5137 // If the operand is a float, integer, or vector constant, spill to a
5138 // constant pool entry to get its address.
5139 Value *OpVal = OpInfo.CallOperandVal;
5140 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5141 isa<ConstantVector>(OpVal)) {
5142 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5143 TLI.getPointerTy());
5145 // Otherwise, create a stack slot and emit a store to it before the
5147 const Type *Ty = OpVal->getType();
5148 uint64_t TySize = TLI.getTargetData()->getTypePaddedSize(Ty);
5149 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5150 MachineFunction &MF = DAG.getMachineFunction();
5151 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align);
5152 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5153 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5154 OpInfo.CallOperand, StackSlot, NULL, 0);
5155 OpInfo.CallOperand = StackSlot;
5158 // There is no longer a Value* corresponding to this operand.
5159 OpInfo.CallOperandVal = 0;
5160 // It is now an indirect operand.
5161 OpInfo.isIndirect = true;
5164 // If this constraint is for a specific register, allocate it before
5166 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5167 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5169 ConstraintInfos.clear();
5172 // Second pass - Loop over all of the operands, assigning virtual or physregs
5173 // to register class operands.
5174 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5175 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5177 // C_Register operands have already been allocated, Other/Memory don't need
5179 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5180 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5183 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5184 std::vector<SDValue> AsmNodeOperands;
5185 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5186 AsmNodeOperands.push_back(
5187 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other));
5190 // Loop over all of the inputs, copying the operand values into the
5191 // appropriate registers and processing the output regs.
5192 RegsForValue RetValRegs;
5194 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5195 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5197 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5198 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5200 switch (OpInfo.Type) {
5201 case InlineAsm::isOutput: {
5202 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5203 OpInfo.ConstraintType != TargetLowering::C_Register) {
5204 // Memory output, or 'other' output (e.g. 'X' constraint).
5205 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5207 // Add information to the INLINEASM node to know about this output.
5208 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5209 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5210 TLI.getPointerTy()));
5211 AsmNodeOperands.push_back(OpInfo.CallOperand);
5215 // Otherwise, this is a register or register class output.
5217 // Copy the output from the appropriate register. Find a register that
5219 if (OpInfo.AssignedRegs.Regs.empty()) {
5220 cerr << "Couldn't allocate output reg for constraint '"
5221 << OpInfo.ConstraintCode << "'!\n";
5225 // If this is an indirect operand, store through the pointer after the
5227 if (OpInfo.isIndirect) {
5228 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5229 OpInfo.CallOperandVal));
5231 // This is the result value of the call.
5232 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
5233 // Concatenate this output onto the outputs list.
5234 RetValRegs.append(OpInfo.AssignedRegs);
5237 // Add information to the INLINEASM node to know that this register is
5239 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5240 6 /* EARLYCLOBBER REGDEF */ :
5242 DAG, AsmNodeOperands);
5245 case InlineAsm::isInput: {
5246 SDValue InOperandVal = OpInfo.CallOperand;
5248 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5249 // If this is required to match an output register we have already set,
5250 // just use its register.
5251 unsigned OperandNo = OpInfo.getMatchedOperand();
5253 // Scan until we find the definition we already emitted of this operand.
5254 // When we find it, create a RegsForValue operand.
5255 unsigned CurOp = 2; // The first operand.
5256 for (; OperandNo; --OperandNo) {
5257 // Advance to the next operand.
5259 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5260 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
5261 (NumOps & 7) == 6 /*EARLYCLOBBER REGDEF*/ ||
5262 (NumOps & 7) == 4 /*MEM*/) &&
5263 "Skipped past definitions?");
5264 CurOp += (NumOps>>3)+1;
5268 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5269 if ((NumOps & 7) == 2 /*REGDEF*/
5270 || (NumOps & 7) == 6 /* EARLYCLOBBER REGDEF */) {
5271 // Add NumOps>>3 registers to MatchedRegs.
