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/Target/TargetRegisterInfo.h"
41 #include "llvm/Target/TargetData.h"
42 #include "llvm/Target/TargetFrameInfo.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include "llvm/Target/TargetLowering.h"
45 #include "llvm/Target/TargetMachine.h"
46 #include "llvm/Target/TargetOptions.h"
47 #include "llvm/Support/Compiler.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/MathExtras.h"
53 /// LimitFloatPrecision - Generate low-precision inline sequences for
54 /// some float libcalls (6, 8 or 12 bits).
55 static unsigned LimitFloatPrecision;
57 static cl::opt<unsigned, true>
58 LimitFPPrecision("limit-float-precision",
59 cl::desc("Generate low-precision inline sequences "
60 "for some float libcalls"),
61 cl::location(LimitFloatPrecision),
64 /// ComputeLinearIndex - Given an LLVM IR aggregate type and a sequence
65 /// insertvalue or extractvalue indices that identify a member, return
66 /// the linearized index of the start of the member.
68 static unsigned ComputeLinearIndex(const TargetLowering &TLI, const Type *Ty,
69 const unsigned *Indices,
70 const unsigned *IndicesEnd,
71 unsigned CurIndex = 0) {
72 // Base case: We're done.
73 if (Indices && Indices == IndicesEnd)
76 // Given a struct type, recursively traverse the elements.
77 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
78 for (StructType::element_iterator EB = STy->element_begin(),
80 EE = STy->element_end();
82 if (Indices && *Indices == unsigned(EI - EB))
83 return ComputeLinearIndex(TLI, *EI, Indices+1, IndicesEnd, CurIndex);
84 CurIndex = ComputeLinearIndex(TLI, *EI, 0, 0, CurIndex);
87 // Given an array type, recursively traverse the elements.
88 else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
89 const Type *EltTy = ATy->getElementType();
90 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i) {
91 if (Indices && *Indices == i)
92 return ComputeLinearIndex(TLI, EltTy, Indices+1, IndicesEnd, CurIndex);
93 CurIndex = ComputeLinearIndex(TLI, EltTy, 0, 0, CurIndex);
96 // We haven't found the type we're looking for, so keep searching.
100 /// ComputeValueVTs - Given an LLVM IR type, compute a sequence of
101 /// MVTs that represent all the individual underlying
102 /// non-aggregate types that comprise it.
104 /// If Offsets is non-null, it points to a vector to be filled in
105 /// with the in-memory offsets of each of the individual values.
107 static void ComputeValueVTs(const TargetLowering &TLI, const Type *Ty,
108 SmallVectorImpl<MVT> &ValueVTs,
109 SmallVectorImpl<uint64_t> *Offsets = 0,
110 uint64_t StartingOffset = 0) {
111 // Given a struct type, recursively traverse the elements.
112 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
113 const StructLayout *SL = TLI.getTargetData()->getStructLayout(STy);
114 for (StructType::element_iterator EB = STy->element_begin(),
116 EE = STy->element_end();
118 ComputeValueVTs(TLI, *EI, ValueVTs, Offsets,
119 StartingOffset + SL->getElementOffset(EI - EB));
122 // Given an array type, recursively traverse the elements.
123 if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
124 const Type *EltTy = ATy->getElementType();
125 uint64_t EltSize = TLI.getTargetData()->getABITypeSize(EltTy);
126 for (unsigned i = 0, e = ATy->getNumElements(); i != e; ++i)
127 ComputeValueVTs(TLI, EltTy, ValueVTs, Offsets,
128 StartingOffset + i * EltSize);
131 // Base case: we can get an MVT for this LLVM IR type.
132 ValueVTs.push_back(TLI.getValueType(Ty));
134 Offsets->push_back(StartingOffset);
138 /// RegsForValue - This struct represents the registers (physical or virtual)
139 /// that a particular set of values is assigned, and the type information about
140 /// the value. The most common situation is to represent one value at a time,
141 /// but struct or array values are handled element-wise as multiple values.
142 /// The splitting of aggregates is performed recursively, so that we never
143 /// have aggregate-typed registers. The values at this point do not necessarily
144 /// have legal types, so each value may require one or more registers of some
147 struct VISIBILITY_HIDDEN RegsForValue {
148 /// TLI - The TargetLowering object.
150 const TargetLowering *TLI;
152 /// ValueVTs - The value types of the values, which may not be legal, and
153 /// may need be promoted or synthesized from one or more registers.
155 SmallVector<MVT, 4> ValueVTs;
157 /// RegVTs - The value types of the registers. This is the same size as
158 /// ValueVTs and it records, for each value, what the type of the assigned
159 /// register or registers are. (Individual values are never synthesized
160 /// from more than one type of register.)
162 /// With virtual registers, the contents of RegVTs is redundant with TLI's
163 /// getRegisterType member function, however when with physical registers
164 /// it is necessary to have a separate record of the types.
166 SmallVector<MVT, 4> RegVTs;
168 /// Regs - This list holds the registers assigned to the values.
169 /// Each legal or promoted value requires one register, and each
170 /// expanded value requires multiple registers.
172 SmallVector<unsigned, 4> Regs;
174 RegsForValue() : TLI(0) {}
176 RegsForValue(const TargetLowering &tli,
177 const SmallVector<unsigned, 4> ®s,
178 MVT regvt, MVT valuevt)
179 : TLI(&tli), ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
180 RegsForValue(const TargetLowering &tli,
181 const SmallVector<unsigned, 4> ®s,
182 const SmallVector<MVT, 4> ®vts,
183 const SmallVector<MVT, 4> &valuevts)
184 : TLI(&tli), ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {}
185 RegsForValue(const TargetLowering &tli,
186 unsigned Reg, const Type *Ty) : TLI(&tli) {
187 ComputeValueVTs(tli, Ty, ValueVTs);
189 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
190 MVT ValueVT = ValueVTs[Value];
191 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
192 MVT RegisterVT = TLI->getRegisterType(ValueVT);
193 for (unsigned i = 0; i != NumRegs; ++i)
194 Regs.push_back(Reg + i);
195 RegVTs.push_back(RegisterVT);
200 /// append - Add the specified values to this one.
201 void append(const RegsForValue &RHS) {
203 ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
204 RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
205 Regs.append(RHS.Regs.begin(), RHS.Regs.end());
209 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
210 /// this value and returns the result as a ValueVTs value. This uses
211 /// Chain/Flag as the input and updates them for the output Chain/Flag.
212 /// If the Flag pointer is NULL, no flag is used.
213 SDValue getCopyFromRegs(SelectionDAG &DAG,
214 SDValue &Chain, SDValue *Flag) const;
216 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
217 /// specified value into the registers specified by this object. This uses
218 /// Chain/Flag as the input and updates them for the output Chain/Flag.
219 /// If the Flag pointer is NULL, no flag is used.
220 void getCopyToRegs(SDValue Val, SelectionDAG &DAG,
221 SDValue &Chain, SDValue *Flag) const;
223 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
224 /// operand list. This adds the code marker and includes the number of
225 /// values added into it.
226 void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
227 std::vector<SDValue> &Ops) const;
231 /// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
232 /// PHI nodes or outside of the basic block that defines it, or used by a
233 /// switch or atomic instruction, which may expand to multiple basic blocks.
234 static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
235 if (isa<PHINode>(I)) return true;
236 BasicBlock *BB = I->getParent();
237 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
238 if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
239 // FIXME: Remove switchinst special case.
240 isa<SwitchInst>(*UI))
245 /// isOnlyUsedInEntryBlock - If the specified argument is only used in the
246 /// entry block, return true. This includes arguments used by switches, since
247 /// the switch may expand into multiple basic blocks.
248 static bool isOnlyUsedInEntryBlock(Argument *A, bool EnableFastISel) {
249 // With FastISel active, we may be splitting blocks, so force creation
250 // of virtual registers for all non-dead arguments.
251 // Don't force virtual registers for byval arguments though, because
252 // fast-isel can't handle those in all cases.
253 if (EnableFastISel && !A->hasByValAttr())
254 return A->use_empty();
256 BasicBlock *Entry = A->getParent()->begin();
257 for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
258 if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
259 return false; // Use not in entry block.
263 FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli)
267 void FunctionLoweringInfo::set(Function &fn, MachineFunction &mf,
268 bool EnableFastISel) {
271 RegInfo = &MF->getRegInfo();
273 // Create a vreg for each argument register that is not dead and is used
274 // outside of the entry block for the function.
275 for (Function::arg_iterator AI = Fn->arg_begin(), E = Fn->arg_end();
277 if (!isOnlyUsedInEntryBlock(AI, EnableFastISel))
278 InitializeRegForValue(AI);
280 // Initialize the mapping of values to registers. This is only set up for
281 // instruction values that are used outside of the block that defines
283 Function::iterator BB = Fn->begin(), EB = Fn->end();
284 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
285 if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
286 if (ConstantInt *CUI = dyn_cast<ConstantInt>(AI->getArraySize())) {
287 const Type *Ty = AI->getAllocatedType();
288 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
290 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
293 TySize *= CUI->getZExtValue(); // Get total allocated size.
294 if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
295 StaticAllocaMap[AI] =
296 MF->getFrameInfo()->CreateStackObject(TySize, Align);
299 for (; BB != EB; ++BB)
300 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
301 if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
302 if (!isa<AllocaInst>(I) ||
303 !StaticAllocaMap.count(cast<AllocaInst>(I)))
304 InitializeRegForValue(I);
306 // Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
307 // also creates the initial PHI MachineInstrs, though none of the input
308 // operands are populated.
309 for (BB = Fn->begin(), EB = Fn->end(); BB != EB; ++BB) {
310 MachineBasicBlock *MBB = mf.CreateMachineBasicBlock(BB);
314 // Create Machine PHI nodes for LLVM PHI nodes, lowering them as
317 for (BasicBlock::iterator I = BB->begin();(PN = dyn_cast<PHINode>(I)); ++I){
318 if (PN->use_empty()) continue;
320 unsigned PHIReg = ValueMap[PN];
321 assert(PHIReg && "PHI node does not have an assigned virtual register!");
323 SmallVector<MVT, 4> ValueVTs;
324 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
325 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
326 MVT VT = ValueVTs[vti];
327 unsigned NumRegisters = TLI.getNumRegisters(VT);
328 const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
329 for (unsigned i = 0; i != NumRegisters; ++i)
330 BuildMI(MBB, TII->get(TargetInstrInfo::PHI), PHIReg+i);
331 PHIReg += NumRegisters;
337 unsigned FunctionLoweringInfo::MakeReg(MVT VT) {
338 return RegInfo->createVirtualRegister(TLI.getRegClassFor(VT));
341 /// CreateRegForValue - Allocate the appropriate number of virtual registers of
342 /// the correctly promoted or expanded types. Assign these registers
343 /// consecutive vreg numbers and return the first assigned number.
345 /// In the case that the given value has struct or array type, this function
346 /// will assign registers for each member or element.
348 unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
349 SmallVector<MVT, 4> ValueVTs;
350 ComputeValueVTs(TLI, V->getType(), ValueVTs);
352 unsigned FirstReg = 0;
353 for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
354 MVT ValueVT = ValueVTs[Value];
355 MVT RegisterVT = TLI.getRegisterType(ValueVT);
357 unsigned NumRegs = TLI.getNumRegisters(ValueVT);
358 for (unsigned i = 0; i != NumRegs; ++i) {
359 unsigned R = MakeReg(RegisterVT);
360 if (!FirstReg) FirstReg = R;
366 /// getCopyFromParts - Create a value that contains the specified legal parts
367 /// combined into the value they represent. If the parts combine to a type
368 /// larger then ValueVT then AssertOp can be used to specify whether the extra
369 /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
370 /// (ISD::AssertSext).
371 static SDValue getCopyFromParts(SelectionDAG &DAG,
372 const SDValue *Parts,
376 ISD::NodeType AssertOp = ISD::DELETED_NODE) {
377 assert(NumParts > 0 && "No parts to assemble!");
378 TargetLowering &TLI = DAG.getTargetLoweringInfo();
379 SDValue Val = Parts[0];
382 // Assemble the value from multiple parts.
383 if (!ValueVT.isVector()) {
384 unsigned PartBits = PartVT.getSizeInBits();
385 unsigned ValueBits = ValueVT.getSizeInBits();
387 // Assemble the power of 2 part.
388 unsigned RoundParts = NumParts & (NumParts - 1) ?
389 1 << Log2_32(NumParts) : NumParts;
390 unsigned RoundBits = PartBits * RoundParts;
391 MVT RoundVT = RoundBits == ValueBits ?
392 ValueVT : MVT::getIntegerVT(RoundBits);
395 MVT HalfVT = ValueVT.isInteger() ?
396 MVT::getIntegerVT(RoundBits/2) :
397 MVT::getFloatingPointVT(RoundBits/2);
399 if (RoundParts > 2) {
400 Lo = getCopyFromParts(DAG, Parts, RoundParts/2, PartVT, HalfVT);
401 Hi = getCopyFromParts(DAG, Parts+RoundParts/2, RoundParts/2,
404 Lo = DAG.getNode(ISD::BIT_CONVERT, HalfVT, Parts[0]);
405 Hi = DAG.getNode(ISD::BIT_CONVERT, HalfVT, Parts[1]);
407 if (TLI.isBigEndian())
409 Val = DAG.getNode(ISD::BUILD_PAIR, RoundVT, Lo, Hi);
411 if (RoundParts < NumParts) {
412 // Assemble the trailing non-power-of-2 part.
413 unsigned OddParts = NumParts - RoundParts;
414 MVT OddVT = MVT::getIntegerVT(OddParts * PartBits);
415 Hi = getCopyFromParts(DAG, Parts+RoundParts, OddParts, PartVT, OddVT);
417 // Combine the round and odd parts.
419 if (TLI.isBigEndian())
421 MVT TotalVT = MVT::getIntegerVT(NumParts * PartBits);
422 Hi = DAG.getNode(ISD::ANY_EXTEND, TotalVT, Hi);
423 Hi = DAG.getNode(ISD::SHL, TotalVT, Hi,
424 DAG.getConstant(Lo.getValueType().getSizeInBits(),
425 TLI.getShiftAmountTy()));
426 Lo = DAG.getNode(ISD::ZERO_EXTEND, TotalVT, Lo);
427 Val = DAG.getNode(ISD::OR, TotalVT, Lo, Hi);
430 // Handle a multi-element vector.
431 MVT IntermediateVT, RegisterVT;
432 unsigned NumIntermediates;
434 TLI.getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
436 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
437 NumParts = NumRegs; // Silence a compiler warning.
438 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
439 assert(RegisterVT == Parts[0].getValueType() &&
440 "Part type doesn't match part!");
442 // Assemble the parts into intermediate operands.
443 SmallVector<SDValue, 8> Ops(NumIntermediates);
444 if (NumIntermediates == NumParts) {
445 // If the register was not expanded, truncate or copy the value,
447 for (unsigned i = 0; i != NumParts; ++i)
448 Ops[i] = getCopyFromParts(DAG, &Parts[i], 1,
449 PartVT, IntermediateVT);
450 } else if (NumParts > 0) {
451 // If the intermediate type was expanded, build the intermediate operands
453 assert(NumParts % NumIntermediates == 0 &&
454 "Must expand into a divisible number of parts!");
455 unsigned Factor = NumParts / NumIntermediates;
456 for (unsigned i = 0; i != NumIntermediates; ++i)
457 Ops[i] = getCopyFromParts(DAG, &Parts[i * Factor], Factor,
458 PartVT, IntermediateVT);
461 // Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the intermediate
463 Val = DAG.getNode(IntermediateVT.isVector() ?
464 ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR,
465 ValueVT, &Ops[0], NumIntermediates);
469 // There is now one part, held in Val. Correct it to match ValueVT.
470 PartVT = Val.getValueType();
472 if (PartVT == ValueVT)
475 if (PartVT.isVector()) {
476 assert(ValueVT.isVector() && "Unknown vector conversion!");
477 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
480 if (ValueVT.isVector()) {
481 assert(ValueVT.getVectorElementType() == PartVT &&
482 ValueVT.getVectorNumElements() == 1 &&
483 "Only trivial scalar-to-vector conversions should get here!");
484 return DAG.getNode(ISD::BUILD_VECTOR, ValueVT, Val);
487 if (PartVT.isInteger() &&
488 ValueVT.isInteger()) {
489 if (ValueVT.bitsLT(PartVT)) {
490 // For a truncate, see if we have any information to
491 // indicate whether the truncated bits will always be
492 // zero or sign-extension.
