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/TargetOptions.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.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()->getTypeAllocSize(EltTy);
133 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
134 ComputeValueVTs(TLI, EltTy, ValueVTs, Offsets,
135 StartingOffset + i * EltSize);
138 // Interpret void as zero return values.
139 if (Ty == Type::VoidTy)
141 // Base case: we can get an MVT for this LLVM IR type.
142 ValueVTs.push_back(TLI.getValueType(Ty));
144 Offsets->push_back(StartingOffset);
148 /// RegsForValue - This struct represents the registers (physical or virtual)
149 /// that a particular set of values is assigned, and the type information about
150 /// the value. The most common situation is to represent one value at a time,
151 /// but struct or array values are handled element-wise as multiple values.
152 /// The splitting of aggregates is performed recursively, so that we never
153 /// have aggregate-typed registers. The values at this point do not necessarily
154 /// have legal types, so each value may require one or more registers of some
157 struct VISIBILITY_HIDDEN RegsForValue {
158 /// TLI - The TargetLowering object.
160 const TargetLowering *TLI;
162 /// ValueVTs - The value types of the values, which may not be legal, and
163 /// may need be promoted or synthesized from one or more registers.
165 SmallVector<MVT, 4> ValueVTs;
167 /// RegVTs - The value types of the registers. This is the same size as
168 /// ValueVTs and it records, for each value, what the type of the assigned
169 /// register or registers are. (Individual values are never synthesized
170 /// from more than one type of register.)
172 /// With virtual registers, the contents of RegVTs is redundant with TLI's
173 /// getRegisterType member function, however when with physical registers
174 /// it is necessary to have a separate record of the types.
176 SmallVector<MVT, 4> RegVTs;
178 /// Regs - This list holds the registers assigned to the values.
179 /// Each legal or promoted value requires one register, and each
180 /// expanded value requires multiple registers.
182 SmallVector<unsigned, 4> Regs;
184 RegsForValue() : TLI(0) {}
186 RegsForValue(const TargetLowering &tli,
187 const SmallVector<unsigned, 4> ®s,
188 MVT regvt, MVT valuevt)
189 : TLI(&tli), ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
190 RegsForValue(const TargetLowering &tli,
191 const SmallVector<unsigned, 4> ®s,
192 const SmallVector<MVT, 4> ®vts,
193 const SmallVector<MVT, 4> &valuevts)
194 : TLI(&tli), ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {}
195 RegsForValue(const TargetLowering &tli,
196 unsigned Reg, const Type *Ty) : TLI(&tli) {
197 ComputeValueVTs(tli, Ty, ValueVTs);
199 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
200 MVT ValueVT = ValueVTs[Value];
201 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
202 MVT RegisterVT = TLI->getRegisterType(ValueVT);
203 for (unsigned i = 0; i != NumRegs; ++i)
204 Regs.push_back(Reg + i);
205 RegVTs.push_back(RegisterVT);
210 /// append - Add the specified values to this one.
211 void append(const RegsForValue &RHS) {
213 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
214 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
215 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
219 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
220 /// this value and returns the result as a ValueVTs value. This uses
221 /// Chain/Flag as the input and updates them for the output Chain/Flag.
222 /// If the Flag pointer is NULL, no flag is used.
223 SDValue getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl,
224 SDValue &Chain, SDValue *Flag) const;
226 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
227 /// specified value into the registers specified by this object. This uses
228 /// Chain/Flag as the input and updates them for the output Chain/Flag.
229 /// If the Flag pointer is NULL, no flag is used.
230 void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
231 SDValue &Chain, SDValue *Flag) const;
233 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
234 /// operand list. This adds the code marker, matching input operand index
235 /// (if applicable), and includes the number of values added into it.
236 void AddInlineAsmOperands(unsigned Code,
237 bool HasMatching, unsigned MatchingIdx,
238 SelectionDAG &DAG, std::vector<SDValue> &Ops) const;
242 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
243 /// PHI nodes or outside of the basic block that defines it, or used by a
244 /// switch or atomic instruction, which may expand to multiple basic blocks.
245 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
246 if (isa<PHINode>(I)) return true;
247 BasicBlock *BB = I->getParent();
248 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
249 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI))
254 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
255 /// entry block, return true. This includes arguments used by switches, since
256 /// the switch may expand into multiple basic blocks.
257 static bool isOnlyUsedInEntryBlock(Argument *A, bool EnableFastISel) {
258 // With FastISel active, we may be splitting blocks, so force creation
259 // of virtual registers for all non-dead arguments.
260 // Don't force virtual registers for byval arguments though, because
261 // fast-isel can't handle those in all cases.
262 if (EnableFastISel && !A->hasByValAttr())
263 return A->use_empty();
265 BasicBlock *Entry = A->getParent()->begin();
266 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
267 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
268 return false; // Use not in entry block.
272 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli)
276 void FunctionLoweringInfo::set(Function &fn, MachineFunction &mf,
278 bool EnableFastISel) {
281 RegInfo = &MF->getRegInfo();
283 // Create a vreg for each argument register that is not dead and is used
284 // outside of the entry block for the function.
285 for (Function::arg_iterator AI = Fn->arg_begin(), E = Fn->arg_end();
287 if (!isOnlyUsedInEntryBlock(AI, EnableFastISel))
288 InitializeRegForValue(AI);
290 // Initialize the mapping of values to registers. This is only set up for
291 // instruction values that are used outside of the block that defines
293 Function::iterator BB = Fn->begin(), EB = Fn->end();
294 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
295 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
296 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
297 const Type *Ty = AI->getAllocatedType();
298 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
300 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
303 TySize *= CUI->getZExtValue(); // Get total allocated size.
304 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
305 StaticAllocaMap[AI] =
306 MF->getFrameInfo()->CreateStackObject(TySize, Align);
309 for (; BB != EB; ++BB)
310 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
311 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
312 if (!isa<AllocaInst>(I) ||
313 !StaticAllocaMap.count(cast<AllocaInst>(I)))
314 InitializeRegForValue(I);
316 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
317 // also creates the initial PHI MachineInstrs, though none of the input
318 // operands are populated.
319 for (BB = Fn->begin(), EB = Fn->end(); BB != EB; ++BB) {
320 MachineBasicBlock *MBB = mf.CreateMachineBasicBlock(BB);
324 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
328 for (BasicBlock::iterator
329 I = BB->begin(), E = BB->end(); I != E; ++I) {
330 if (CallInst *CI = dyn_cast<CallInst>(I)) {
331 if (Function *F = CI->getCalledFunction()) {
332 switch (F->getIntrinsicID()) {
334 case Intrinsic::dbg_stoppoint: {
335 DbgStopPointInst *SPI = cast<DbgStopPointInst>(I);
336 if (isValidDebugInfoIntrinsic(*SPI, CodeGenOpt::Default))
337 DL = ExtractDebugLocation(*SPI, MF->getDebugLocInfo());
340 case Intrinsic::dbg_func_start: {
341 DbgFuncStartInst *FSI = cast<DbgFuncStartInst>(I);
342 if (isValidDebugInfoIntrinsic(*FSI, CodeGenOpt::Default))
343 DL = ExtractDebugLocation(*FSI, MF->getDebugLocInfo());
350 PN = dyn_cast<PHINode>(I);
351 if (!PN || PN->use_empty()) continue;
353 unsigned PHIReg = ValueMap[PN];
354 assert(PHIReg && "PHI node does not have an assigned virtual register!");
356 SmallVector<MVT, 4> ValueVTs;
357 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
358 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
359 MVT VT = ValueVTs[vti];
360 unsigned NumRegisters = TLI.getNumRegisters(VT);
361 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
362 for (unsigned i = 0; i != NumRegisters; ++i)
363 BuildMI(MBB, DL, TII->get(TargetInstrInfo::PHI), PHIReg + i);
364 PHIReg += NumRegisters;
370 unsigned FunctionLoweringInfo::MakeReg(MVT VT) {
371 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
374 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
375 /// the correctly promoted or expanded types. Assign these registers
376 /// consecutive vreg numbers and return the first assigned number.
378 /// In the case that the given value has struct or array type, this function
379 /// will assign registers for each member or element.
381 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
382 SmallVector<MVT, 4> ValueVTs;
383 ComputeValueVTs(TLI, V->getType(), ValueVTs);
385 unsigned FirstReg = 0;
386 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
387 MVT ValueVT = ValueVTs[Value];
388 MVT RegisterVT = TLI.getRegisterType(ValueVT);
390 unsigned NumRegs = TLI.getNumRegisters(ValueVT);
391 for (unsigned i = 0; i != NumRegs; ++i) {
392 unsigned R = MakeReg(RegisterVT);
393 if (!FirstReg) FirstReg = R;
399 /// getCopyFromParts - Create a value that contains the specified legal parts
400 /// combined into the value they represent. If the parts combine to a type
401 /// larger then ValueVT then AssertOp can be used to specify whether the extra
402 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
403 /// (ISD::AssertSext).
404 static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl,
405 const SDValue *Parts,
406 unsigned NumParts, MVT PartVT, MVT ValueVT,
407 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
408 assert(NumParts > 0 && "No parts to assemble!");
409 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
410 SDValue Val = Parts[0];
413 // Assemble the value from multiple parts.
414 if (!ValueVT.isVector() && ValueVT.isInteger()) {
415 unsigned PartBits = PartVT.getSizeInBits();
416 unsigned ValueBits = ValueVT.getSizeInBits();
418 // Assemble the power of 2 part.
419 unsigned RoundParts = NumParts & (NumParts - 1) ?
420 1 << Log2_32(NumParts) : NumParts;
421 unsigned RoundBits = PartBits * RoundParts;
422 MVT RoundVT = RoundBits == ValueBits ?
423 ValueVT : MVT::getIntegerVT(RoundBits);
426 MVT HalfVT = MVT::getIntegerVT(RoundBits/2);
428 if (RoundParts > 2) {
429 Lo = getCopyFromParts(DAG, dl, Parts, RoundParts/2, PartVT, HalfVT);
430 Hi = getCopyFromParts(DAG, dl, Parts+RoundParts/2, RoundParts/2,
433 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]);
434 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]);
436 if (TLI.isBigEndian())
438 Val = DAG.getNode(ISD::BUILD_PAIR, dl, RoundVT, Lo, Hi);
440 if (RoundParts < NumParts) {
441 // Assemble the trailing non-power-of-2 part.
442 unsigned OddParts = NumParts - RoundParts;
443 MVT OddVT = MVT::getIntegerVT(OddParts * PartBits);
444 Hi = getCopyFromParts(DAG, dl,
445 Parts+RoundParts, OddParts, PartVT, OddVT);
447 // Combine the round and odd parts.
449 if (TLI.isBigEndian())
451 MVT TotalVT = MVT::getIntegerVT(NumParts * PartBits);
452 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, TotalVT, Hi);
453 Hi = DAG.getNode(ISD::SHL, dl, TotalVT, Hi,
454 DAG.getConstant(Lo.getValueType().getSizeInBits(),
455 TLI.getPointerTy()));
456 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, TotalVT, Lo);
457 Val = DAG.getNode(ISD::OR, dl, TotalVT, Lo, Hi);
459 } else if (ValueVT.isVector()) {
460 // Handle a multi-element vector.
461 MVT IntermediateVT, RegisterVT;
462 unsigned NumIntermediates;
464 TLI.getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
466 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
467 NumParts = NumRegs; // Silence a compiler warning.
468 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
469 assert(RegisterVT == Parts[0].getValueType() &&
470 "Part type doesn't match part!");
472 // Assemble the parts into intermediate operands.
473 SmallVector<SDValue, 8> Ops(NumIntermediates);
474 if (NumIntermediates == NumParts) {
475 // If the register was not expanded, truncate or copy the value,
477 for (unsigned i = 0; i != NumParts; ++i)
478 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1,
479 PartVT, IntermediateVT);
480 } else if (NumParts > 0) {
481 // If the intermediate type was expanded, build the intermediate operands
483 assert(NumParts % NumIntermediates == 0 &&
484 "Must expand into a divisible number of parts!");
485 unsigned Factor = NumParts / NumIntermediates;
486 for (unsigned i = 0; i != NumIntermediates; ++i)
487 Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor,
488 PartVT, IntermediateVT);
491 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the intermediate
493 Val = DAG.getNode(IntermediateVT.isVector() ?
494 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, dl,
495 ValueVT, &Ops[0], NumIntermediates);
496 } else if (PartVT.isFloatingPoint()) {
497 // FP split into multiple FP parts (for ppcf128)
498 assert(ValueVT == MVT(MVT::ppcf128) && PartVT == MVT(MVT::f64) &&
501 Lo = DAG.getNode(ISD::BIT_CONVERT, dl, MVT(MVT::f64), Parts[0]);
502 Hi = DAG.getNode(ISD::BIT_CONVERT, dl, MVT(MVT::f64), Parts[1]);
503 if (TLI.isBigEndian())
505 Val = DAG.getNode(ISD::BUILD_PAIR, dl, ValueVT, Lo, Hi);
507 // FP split into integer parts (soft fp)
508 assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
509 !PartVT.isVector() && "Unexpected split");
510 MVT IntVT = MVT::getIntegerVT(ValueVT.getSizeInBits());
511 Val = getCopyFromParts(DAG, dl, Parts, NumParts, PartVT, IntVT);
515 // There is now one part, held in Val. Correct it to match ValueVT.
516 PartVT = Val.getValueType();
518 if (PartVT == ValueVT)
521 if (PartVT.isVector()) {
522 assert(ValueVT.isVector() && "Unknown vector conversion!");
523 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
526 if (ValueVT.isVector()) {
527 assert(ValueVT.getVectorElementType() == PartVT &&
528 ValueVT.getVectorNumElements() == 1 &&
529 "Only trivial scalar-to-vector conversions should get here!");
530 return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val);
533 if (PartVT.isInteger() &&
534 ValueVT.isInteger()) {
535 if (ValueVT.bitsLT(PartVT)) {
536 // For a truncate, see if we have any information to
537 // indicate whether the truncated bits will always be
538 // zero or sign-extension.
539 if (AssertOp != ISD::DELETED_NODE)
540 Val = DAG.getNode(AssertOp, dl, PartVT, Val,
541 DAG.getValueType(ValueVT));
542 return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
544 return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val);
548 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
549 if (ValueVT.bitsLT(Val.getValueType()))
550 // FP_ROUND's are always exact here.
551 return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val,
552 DAG.getIntPtrConstant(1));
553 return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val);
556 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
557 return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
559 LLVM_UNREACHABLE("Unknown mismatch!");
563 /// getCopyToParts - Create a series of nodes that contain the specified value
564 /// split into legal parts. If the parts contain more bits than Val, then, for
565 /// integers, ExtendKind can be used to specify how to generate the extra bits.
566 static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, SDValue Val,
567 SDValue *Parts, unsigned NumParts, MVT PartVT,
568 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
569 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
570 MVT PtrVT = TLI.getPointerTy();
571 MVT ValueVT = Val.getValueType();
572 unsigned PartBits = PartVT.getSizeInBits();
573 unsigned OrigNumParts = NumParts;
574 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
579 if (!ValueVT.isVector()) {
580 if (PartVT == ValueVT) {
581 assert(NumParts == 1 && "No-op copy with multiple parts!");
586 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
587 // If the parts cover more bits than the value has, promote the value.
588 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
589 assert(NumParts == 1 && "Do not know what to promote to!");
590 Val = DAG.getNode(ISD::FP_EXTEND, dl, PartVT, Val);
591 } else if (PartVT.isInteger() && ValueVT.isInteger()) {
592 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
593 Val = DAG.getNode(ExtendKind, dl, ValueVT, Val);
595 LLVM_UNREACHABLE("Unknown mismatch!");
597 } else if (PartBits == ValueVT.getSizeInBits()) {
598 // Different types of the same size.
599 assert(NumParts == 1 && PartVT != ValueVT);
600 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
601 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
602 // If the parts cover less bits than value has, truncate the value.
603 if (PartVT.isInteger() && ValueVT.isInteger()) {
604 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
605 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
607 LLVM_UNREACHABLE("Unknown mismatch!");
611 // The value may have changed - recompute ValueVT.
612 ValueVT = Val.getValueType();
613 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
614 "Failed to tile the value with PartVT!");
617 assert(PartVT == ValueVT && "Type conversion failed!");
622 // Expand the value into multiple parts.
623 if (NumParts & (NumParts - 1)) {
624 // The number of parts is not a power of 2. Split off and copy the tail.
625 assert(PartVT.isInteger() && ValueVT.isInteger() &&
626 "Do not know what to expand to!");
627 unsigned RoundParts = 1 << Log2_32(NumParts);
628 unsigned RoundBits = RoundParts * PartBits;
629 unsigned OddParts = NumParts - RoundParts;
630 SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val,
631 DAG.getConstant(RoundBits,
632 TLI.getPointerTy()));
633 getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, OddParts, PartVT);
634 if (TLI.isBigEndian())
635 // The odd parts were reversed by getCopyToParts - unreverse them.
636 std::reverse(Parts + RoundParts, Parts + NumParts);
637 NumParts = RoundParts;
638 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
639 Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
642 // The number of parts is a power of 2. Repeatedly bisect the value using
644 Parts[0] = DAG.getNode(ISD::BIT_CONVERT, dl,
645 MVT::getIntegerVT(ValueVT.getSizeInBits()),
647 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
648 for (unsigned i = 0; i < NumParts; i += StepSize) {
649 unsigned ThisBits = StepSize * PartBits / 2;
650 MVT ThisVT = MVT::getIntegerVT (ThisBits);
651 SDValue &Part0 = Parts[i];
652 SDValue &Part1 = Parts[i+StepSize/2];
654 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
656 DAG.getConstant(1, PtrVT));
657 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
659 DAG.getConstant(0, PtrVT));
661 if (ThisBits == PartBits && ThisVT != PartVT) {
662 Part0 = DAG.getNode(ISD::BIT_CONVERT, dl,
664 Part1 = DAG.getNode(ISD::BIT_CONVERT, dl,
670 if (TLI.isBigEndian())
671 std::reverse(Parts, Parts + OrigNumParts);
678 if (PartVT != ValueVT) {
679 if (PartVT.isVector()) {
680 Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
682 assert(ValueVT.getVectorElementType() == PartVT &&
683 ValueVT.getVectorNumElements() == 1 &&
684 "Only trivial vector-to-scalar conversions should get here!");
685 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
687 DAG.getConstant(0, PtrVT));
695 // Handle a multi-element vector.
696 MVT IntermediateVT, RegisterVT;
697 unsigned NumIntermediates;
698 unsigned NumRegs = TLI
699 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
701 unsigned NumElements = ValueVT.getVectorNumElements();
703 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
704 NumParts = NumRegs; // Silence a compiler warning.
705 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
707 // Split the vector into intermediate operands.
708 SmallVector<SDValue, 8> Ops(NumIntermediates);
709 for (unsigned i = 0; i != NumIntermediates; ++i)
710 if (IntermediateVT.isVector())
711 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,
713 DAG.getConstant(i * (NumElements / NumIntermediates),
716 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
718 DAG.getConstant(i, PtrVT));
720 // Split the intermediate operands into legal parts.
721 if (NumParts == NumIntermediates) {
722 // If the register was not expanded, promote or copy the value,
724 for (unsigned i = 0; i != NumParts; ++i)
725 getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT);
726 } else if (NumParts > 0) {
727 // If the intermediate type was expanded, split each the value into
729 assert(NumParts % NumIntermediates == 0 &&
730 "Must expand into a divisible number of parts!");
731 unsigned Factor = NumParts / NumIntermediates;
732 for (unsigned i = 0; i != NumIntermediates; ++i)
733 getCopyToParts(DAG, dl, Ops[i], &Parts[i * Factor], Factor, PartVT);
738 void SelectionDAGLowering::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
741 TD = DAG.getTarget().getTargetData();
744 /// clear - Clear out the curret SelectionDAG and the associated
745 /// state and prepare this SelectionDAGLowering object to be used
746 /// for a new block. This doesn't clear out information about
747 /// additional blocks that are needed to complete switch lowering
748 /// or PHI node updating; that information is cleared out as it is
750 void SelectionDAGLowering::clear() {
752 PendingLoads.clear();
753 PendingExports.clear();
755 CurDebugLoc = DebugLoc::getUnknownLoc();
758 /// getRoot - Return the current virtual root of the Selection DAG,
759 /// flushing any PendingLoad items. This must be done before emitting
760 /// a store or any other node that may need to be ordered after any
761 /// prior load instructions.
763 SDValue SelectionDAGLowering::getRoot() {
764 if (PendingLoads.empty())
765 return DAG.getRoot();
767 if (PendingLoads.size() == 1) {
768 SDValue Root = PendingLoads[0];
770 PendingLoads.clear();
774 // Otherwise, we have to make a token factor node.
775 SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
776 &PendingLoads[0], PendingLoads.size());
777 PendingLoads.clear();
782 /// getControlRoot - Similar to getRoot, but instead of flushing all the
783 /// PendingLoad items, flush all the PendingExports items. It is necessary
784 /// to do this before emitting a terminator instruction.
786 SDValue SelectionDAGLowering::getControlRoot() {
787 SDValue Root = DAG.getRoot();
789 if (PendingExports.empty())
792 // Turn all of the CopyToReg chains into one factored node.
793 if (Root.getOpcode() != ISD::EntryToken) {
794 unsigned i = 0, e = PendingExports.size();
795 for (; i != e; ++i) {
796 assert(PendingExports[i].getNode()->getNumOperands() > 1);
797 if (PendingExports[i].getNode()->getOperand(0) == Root)
798 break; // Don't add the root if we already indirectly depend on it.
802 PendingExports.push_back(Root);
805 Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
807 PendingExports.size());
808 PendingExports.clear();
813 void SelectionDAGLowering::visit(Instruction &I) {
814 visit(I.getOpcode(), I);
817 void SelectionDAGLowering::visit(unsigned Opcode, User &I) {
818 // Note: this doesn't use InstVisitor, because it has to work with
819 // ConstantExpr's in addition to instructions.
821 default: LLVM_UNREACHABLE("Unknown instruction type encountered!");
822 // Build the switch statement using the Instruction.def file.