5272 RegsForValue MatchedRegs;
5273 MatchedRegs.TLI = &TLI;
5274 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5275 MatchedRegs.RegVTs.push_back(AsmNodeOperands[CurOp+1].getValueType());
5276 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
5278 cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
5279 MatchedRegs.Regs.push_back(Reg);
5282 // Use the produced MatchedRegs object to
5283 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5285 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
5288 assert(((NumOps & 7) == 4) && "Unknown matching constraint!");
5289 assert((NumOps >> 3) == 1 && "Unexpected number of operands");
5290 // Add information to the INLINEASM node to know about this input.
5291 AsmNodeOperands.push_back(DAG.getTargetConstant(NumOps,
5292 TLI.getPointerTy()));
5293 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5298 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5299 assert(!OpInfo.isIndirect &&
5300 "Don't know how to handle indirect other inputs yet!");
5302 std::vector<SDValue> Ops;
5303 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5304 hasMemory, Ops, DAG);
5306 cerr << "Invalid operand for inline asm constraint '"
5307 << OpInfo.ConstraintCode << "'!\n";
5311 // Add information to the INLINEASM node to know about this input.
5312 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
5313 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5314 TLI.getPointerTy()));
5315 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5317 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5318 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5319 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5320 "Memory operands expect pointer values");
5322 // Add information to the INLINEASM node to know about this input.
5323 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5324 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5325 TLI.getPointerTy()));
5326 AsmNodeOperands.push_back(InOperandVal);
5330 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5331 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5332 "Unknown constraint type!");
5333 assert(!OpInfo.isIndirect &&
5334 "Don't know how to handle indirect register inputs yet!");
5336 // Copy the input into the appropriate registers.
5337 if (OpInfo.AssignedRegs.Regs.empty()) {
5338 cerr << "Couldn't allocate output reg for constraint '"
5339 << OpInfo.ConstraintCode << "'!\n";
5343 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5346 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/,
5347 DAG, AsmNodeOperands);
5350 case InlineAsm::isClobber: {
5351 // Add the clobbered value to the operand list, so that the register
5352 // allocator is aware that the physreg got clobbered.
5353 if (!OpInfo.AssignedRegs.Regs.empty())
5354 OpInfo.AssignedRegs.AddInlineAsmOperands(6 /* EARLYCLOBBER REGDEF */,
5355 DAG, AsmNodeOperands);
5361 // Finish up input operands.
5362 AsmNodeOperands[0] = Chain;
5363 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5365 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5366 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
5367 &AsmNodeOperands[0], AsmNodeOperands.size());
5368 Flag = Chain.getValue(1);
5370 // If this asm returns a register value, copy the result from that register
5371 // and set it as the value of the call.
5372 if (!RetValRegs.Regs.empty()) {
5373 SDValue Val = RetValRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
5376 // FIXME: Why don't we do this for inline asms with MRVs?
5377 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5378 MVT ResultType = TLI.getValueType(CS.getType());
5380 // If any of the results of the inline asm is a vector, it may have the
5381 // wrong width/num elts. This can happen for register classes that can
5382 // contain multiple different value types. The preg or vreg allocated may
5383 // not have the same VT as was expected. Convert it to the right type
5384 // with bit_convert.
5385 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5386 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5389 } else if (ResultType != Val.getValueType() &&
5390 ResultType.isInteger() && Val.getValueType().isInteger()) {
5391 // If a result value was tied to an input value, the computed result may
5392 // have a wider width than the expected result. Extract the relevant
5394 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5397 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5400 setValue(CS.getInstruction(), Val);
5403 std::vector<std::pair<SDValue, Value*> > StoresToEmit;
5405 // Process indirect outputs, first output all of the flagged copies out of
5407 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5408 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5409 Value *Ptr = IndirectStoresToEmit[i].second;
5410 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
5412 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5415 // Emit the non-flagged stores from the physregs.