493 if (AssertOp != ISD::DELETED_NODE)
494 Val = DAG.getNode(AssertOp, PartVT, Val,
495 DAG.getValueType(ValueVT));
496 return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
498 return DAG.getNode(ISD::ANY_EXTEND, ValueVT, Val);
502 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
503 if (ValueVT.bitsLT(Val.getValueType()))
504 // FP_ROUND's are always exact here.
505 return DAG.getNode(ISD::FP_ROUND, ValueVT, Val,
506 DAG.getIntPtrConstant(1));
507 return DAG.getNode(ISD::FP_EXTEND, ValueVT, Val);
510 if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
511 return DAG.getNode(ISD::BIT_CONVERT, ValueVT, Val);
513 assert(0 && "Unknown mismatch!");
517 /// getCopyToParts - Create a series of nodes that contain the specified value
518 /// split into legal parts. If the parts contain more bits than Val, then, for
519 /// integers, ExtendKind can be used to specify how to generate the extra bits.
520 static void getCopyToParts(SelectionDAG &DAG, SDValue Val,
521 SDValue *Parts, unsigned NumParts, MVT PartVT,
522 ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
523 TargetLowering &TLI = DAG.getTargetLoweringInfo();
524 MVT PtrVT = TLI.getPointerTy();
525 MVT ValueVT = Val.getValueType();
526 unsigned PartBits = PartVT.getSizeInBits();
527 assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
532 if (!ValueVT.isVector()) {
533 if (PartVT == ValueVT) {
534 assert(NumParts == 1 && "No-op copy with multiple parts!");
539 if (NumParts * PartBits > ValueVT.getSizeInBits()) {
540 // If the parts cover more bits than the value has, promote the value.
541 if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
542 assert(NumParts == 1 && "Do not know what to promote to!");
543 Val = DAG.getNode(ISD::FP_EXTEND, PartVT, Val);
544 } else if (PartVT.isInteger() && ValueVT.isInteger()) {
545 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
546 Val = DAG.getNode(ExtendKind, ValueVT, Val);
548 assert(0 && "Unknown mismatch!");
550 } else if (PartBits == ValueVT.getSizeInBits()) {
551 // Different types of the same size.
552 assert(NumParts == 1 && PartVT != ValueVT);
553 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
554 } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
555 // If the parts cover less bits than value has, truncate the value.
556 if (PartVT.isInteger() && ValueVT.isInteger()) {
557 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
558 Val = DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
560 assert(0 && "Unknown mismatch!");
564 // The value may have changed - recompute ValueVT.
565 ValueVT = Val.getValueType();
566 assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
567 "Failed to tile the value with PartVT!");
570 assert(PartVT == ValueVT && "Type conversion failed!");
575 // Expand the value into multiple parts.
576 if (NumParts & (NumParts - 1)) {
577 // The number of parts is not a power of 2. Split off and copy the tail.
578 assert(PartVT.isInteger() && ValueVT.isInteger() &&
579 "Do not know what to expand to!");
580 unsigned RoundParts = 1 << Log2_32(NumParts);
581 unsigned RoundBits = RoundParts * PartBits;
582 unsigned OddParts = NumParts - RoundParts;
583 SDValue OddVal = DAG.getNode(ISD::SRL, ValueVT, Val,
584 DAG.getConstant(RoundBits,
585 TLI.getShiftAmountTy()));
586 getCopyToParts(DAG, OddVal, Parts + RoundParts, OddParts, PartVT);
587 if (TLI.isBigEndian())
588 // The odd parts were reversed by getCopyToParts - unreverse them.
589 std::reverse(Parts + RoundParts, Parts + NumParts);
590 NumParts = RoundParts;
591 ValueVT = MVT::getIntegerVT(NumParts * PartBits);
592 Val = DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
595 // The number of parts is a power of 2. Repeatedly bisect the value using
597 Parts[0] = DAG.getNode(ISD::BIT_CONVERT,
598 MVT::getIntegerVT(ValueVT.getSizeInBits()),
600 for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
601 for (unsigned i = 0; i < NumParts; i += StepSize) {
602 unsigned ThisBits = StepSize * PartBits / 2;
603 MVT ThisVT = MVT::getIntegerVT (ThisBits);
604 SDValue &Part0 = Parts[i];
605 SDValue &Part1 = Parts[i+StepSize/2];
607 Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, ThisVT, Part0,
608 DAG.getConstant(1, PtrVT));
609 Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, ThisVT, Part0,
610 DAG.getConstant(0, PtrVT));
612 if (ThisBits == PartBits && ThisVT != PartVT) {
613 Part0 = DAG.getNode(ISD::BIT_CONVERT, PartVT, Part0);
614 Part1 = DAG.getNode(ISD::BIT_CONVERT, PartVT, Part1);
619 if (TLI.isBigEndian())
620 std::reverse(Parts, Parts + NumParts);
627 if (PartVT != ValueVT) {
628 if (PartVT.isVector()) {
629 Val = DAG.getNode(ISD::BIT_CONVERT, PartVT, Val);
631 assert(ValueVT.getVectorElementType() == PartVT &&
632 ValueVT.getVectorNumElements() == 1 &&
633 "Only trivial vector-to-scalar conversions should get here!");
634 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, PartVT, Val,
635 DAG.getConstant(0, PtrVT));
643 // Handle a multi-element vector.
644 MVT IntermediateVT, RegisterVT;
645 unsigned NumIntermediates;
647 DAG.getTargetLoweringInfo()
648 .getVectorTypeBreakdown(ValueVT, IntermediateVT, NumIntermediates,
650 unsigned NumElements = ValueVT.getVectorNumElements();
652 assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
653 NumParts = NumRegs; // Silence a compiler warning.
654 assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
656 // Split the vector into intermediate operands.
657 SmallVector<SDValue, 8> Ops(NumIntermediates);
658 for (unsigned i = 0; i != NumIntermediates; ++i)
659 if (IntermediateVT.isVector())
660 Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR,
662 DAG.getConstant(i * (NumElements / NumIntermediates),
665 Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
667 DAG.getConstant(i, PtrVT));
669 // Split the intermediate operands into legal parts.
670 if (NumParts == NumIntermediates) {
671 // If the register was not expanded, promote or copy the value,
673 for (unsigned i = 0; i != NumParts; ++i)
674 getCopyToParts(DAG, Ops[i], &Parts[i], 1, PartVT);
675 } else if (NumParts > 0) {
676 // If the intermediate type was expanded, split each the value into
678 assert(NumParts % NumIntermediates == 0 &&
679 "Must expand into a divisible number of parts!");
680 unsigned Factor = NumParts / NumIntermediates;
681 for (unsigned i = 0; i != NumIntermediates; ++i)
682 getCopyToParts(DAG, Ops[i], &Parts[i * Factor], Factor, PartVT);
687 void SelectionDAGLowering::init(GCFunctionInfo *gfi, AliasAnalysis &aa) {
690 TD = DAG.getTarget().getTargetData();
693 /// clear - Clear out the curret SelectionDAG and the associated
694 /// state and prepare this SelectionDAGLowering object to be used
695 /// for a new block. This doesn't clear out information about
696 /// additional blocks that are needed to complete switch lowering
697 /// or PHI node updating; that information is cleared out as it is
699 void SelectionDAGLowering::clear() {
701 PendingLoads.clear();
702 PendingExports.clear();
706 /// getRoot - Return the current virtual root of the Selection DAG,
707 /// flushing any PendingLoad items. This must be done before emitting
708 /// a store or any other node that may need to be ordered after any
709 /// prior load instructions.
711 SDValue SelectionDAGLowering::getRoot() {
712 if (PendingLoads.empty())
713 return DAG.getRoot();
715 if (PendingLoads.size() == 1) {
716 SDValue Root = PendingLoads[0];
718 PendingLoads.clear();
722 // Otherwise, we have to make a token factor node.
723 SDValue Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
724 &PendingLoads[0], PendingLoads.size());
725 PendingLoads.clear();
730 /// getControlRoot - Similar to getRoot, but instead of flushing all the
731 /// PendingLoad items, flush all the PendingExports items. It is necessary
732 /// to do this before emitting a terminator instruction.
734 SDValue SelectionDAGLowering::getControlRoot() {
735 SDValue Root = DAG.getRoot();
737 if (PendingExports.empty())
740 // Turn all of the CopyToReg chains into one factored node.
741 if (Root.getOpcode() != ISD::EntryToken) {
742 unsigned i = 0, e = PendingExports.size();
743 for (; i != e; ++i) {
744 assert(PendingExports[i].getNode()->getNumOperands() > 1);
745 if (PendingExports[i].getNode()->getOperand(0) == Root)
746 break; // Don't add the root if we already indirectly depend on it.
750 PendingExports.push_back(Root);
753 Root = DAG.getNode(ISD::TokenFactor, MVT::Other,
755 PendingExports.size());
756 PendingExports.clear();
761 void SelectionDAGLowering::visit(Instruction &I) {
762 visit(I.getOpcode(), I);
765 void SelectionDAGLowering::visit(unsigned Opcode, User &I) {
766 // Note: this doesn't use InstVisitor, because it has to work with
767 // ConstantExpr's in addition to instructions.
769 default: assert(0 && "Unknown instruction type encountered!");
771 // Build the switch statement using the Instruction.def file.
772 #define HANDLE_INST(NUM, OPCODE, CLASS) \
773 case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
774 #include "llvm/Instruction.def"
778 void SelectionDAGLowering::visitAdd(User &I) {
779 if (I.getType()->isFPOrFPVector())
780 visitBinary(I, ISD::FADD);
782 visitBinary(I, ISD::ADD);
785 void SelectionDAGLowering::visitMul(User &I) {
786 if (I.getType()->isFPOrFPVector())
787 visitBinary(I, ISD::FMUL);
789 visitBinary(I, ISD::MUL);
792 SDValue SelectionDAGLowering::getValue(const Value *V) {
793 SDValue &N = NodeMap[V];
794 if (N.getNode()) return N;
796 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
797 MVT VT = TLI.getValueType(V->getType(), true);
799 if (ConstantInt *CI = dyn_cast<ConstantInt>(C))
800 return N = DAG.getConstant(*CI, VT);
802 if (GlobalValue *GV = dyn_cast<GlobalValue>(C))
803 return N = DAG.getGlobalAddress(GV, VT);
805 if (isa<ConstantPointerNull>(C))
806 return N = DAG.getConstant(0, TLI.getPointerTy());
808 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C))
809 return N = DAG.getConstantFP(*CFP, VT);
811 if (isa<UndefValue>(C) && !isa<VectorType>(V->getType()) &&
812 !V->getType()->isAggregateType())
813 return N = DAG.getNode(ISD::UNDEF, VT);
815 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
816 visit(CE->getOpcode(), *CE);
817 SDValue N1 = NodeMap[V];
818 assert(N1.getNode() && "visit didn't populate the ValueMap!");
822 if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
823 SmallVector<SDValue, 4> Constants;
824 for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
826 SDNode *Val = getValue(*OI).getNode();
827 for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
828 Constants.push_back(SDValue(Val, i));
830 return DAG.getMergeValues(&Constants[0], Constants.size());
833 if (isa<StructType>(C->getType()) || isa<ArrayType>(C->getType())) {
834 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
835 "Unknown struct or array constant!");
837 SmallVector<MVT, 4> ValueVTs;
838 ComputeValueVTs(TLI, C->getType(), ValueVTs);
839 unsigned NumElts = ValueVTs.size();
841 return SDValue(); // empty struct
842 SmallVector<SDValue, 4> Constants(NumElts);
843 for (unsigned i = 0; i != NumElts; ++i) {
844 MVT EltVT = ValueVTs[i];
845 if (isa<UndefValue>(C))
846 Constants[i] = DAG.getNode(ISD::UNDEF, EltVT);
847 else if (EltVT.isFloatingPoint())
848 Constants[i] = DAG.getConstantFP(0, EltVT);
850 Constants[i] = DAG.getConstant(0, EltVT);
852 return DAG.getMergeValues(&Constants[0], NumElts);
855 const VectorType *VecTy = cast<VectorType>(V->getType());
856 unsigned NumElements = VecTy->getNumElements();
858 // Now that we know the number and type of the elements, get that number of
859 // elements into the Ops array based on what kind of constant it is.
860 SmallVector<SDValue, 16> Ops;
861 if (ConstantVector *CP = dyn_cast<ConstantVector>(C)) {
862 for (unsigned i = 0; i != NumElements; ++i)
863 Ops.push_back(getValue(CP->getOperand(i)));
865 assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
866 "Unknown vector constant!");
867 MVT EltVT = TLI.getValueType(VecTy->getElementType());
870 if (isa<UndefValue>(C))
871 Op = DAG.getNode(ISD::UNDEF, EltVT);
872 else if (EltVT.isFloatingPoint())
873 Op = DAG.getConstantFP(0, EltVT);
875 Op = DAG.getConstant(0, EltVT);
876 Ops.assign(NumElements, Op);
879 // Create a BUILD_VECTOR node.
880 return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size());
883 // If this is a static alloca, generate it as the frameindex instead of
885 if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
886 DenseMap<const AllocaInst*, int>::iterator SI =
887 FuncInfo.StaticAllocaMap.find(AI);
888 if (SI != FuncInfo.StaticAllocaMap.end())
889 return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
892 unsigned InReg = FuncInfo.ValueMap[V];
893 assert(InReg && "Value not in map!");
895 RegsForValue RFV(TLI, InReg, V->getType());
896 SDValue Chain = DAG.getEntryNode();
897 return RFV.getCopyFromRegs(DAG, Chain, NULL);
901 void SelectionDAGLowering::visitRet(ReturnInst &I) {
902 if (I.getNumOperands() == 0) {
903 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getControlRoot()));
907 SmallVector<SDValue, 8> NewValues;
908 NewValues.push_back(getControlRoot());
909 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
910 SmallVector<MVT, 4> ValueVTs;
911 ComputeValueVTs(TLI, I.getOperand(i)->getType(), ValueVTs);
912 unsigned NumValues = ValueVTs.size();
913 if (NumValues == 0) continue;
915 SDValue RetOp = getValue(I.getOperand(i));
916 for (unsigned j = 0, f = NumValues; j != f; ++j) {
917 MVT VT = ValueVTs[j];
919 // FIXME: C calling convention requires the return type to be promoted to
920 // at least 32-bit. But this is not necessary for non-C calling
922 if (VT.isInteger()) {
923 MVT MinVT = TLI.getRegisterType(MVT::i32);
924 if (VT.bitsLT(MinVT))
928 unsigned NumParts = TLI.getNumRegisters(VT);
929 MVT PartVT = TLI.getRegisterType(VT);
930 SmallVector<SDValue, 4> Parts(NumParts);
931 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
933 const Function *F = I.getParent()->getParent();
934 if (F->paramHasAttr(0, Attribute::SExt))
935 ExtendKind = ISD::SIGN_EXTEND;
936 else if (F->paramHasAttr(0, Attribute::ZExt))
937 ExtendKind = ISD::ZERO_EXTEND;
939 getCopyToParts(DAG, SDValue(RetOp.getNode(), RetOp.getResNo() + j),
940 &Parts[0], NumParts, PartVT, ExtendKind);
942 // 'inreg' on function refers to return value
943 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
944 if (F->paramHasAttr(0, Attribute::InReg))
946 for (unsigned i = 0; i < NumParts; ++i) {
947 NewValues.push_back(Parts[i]);
948 NewValues.push_back(DAG.getArgFlags(Flags));
952 DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other,
953 &NewValues[0], NewValues.size()));
956 /// ExportFromCurrentBlock - If this condition isn't known to be exported from
957 /// the current basic block, add it to ValueMap now so that we'll get a
959 void SelectionDAGLowering::ExportFromCurrentBlock(Value *V) {
960 // No need to export constants.
961 if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
964 if (FuncInfo.isExportedInst(V)) return;
966 unsigned Reg = FuncInfo.InitializeRegForValue(V);
967 CopyValueToVirtualRegister(V, Reg);
970 bool SelectionDAGLowering::isExportableFromCurrentBlock(Value *V,
971 const BasicBlock *FromBB) {
972 // The operands of the setcc have to be in this block. We don't know
973 // how to export them from some other block.
974 if (Instruction *VI = dyn_cast<Instruction>(V)) {
975 // Can export from current BB.
976 if (VI->getParent() == FromBB)
979 // Is already exported, noop.