823 #define HANDLE_INST(NUM, OPCODE, CLASS) \
824 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
825 #include "llvm/Instruction.def"
829 SDValue SelectionDAGLowering::getValue(const Value *V) {
830 SDValue &N = NodeMap[V];
831 if (N.getNode()) return N;
833 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
834 MVT VT = TLI.getValueType(V->getType(), true);
836 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
837 return N = DAG.getConstant(*CI, VT);
839 if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
840 return N = DAG.getGlobalAddress(GV, VT);
842 if (isa<ConstantPointerNull>(C))
843 return N = DAG.getConstant(0, TLI.getPointerTy());
845 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C))
846 return N = DAG.getConstantFP(*CFP, VT);
848 if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
849 return N = DAG.getUNDEF(VT);
851 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
852 visit(CE->getOpcode(), *CE);
853 SDValue N1 = NodeMap[V];
854 assert(N1.getNode() && "visit didn't populate the ValueMap!");
858 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
859 SmallVector<SDValue, 4> Constants;
860 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
862 SDNode *Val = getValue(*OI).getNode();
863 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
864 Constants.push_back(SDValue(Val, i));
866 return DAG.getMergeValues(&Constants[0], Constants.size(),
870 if (isa<StructType>(C->getType()) || isa<ArrayType>(C->getType())) {
871 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
872 "Unknown struct or array constant!");
874 SmallVector<MVT, 4> ValueVTs;
875 ComputeValueVTs(TLI, C->getType(), ValueVTs);
876 unsigned NumElts = ValueVTs.size();
878 return SDValue(); // empty struct
879 SmallVector<SDValue, 4> Constants(NumElts);
880 for (unsigned i = 0; i != NumElts; ++i) {
881 MVT EltVT = ValueVTs[i];
882 if (isa<UndefValue>(C))
883 Constants[i] = DAG.getUNDEF(EltVT);
884 else if (EltVT.isFloatingPoint())
885 Constants[i] = DAG.getConstantFP(0, EltVT);
887 Constants[i] = DAG.getConstant(0, EltVT);
889 return DAG.getMergeValues(&Constants[0], NumElts, getCurDebugLoc());
892 const VectorType *VecTy = cast<VectorType>(V->getType());
893 unsigned NumElements = VecTy->getNumElements();
895 // Now that we know the number and type of the elements, get that number of
896 // elements into the Ops array based on what kind of constant it is.
897 SmallVector<SDValue, 16> Ops;
898 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
899 for (unsigned i = 0; i != NumElements; ++i)
900 Ops.push_back(getValue(CP->getOperand(i)));
902 assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
903 MVT EltVT = TLI.getValueType(VecTy->getElementType());
906 if (EltVT.isFloatingPoint())
907 Op = DAG.getConstantFP(0, EltVT);
909 Op = DAG.getConstant(0, EltVT);
910 Ops.assign(NumElements, Op);
913 // Create a BUILD_VECTOR node.
914 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
915 VT, &Ops[0], Ops.size());
918 // If this is a static alloca, generate it as the frameindex instead of
920 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
921 DenseMap<const AllocaInst*, int>::iterator SI =
922 FuncInfo.StaticAllocaMap.find(AI);
923 if (SI != FuncInfo.StaticAllocaMap.end())
924 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
927 unsigned InReg = FuncInfo.ValueMap[V];
928 assert(InReg && "Value not in map!");
930 RegsForValue RFV(TLI, InReg, V->getType());
931 SDValue Chain = DAG.getEntryNode();
932 return RFV.getCopyFromRegs(DAG, getCurDebugLoc(), Chain, NULL);
936 void SelectionDAGLowering::visitRet(ReturnInst &I) {
937 if (I.getNumOperands() == 0) {
938 DAG.setRoot(DAG.getNode(ISD::RET, getCurDebugLoc(),
939 MVT::Other, getControlRoot()));
943 SmallVector<SDValue, 8> NewValues;
944 NewValues.push_back(getControlRoot());
945 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
946 SmallVector<MVT, 4> ValueVTs;
947 ComputeValueVTs(TLI, I.getOperand(i)->getType(), ValueVTs);
948 unsigned NumValues = ValueVTs.size();
949 if (NumValues == 0) continue;
951 SDValue RetOp = getValue(I.getOperand(i));
952 for (unsigned j = 0, f = NumValues; j != f; ++j) {
953 MVT VT = ValueVTs[j];
955 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
957 const Function *F = I.getParent()->getParent();
958 if (F->paramHasAttr(0, Attribute::SExt))
959 ExtendKind = ISD::SIGN_EXTEND;
960 else if (F->paramHasAttr(0, Attribute::ZExt))
961 ExtendKind = ISD::ZERO_EXTEND;
963 // FIXME: C calling convention requires the return type to be promoted to
964 // at least 32-bit. But this is not necessary for non-C calling
965 // conventions. The frontend should mark functions whose return values
966 // require promoting with signext or zeroext attributes.
967 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) {
968 MVT MinVT = TLI.getRegisterType(MVT::i32);
969 if (VT.bitsLT(MinVT))
973 unsigned NumParts = TLI.getNumRegisters(VT);
974 MVT PartVT = TLI.getRegisterType(VT);
975 SmallVector<SDValue, 4> Parts(NumParts);
976 getCopyToParts(DAG, getCurDebugLoc(),
977 SDValue(RetOp.getNode(), RetOp.getResNo() + j),
978 &Parts[0], NumParts, PartVT, ExtendKind);
980 // 'inreg' on function refers to return value
981 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
982 if (F->paramHasAttr(0, Attribute::InReg))
984 for (unsigned i = 0; i < NumParts; ++i) {
985 NewValues.push_back(Parts[i]);
986 NewValues.push_back(DAG.getArgFlags(Flags));
990 DAG.setRoot(DAG.getNode(ISD::RET, getCurDebugLoc(), MVT::Other,
991 &NewValues[0], NewValues.size()));
994 /// CopyToExportRegsIfNeeded - If the given value has virtual registers
995 /// created for it, emit nodes to copy the value into the virtual
997 void SelectionDAGLowering::CopyToExportRegsIfNeeded(Value *V) {
998 if (!V->use_empty()) {
999 DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
1000 if (VMI != FuncInfo.ValueMap.end())
1001 CopyValueToVirtualRegister(V, VMI->second);
1005 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
1006 /// the current basic block, add it to ValueMap now so that we'll get a
1008 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
1009 // No need to export constants.
1010 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
1012 // Already exported?
1013 if (FuncInfo.isExportedInst(V)) return;
1015 unsigned Reg = FuncInfo.InitializeRegForValue(V);
1016 CopyValueToVirtualRegister(V, Reg);
1019 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
1020 const BasicBlock *FromBB) {
1021 // The operands of the setcc have to be in this block. We don't know
1022 // how to export them from some other block.
1023 if (Instruction *VI = dyn_cast<Instruction>(V)) {
1024 // Can export from current BB.
1025 if (VI->getParent() == FromBB)
1028 // Is already exported, noop.
1029 return FuncInfo.isExportedInst(V);
1032 // If this is an argument, we can export it if the BB is the entry block or
1033 // if it is already exported.
1034 if (isa<Argument>(V)) {
1035 if (FromBB == &FromBB->getParent()->getEntryBlock())
1038 // Otherwise, can only export this if it is already exported.
1039 return FuncInfo.isExportedInst(V);
1042 // Otherwise, constants can always be exported.
1046 static bool InBlock(const Value *V, const BasicBlock *BB) {
1047 if (const Instruction *I = dyn_cast<Instruction>(V))
1048 return I->getParent() == BB;
1052 /// getFCmpCondCode - Return the ISD condition code corresponding to
1053 /// the given LLVM IR floating-point condition code. This includes
1054 /// consideration of global floating-point math flags.
1056 static ISD::CondCode getFCmpCondCode(FCmpInst::Predicate Pred) {
1057 ISD::CondCode FPC, FOC;
1059 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1060 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1061 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1062 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1063 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1064 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1065 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1066 case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
1067 case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
1068 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1069 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1070 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1071 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1072 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1073 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1074 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1076 LLVM_UNREACHABLE("Invalid FCmp predicate opcode!");
1077 FOC = FPC = ISD::SETFALSE;
1080 if (FiniteOnlyFPMath())
1086 /// getICmpCondCode - Return the ISD condition code corresponding to
1087 /// the given LLVM IR integer condition code.
1089 static ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred) {
1091 case ICmpInst::ICMP_EQ: return ISD::SETEQ;
1092 case ICmpInst::ICMP_NE: return ISD::SETNE;
1093 case ICmpInst::ICMP_SLE: return ISD::SETLE;
1094 case ICmpInst::ICMP_ULE: return ISD::SETULE;
1095 case ICmpInst::ICMP_SGE: return ISD::SETGE;
1096 case ICmpInst::ICMP_UGE: return ISD::SETUGE;
1097 case ICmpInst::ICMP_SLT: return ISD::SETLT;
1098 case ICmpInst::ICMP_ULT: return ISD::SETULT;
1099 case ICmpInst::ICMP_SGT: return ISD::SETGT;
1100 case ICmpInst::ICMP_UGT: return ISD::SETUGT;
1102 LLVM_UNREACHABLE("Invalid ICmp predicate opcode!");
1107 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1108 /// This function emits a branch and is used at the leaves of an OR or an
1109 /// AND operator tree.
1112 SelectionDAGLowering::EmitBranchForMergedCondition(Value *Cond,
1113 MachineBasicBlock *TBB,
1114 MachineBasicBlock *FBB,
1115 MachineBasicBlock *CurBB) {
1116 const BasicBlock *BB = CurBB->getBasicBlock();
1118 // If the leaf of the tree is a comparison, merge the condition into
1120 if (CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1121 // The operands of the cmp have to be in this block. We don't know
1122 // how to export them from some other block. If this is the first block
1123 // of the sequence, no exporting is needed.
1124 if (CurBB == CurMBB ||
1125 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1126 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1127 ISD::CondCode Condition;
1128 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1129 Condition = getICmpCondCode(IC->getPredicate());
1130 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1131 Condition = getFCmpCondCode(FC->getPredicate());
1133 Condition = ISD::SETEQ; // silence warning.
1134 LLVM_UNREACHABLE("Unknown compare instruction");
1137 CaseBlock CB(Condition, BOp->getOperand(0),
1138 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1139 SwitchCases.push_back(CB);
1144 // Create a CaseBlock record representing this branch.
1145 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(),
1146 NULL, TBB, FBB, CurBB);
1147 SwitchCases.push_back(CB);
1150 /// FindMergedConditions - If Cond is an expression like
1151 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
1152 MachineBasicBlock *TBB,
1153 MachineBasicBlock *FBB,
1154 MachineBasicBlock *CurBB,
1156 // If this node is not part of the or/and tree, emit it as a branch.
1157 Instruction *BOp = dyn_cast<Instruction>(Cond);
1158 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1159 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1160 BOp->getParent() != CurBB->getBasicBlock() ||
1161 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1162 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1163 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB);
1167 // Create TmpBB after CurBB.
1168 MachineFunction::iterator BBI = CurBB;
1169 MachineFunction &MF = DAG.getMachineFunction();
1170 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1171 CurBB->getParent()->insert(++BBI, TmpBB);
1173 if (Opc == Instruction::Or) {
1174 // Codegen X | Y as:
1182 // Emit the LHS condition.
1183 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1185 // Emit the RHS condition into TmpBB.
1186 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1188 assert(Opc == Instruction::And && "Unknown merge op!");
1189 // Codegen X & Y as:
1196 // This requires creation of TmpBB after CurBB.
1198 // Emit the LHS condition.
1199 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1201 // Emit the RHS condition into TmpBB.
1202 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1206 /// If the set of cases should be emitted as a series of branches, return true.
1207 /// If we should emit this as a bunch of and/or'd together conditions, return
1210 SelectionDAGLowering::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1211 if (Cases.size() != 2) return true;
1213 // If this is two comparisons of the same values or'd or and'd together, they
1214 // will get folded into a single comparison, so don't emit two blocks.
1215 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1216 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1217 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1218 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1225 void SelectionDAGLowering::visitBr(BranchInst &I) {
1226 // Update machine-CFG edges.
1227 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1229 // Figure out which block is immediately after the current one.
1230 MachineBasicBlock *NextBlock = 0;
1231 MachineFunction::iterator BBI = CurMBB;
1232 if (++BBI != CurMBB->getParent()->end())
1235 if (I.isUnconditional()) {
1236 // Update machine-CFG edges.
1237 CurMBB->addSuccessor(Succ0MBB);
1239 // If this is not a fall-through branch, emit the branch.
1240 if (Succ0MBB != NextBlock)
1241 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1242 MVT::Other, getControlRoot(),
1243 DAG.getBasicBlock(Succ0MBB)));
1247 // If this condition is one of the special cases we handle, do special stuff
1249 Value *CondVal = I.getCondition();
1250 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1252 // If this is a series of conditions that are or'd or and'd together, emit
1253 // this as a sequence of branches instead of setcc's with and/or operations.
1254 // For example, instead of something like:
1267 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1268 if (BOp->hasOneUse() &&
1269 (BOp->getOpcode() == Instruction::And ||
1270 BOp->getOpcode() == Instruction::Or)) {
1271 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1272 // If the compares in later blocks need to use values not currently
1273 // exported from this block, export them now. This block should always
1274 // be the first entry.
1275 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1277 // Allow some cases to be rejected.
1278 if (ShouldEmitAsBranches(SwitchCases)) {
1279 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1280 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1281 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1284 // Emit the branch for this block.
1285 visitSwitchCase(SwitchCases[0]);
1286 SwitchCases.erase(SwitchCases.begin());
1290 // Okay, we decided not to do this, remove any inserted MBB's and clear
1292 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1293 CurMBB->getParent()->erase(SwitchCases[i].ThisBB);
1295 SwitchCases.clear();
1299 // Create a CaseBlock record representing this branch.
1300 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(),
1301 NULL, Succ0MBB, Succ1MBB, CurMBB);
1302 // Use visitSwitchCase to actually insert the fast branch sequence for this
1304 visitSwitchCase(CB);
1307 /// visitSwitchCase - Emits the necessary code to represent a single node in
1308 /// the binary search tree resulting from lowering a switch instruction.
1309 void SelectionDAGLowering::visitSwitchCase(CaseBlock &CB) {
1311 SDValue CondLHS = getValue(CB.CmpLHS);
1312 DebugLoc dl = getCurDebugLoc();
1314 // Build the setcc now.
1315 if (CB.CmpMHS == NULL) {
1316 // Fold "(X == true)" to X and "(X == false)" to !X to
1317 // handle common cases produced by branch lowering.
1318 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ)
1320 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) {
1321 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1322 Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
1324 Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1326 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1328 const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
1329 const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
1331 SDValue CmpOp = getValue(CB.CmpMHS);
1332 MVT VT = CmpOp.getValueType();
1334 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1335 Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
1338 SDValue SUB = DAG.getNode(ISD::SUB, dl,
1339 VT, CmpOp, DAG.getConstant(Low, VT));
1340 Cond = DAG.getSetCC(dl, MVT::i1, SUB,
1341 DAG.getConstant(High-Low, VT), ISD::SETULE);
1345 // Update successor info
1346 CurMBB->addSuccessor(CB.TrueBB);
1347 CurMBB->addSuccessor(CB.FalseBB);
1349 // Set NextBlock to be the MBB immediately after the current one, if any.
1350 // This is used to avoid emitting unnecessary branches to the next block.
1351 MachineBasicBlock *NextBlock = 0;
1352 MachineFunction::iterator BBI = CurMBB;
1353 if (++BBI != CurMBB->getParent()->end())
1356 // If the lhs block is the next block, invert the condition so that we can
1357 // fall through to the lhs instead of the rhs block.
1358 if (CB.TrueBB == NextBlock) {
1359 std::swap(CB.TrueBB, CB.FalseBB);
1360 SDValue True = DAG.getConstant(1, Cond.getValueType());
1361 Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
1363 SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
1364 MVT::Other, getControlRoot(), Cond,
1365 DAG.getBasicBlock(CB.TrueBB));
1367 // If the branch was constant folded, fix up the CFG.
1368 if (BrCond.getOpcode() == ISD::BR) {
1369 CurMBB->removeSuccessor(CB.FalseBB);
1370 DAG.setRoot(BrCond);
1372 // Otherwise, go ahead and insert the false branch.
1373 if (BrCond == getControlRoot())
1374 CurMBB->removeSuccessor(CB.TrueBB);
1376 if (CB.FalseBB == NextBlock)
1377 DAG.setRoot(BrCond);
1379 DAG.setRoot(DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
1380 DAG.getBasicBlock(CB.FalseBB)));
1384 /// visitJumpTable - Emit JumpTable node in the current MBB
1385 void SelectionDAGLowering::visitJumpTable(JumpTable &JT) {
1386 // Emit the code for the jump table
1387 assert(JT.Reg != -1U && "Should lower JT Header first!");
1388 MVT PTy = TLI.getPointerTy();
1389 SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
1391 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1392 DAG.setRoot(DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
1393 MVT::Other, Index.getValue(1),
1397 /// visitJumpTableHeader - This function emits necessary code to produce index
1398 /// in the JumpTable from switch case.
1399 void SelectionDAGLowering::visitJumpTableHeader(JumpTable &JT,
1400 JumpTableHeader &JTH) {
1401 // Subtract the lowest switch case value from the value being switched on and
1402 // conditional branch to default mbb if the result is greater than the
1403 // difference between smallest and largest cases.
1404 SDValue SwitchOp = getValue(JTH.SValue);
1405 MVT VT = SwitchOp.getValueType();
1406 SDValue SUB = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1407 DAG.getConstant(JTH.First, VT));
1409 // The SDNode we just created, which holds the value being switched on minus
1410 // the the smallest case value, needs to be copied to a virtual register so it
1411 // can be used as an index into the jump table in a subsequent basic block.
1412 // This value may be smaller or larger than the target's pointer type, and
1413 // therefore require extension or truncating.
1414 if (VT.bitsGT(TLI.getPointerTy()))
1415 SwitchOp = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
1416 TLI.getPointerTy(), SUB);
1418 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
1419 TLI.getPointerTy(), SUB);
1421 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1422 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1423 JumpTableReg, SwitchOp);
1424 JT.Reg = JumpTableReg;
1426 // Emit the range check for the jump table, and branch to the default block
1427 // for the switch statement if the value being switched on exceeds the largest
1428 // case in the switch.
1429 SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
1430 TLI.getSetCCResultType(SUB.getValueType()), SUB,
1431 DAG.getConstant(JTH.Last-JTH.First,VT),
1434 // Set NextBlock to be the MBB immediately after the current one, if any.
1435 // This is used to avoid emitting unnecessary branches to the next block.
1436 MachineBasicBlock *NextBlock = 0;
1437 MachineFunction::iterator BBI = CurMBB;
1438 if (++BBI != CurMBB->getParent()->end())
1441 SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1442 MVT::Other, CopyTo, CMP,
1443 DAG.getBasicBlock(JT.Default));
1445 if (JT.MBB == NextBlock)
1446 DAG.setRoot(BrCond);
1448 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
1449 DAG.getBasicBlock(JT.MBB)));
1452 /// visitBitTestHeader - This function emits necessary code to produce value
1453 /// suitable for "bit tests"
1454 void SelectionDAGLowering::visitBitTestHeader(BitTestBlock &B) {
1455 // Subtract the minimum value
1456 SDValue SwitchOp = getValue(B.SValue);
1457 MVT VT = SwitchOp.getValueType();
1458 SDValue SUB = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
1459 DAG.getConstant(B.First, VT));
1462 SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
1463 TLI.getSetCCResultType(SUB.getValueType()),
1464 SUB, DAG.getConstant(B.Range, VT),
1468 if (VT.bitsGT(TLI.getPointerTy()))
1469 ShiftOp = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
1470 TLI.getPointerTy(), SUB);
1472 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
1473 TLI.getPointerTy(), SUB);
1475 B.Reg = FuncInfo.MakeReg(TLI.getPointerTy());
1476 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
1479 // Set NextBlock to be the MBB immediately after the current one, if any.
1480 // This is used to avoid emitting unnecessary branches to the next block.
1481 MachineBasicBlock *NextBlock = 0;
1482 MachineFunction::iterator BBI = CurMBB;
1483 if (++BBI != CurMBB->getParent()->end())
1486 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1488 CurMBB->addSuccessor(B.Default);
1489 CurMBB->addSuccessor(MBB);
1491 SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1492 MVT::Other, CopyTo, RangeCmp,
1493 DAG.getBasicBlock(B.Default));
1495 if (MBB == NextBlock)
1496 DAG.setRoot(BrRange);
1498 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
1499 DAG.getBasicBlock(MBB)));
1502 /// visitBitTestCase - this function produces one "bit test"
1503 void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB,
1506 // Make desired shift
1507 SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
1508 TLI.getPointerTy());
1509 SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
1511 DAG.getConstant(1, TLI.getPointerTy()),
1514 // Emit bit tests and jumps
1515 SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
1516 TLI.getPointerTy(), SwitchVal,
1517 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1518 SDValue AndCmp = DAG.getSetCC(getCurDebugLoc(),
1519 TLI.getSetCCResultType(AndOp.getValueType()),
1520 AndOp, DAG.getConstant(0, TLI.getPointerTy()),
1523 CurMBB->addSuccessor(B.TargetBB);
1524 CurMBB->addSuccessor(NextMBB);
1526 SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
1527 MVT::Other, getControlRoot(),
1528 AndCmp, DAG.getBasicBlock(B.TargetBB));
1530 // Set NextBlock to be the MBB immediately after the current one, if any.
1531 // This is used to avoid emitting unnecessary branches to the next block.
1532 MachineBasicBlock *NextBlock = 0;
1533 MachineFunction::iterator BBI = CurMBB;
1534 if (++BBI != CurMBB->getParent()->end())
1537 if (NextMBB == NextBlock)
1540 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
1541 DAG.getBasicBlock(NextMBB)));
1544 void SelectionDAGLowering::visitInvoke(InvokeInst &I) {
1545 // Retrieve successors.
1546 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1547 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1549 const Value *Callee(I.getCalledValue());
1550 if (isa<InlineAsm>(Callee))
1553 LowerCallTo(&I, getValue(Callee), false, LandingPad);
1555 // If the value of the invoke is used outside of its defining block, make it
1556 // available as a virtual register.
1557 CopyToExportRegsIfNeeded(&I);
1559 // Update successor info
1560 CurMBB->addSuccessor(Return);
1561 CurMBB->addSuccessor(LandingPad);
1563 // Drop into normal successor.
1564 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
1565 MVT::Other, getControlRoot(),
1566 DAG.getBasicBlock(Return)));
1569 void SelectionDAGLowering::visitUnwind(UnwindInst &I) {
1572 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1573 /// small case ranges).
1574 bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR,
1575 CaseRecVector& WorkList,
1577 MachineBasicBlock* Default) {
1578 Case& BackCase = *(CR.Range.second-1);
1580 // Size is the number of Cases represented by this range.
1581 size_t Size = CR.Range.second - CR.Range.first;
1585 // Get the MachineFunction which holds the current MBB. This is used when
1586 // inserting any additional MBBs necessary to represent the switch.
1587 MachineFunction *CurMF = CurMBB->getParent();
1589 // Figure out which block is immediately after the current one.
1590 MachineBasicBlock *NextBlock = 0;
1591 MachineFunction::iterator BBI = CR.CaseBB;
1593 if (++BBI != CurMBB->getParent()->end())
1596 // TODO: If any two of the cases has the same destination, and if one value
1597 // is the same as the other, but has one bit unset that the other has set,
1598 // use bit manipulation to do two compares at once. For example:
1599 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1601 // Rearrange the case blocks so that the last one falls through if possible.