5416 SmallVector<SDValue, 8> OutChains;
5417 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
5418 OutChains.push_back(DAG.getStore(Chain, getCurDebugLoc(),
5419 StoresToEmit[i].first,
5420 getValue(StoresToEmit[i].second),
5421 StoresToEmit[i].second, 0));
5422 if (!OutChains.empty())
5423 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5424 &OutChains[0], OutChains.size());
5429 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
5430 SDValue Src = getValue(I.getOperand(0));
5432 // Scale up by the type size in the original i32 type width. Various
5433 // mid-level optimizers may make assumptions about demanded bits etc from the
5434 // i32-ness of the optimizer: we do not want to promote to i64 and then
5435 // multiply on 64-bit targets.
5436 // FIXME: Malloc inst should go away: PR715.
5437 uint64_t ElementSize = TD->getTypePaddedSize(I.getType()->getElementType());
5438 if (ElementSize != 1)
5439 Src = DAG.getNode(ISD::MUL, getCurDebugLoc(), Src.getValueType(),
5440 Src, DAG.getConstant(ElementSize, Src.getValueType()));
5442 MVT IntPtr = TLI.getPointerTy();
5444 if (IntPtr.bitsLT(Src.getValueType()))
5445 Src = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), IntPtr, Src);
5446 else if (IntPtr.bitsGT(Src.getValueType()))
5447 Src = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), IntPtr, Src);
5449 TargetLowering::ArgListTy Args;
5450 TargetLowering::ArgListEntry Entry;
5452 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5453 Args.push_back(Entry);
5455 std::pair<SDValue,SDValue> Result =
5456 TLI.LowerCallTo(getRoot(), I.getType(), false, false, false, false,
5457 CallingConv::C, PerformTailCallOpt,
5458 DAG.getExternalSymbol("malloc", IntPtr),
5459 Args, DAG, getCurDebugLoc());
5460 setValue(&I, Result.first); // Pointers always fit in registers
5461 DAG.setRoot(Result.second);
5464 void SelectionDAGLowering::visitFree(FreeInst &I) {
5465 TargetLowering::ArgListTy Args;
5466 TargetLowering::ArgListEntry Entry;
5467 Entry.Node = getValue(I.getOperand(0));
5468 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5469 Args.push_back(Entry);
5470 MVT IntPtr = TLI.getPointerTy();
5471 std::pair<SDValue,SDValue> Result =
5472 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, false, false,
5473 CallingConv::C, PerformTailCallOpt,
5474 DAG.getExternalSymbol("free", IntPtr), Args, DAG,
5476 DAG.setRoot(Result.second);
5479 void SelectionDAGLowering::visitVAStart(CallInst &I) {
5480 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5481 MVT::Other, getRoot(),
5482 getValue(I.getOperand(1)),
5483 DAG.getSrcValue(I.getOperand(1))));
5486 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
5487 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5488 getRoot(), getValue(I.getOperand(0)),
5489 DAG.getSrcValue(I.getOperand(0)));
5491 DAG.setRoot(V.getValue(1));
5494 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
5495 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5496 MVT::Other, getRoot(),
5497 getValue(I.getOperand(1)),
5498 DAG.getSrcValue(I.getOperand(1))));
5501 void SelectionDAGLowering::visitVACopy(CallInst &I) {
5502 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5503 MVT::Other, getRoot(),
5504 getValue(I.getOperand(1)),
5505 getValue(I.getOperand(2)),
5506 DAG.getSrcValue(I.getOperand(1)),
5507 DAG.getSrcValue(I.getOperand(2))));
5510 /// TargetLowering::LowerArguments - This is the default LowerArguments
5511 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
5512 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
5513 /// integrated into SDISel.
5514 void TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG,
5515 SmallVectorImpl<SDValue> &ArgValues,
5517 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
5518 SmallVector<SDValue, 3+16> Ops;
5519 Ops.push_back(DAG.getRoot());
5520 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
5521 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
5523 // Add one result value for each formal argument.