980 return FuncInfo.isExportedInst(V);
983 // If this is an argument, we can export it if the BB is the entry block or
984 // if it is already exported.
985 if (isa<Argument>(V)) {
986 if (FromBB == &FromBB->getParent()->getEntryBlock())
989 // Otherwise, can only export this if it is already exported.
990 return FuncInfo.isExportedInst(V);
993 // Otherwise, constants can always be exported.
997 static bool InBlock(const Value *V, const BasicBlock *BB) {
998 if (const Instruction *I = dyn_cast<Instruction>(V))
999 return I->getParent() == BB;
1003 /// getFCmpCondCode - Return the ISD condition code corresponding to
1004 /// the given LLVM IR floating-point condition code. This includes
1005 /// consideration of global floating-point math flags.
1007 static ISD::CondCode getFCmpCondCode(FCmpInst::Predicate Pred) {
1008 ISD::CondCode FPC, FOC;
1010 case FCmpInst::FCMP_FALSE: FOC = FPC = ISD::SETFALSE; break;
1011 case FCmpInst::FCMP_OEQ: FOC = ISD::SETEQ; FPC = ISD::SETOEQ; break;
1012 case FCmpInst::FCMP_OGT: FOC = ISD::SETGT; FPC = ISD::SETOGT; break;
1013 case FCmpInst::FCMP_OGE: FOC = ISD::SETGE; FPC = ISD::SETOGE; break;
1014 case FCmpInst::FCMP_OLT: FOC = ISD::SETLT; FPC = ISD::SETOLT; break;
1015 case FCmpInst::FCMP_OLE: FOC = ISD::SETLE; FPC = ISD::SETOLE; break;
1016 case FCmpInst::FCMP_ONE: FOC = ISD::SETNE; FPC = ISD::SETONE; break;
1017 case FCmpInst::FCMP_ORD: FOC = FPC = ISD::SETO; break;
1018 case FCmpInst::FCMP_UNO: FOC = FPC = ISD::SETUO; break;
1019 case FCmpInst::FCMP_UEQ: FOC = ISD::SETEQ; FPC = ISD::SETUEQ; break;
1020 case FCmpInst::FCMP_UGT: FOC = ISD::SETGT; FPC = ISD::SETUGT; break;
1021 case FCmpInst::FCMP_UGE: FOC = ISD::SETGE; FPC = ISD::SETUGE; break;
1022 case FCmpInst::FCMP_ULT: FOC = ISD::SETLT; FPC = ISD::SETULT; break;
1023 case FCmpInst::FCMP_ULE: FOC = ISD::SETLE; FPC = ISD::SETULE; break;
1024 case FCmpInst::FCMP_UNE: FOC = ISD::SETNE; FPC = ISD::SETUNE; break;
1025 case FCmpInst::FCMP_TRUE: FOC = FPC = ISD::SETTRUE; break;
1027 assert(0 && "Invalid FCmp predicate opcode!");
1028 FOC = FPC = ISD::SETFALSE;
1031 if (FiniteOnlyFPMath())
1037 /// getICmpCondCode - Return the ISD condition code corresponding to
1038 /// the given LLVM IR integer condition code.
1040 static ISD::CondCode getICmpCondCode(ICmpInst::Predicate Pred) {
1042 case ICmpInst::ICMP_EQ: return ISD::SETEQ;
1043 case ICmpInst::ICMP_NE: return ISD::SETNE;
1044 case ICmpInst::ICMP_SLE: return ISD::SETLE;
1045 case ICmpInst::ICMP_ULE: return ISD::SETULE;
1046 case ICmpInst::ICMP_SGE: return ISD::SETGE;
1047 case ICmpInst::ICMP_UGE: return ISD::SETUGE;
1048 case ICmpInst::ICMP_SLT: return ISD::SETLT;
1049 case ICmpInst::ICMP_ULT: return ISD::SETULT;
1050 case ICmpInst::ICMP_SGT: return ISD::SETGT;
1051 case ICmpInst::ICMP_UGT: return ISD::SETUGT;
1053 assert(0 && "Invalid ICmp predicate opcode!");
1058 /// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
1059 /// This function emits a branch and is used at the leaves of an OR or an
1060 /// AND operator tree.
1063 SelectionDAGLowering::EmitBranchForMergedCondition(Value *Cond,
1064 MachineBasicBlock *TBB,
1065 MachineBasicBlock *FBB,
1066 MachineBasicBlock *CurBB) {
1067 const BasicBlock *BB = CurBB->getBasicBlock();
1069 // If the leaf of the tree is a comparison, merge the condition into
1071 if (CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
1072 // The operands of the cmp have to be in this block. We don't know
1073 // how to export them from some other block. If this is the first block
1074 // of the sequence, no exporting is needed.
1075 if (CurBB == CurMBB ||
1076 (isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
1077 isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
1078 ISD::CondCode Condition;
1079 if (ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
1080 Condition = getICmpCondCode(IC->getPredicate());
1081 } else if (FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
1082 Condition = getFCmpCondCode(FC->getPredicate());
1084 Condition = ISD::SETEQ; // silence warning.
1085 assert(0 && "Unknown compare instruction");
1088 CaseBlock CB(Condition, BOp->getOperand(0),
1089 BOp->getOperand(1), NULL, TBB, FBB, CurBB);
1090 SwitchCases.push_back(CB);
1095 // Create a CaseBlock record representing this branch.
1096 CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(),
1097 NULL, TBB, FBB, CurBB);
1098 SwitchCases.push_back(CB);
1101 /// FindMergedConditions - If Cond is an expression like
1102 void SelectionDAGLowering::FindMergedConditions(Value *Cond,
1103 MachineBasicBlock *TBB,
1104 MachineBasicBlock *FBB,
1105 MachineBasicBlock *CurBB,
1107 // If this node is not part of the or/and tree, emit it as a branch.
1108 Instruction *BOp = dyn_cast<Instruction>(Cond);
1109 if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
1110 (unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
1111 BOp->getParent() != CurBB->getBasicBlock() ||
1112 !InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
1113 !InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
1114 EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB);
1118 // Create TmpBB after CurBB.
1119 MachineFunction::iterator BBI = CurBB;
1120 MachineFunction &MF = DAG.getMachineFunction();
1121 MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
1122 CurBB->getParent()->insert(++BBI, TmpBB);
1124 if (Opc == Instruction::Or) {
1125 // Codegen X | Y as:
1133 // Emit the LHS condition.
1134 FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, Opc);
1136 // Emit the RHS condition into TmpBB.
1137 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1139 assert(Opc == Instruction::And && "Unknown merge op!");
1140 // Codegen X & Y as:
1147 // This requires creation of TmpBB after CurBB.
1149 // Emit the LHS condition.
1150 FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, Opc);
1152 // Emit the RHS condition into TmpBB.
1153 FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, Opc);
1157 /// If the set of cases should be emitted as a series of branches, return true.
1158 /// If we should emit this as a bunch of and/or'd together conditions, return
1161 SelectionDAGLowering::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
1162 if (Cases.size() != 2) return true;
1164 // If this is two comparisons of the same values or'd or and'd together, they
1165 // will get folded into a single comparison, so don't emit two blocks.
1166 if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
1167 Cases[0].CmpRHS == Cases[1].CmpRHS) ||
1168 (Cases[0].CmpRHS == Cases[1].CmpLHS &&
1169 Cases[0].CmpLHS == Cases[1].CmpRHS)) {
1176 void SelectionDAGLowering::visitBr(BranchInst &I) {
1177 // Update machine-CFG edges.
1178 MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
1180 // Figure out which block is immediately after the current one.
1181 MachineBasicBlock *NextBlock = 0;
1182 MachineFunction::iterator BBI = CurMBB;
1183 if (++BBI != CurMBB->getParent()->end())
1186 if (I.isUnconditional()) {
1187 // Update machine-CFG edges.
1188 CurMBB->addSuccessor(Succ0MBB);
1190 // If this is not a fall-through branch, emit the branch.
1191 if (Succ0MBB != NextBlock)
1192 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getControlRoot(),
1193 DAG.getBasicBlock(Succ0MBB)));
1197 // If this condition is one of the special cases we handle, do special stuff
1199 Value *CondVal = I.getCondition();
1200 MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
1202 // If this is a series of conditions that are or'd or and'd together, emit
1203 // this as a sequence of branches instead of setcc's with and/or operations.
1204 // For example, instead of something like:
1217 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
1218 if (BOp->hasOneUse() &&
1219 (BOp->getOpcode() == Instruction::And ||
1220 BOp->getOpcode() == Instruction::Or)) {
1221 FindMergedConditions(BOp, Succ0MBB, Succ1MBB, CurMBB, BOp->getOpcode());
1222 // If the compares in later blocks need to use values not currently
1223 // exported from this block, export them now. This block should always
1224 // be the first entry.
1225 assert(SwitchCases[0].ThisBB == CurMBB && "Unexpected lowering!");
1227 // Allow some cases to be rejected.
1228 if (ShouldEmitAsBranches(SwitchCases)) {
1229 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
1230 ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
1231 ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
1234 // Emit the branch for this block.
1235 visitSwitchCase(SwitchCases[0]);
1236 SwitchCases.erase(SwitchCases.begin());
1240 // Okay, we decided not to do this, remove any inserted MBB's and clear
1242 for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
1243 CurMBB->getParent()->erase(SwitchCases[i].ThisBB);
1245 SwitchCases.clear();
1249 // Create a CaseBlock record representing this branch.
1250 CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(),
1251 NULL, Succ0MBB, Succ1MBB, CurMBB);
1252 // Use visitSwitchCase to actually insert the fast branch sequence for this
1254 visitSwitchCase(CB);
1257 /// visitSwitchCase - Emits the necessary code to represent a single node in
1258 /// the binary search tree resulting from lowering a switch instruction.
1259 void SelectionDAGLowering::visitSwitchCase(CaseBlock &CB) {
1261 SDValue CondLHS = getValue(CB.CmpLHS);
1263 // Build the setcc now.
1264 if (CB.CmpMHS == NULL) {
1265 // Fold "(X == true)" to X and "(X == false)" to !X to
1266 // handle common cases produced by branch lowering.
1267 if (CB.CmpRHS == ConstantInt::getTrue() && CB.CC == ISD::SETEQ)
1269 else if (CB.CmpRHS == ConstantInt::getFalse() && CB.CC == ISD::SETEQ) {
1270 SDValue True = DAG.getConstant(1, CondLHS.getValueType());
1271 Cond = DAG.getNode(ISD::XOR, CondLHS.getValueType(), CondLHS, True);
1273 Cond = DAG.getSetCC(MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
1275 assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
1277 uint64_t Low = cast<ConstantInt>(CB.CmpLHS)->getSExtValue();
1278 uint64_t High = cast<ConstantInt>(CB.CmpRHS)->getSExtValue();
1280 SDValue CmpOp = getValue(CB.CmpMHS);
1281 MVT VT = CmpOp.getValueType();
1283 if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
1284 Cond = DAG.getSetCC(MVT::i1, CmpOp, DAG.getConstant(High, VT), ISD::SETLE);
1286 SDValue SUB = DAG.getNode(ISD::SUB, VT, CmpOp, DAG.getConstant(Low, VT));
1287 Cond = DAG.getSetCC(MVT::i1, SUB,
1288 DAG.getConstant(High-Low, VT), ISD::SETULE);
1292 // Update successor info
1293 CurMBB->addSuccessor(CB.TrueBB);
1294 CurMBB->addSuccessor(CB.FalseBB);
1296 // Set NextBlock to be the MBB immediately after the current one, if any.
1297 // This is used to avoid emitting unnecessary branches to the next block.
1298 MachineBasicBlock *NextBlock = 0;
1299 MachineFunction::iterator BBI = CurMBB;
1300 if (++BBI != CurMBB->getParent()->end())
1303 // If the lhs block is the next block, invert the condition so that we can
1304 // fall through to the lhs instead of the rhs block.
1305 if (CB.TrueBB == NextBlock) {
1306 std::swap(CB.TrueBB, CB.FalseBB);
1307 SDValue True = DAG.getConstant(1, Cond.getValueType());
1308 Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
1310 SDValue BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getControlRoot(), Cond,
1311 DAG.getBasicBlock(CB.TrueBB));
1313 // If the branch was constant folded, fix up the CFG.
1314 if (BrCond.getOpcode() == ISD::BR) {
1315 CurMBB->removeSuccessor(CB.FalseBB);
1316 DAG.setRoot(BrCond);
1318 // Otherwise, go ahead and insert the false branch.
1319 if (BrCond == getControlRoot())
1320 CurMBB->removeSuccessor(CB.TrueBB);
1322 if (CB.FalseBB == NextBlock)
1323 DAG.setRoot(BrCond);
1325 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1326 DAG.getBasicBlock(CB.FalseBB)));
1330 /// visitJumpTable - Emit JumpTable node in the current MBB
1331 void SelectionDAGLowering::visitJumpTable(JumpTable &JT) {
1332 // Emit the code for the jump table
1333 assert(JT.Reg != -1U && "Should lower JT Header first!");
1334 MVT PTy = TLI.getPointerTy();
1335 SDValue Index = DAG.getCopyFromReg(getControlRoot(), JT.Reg, PTy);
1336 SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
1337 DAG.setRoot(DAG.getNode(ISD::BR_JT, MVT::Other, Index.getValue(1),
1342 /// visitJumpTableHeader - This function emits necessary code to produce index
1343 /// in the JumpTable from switch case.
1344 void SelectionDAGLowering::visitJumpTableHeader(JumpTable &JT,
1345 JumpTableHeader &JTH) {
1346 // Subtract the lowest switch case value from the value being switched on
1347 // and conditional branch to default mbb if the result is greater than the
1348 // difference between smallest and largest cases.
1349 SDValue SwitchOp = getValue(JTH.SValue);
1350 MVT VT = SwitchOp.getValueType();
1351 SDValue SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1352 DAG.getConstant(JTH.First, VT));
1354 // The SDNode we just created, which holds the value being switched on
1355 // minus the the smallest case value, needs to be copied to a virtual
1356 // register so it can be used as an index into the jump table in a
1357 // subsequent basic block. This value may be smaller or larger than the
1358 // target's pointer type, and therefore require extension or truncating.
1359 if (VT.bitsGT(TLI.getPointerTy()))
1360 SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
1362 SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
1364 unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
1365 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), JumpTableReg, SwitchOp);
1366 JT.Reg = JumpTableReg;
1368 // Emit the range check for the jump table, and branch to the default
1369 // block for the switch statement if the value being switched on exceeds
1370 // the largest case in the switch.
1371 SDValue CMP = DAG.getSetCC(TLI.getSetCCResultType(SUB), SUB,
1372 DAG.getConstant(JTH.Last-JTH.First,VT),
1375 // Set NextBlock to be the MBB immediately after the current one, if any.
1376 // This is used to avoid emitting unnecessary branches to the next block.
1377 MachineBasicBlock *NextBlock = 0;
1378 MachineFunction::iterator BBI = CurMBB;
1379 if (++BBI != CurMBB->getParent()->end())
1382 SDValue BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
1383 DAG.getBasicBlock(JT.Default));
1385 if (JT.MBB == NextBlock)
1386 DAG.setRoot(BrCond);
1388 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
1389 DAG.getBasicBlock(JT.MBB)));
1394 /// visitBitTestHeader - This function emits necessary code to produce value
1395 /// suitable for "bit tests"
1396 void SelectionDAGLowering::visitBitTestHeader(BitTestBlock &B) {
1397 // Subtract the minimum value
1398 SDValue SwitchOp = getValue(B.SValue);
1399 MVT VT = SwitchOp.getValueType();
1400 SDValue SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
1401 DAG.getConstant(B.First, VT));
1404 SDValue RangeCmp = DAG.getSetCC(TLI.getSetCCResultType(SUB), SUB,
1405 DAG.getConstant(B.Range, VT),
1409 if (VT.bitsGT(TLI.getShiftAmountTy()))
1410 ShiftOp = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), SUB);
1412 ShiftOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getShiftAmountTy(), SUB);
1414 // Make desired shift
1415 SDValue SwitchVal = DAG.getNode(ISD::SHL, TLI.getPointerTy(),
1416 DAG.getConstant(1, TLI.getPointerTy()),
1419 unsigned SwitchReg = FuncInfo.MakeReg(TLI.getPointerTy());
1420 SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), SwitchReg, SwitchVal);
1423 // Set NextBlock to be the MBB immediately after the current one, if any.
1424 // This is used to avoid emitting unnecessary branches to the next block.