1602 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1603 // The last case block won't fall through into 'NextBlock' if we emit the
1604 // branches in this order. See if rearranging a case value would help.
1605 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1606 if (I->BB == NextBlock) {
1607 std::swap(*I, BackCase);
1613 // Create a CaseBlock record representing a conditional branch to
1614 // the Case's target mbb if the value being switched on SV is equal
1616 MachineBasicBlock *CurBlock = CR.CaseBB;
1617 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1618 MachineBasicBlock *FallThrough;
1620 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1621 CurMF->insert(BBI, FallThrough);
1623 // Put SV in a virtual register to make it available from the new blocks.
1624 ExportFromCurrentBlock(SV);
1626 // If the last case doesn't match, go to the default block.
1627 FallThrough = Default;
1630 Value *RHS, *LHS, *MHS;
1632 if (I->High == I->Low) {
1633 // This is just small small case range :) containing exactly 1 case
1635 LHS = SV; RHS = I->High; MHS = NULL;
1638 LHS = I->Low; MHS = SV; RHS = I->High;
1640 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1642 // If emitting the first comparison, just call visitSwitchCase to emit the
1643 // code into the current block. Otherwise, push the CaseBlock onto the
1644 // vector to be later processed by SDISel, and insert the node's MBB
1645 // before the next MBB.
1646 if (CurBlock == CurMBB)
1647 visitSwitchCase(CB);
1649 SwitchCases.push_back(CB);
1651 CurBlock = FallThrough;
1657 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1658 return !DisableJumpTables &&
1659 (TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
1660 TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
1663 static APInt ComputeRange(const APInt &First, const APInt &Last) {
1664 APInt LastExt(Last), FirstExt(First);
1665 uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
1666 LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
1667 return (LastExt - FirstExt + 1ULL);
1670 /// handleJTSwitchCase - Emit jumptable for current switch case range
1671 bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR,
1672 CaseRecVector& WorkList,
1674 MachineBasicBlock* Default) {
1675 Case& FrontCase = *CR.Range.first;
1676 Case& BackCase = *(CR.Range.second-1);
1678 const APInt& First = cast<ConstantInt>(FrontCase.Low)->getValue();
1679 const APInt& Last = cast<ConstantInt>(BackCase.High)->getValue();
1682 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1686 if (!areJTsAllowed(TLI) || TSize <= 3)
1689 APInt Range = ComputeRange(First, Last);
1690 double Density = (double)TSize / Range.roundToDouble();
1694 DEBUG(errs() << "Lowering jump table\n"
1695 << "First entry: " << First << ". Last entry: " << Last << '\n'
1696 << "Range: " << Range
1697 << "Size: " << TSize << ". Density: " << Density << "\n\n");
1699 // Get the MachineFunction which holds the current MBB. This is used when
1700 // inserting any additional MBBs necessary to represent the switch.
1701 MachineFunction *CurMF = CurMBB->getParent();
1703 // Figure out which block is immediately after the current one.
1704 MachineBasicBlock *NextBlock = 0;
1705 MachineFunction::iterator BBI = CR.CaseBB;
1707 if (++BBI != CurMBB->getParent()->end())
1710 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1712 // Create a new basic block to hold the code for loading the address
1713 // of the jump table, and jumping to it. Update successor information;
1714 // we will either branch to the default case for the switch, or the jump
1716 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1717 CurMF->insert(BBI, JumpTableBB);
1718 CR.CaseBB->addSuccessor(Default);
1719 CR.CaseBB->addSuccessor(JumpTableBB);
1721 // Build a vector of destination BBs, corresponding to each target
1722 // of the jump table. If the value of the jump table slot corresponds to
1723 // a case statement, push the case's BB onto the vector, otherwise, push
1725 std::vector<MachineBasicBlock*> DestBBs;
1727 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1728 const APInt& Low = cast<ConstantInt>(I->Low)->getValue();
1729 const APInt& High = cast<ConstantInt>(I->High)->getValue();
1731 if (Low.sle(TEI) && TEI.sle(High)) {
1732 DestBBs.push_back(I->BB);
1736 DestBBs.push_back(Default);
1740 // Update successor info. Add one edge to each unique successor.
1741 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1742 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1743 E = DestBBs.end(); I != E; ++I) {
1744 if (!SuccsHandled[(*I)->getNumber()]) {
1745 SuccsHandled[(*I)->getNumber()] = true;
1746 JumpTableBB->addSuccessor(*I);
1750 // Create a jump table index for this jump table, or return an existing
1752 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1754 // Set the jump table information so that we can codegen it as a second
1755 // MachineBasicBlock
1756 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1757 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == CurMBB));
1758 if (CR.CaseBB == CurMBB)
1759 visitJumpTableHeader(JT, JTH);
1761 JTCases.push_back(JumpTableBlock(JTH, JT));
1766 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1768 bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR,
1769 CaseRecVector& WorkList,
1771 MachineBasicBlock* Default) {
1772 // Get the MachineFunction which holds the current MBB. This is used when
1773 // inserting any additional MBBs necessary to represent the switch.
1774 MachineFunction *CurMF = CurMBB->getParent();
1776 // Figure out which block is immediately after the current one.
1777 MachineBasicBlock *NextBlock = 0;
1778 MachineFunction::iterator BBI = CR.CaseBB;
1780 if (++BBI != CurMBB->getParent()->end())
1783 Case& FrontCase = *CR.Range.first;
1784 Case& BackCase = *(CR.Range.second-1);
1785 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1787 // Size is the number of Cases represented by this range.
1788 unsigned Size = CR.Range.second - CR.Range.first;
1790 const APInt& First = cast<ConstantInt>(FrontCase.Low)->getValue();
1791 const APInt& Last = cast<ConstantInt>(BackCase.High)->getValue();
1793 CaseItr Pivot = CR.Range.first + Size/2;
1795 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1796 // (heuristically) allow us to emit JumpTable's later.
1798 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1802 size_t LSize = FrontCase.size();
1803 size_t RSize = TSize-LSize;
1804 DEBUG(errs() << "Selecting best pivot: \n"
1805 << "First: " << First << ", Last: " << Last <<'\n'
1806 << "LSize: " << LSize << ", RSize: " << RSize << '\n');
1807 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1809 const APInt& LEnd = cast<ConstantInt>(I->High)->getValue();
1810 const APInt& RBegin = cast<ConstantInt>(J->Low)->getValue();
1811 APInt Range = ComputeRange(LEnd, RBegin);
1812 assert((Range - 2ULL).isNonNegative() &&
1813 "Invalid case distance");
1814 double LDensity = (double)LSize / (LEnd - First + 1ULL).roundToDouble();
1815 double RDensity = (double)RSize / (Last - RBegin + 1ULL).roundToDouble();
1816 double Metric = Range.logBase2()*(LDensity+RDensity);
1817 // Should always split in some non-trivial place
1818 DEBUG(errs() <<"=>Step\n"
1819 << "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
1820 << "LDensity: " << LDensity
1821 << ", RDensity: " << RDensity << '\n'
1822 << "Metric: " << Metric << '\n');
1823 if (FMetric < Metric) {
1826 DEBUG(errs() << "Current metric set to: " << FMetric << '\n');
1832 if (areJTsAllowed(TLI)) {
1833 // If our case is dense we *really* should handle it earlier!
1834 assert((FMetric > 0) && "Should handle dense range earlier!");
1836 Pivot = CR.Range.first + Size/2;
1839 CaseRange LHSR(CR.Range.first, Pivot);
1840 CaseRange RHSR(Pivot, CR.Range.second);
1841 Constant *C = Pivot->Low;
1842 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1844 // We know that we branch to the LHS if the Value being switched on is
1845 // less than the Pivot value, C. We use this to optimize our binary
1846 // tree a bit, by recognizing that if SV is greater than or equal to the
1847 // LHS's Case Value, and that Case Value is exactly one less than the
1848 // Pivot's Value, then we can branch directly to the LHS's Target,
1849 // rather than creating a leaf node for it.
1850 if ((LHSR.second - LHSR.first) == 1 &&
1851 LHSR.first->High == CR.GE &&
1852 cast<ConstantInt>(C)->getValue() ==
1853 (cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
1854 TrueBB = LHSR.first->BB;
1856 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1857 CurMF->insert(BBI, TrueBB);
1858 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1860 // Put SV in a virtual register to make it available from the new blocks.
1861 ExportFromCurrentBlock(SV);
1864 // Similar to the optimization above, if the Value being switched on is
1865 // known to be less than the Constant CR.LT, and the current Case Value
1866 // is CR.LT - 1, then we can branch directly to the target block for
1867 // the current Case Value, rather than emitting a RHS leaf node for it.
1868 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1869 cast<ConstantInt>(RHSR.first->Low)->getValue() ==
1870 (cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
1871 FalseBB = RHSR.first->BB;
1873 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1874 CurMF->insert(BBI, FalseBB);
1875 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1877 // Put SV in a virtual register to make it available from the new blocks.
1878 ExportFromCurrentBlock(SV);
1881 // Create a CaseBlock record representing a conditional branch to
1882 // the LHS node if the value being switched on SV is less than C.
1883 // Otherwise, branch to LHS.
1884 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
1886 if (CR.CaseBB == CurMBB)
1887 visitSwitchCase(CB);
1889 SwitchCases.push_back(CB);
1894 /// handleBitTestsSwitchCase - if current case range has few destination and
1895 /// range span less, than machine word bitwidth, encode case range into series
1896 /// of masks and emit bit tests with these masks.
1897 bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR,
1898 CaseRecVector& WorkList,
1900 MachineBasicBlock* Default){
1901 unsigned IntPtrBits = TLI.getPointerTy().getSizeInBits();
1903 Case& FrontCase = *CR.Range.first;
1904 Case& BackCase = *(CR.Range.second-1);
1906 // Get the MachineFunction which holds the current MBB. This is used when
1907 // inserting any additional MBBs necessary to represent the switch.
1908 MachineFunction *CurMF = CurMBB->getParent();
1910 // If target does not have legal shift left, do not emit bit tests at all.
1911 if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
1915 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1917 // Single case counts one, case range - two.
1918 numCmps += (I->Low == I->High ? 1 : 2);
1921 // Count unique destinations
1922 SmallSet<MachineBasicBlock*, 4> Dests;
1923 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1924 Dests.insert(I->BB);
1925 if (Dests.size() > 3)
1926 // Don't bother the code below, if there are too much unique destinations
1929 DEBUG(errs() << "Total number of unique destinations: " << Dests.size() << '\n'
1930 << "Total number of comparisons: " << numCmps << '\n');
1932 // Compute span of values.
1933 const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
1934 const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();
1935 APInt cmpRange = maxValue - minValue;
1937 DEBUG(errs() << "Compare range: " << cmpRange << '\n'
1938 << "Low bound: " << minValue << '\n'
1939 << "High bound: " << maxValue << '\n');
1941 if (cmpRange.uge(APInt(cmpRange.getBitWidth(), IntPtrBits)) ||
1942 (!(Dests.size() == 1 && numCmps >= 3) &&
1943 !(Dests.size() == 2 && numCmps >= 5) &&
1944 !(Dests.size() >= 3 && numCmps >= 6)))
1947 DEBUG(errs() << "Emitting bit tests\n");
1948 APInt lowBound = APInt::getNullValue(cmpRange.getBitWidth());
1950 // Optimize the case where all the case values fit in a
1951 // word without having to subtract minValue. In this case,
1952 // we can optimize away the subtraction.
1953 if (minValue.isNonNegative() &&
1954 maxValue.slt(APInt(maxValue.getBitWidth(), IntPtrBits))) {
1955 cmpRange = maxValue;
1957 lowBound = minValue;
1960 CaseBitsVector CasesBits;
1961 unsigned i, count = 0;
1963 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1964 MachineBasicBlock* Dest = I->BB;
1965 for (i = 0; i < count; ++i)
1966 if (Dest == CasesBits[i].BB)
1970 assert((count < 3) && "Too much destinations to test!");
1971 CasesBits.push_back(CaseBits(0, Dest, 0));
1975 const APInt& lowValue = cast<ConstantInt>(I->Low)->getValue();
1976 const APInt& highValue = cast<ConstantInt>(I->High)->getValue();
1978 uint64_t lo = (lowValue - lowBound).getZExtValue();
1979 uint64_t hi = (highValue - lowBound).getZExtValue();
1981 for (uint64_t j = lo; j <= hi; j++) {
1982 CasesBits[i].Mask |= 1ULL << j;
1983 CasesBits[i].Bits++;
1987 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
1991 // Figure out which block is immediately after the current one.
1992 MachineFunction::iterator BBI = CR.CaseBB;
1995 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1997 DEBUG(errs() << "Cases:\n");
1998 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
1999 DEBUG(errs() << "Mask: " << CasesBits[i].Mask
2000 << ", Bits: " << CasesBits[i].Bits
2001 << ", BB: " << CasesBits[i].BB << '\n');
2003 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
2004 CurMF->insert(BBI, CaseBB);
2005 BTC.push_back(BitTestCase(CasesBits[i].Mask,
2009 // Put SV in a virtual register to make it available from the new blocks.
2010 ExportFromCurrentBlock(SV);
2013 BitTestBlock BTB(lowBound, cmpRange, SV,
2014 -1U, (CR.CaseBB == CurMBB),
2015 CR.CaseBB, Default, BTC);
2017 if (CR.CaseBB == CurMBB)
2018 visitBitTestHeader(BTB);
2020 BitTestCases.push_back(BTB);
2026 /// Clusterify - Transform simple list of Cases into list of CaseRange's
2027 size_t SelectionDAGLowering::Clusterify(CaseVector& Cases,
2028 const SwitchInst& SI) {
2031 // Start with "simple" cases
2032 for (size_t i = 1; i < SI.getNumSuccessors(); ++i) {
2033 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
2034 Cases.push_back(Case(SI.getSuccessorValue(i),
2035 SI.getSuccessorValue(i),
2038 std::sort(Cases.begin(), Cases.end(), CaseCmp());
2040 // Merge case into clusters
2041 if (Cases.size() >= 2)
2042 // Must recompute end() each iteration because it may be
2043 // invalidated by erase if we hold on to it
2044 for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
2045 const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
2046 const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
2047 MachineBasicBlock* nextBB = J->BB;
2048 MachineBasicBlock* currentBB = I->BB;
2050 // If the two neighboring cases go to the same destination, merge them
2051 // into a single case.
2052 if ((nextValue - currentValue == 1) && (currentBB == nextBB)) {
2060 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
2061 if (I->Low != I->High)
2062 // A range counts double, since it requires two compares.
2069 void SelectionDAGLowering::visitSwitch(SwitchInst &SI) {
2070 // Figure out which block is immediately after the current one.
2071 MachineBasicBlock *NextBlock = 0;
2072 MachineFunction::iterator BBI = CurMBB;
2074 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
2076 // If there is only the default destination, branch to it if it is not the
2077 // next basic block. Otherwise, just fall through.
2078 if (SI.getNumOperands() == 2) {
2079 // Update machine-CFG edges.
2081 // If this is not a fall-through branch, emit the branch.
2082 CurMBB->addSuccessor(Default);
2083 if (Default != NextBlock)
2084 DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
2085 MVT::Other, getControlRoot(),
2086 DAG.getBasicBlock(Default)));
2090 // If there are any non-default case statements, create a vector of Cases
2091 // representing each one, and sort the vector so that we can efficiently
2092 // create a binary search tree from them.
2094 size_t numCmps = Clusterify(Cases, SI);
2095 DEBUG(errs() << "Clusterify finished. Total clusters: " << Cases.size()
2096 << ". Total compares: " << numCmps << '\n');
2099 // Get the Value to be switched on and default basic blocks, which will be
2100 // inserted into CaseBlock records, representing basic blocks in the binary
2102 Value *SV = SI.getOperand(0);
2104 // Push the initial CaseRec onto the worklist
2105 CaseRecVector WorkList;
2106 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
2108 while (!WorkList.empty()) {
2109 // Grab a record representing a case range to process off the worklist
2110 CaseRec CR = WorkList.back();
2111 WorkList.pop_back();
2113 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
2116 // If the range has few cases (two or less) emit a series of specific
2118 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2121 // If the switch has more than 5 blocks, and at least 40% dense, and the
2122 // target supports indirect branches, then emit a jump table rather than
2123 // lowering the switch to a binary tree of conditional branches.
2124 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2127 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2128 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2129 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2134 void SelectionDAGLowering::visitFSub(User &I) {
2135 // -0.0 - X --> fneg
2136 const Type *Ty = I.getType();
2137 if (isa<VectorType>(Ty)) {
2138 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2139 const VectorType *DestTy = cast<VectorType>(I.getType());
2140 const Type *ElTy = DestTy->getElementType();
2141 unsigned VL = DestTy->getNumElements();
2142 std::vector<Constant*> NZ(VL, Context->getConstantFPNegativeZero(ElTy));
2143 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2145 SDValue Op2 = getValue(I.getOperand(1));
2146 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2147 Op2.getValueType(), Op2));
2152 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2153 if (CFP->isExactlyValue(
2154 Context->getConstantFPNegativeZero(Ty)->getValueAPF())) {
2155 SDValue Op2 = getValue(I.getOperand(1));
2156 setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
2157 Op2.getValueType(), Op2));
2161 visitBinary(I, ISD::FSUB);
2164 void SelectionDAGLowering::visitBinary(User &I, unsigned OpCode) {
2165 SDValue Op1 = getValue(I.getOperand(0));
2166 SDValue Op2 = getValue(I.getOperand(1));
2168 setValue(&I, DAG.getNode(OpCode, getCurDebugLoc(),
2169 Op1.getValueType(), Op1, Op2));
2172 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
2173 SDValue Op1 = getValue(I.getOperand(0));
2174 SDValue Op2 = getValue(I.getOperand(1));
2175 if (!isa<VectorType>(I.getType()) &&
2176 Op2.getValueType() != TLI.getShiftAmountTy()) {
2177 // If the operand is smaller than the shift count type, promote it.
2178 if (TLI.getShiftAmountTy().bitsGT(Op2.getValueType()))
2179 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2180 TLI.getShiftAmountTy(), Op2);
2181 // If the operand is larger than the shift count type but the shift
2182 // count type has enough bits to represent any shift value, truncate
2183 // it now. This is a common case and it exposes the truncate to
2184 // optimization early.
2185 else if (TLI.getShiftAmountTy().getSizeInBits() >=
2186 Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
2187 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2188 TLI.getShiftAmountTy(), Op2);
2189 // Otherwise we'll need to temporarily settle for some other
2190 // convenient type; type legalization will make adjustments as
2192 else if (TLI.getPointerTy().bitsLT(Op2.getValueType()))
2193 Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2194 TLI.getPointerTy(), Op2);
2195 else if (TLI.getPointerTy().bitsGT(Op2.getValueType()))
2196 Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
2197 TLI.getPointerTy(), Op2);
2200 setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
2201 Op1.getValueType(), Op1, Op2));
2204 void SelectionDAGLowering::visitICmp(User &I) {
2205 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2206 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2207 predicate = IC->getPredicate();
2208 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2209 predicate = ICmpInst::Predicate(IC->getPredicate());
2210 SDValue Op1 = getValue(I.getOperand(0));
2211 SDValue Op2 = getValue(I.getOperand(1));
2212 ISD::CondCode Opcode = getICmpCondCode(predicate);
2214 MVT DestVT = TLI.getValueType(I.getType());
2215 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Opcode));
2218 void SelectionDAGLowering::visitFCmp(User &I) {
2219 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2220 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2221 predicate = FC->getPredicate();
2222 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2223 predicate = FCmpInst::Predicate(FC->getPredicate());
2224 SDValue Op1 = getValue(I.getOperand(0));
2225 SDValue Op2 = getValue(I.getOperand(1));
2226 ISD::CondCode Condition = getFCmpCondCode(predicate);
2227 MVT DestVT = TLI.getValueType(I.getType());
2228 setValue(&I, DAG.getSetCC(getCurDebugLoc(), DestVT, Op1, Op2, Condition));
2231 void SelectionDAGLowering::visitSelect(User &I) {
2232 SmallVector<MVT, 4> ValueVTs;
2233 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2234 unsigned NumValues = ValueVTs.size();
2235 if (NumValues != 0) {
2236 SmallVector<SDValue, 4> Values(NumValues);
2237 SDValue Cond = getValue(I.getOperand(0));
2238 SDValue TrueVal = getValue(I.getOperand(1));
2239 SDValue FalseVal = getValue(I.getOperand(2));
2241 for (unsigned i = 0; i != NumValues; ++i)
2242 Values[i] = DAG.getNode(ISD::SELECT, getCurDebugLoc(),
2243 TrueVal.getValueType(), Cond,
2244 SDValue(TrueVal.getNode(), TrueVal.getResNo() + i),
2245 SDValue(FalseVal.getNode(), FalseVal.getResNo() + i));
2247 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2248 DAG.getVTList(&ValueVTs[0], NumValues),
2249 &Values[0], NumValues));
2254 void SelectionDAGLowering::visitTrunc(User &I) {
2255 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2256 SDValue N = getValue(I.getOperand(0));
2257 MVT DestVT = TLI.getValueType(I.getType());
2258 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2261 void SelectionDAGLowering::visitZExt(User &I) {
2262 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2263 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2264 SDValue N = getValue(I.getOperand(0));
2265 MVT DestVT = TLI.getValueType(I.getType());
2266 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N));
2269 void SelectionDAGLowering::visitSExt(User &I) {
2270 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2271 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2272 SDValue N = getValue(I.getOperand(0));
2273 MVT DestVT = TLI.getValueType(I.getType());
2274 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(), DestVT, N));
2277 void SelectionDAGLowering::visitFPTrunc(User &I) {
2278 // FPTrunc is never a no-op cast, no need to check
2279 SDValue N = getValue(I.getOperand(0));
2280 MVT DestVT = TLI.getValueType(I.getType());
2281 setValue(&I, DAG.getNode(ISD::FP_ROUND, getCurDebugLoc(),
2282 DestVT, N, DAG.getIntPtrConstant(0)));
2285 void SelectionDAGLowering::visitFPExt(User &I){
2286 // FPTrunc is never a no-op cast, no need to check
2287 SDValue N = getValue(I.getOperand(0));
2288 MVT DestVT = TLI.getValueType(I.getType());
2289 setValue(&I, DAG.getNode(ISD::FP_EXTEND, getCurDebugLoc(), DestVT, N));
2292 void SelectionDAGLowering::visitFPToUI(User &I) {
2293 // FPToUI is never a no-op cast, no need to check
2294 SDValue N = getValue(I.getOperand(0));
2295 MVT DestVT = TLI.getValueType(I.getType());
2296 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, getCurDebugLoc(), DestVT, N));
2299 void SelectionDAGLowering::visitFPToSI(User &I) {
2300 // FPToSI 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_TO_SINT, getCurDebugLoc(), DestVT, N));
2306 void SelectionDAGLowering::visitUIToFP(User &I) {
2307 // UIToFP 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::UINT_TO_FP, getCurDebugLoc(), DestVT, N));
2313 void SelectionDAGLowering::visitSIToFP(User &I){
2314 // SIToFP 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::SINT_TO_FP, getCurDebugLoc(), DestVT, N));
2320 void SelectionDAGLowering::visitPtrToInt(User &I) {
2321 // What to do depends on the size of the integer and the size of the pointer.