5524 SmallVector<MVT, 16> RetVals;
5526 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
5528 SmallVector<MVT, 4> ValueVTs;
5529 ComputeValueVTs(*this, I->getType(), ValueVTs);
5530 for (unsigned Value = 0, NumValues = ValueVTs.size();
5531 Value != NumValues; ++Value) {
5532 MVT VT = ValueVTs[Value];
5533 const Type *ArgTy = VT.getTypeForMVT();
5534 ISD::ArgFlagsTy Flags;
5535 unsigned OriginalAlignment =
5536 getTargetData()->getABITypeAlignment(ArgTy);
5538 if (F.paramHasAttr(j, Attribute::ZExt))
5540 if (F.paramHasAttr(j, Attribute::SExt))
5542 if (F.paramHasAttr(j, Attribute::InReg))
5544 if (F.paramHasAttr(j, Attribute::StructRet))
5546 if (F.paramHasAttr(j, Attribute::ByVal)) {
5548 const PointerType *Ty = cast<PointerType>(I->getType());
5549 const Type *ElementTy = Ty->getElementType();
5550 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5551 unsigned FrameSize = getTargetData()->getTypePaddedSize(ElementTy);
5552 // For ByVal, alignment should be passed from FE. BE will guess if
5553 // this info is not there but there are cases it cannot get right.
5554 if (F.getParamAlignment(j))
5555 FrameAlign = F.getParamAlignment(j);
5556 Flags.setByValAlign(FrameAlign);
5557 Flags.setByValSize(FrameSize);
5559 if (F.paramHasAttr(j, Attribute::Nest))
5561 Flags.setOrigAlign(OriginalAlignment);
5563 MVT RegisterVT = getRegisterType(VT);
5564 unsigned NumRegs = getNumRegisters(VT);
5565 for (unsigned i = 0; i != NumRegs; ++i) {
5566 RetVals.push_back(RegisterVT);
5567 ISD::ArgFlagsTy MyFlags = Flags;
5568 if (NumRegs > 1 && i == 0)
5570 // if it isn't first piece, alignment must be 1
5572 MyFlags.setOrigAlign(1);
5573 Ops.push_back(DAG.getArgFlags(MyFlags));
5578 RetVals.push_back(MVT::Other);
5581 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS, dl,
5582 DAG.getVTList(&RetVals[0], RetVals.size()),
5583 &Ops[0], Ops.size()).getNode();
5585 // Prelower FORMAL_ARGUMENTS. This isn't required for functionality, but
5586 // allows exposing the loads that may be part of the argument access to the
5587 // first DAGCombiner pass.
5588 SDValue TmpRes = LowerOperation(SDValue(Result, 0), DAG);
5590 // The number of results should match up, except that the lowered one may have
5591 // an extra flag result.
5592 assert((Result->getNumValues() == TmpRes.getNode()->getNumValues() ||
5593 (Result->getNumValues()+1 == TmpRes.getNode()->getNumValues() &&
5594 TmpRes.getValue(Result->getNumValues()).getValueType() == MVT::Flag))
5595 && "Lowering produced unexpected number of results!");
5597 // The FORMAL_ARGUMENTS node itself is likely no longer needed.
5598 if (Result != TmpRes.getNode() && Result->use_empty()) {
5599 HandleSDNode Dummy(DAG.getRoot());
5600 DAG.RemoveDeadNode(Result);
5603 Result = TmpRes.getNode();
5605 unsigned NumArgRegs = Result->getNumValues() - 1;
5606 DAG.setRoot(SDValue(Result, NumArgRegs));
5608 // Set up the return result vector.