1425 MachineBasicBlock *NextBlock = 0;
1426 MachineFunction::iterator BBI = CurMBB;
1427 if (++BBI != CurMBB->getParent()->end())
1430 MachineBasicBlock* MBB = B.Cases[0].ThisBB;
1432 CurMBB->addSuccessor(B.Default);
1433 CurMBB->addSuccessor(MBB);
1435 SDValue BrRange = DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, RangeCmp,
1436 DAG.getBasicBlock(B.Default));
1438 if (MBB == NextBlock)
1439 DAG.setRoot(BrRange);
1441 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, CopyTo,
1442 DAG.getBasicBlock(MBB)));
1447 /// visitBitTestCase - this function produces one "bit test"
1448 void SelectionDAGLowering::visitBitTestCase(MachineBasicBlock* NextMBB,
1451 // Emit bit tests and jumps
1452 SDValue SwitchVal = DAG.getCopyFromReg(getControlRoot(), Reg,
1453 TLI.getPointerTy());
1455 SDValue AndOp = DAG.getNode(ISD::AND, TLI.getPointerTy(), SwitchVal,
1456 DAG.getConstant(B.Mask, TLI.getPointerTy()));
1457 SDValue AndCmp = DAG.getSetCC(TLI.getSetCCResultType(AndOp), AndOp,
1458 DAG.getConstant(0, TLI.getPointerTy()),
1461 CurMBB->addSuccessor(B.TargetBB);
1462 CurMBB->addSuccessor(NextMBB);
1464 SDValue BrAnd = DAG.getNode(ISD::BRCOND, MVT::Other, getControlRoot(),
1465 AndCmp, DAG.getBasicBlock(B.TargetBB));
1467 // Set NextBlock to be the MBB immediately after the current one, if any.
1468 // This is used to avoid emitting unnecessary branches to the next block.
1469 MachineBasicBlock *NextBlock = 0;
1470 MachineFunction::iterator BBI = CurMBB;
1471 if (++BBI != CurMBB->getParent()->end())
1474 if (NextMBB == NextBlock)
1477 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrAnd,
1478 DAG.getBasicBlock(NextMBB)));
1483 void SelectionDAGLowering::visitInvoke(InvokeInst &I) {
1484 // Retrieve successors.
1485 MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
1486 MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
1488 if (isa<InlineAsm>(I.getCalledValue()))
1491 LowerCallTo(&I, getValue(I.getOperand(0)), false, LandingPad);
1493 // If the value of the invoke is used outside of its defining block, make it
1494 // available as a virtual register.
1495 if (!I.use_empty()) {
1496 DenseMap<const Value*, unsigned>::iterator VMI = FuncInfo.ValueMap.find(&I);
1497 if (VMI != FuncInfo.ValueMap.end())
1498 CopyValueToVirtualRegister(&I, VMI->second);
1501 // Update successor info
1502 CurMBB->addSuccessor(Return);
1503 CurMBB->addSuccessor(LandingPad);
1505 // Drop into normal successor.
1506 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getControlRoot(),
1507 DAG.getBasicBlock(Return)));
1510 void SelectionDAGLowering::visitUnwind(UnwindInst &I) {
1513 /// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
1514 /// small case ranges).
1515 bool SelectionDAGLowering::handleSmallSwitchRange(CaseRec& CR,
1516 CaseRecVector& WorkList,
1518 MachineBasicBlock* Default) {
1519 Case& BackCase = *(CR.Range.second-1);
1521 // Size is the number of Cases represented by this range.
1522 unsigned Size = CR.Range.second - CR.Range.first;
1526 // Get the MachineFunction which holds the current MBB. This is used when
1527 // inserting any additional MBBs necessary to represent the switch.
1528 MachineFunction *CurMF = CurMBB->getParent();
1530 // Figure out which block is immediately after the current one.
1531 MachineBasicBlock *NextBlock = 0;
1532 MachineFunction::iterator BBI = CR.CaseBB;
1534 if (++BBI != CurMBB->getParent()->end())
1537 // TODO: If any two of the cases has the same destination, and if one value
1538 // is the same as the other, but has one bit unset that the other has set,
1539 // use bit manipulation to do two compares at once. For example:
1540 // "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
1542 // Rearrange the case blocks so that the last one falls through if possible.
1543 if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
1544 // The last case block won't fall through into 'NextBlock' if we emit the
1545 // branches in this order. See if rearranging a case value would help.
1546 for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
1547 if (I->BB == NextBlock) {
1548 std::swap(*I, BackCase);
1554 // Create a CaseBlock record representing a conditional branch to
1555 // the Case's target mbb if the value being switched on SV is equal
1557 MachineBasicBlock *CurBlock = CR.CaseBB;
1558 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
1559 MachineBasicBlock *FallThrough;
1561 FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
1562 CurMF->insert(BBI, FallThrough);
1564 // If the last case doesn't match, go to the default block.
1565 FallThrough = Default;
1568 Value *RHS, *LHS, *MHS;
1570 if (I->High == I->Low) {
1571 // This is just small small case range :) containing exactly 1 case
1573 LHS = SV; RHS = I->High; MHS = NULL;
1576 LHS = I->Low; MHS = SV; RHS = I->High;
1578 CaseBlock CB(CC, LHS, RHS, MHS, I->BB, FallThrough, CurBlock);
1580 // If emitting the first comparison, just call visitSwitchCase to emit the
1581 // code into the current block. Otherwise, push the CaseBlock onto the
1582 // vector to be later processed by SDISel, and insert the node's MBB
1583 // before the next MBB.
1584 if (CurBlock == CurMBB)
1585 visitSwitchCase(CB);
1587 SwitchCases.push_back(CB);
1589 CurBlock = FallThrough;
1595 static inline bool areJTsAllowed(const TargetLowering &TLI) {
1596 return !DisableJumpTables &&
1597 (TLI.isOperationLegal(ISD::BR_JT, MVT::Other) ||
1598 TLI.isOperationLegal(ISD::BRIND, MVT::Other));
1601 /// handleJTSwitchCase - Emit jumptable for current switch case range
1602 bool SelectionDAGLowering::handleJTSwitchCase(CaseRec& CR,
1603 CaseRecVector& WorkList,
1605 MachineBasicBlock* Default) {
1606 Case& FrontCase = *CR.Range.first;
1607 Case& BackCase = *(CR.Range.second-1);
1609 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1610 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1613 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1617 if (!areJTsAllowed(TLI) || TSize <= 3)
1620 double Density = (double)TSize / (double)((Last - First) + 1ULL);
1624 DOUT << "Lowering jump table\n"
1625 << "First entry: " << First << ". Last entry: " << Last << "\n"
1626 << "Size: " << TSize << ". Density: " << Density << "\n\n";
1628 // Get the MachineFunction which holds the current MBB. This is used when
1629 // inserting any additional MBBs necessary to represent the switch.
1630 MachineFunction *CurMF = CurMBB->getParent();
1632 // Figure out which block is immediately after the current one.
1633 MachineBasicBlock *NextBlock = 0;
1634 MachineFunction::iterator BBI = CR.CaseBB;
1636 if (++BBI != CurMBB->getParent()->end())
1639 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1641 // Create a new basic block to hold the code for loading the address
1642 // of the jump table, and jumping to it. Update successor information;
1643 // we will either branch to the default case for the switch, or the jump
1645 MachineBasicBlock *JumpTableBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1646 CurMF->insert(BBI, JumpTableBB);
1647 CR.CaseBB->addSuccessor(Default);
1648 CR.CaseBB->addSuccessor(JumpTableBB);
1650 // Build a vector of destination BBs, corresponding to each target
1651 // of the jump table. If the value of the jump table slot corresponds to
1652 // a case statement, push the case's BB onto the vector, otherwise, push
1654 std::vector<MachineBasicBlock*> DestBBs;
1655 int64_t TEI = First;
1656 for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
1657 int64_t Low = cast<ConstantInt>(I->Low)->getSExtValue();
1658 int64_t High = cast<ConstantInt>(I->High)->getSExtValue();
1660 if ((Low <= TEI) && (TEI <= High)) {
1661 DestBBs.push_back(I->BB);
1665 DestBBs.push_back(Default);
1669 // Update successor info. Add one edge to each unique successor.
1670 BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
1671 for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
1672 E = DestBBs.end(); I != E; ++I) {
1673 if (!SuccsHandled[(*I)->getNumber()]) {
1674 SuccsHandled[(*I)->getNumber()] = true;
1675 JumpTableBB->addSuccessor(*I);
1679 // Create a jump table index for this jump table, or return an existing
1681 unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
1683 // Set the jump table information so that we can codegen it as a second
1684 // MachineBasicBlock
1685 JumpTable JT(-1U, JTI, JumpTableBB, Default);
1686 JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == CurMBB));
1687 if (CR.CaseBB == CurMBB)
1688 visitJumpTableHeader(JT, JTH);
1690 JTCases.push_back(JumpTableBlock(JTH, JT));
1695 /// handleBTSplitSwitchCase - emit comparison and split binary search tree into
1697 bool SelectionDAGLowering::handleBTSplitSwitchCase(CaseRec& CR,
1698 CaseRecVector& WorkList,
1700 MachineBasicBlock* Default) {
1701 // Get the MachineFunction which holds the current MBB. This is used when
1702 // inserting any additional MBBs necessary to represent the switch.
1703 MachineFunction *CurMF = CurMBB->getParent();
1705 // Figure out which block is immediately after the current one.
1706 MachineBasicBlock *NextBlock = 0;
1707 MachineFunction::iterator BBI = CR.CaseBB;
1709 if (++BBI != CurMBB->getParent()->end())
1712 Case& FrontCase = *CR.Range.first;
1713 Case& BackCase = *(CR.Range.second-1);
1714 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1716 // Size is the number of Cases represented by this range.
1717 unsigned Size = CR.Range.second - CR.Range.first;
1719 int64_t First = cast<ConstantInt>(FrontCase.Low)->getSExtValue();
1720 int64_t Last = cast<ConstantInt>(BackCase.High)->getSExtValue();
1722 CaseItr Pivot = CR.Range.first + Size/2;
1724 // Select optimal pivot, maximizing sum density of LHS and RHS. This will
1725 // (heuristically) allow us to emit JumpTable's later.
1727 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1731 uint64_t LSize = FrontCase.size();
1732 uint64_t RSize = TSize-LSize;
1733 DOUT << "Selecting best pivot: \n"
1734 << "First: " << First << ", Last: " << Last <<"\n"
1735 << "LSize: " << LSize << ", RSize: " << RSize << "\n";
1736 for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
1738 int64_t LEnd = cast<ConstantInt>(I->High)->getSExtValue();
1739 int64_t RBegin = cast<ConstantInt>(J->Low)->getSExtValue();
1740 assert((RBegin-LEnd>=1) && "Invalid case distance");
1741 double LDensity = (double)LSize / (double)((LEnd - First) + 1ULL);
1742 double RDensity = (double)RSize / (double)((Last - RBegin) + 1ULL);
1743 double Metric = Log2_64(RBegin-LEnd)*(LDensity+RDensity);
1744 // Should always split in some non-trivial place
1746 << "LEnd: " << LEnd << ", RBegin: " << RBegin << "\n"
1747 << "LDensity: " << LDensity << ", RDensity: " << RDensity << "\n"
1748 << "Metric: " << Metric << "\n";
1749 if (FMetric < Metric) {
1752 DOUT << "Current metric set to: " << FMetric << "\n";
1758 if (areJTsAllowed(TLI)) {
1759 // If our case is dense we *really* should handle it earlier!
1760 assert((FMetric > 0) && "Should handle dense range earlier!");
1762 Pivot = CR.Range.first + Size/2;
1765 CaseRange LHSR(CR.Range.first, Pivot);
1766 CaseRange RHSR(Pivot, CR.Range.second);
1767 Constant *C = Pivot->Low;
1768 MachineBasicBlock *FalseBB = 0, *TrueBB = 0;
1770 // We know that we branch to the LHS if the Value being switched on is
1771 // less than the Pivot value, C. We use this to optimize our binary
1772 // tree a bit, by recognizing that if SV is greater than or equal to the
1773 // LHS's Case Value, and that Case Value is exactly one less than the
1774 // Pivot's Value, then we can branch directly to the LHS's Target,
1775 // rather than creating a leaf node for it.
1776 if ((LHSR.second - LHSR.first) == 1 &&
1777 LHSR.first->High == CR.GE &&
1778 cast<ConstantInt>(C)->getSExtValue() ==
1779 (cast<ConstantInt>(CR.GE)->getSExtValue() + 1LL)) {
1780 TrueBB = LHSR.first->BB;
1782 TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1783 CurMF->insert(BBI, TrueBB);
1784 WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
1787 // Similar to the optimization above, if the Value being switched on is
1788 // known to be less than the Constant CR.LT, and the current Case Value
1789 // is CR.LT - 1, then we can branch directly to the target block for
1790 // the current Case Value, rather than emitting a RHS leaf node for it.
1791 if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
1792 cast<ConstantInt>(RHSR.first->Low)->getSExtValue() ==
1793 (cast<ConstantInt>(CR.LT)->getSExtValue() - 1LL)) {
1794 FalseBB = RHSR.first->BB;
1796 FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1797 CurMF->insert(BBI, FalseBB);
1798 WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
1801 // Create a CaseBlock record representing a conditional branch to
1802 // the LHS node if the value being switched on SV is less than C.
1803 // Otherwise, branch to LHS.
1804 CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
1806 if (CR.CaseBB == CurMBB)
1807 visitSwitchCase(CB);
1809 SwitchCases.push_back(CB);
1814 /// handleBitTestsSwitchCase - if current case range has few destination and
1815 /// range span less, than machine word bitwidth, encode case range into series
1816 /// of masks and emit bit tests with these masks.
1817 bool SelectionDAGLowering::handleBitTestsSwitchCase(CaseRec& CR,
1818 CaseRecVector& WorkList,
1820 MachineBasicBlock* Default){
1821 unsigned IntPtrBits = TLI.getPointerTy().getSizeInBits();
1823 Case& FrontCase = *CR.Range.first;
1824 Case& BackCase = *(CR.Range.second-1);
1826 // Get the MachineFunction which holds the current MBB. This is used when
1827 // inserting any additional MBBs necessary to represent the switch.
1828 MachineFunction *CurMF = CurMBB->getParent();
1830 unsigned numCmps = 0;
1831 for (CaseItr I = CR.Range.first, E = CR.Range.second;
1833 // Single case counts one, case range - two.
1834 if (I->Low == I->High)
1840 // Count unique destinations
1841 SmallSet<MachineBasicBlock*, 4> Dests;
1842 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1843 Dests.insert(I->BB);
1844 if (Dests.size() > 3)
1845 // Don't bother the code below, if there are too much unique destinations
1848 DOUT << "Total number of unique destinations: " << Dests.size() << "\n"
1849 << "Total number of comparisons: " << numCmps << "\n";
1851 // Compute span of values.
1852 Constant* minValue = FrontCase.Low;
1853 Constant* maxValue = BackCase.High;
1854 uint64_t range = cast<ConstantInt>(maxValue)->getSExtValue() -
1855 cast<ConstantInt>(minValue)->getSExtValue();
1856 DOUT << "Compare range: " << range << "\n"
1857 << "Low bound: " << cast<ConstantInt>(minValue)->getSExtValue() << "\n"
1858 << "High bound: " << cast<ConstantInt>(maxValue)->getSExtValue() << "\n";
1860 if (range>=IntPtrBits ||
1861 (!(Dests.size() == 1 && numCmps >= 3) &&
1862 !(Dests.size() == 2 && numCmps >= 5) &&
1863 !(Dests.size() >= 3 && numCmps >= 6)))
1866 DOUT << "Emitting bit tests\n";
1867 int64_t lowBound = 0;
1869 // Optimize the case where all the case values fit in a
1870 // word without having to subtract minValue. In this case,
1871 // we can optimize away the subtraction.