2322 // We can either truncate, zero extend, or no-op, accordingly.
2323 SDValue N = getValue(I.getOperand(0));
2324 MVT SrcVT = N.getValueType();
2325 MVT DestVT = TLI.getValueType(I.getType());
2327 if (DestVT.bitsLT(SrcVT))
2328 Result = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N);
2330 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2331 Result = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), DestVT, N);
2332 setValue(&I, Result);
2335 void SelectionDAGLowering::visitIntToPtr(User &I) {
2336 // What to do depends on the size of the integer and the size of the pointer.
2337 // We can either truncate, zero extend, or no-op, accordingly.
2338 SDValue N = getValue(I.getOperand(0));
2339 MVT SrcVT = N.getValueType();
2340 MVT DestVT = TLI.getValueType(I.getType());
2341 if (DestVT.bitsLT(SrcVT))
2342 setValue(&I, DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), DestVT, N));
2344 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2345 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2349 void SelectionDAGLowering::visitBitCast(User &I) {
2350 SDValue N = getValue(I.getOperand(0));
2351 MVT DestVT = TLI.getValueType(I.getType());
2353 // BitCast assures us that source and destination are the same size so this
2354 // is either a BIT_CONVERT or a no-op.
2355 if (DestVT != N.getValueType())
2356 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
2357 DestVT, N)); // convert types
2359 setValue(&I, N); // noop cast.
2362 void SelectionDAGLowering::visitInsertElement(User &I) {
2363 SDValue InVec = getValue(I.getOperand(0));
2364 SDValue InVal = getValue(I.getOperand(1));
2365 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2367 getValue(I.getOperand(2)));
2369 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT, getCurDebugLoc(),
2370 TLI.getValueType(I.getType()),
2371 InVec, InVal, InIdx));
2374 void SelectionDAGLowering::visitExtractElement(User &I) {
2375 SDValue InVec = getValue(I.getOperand(0));
2376 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2378 getValue(I.getOperand(1)));
2379 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2380 TLI.getValueType(I.getType()), InVec, InIdx));
2384 // Utility for visitShuffleVector - Returns true if the mask is mask starting
2385 // from SIndx and increasing to the element length (undefs are allowed).
2386 static bool SequentialMask(SmallVectorImpl<int> &Mask, unsigned SIndx) {
2387 unsigned MaskNumElts = Mask.size();
2388 for (unsigned i = 0; i != MaskNumElts; ++i)
2389 if ((Mask[i] >= 0) && (Mask[i] != (int)(i + SIndx)))
2394 void SelectionDAGLowering::visitShuffleVector(User &I) {
2395 SmallVector<int, 8> Mask;
2396 SDValue Src1 = getValue(I.getOperand(0));
2397 SDValue Src2 = getValue(I.getOperand(1));
2399 // Convert the ConstantVector mask operand into an array of ints, with -1
2400 // representing undef values.
2401 SmallVector<Constant*, 8> MaskElts;
2402 cast<Constant>(I.getOperand(2))->getVectorElements(*Context, MaskElts);
2403 unsigned MaskNumElts = MaskElts.size();
2404 for (unsigned i = 0; i != MaskNumElts; ++i) {
2405 if (isa<UndefValue>(MaskElts[i]))
2408 Mask.push_back(cast<ConstantInt>(MaskElts[i])->getSExtValue());
2411 MVT VT = TLI.getValueType(I.getType());
2412 MVT SrcVT = Src1.getValueType();
2413 unsigned SrcNumElts = SrcVT.getVectorNumElements();
2415 if (SrcNumElts == MaskNumElts) {
2416 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2421 // Normalize the shuffle vector since mask and vector length don't match.
2422 if (SrcNumElts < MaskNumElts && MaskNumElts % SrcNumElts == 0) {
2423 // Mask is longer than the source vectors and is a multiple of the source
2424 // vectors. We can use concatenate vector to make the mask and vectors
2426 if (SrcNumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
2427 // The shuffle is concatenating two vectors together.
2428 setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, getCurDebugLoc(),
2433 // Pad both vectors with undefs to make them the same length as the mask.
2434 unsigned NumConcat = MaskNumElts / SrcNumElts;
2435 bool Src1U = Src1.getOpcode() == ISD::UNDEF;
2436 bool Src2U = Src2.getOpcode() == ISD::UNDEF;
2437 SDValue UndefVal = DAG.getUNDEF(SrcVT);
2439 SmallVector<SDValue, 8> MOps1(NumConcat, UndefVal);
2440 SmallVector<SDValue, 8> MOps2(NumConcat, UndefVal);
2444 Src1 = Src1U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2445 getCurDebugLoc(), VT,
2446 &MOps1[0], NumConcat);
2447 Src2 = Src2U ? DAG.getUNDEF(VT) : DAG.getNode(ISD::CONCAT_VECTORS,
2448 getCurDebugLoc(), VT,
2449 &MOps2[0], NumConcat);
2451 // Readjust mask for new input vector length.
2452 SmallVector<int, 8> MappedOps;
2453 for (unsigned i = 0; i != MaskNumElts; ++i) {
2455 if (Idx < (int)SrcNumElts)
2456 MappedOps.push_back(Idx);
2458 MappedOps.push_back(Idx + MaskNumElts - SrcNumElts);
2460 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2465 if (SrcNumElts > MaskNumElts) {
2466 // Analyze the access pattern of the vector to see if we can extract
2467 // two subvectors and do the shuffle. The analysis is done by calculating
2468 // the range of elements the mask access on both vectors.
2469 int MinRange[2] = { SrcNumElts+1, SrcNumElts+1};
2470 int MaxRange[2] = {-1, -1};
2472 for (unsigned i = 0; i != MaskNumElts; ++i) {
2478 if (Idx >= (int)SrcNumElts) {
2482 if (Idx > MaxRange[Input])
2483 MaxRange[Input] = Idx;
2484 if (Idx < MinRange[Input])
2485 MinRange[Input] = Idx;
2488 // Check if the access is smaller than the vector size and can we find
2489 // a reasonable extract index.
2490 int RangeUse[2] = { 2, 2 }; // 0 = Unused, 1 = Extract, 2 = Can not Extract.
2491 int StartIdx[2]; // StartIdx to extract from
2492 for (int Input=0; Input < 2; ++Input) {
2493 if (MinRange[Input] == (int)(SrcNumElts+1) && MaxRange[Input] == -1) {
2494 RangeUse[Input] = 0; // Unused
2495 StartIdx[Input] = 0;
2496 } else if (MaxRange[Input] - MinRange[Input] < (int)MaskNumElts) {
2497 // Fits within range but we should see if we can find a good
2498 // start index that is a multiple of the mask length.
2499 if (MaxRange[Input] < (int)MaskNumElts) {
2500 RangeUse[Input] = 1; // Extract from beginning of the vector
2501 StartIdx[Input] = 0;
2503 StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
2504 if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
2505 StartIdx[Input] + MaskNumElts < SrcNumElts)
2506 RangeUse[Input] = 1; // Extract from a multiple of the mask length.
2511 if (RangeUse[0] == 0 && RangeUse[0] == 0) {
2512 setValue(&I, DAG.getUNDEF(VT)); // Vectors are not used.
2515 else if (RangeUse[0] < 2 && RangeUse[1] < 2) {
2516 // Extract appropriate subvector and generate a vector shuffle
2517 for (int Input=0; Input < 2; ++Input) {
2518 SDValue& Src = Input == 0 ? Src1 : Src2;
2519 if (RangeUse[Input] == 0) {
2520 Src = DAG.getUNDEF(VT);
2522 Src = DAG.getNode(ISD::EXTRACT_SUBVECTOR, getCurDebugLoc(), VT,
2523 Src, DAG.getIntPtrConstant(StartIdx[Input]));
2526 // Calculate new mask.
2527 SmallVector<int, 8> MappedOps;
2528 for (unsigned i = 0; i != MaskNumElts; ++i) {
2531 MappedOps.push_back(Idx);
2532 else if (Idx < (int)SrcNumElts)
2533 MappedOps.push_back(Idx - StartIdx[0]);
2535 MappedOps.push_back(Idx - SrcNumElts - StartIdx[1] + MaskNumElts);
2537 setValue(&I, DAG.getVectorShuffle(VT, getCurDebugLoc(), Src1, Src2,
2543 // We can't use either concat vectors or extract subvectors so fall back to
2544 // replacing the shuffle with extract and build vector.
2545 // to insert and build vector.
2546 MVT EltVT = VT.getVectorElementType();
2547 MVT PtrVT = TLI.getPointerTy();
2548 SmallVector<SDValue,8> Ops;
2549 for (unsigned i = 0; i != MaskNumElts; ++i) {
2551 Ops.push_back(DAG.getUNDEF(EltVT));
2554 if (Idx < (int)SrcNumElts)
2555 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2556 EltVT, Src1, DAG.getConstant(Idx, PtrVT)));
2558 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, getCurDebugLoc(),
2560 DAG.getConstant(Idx - SrcNumElts, PtrVT)));
2563 setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
2564 VT, &Ops[0], Ops.size()));
2567 void SelectionDAGLowering::visitInsertValue(InsertValueInst &I) {
2568 const Value *Op0 = I.getOperand(0);
2569 const Value *Op1 = I.getOperand(1);
2570 const Type *AggTy = I.getType();
2571 const Type *ValTy = Op1->getType();
2572 bool IntoUndef = isa<UndefValue>(Op0);
2573 bool FromUndef = isa<UndefValue>(Op1);
2575 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2576 I.idx_begin(), I.idx_end());
2578 SmallVector<MVT, 4> AggValueVTs;
2579 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2580 SmallVector<MVT, 4> ValValueVTs;
2581 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2583 unsigned NumAggValues = AggValueVTs.size();
2584 unsigned NumValValues = ValValueVTs.size();
2585 SmallVector<SDValue, 4> Values(NumAggValues);
2587 SDValue Agg = getValue(Op0);
2588 SDValue Val = getValue(Op1);
2590 // Copy the beginning value(s) from the original aggregate.
2591 for (; i != LinearIndex; ++i)
2592 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2593 SDValue(Agg.getNode(), Agg.getResNo() + i);
2594 // Copy values from the inserted value(s).
2595 for (; i != LinearIndex + NumValValues; ++i)
2596 Values[i] = FromUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2597 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2598 // Copy remaining value(s) from the original aggregate.
2599 for (; i != NumAggValues; ++i)
2600 Values[i] = IntoUndef ? DAG.getUNDEF(AggValueVTs[i]) :
2601 SDValue(Agg.getNode(), Agg.getResNo() + i);
2603 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2604 DAG.getVTList(&AggValueVTs[0], NumAggValues),
2605 &Values[0], NumAggValues));
2608 void SelectionDAGLowering::visitExtractValue(ExtractValueInst &I) {
2609 const Value *Op0 = I.getOperand(0);
2610 const Type *AggTy = Op0->getType();
2611 const Type *ValTy = I.getType();
2612 bool OutOfUndef = isa<UndefValue>(Op0);
2614 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2615 I.idx_begin(), I.idx_end());
2617 SmallVector<MVT, 4> ValValueVTs;
2618 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2620 unsigned NumValValues = ValValueVTs.size();
2621 SmallVector<SDValue, 4> Values(NumValValues);
2623 SDValue Agg = getValue(Op0);
2624 // Copy out the selected value(s).
2625 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2626 Values[i - LinearIndex] =
2628 DAG.getUNDEF(Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2629 SDValue(Agg.getNode(), Agg.getResNo() + i);
2631 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2632 DAG.getVTList(&ValValueVTs[0], NumValValues),
2633 &Values[0], NumValValues));
2637 void SelectionDAGLowering::visitGetElementPtr(User &I) {
2638 SDValue N = getValue(I.getOperand(0));
2639 const Type *Ty = I.getOperand(0)->getType();
2641 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2644 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2645 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2648 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2649 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2650 DAG.getIntPtrConstant(Offset));
2652 Ty = StTy->getElementType(Field);
2654 Ty = cast<SequentialType>(Ty)->getElementType();
2656 // If this is a constant subscript, handle it quickly.
2657 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2658 if (CI->getZExtValue() == 0) continue;
2660 TD->getTypeAllocSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2662 unsigned PtrBits = TLI.getPointerTy().getSizeInBits();
2664 OffsVal = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2666 DAG.getConstant(Offs, MVT::i64));
2668 OffsVal = DAG.getIntPtrConstant(Offs);
2669 N = DAG.getNode(ISD::ADD, getCurDebugLoc(), N.getValueType(), N,
2674 // N = N + Idx * ElementSize;
2675 uint64_t ElementSize = TD->getTypeAllocSize(Ty);
2676 SDValue IdxN = getValue(Idx);
2678 // If the index is smaller or larger than intptr_t, truncate or extend
2680 if (IdxN.getValueType().bitsLT(N.getValueType()))
2681 IdxN = DAG.getNode(ISD::SIGN_EXTEND, getCurDebugLoc(),
2682 N.getValueType(), IdxN);
2683 else if (IdxN.getValueType().bitsGT(N.getValueType()))
2684 IdxN = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2685 N.getValueType(), IdxN);
2687 // If this is a multiply by a power of two, turn it into a shl
2688 // immediately. This is a very common case.
2689 if (ElementSize != 1) {
2690 if (isPowerOf2_64(ElementSize)) {
2691 unsigned Amt = Log2_64(ElementSize);
2692 IdxN = DAG.getNode(ISD::SHL, getCurDebugLoc(),
2693 N.getValueType(), IdxN,
2694 DAG.getConstant(Amt, TLI.getPointerTy()));
2696 SDValue Scale = DAG.getIntPtrConstant(ElementSize);
2697 IdxN = DAG.getNode(ISD::MUL, getCurDebugLoc(),
2698 N.getValueType(), IdxN, Scale);
2702 N = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2703 N.getValueType(), N, IdxN);
2709 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
2710 // If this is a fixed sized alloca in the entry block of the function,
2711 // allocate it statically on the stack.
2712 if (FuncInfo.StaticAllocaMap.count(&I))
2713 return; // getValue will auto-populate this.
2715 const Type *Ty = I.getAllocatedType();
2716 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
2718 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2721 SDValue AllocSize = getValue(I.getArraySize());
2723 AllocSize = DAG.getNode(ISD::MUL, getCurDebugLoc(), AllocSize.getValueType(),
2725 DAG.getConstant(TySize, AllocSize.getValueType()));
2729 MVT IntPtr = TLI.getPointerTy();
2730 if (IntPtr.bitsLT(AllocSize.getValueType()))
2731 AllocSize = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
2733 else if (IntPtr.bitsGT(AllocSize.getValueType()))
2734 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(),
2737 // Handle alignment. If the requested alignment is less than or equal to
2738 // the stack alignment, ignore it. If the size is greater than or equal to
2739 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2740 unsigned StackAlign =
2741 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2742 if (Align <= StackAlign)
2745 // Round the size of the allocation up to the stack alignment size
2746 // by add SA-1 to the size.
2747 AllocSize = DAG.getNode(ISD::ADD, getCurDebugLoc(),
2748 AllocSize.getValueType(), AllocSize,
2749 DAG.getIntPtrConstant(StackAlign-1));
2750 // Mask out the low bits for alignment purposes.
2751 AllocSize = DAG.getNode(ISD::AND, getCurDebugLoc(),
2752 AllocSize.getValueType(), AllocSize,
2753 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2755 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2756 SDVTList VTs = DAG.getVTList(AllocSize.getValueType(), MVT::Other);
2757 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, getCurDebugLoc(),
2760 DAG.setRoot(DSA.getValue(1));
2762 // Inform the Frame Information that we have just allocated a variable-sized
2764 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
2767 void SelectionDAGLowering::visitLoad(LoadInst &I) {
2768 const Value *SV = I.getOperand(0);
2769 SDValue Ptr = getValue(SV);
2771 const Type *Ty = I.getType();
2772 bool isVolatile = I.isVolatile();
2773 unsigned Alignment = I.getAlignment();
2775 SmallVector<MVT, 4> ValueVTs;
2776 SmallVector<uint64_t, 4> Offsets;
2777 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2778 unsigned NumValues = ValueVTs.size();
2783 bool ConstantMemory = false;
2785 // Serialize volatile loads with other side effects.
2787 else if (AA->pointsToConstantMemory(SV)) {
2788 // Do not serialize (non-volatile) loads of constant memory with anything.
2789 Root = DAG.getEntryNode();
2790 ConstantMemory = true;
2792 // Do not serialize non-volatile loads against each other.
2793 Root = DAG.getRoot();
2796 SmallVector<SDValue, 4> Values(NumValues);
2797 SmallVector<SDValue, 4> Chains(NumValues);
2798 MVT PtrVT = Ptr.getValueType();
2799 for (unsigned i = 0; i != NumValues; ++i) {
2800 SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
2801 DAG.getNode(ISD::ADD, getCurDebugLoc(),
2803 DAG.getConstant(Offsets[i], PtrVT)),
2805 isVolatile, Alignment);
2807 Chains[i] = L.getValue(1);
2810 if (!ConstantMemory) {
2811 SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2813 &Chains[0], NumValues);
2817 PendingLoads.push_back(Chain);
2820 setValue(&I, DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
2821 DAG.getVTList(&ValueVTs[0], NumValues),
2822 &Values[0], NumValues));
2826 void SelectionDAGLowering::visitStore(StoreInst &I) {
2827 Value *SrcV = I.getOperand(0);
2828 Value *PtrV = I.getOperand(1);
2830 SmallVector<MVT, 4> ValueVTs;
2831 SmallVector<uint64_t, 4> Offsets;
2832 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2833 unsigned NumValues = ValueVTs.size();
2837 // Get the lowered operands. Note that we do this after
2838 // checking if NumResults is zero, because with zero results
2839 // the operands won't have values in the map.
2840 SDValue Src = getValue(SrcV);
2841 SDValue Ptr = getValue(PtrV);
2843 SDValue Root = getRoot();
2844 SmallVector<SDValue, 4> Chains(NumValues);
2845 MVT PtrVT = Ptr.getValueType();
2846 bool isVolatile = I.isVolatile();
2847 unsigned Alignment = I.getAlignment();
2848 for (unsigned i = 0; i != NumValues; ++i)
2849 Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
2850 SDValue(Src.getNode(), Src.getResNo() + i),
2851 DAG.getNode(ISD::ADD, getCurDebugLoc(),
2853 DAG.getConstant(Offsets[i], PtrVT)),
2855 isVolatile, Alignment);
2857 DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
2858 MVT::Other, &Chains[0], NumValues));
2861 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2863 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
2864 unsigned Intrinsic) {
2865 bool HasChain = !I.doesNotAccessMemory();
2866 bool OnlyLoad = HasChain && I.onlyReadsMemory();
2868 // Build the operand list.
2869 SmallVector<SDValue, 8> Ops;
2870 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2872 // We don't need to serialize loads against other loads.
2873 Ops.push_back(DAG.getRoot());
2875 Ops.push_back(getRoot());
2879 // Info is set by getTgtMemInstrinsic
2880 TargetLowering::IntrinsicInfo Info;
2881 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
2883 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
2884 if (!IsTgtIntrinsic)
2885 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2887 // Add all operands of the call to the operand list.
2888 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2889 SDValue Op = getValue(I.getOperand(i));
2890 assert(TLI.isTypeLegal(Op.getValueType()) &&
2891 "Intrinsic uses a non-legal type?");
2895 std::vector<MVT> VTArray;
2896 if (I.getType() != Type::VoidTy) {
2897 MVT VT = TLI.getValueType(I.getType());
2898 if (VT.isVector()) {
2899 const VectorType *DestTy = cast<VectorType>(I.getType());
2900 MVT EltVT = TLI.getValueType(DestTy->getElementType());
2902 VT = MVT::getVectorVT(EltVT, DestTy->getNumElements());
2903 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
2906 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
2907 VTArray.push_back(VT);
2910 VTArray.push_back(MVT::Other);
2912 SDVTList VTs = DAG.getVTList(&VTArray[0], VTArray.size());
2916 if (IsTgtIntrinsic) {
2917 // This is target intrinsic that touches memory
2918 Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
2919 VTs, &Ops[0], Ops.size(),
2920 Info.memVT, Info.ptrVal, Info.offset,
2921 Info.align, Info.vol,
2922 Info.readMem, Info.writeMem);
2925 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, getCurDebugLoc(),
2926 VTs, &Ops[0], Ops.size());
2927 else if (I.getType() != Type::VoidTy)
2928 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, getCurDebugLoc(),
2929 VTs, &Ops[0], Ops.size());
2931 Result = DAG.getNode(ISD::INTRINSIC_VOID, getCurDebugLoc(),
2932 VTs, &Ops[0], Ops.size());
2935 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
2937 PendingLoads.push_back(Chain);
2941 if (I.getType() != Type::VoidTy) {
2942 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
2943 MVT VT = TLI.getValueType(PTy);
2944 Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
2946 setValue(&I, Result);
2950 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
2951 static GlobalVariable *ExtractTypeInfo(Value *V) {
2952 V = V->stripPointerCasts();
2953 GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2954 assert ((GV || isa<ConstantPointerNull>(V)) &&
2955 "TypeInfo must be a global variable or NULL");
2961 /// AddCatchInfo - Extract the personality and type infos from an eh.selector
2962 /// call, and add them to the specified machine basic block.
2963 void AddCatchInfo(CallInst &I, MachineModuleInfo *MMI,
2964 MachineBasicBlock *MBB) {
2965 // Inform the MachineModuleInfo of the personality for this landing pad.
2966 ConstantExpr *CE = cast<ConstantExpr>(I.getOperand(2));
2967 assert(CE->getOpcode() == Instruction::BitCast &&
2968 isa<Function>(CE->getOperand(0)) &&
2969 "Personality should be a function");
2970 MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
2972 // Gather all the type infos for this landing pad and pass them along to
2973 // MachineModuleInfo.
2974 std::vector<GlobalVariable *> TyInfo;
2975 unsigned N = I.getNumOperands();
2977 for (unsigned i = N - 1; i > 2; --i) {
2978 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(i))) {
2979 unsigned FilterLength = CI->getZExtValue();
2980 unsigned FirstCatch = i + FilterLength + !FilterLength;
2981 assert (FirstCatch <= N && "Invalid filter length");
2983 if (FirstCatch < N) {
2984 TyInfo.reserve(N - FirstCatch);
2985 for (unsigned j = FirstCatch; j < N; ++j)
2986 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2987 MMI->addCatchTypeInfo(MBB, TyInfo);
2991 if (!FilterLength) {
2993 MMI->addCleanup(MBB);
2996 TyInfo.reserve(FilterLength - 1);
2997 for (unsigned j = i + 1; j < FirstCatch; ++j)
2998 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2999 MMI->addFilterTypeInfo(MBB, TyInfo);
3008 TyInfo.reserve(N - 3);
3009 for (unsigned j = 3; j < N; ++j)
3010 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
3011 MMI->addCatchTypeInfo(MBB, TyInfo);
3017 /// GetSignificand - Get the significand and build it into a floating-point
3018 /// number with exponent of 1:
3020 /// Op = (Op & 0x007fffff) | 0x3f800000;
3022 /// where Op is the hexidecimal representation of floating point value.