5611 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
5613 SmallVector<MVT, 4> ValueVTs;
5614 ComputeValueVTs(*this, I->getType(), ValueVTs);
5615 for (unsigned Value = 0, NumValues = ValueVTs.size();
5616 Value != NumValues; ++Value) {
5617 MVT VT = ValueVTs[Value];
5618 MVT PartVT = getRegisterType(VT);
5620 unsigned NumParts = getNumRegisters(VT);
5621 SmallVector<SDValue, 4> Parts(NumParts);
5622 for (unsigned j = 0; j != NumParts; ++j)
5623 Parts[j] = SDValue(Result, i++);
5625 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5626 if (F.paramHasAttr(Idx, Attribute::SExt))
5627 AssertOp = ISD::AssertSext;
5628 else if (F.paramHasAttr(Idx, Attribute::ZExt))
5629 AssertOp = ISD::AssertZext;
5631 ArgValues.push_back(getCopyFromParts(DAG, dl, &Parts[0], NumParts,
5632 PartVT, VT, AssertOp));
5635 assert(i == NumArgRegs && "Argument register count mismatch!");
5639 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5640 /// implementation, which just inserts an ISD::CALL node, which is later custom
5641 /// lowered by the target to something concrete. FIXME: When all targets are
5642 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
5643 std::pair<SDValue, SDValue>
5644 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5645 bool RetSExt, bool RetZExt, bool isVarArg,
5647 unsigned CallingConv, bool isTailCall,
5649 ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl) {
5650 assert((!isTailCall || PerformTailCallOpt) &&
5651 "isTailCall set when tail-call optimizations are disabled!");
5653 SmallVector<SDValue, 32> Ops;
5654 Ops.push_back(Chain); // Op#0 - Chain
5655 Ops.push_back(Callee);
5657 // Handle all of the outgoing arguments.
5658 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5659 SmallVector<MVT, 4> ValueVTs;
5660 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5661 for (unsigned Value = 0, NumValues = ValueVTs.size();
5662 Value != NumValues; ++Value) {
5663 MVT VT = ValueVTs[Value];
5664 const Type *ArgTy = VT.getTypeForMVT();
5665 SDValue Op = SDValue(Args[i].Node.getNode(),
5666 Args[i].Node.getResNo() + Value);
5667 ISD::ArgFlagsTy Flags;
5668 unsigned OriginalAlignment =
5669 getTargetData()->getABITypeAlignment(ArgTy);
5675 if (Args[i].isInReg)
5679 if (Args[i].isByVal) {
5681 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5682 const Type *ElementTy = Ty->getElementType();
5683 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5684 unsigned FrameSize = getTargetData()->getTypePaddedSize(ElementTy);
5685 // For ByVal, alignment should come from FE. BE will guess if this
5686 // info is not there but there are cases it cannot get right.
5687 if (Args[i].Alignment)
5688 FrameAlign = Args[i].Alignment;
5689 Flags.setByValAlign(FrameAlign);
5690 Flags.setByValSize(FrameSize);
5694 Flags.setOrigAlign(OriginalAlignment);
5696 MVT PartVT = getRegisterType(VT);
5697 unsigned NumParts = getNumRegisters(VT);
5698 SmallVector<SDValue, 4> Parts(NumParts);
5699 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5702 ExtendKind = ISD::SIGN_EXTEND;
5703 else if (Args[i].isZExt)
5704 ExtendKind = ISD::ZERO_EXTEND;
5706 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts, PartVT, ExtendKind);
5708 for (unsigned i = 0; i != NumParts; ++i) {
5709 // if it isn't first piece, alignment must be 1
5710 ISD::ArgFlagsTy MyFlags = Flags;
5711 if (NumParts > 1 && i == 0)
5714 MyFlags.setOrigAlign(1);
5716 Ops.push_back(Parts[i]);
5717 Ops.push_back(DAG.getArgFlags(MyFlags));
5722 // Figure out the result value types. We start by making a list of
5723 // the potentially illegal return value types.
5724 SmallVector<MVT, 4> LoweredRetTys;
5725 SmallVector<MVT, 4> RetTys;
5726 ComputeValueVTs(*this, RetTy, RetTys);
5728 // Then we translate that to a list of legal types.