1872 if (cast<ConstantInt>(minValue)->getSExtValue() >= 0 &&
1873 cast<ConstantInt>(maxValue)->getSExtValue() < IntPtrBits) {
1874 range = cast<ConstantInt>(maxValue)->getSExtValue();
1876 lowBound = cast<ConstantInt>(minValue)->getSExtValue();
1879 CaseBitsVector CasesBits;
1880 unsigned i, count = 0;
1882 for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
1883 MachineBasicBlock* Dest = I->BB;
1884 for (i = 0; i < count; ++i)
1885 if (Dest == CasesBits[i].BB)
1889 assert((count < 3) && "Too much destinations to test!");
1890 CasesBits.push_back(CaseBits(0, Dest, 0));
1894 uint64_t lo = cast<ConstantInt>(I->Low)->getSExtValue() - lowBound;
1895 uint64_t hi = cast<ConstantInt>(I->High)->getSExtValue() - lowBound;
1897 for (uint64_t j = lo; j <= hi; j++) {
1898 CasesBits[i].Mask |= 1ULL << j;
1899 CasesBits[i].Bits++;
1903 std::sort(CasesBits.begin(), CasesBits.end(), CaseBitsCmp());
1907 // Figure out which block is immediately after the current one.
1908 MachineFunction::iterator BBI = CR.CaseBB;
1911 const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
1914 for (unsigned i = 0, e = CasesBits.size(); i!=e; ++i) {
1915 DOUT << "Mask: " << CasesBits[i].Mask << ", Bits: " << CasesBits[i].Bits
1916 << ", BB: " << CasesBits[i].BB << "\n";
1918 MachineBasicBlock *CaseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
1919 CurMF->insert(BBI, CaseBB);
1920 BTC.push_back(BitTestCase(CasesBits[i].Mask,
1925 BitTestBlock BTB(lowBound, range, SV,
1926 -1U, (CR.CaseBB == CurMBB),
1927 CR.CaseBB, Default, BTC);
1929 if (CR.CaseBB == CurMBB)
1930 visitBitTestHeader(BTB);
1932 BitTestCases.push_back(BTB);
1938 /// Clusterify - Transform simple list of Cases into list of CaseRange's
1939 unsigned SelectionDAGLowering::Clusterify(CaseVector& Cases,
1940 const SwitchInst& SI) {
1941 unsigned numCmps = 0;
1943 // Start with "simple" cases
1944 for (unsigned i = 1; i < SI.getNumSuccessors(); ++i) {
1945 MachineBasicBlock *SMBB = FuncInfo.MBBMap[SI.getSuccessor(i)];
1946 Cases.push_back(Case(SI.getSuccessorValue(i),
1947 SI.getSuccessorValue(i),
1950 std::sort(Cases.begin(), Cases.end(), CaseCmp());
1952 // Merge case into clusters
1953 if (Cases.size()>=2)
1954 // Must recompute end() each iteration because it may be
1955 // invalidated by erase if we hold on to it
1956 for (CaseItr I=Cases.begin(), J=++(Cases.begin()); J!=Cases.end(); ) {
1957 int64_t nextValue = cast<ConstantInt>(J->Low)->getSExtValue();
1958 int64_t currentValue = cast<ConstantInt>(I->High)->getSExtValue();
1959 MachineBasicBlock* nextBB = J->BB;
1960 MachineBasicBlock* currentBB = I->BB;
1962 // If the two neighboring cases go to the same destination, merge them
1963 // into a single case.
1964 if ((nextValue-currentValue==1) && (currentBB == nextBB)) {
1972 for (CaseItr I=Cases.begin(), E=Cases.end(); I!=E; ++I, ++numCmps) {
1973 if (I->Low != I->High)
1974 // A range counts double, since it requires two compares.
1981 void SelectionDAGLowering::visitSwitch(SwitchInst &SI) {
1982 // Figure out which block is immediately after the current one.
1983 MachineBasicBlock *NextBlock = 0;
1984 MachineFunction::iterator BBI = CurMBB;
1986 MachineBasicBlock *Default = FuncInfo.MBBMap[SI.getDefaultDest()];
1988 // If there is only the default destination, branch to it if it is not the
1989 // next basic block. Otherwise, just fall through.
1990 if (SI.getNumOperands() == 2) {
1991 // Update machine-CFG edges.
1993 // If this is not a fall-through branch, emit the branch.
1994 CurMBB->addSuccessor(Default);
1995 if (Default != NextBlock)
1996 DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getControlRoot(),
1997 DAG.getBasicBlock(Default)));
2002 // If there are any non-default case statements, create a vector of Cases
2003 // representing each one, and sort the vector so that we can efficiently
2004 // create a binary search tree from them.
2006 unsigned numCmps = Clusterify(Cases, SI);
2007 DOUT << "Clusterify finished. Total clusters: " << Cases.size()
2008 << ". Total compares: " << numCmps << "\n";
2010 // Get the Value to be switched on and default basic blocks, which will be
2011 // inserted into CaseBlock records, representing basic blocks in the binary
2013 Value *SV = SI.getOperand(0);
2015 // Push the initial CaseRec onto the worklist
2016 CaseRecVector WorkList;
2017 WorkList.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
2019 while (!WorkList.empty()) {
2020 // Grab a record representing a case range to process off the worklist
2021 CaseRec CR = WorkList.back();
2022 WorkList.pop_back();
2024 if (handleBitTestsSwitchCase(CR, WorkList, SV, Default))
2027 // If the range has few cases (two or less) emit a series of specific
2029 if (handleSmallSwitchRange(CR, WorkList, SV, Default))
2032 // If the switch has more than 5 blocks, and at least 40% dense, and the
2033 // target supports indirect branches, then emit a jump table rather than
2034 // lowering the switch to a binary tree of conditional branches.
2035 if (handleJTSwitchCase(CR, WorkList, SV, Default))
2038 // Emit binary tree. We need to pick a pivot, and push left and right ranges
2039 // onto the worklist. Leafs are handled via handleSmallSwitchRange() call.
2040 handleBTSplitSwitchCase(CR, WorkList, SV, Default);
2045 void SelectionDAGLowering::visitSub(User &I) {
2046 // -0.0 - X --> fneg
2047 const Type *Ty = I.getType();
2048 if (isa<VectorType>(Ty)) {
2049 if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
2050 const VectorType *DestTy = cast<VectorType>(I.getType());
2051 const Type *ElTy = DestTy->getElementType();
2052 if (ElTy->isFloatingPoint()) {
2053 unsigned VL = DestTy->getNumElements();
2054 std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
2055 Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
2057 SDValue Op2 = getValue(I.getOperand(1));
2058 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2064 if (Ty->isFloatingPoint()) {
2065 if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
2066 if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
2067 SDValue Op2 = getValue(I.getOperand(1));
2068 setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
2073 visitBinary(I, Ty->isFPOrFPVector() ? ISD::FSUB : ISD::SUB);
2076 void SelectionDAGLowering::visitBinary(User &I, unsigned OpCode) {
2077 SDValue Op1 = getValue(I.getOperand(0));
2078 SDValue Op2 = getValue(I.getOperand(1));
2080 setValue(&I, DAG.getNode(OpCode, Op1.getValueType(), Op1, Op2));
2083 void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
2084 SDValue Op1 = getValue(I.getOperand(0));
2085 SDValue Op2 = getValue(I.getOperand(1));
2086 if (!isa<VectorType>(I.getType())) {
2087 if (TLI.getShiftAmountTy().bitsLT(Op2.getValueType()))
2088 Op2 = DAG.getNode(ISD::TRUNCATE, TLI.getShiftAmountTy(), Op2);
2089 else if (TLI.getShiftAmountTy().bitsGT(Op2.getValueType()))
2090 Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
2093 setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
2096 void SelectionDAGLowering::visitICmp(User &I) {
2097 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2098 if (ICmpInst *IC = dyn_cast<ICmpInst>(&I))
2099 predicate = IC->getPredicate();
2100 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2101 predicate = ICmpInst::Predicate(IC->getPredicate());
2102 SDValue Op1 = getValue(I.getOperand(0));
2103 SDValue Op2 = getValue(I.getOperand(1));
2104 ISD::CondCode Opcode = getICmpCondCode(predicate);
2105 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
2108 void SelectionDAGLowering::visitFCmp(User &I) {
2109 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2110 if (FCmpInst *FC = dyn_cast<FCmpInst>(&I))
2111 predicate = FC->getPredicate();
2112 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2113 predicate = FCmpInst::Predicate(FC->getPredicate());
2114 SDValue Op1 = getValue(I.getOperand(0));
2115 SDValue Op2 = getValue(I.getOperand(1));
2116 ISD::CondCode Condition = getFCmpCondCode(predicate);
2117 setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Condition));
2120 void SelectionDAGLowering::visitVICmp(User &I) {
2121 ICmpInst::Predicate predicate = ICmpInst::BAD_ICMP_PREDICATE;
2122 if (VICmpInst *IC = dyn_cast<VICmpInst>(&I))
2123 predicate = IC->getPredicate();
2124 else if (ConstantExpr *IC = dyn_cast<ConstantExpr>(&I))
2125 predicate = ICmpInst::Predicate(IC->getPredicate());
2126 SDValue Op1 = getValue(I.getOperand(0));
2127 SDValue Op2 = getValue(I.getOperand(1));
2128 ISD::CondCode Opcode = getICmpCondCode(predicate);
2129 setValue(&I, DAG.getVSetCC(Op1.getValueType(), Op1, Op2, Opcode));
2132 void SelectionDAGLowering::visitVFCmp(User &I) {
2133 FCmpInst::Predicate predicate = FCmpInst::BAD_FCMP_PREDICATE;
2134 if (VFCmpInst *FC = dyn_cast<VFCmpInst>(&I))
2135 predicate = FC->getPredicate();
2136 else if (ConstantExpr *FC = dyn_cast<ConstantExpr>(&I))
2137 predicate = FCmpInst::Predicate(FC->getPredicate());
2138 SDValue Op1 = getValue(I.getOperand(0));
2139 SDValue Op2 = getValue(I.getOperand(1));
2140 ISD::CondCode Condition = getFCmpCondCode(predicate);
2141 MVT DestVT = TLI.getValueType(I.getType());
2143 setValue(&I, DAG.getVSetCC(DestVT, Op1, Op2, Condition));
2146 void SelectionDAGLowering::visitSelect(User &I) {
2147 SmallVector<MVT, 4> ValueVTs;
2148 ComputeValueVTs(TLI, I.getType(), ValueVTs);
2149 unsigned NumValues = ValueVTs.size();
2150 if (NumValues != 0) {
2151 SmallVector<SDValue, 4> Values(NumValues);
2152 SDValue Cond = getValue(I.getOperand(0));
2153 SDValue TrueVal = getValue(I.getOperand(1));
2154 SDValue FalseVal = getValue(I.getOperand(2));
2156 for (unsigned i = 0; i != NumValues; ++i)
2157 Values[i] = DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
2158 SDValue(TrueVal.getNode(), TrueVal.getResNo() + i),
2159 SDValue(FalseVal.getNode(), FalseVal.getResNo() + i));
2161 setValue(&I, DAG.getMergeValues(DAG.getVTList(&ValueVTs[0], NumValues),
2162 &Values[0], NumValues));
2167 void SelectionDAGLowering::visitTrunc(User &I) {
2168 // TruncInst cannot be a no-op cast because sizeof(src) > sizeof(dest).
2169 SDValue N = getValue(I.getOperand(0));
2170 MVT DestVT = TLI.getValueType(I.getType());
2171 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2174 void SelectionDAGLowering::visitZExt(User &I) {
2175 // ZExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2176 // ZExt also can't be a cast to bool for same reason. So, nothing much to do
2177 SDValue N = getValue(I.getOperand(0));
2178 MVT DestVT = TLI.getValueType(I.getType());
2179 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2182 void SelectionDAGLowering::visitSExt(User &I) {
2183 // SExt cannot be a no-op cast because sizeof(src) < sizeof(dest).
2184 // SExt also can't be a cast to bool for same reason. So, nothing much to do
2185 SDValue N = getValue(I.getOperand(0));
2186 MVT DestVT = TLI.getValueType(I.getType());
2187 setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
2190 void SelectionDAGLowering::visitFPTrunc(User &I) {
2191 // FPTrunc is never a no-op cast, no need to check
2192 SDValue N = getValue(I.getOperand(0));
2193 MVT DestVT = TLI.getValueType(I.getType());
2194 setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N, DAG.getIntPtrConstant(0)));
2197 void SelectionDAGLowering::visitFPExt(User &I){
2198 // FPTrunc is never a no-op cast, no need to check
2199 SDValue N = getValue(I.getOperand(0));
2200 MVT DestVT = TLI.getValueType(I.getType());
2201 setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
2204 void SelectionDAGLowering::visitFPToUI(User &I) {
2205 // FPToUI is never a no-op cast, no need to check
2206 SDValue N = getValue(I.getOperand(0));
2207 MVT DestVT = TLI.getValueType(I.getType());
2208 setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
2211 void SelectionDAGLowering::visitFPToSI(User &I) {
2212 // FPToSI is never a no-op cast, no need to check
2213 SDValue N = getValue(I.getOperand(0));
2214 MVT DestVT = TLI.getValueType(I.getType());
2215 setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
2218 void SelectionDAGLowering::visitUIToFP(User &I) {
2219 // UIToFP is never a no-op cast, no need to check
2220 SDValue N = getValue(I.getOperand(0));
2221 MVT DestVT = TLI.getValueType(I.getType());
2222 setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
2225 void SelectionDAGLowering::visitSIToFP(User &I){
2226 // SIToFP is never a no-op cast, no need to check
2227 SDValue N = getValue(I.getOperand(0));
2228 MVT DestVT = TLI.getValueType(I.getType());
2229 setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
2232 void SelectionDAGLowering::visitPtrToInt(User &I) {
2233 // What to do depends on the size of the integer and the size of the pointer.
2234 // We can either truncate, zero extend, or no-op, accordingly.
2235 SDValue N = getValue(I.getOperand(0));
2236 MVT SrcVT = N.getValueType();
2237 MVT DestVT = TLI.getValueType(I.getType());
2239 if (DestVT.bitsLT(SrcVT))
2240 Result = DAG.getNode(ISD::TRUNCATE, DestVT, N);
2242 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2243 Result = DAG.getNode(ISD::ZERO_EXTEND, DestVT, N);
2244 setValue(&I, Result);
2247 void SelectionDAGLowering::visitIntToPtr(User &I) {
2248 // What to do depends on the size of the integer and the size of the pointer.
2249 // We can either truncate, zero extend, or no-op, accordingly.
2250 SDValue N = getValue(I.getOperand(0));
2251 MVT SrcVT = N.getValueType();
2252 MVT DestVT = TLI.getValueType(I.getType());
2253 if (DestVT.bitsLT(SrcVT))
2254 setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
2256 // Note: ZERO_EXTEND can handle cases where the sizes are equal too
2257 setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
2260 void SelectionDAGLowering::visitBitCast(User &I) {
2261 SDValue N = getValue(I.getOperand(0));
2262 MVT DestVT = TLI.getValueType(I.getType());
2264 // BitCast assures us that source and destination are the same size so this
2265 // is either a BIT_CONVERT or a no-op.
2266 if (DestVT != N.getValueType())
2267 setValue(&I, DAG.getNode(ISD::BIT_CONVERT, DestVT, N)); // convert types
2269 setValue(&I, N); // noop cast.
2272 void SelectionDAGLowering::visitInsertElement(User &I) {
2273 SDValue InVec = getValue(I.getOperand(0));
2274 SDValue InVal = getValue(I.getOperand(1));
2275 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2276 getValue(I.getOperand(2)));
2278 setValue(&I, DAG.getNode(ISD::INSERT_VECTOR_ELT,
2279 TLI.getValueType(I.getType()),
2280 InVec, InVal, InIdx));
2283 void SelectionDAGLowering::visitExtractElement(User &I) {
2284 SDValue InVec = getValue(I.getOperand(0));
2285 SDValue InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
2286 getValue(I.getOperand(1)));
2287 setValue(&I, DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
2288 TLI.getValueType(I.getType()), InVec, InIdx));
2291 void SelectionDAGLowering::visitShuffleVector(User &I) {
2292 SDValue V1 = getValue(I.getOperand(0));
2293 SDValue V2 = getValue(I.getOperand(1));
2294 SDValue Mask = getValue(I.getOperand(2));
2296 setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE,
2297 TLI.getValueType(I.getType()),
2301 void SelectionDAGLowering::visitInsertValue(InsertValueInst &I) {
2302 const Value *Op0 = I.getOperand(0);
2303 const Value *Op1 = I.getOperand(1);
2304 const Type *AggTy = I.getType();
2305 const Type *ValTy = Op1->getType();
2306 bool IntoUndef = isa<UndefValue>(Op0);
2307 bool FromUndef = isa<UndefValue>(Op1);
2309 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2310 I.idx_begin(), I.idx_end());
2312 SmallVector<MVT, 4> AggValueVTs;
2313 ComputeValueVTs(TLI, AggTy, AggValueVTs);
2314 SmallVector<MVT, 4> ValValueVTs;
2315 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2317 unsigned NumAggValues = AggValueVTs.size();
2318 unsigned NumValValues = ValValueVTs.size();
2319 SmallVector<SDValue, 4> Values(NumAggValues);
2321 SDValue Agg = getValue(Op0);
2322 SDValue Val = getValue(Op1);
2324 // Copy the beginning value(s) from the original aggregate.