3024 GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) {
3025 SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3026 DAG.getConstant(0x007fffff, MVT::i32));
3027 SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
3028 DAG.getConstant(0x3f800000, MVT::i32));
3029 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
3032 /// GetExponent - Get the exponent:
3034 /// (float)(int)(((Op & 0x7f800000) >> 23) - 127);
3036 /// where Op is the hexidecimal representation of floating point value.
3038 GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI,
3040 SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op,
3041 DAG.getConstant(0x7f800000, MVT::i32));
3042 SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0,
3043 DAG.getConstant(23, TLI.getPointerTy()));
3044 SDValue t2 = DAG.getNode(ISD::SUB, dl, MVT::i32, t1,
3045 DAG.getConstant(127, MVT::i32));
3046 return DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, t2);
3049 /// getF32Constant - Get 32-bit floating point constant.
3051 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
3052 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
3055 /// Inlined utility function to implement binary input atomic intrinsics for
3056 /// visitIntrinsicCall: I is a call instruction
3057 /// Op is the associated NodeType for I
3059 SelectionDAGLowering::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) {
3060 SDValue Root = getRoot();
3062 DAG.getAtomic(Op, getCurDebugLoc(),
3063 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
3065 getValue(I.getOperand(1)),
3066 getValue(I.getOperand(2)),
3069 DAG.setRoot(L.getValue(1));
3073 // implVisitAluOverflow - Lower arithmetic overflow instrinsics.
3075 SelectionDAGLowering::implVisitAluOverflow(CallInst &I, ISD::NodeType Op) {
3076 SDValue Op1 = getValue(I.getOperand(1));
3077 SDValue Op2 = getValue(I.getOperand(2));
3079 SDVTList VTs = DAG.getVTList(Op1.getValueType(), MVT::i1);
3080 SDValue Result = DAG.getNode(Op, getCurDebugLoc(), VTs, Op1, Op2);
3082 setValue(&I, Result);
3086 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
3087 /// limited-precision mode.
3089 SelectionDAGLowering::visitExp(CallInst &I) {
3091 DebugLoc dl = getCurDebugLoc();
3093 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3094 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3095 SDValue Op = getValue(I.getOperand(1));
3097 // Put the exponent in the right bit position for later addition to the
3100 // #define LOG2OFe 1.4426950f
3101 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
3102 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3103 getF32Constant(DAG, 0x3fb8aa3b));
3104 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3106 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
3107 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3108 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3110 // IntegerPartOfX <<= 23;
3111 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3112 DAG.getConstant(23, TLI.getPointerTy()));
3114 if (LimitFloatPrecision <= 6) {
3115 // For floating-point precision of 6:
3117 // TwoToFractionalPartOfX =
3119 // (0.735607626f + 0.252464424f * x) * x;
3121 // error 0.0144103317, which is 6 bits
3122 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3123 getF32Constant(DAG, 0x3e814304));
3124 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3125 getF32Constant(DAG, 0x3f3c50c8));
3126 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3127 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3128 getF32Constant(DAG, 0x3f7f5e7e));
3129 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
3131 // Add the exponent into the result in integer domain.
3132 SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3133 TwoToFracPartOfX, IntegerPartOfX);
3135 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
3136 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3137 // For floating-point precision of 12:
3139 // TwoToFractionalPartOfX =
3142 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3144 // 0.000107046256 error, which is 13 to 14 bits
3145 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3146 getF32Constant(DAG, 0x3da235e3));
3147 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3148 getF32Constant(DAG, 0x3e65b8f3));
3149 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3150 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3151 getF32Constant(DAG, 0x3f324b07));
3152 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3153 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3154 getF32Constant(DAG, 0x3f7ff8fd));
3155 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
3157 // Add the exponent into the result in integer domain.
3158 SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3159 TwoToFracPartOfX, IntegerPartOfX);
3161 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
3162 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3163 // For floating-point precision of 18:
3165 // TwoToFractionalPartOfX =
3169 // (0.554906021e-1f +
3170 // (0.961591928e-2f +
3171 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3173 // error 2.47208000*10^(-7), which is better than 18 bits
3174 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3175 getF32Constant(DAG, 0x3924b03e));
3176 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3177 getF32Constant(DAG, 0x3ab24b87));
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, 0x3c1d8c17));
3181 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3182 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3183 getF32Constant(DAG, 0x3d634a1d));
3184 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3185 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3186 getF32Constant(DAG, 0x3e75fe14));
3187 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3188 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3189 getF32Constant(DAG, 0x3f317234));
3190 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3191 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3192 getF32Constant(DAG, 0x3f800000));
3193 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
3196 // Add the exponent into the result in integer domain.
3197 SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
3198 TwoToFracPartOfX, IntegerPartOfX);
3200 result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
3203 // No special expansion.
3204 result = DAG.getNode(ISD::FEXP, dl,
3205 getValue(I.getOperand(1)).getValueType(),
3206 getValue(I.getOperand(1)));
3209 setValue(&I, result);
3212 /// visitLog - Lower a log intrinsic. Handles the special sequences for
3213 /// limited-precision mode.
3215 SelectionDAGLowering::visitLog(CallInst &I) {
3217 DebugLoc dl = getCurDebugLoc();
3219 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3220 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3221 SDValue Op = getValue(I.getOperand(1));
3222 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3224 // Scale the exponent by log(2) [0.69314718f].
3225 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3226 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3227 getF32Constant(DAG, 0x3f317218));
3229 // Get the significand and build it into a floating-point number with
3231 SDValue X = GetSignificand(DAG, Op1, dl);
3233 if (LimitFloatPrecision <= 6) {
3234 // For floating-point precision of 6:
3238 // (1.4034025f - 0.23903021f * x) * x;
3240 // error 0.0034276066, which is better than 8 bits
3241 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3242 getF32Constant(DAG, 0xbe74c456));
3243 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3244 getF32Constant(DAG, 0x3fb3a2b1));
3245 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3246 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3247 getF32Constant(DAG, 0x3f949a29));
3249 result = DAG.getNode(ISD::FADD, dl,
3250 MVT::f32, LogOfExponent, LogOfMantissa);
3251 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3252 // For floating-point precision of 12:
3258 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
3260 // error 0.000061011436, which is 14 bits
3261 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3262 getF32Constant(DAG, 0xbd67b6d6));
3263 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3264 getF32Constant(DAG, 0x3ee4f4b8));
3265 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3266 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3267 getF32Constant(DAG, 0x3fbc278b));
3268 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3269 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3270 getF32Constant(DAG, 0x40348e95));
3271 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3272 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3273 getF32Constant(DAG, 0x3fdef31a));
3275 result = DAG.getNode(ISD::FADD, dl,
3276 MVT::f32, LogOfExponent, LogOfMantissa);
3277 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3278 // For floating-point precision of 18:
3286 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
3288 // error 0.0000023660568, which is better than 18 bits
3289 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3290 getF32Constant(DAG, 0xbc91e5ac));
3291 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3292 getF32Constant(DAG, 0x3e4350aa));
3293 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3294 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3295 getF32Constant(DAG, 0x3f60d3e3));
3296 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3297 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3298 getF32Constant(DAG, 0x4011cdf0));
3299 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3300 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3301 getF32Constant(DAG, 0x406cfd1c));
3302 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3303 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3304 getF32Constant(DAG, 0x408797cb));
3305 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3306 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3307 getF32Constant(DAG, 0x4006dcab));
3309 result = DAG.getNode(ISD::FADD, dl,
3310 MVT::f32, LogOfExponent, LogOfMantissa);
3313 // No special expansion.
3314 result = DAG.getNode(ISD::FLOG, dl,
3315 getValue(I.getOperand(1)).getValueType(),
3316 getValue(I.getOperand(1)));
3319 setValue(&I, result);
3322 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3323 /// limited-precision mode.
3325 SelectionDAGLowering::visitLog2(CallInst &I) {
3327 DebugLoc dl = getCurDebugLoc();
3329 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3330 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3331 SDValue Op = getValue(I.getOperand(1));
3332 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3334 // Get the exponent.
3335 SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
3337 // Get the significand and build it into a floating-point number with
3339 SDValue X = GetSignificand(DAG, Op1, dl);
3341 // Different possible minimax approximations of significand in
3342 // floating-point for various degrees of accuracy over [1,2].
3343 if (LimitFloatPrecision <= 6) {
3344 // For floating-point precision of 6:
3346 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3348 // error 0.0049451742, which is more than 7 bits
3349 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3350 getF32Constant(DAG, 0xbeb08fe0));
3351 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3352 getF32Constant(DAG, 0x40019463));
3353 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3354 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3355 getF32Constant(DAG, 0x3fd6633d));
3357 result = DAG.getNode(ISD::FADD, dl,
3358 MVT::f32, LogOfExponent, Log2ofMantissa);
3359 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3360 // For floating-point precision of 12:
3366 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3368 // error 0.0000876136000, which is better than 13 bits
3369 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3370 getF32Constant(DAG, 0xbda7262e));
3371 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3372 getF32Constant(DAG, 0x3f25280b));
3373 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3374 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3375 getF32Constant(DAG, 0x4007b923));
3376 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3377 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3378 getF32Constant(DAG, 0x40823e2f));
3379 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3380 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3381 getF32Constant(DAG, 0x4020d29c));
3383 result = DAG.getNode(ISD::FADD, dl,
3384 MVT::f32, LogOfExponent, Log2ofMantissa);
3385 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3386 // For floating-point precision of 18:
3395 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3397 // error 0.0000018516, which is better than 18 bits
3398 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3399 getF32Constant(DAG, 0xbcd2769e));
3400 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3401 getF32Constant(DAG, 0x3e8ce0b9));
3402 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3403 SDValue t3 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3404 getF32Constant(DAG, 0x3fa22ae7));
3405 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3406 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3407 getF32Constant(DAG, 0x40525723));
3408 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3409 SDValue t7 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t6,
3410 getF32Constant(DAG, 0x40aaf200));
3411 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3412 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3413 getF32Constant(DAG, 0x40c39dad));
3414 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3415 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t10,
3416 getF32Constant(DAG, 0x4042902c));
3418 result = DAG.getNode(ISD::FADD, dl,
3419 MVT::f32, LogOfExponent, Log2ofMantissa);
3422 // No special expansion.
3423 result = DAG.getNode(ISD::FLOG2, dl,
3424 getValue(I.getOperand(1)).getValueType(),
3425 getValue(I.getOperand(1)));
3428 setValue(&I, result);
3431 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3432 /// limited-precision mode.
3434 SelectionDAGLowering::visitLog10(CallInst &I) {
3436 DebugLoc dl = getCurDebugLoc();
3438 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3439 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3440 SDValue Op = getValue(I.getOperand(1));
3441 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
3443 // Scale the exponent by log10(2) [0.30102999f].
3444 SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
3445 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp,
3446 getF32Constant(DAG, 0x3e9a209a));
3448 // Get the significand and build it into a floating-point number with
3450 SDValue X = GetSignificand(DAG, Op1, dl);
3452 if (LimitFloatPrecision <= 6) {
3453 // For floating-point precision of 6:
3455 // Log10ofMantissa =
3457 // (0.60948995f - 0.10380950f * x) * x;
3459 // error 0.0014886165, which is 6 bits
3460 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3461 getF32Constant(DAG, 0xbdd49a13));
3462 SDValue t1 = DAG.getNode(ISD::FADD, dl, MVT::f32, t0,
3463 getF32Constant(DAG, 0x3f1c0789));
3464 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3465 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t2,
3466 getF32Constant(DAG, 0x3f011300));
3468 result = DAG.getNode(ISD::FADD, dl,
3469 MVT::f32, LogOfExponent, Log10ofMantissa);
3470 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3471 // For floating-point precision of 12:
3473 // Log10ofMantissa =
3476 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3478 // error 0.00019228036, which is better than 12 bits
3479 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3480 getF32Constant(DAG, 0x3d431f31));
3481 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3482 getF32Constant(DAG, 0x3ea21fb2));
3483 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3484 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3485 getF32Constant(DAG, 0x3f6ae232));
3486 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3487 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3488 getF32Constant(DAG, 0x3f25f7c3));
3490 result = DAG.getNode(ISD::FADD, dl,
3491 MVT::f32, LogOfExponent, Log10ofMantissa);
3492 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3493 // For floating-point precision of 18:
3495 // Log10ofMantissa =
3500 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3502 // error 0.0000037995730, which is better than 18 bits
3503 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3504 getF32Constant(DAG, 0x3c5d51ce));
3505 SDValue t1 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0,
3506 getF32Constant(DAG, 0x3e00685a));
3507 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t1, X);
3508 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3509 getF32Constant(DAG, 0x3efb6798));
3510 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3511 SDValue t5 = DAG.getNode(ISD::FSUB, dl, MVT::f32, t4,
3512 getF32Constant(DAG, 0x3f88d192));
3513 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3514 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3515 getF32Constant(DAG, 0x3fc4316c));
3516 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3517 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, dl, MVT::f32, t8,
3518 getF32Constant(DAG, 0x3f57ce70));
3520 result = DAG.getNode(ISD::FADD, dl,
3521 MVT::f32, LogOfExponent, Log10ofMantissa);
3524 // No special expansion.
3525 result = DAG.getNode(ISD::FLOG10, dl,
3526 getValue(I.getOperand(1)).getValueType(),
3527 getValue(I.getOperand(1)));
3530 setValue(&I, result);
3533 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3534 /// limited-precision mode.
3536 SelectionDAGLowering::visitExp2(CallInst &I) {
3538 DebugLoc dl = getCurDebugLoc();
3540 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3541 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3542 SDValue Op = getValue(I.getOperand(1));
3544 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, Op);
3546 // FractionalPartOfX = x - (float)IntegerPartOfX;
3547 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3548 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, Op, t1);
3550 // IntegerPartOfX <<= 23;
3551 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3552 DAG.getConstant(23, TLI.getPointerTy()));
3554 if (LimitFloatPrecision <= 6) {
3555 // For floating-point precision of 6:
3557 // TwoToFractionalPartOfX =
3559 // (0.735607626f + 0.252464424f * x) * x;
3561 // error 0.0144103317, which is 6 bits
3562 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3563 getF32Constant(DAG, 0x3e814304));
3564 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3565 getF32Constant(DAG, 0x3f3c50c8));
3566 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3567 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3568 getF32Constant(DAG, 0x3f7f5e7e));
3569 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3570 SDValue TwoToFractionalPartOfX =
3571 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3573 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3574 MVT::f32, TwoToFractionalPartOfX);
3575 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3576 // For floating-point precision of 12:
3578 // TwoToFractionalPartOfX =
3581 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3583 // error 0.000107046256, which is 13 to 14 bits
3584 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3585 getF32Constant(DAG, 0x3da235e3));
3586 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3587 getF32Constant(DAG, 0x3e65b8f3));
3588 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3589 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3590 getF32Constant(DAG, 0x3f324b07));
3591 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3592 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3593 getF32Constant(DAG, 0x3f7ff8fd));
3594 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3595 SDValue TwoToFractionalPartOfX =
3596 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3598 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3599 MVT::f32, TwoToFractionalPartOfX);
3600 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3601 // For floating-point precision of 18:
3603 // TwoToFractionalPartOfX =
3607 // (0.554906021e-1f +
3608 // (0.961591928e-2f +
3609 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3610 // error 2.47208000*10^(-7), which is better than 18 bits
3611 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3612 getF32Constant(DAG, 0x3924b03e));
3613 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3614 getF32Constant(DAG, 0x3ab24b87));
3615 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3616 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3617 getF32Constant(DAG, 0x3c1d8c17));
3618 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3619 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3620 getF32Constant(DAG, 0x3d634a1d));
3621 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3622 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3623 getF32Constant(DAG, 0x3e75fe14));
3624 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3625 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3626 getF32Constant(DAG, 0x3f317234));
3627 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3628 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3629 getF32Constant(DAG, 0x3f800000));
3630 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3631 SDValue TwoToFractionalPartOfX =
3632 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3634 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3635 MVT::f32, TwoToFractionalPartOfX);
3638 // No special expansion.
3639 result = DAG.getNode(ISD::FEXP2, dl,
3640 getValue(I.getOperand(1)).getValueType(),
3641 getValue(I.getOperand(1)));
3644 setValue(&I, result);
3647 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3648 /// limited-precision mode with x == 10.0f.
3650 SelectionDAGLowering::visitPow(CallInst &I) {
3652 Value *Val = I.getOperand(1);
3653 DebugLoc dl = getCurDebugLoc();
3654 bool IsExp10 = false;
3656 if (getValue(Val).getValueType() == MVT::f32 &&
3657 getValue(I.getOperand(2)).getValueType() == MVT::f32 &&
3658 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3659 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3660 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3662 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3667 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3668 SDValue Op = getValue(I.getOperand(2));
3670 // Put the exponent in the right bit position for later addition to the
3673 // #define LOG2OF10 3.3219281f
3674 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3675 SDValue t0 = DAG.getNode(ISD::FMUL, dl, MVT::f32, Op,
3676 getF32Constant(DAG, 0x40549a78));
3677 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, dl, MVT::i32, t0);
3679 // FractionalPartOfX = x - (float)IntegerPartOfX;
3680 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, IntegerPartOfX);
3681 SDValue X = DAG.getNode(ISD::FSUB, dl, MVT::f32, t0, t1);
3683 // IntegerPartOfX <<= 23;
3684 IntegerPartOfX = DAG.getNode(ISD::SHL, dl, MVT::i32, IntegerPartOfX,
3685 DAG.getConstant(23, TLI.getPointerTy()));
3687 if (LimitFloatPrecision <= 6) {
3688 // For floating-point precision of 6:
3690 // twoToFractionalPartOfX =
3692 // (0.735607626f + 0.252464424f * x) * x;
3694 // error 0.0144103317, which is 6 bits
3695 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3696 getF32Constant(DAG, 0x3e814304));
3697 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3698 getF32Constant(DAG, 0x3f3c50c8));
3699 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3700 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3701 getF32Constant(DAG, 0x3f7f5e7e));
3702 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
3703 SDValue TwoToFractionalPartOfX =
3704 DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
3706 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3707 MVT::f32, TwoToFractionalPartOfX);
3708 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3709 // For floating-point precision of 12:
3711 // TwoToFractionalPartOfX =
3714 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3716 // error 0.000107046256, which is 13 to 14 bits
3717 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3718 getF32Constant(DAG, 0x3da235e3));
3719 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3720 getF32Constant(DAG, 0x3e65b8f3));
3721 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3722 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3723 getF32Constant(DAG, 0x3f324b07));
3724 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3725 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3726 getF32Constant(DAG, 0x3f7ff8fd));
3727 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
3728 SDValue TwoToFractionalPartOfX =
3729 DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
3731 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3732 MVT::f32, TwoToFractionalPartOfX);
3733 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3734 // For floating-point precision of 18:
3736 // TwoToFractionalPartOfX =
3740 // (0.554906021e-1f +
3741 // (0.961591928e-2f +
3742 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3743 // error 2.47208000*10^(-7), which is better than 18 bits
3744 SDValue t2 = DAG.getNode(ISD::FMUL, dl, MVT::f32, X,
3745 getF32Constant(DAG, 0x3924b03e));
3746 SDValue t3 = DAG.getNode(ISD::FADD, dl, MVT::f32, t2,
3747 getF32Constant(DAG, 0x3ab24b87));
3748 SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
3749 SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
3750 getF32Constant(DAG, 0x3c1d8c17));
3751 SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
3752 SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
3753 getF32Constant(DAG, 0x3d634a1d));
3754 SDValue t8 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t7, X);
3755 SDValue t9 = DAG.getNode(ISD::FADD, dl, MVT::f32, t8,
3756 getF32Constant(DAG, 0x3e75fe14));
3757 SDValue t10 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t9, X);
3758 SDValue t11 = DAG.getNode(ISD::FADD, dl, MVT::f32, t10,
3759 getF32Constant(DAG, 0x3f317234));
3760 SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
3761 SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
3762 getF32Constant(DAG, 0x3f800000));
3763 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
3764 SDValue TwoToFractionalPartOfX =
3765 DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
3767 result = DAG.getNode(ISD::BIT_CONVERT, dl,
3768 MVT::f32, TwoToFractionalPartOfX);
3771 // No special expansion.
3772 result = DAG.getNode(ISD::FPOW, dl,
3773 getValue(I.getOperand(1)).getValueType(),
3774 getValue(I.getOperand(1)),
3775 getValue(I.getOperand(2)));
3778 setValue(&I, result);
3781 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3782 /// we want to emit this as a call to a named external function, return the name
3783 /// otherwise lower it and return null.
3785 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
3786 DebugLoc dl = getCurDebugLoc();
3787 switch (Intrinsic) {
3789 // By default, turn this into a target intrinsic node.
3790 visitTargetIntrinsic(I, Intrinsic);
3792 case Intrinsic::vastart: visitVAStart(I); return 0;
3793 case Intrinsic::vaend: visitVAEnd(I); return 0;
3794 case Intrinsic::vacopy: visitVACopy(I); return 0;
3795 case Intrinsic::returnaddress:
3796 setValue(&I, DAG.getNode(ISD::RETURNADDR, dl, TLI.getPointerTy(),
3797 getValue(I.getOperand(1))));
3799 case Intrinsic::frameaddress:
3800 setValue(&I, DAG.getNode(ISD::FRAMEADDR, dl, TLI.getPointerTy(),
3801 getValue(I.getOperand(1))));
3803 case Intrinsic::setjmp:
3804 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
3806 case Intrinsic::longjmp:
3807 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
3809 case Intrinsic::memcpy: {
3810 SDValue Op1 = getValue(I.getOperand(1));
3811 SDValue Op2 = getValue(I.getOperand(2));
3812 SDValue Op3 = getValue(I.getOperand(3));
3813 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3814 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
3815 I.getOperand(1), 0, I.getOperand(2), 0));
3818 case Intrinsic::memset: {
3819 SDValue Op1 = getValue(I.getOperand(1));
3820 SDValue Op2 = getValue(I.getOperand(2));
3821 SDValue Op3 = getValue(I.getOperand(3));
3822 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3823 DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align,
3824 I.getOperand(1), 0));
3827 case Intrinsic::memmove: {
3828 SDValue Op1 = getValue(I.getOperand(1));
3829 SDValue Op2 = getValue(I.getOperand(2));
3830 SDValue Op3 = getValue(I.getOperand(3));
3831 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3833 // If the source and destination are known to not be aliases, we can
3834 // lower memmove as memcpy.