5729 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5731 MVT RegisterVT = getRegisterType(VT);
5732 unsigned NumRegs = getNumRegisters(VT);
5733 for (unsigned i = 0; i != NumRegs; ++i)
5734 LoweredRetTys.push_back(RegisterVT);
5737 LoweredRetTys.push_back(MVT::Other); // Always has a chain.
5739 // Create the CALL node.
5740 SDValue Res = DAG.getCall(CallingConv, dl,
5741 isVarArg, isTailCall, isInreg,
5742 DAG.getVTList(&LoweredRetTys[0],
5743 LoweredRetTys.size()),
5746 Chain = Res.getValue(LoweredRetTys.size() - 1);
5748 // Gather up the call result into a single value.
5749 if (RetTy != Type::VoidTy && !RetTys.empty()) {
5750 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5753 AssertOp = ISD::AssertSext;
5755 AssertOp = ISD::AssertZext;
5757 SmallVector<SDValue, 4> ReturnValues;
5759 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5761 MVT RegisterVT = getRegisterType(VT);
5762 unsigned NumRegs = getNumRegisters(VT);
5763 unsigned RegNoEnd = NumRegs + RegNo;
5764 SmallVector<SDValue, 4> Results;
5765 for (; RegNo != RegNoEnd; ++RegNo)
5766 Results.push_back(Res.getValue(RegNo));
5767 SDValue ReturnValue =
5768 getCopyFromParts(DAG, dl, &Results[0], NumRegs, RegisterVT, VT,
5770 ReturnValues.push_back(ReturnValue);
5772 Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5773 DAG.getVTList(&RetTys[0], RetTys.size()),
5774 &ReturnValues[0], ReturnValues.size());
5777 return std::make_pair(Res, Chain);
5780 void TargetLowering::LowerOperationWrapper(SDNode *N,
5781 SmallVectorImpl<SDValue> &Results,
5782 SelectionDAG &DAG) {
5783 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5785 Results.push_back(Res);
5788 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5789 assert(0 && "LowerOperation not implemented for this target!");
5795 void SelectionDAGLowering::CopyValueToVirtualRegister(Value *V, unsigned Reg) {
5796 SDValue Op = getValue(V);
5797 assert((Op.getOpcode() != ISD::CopyFromReg ||
5798 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5799 "Copy from a reg to the same reg!");
5800 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5802 RegsForValue RFV(TLI, Reg, V->getType());
5803 SDValue Chain = DAG.getEntryNode();
5804 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5805 PendingExports.push_back(Chain);
5808 #include "llvm/CodeGen/SelectionDAGISel.h"
5810 void SelectionDAGISel::
5811 LowerArguments(BasicBlock *LLVMBB) {
5812 // If this is the entry block, emit arguments.
5813 Function &F = *LLVMBB->getParent();
5814 SDValue OldRoot = SDL->DAG.getRoot();
5815 SmallVector<SDValue, 16> Args;
5816 TLI.LowerArguments(F, SDL->DAG, Args, SDL->getCurDebugLoc());
5819 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
5821 SmallVector<MVT, 4> ValueVTs;
5822 ComputeValueVTs(TLI, AI->getType(), ValueVTs);
5823 unsigned NumValues = ValueVTs.size();
5824 if (!AI->use_empty()) {
5825 SDL->setValue(AI, SDL->DAG.getMergeValues(&Args[a], NumValues,
5826 SDL->getCurDebugLoc()));
5827 // If this argument is live outside of the entry block, insert a copy from
5828 // whereever we got it to the vreg that other BB's will reference it as.
5829 DenseMap<const Value*, unsigned>::iterator VMI=FuncInfo->ValueMap.find(AI);
5830 if (VMI != FuncInfo->ValueMap.end()) {
5831 SDL->CopyValueToVirtualRegister(AI, VMI->second);
5837 // Finally, if the target has anything special to do, allow it to do so.
5838 // FIXME: this should insert code into the DAG!