2325 for (; i != LinearIndex; ++i)
2326 Values[i] = IntoUndef ? DAG.getNode(ISD::UNDEF, AggValueVTs[i]) :
2327 SDValue(Agg.getNode(), Agg.getResNo() + i);
2328 // Copy values from the inserted value(s).
2329 for (; i != LinearIndex + NumValValues; ++i)
2330 Values[i] = FromUndef ? DAG.getNode(ISD::UNDEF, AggValueVTs[i]) :
2331 SDValue(Val.getNode(), Val.getResNo() + i - LinearIndex);
2332 // Copy remaining value(s) from the original aggregate.
2333 for (; i != NumAggValues; ++i)
2334 Values[i] = IntoUndef ? DAG.getNode(ISD::UNDEF, AggValueVTs[i]) :
2335 SDValue(Agg.getNode(), Agg.getResNo() + i);
2337 setValue(&I, DAG.getMergeValues(DAG.getVTList(&AggValueVTs[0], NumAggValues),
2338 &Values[0], NumAggValues));
2341 void SelectionDAGLowering::visitExtractValue(ExtractValueInst &I) {
2342 const Value *Op0 = I.getOperand(0);
2343 const Type *AggTy = Op0->getType();
2344 const Type *ValTy = I.getType();
2345 bool OutOfUndef = isa<UndefValue>(Op0);
2347 unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
2348 I.idx_begin(), I.idx_end());
2350 SmallVector<MVT, 4> ValValueVTs;
2351 ComputeValueVTs(TLI, ValTy, ValValueVTs);
2353 unsigned NumValValues = ValValueVTs.size();
2354 SmallVector<SDValue, 4> Values(NumValValues);
2356 SDValue Agg = getValue(Op0);
2357 // Copy out the selected value(s).
2358 for (unsigned i = LinearIndex; i != LinearIndex + NumValValues; ++i)
2359 Values[i - LinearIndex] =
2360 OutOfUndef ? DAG.getNode(ISD::UNDEF, Agg.getNode()->getValueType(Agg.getResNo() + i)) :
2361 SDValue(Agg.getNode(), Agg.getResNo() + i);
2363 setValue(&I, DAG.getMergeValues(DAG.getVTList(&ValValueVTs[0], NumValValues),
2364 &Values[0], NumValValues));
2368 void SelectionDAGLowering::visitGetElementPtr(User &I) {
2369 SDValue N = getValue(I.getOperand(0));
2370 const Type *Ty = I.getOperand(0)->getType();
2372 for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
2375 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
2376 unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
2379 uint64_t Offset = TD->getStructLayout(StTy)->getElementOffset(Field);
2380 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
2381 DAG.getIntPtrConstant(Offset));
2383 Ty = StTy->getElementType(Field);
2385 Ty = cast<SequentialType>(Ty)->getElementType();
2387 // If this is a constant subscript, handle it quickly.
2388 if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
2389 if (CI->getZExtValue() == 0) continue;
2391 TD->getABITypeSize(Ty)*cast<ConstantInt>(CI)->getSExtValue();
2392 N = DAG.getNode(ISD::ADD, N.getValueType(), N,
2393 DAG.getIntPtrConstant(Offs));
2397 // N = N + Idx * ElementSize;
2398 uint64_t ElementSize = TD->getABITypeSize(Ty);
2399 SDValue IdxN = getValue(Idx);
2401 // If the index is smaller or larger than intptr_t, truncate or extend
2403 if (IdxN.getValueType().bitsLT(N.getValueType()))
2404 IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
2405 else if (IdxN.getValueType().bitsGT(N.getValueType()))
2406 IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
2408 // If this is a multiply by a power of two, turn it into a shl
2409 // immediately. This is a very common case.
2410 if (ElementSize != 1) {
2411 if (isPowerOf2_64(ElementSize)) {
2412 unsigned Amt = Log2_64(ElementSize);
2413 IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
2414 DAG.getConstant(Amt, TLI.getShiftAmountTy()));
2416 SDValue Scale = DAG.getIntPtrConstant(ElementSize);
2417 IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
2421 N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
2427 void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
2428 // If this is a fixed sized alloca in the entry block of the function,
2429 // allocate it statically on the stack.
2430 if (FuncInfo.StaticAllocaMap.count(&I))
2431 return; // getValue will auto-populate this.
2433 const Type *Ty = I.getAllocatedType();
2434 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
2436 std::max((unsigned)TLI.getTargetData()->getPrefTypeAlignment(Ty),
2439 SDValue AllocSize = getValue(I.getArraySize());
2440 MVT IntPtr = TLI.getPointerTy();
2441 if (IntPtr.bitsLT(AllocSize.getValueType()))
2442 AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
2443 else if (IntPtr.bitsGT(AllocSize.getValueType()))
2444 AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
2446 AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
2447 DAG.getIntPtrConstant(TySize));
2449 // Handle alignment. If the requested alignment is less than or equal to
2450 // the stack alignment, ignore it. If the size is greater than or equal to
2451 // the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
2452 unsigned StackAlign =
2453 TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
2454 if (Align <= StackAlign)
2457 // Round the size of the allocation up to the stack alignment size
2458 // by add SA-1 to the size.
2459 AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
2460 DAG.getIntPtrConstant(StackAlign-1));
2461 // Mask out the low bits for alignment purposes.
2462 AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
2463 DAG.getIntPtrConstant(~(uint64_t)(StackAlign-1)));
2465 SDValue Ops[] = { getRoot(), AllocSize, DAG.getIntPtrConstant(Align) };
2466 const MVT *VTs = DAG.getNodeValueTypes(AllocSize.getValueType(),
2468 SDValue DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, 2, Ops, 3);
2470 DAG.setRoot(DSA.getValue(1));
2472 // Inform the Frame Information that we have just allocated a variable-sized
2474 CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
2477 void SelectionDAGLowering::visitLoad(LoadInst &I) {
2478 const Value *SV = I.getOperand(0);
2479 SDValue Ptr = getValue(SV);
2481 const Type *Ty = I.getType();
2482 bool isVolatile = I.isVolatile();
2483 unsigned Alignment = I.getAlignment();
2485 SmallVector<MVT, 4> ValueVTs;
2486 SmallVector<uint64_t, 4> Offsets;
2487 ComputeValueVTs(TLI, Ty, ValueVTs, &Offsets);
2488 unsigned NumValues = ValueVTs.size();
2493 bool ConstantMemory = false;
2495 // Serialize volatile loads with other side effects.
2497 else if (AA->pointsToConstantMemory(SV)) {
2498 // Do not serialize (non-volatile) loads of constant memory with anything.
2499 Root = DAG.getEntryNode();
2500 ConstantMemory = true;
2502 // Do not serialize non-volatile loads against each other.
2503 Root = DAG.getRoot();
2506 SmallVector<SDValue, 4> Values(NumValues);
2507 SmallVector<SDValue, 4> Chains(NumValues);
2508 MVT PtrVT = Ptr.getValueType();
2509 for (unsigned i = 0; i != NumValues; ++i) {
2510 SDValue L = DAG.getLoad(ValueVTs[i], Root,
2511 DAG.getNode(ISD::ADD, PtrVT, Ptr,
2512 DAG.getConstant(Offsets[i], PtrVT)),
2514 isVolatile, Alignment);
2516 Chains[i] = L.getValue(1);
2519 if (!ConstantMemory) {
2520 SDValue Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
2521 &Chains[0], NumValues);
2525 PendingLoads.push_back(Chain);
2528 setValue(&I, DAG.getMergeValues(DAG.getVTList(&ValueVTs[0], NumValues),
2529 &Values[0], NumValues));
2533 void SelectionDAGLowering::visitStore(StoreInst &I) {
2534 Value *SrcV = I.getOperand(0);
2535 Value *PtrV = I.getOperand(1);
2537 SmallVector<MVT, 4> ValueVTs;
2538 SmallVector<uint64_t, 4> Offsets;
2539 ComputeValueVTs(TLI, SrcV->getType(), ValueVTs, &Offsets);
2540 unsigned NumValues = ValueVTs.size();
2544 // Get the lowered operands. Note that we do this after
2545 // checking if NumResults is zero, because with zero results
2546 // the operands won't have values in the map.
2547 SDValue Src = getValue(SrcV);
2548 SDValue Ptr = getValue(PtrV);
2550 SDValue Root = getRoot();
2551 SmallVector<SDValue, 4> Chains(NumValues);
2552 MVT PtrVT = Ptr.getValueType();
2553 bool isVolatile = I.isVolatile();
2554 unsigned Alignment = I.getAlignment();
2555 for (unsigned i = 0; i != NumValues; ++i)
2556 Chains[i] = DAG.getStore(Root, SDValue(Src.getNode(), Src.getResNo() + i),
2557 DAG.getNode(ISD::ADD, PtrVT, Ptr,
2558 DAG.getConstant(Offsets[i], PtrVT)),
2560 isVolatile, Alignment);
2562 DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, &Chains[0], NumValues));
2565 /// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
2567 void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
2568 unsigned Intrinsic) {
2569 bool HasChain = !I.doesNotAccessMemory();
2570 bool OnlyLoad = HasChain && I.onlyReadsMemory();
2572 // Build the operand list.
2573 SmallVector<SDValue, 8> Ops;
2574 if (HasChain) { // If this intrinsic has side-effects, chainify it.
2576 // We don't need to serialize loads against other loads.
2577 Ops.push_back(DAG.getRoot());
2579 Ops.push_back(getRoot());
2583 // Info is set by getTgtMemInstrinsic
2584 TargetLowering::IntrinsicInfo Info;
2585 bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
2587 // Add the intrinsic ID as an integer operand if it's not a target intrinsic.
2588 if (!IsTgtIntrinsic)
2589 Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
2591 // Add all operands of the call to the operand list.
2592 for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
2593 SDValue Op = getValue(I.getOperand(i));
2594 assert(TLI.isTypeLegal(Op.getValueType()) &&
2595 "Intrinsic uses a non-legal type?");
2599 std::vector<MVT> VTs;
2600 if (I.getType() != Type::VoidTy) {
2601 MVT VT = TLI.getValueType(I.getType());
2602 if (VT.isVector()) {
2603 const VectorType *DestTy = cast<VectorType>(I.getType());
2604 MVT EltVT = TLI.getValueType(DestTy->getElementType());
2606 VT = MVT::getVectorVT(EltVT, DestTy->getNumElements());
2607 assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
2610 assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
2614 VTs.push_back(MVT::Other);
2616 const MVT *VTList = DAG.getNodeValueTypes(VTs);
2620 if (IsTgtIntrinsic) {
2621 // This is target intrinsic that touches memory
2622 Result = DAG.getMemIntrinsicNode(Info.opc, VTList, VTs.size(),
2623 &Ops[0], Ops.size(),
2624 Info.memVT, Info.ptrVal, Info.offset,
2625 Info.align, Info.vol,
2626 Info.readMem, Info.writeMem);
2629 Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTList, VTs.size(),
2630 &Ops[0], Ops.size());
2631 else if (I.getType() != Type::VoidTy)
2632 Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTList, VTs.size(),
2633 &Ops[0], Ops.size());
2635 Result = DAG.getNode(ISD::INTRINSIC_VOID, VTList, VTs.size(),
2636 &Ops[0], Ops.size());
2639 SDValue Chain = Result.getValue(Result.getNode()->getNumValues()-1);
2641 PendingLoads.push_back(Chain);
2645 if (I.getType() != Type::VoidTy) {
2646 if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
2647 MVT VT = TLI.getValueType(PTy);
2648 Result = DAG.getNode(ISD::BIT_CONVERT, VT, Result);
2650 setValue(&I, Result);
2654 /// ExtractTypeInfo - Returns the type info, possibly bitcast, encoded in V.
2655 static GlobalVariable *ExtractTypeInfo(Value *V) {
2656 V = V->stripPointerCasts();
2657 GlobalVariable *GV = dyn_cast<GlobalVariable>(V);
2658 assert ((GV || isa<ConstantPointerNull>(V)) &&
2659 "TypeInfo must be a global variable or NULL");
2665 /// AddCatchInfo - Extract the personality and type infos from an eh.selector
2666 /// call, and add them to the specified machine basic block.
2667 void AddCatchInfo(CallInst &I, MachineModuleInfo *MMI,
2668 MachineBasicBlock *MBB) {
2669 // Inform the MachineModuleInfo of the personality for this landing pad.
2670 ConstantExpr *CE = cast<ConstantExpr>(I.getOperand(2));
2671 assert(CE->getOpcode() == Instruction::BitCast &&
2672 isa<Function>(CE->getOperand(0)) &&
2673 "Personality should be a function");
2674 MMI->addPersonality(MBB, cast<Function>(CE->getOperand(0)));
2676 // Gather all the type infos for this landing pad and pass them along to
2677 // MachineModuleInfo.
2678 std::vector<GlobalVariable *> TyInfo;
2679 unsigned N = I.getNumOperands();
2681 for (unsigned i = N - 1; i > 2; --i) {
2682 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(i))) {
2683 unsigned FilterLength = CI->getZExtValue();
2684 unsigned FirstCatch = i + FilterLength + !FilterLength;
2685 assert (FirstCatch <= N && "Invalid filter length");
2687 if (FirstCatch < N) {
2688 TyInfo.reserve(N - FirstCatch);
2689 for (unsigned j = FirstCatch; j < N; ++j)
2690 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2691 MMI->addCatchTypeInfo(MBB, TyInfo);
2695 if (!FilterLength) {
2697 MMI->addCleanup(MBB);
2700 TyInfo.reserve(FilterLength - 1);
2701 for (unsigned j = i + 1; j < FirstCatch; ++j)
2702 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2703 MMI->addFilterTypeInfo(MBB, TyInfo);
2712 TyInfo.reserve(N - 3);
2713 for (unsigned j = 3; j < N; ++j)
2714 TyInfo.push_back(ExtractTypeInfo(I.getOperand(j)));
2715 MMI->addCatchTypeInfo(MBB, TyInfo);
2721 /// GetSignificand - Get the significand and build it into a floating-point
2722 /// number with exponent of 1:
2724 /// Op = (Op & 0x007fffff) | 0x3f800000;
2726 /// where Op is the hexidecimal representation of floating point value.
2728 GetSignificand(SelectionDAG &DAG, SDValue Op) {
2729 SDValue t1 = DAG.getNode(ISD::AND, MVT::i32, Op,
2730 DAG.getConstant(0x007fffff, MVT::i32));
2731 SDValue t2 = DAG.getNode(ISD::OR, MVT::i32, t1,
2732 DAG.getConstant(0x3f800000, MVT::i32));
2733 return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t2);
2736 /// GetExponent - Get the exponent:
2738 /// (float)((Op1 >> 23) - 127);
2740 /// where Op is the hexidecimal representation of floating point value.
2742 GetExponent(SelectionDAG &DAG, SDValue Op) {
2743 SDValue t1 = DAG.getNode(ISD::SRL, MVT::i32, Op,
2744 DAG.getConstant(23, MVT::i32));
2745 SDValue t2 = DAG.getNode(ISD::SUB, MVT::i32, t1,
2746 DAG.getConstant(127, MVT::i32));
2747 return DAG.getNode(ISD::UINT_TO_FP, MVT::f32, t2);
2750 /// getF32Constant - Get 32-bit floating point constant.
2752 getF32Constant(SelectionDAG &DAG, unsigned Flt) {
2753 return DAG.getConstantFP(APFloat(APInt(32, Flt)), MVT::f32);
2756 /// Inlined utility function to implement binary input atomic intrinsics for
2757 /// visitIntrinsicCall: I is a call instruction
2758 /// Op is the associated NodeType for I
2760 SelectionDAGLowering::implVisitBinaryAtomic(CallInst& I, ISD::NodeType Op) {
2761 SDValue Root = getRoot();
2762 SDValue L = DAG.getAtomic(Op, Root,
2763 getValue(I.getOperand(1)),
2764 getValue(I.getOperand(2)),
2767 DAG.setRoot(L.getValue(1));
2771 /// visitExp - Lower an exp intrinsic. Handles the special sequences for
2772 /// limited-precision mode.