3835 uint64_t Size = -1ULL;
3836 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
3837 Size = C->getZExtValue();
3838 if (AA->alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
3839 AliasAnalysis::NoAlias) {
3840 DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, false,
3841 I.getOperand(1), 0, I.getOperand(2), 0));
3845 DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align,
3846 I.getOperand(1), 0, I.getOperand(2), 0));
3849 case Intrinsic::dbg_stoppoint: {
3850 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
3851 if (isValidDebugInfoIntrinsic(SPI, CodeGenOpt::Default)) {
3852 MachineFunction &MF = DAG.getMachineFunction();
3853 DebugLoc Loc = ExtractDebugLocation(SPI, MF.getDebugLocInfo());
3854 setCurDebugLoc(Loc);
3856 if (OptLevel == CodeGenOpt::None)
3857 DAG.setRoot(DAG.getDbgStopPoint(Loc, getRoot(),
3864 case Intrinsic::dbg_region_start: {
3865 DwarfWriter *DW = DAG.getDwarfWriter();
3866 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
3867 if (isValidDebugInfoIntrinsic(RSI, OptLevel) && DW
3868 && DW->ShouldEmitDwarfDebug()) {
3870 DW->RecordRegionStart(cast<GlobalVariable>(RSI.getContext()));
3871 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3872 getRoot(), LabelID));
3876 case Intrinsic::dbg_region_end: {
3877 DwarfWriter *DW = DAG.getDwarfWriter();
3878 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
3880 if (!isValidDebugInfoIntrinsic(REI, OptLevel) || !DW
3881 || !DW->ShouldEmitDwarfDebug())
3884 MachineFunction &MF = DAG.getMachineFunction();
3885 DISubprogram Subprogram(cast<GlobalVariable>(REI.getContext()));
3887 if (isInlinedFnEnd(REI, MF.getFunction())) {
3888 // This is end of inlined function. Debugging information for inlined
3889 // function is not handled yet (only supported by FastISel).
3890 if (OptLevel == CodeGenOpt::None) {
3891 unsigned ID = DW->RecordInlinedFnEnd(Subprogram);
3893 // Returned ID is 0 if this is unbalanced "end of inlined
3894 // scope". This could happen if optimizer eats dbg intrinsics or
3895 // "beginning of inlined scope" is not recoginized due to missing
3896 // location info. In such cases, do ignore this region.end.
3897 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3904 DW->RecordRegionEnd(cast<GlobalVariable>(REI.getContext()));
3905 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3906 getRoot(), LabelID));
3909 case Intrinsic::dbg_func_start: {
3910 DwarfWriter *DW = DAG.getDwarfWriter();
3911 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
3912 if (!isValidDebugInfoIntrinsic(FSI, CodeGenOpt::None) || !DW
3913 || !DW->ShouldEmitDwarfDebug())
3916 MachineFunction &MF = DAG.getMachineFunction();
3917 // This is a beginning of an inlined function.
3918 if (isInlinedFnStart(FSI, MF.getFunction())) {
3919 if (OptLevel != CodeGenOpt::None)
3920 // FIXME: Debugging informaation for inlined function is only
3921 // supported at CodeGenOpt::Node.
3924 DebugLoc PrevLoc = CurDebugLoc;
3925 // If llvm.dbg.func.start is seen in a new block before any
3926 // llvm.dbg.stoppoint intrinsic then the location info is unknown.
3927 // FIXME : Why DebugLoc is reset at the beginning of each block ?
3928 if (PrevLoc.isUnknown())
3931 // Record the source line.
3932 setCurDebugLoc(ExtractDebugLocation(FSI, MF.getDebugLocInfo()));
3934 DebugLocTuple PrevLocTpl = MF.getDebugLocTuple(PrevLoc);
3935 DISubprogram SP(cast<GlobalVariable>(FSI.getSubprogram()));
3936 DICompileUnit CU(PrevLocTpl.CompileUnit);
3937 unsigned LabelID = DW->RecordInlinedFnStart(SP, CU,
3940 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getCurDebugLoc(),
3941 getRoot(), LabelID));
3945 // This is a beginning of a new function.
3946 MF.setDefaultDebugLoc(ExtractDebugLocation(FSI, MF.getDebugLocInfo()));
3948 // llvm.dbg.func_start also defines beginning of function scope.
3949 DW->RecordRegionStart(cast<GlobalVariable>(FSI.getSubprogram()));
3952 case Intrinsic::dbg_declare: {
3953 if (OptLevel != CodeGenOpt::None)
3954 // FIXME: Variable debug info is not supported here.
3957 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
3958 if (!isValidDebugInfoIntrinsic(DI, CodeGenOpt::None))
3961 Value *Variable = DI.getVariable();
3962 DAG.setRoot(DAG.getNode(ISD::DECLARE, dl, MVT::Other, getRoot(),
3963 getValue(DI.getAddress()), getValue(Variable)));
3966 case Intrinsic::eh_exception: {
3967 // Insert the EXCEPTIONADDR instruction.
3968 assert(CurMBB->isLandingPad() &&"Call to eh.exception not in landing pad!");
3969 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
3971 Ops[0] = DAG.getRoot();
3972 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, dl, VTs, Ops, 1);
3974 DAG.setRoot(Op.getValue(1));
3978 case Intrinsic::eh_selector_i32:
3979 case Intrinsic::eh_selector_i64: {
3980 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3981 MVT VT = (Intrinsic == Intrinsic::eh_selector_i32 ?
3982 MVT::i32 : MVT::i64);
3985 if (CurMBB->isLandingPad())
3986 AddCatchInfo(I, MMI, CurMBB);
3989 FuncInfo.CatchInfoLost.insert(&I);
3991 // FIXME: Mark exception selector register as live in. Hack for PR1508.
3992 unsigned Reg = TLI.getExceptionSelectorRegister();
3993 if (Reg) CurMBB->addLiveIn(Reg);
3996 // Insert the EHSELECTION instruction.
3997 SDVTList VTs = DAG.getVTList(VT, MVT::Other);
3999 Ops[0] = getValue(I.getOperand(1));
4001 SDValue Op = DAG.getNode(ISD::EHSELECTION, dl, VTs, Ops, 2);
4003 DAG.setRoot(Op.getValue(1));
4005 setValue(&I, DAG.getConstant(0, VT));
4011 case Intrinsic::eh_typeid_for_i32:
4012 case Intrinsic::eh_typeid_for_i64: {
4013 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4014 MVT VT = (Intrinsic == Intrinsic::eh_typeid_for_i32 ?
4015 MVT::i32 : MVT::i64);
4018 // Find the type id for the given typeinfo.
4019 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
4021 unsigned TypeID = MMI->getTypeIDFor(GV);
4022 setValue(&I, DAG.getConstant(TypeID, VT));
4024 // Return something different to eh_selector.
4025 setValue(&I, DAG.getConstant(1, VT));
4031 case Intrinsic::eh_return_i32:
4032 case Intrinsic::eh_return_i64:
4033 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4034 MMI->setCallsEHReturn(true);
4035 DAG.setRoot(DAG.getNode(ISD::EH_RETURN, dl,
4038 getValue(I.getOperand(1)),
4039 getValue(I.getOperand(2))));
4041 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
4045 case Intrinsic::eh_unwind_init:
4046 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
4047 MMI->setCallsUnwindInit(true);
4052 case Intrinsic::eh_dwarf_cfa: {
4053 MVT VT = getValue(I.getOperand(1)).getValueType();
4055 if (VT.bitsGT(TLI.getPointerTy()))
4056 CfaArg = DAG.getNode(ISD::TRUNCATE, dl,
4057 TLI.getPointerTy(), getValue(I.getOperand(1)));
4059 CfaArg = DAG.getNode(ISD::SIGN_EXTEND, dl,
4060 TLI.getPointerTy(), getValue(I.getOperand(1)));
4062 SDValue Offset = DAG.getNode(ISD::ADD, dl,
4064 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET, dl,
4065 TLI.getPointerTy()),
4067 setValue(&I, DAG.getNode(ISD::ADD, dl,
4069 DAG.getNode(ISD::FRAMEADDR, dl,
4072 TLI.getPointerTy())),
4077 case Intrinsic::convertff:
4078 case Intrinsic::convertfsi:
4079 case Intrinsic::convertfui:
4080 case Intrinsic::convertsif:
4081 case Intrinsic::convertuif:
4082 case Intrinsic::convertss:
4083 case Intrinsic::convertsu:
4084 case Intrinsic::convertus:
4085 case Intrinsic::convertuu: {
4086 ISD::CvtCode Code = ISD::CVT_INVALID;
4087 switch (Intrinsic) {
4088 case Intrinsic::convertff: Code = ISD::CVT_FF; break;
4089 case Intrinsic::convertfsi: Code = ISD::CVT_FS; break;
4090 case Intrinsic::convertfui: Code = ISD::CVT_FU; break;
4091 case Intrinsic::convertsif: Code = ISD::CVT_SF; break;
4092 case Intrinsic::convertuif: Code = ISD::CVT_UF; break;
4093 case Intrinsic::convertss: Code = ISD::CVT_SS; break;
4094 case Intrinsic::convertsu: Code = ISD::CVT_SU; break;
4095 case Intrinsic::convertus: Code = ISD::CVT_US; break;
4096 case Intrinsic::convertuu: Code = ISD::CVT_UU; break;
4098 MVT DestVT = TLI.getValueType(I.getType());
4099 Value* Op1 = I.getOperand(1);
4100 setValue(&I, DAG.getConvertRndSat(DestVT, getCurDebugLoc(), getValue(Op1),
4101 DAG.getValueType(DestVT),
4102 DAG.getValueType(getValue(Op1).getValueType()),
4103 getValue(I.getOperand(2)),
4104 getValue(I.getOperand(3)),
4109 case Intrinsic::sqrt:
4110 setValue(&I, DAG.getNode(ISD::FSQRT, dl,
4111 getValue(I.getOperand(1)).getValueType(),
4112 getValue(I.getOperand(1))));
4114 case Intrinsic::powi:
4115 setValue(&I, DAG.getNode(ISD::FPOWI, dl,
4116 getValue(I.getOperand(1)).getValueType(),
4117 getValue(I.getOperand(1)),
4118 getValue(I.getOperand(2))));
4120 case Intrinsic::sin:
4121 setValue(&I, DAG.getNode(ISD::FSIN, dl,
4122 getValue(I.getOperand(1)).getValueType(),
4123 getValue(I.getOperand(1))));
4125 case Intrinsic::cos:
4126 setValue(&I, DAG.getNode(ISD::FCOS, dl,
4127 getValue(I.getOperand(1)).getValueType(),
4128 getValue(I.getOperand(1))));
4130 case Intrinsic::log:
4133 case Intrinsic::log2:
4136 case Intrinsic::log10:
4139 case Intrinsic::exp:
4142 case Intrinsic::exp2:
4145 case Intrinsic::pow:
4148 case Intrinsic::pcmarker: {
4149 SDValue Tmp = getValue(I.getOperand(1));
4150 DAG.setRoot(DAG.getNode(ISD::PCMARKER, dl, MVT::Other, getRoot(), Tmp));
4153 case Intrinsic::readcyclecounter: {
4154 SDValue Op = getRoot();
4155 SDValue Tmp = DAG.getNode(ISD::READCYCLECOUNTER, dl,
4156 DAG.getVTList(MVT::i64, MVT::Other),
4159 DAG.setRoot(Tmp.getValue(1));
4162 case Intrinsic::bswap:
4163 setValue(&I, DAG.getNode(ISD::BSWAP, dl,
4164 getValue(I.getOperand(1)).getValueType(),
4165 getValue(I.getOperand(1))));
4167 case Intrinsic::cttz: {
4168 SDValue Arg = getValue(I.getOperand(1));
4169 MVT Ty = Arg.getValueType();
4170 SDValue result = DAG.getNode(ISD::CTTZ, dl, Ty, Arg);
4171 setValue(&I, result);
4174 case Intrinsic::ctlz: {
4175 SDValue Arg = getValue(I.getOperand(1));
4176 MVT Ty = Arg.getValueType();
4177 SDValue result = DAG.getNode(ISD::CTLZ, dl, Ty, Arg);
4178 setValue(&I, result);
4181 case Intrinsic::ctpop: {
4182 SDValue Arg = getValue(I.getOperand(1));
4183 MVT Ty = Arg.getValueType();
4184 SDValue result = DAG.getNode(ISD::CTPOP, dl, Ty, Arg);
4185 setValue(&I, result);
4188 case Intrinsic::stacksave: {
4189 SDValue Op = getRoot();
4190 SDValue Tmp = DAG.getNode(ISD::STACKSAVE, dl,
4191 DAG.getVTList(TLI.getPointerTy(), MVT::Other), &Op, 1);
4193 DAG.setRoot(Tmp.getValue(1));
4196 case Intrinsic::stackrestore: {
4197 SDValue Tmp = getValue(I.getOperand(1));
4198 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, dl, MVT::Other, getRoot(), Tmp));
4201 case Intrinsic::stackprotector: {
4202 // Emit code into the DAG to store the stack guard onto the stack.
4203 MachineFunction &MF = DAG.getMachineFunction();
4204 MachineFrameInfo *MFI = MF.getFrameInfo();
4205 MVT PtrTy = TLI.getPointerTy();
4207 SDValue Src = getValue(I.getOperand(1)); // The guard's value.
4208 AllocaInst *Slot = cast<AllocaInst>(I.getOperand(2));
4210 int FI = FuncInfo.StaticAllocaMap[Slot];
4211 MFI->setStackProtectorIndex(FI);
4213 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
4215 // Store the stack protector onto the stack.
4216 SDValue Result = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
4217 PseudoSourceValue::getFixedStack(FI),
4219 setValue(&I, Result);
4220 DAG.setRoot(Result);
4223 case Intrinsic::var_annotation:
4224 // Discard annotate attributes
4227 case Intrinsic::init_trampoline: {
4228 const Function *F = cast<Function>(I.getOperand(2)->stripPointerCasts());
4232 Ops[1] = getValue(I.getOperand(1));
4233 Ops[2] = getValue(I.getOperand(2));
4234 Ops[3] = getValue(I.getOperand(3));
4235 Ops[4] = DAG.getSrcValue(I.getOperand(1));
4236 Ops[5] = DAG.getSrcValue(F);
4238 SDValue Tmp = DAG.getNode(ISD::TRAMPOLINE, dl,
4239 DAG.getVTList(TLI.getPointerTy(), MVT::Other),
4243 DAG.setRoot(Tmp.getValue(1));
4247 case Intrinsic::gcroot:
4249 Value *Alloca = I.getOperand(1);
4250 Constant *TypeMap = cast<Constant>(I.getOperand(2));
4252 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
4253 GFI->addStackRoot(FI->getIndex(), TypeMap);
4257 case Intrinsic::gcread:
4258 case Intrinsic::gcwrite:
4259 LLVM_UNREACHABLE("GC failed to lower gcread/gcwrite intrinsics!");
4262 case Intrinsic::flt_rounds: {
4263 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
4267 case Intrinsic::trap: {
4268 DAG.setRoot(DAG.getNode(ISD::TRAP, dl,MVT::Other, getRoot()));
4272 case Intrinsic::uadd_with_overflow:
4273 return implVisitAluOverflow(I, ISD::UADDO);
4274 case Intrinsic::sadd_with_overflow:
4275 return implVisitAluOverflow(I, ISD::SADDO);
4276 case Intrinsic::usub_with_overflow:
4277 return implVisitAluOverflow(I, ISD::USUBO);
4278 case Intrinsic::ssub_with_overflow:
4279 return implVisitAluOverflow(I, ISD::SSUBO);
4280 case Intrinsic::umul_with_overflow:
4281 return implVisitAluOverflow(I, ISD::UMULO);
4282 case Intrinsic::smul_with_overflow:
4283 return implVisitAluOverflow(I, ISD::SMULO);
4285 case Intrinsic::prefetch: {
4288 Ops[1] = getValue(I.getOperand(1));
4289 Ops[2] = getValue(I.getOperand(2));
4290 Ops[3] = getValue(I.getOperand(3));
4291 DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
4295 case Intrinsic::memory_barrier: {
4298 for (int x = 1; x < 6; ++x)
4299 Ops[x] = getValue(I.getOperand(x));
4301 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, dl, MVT::Other, &Ops[0], 6));
4304 case Intrinsic::atomic_cmp_swap: {
4305 SDValue Root = getRoot();
4307 DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, getCurDebugLoc(),
4308 getValue(I.getOperand(2)).getValueType().getSimpleVT(),
4310 getValue(I.getOperand(1)),
4311 getValue(I.getOperand(2)),
4312 getValue(I.getOperand(3)),
4315 DAG.setRoot(L.getValue(1));
4318 case Intrinsic::atomic_load_add:
4319 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD);
4320 case Intrinsic::atomic_load_sub:
4321 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB);
4322 case Intrinsic::atomic_load_or:
4323 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR);
4324 case Intrinsic::atomic_load_xor:
4325 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR);
4326 case Intrinsic::atomic_load_and:
4327 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND);
4328 case Intrinsic::atomic_load_nand:
4329 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND);
4330 case Intrinsic::atomic_load_max:
4331 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX);
4332 case Intrinsic::atomic_load_min:
4333 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN);
4334 case Intrinsic::atomic_load_umin:
4335 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN);
4336 case Intrinsic::atomic_load_umax:
4337 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX);
4338 case Intrinsic::atomic_swap:
4339 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP);
4344 void SelectionDAGLowering::LowerCallTo(CallSite CS, SDValue Callee,
4346 MachineBasicBlock *LandingPad) {
4347 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4348 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4349 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4350 unsigned BeginLabel = 0, EndLabel = 0;
4352 TargetLowering::ArgListTy Args;
4353 TargetLowering::ArgListEntry Entry;
4354 Args.reserve(CS.arg_size());
4355 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4357 SDValue ArgNode = getValue(*i);
4358 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4360 unsigned attrInd = i - CS.arg_begin() + 1;
4361 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4362 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4363 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4364 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4365 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4366 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4367 Entry.Alignment = CS.getParamAlignment(attrInd);
4368 Args.push_back(Entry);
4371 if (LandingPad && MMI) {
4372 // Insert a label before the invoke call to mark the try range. This can be
4373 // used to detect deletion of the invoke via the MachineModuleInfo.
4374 BeginLabel = MMI->NextLabelID();
4375 // Both PendingLoads and PendingExports must be flushed here;
4376 // this call might not return.
4378 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4379 getControlRoot(), BeginLabel));
4382 std::pair<SDValue,SDValue> Result =
4383 TLI.LowerCallTo(getRoot(), CS.getType(),
4384 CS.paramHasAttr(0, Attribute::SExt),
4385 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4386 CS.paramHasAttr(0, Attribute::InReg), FTy->getNumParams(),
4387 CS.getCallingConv(),
4388 IsTailCall && PerformTailCallOpt,
4389 Callee, Args, DAG, getCurDebugLoc());
4390 if (CS.getType() != Type::VoidTy)
4391 setValue(CS.getInstruction(), Result.first);
4392 DAG.setRoot(Result.second);
4394 if (LandingPad && MMI) {
4395 // Insert a label at the end of the invoke call to mark the try range. This
4396 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4397 EndLabel = MMI->NextLabelID();
4398 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getCurDebugLoc(),
4399 getRoot(), EndLabel));
4401 // Inform MachineModuleInfo of range.
4402 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
4407 void SelectionDAGLowering::visitCall(CallInst &I) {
4408 const char *RenameFn = 0;
4409 if (Function *F = I.getCalledFunction()) {
4410 if (F->isDeclaration()) {
4411 const TargetIntrinsicInfo *II = TLI.getTargetMachine().getIntrinsicInfo();
4413 if (unsigned IID = II->getIntrinsicID(F)) {
4414 RenameFn = visitIntrinsicCall(I, IID);
4419 if (unsigned IID = F->getIntrinsicID()) {
4420 RenameFn = visitIntrinsicCall(I, IID);
4426 // Check for well-known libc/libm calls. If the function is internal, it
4427 // can't be a library call.
4428 unsigned NameLen = F->getNameLen();
4429 if (!F->hasLocalLinkage() && NameLen) {
4430 const char *NameStr = F->getNameStart();
4431 if (NameStr[0] == 'c' &&
4432 ((NameLen == 8 && !strcmp(NameStr, "copysign")) ||
4433 (NameLen == 9 && !strcmp(NameStr, "copysignf")))) {
4434 if (I.getNumOperands() == 3 && // Basic sanity checks.
4435 I.getOperand(1)->getType()->isFloatingPoint() &&
4436 I.getType() == I.getOperand(1)->getType() &&
4437 I.getType() == I.getOperand(2)->getType()) {
4438 SDValue LHS = getValue(I.getOperand(1));
4439 SDValue RHS = getValue(I.getOperand(2));
4440 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, getCurDebugLoc(),
4441 LHS.getValueType(), LHS, RHS));
4444 } else if (NameStr[0] == 'f' &&
4445 ((NameLen == 4 && !strcmp(NameStr, "fabs")) ||
4446 (NameLen == 5 && !strcmp(NameStr, "fabsf")) ||
4447 (NameLen == 5 && !strcmp(NameStr, "fabsl")))) {
4448 if (I.getNumOperands() == 2 && // Basic sanity checks.
4449 I.getOperand(1)->getType()->isFloatingPoint() &&
4450 I.getType() == I.getOperand(1)->getType()) {
4451 SDValue Tmp = getValue(I.getOperand(1));
4452 setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(),
4453 Tmp.getValueType(), Tmp));
4456 } else if (NameStr[0] == 's' &&
4457 ((NameLen == 3 && !strcmp(NameStr, "sin")) ||
4458 (NameLen == 4 && !strcmp(NameStr, "sinf")) ||
4459 (NameLen == 4 && !strcmp(NameStr, "sinl")))) {
4460 if (I.getNumOperands() == 2 && // Basic sanity checks.
4461 I.getOperand(1)->getType()->isFloatingPoint() &&
4462 I.getType() == I.getOperand(1)->getType()) {
4463 SDValue Tmp = getValue(I.getOperand(1));
4464 setValue(&I, DAG.getNode(ISD::FSIN, getCurDebugLoc(),
4465 Tmp.getValueType(), Tmp));
4468 } else if (NameStr[0] == 'c' &&
4469 ((NameLen == 3 && !strcmp(NameStr, "cos")) ||
4470 (NameLen == 4 && !strcmp(NameStr, "cosf")) ||
4471 (NameLen == 4 && !strcmp(NameStr, "cosl")))) {
4472 if (I.getNumOperands() == 2 && // Basic sanity checks.