5839 EmitFunctionEntryCode(F, SDL->DAG.getMachineFunction());
5842 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
5843 /// ensure constants are generated when needed. Remember the virtual registers
5844 /// that need to be added to the Machine PHI nodes as input. We cannot just
5845 /// directly add them, because expansion might result in multiple MBB's for one
5846 /// BB. As such, the start of the BB might correspond to a different MBB than
5850 SelectionDAGISel::HandlePHINodesInSuccessorBlocks(BasicBlock *LLVMBB) {
5851 TerminatorInst *TI = LLVMBB->getTerminator();
5853 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5855 // Check successor nodes' PHI nodes that expect a constant to be available
5857 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5858 BasicBlock *SuccBB = TI->getSuccessor(succ);
5859 if (!isa<PHINode>(SuccBB->begin())) continue;
5860 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5862 // If this terminator has multiple identical successors (common for
5863 // switches), only handle each succ once.
5864 if (!SuccsHandled.insert(SuccMBB)) continue;
5866 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5869 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5870 // nodes and Machine PHI nodes, but the incoming operands have not been
5872 for (BasicBlock::iterator I = SuccBB->begin();
5873 (PN = dyn_cast<PHINode>(I)); ++I) {
5874 // Ignore dead phi's.
5875 if (PN->use_empty()) continue;
5878 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5880 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
5881 unsigned &RegOut = SDL->ConstantsOut[C];
5883 RegOut = FuncInfo->CreateRegForValue(C);
5884 SDL->CopyValueToVirtualRegister(C, RegOut);
5888 Reg = FuncInfo->ValueMap[PHIOp];
5890 assert(isa<AllocaInst>(PHIOp) &&
5891 FuncInfo->StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
5892 "Didn't codegen value into a register!??");
5893 Reg = FuncInfo->CreateRegForValue(PHIOp);
5894 SDL->CopyValueToVirtualRegister(PHIOp, Reg);
5898 // Remember that this register needs to added to the machine PHI node as
5899 // the input for this MBB.
5900 SmallVector<MVT, 4> ValueVTs;
5901 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
5902 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
5903 MVT VT = ValueVTs[vti];
5904 unsigned NumRegisters = TLI.getNumRegisters(VT);
5905 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
5906 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
5907 Reg += NumRegisters;
5911 SDL->ConstantsOut.clear();
5914 /// This is the Fast-ISel version of HandlePHINodesInSuccessorBlocks. It only
5915 /// supports legal types, and it emits MachineInstrs directly instead of
5916 /// creating SelectionDAG nodes.
5919 SelectionDAGISel::HandlePHINodesInSuccessorBlocksFast(BasicBlock *LLVMBB,
5921 TerminatorInst *TI = LLVMBB->getTerminator();
5923 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5924 unsigned OrigNumPHINodesToUpdate = SDL->PHINodesToUpdate.size();
5926 // Check successor nodes' PHI nodes that expect a constant to be available
5928 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5929 BasicBlock *SuccBB = TI->getSuccessor(succ);
5930 if (!isa<PHINode>(SuccBB->begin())) continue;
5931 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5933 // If this terminator has multiple identical successors (common for
5934 // switches), only handle each succ once.
5935 if (!SuccsHandled.insert(SuccMBB)) continue;
5937 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5940 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5941 // nodes and Machine PHI nodes, but the incoming operands have not been
5943 for (BasicBlock::iterator I = SuccBB->begin();
5944 (PN = dyn_cast<PHINode>(I)); ++I) {
5945 // Ignore dead phi's.
5946 if (PN->use_empty()) continue;
5948 // Only handle legal types. Two interesting things to note here. First,
5949 // by bailing out early, we may leave behind some dead instructions,
5950 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
5951 // own moves. Second, this check is necessary becuase FastISel doesn't
5952 // use CreateRegForValue to create registers, so it always creates
5953 // exactly one register for each non-void instruction.
5954 MVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
5955 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
5958 VT = TLI.getTypeToTransformTo(VT);
5960 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5965 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5967 unsigned Reg = F->getRegForValue(PHIOp);
5969 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5972 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));