2774 SelectionDAGLowering::visitExp(CallInst &I) {
2777 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
2778 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
2779 SDValue Op = getValue(I.getOperand(1));
2781 // Put the exponent in the right bit position for later addition to the
2784 // #define LOG2OFe 1.4426950f
2785 // IntegerPartOfX = ((int32_t)(X * LOG2OFe));
2786 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, Op,
2787 getF32Constant(DAG, 0x3fb8aa3b));
2788 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, t0);
2790 // FractionalPartOfX = (X * LOG2OFe) - (float)IntegerPartOfX;
2791 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, IntegerPartOfX);
2792 SDValue X = DAG.getNode(ISD::FSUB, MVT::f32, t0, t1);
2794 // IntegerPartOfX <<= 23;
2795 IntegerPartOfX = DAG.getNode(ISD::SHL, MVT::i32, IntegerPartOfX,
2796 DAG.getConstant(23, MVT::i32));
2798 if (LimitFloatPrecision <= 6) {
2799 // For floating-point precision of 6:
2801 // TwoToFractionalPartOfX =
2803 // (0.735607626f + 0.252464424f * x) * x;
2805 // error 0.0144103317, which is 6 bits
2806 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2807 getF32Constant(DAG, 0x3e814304));
2808 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
2809 getF32Constant(DAG, 0x3f3c50c8));
2810 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
2811 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
2812 getF32Constant(DAG, 0x3f7f5e7e));
2813 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t5);
2815 // Add the exponent into the result in integer domain.
2816 SDValue t6 = DAG.getNode(ISD::ADD, MVT::i32,
2817 TwoToFracPartOfX, IntegerPartOfX);
2819 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t6);
2820 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
2821 // For floating-point precision of 12:
2823 // TwoToFractionalPartOfX =
2826 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
2828 // 0.000107046256 error, which is 13 to 14 bits
2829 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2830 getF32Constant(DAG, 0x3da235e3));
2831 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
2832 getF32Constant(DAG, 0x3e65b8f3));
2833 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
2834 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
2835 getF32Constant(DAG, 0x3f324b07));
2836 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
2837 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
2838 getF32Constant(DAG, 0x3f7ff8fd));
2839 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t7);
2841 // Add the exponent into the result in integer domain.
2842 SDValue t8 = DAG.getNode(ISD::ADD, MVT::i32,
2843 TwoToFracPartOfX, IntegerPartOfX);
2845 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t8);
2846 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
2847 // For floating-point precision of 18:
2849 // TwoToFractionalPartOfX =
2853 // (0.554906021e-1f +
2854 // (0.961591928e-2f +
2855 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
2857 // error 2.47208000*10^(-7), which is better than 18 bits
2858 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2859 getF32Constant(DAG, 0x3924b03e));
2860 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
2861 getF32Constant(DAG, 0x3ab24b87));
2862 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
2863 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
2864 getF32Constant(DAG, 0x3c1d8c17));
2865 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
2866 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
2867 getF32Constant(DAG, 0x3d634a1d));
2868 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
2869 SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
2870 getF32Constant(DAG, 0x3e75fe14));
2871 SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X);
2872 SDValue t11 = DAG.getNode(ISD::FADD, MVT::f32, t10,
2873 getF32Constant(DAG, 0x3f317234));
2874 SDValue t12 = DAG.getNode(ISD::FMUL, MVT::f32, t11, X);
2875 SDValue t13 = DAG.getNode(ISD::FADD, MVT::f32, t12,
2876 getF32Constant(DAG, 0x3f800000));
2877 SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t13);
2879 // Add the exponent into the result in integer domain.
2880 SDValue t14 = DAG.getNode(ISD::ADD, MVT::i32,
2881 TwoToFracPartOfX, IntegerPartOfX);
2883 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, t14);
2886 // No special expansion.
2887 result = DAG.getNode(ISD::FEXP,
2888 getValue(I.getOperand(1)).getValueType(),
2889 getValue(I.getOperand(1)));
2892 setValue(&I, result);
2895 /// visitLog - Lower a log intrinsic. Handles the special sequences for
2896 /// limited-precision mode.
2898 SelectionDAGLowering::visitLog(CallInst &I) {
2901 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
2902 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
2903 SDValue Op = getValue(I.getOperand(1));
2904 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
2906 // Scale the exponent by log(2) [0.69314718f].
2907 SDValue Exp = GetExponent(DAG, Op1);
2908 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, MVT::f32, Exp,
2909 getF32Constant(DAG, 0x3f317218));
2911 // Get the significand and build it into a floating-point number with
2913 SDValue X = GetSignificand(DAG, Op1);
2915 if (LimitFloatPrecision <= 6) {
2916 // For floating-point precision of 6:
2920 // (1.4034025f - 0.23903021f * x) * x;
2922 // error 0.0034276066, which is better than 8 bits
2923 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2924 getF32Constant(DAG, 0xbe74c456));
2925 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
2926 getF32Constant(DAG, 0x3fb3a2b1));
2927 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
2928 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t2,
2929 getF32Constant(DAG, 0x3f949a29));
2931 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, LogOfMantissa);
2932 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
2933 // For floating-point precision of 12:
2939 // (0.44717955f - 0.56570851e-1f * x) * x) * x) * x;
2941 // error 0.000061011436, which is 14 bits
2942 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2943 getF32Constant(DAG, 0xbd67b6d6));
2944 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
2945 getF32Constant(DAG, 0x3ee4f4b8));
2946 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
2947 SDValue t3 = DAG.getNode(ISD::FSUB, MVT::f32, t2,
2948 getF32Constant(DAG, 0x3fbc278b));
2949 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
2950 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
2951 getF32Constant(DAG, 0x40348e95));
2952 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
2953 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t6,
2954 getF32Constant(DAG, 0x3fdef31a));
2956 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, LogOfMantissa);
2957 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
2958 // For floating-point precision of 18:
2966 // (0.19073739f - 0.17809712e-1f * x) * x) * x) * x) * x)*x;
2968 // error 0.0000023660568, which is better than 18 bits
2969 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
2970 getF32Constant(DAG, 0xbc91e5ac));
2971 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
2972 getF32Constant(DAG, 0x3e4350aa));
2973 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
2974 SDValue t3 = DAG.getNode(ISD::FSUB, MVT::f32, t2,
2975 getF32Constant(DAG, 0x3f60d3e3));
2976 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
2977 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
2978 getF32Constant(DAG, 0x4011cdf0));
2979 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
2980 SDValue t7 = DAG.getNode(ISD::FSUB, MVT::f32, t6,
2981 getF32Constant(DAG, 0x406cfd1c));
2982 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
2983 SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
2984 getF32Constant(DAG, 0x408797cb));
2985 SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X);
2986 SDValue LogOfMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t10,
2987 getF32Constant(DAG, 0x4006dcab));
2989 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, LogOfMantissa);
2992 // No special expansion.
2993 result = DAG.getNode(ISD::FLOG,
2994 getValue(I.getOperand(1)).getValueType(),
2995 getValue(I.getOperand(1)));
2998 setValue(&I, result);
3001 /// visitLog2 - Lower a log2 intrinsic. Handles the special sequences for
3002 /// limited-precision mode.
3004 SelectionDAGLowering::visitLog2(CallInst &I) {
3007 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3008 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3009 SDValue Op = getValue(I.getOperand(1));
3010 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
3012 // Get the exponent.
3013 SDValue LogOfExponent = GetExponent(DAG, Op1);
3015 // Get the significand and build it into a floating-point number with
3017 SDValue X = GetSignificand(DAG, Op1);
3019 // Different possible minimax approximations of significand in
3020 // floating-point for various degrees of accuracy over [1,2].
3021 if (LimitFloatPrecision <= 6) {
3022 // For floating-point precision of 6:
3024 // Log2ofMantissa = -1.6749035f + (2.0246817f - .34484768f * x) * x;
3026 // error 0.0049451742, which is more than 7 bits
3027 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3028 getF32Constant(DAG, 0xbeb08fe0));
3029 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
3030 getF32Constant(DAG, 0x40019463));
3031 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3032 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t2,
3033 getF32Constant(DAG, 0x3fd6633d));
3035 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log2ofMantissa);
3036 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3037 // For floating-point precision of 12:
3043 // (.645142248f - 0.816157886e-1f * x) * x) * x) * x;
3045 // error 0.0000876136000, which is better than 13 bits
3046 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3047 getF32Constant(DAG, 0xbda7262e));
3048 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
3049 getF32Constant(DAG, 0x3f25280b));
3050 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3051 SDValue t3 = DAG.getNode(ISD::FSUB, MVT::f32, t2,
3052 getF32Constant(DAG, 0x4007b923));
3053 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3054 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3055 getF32Constant(DAG, 0x40823e2f));
3056 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3057 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t6,
3058 getF32Constant(DAG, 0x4020d29c));
3060 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log2ofMantissa);
3061 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3062 // For floating-point precision of 18:
3071 // 0.25691327e-1f * x) * x) * x) * x) * x) * x;
3073 // error 0.0000018516, which is better than 18 bits
3074 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3075 getF32Constant(DAG, 0xbcd2769e));
3076 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
3077 getF32Constant(DAG, 0x3e8ce0b9));
3078 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3079 SDValue t3 = DAG.getNode(ISD::FSUB, MVT::f32, t2,
3080 getF32Constant(DAG, 0x3fa22ae7));
3081 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3082 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3083 getF32Constant(DAG, 0x40525723));
3084 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3085 SDValue t7 = DAG.getNode(ISD::FSUB, MVT::f32, t6,
3086 getF32Constant(DAG, 0x40aaf200));
3087 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
3088 SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
3089 getF32Constant(DAG, 0x40c39dad));
3090 SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X);
3091 SDValue Log2ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t10,
3092 getF32Constant(DAG, 0x4042902c));
3094 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log2ofMantissa);
3097 // No special expansion.
3098 result = DAG.getNode(ISD::FLOG2,
3099 getValue(I.getOperand(1)).getValueType(),
3100 getValue(I.getOperand(1)));
3103 setValue(&I, result);
3106 /// visitLog10 - Lower a log10 intrinsic. Handles the special sequences for
3107 /// limited-precision mode.
3109 SelectionDAGLowering::visitLog10(CallInst &I) {
3112 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3113 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3114 SDValue Op = getValue(I.getOperand(1));
3115 SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op);
3117 // Scale the exponent by log10(2) [0.30102999f].
3118 SDValue Exp = GetExponent(DAG, Op1);
3119 SDValue LogOfExponent = DAG.getNode(ISD::FMUL, MVT::f32, Exp,
3120 getF32Constant(DAG, 0x3e9a209a));
3122 // Get the significand and build it into a floating-point number with
3124 SDValue X = GetSignificand(DAG, Op1);
3126 if (LimitFloatPrecision <= 6) {
3127 // For floating-point precision of 6:
3129 // Log10ofMantissa =
3131 // (0.60948995f - 0.10380950f * x) * x;
3133 // error 0.0014886165, which is 6 bits
3134 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3135 getF32Constant(DAG, 0xbdd49a13));
3136 SDValue t1 = DAG.getNode(ISD::FADD, MVT::f32, t0,
3137 getF32Constant(DAG, 0x3f1c0789));
3138 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3139 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t2,
3140 getF32Constant(DAG, 0x3f011300));
3142 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log10ofMantissa);
3143 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3144 // For floating-point precision of 12:
3146 // Log10ofMantissa =
3149 // (-0.31664806f + 0.47637168e-1f * x) * x) * x;
3151 // error 0.00019228036, which is better than 12 bits
3152 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3153 getF32Constant(DAG, 0x3d431f31));
3154 SDValue t1 = DAG.getNode(ISD::FSUB, MVT::f32, t0,
3155 getF32Constant(DAG, 0x3ea21fb2));
3156 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3157 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3158 getF32Constant(DAG, 0x3f6ae232));
3159 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3160 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t4,
3161 getF32Constant(DAG, 0x3f25f7c3));
3163 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log10ofMantissa);
3164 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3165 // For floating-point precision of 18:
3167 // Log10ofMantissa =
3172 // (-0.12539807f + 0.13508273e-1f * x) * x) * x) * x) * x;
3174 // error 0.0000037995730, which is better than 18 bits
3175 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3176 getF32Constant(DAG, 0x3c5d51ce));
3177 SDValue t1 = DAG.getNode(ISD::FSUB, MVT::f32, t0,
3178 getF32Constant(DAG, 0x3e00685a));
3179 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, t1, X);
3180 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3181 getF32Constant(DAG, 0x3efb6798));
3182 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3183 SDValue t5 = DAG.getNode(ISD::FSUB, MVT::f32, t4,
3184 getF32Constant(DAG, 0x3f88d192));
3185 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3186 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
3187 getF32Constant(DAG, 0x3fc4316c));
3188 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
3189 SDValue Log10ofMantissa = DAG.getNode(ISD::FSUB, MVT::f32, t8,
3190 getF32Constant(DAG, 0x3f57ce70));
3192 result = DAG.getNode(ISD::FADD, MVT::f32, LogOfExponent, Log10ofMantissa);
3195 // No special expansion.
3196 result = DAG.getNode(ISD::FLOG10,
3197 getValue(I.getOperand(1)).getValueType(),
3198 getValue(I.getOperand(1)));
3201 setValue(&I, result);
3204 /// visitExp2 - Lower an exp2 intrinsic. Handles the special sequences for
3205 /// limited-precision mode.
3207 SelectionDAGLowering::visitExp2(CallInst &I) {
3210 if (getValue(I.getOperand(1)).getValueType() == MVT::f32 &&
3211 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3212 SDValue Op = getValue(I.getOperand(1));
3214 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, Op);
3216 // FractionalPartOfX = x - (float)IntegerPartOfX;
3217 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, IntegerPartOfX);
3218 SDValue X = DAG.getNode(ISD::FSUB, MVT::f32, Op, t1);
3220 // IntegerPartOfX <<= 23;
3221 IntegerPartOfX = DAG.getNode(ISD::SHL, MVT::i32, IntegerPartOfX,
3222 DAG.getConstant(23, MVT::i32));
3224 if (LimitFloatPrecision <= 6) {
3225 // For floating-point precision of 6:
3227 // TwoToFractionalPartOfX =
3229 // (0.735607626f + 0.252464424f * x) * x;
3231 // error 0.0144103317, which is 6 bits
3232 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3233 getF32Constant(DAG, 0x3e814304));
3234 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3235 getF32Constant(DAG, 0x3f3c50c8));
3236 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3237 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3238 getF32Constant(DAG, 0x3f7f5e7e));
3239 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t5);
3240 SDValue TwoToFractionalPartOfX =
3241 DAG.getNode(ISD::ADD, MVT::i32, t6, IntegerPartOfX);
3243 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3244 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3245 // For floating-point precision of 12:
3247 // TwoToFractionalPartOfX =
3250 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3252 // error 0.000107046256, which is 13 to 14 bits
3253 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3254 getF32Constant(DAG, 0x3da235e3));
3255 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3256 getF32Constant(DAG, 0x3e65b8f3));
3257 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3258 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3259 getF32Constant(DAG, 0x3f324b07));
3260 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3261 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
3262 getF32Constant(DAG, 0x3f7ff8fd));
3263 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t7);
3264 SDValue TwoToFractionalPartOfX =
3265 DAG.getNode(ISD::ADD, MVT::i32, t8, IntegerPartOfX);
3267 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3268 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3269 // For floating-point precision of 18:
3271 // TwoToFractionalPartOfX =
3275 // (0.554906021e-1f +
3276 // (0.961591928e-2f +
3277 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3278 // error 2.47208000*10^(-7), which is better than 18 bits
3279 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3280 getF32Constant(DAG, 0x3924b03e));
3281 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3282 getF32Constant(DAG, 0x3ab24b87));
3283 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3284 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3285 getF32Constant(DAG, 0x3c1d8c17));
3286 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3287 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
3288 getF32Constant(DAG, 0x3d634a1d));
3289 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
3290 SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
3291 getF32Constant(DAG, 0x3e75fe14));
3292 SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X);
3293 SDValue t11 = DAG.getNode(ISD::FADD, MVT::f32, t10,
3294 getF32Constant(DAG, 0x3f317234));
3295 SDValue t12 = DAG.getNode(ISD::FMUL, MVT::f32, t11, X);
3296 SDValue t13 = DAG.getNode(ISD::FADD, MVT::f32, t12,
3297 getF32Constant(DAG, 0x3f800000));
3298 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t13);
3299 SDValue TwoToFractionalPartOfX =
3300 DAG.getNode(ISD::ADD, MVT::i32, t14, IntegerPartOfX);
3302 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3305 // No special expansion.