4473 I.getOperand(1)->getType()->isFloatingPoint() &&
4474 I.getType() == I.getOperand(1)->getType()) {
4475 SDValue Tmp = getValue(I.getOperand(1));
4476 setValue(&I, DAG.getNode(ISD::FCOS, getCurDebugLoc(),
4477 Tmp.getValueType(), Tmp));
4482 } else if (isa<InlineAsm>(I.getOperand(0))) {
4489 Callee = getValue(I.getOperand(0));
4491 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4493 LowerCallTo(&I, Callee, I.isTailCall());
4497 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
4498 /// this value and returns the result as a ValueVT value. This uses
4499 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4500 /// If the Flag pointer is NULL, no flag is used.
4501 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl,
4503 SDValue *Flag) const {
4504 // Assemble the legal parts into the final values.
4505 SmallVector<SDValue, 4> Values(ValueVTs.size());
4506 SmallVector<SDValue, 8> Parts;
4507 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4508 // Copy the legal parts from the registers.
4509 MVT ValueVT = ValueVTs[Value];
4510 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
4511 MVT RegisterVT = RegVTs[Value];
4513 Parts.resize(NumRegs);
4514 for (unsigned i = 0; i != NumRegs; ++i) {
4517 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
4519 P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
4520 *Flag = P.getValue(2);
4522 Chain = P.getValue(1);
4524 // If the source register was virtual and if we know something about it,
4525 // add an assert node.
4526 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
4527 RegisterVT.isInteger() && !RegisterVT.isVector()) {
4528 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
4529 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
4530 if (FLI.LiveOutRegInfo.size() > SlotNo) {
4531 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[SlotNo];
4533 unsigned RegSize = RegisterVT.getSizeInBits();
4534 unsigned NumSignBits = LOI.NumSignBits;
4535 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
4537 // FIXME: We capture more information than the dag can represent. For
4538 // now, just use the tightest assertzext/assertsext possible.
4540 MVT FromVT(MVT::Other);
4541 if (NumSignBits == RegSize)
4542 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
4543 else if (NumZeroBits >= RegSize-1)
4544 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
4545 else if (NumSignBits > RegSize-8)
4546 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
4547 else if (NumZeroBits >= RegSize-8)
4548 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
4549 else if (NumSignBits > RegSize-16)
4550 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
4551 else if (NumZeroBits >= RegSize-16)
4552 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
4553 else if (NumSignBits > RegSize-32)
4554 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
4555 else if (NumZeroBits >= RegSize-32)
4556 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
4558 if (FromVT != MVT::Other) {
4559 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
4560 RegisterVT, P, DAG.getValueType(FromVT));
4569 Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
4570 NumRegs, RegisterVT, ValueVT);
4575 return DAG.getNode(ISD::MERGE_VALUES, dl,
4576 DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
4577 &Values[0], ValueVTs.size());
4580 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
4581 /// specified value into the registers specified by this object. This uses
4582 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4583 /// If the Flag pointer is NULL, no flag is used.
4584 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
4585 SDValue &Chain, SDValue *Flag) const {
4586 // Get the list of the values's legal parts.
4587 unsigned NumRegs = Regs.size();
4588 SmallVector<SDValue, 8> Parts(NumRegs);
4589 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4590 MVT ValueVT = ValueVTs[Value];
4591 unsigned NumParts = TLI->getNumRegisters(ValueVT);
4592 MVT RegisterVT = RegVTs[Value];
4594 getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
4595 &Parts[Part], NumParts, RegisterVT);
4599 // Copy the parts into the registers.
4600 SmallVector<SDValue, 8> Chains(NumRegs);
4601 for (unsigned i = 0; i != NumRegs; ++i) {
4604 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
4606 Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
4607 *Flag = Part.getValue(1);
4609 Chains[i] = Part.getValue(0);
4612 if (NumRegs == 1 || Flag)
4613 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
4614 // flagged to it. That is the CopyToReg nodes and the user are considered
4615 // a single scheduling unit. If we create a TokenFactor and return it as
4616 // chain, then the TokenFactor is both a predecessor (operand) of the
4617 // user as well as a successor (the TF operands are flagged to the user).
4618 // c1, f1 = CopyToReg
4619 // c2, f2 = CopyToReg
4620 // c3 = TokenFactor c1, c2
4623 Chain = Chains[NumRegs-1];
4625 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
4628 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
4629 /// operand list. This adds the code marker and includes the number of
4630 /// values added into it.
4631 void RegsForValue::AddInlineAsmOperands(unsigned Code,
4632 bool HasMatching,unsigned MatchingIdx,
4634 std::vector<SDValue> &Ops) const {
4635 MVT IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
4636 assert(Regs.size() < (1 << 13) && "Too many inline asm outputs!");
4637 unsigned Flag = Code | (Regs.size() << 3);
4639 Flag |= 0x80000000 | (MatchingIdx << 16);
4640 Ops.push_back(DAG.getTargetConstant(Flag, IntPtrTy));
4641 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
4642 unsigned NumRegs = TLI->getNumRegisters(ValueVTs[Value]);
4643 MVT RegisterVT = RegVTs[Value];
4644 for (unsigned i = 0; i != NumRegs; ++i) {
4645 assert(Reg < Regs.size() && "Mismatch in # registers expected");
4646 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
4651 /// isAllocatableRegister - If the specified register is safe to allocate,
4652 /// i.e. it isn't a stack pointer or some other special register, return the
4653 /// register class for the register. Otherwise, return null.
4654 static const TargetRegisterClass *
4655 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
4656 const TargetLowering &TLI,
4657 const TargetRegisterInfo *TRI) {
4658 MVT FoundVT = MVT::Other;
4659 const TargetRegisterClass *FoundRC = 0;
4660 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
4661 E = TRI->regclass_end(); RCI != E; ++RCI) {
4662 MVT ThisVT = MVT::Other;
4664 const TargetRegisterClass *RC = *RCI;
4665 // If none of the the value types for this register class are valid, we
4666 // can't use it. For example, 64-bit reg classes on 32-bit targets.
4667 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
4669 if (TLI.isTypeLegal(*I)) {
4670 // If we have already found this register in a different register class,
4671 // choose the one with the largest VT specified. For example, on
4672 // PowerPC, we favor f64 register classes over f32.
4673 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
4680 if (ThisVT == MVT::Other) continue;
4682 // NOTE: This isn't ideal. In particular, this might allocate the
4683 // frame pointer in functions that need it (due to them not being taken
4684 // out of allocation, because a variable sized allocation hasn't been seen
4685 // yet). This is a slight code pessimization, but should still work.
4686 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
4687 E = RC->allocation_order_end(MF); I != E; ++I)
4689 // We found a matching register class. Keep looking at others in case
4690 // we find one with larger registers that this physreg is also in.
4701 /// AsmOperandInfo - This contains information for each constraint that we are
4703 class VISIBILITY_HIDDEN SDISelAsmOperandInfo :
4704 public TargetLowering::AsmOperandInfo {
4706 /// CallOperand - If this is the result output operand or a clobber
4707 /// this is null, otherwise it is the incoming operand to the CallInst.
4708 /// This gets modified as the asm is processed.
4709 SDValue CallOperand;
4711 /// AssignedRegs - If this is a register or register class operand, this
4712 /// contains the set of register corresponding to the operand.
4713 RegsForValue AssignedRegs;
4715 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
4716 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
4719 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
4720 /// busy in OutputRegs/InputRegs.
4721 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
4722 std::set<unsigned> &OutputRegs,
4723 std::set<unsigned> &InputRegs,
4724 const TargetRegisterInfo &TRI) const {
4726 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4727 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
4730 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4731 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
4735 /// getCallOperandValMVT - Return the MVT of the Value* that this operand
4736 /// corresponds to. If there is no Value* for this operand, it returns
4738 MVT getCallOperandValMVT(const TargetLowering &TLI,
4739 const TargetData *TD) const {
4740 if (CallOperandVal == 0) return MVT::Other;
4742 if (isa<BasicBlock>(CallOperandVal))
4743 return TLI.getPointerTy();
4745 const llvm::Type *OpTy = CallOperandVal->getType();
4747 // If this is an indirect operand, the operand is a pointer to the
4750 OpTy = cast<PointerType>(OpTy)->getElementType();
4752 // If OpTy is not a single value, it may be a struct/union that we
4753 // can tile with integers.
4754 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
4755 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
4764 OpTy = IntegerType::get(BitSize);
4769 return TLI.getValueType(OpTy, true);
4773 /// MarkRegAndAliases - Mark the specified register and all aliases in the
4775 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
4776 const TargetRegisterInfo &TRI) {
4777 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
4779 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
4780 for (; *Aliases; ++Aliases)
4781 Regs.insert(*Aliases);
4784 } // end llvm namespace.
4787 /// GetRegistersForValue - Assign registers (virtual or physical) for the
4788 /// specified operand. We prefer to assign virtual registers, to allow the
4789 /// register allocator handle the assignment process. However, if the asm uses
4790 /// features that we can't model on machineinstrs, we have SDISel do the
4791 /// allocation. This produces generally horrible, but correct, code.
4793 /// OpInfo describes the operand.
4794 /// Input and OutputRegs are the set of already allocated physical registers.
4796 void SelectionDAGLowering::
4797 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
4798 std::set<unsigned> &OutputRegs,
4799 std::set<unsigned> &InputRegs) {
4800 // Compute whether this value requires an input register, an output register,
4802 bool isOutReg = false;
4803 bool isInReg = false;
4804 switch (OpInfo.Type) {
4805 case InlineAsm::isOutput:
4808 // If there is an input constraint that matches this, we need to reserve
4809 // the input register so no other inputs allocate to it.
4810 isInReg = OpInfo.hasMatchingInput();
4812 case InlineAsm::isInput:
4816 case InlineAsm::isClobber:
4823 MachineFunction &MF = DAG.getMachineFunction();
4824 SmallVector<unsigned, 4> Regs;
4826 // If this is a constraint for a single physreg, or a constraint for a
4827 // register class, find it.
4828 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
4829 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
4830 OpInfo.ConstraintVT);
4832 unsigned NumRegs = 1;
4833 if (OpInfo.ConstraintVT != MVT::Other) {
4834 // If this is a FP input in an integer register (or visa versa) insert a bit
4835 // cast of the input value. More generally, handle any case where the input
4836 // value disagrees with the register class we plan to stick this in.
4837 if (OpInfo.Type == InlineAsm::isInput &&
4838 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
4839 // Try to convert to the first MVT that the reg class contains. If the
4840 // types are identical size, use a bitcast to convert (e.g. two differing
4842 MVT RegVT = *PhysReg.second->vt_begin();
4843 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
4844 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
4845 RegVT, OpInfo.CallOperand);
4846 OpInfo.ConstraintVT = RegVT;
4847 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
4848 // If the input is a FP value and we want it in FP registers, do a
4849 // bitcast to the corresponding integer type. This turns an f64 value
4850 // into i64, which can be passed with two i32 values on a 32-bit
4852 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
4853 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
4854 RegVT, OpInfo.CallOperand);
4855 OpInfo.ConstraintVT = RegVT;
4859 NumRegs = TLI.getNumRegisters(OpInfo.ConstraintVT);
4863 MVT ValueVT = OpInfo.ConstraintVT;
4865 // If this is a constraint for a specific physical register, like {r17},
4867 if (unsigned AssignedReg = PhysReg.first) {
4868 const TargetRegisterClass *RC = PhysReg.second;
4869 if (OpInfo.ConstraintVT == MVT::Other)
4870 ValueVT = *RC->vt_begin();
4872 // Get the actual register value type. This is important, because the user
4873 // may have asked for (e.g.) the AX register in i32 type. We need to
4874 // remember that AX is actually i16 to get the right extension.
4875 RegVT = *RC->vt_begin();
4877 // This is a explicit reference to a physical register.
4878 Regs.push_back(AssignedReg);
4880 // If this is an expanded reference, add the rest of the regs to Regs.
4882 TargetRegisterClass::iterator I = RC->begin();
4883 for (; *I != AssignedReg; ++I)
4884 assert(I != RC->end() && "Didn't find reg!");
4886 // Already added the first reg.
4888 for (; NumRegs; --NumRegs, ++I) {
4889 assert(I != RC->end() && "Ran out of registers to allocate!");
4893 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4894 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4895 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
4899 // Otherwise, if this was a reference to an LLVM register class, create vregs
4900 // for this reference.
4901 if (const TargetRegisterClass *RC = PhysReg.second) {
4902 RegVT = *RC->vt_begin();
4903 if (OpInfo.ConstraintVT == MVT::Other)
4906 // Create the appropriate number of virtual registers.
4907 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4908 for (; NumRegs; --NumRegs)
4909 Regs.push_back(RegInfo.createVirtualRegister(RC));
4911 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4915 // This is a reference to a register class that doesn't directly correspond
4916 // to an LLVM register class. Allocate NumRegs consecutive, available,
4917 // registers from the class.
4918 std::vector<unsigned> RegClassRegs
4919 = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
4920 OpInfo.ConstraintVT);
4922 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4923 unsigned NumAllocated = 0;
4924 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
4925 unsigned Reg = RegClassRegs[i];
4926 // See if this register is available.
4927 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
4928 (isInReg && InputRegs.count(Reg))) { // Already used.
4929 // Make sure we find consecutive registers.
4934 // Check to see if this register is allocatable (i.e. don't give out the
4936 const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, TRI);
4937 if (!RC) { // Couldn't allocate this register.
4938 // Reset NumAllocated to make sure we return consecutive registers.
4943 // Okay, this register is good, we can use it.
4946 // If we allocated enough consecutive registers, succeed.
4947 if (NumAllocated == NumRegs) {
4948 unsigned RegStart = (i-NumAllocated)+1;
4949 unsigned RegEnd = i+1;
4950 // Mark all of the allocated registers used.
4951 for (unsigned i = RegStart; i != RegEnd; ++i)
4952 Regs.push_back(RegClassRegs[i]);
4954 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, *RC->vt_begin(),
4955 OpInfo.ConstraintVT);
4956 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
4961 // Otherwise, we couldn't allocate enough registers for this.
4964 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
4965 /// processed uses a memory 'm' constraint.
4967 hasInlineAsmMemConstraint(std::vector<InlineAsm::ConstraintInfo> &CInfos,
4968 const TargetLowering &TLI) {
4969 for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
4970 InlineAsm::ConstraintInfo &CI = CInfos[i];
4971 for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
4972 TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
4973 if (CType == TargetLowering::C_Memory)
4977 // Indirect operand accesses access memory.
4985 /// visitInlineAsm - Handle a call to an InlineAsm object.
4987 void SelectionDAGLowering::visitInlineAsm(CallSite CS) {
4988 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
4990 /// ConstraintOperands - Information about all of the constraints.
4991 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
4993 std::set<unsigned> OutputRegs, InputRegs;
4995 // Do a prepass over the constraints, canonicalizing them, and building up the
4996 // ConstraintOperands list.
4997 std::vector<InlineAsm::ConstraintInfo>
4998 ConstraintInfos = IA->ParseConstraints();
5000 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
5002 SDValue Chain, Flag;
5004 // We won't need to flush pending loads if this asm doesn't touch
5005 // memory and is nonvolatile.
5006 if (hasMemory || IA->hasSideEffects())
5009 Chain = DAG.getRoot();
5011 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
5012 unsigned ResNo = 0; // ResNo - The result number of the next output.
5013 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5014 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
5015 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
5017 MVT OpVT = MVT::Other;
5019 // Compute the value type for each operand.
5020 switch (OpInfo.Type) {
5021 case InlineAsm::isOutput:
5022 // Indirect outputs just consume an argument.
5023 if (OpInfo.isIndirect) {
5024 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5028 // The return value of the call is this value. As such, there is no
5029 // corresponding argument.
5030 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
5031 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
5032 OpVT = TLI.getValueType(STy->getElementType(ResNo));
5034 assert(ResNo == 0 && "Asm only has one result!");
5035 OpVT = TLI.getValueType(CS.getType());
5039 case InlineAsm::isInput:
5040 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
5042 case InlineAsm::isClobber:
5047 // If this is an input or an indirect output, process the call argument.
5048 // BasicBlocks are labels, currently appearing only in asm's.
5049 if (OpInfo.CallOperandVal) {
5050 if (BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
5051 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
5053 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
5056 OpVT = OpInfo.getCallOperandValMVT(TLI, TD);
5059 OpInfo.ConstraintVT = OpVT;
5062 // Second pass over the constraints: compute which constraint option to use
5063 // and assign registers to constraints that want a specific physreg.
5064 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
5065 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5067 // If this is an output operand with a matching input operand, look up the
5068 // matching input. If their types mismatch, e.g. one is an integer, the
5069 // other is floating point, or their sizes are different, flag it as an
5071 if (OpInfo.hasMatchingInput()) {
5072 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
5073 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
5074 if ((OpInfo.ConstraintVT.isInteger() !=
5075 Input.ConstraintVT.isInteger()) ||
5076 (OpInfo.ConstraintVT.getSizeInBits() !=
5077 Input.ConstraintVT.getSizeInBits())) {
5078 llvm_report_error("llvm: error: Unsupported asm: input constraint"
5079 " with a matching output constraint of incompatible"
5082 Input.ConstraintVT = OpInfo.ConstraintVT;
5086 // Compute the constraint code and ConstraintType to use.
5087 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, hasMemory, &DAG);
5089 // If this is a memory input, and if the operand is not indirect, do what we
5090 // need to to provide an address for the memory input.
5091 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
5092 !OpInfo.isIndirect) {
5093 assert(OpInfo.Type == InlineAsm::isInput &&
5094 "Can only indirectify direct input operands!");
5096 // Memory operands really want the address of the value. If we don't have
5097 // an indirect input, put it in the constpool if we can, otherwise spill
5098 // it to a stack slot.
5100 // If the operand is a float, integer, or vector constant, spill to a
5101 // constant pool entry to get its address.
5102 Value *OpVal = OpInfo.CallOperandVal;
5103 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
5104 isa<ConstantVector>(OpVal)) {
5105 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
5106 TLI.getPointerTy());
5108 // Otherwise, create a stack slot and emit a store to it before the
5110 const Type *Ty = OpVal->getType();
5111 uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(Ty);
5112 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
5113 MachineFunction &MF = DAG.getMachineFunction();
5114 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align);
5115 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
5116 Chain = DAG.getStore(Chain, getCurDebugLoc(),
5117 OpInfo.CallOperand, StackSlot, NULL, 0);
5118 OpInfo.CallOperand = StackSlot;
5121 // There is no longer a Value* corresponding to this operand.
5122 OpInfo.CallOperandVal = 0;
5123 // It is now an indirect operand.
5124 OpInfo.isIndirect = true;
5127 // If this constraint is for a specific register, allocate it before
5129 if (OpInfo.ConstraintType == TargetLowering::C_Register)
5130 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5132 ConstraintInfos.clear();
5135 // Second pass - Loop over all of the operands, assigning virtual or physregs
5136 // to register class operands.
5137 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5138 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5140 // C_Register operands have already been allocated, Other/Memory don't need
5142 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
5143 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
5146 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
5147 std::vector<SDValue> AsmNodeOperands;
5148 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
5149 AsmNodeOperands.push_back(
5150 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other));
5153 // Loop over all of the inputs, copying the operand values into the
5154 // appropriate registers and processing the output regs.
5155 RegsForValue RetValRegs;
5157 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
5158 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
5160 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
5161 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
5163 switch (OpInfo.Type) {
5164 case InlineAsm::isOutput: {
5165 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
5166 OpInfo.ConstraintType != TargetLowering::C_Register) {
5167 // Memory output, or 'other' output (e.g. 'X' constraint).
5168 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
5170 // Add information to the INLINEASM node to know about this output.
5171 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5172 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5173 TLI.getPointerTy()));
5174 AsmNodeOperands.push_back(OpInfo.CallOperand);
5178 // Otherwise, this is a register or register class output.
5180 // Copy the output from the appropriate register. Find a register that
5182 if (OpInfo.AssignedRegs.Regs.empty()) {
5183 llvm_report_error("llvm: error: Couldn't allocate output reg for"
5184 " constraint '" + OpInfo.ConstraintCode + "'!");
5187 // If this is an indirect operand, store through the pointer after the
5189 if (OpInfo.isIndirect) {
5190 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
5191 OpInfo.CallOperandVal));
5193 // This is the result value of the call.
5194 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
5195 // Concatenate this output onto the outputs list.
5196 RetValRegs.append(OpInfo.AssignedRegs);
5199 // Add information to the INLINEASM node to know that this register is
5201 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
5202 6 /* EARLYCLOBBER REGDEF */ :
5206 DAG, AsmNodeOperands);
5209 case InlineAsm::isInput: {
5210 SDValue InOperandVal = OpInfo.CallOperand;
5212 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
5213 // If this is required to match an output register we have already set,
5214 // just use its register.
5215 unsigned OperandNo = OpInfo.getMatchedOperand();
5217 // Scan until we find the definition we already emitted of this operand.
5218 // When we find it, create a RegsForValue operand.
5219 unsigned CurOp = 2; // The first operand.
5220 for (; OperandNo; --OperandNo) {
5221 // Advance to the next operand.
5223 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5224 assert(((OpFlag & 7) == 2 /*REGDEF*/ ||
5225 (OpFlag & 7) == 6 /*EARLYCLOBBER REGDEF*/ ||
5226 (OpFlag & 7) == 4 /*MEM*/) &&
5227 "Skipped past definitions?");
5228 CurOp += InlineAsm::getNumOperandRegisters(OpFlag)+1;
5232 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
5233 if ((OpFlag & 7) == 2 /*REGDEF*/
5234 || (OpFlag & 7) == 6 /* EARLYCLOBBER REGDEF */) {
5235 // Add (OpFlag&0xffff)>>3 registers to MatchedRegs.
5236 if (OpInfo.isIndirect) {
5237 llvm_report_error("llvm: error: "
5238 "Don't know how to handle tied indirect "
5239 "register inputs yet!");
5241 RegsForValue MatchedRegs;
5242 MatchedRegs.TLI = &TLI;
5243 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
5244 MVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
5245 MatchedRegs.RegVTs.push_back(RegVT);
5246 MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
5247 for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
5250 push_back(RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
5252 // Use the produced MatchedRegs object to
5253 MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5255 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/,
5256 true, OpInfo.getMatchedOperand(),
5257 DAG, AsmNodeOperands);
5260 assert(((OpFlag & 7) == 4) && "Unknown matching constraint!");
5261 assert((InlineAsm::getNumOperandRegisters(OpFlag)) == 1 &&
5262 "Unexpected number of operands");
5263 // Add information to the INLINEASM node to know about this input.
5264 // See InlineAsm.h isUseOperandTiedToDef.
5265 OpFlag |= 0x80000000 | (OpInfo.getMatchedOperand() << 16);
5266 AsmNodeOperands.push_back(DAG.getTargetConstant(OpFlag,
5267 TLI.getPointerTy()));
5268 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
5273 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
5274 assert(!OpInfo.isIndirect &&
5275 "Don't know how to handle indirect other inputs yet!");
5277 std::vector<SDValue> Ops;
5278 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5279 hasMemory, Ops, DAG);
5281 llvm_report_error("llvm: error: Invalid operand for inline asm"
5282 " constraint '" + OpInfo.ConstraintCode + "'!");
5285 // Add information to the INLINEASM node to know about this input.
5286 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
5287 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5288 TLI.getPointerTy()));
5289 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5291 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5292 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5293 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5294 "Memory operands expect pointer values");
5296 // Add information to the INLINEASM node to know about this input.
5297 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5298 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5299 TLI.getPointerTy()));
5300 AsmNodeOperands.push_back(InOperandVal);
5304 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5305 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5306 "Unknown constraint type!");
5307 assert(!OpInfo.isIndirect &&
5308 "Don't know how to handle indirect register inputs yet!");
5310 // Copy the input into the appropriate registers.
5311 if (OpInfo.AssignedRegs.Regs.empty()) {
5312 llvm_report_error("llvm: error: Couldn't allocate input reg for"
5313 " constraint '"+ OpInfo.ConstraintCode +"'!");
5316 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
5319 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, false, 0,
5320 DAG, AsmNodeOperands);
5323 case InlineAsm::isClobber: {
5324 // Add the clobbered value to the operand list, so that the register
5325 // allocator is aware that the physreg got clobbered.
5326 if (!OpInfo.AssignedRegs.Regs.empty())
5327 OpInfo.AssignedRegs.AddInlineAsmOperands(6 /* EARLYCLOBBER REGDEF */,
5328 false, 0, DAG,AsmNodeOperands);
5334 // Finish up input operands.
5335 AsmNodeOperands[0] = Chain;
5336 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5338 Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
5339 DAG.getVTList(MVT::Other, MVT::Flag),
5340 &AsmNodeOperands[0], AsmNodeOperands.size());
5341 Flag = Chain.getValue(1);
5343 // If this asm returns a register value, copy the result from that register
5344 // and set it as the value of the call.
5345 if (!RetValRegs.Regs.empty()) {
5346 SDValue Val = RetValRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
5349 // FIXME: Why don't we do this for inline asms with MRVs?
5350 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5351 MVT ResultType = TLI.getValueType(CS.getType());
5353 // If any of the results of the inline asm is a vector, it may have the
5354 // wrong width/num elts. This can happen for register classes that can
5355 // contain multiple different value types. The preg or vreg allocated may
5356 // not have the same VT as was expected. Convert it to the right type
5357 // with bit_convert.
5358 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5359 Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
5362 } else if (ResultType != Val.getValueType() &&
5363 ResultType.isInteger() && Val.getValueType().isInteger()) {
5364 // If a result value was tied to an input value, the computed result may
5365 // have a wider width than the expected result. Extract the relevant
5367 Val = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), ResultType, Val);
5370 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5373 setValue(CS.getInstruction(), Val);
5374 // Don't need to use this as a chain in this case.
5375 if (!IA->hasSideEffects() && !hasMemory && IndirectStoresToEmit.empty())
5379 std::vector<std::pair<SDValue, Value*> > StoresToEmit;
5381 // Process indirect outputs, first output all of the flagged copies out of
5383 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5384 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5385 Value *Ptr = IndirectStoresToEmit[i].second;
5386 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, getCurDebugLoc(),
5388 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5392 // Emit the non-flagged stores from the physregs.
5393 SmallVector<SDValue, 8> OutChains;
5394 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
5395 OutChains.push_back(DAG.getStore(Chain, getCurDebugLoc(),
5396 StoresToEmit[i].first,
5397 getValue(StoresToEmit[i].second),
5398 StoresToEmit[i].second, 0));
5399 if (!OutChains.empty())
5400 Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
5401 &OutChains[0], OutChains.size());
5406 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
5407 SDValue Src = getValue(I.getOperand(0));
5409 // Scale up by the type size in the original i32 type width. Various
5410 // mid-level optimizers may make assumptions about demanded bits etc from the
5411 // i32-ness of the optimizer: we do not want to promote to i64 and then
5412 // multiply on 64-bit targets.
5413 // FIXME: Malloc inst should go away: PR715.
5414 uint64_t ElementSize = TD->getTypeAllocSize(I.getType()->getElementType());
5415 if (ElementSize != 1)
5416 Src = DAG.getNode(ISD::MUL, getCurDebugLoc(), Src.getValueType(),
5417 Src, DAG.getConstant(ElementSize, Src.getValueType()));
5419 MVT IntPtr = TLI.getPointerTy();
5421 if (IntPtr.bitsLT(Src.getValueType()))
5422 Src = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(), IntPtr, Src);
5423 else if (IntPtr.bitsGT(Src.getValueType()))
5424 Src = DAG.getNode(ISD::ZERO_EXTEND, getCurDebugLoc(), IntPtr, Src);
5426 TargetLowering::ArgListTy Args;
5427 TargetLowering::ArgListEntry Entry;
5429 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5430 Args.push_back(Entry);
5432 std::pair<SDValue,SDValue> Result =
5433 TLI.LowerCallTo(getRoot(), I.getType(), false, false, false, false,
5434 0, CallingConv::C, PerformTailCallOpt,
5435 DAG.getExternalSymbol("malloc", IntPtr),
5436 Args, DAG, getCurDebugLoc());
5437 setValue(&I, Result.first); // Pointers always fit in registers
5438 DAG.setRoot(Result.second);
5441 void SelectionDAGLowering::visitFree(FreeInst &I) {
5442 TargetLowering::ArgListTy Args;
5443 TargetLowering::ArgListEntry Entry;
5444 Entry.Node = getValue(I.getOperand(0));
5445 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5446 Args.push_back(Entry);
5447 MVT IntPtr = TLI.getPointerTy();
5448 std::pair<SDValue,SDValue> Result =
5449 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, false, false,
5450 0, CallingConv::C, PerformTailCallOpt,
5451 DAG.getExternalSymbol("free", IntPtr), Args, DAG,
5453 DAG.setRoot(Result.second);
5456 void SelectionDAGLowering::visitVAStart(CallInst &I) {
5457 DAG.setRoot(DAG.getNode(ISD::VASTART, getCurDebugLoc(),
5458 MVT::Other, getRoot(),
5459 getValue(I.getOperand(1)),
5460 DAG.getSrcValue(I.getOperand(1))));
5463 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
5464 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
5465 getRoot(), getValue(I.getOperand(0)),
5466 DAG.getSrcValue(I.getOperand(0)));
5468 DAG.setRoot(V.getValue(1));
5471 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
5472 DAG.setRoot(DAG.getNode(ISD::VAEND, getCurDebugLoc(),
5473 MVT::Other, getRoot(),
5474 getValue(I.getOperand(1)),
5475 DAG.getSrcValue(I.getOperand(1))));
5478 void SelectionDAGLowering::visitVACopy(CallInst &I) {
5479 DAG.setRoot(DAG.getNode(ISD::VACOPY, getCurDebugLoc(),
5480 MVT::Other, getRoot(),
5481 getValue(I.getOperand(1)),
5482 getValue(I.getOperand(2)),
5483 DAG.getSrcValue(I.getOperand(1)),
5484 DAG.getSrcValue(I.getOperand(2))));
5487 /// TargetLowering::LowerArguments - This is the default LowerArguments
5488 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
5489 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
5490 /// integrated into SDISel.
5491 void TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG,
5492 SmallVectorImpl<SDValue> &ArgValues,
5494 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
5495 SmallVector<SDValue, 3+16> Ops;
5496 Ops.push_back(DAG.getRoot());
5497 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
5498 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
5500 // Add one result value for each formal argument.
5501 SmallVector<MVT, 16> RetVals;
5503 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
5505 SmallVector<MVT, 4> ValueVTs;
5506 ComputeValueVTs(*this, I->getType(), ValueVTs);
5507 for (unsigned Value = 0, NumValues = ValueVTs.size();
5508 Value != NumValues; ++Value) {
5509 MVT VT = ValueVTs[Value];
5510 const Type *ArgTy = VT.getTypeForMVT(*DAG.getContext());
5511 ISD::ArgFlagsTy Flags;
5512 unsigned OriginalAlignment =
5513 getTargetData()->getABITypeAlignment(ArgTy);
5515 if (F.paramHasAttr(j, Attribute::ZExt))
5517 if (F.paramHasAttr(j, Attribute::SExt))
5519 if (F.paramHasAttr(j, Attribute::InReg))
5521 if (F.paramHasAttr(j, Attribute::StructRet))
5523 if (F.paramHasAttr(j, Attribute::ByVal)) {
5525 const PointerType *Ty = cast<PointerType>(I->getType());
5526 const Type *ElementTy = Ty->getElementType();
5527 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5528 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5529 // For ByVal, alignment should be passed from FE. BE will guess if
5530 // this info is not there but there are cases it cannot get right.
5531 if (F.getParamAlignment(j))
5532 FrameAlign = F.getParamAlignment(j);
5533 Flags.setByValAlign(FrameAlign);
5534 Flags.setByValSize(FrameSize);
5536 if (F.paramHasAttr(j, Attribute::Nest))
5538 Flags.setOrigAlign(OriginalAlignment);
5540 MVT RegisterVT = getRegisterType(VT);
5541 unsigned NumRegs = getNumRegisters(VT);
5542 for (unsigned i = 0; i != NumRegs; ++i) {
5543 RetVals.push_back(RegisterVT);
5544 ISD::ArgFlagsTy MyFlags = Flags;
5545 if (NumRegs > 1 && i == 0)
5547 // if it isn't first piece, alignment must be 1
5549 MyFlags.setOrigAlign(1);
5550 Ops.push_back(DAG.getArgFlags(MyFlags));
5555 RetVals.push_back(MVT::Other);
5558 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS, dl,
5559 DAG.getVTList(&RetVals[0], RetVals.size()),
5560 &Ops[0], Ops.size()).getNode();
5562 // Prelower FORMAL_ARGUMENTS. This isn't required for functionality, but
5563 // allows exposing the loads that may be part of the argument access to the
5564 // first DAGCombiner pass.
5565 SDValue TmpRes = LowerOperation(SDValue(Result, 0), DAG);
5567 // The number of results should match up, except that the lowered one may have
5568 // an extra flag result.
5569 assert((Result->getNumValues() == TmpRes.getNode()->getNumValues() ||
5570 (Result->getNumValues()+1 == TmpRes.getNode()->getNumValues() &&
5571 TmpRes.getValue(Result->getNumValues()).getValueType() == MVT::Flag))
5572 && "Lowering produced unexpected number of results!");
5574 // The FORMAL_ARGUMENTS node itself is likely no longer needed.
5575 if (Result != TmpRes.getNode() && Result->use_empty()) {
5576 HandleSDNode Dummy(DAG.getRoot());
5577 DAG.RemoveDeadNode(Result);
5580 Result = TmpRes.getNode();
5582 unsigned NumArgRegs = Result->getNumValues() - 1;
5583 DAG.setRoot(SDValue(Result, NumArgRegs));
5585 // Set up the return result vector.
5588 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
5590 SmallVector<MVT, 4> ValueVTs;
5591 ComputeValueVTs(*this, I->getType(), ValueVTs);
5592 for (unsigned Value = 0, NumValues = ValueVTs.size();
5593 Value != NumValues; ++Value) {
5594 MVT VT = ValueVTs[Value];
5595 MVT PartVT = getRegisterType(VT);
5597 unsigned NumParts = getNumRegisters(VT);
5598 SmallVector<SDValue, 4> Parts(NumParts);
5599 for (unsigned j = 0; j != NumParts; ++j)
5600 Parts[j] = SDValue(Result, i++);
5602 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5603 if (F.paramHasAttr(Idx, Attribute::SExt))
5604 AssertOp = ISD::AssertSext;
5605 else if (F.paramHasAttr(Idx, Attribute::ZExt))
5606 AssertOp = ISD::AssertZext;
5608 ArgValues.push_back(getCopyFromParts(DAG, dl, &Parts[0], NumParts,
5609 PartVT, VT, AssertOp));
5612 assert(i == NumArgRegs && "Argument register count mismatch!");
5616 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5617 /// implementation, which just inserts an ISD::CALL node, which is later custom
5618 /// lowered by the target to something concrete. FIXME: When all targets are
5619 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
5620 std::pair<SDValue, SDValue>
5621 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5622 bool RetSExt, bool RetZExt, bool isVarArg,
5623 bool isInreg, unsigned NumFixedArgs,
5624 unsigned CallingConv, bool isTailCall,
5626 ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl) {
5627 assert((!isTailCall || PerformTailCallOpt) &&
5628 "isTailCall set when tail-call optimizations are disabled!");
5630 SmallVector<SDValue, 32> Ops;
5631 Ops.push_back(Chain); // Op#0 - Chain
5632 Ops.push_back(Callee);
5634 // Handle all of the outgoing arguments.
5635 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5636 SmallVector<MVT, 4> ValueVTs;
5637 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5638 for (unsigned Value = 0, NumValues = ValueVTs.size();
5639 Value != NumValues; ++Value) {
5640 MVT VT = ValueVTs[Value];
5641 const Type *ArgTy = VT.getTypeForMVT(*DAG.getContext());
5642 SDValue Op = SDValue(Args[i].Node.getNode(),
5643 Args[i].Node.getResNo() + Value);
5644 ISD::ArgFlagsTy Flags;
5645 unsigned OriginalAlignment =
5646 getTargetData()->getABITypeAlignment(ArgTy);
5652 if (Args[i].isInReg)
5656 if (Args[i].isByVal) {
5658 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5659 const Type *ElementTy = Ty->getElementType();
5660 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5661 unsigned FrameSize = getTargetData()->getTypeAllocSize(ElementTy);
5662 // For ByVal, alignment should come from FE. BE will guess if this
5663 // info is not there but there are cases it cannot get right.
5664 if (Args[i].Alignment)
5665 FrameAlign = Args[i].Alignment;
5666 Flags.setByValAlign(FrameAlign);
5667 Flags.setByValSize(FrameSize);
5671 Flags.setOrigAlign(OriginalAlignment);
5673 MVT PartVT = getRegisterType(VT);
5674 unsigned NumParts = getNumRegisters(VT);
5675 SmallVector<SDValue, 4> Parts(NumParts);
5676 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5679 ExtendKind = ISD::SIGN_EXTEND;
5680 else if (Args[i].isZExt)
5681 ExtendKind = ISD::ZERO_EXTEND;
5683 getCopyToParts(DAG, dl, Op, &Parts[0], NumParts, PartVT, ExtendKind);
5685 for (unsigned i = 0; i != NumParts; ++i) {
5686 // if it isn't first piece, alignment must be 1
5687 ISD::ArgFlagsTy MyFlags = Flags;
5688 if (NumParts > 1 && i == 0)
5691 MyFlags.setOrigAlign(1);
5693 Ops.push_back(Parts[i]);
5694 Ops.push_back(DAG.getArgFlags(MyFlags));
5699 // Figure out the result value types. We start by making a list of
5700 // the potentially illegal return value types.
5701 SmallVector<MVT, 4> LoweredRetTys;
5702 SmallVector<MVT, 4> RetTys;
5703 ComputeValueVTs(*this, RetTy, RetTys);
5705 // Then we translate that to a list of legal types.
5706 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5708 MVT RegisterVT = getRegisterType(VT);
5709 unsigned NumRegs = getNumRegisters(VT);
5710 for (unsigned i = 0; i != NumRegs; ++i)
5711 LoweredRetTys.push_back(RegisterVT);
5714 LoweredRetTys.push_back(MVT::Other); // Always has a chain.
5716 // Create the CALL node.
5717 SDValue Res = DAG.getCall(CallingConv, dl,
5718 isVarArg, isTailCall, isInreg,
5719 DAG.getVTList(&LoweredRetTys[0],
5720 LoweredRetTys.size()),
5721 &Ops[0], Ops.size(), NumFixedArgs
5723 Chain = Res.getValue(LoweredRetTys.size() - 1);
5725 // Gather up the call result into a single value.
5726 if (RetTy != Type::VoidTy && !RetTys.empty()) {
5727 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5730 AssertOp = ISD::AssertSext;
5732 AssertOp = ISD::AssertZext;
5734 SmallVector<SDValue, 4> ReturnValues;
5736 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5738 MVT RegisterVT = getRegisterType(VT);
5739 unsigned NumRegs = getNumRegisters(VT);
5740 unsigned RegNoEnd = NumRegs + RegNo;
5741 SmallVector<SDValue, 4> Results;
5742 for (; RegNo != RegNoEnd; ++RegNo)
5743 Results.push_back(Res.getValue(RegNo));
5744 SDValue ReturnValue =
5745 getCopyFromParts(DAG, dl, &Results[0], NumRegs, RegisterVT, VT,
5747 ReturnValues.push_back(ReturnValue);
5749 Res = DAG.getNode(ISD::MERGE_VALUES, dl,
5750 DAG.getVTList(&RetTys[0], RetTys.size()),
5751 &ReturnValues[0], ReturnValues.size());
5754 return std::make_pair(Res, Chain);
5757 void TargetLowering::LowerOperationWrapper(SDNode *N,
5758 SmallVectorImpl<SDValue> &Results,
5759 SelectionDAG &DAG) {
5760 SDValue Res = LowerOperation(SDValue(N, 0), DAG);
5762 Results.push_back(Res);
5765 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5766 LLVM_UNREACHABLE("LowerOperation not implemented for this target!");
5771 void SelectionDAGLowering::CopyValueToVirtualRegister(Value *V, unsigned Reg) {
5772 SDValue Op = getValue(V);
5773 assert((Op.getOpcode() != ISD::CopyFromReg ||
5774 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5775 "Copy from a reg to the same reg!");
5776 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5778 RegsForValue RFV(TLI, Reg, V->getType());
5779 SDValue Chain = DAG.getEntryNode();
5780 RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0);
5781 PendingExports.push_back(Chain);
5784 #include "llvm/CodeGen/SelectionDAGISel.h"
5786 void SelectionDAGISel::
5787 LowerArguments(BasicBlock *LLVMBB) {
5788 // If this is the entry block, emit arguments.
5789 Function &F = *LLVMBB->getParent();
5790 SDValue OldRoot = SDL->DAG.getRoot();
5791 SmallVector<SDValue, 16> Args;
5792 TLI.LowerArguments(F, SDL->DAG, Args, SDL->getCurDebugLoc());
5795 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
5797 SmallVector<MVT, 4> ValueVTs;
5798 ComputeValueVTs(TLI, AI->getType(), ValueVTs);
5799 unsigned NumValues = ValueVTs.size();
5800 if (!AI->use_empty()) {
5801 SDL->setValue(AI, SDL->DAG.getMergeValues(&Args[a], NumValues,
5802 SDL->getCurDebugLoc()));
5803 // If this argument is live outside of the entry block, insert a copy from
5804 // whereever we got it to the vreg that other BB's will reference it as.
5805 SDL->CopyToExportRegsIfNeeded(AI);
5810 // Finally, if the target has anything special to do, allow it to do so.
5811 // FIXME: this should insert code into the DAG!
5812 EmitFunctionEntryCode(F, SDL->DAG.getMachineFunction());
5815 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
5816 /// ensure constants are generated when needed. Remember the virtual registers
5817 /// that need to be added to the Machine PHI nodes as input. We cannot just
5818 /// directly add them, because expansion might result in multiple MBB's for one
5819 /// BB. As such, the start of the BB might correspond to a different MBB than
5823 SelectionDAGISel::HandlePHINodesInSuccessorBlocks(BasicBlock *LLVMBB) {
5824 TerminatorInst *TI = LLVMBB->getTerminator();
5826 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5828 // Check successor nodes' PHI nodes that expect a constant to be available
5830 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5831 BasicBlock *SuccBB = TI->getSuccessor(succ);
5832 if (!isa<PHINode>(SuccBB->begin())) continue;
5833 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5835 // If this terminator has multiple identical successors (common for
5836 // switches), only handle each succ once.
5837 if (!SuccsHandled.insert(SuccMBB)) continue;
5839 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5842 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5843 // nodes and Machine PHI nodes, but the incoming operands have not been
5845 for (BasicBlock::iterator I = SuccBB->begin();
5846 (PN = dyn_cast<PHINode>(I)); ++I) {
5847 // Ignore dead phi's.
5848 if (PN->use_empty()) continue;
5851 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5853 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
5854 unsigned &RegOut = SDL->ConstantsOut[C];
5856 RegOut = FuncInfo->CreateRegForValue(C);
5857 SDL->CopyValueToVirtualRegister(C, RegOut);
5861 Reg = FuncInfo->ValueMap[PHIOp];
5863 assert(isa<AllocaInst>(PHIOp) &&
5864 FuncInfo->StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
5865 "Didn't codegen value into a register!??");
5866 Reg = FuncInfo->CreateRegForValue(PHIOp);
5867 SDL->CopyValueToVirtualRegister(PHIOp, Reg);
5871 // Remember that this register needs to added to the machine PHI node as
5872 // the input for this MBB.
5873 SmallVector<MVT, 4> ValueVTs;
5874 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
5875 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
5876 MVT VT = ValueVTs[vti];
5877 unsigned NumRegisters = TLI.getNumRegisters(VT);
5878 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
5879 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
5880 Reg += NumRegisters;
5884 SDL->ConstantsOut.clear();
5887 /// This is the Fast-ISel version of HandlePHINodesInSuccessorBlocks. It only
5888 /// supports legal types, and it emits MachineInstrs directly instead of
5889 /// creating SelectionDAG nodes.
5892 SelectionDAGISel::HandlePHINodesInSuccessorBlocksFast(BasicBlock *LLVMBB,
5894 TerminatorInst *TI = LLVMBB->getTerminator();
5896 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5897 unsigned OrigNumPHINodesToUpdate = SDL->PHINodesToUpdate.size();
5899 // Check successor nodes' PHI nodes that expect a constant to be available
5901 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5902 BasicBlock *SuccBB = TI->getSuccessor(succ);
5903 if (!isa<PHINode>(SuccBB->begin())) continue;
5904 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5906 // If this terminator has multiple identical successors (common for
5907 // switches), only handle each succ once.
5908 if (!SuccsHandled.insert(SuccMBB)) continue;
5910 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5913 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5914 // nodes and Machine PHI nodes, but the incoming operands have not been
5916 for (BasicBlock::iterator I = SuccBB->begin();
5917 (PN = dyn_cast<PHINode>(I)); ++I) {
5918 // Ignore dead phi's.
5919 if (PN->use_empty()) continue;
5921 // Only handle legal types. Two interesting things to note here. First,
5922 // by bailing out early, we may leave behind some dead instructions,
5923 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
5924 // own moves. Second, this check is necessary becuase FastISel doesn't
5925 // use CreateRegForValue to create registers, so it always creates
5926 // exactly one register for each non-void instruction.
5927 MVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
5928 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
5931 VT = TLI.getTypeToTransformTo(VT);
5933 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5938 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5940 unsigned Reg = F->getRegForValue(PHIOp);
5942 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5945 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));