3306 result = DAG.getNode(ISD::FEXP2,
3307 getValue(I.getOperand(1)).getValueType(),
3308 getValue(I.getOperand(1)));
3311 setValue(&I, result);
3314 /// visitPow - Lower a pow intrinsic. Handles the special sequences for
3315 /// limited-precision mode with x == 10.0f.
3317 SelectionDAGLowering::visitPow(CallInst &I) {
3319 Value *Val = I.getOperand(1);
3320 bool IsExp10 = false;
3322 if (getValue(Val).getValueType() == MVT::f32 &&
3323 getValue(I.getOperand(2)).getValueType() == MVT::f32 &&
3324 LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3325 if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(Val))) {
3326 if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
3328 IsExp10 = CFP->getValueAPF().bitwiseIsEqual(Ten);
3333 if (IsExp10 && LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
3334 SDValue Op = getValue(I.getOperand(2));
3336 // Put the exponent in the right bit position for later addition to the
3339 // #define LOG2OF10 3.3219281f
3340 // IntegerPartOfX = (int32_t)(x * LOG2OF10);
3341 SDValue t0 = DAG.getNode(ISD::FMUL, MVT::f32, Op,
3342 getF32Constant(DAG, 0x40549a78));
3343 SDValue IntegerPartOfX = DAG.getNode(ISD::FP_TO_SINT, MVT::i32, t0);
3345 // FractionalPartOfX = x - (float)IntegerPartOfX;
3346 SDValue t1 = DAG.getNode(ISD::SINT_TO_FP, MVT::f32, IntegerPartOfX);
3347 SDValue X = DAG.getNode(ISD::FSUB, MVT::f32, t0, t1);
3349 // IntegerPartOfX <<= 23;
3350 IntegerPartOfX = DAG.getNode(ISD::SHL, MVT::i32, IntegerPartOfX,
3351 DAG.getConstant(23, MVT::i32));
3353 if (LimitFloatPrecision <= 6) {
3354 // For floating-point precision of 6:
3356 // twoToFractionalPartOfX =
3358 // (0.735607626f + 0.252464424f * x) * x;
3360 // error 0.0144103317, which is 6 bits
3361 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3362 getF32Constant(DAG, 0x3e814304));
3363 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3364 getF32Constant(DAG, 0x3f3c50c8));
3365 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3366 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3367 getF32Constant(DAG, 0x3f7f5e7e));
3368 SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t5);
3369 SDValue TwoToFractionalPartOfX =
3370 DAG.getNode(ISD::ADD, MVT::i32, t6, IntegerPartOfX);
3372 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3373 } else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
3374 // For floating-point precision of 12:
3376 // TwoToFractionalPartOfX =
3379 // (0.224338339f + 0.792043434e-1f * x) * x) * x;
3381 // error 0.000107046256, which is 13 to 14 bits
3382 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3383 getF32Constant(DAG, 0x3da235e3));
3384 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3385 getF32Constant(DAG, 0x3e65b8f3));
3386 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3387 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3388 getF32Constant(DAG, 0x3f324b07));
3389 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3390 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
3391 getF32Constant(DAG, 0x3f7ff8fd));
3392 SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t7);
3393 SDValue TwoToFractionalPartOfX =
3394 DAG.getNode(ISD::ADD, MVT::i32, t8, IntegerPartOfX);
3396 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3397 } else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
3398 // For floating-point precision of 18:
3400 // TwoToFractionalPartOfX =
3404 // (0.554906021e-1f +
3405 // (0.961591928e-2f +
3406 // (0.136028312e-2f + 0.157059148e-3f *x)*x)*x)*x)*x)*x;
3407 // error 2.47208000*10^(-7), which is better than 18 bits
3408 SDValue t2 = DAG.getNode(ISD::FMUL, MVT::f32, X,
3409 getF32Constant(DAG, 0x3924b03e));
3410 SDValue t3 = DAG.getNode(ISD::FADD, MVT::f32, t2,
3411 getF32Constant(DAG, 0x3ab24b87));
3412 SDValue t4 = DAG.getNode(ISD::FMUL, MVT::f32, t3, X);
3413 SDValue t5 = DAG.getNode(ISD::FADD, MVT::f32, t4,
3414 getF32Constant(DAG, 0x3c1d8c17));
3415 SDValue t6 = DAG.getNode(ISD::FMUL, MVT::f32, t5, X);
3416 SDValue t7 = DAG.getNode(ISD::FADD, MVT::f32, t6,
3417 getF32Constant(DAG, 0x3d634a1d));
3418 SDValue t8 = DAG.getNode(ISD::FMUL, MVT::f32, t7, X);
3419 SDValue t9 = DAG.getNode(ISD::FADD, MVT::f32, t8,
3420 getF32Constant(DAG, 0x3e75fe14));
3421 SDValue t10 = DAG.getNode(ISD::FMUL, MVT::f32, t9, X);
3422 SDValue t11 = DAG.getNode(ISD::FADD, MVT::f32, t10,
3423 getF32Constant(DAG, 0x3f317234));
3424 SDValue t12 = DAG.getNode(ISD::FMUL, MVT::f32, t11, X);
3425 SDValue t13 = DAG.getNode(ISD::FADD, MVT::f32, t12,
3426 getF32Constant(DAG, 0x3f800000));
3427 SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, MVT::i32, t13);
3428 SDValue TwoToFractionalPartOfX =
3429 DAG.getNode(ISD::ADD, MVT::i32, t14, IntegerPartOfX);
3431 result = DAG.getNode(ISD::BIT_CONVERT, MVT::f32, TwoToFractionalPartOfX);
3434 // No special expansion.
3435 result = DAG.getNode(ISD::FPOW,
3436 getValue(I.getOperand(1)).getValueType(),
3437 getValue(I.getOperand(1)),
3438 getValue(I.getOperand(2)));
3441 setValue(&I, result);
3444 /// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
3445 /// we want to emit this as a call to a named external function, return the name
3446 /// otherwise lower it and return null.
3448 SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
3449 switch (Intrinsic) {
3451 // By default, turn this into a target intrinsic node.
3452 visitTargetIntrinsic(I, Intrinsic);
3454 case Intrinsic::vastart: visitVAStart(I); return 0;
3455 case Intrinsic::vaend: visitVAEnd(I); return 0;
3456 case Intrinsic::vacopy: visitVACopy(I); return 0;
3457 case Intrinsic::returnaddress:
3458 setValue(&I, DAG.getNode(ISD::RETURNADDR, TLI.getPointerTy(),
3459 getValue(I.getOperand(1))));
3461 case Intrinsic::frameaddress:
3462 setValue(&I, DAG.getNode(ISD::FRAMEADDR, TLI.getPointerTy(),
3463 getValue(I.getOperand(1))));
3465 case Intrinsic::setjmp:
3466 return "_setjmp"+!TLI.usesUnderscoreSetJmp();
3468 case Intrinsic::longjmp:
3469 return "_longjmp"+!TLI.usesUnderscoreLongJmp();
3471 case Intrinsic::memcpy_i32:
3472 case Intrinsic::memcpy_i64: {
3473 SDValue Op1 = getValue(I.getOperand(1));
3474 SDValue Op2 = getValue(I.getOperand(2));
3475 SDValue Op3 = getValue(I.getOperand(3));
3476 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3477 DAG.setRoot(DAG.getMemcpy(getRoot(), Op1, Op2, Op3, Align, false,
3478 I.getOperand(1), 0, I.getOperand(2), 0));
3481 case Intrinsic::memset_i32:
3482 case Intrinsic::memset_i64: {
3483 SDValue Op1 = getValue(I.getOperand(1));
3484 SDValue Op2 = getValue(I.getOperand(2));
3485 SDValue Op3 = getValue(I.getOperand(3));
3486 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3487 DAG.setRoot(DAG.getMemset(getRoot(), Op1, Op2, Op3, Align,
3488 I.getOperand(1), 0));
3491 case Intrinsic::memmove_i32:
3492 case Intrinsic::memmove_i64: {
3493 SDValue Op1 = getValue(I.getOperand(1));
3494 SDValue Op2 = getValue(I.getOperand(2));
3495 SDValue Op3 = getValue(I.getOperand(3));
3496 unsigned Align = cast<ConstantInt>(I.getOperand(4))->getZExtValue();
3498 // If the source and destination are known to not be aliases, we can
3499 // lower memmove as memcpy.
3500 uint64_t Size = -1ULL;
3501 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
3502 Size = C->getZExtValue();
3503 if (AA->alias(I.getOperand(1), Size, I.getOperand(2), Size) ==
3504 AliasAnalysis::NoAlias) {
3505 DAG.setRoot(DAG.getMemcpy(getRoot(), Op1, Op2, Op3, Align, false,
3506 I.getOperand(1), 0, I.getOperand(2), 0));
3510 DAG.setRoot(DAG.getMemmove(getRoot(), Op1, Op2, Op3, Align,
3511 I.getOperand(1), 0, I.getOperand(2), 0));
3514 case Intrinsic::dbg_stoppoint: {
3515 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3516 DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
3517 if (MMI && SPI.getContext() && MMI->Verify(SPI.getContext())) {
3518 DebugInfoDesc *DD = MMI->getDescFor(SPI.getContext());
3519 assert(DD && "Not a debug information descriptor");
3520 DAG.setRoot(DAG.getDbgStopPoint(getRoot(),
3523 cast<CompileUnitDesc>(DD)));
3528 case Intrinsic::dbg_region_start: {
3529 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3530 DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
3531 if (MMI && RSI.getContext() && MMI->Verify(RSI.getContext())) {
3532 unsigned LabelID = MMI->RecordRegionStart(RSI.getContext());
3533 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getRoot(), LabelID));
3538 case Intrinsic::dbg_region_end: {
3539 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3540 DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
3541 if (MMI && REI.getContext() && MMI->Verify(REI.getContext())) {
3542 unsigned LabelID = MMI->RecordRegionEnd(REI.getContext());
3543 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getRoot(), LabelID));
3548 case Intrinsic::dbg_func_start: {
3549 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3551 DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
3552 Value *SP = FSI.getSubprogram();
3553 if (SP && MMI->Verify(SP)) {
3554 // llvm.dbg.func.start implicitly defines a dbg_stoppoint which is
3555 // what (most?) gdb expects.
3556 DebugInfoDesc *DD = MMI->getDescFor(SP);
3557 assert(DD && "Not a debug information descriptor");
3558 SubprogramDesc *Subprogram = cast<SubprogramDesc>(DD);
3559 const CompileUnitDesc *CompileUnit = Subprogram->getFile();
3560 unsigned SrcFile = MMI->RecordSource(CompileUnit);
3561 // Record the source line but does not create a label for the normal
3562 // function start. It will be emitted at asm emission time. However,
3563 // create a label if this is a beginning of inlined function.
3564 unsigned LabelID = MMI->RecordSourceLine(Subprogram->getLine(), 0, SrcFile);
3565 if (MMI->getSourceLines().size() != 1)
3566 DAG.setRoot(DAG.getLabel(ISD::DBG_LABEL, getRoot(), LabelID));
3571 case Intrinsic::dbg_declare: {
3572 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3573 DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
3574 Value *Variable = DI.getVariable();
3575 if (MMI && Variable && MMI->Verify(Variable))
3576 DAG.setRoot(DAG.getNode(ISD::DECLARE, MVT::Other, getRoot(),
3577 getValue(DI.getAddress()), getValue(Variable)));
3581 case Intrinsic::eh_exception: {
3582 if (!CurMBB->isLandingPad()) {
3583 // FIXME: Mark exception register as live in. Hack for PR1508.
3584 unsigned Reg = TLI.getExceptionAddressRegister();
3585 if (Reg) CurMBB->addLiveIn(Reg);
3587 // Insert the EXCEPTIONADDR instruction.
3588 SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
3590 Ops[0] = DAG.getRoot();
3591 SDValue Op = DAG.getNode(ISD::EXCEPTIONADDR, VTs, Ops, 1);
3593 DAG.setRoot(Op.getValue(1));
3597 case Intrinsic::eh_selector_i32:
3598 case Intrinsic::eh_selector_i64: {
3599 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3600 MVT VT = (Intrinsic == Intrinsic::eh_selector_i32 ?
3601 MVT::i32 : MVT::i64);
3604 if (CurMBB->isLandingPad())
3605 AddCatchInfo(I, MMI, CurMBB);
3608 FuncInfo.CatchInfoLost.insert(&I);
3610 // FIXME: Mark exception selector register as live in. Hack for PR1508.
3611 unsigned Reg = TLI.getExceptionSelectorRegister();
3612 if (Reg) CurMBB->addLiveIn(Reg);
3615 // Insert the EHSELECTION instruction.
3616 SDVTList VTs = DAG.getVTList(VT, MVT::Other);
3618 Ops[0] = getValue(I.getOperand(1));
3620 SDValue Op = DAG.getNode(ISD::EHSELECTION, VTs, Ops, 2);
3622 DAG.setRoot(Op.getValue(1));
3624 setValue(&I, DAG.getConstant(0, VT));
3630 case Intrinsic::eh_typeid_for_i32:
3631 case Intrinsic::eh_typeid_for_i64: {
3632 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
3633 MVT VT = (Intrinsic == Intrinsic::eh_typeid_for_i32 ?
3634 MVT::i32 : MVT::i64);
3637 // Find the type id for the given typeinfo.
3638 GlobalVariable *GV = ExtractTypeInfo(I.getOperand(1));
3640 unsigned TypeID = MMI->getTypeIDFor(GV);
3641 setValue(&I, DAG.getConstant(TypeID, VT));
3643 // Return something different to eh_selector.
3644 setValue(&I, DAG.getConstant(1, VT));
3650 case Intrinsic::eh_return_i32:
3651 case Intrinsic::eh_return_i64:
3652 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
3653 MMI->setCallsEHReturn(true);
3654 DAG.setRoot(DAG.getNode(ISD::EH_RETURN,
3657 getValue(I.getOperand(1)),
3658 getValue(I.getOperand(2))));
3660 setValue(&I, DAG.getConstant(0, TLI.getPointerTy()));
3664 case Intrinsic::eh_unwind_init:
3665 if (MachineModuleInfo *MMI = DAG.getMachineModuleInfo()) {
3666 MMI->setCallsUnwindInit(true);
3671 case Intrinsic::eh_dwarf_cfa: {
3672 MVT VT = getValue(I.getOperand(1)).getValueType();
3674 if (VT.bitsGT(TLI.getPointerTy()))
3675 CfaArg = DAG.getNode(ISD::TRUNCATE,
3676 TLI.getPointerTy(), getValue(I.getOperand(1)));
3678 CfaArg = DAG.getNode(ISD::SIGN_EXTEND,
3679 TLI.getPointerTy(), getValue(I.getOperand(1)));
3681 SDValue Offset = DAG.getNode(ISD::ADD,
3683 DAG.getNode(ISD::FRAME_TO_ARGS_OFFSET,
3684 TLI.getPointerTy()),
3686 setValue(&I, DAG.getNode(ISD::ADD,
3688 DAG.getNode(ISD::FRAMEADDR,
3691 TLI.getPointerTy())),
3696 case Intrinsic::sqrt:
3697 setValue(&I, DAG.getNode(ISD::FSQRT,
3698 getValue(I.getOperand(1)).getValueType(),
3699 getValue(I.getOperand(1))));
3701 case Intrinsic::powi:
3702 setValue(&I, DAG.getNode(ISD::FPOWI,
3703 getValue(I.getOperand(1)).getValueType(),
3704 getValue(I.getOperand(1)),
3705 getValue(I.getOperand(2))));
3707 case Intrinsic::sin:
3708 setValue(&I, DAG.getNode(ISD::FSIN,
3709 getValue(I.getOperand(1)).getValueType(),
3710 getValue(I.getOperand(1))));
3712 case Intrinsic::cos:
3713 setValue(&I, DAG.getNode(ISD::FCOS,
3714 getValue(I.getOperand(1)).getValueType(),
3715 getValue(I.getOperand(1))));
3717 case Intrinsic::log:
3720 case Intrinsic::log2:
3723 case Intrinsic::log10:
3726 case Intrinsic::exp:
3729 case Intrinsic::exp2:
3732 case Intrinsic::pow:
3735 case Intrinsic::pcmarker: {
3736 SDValue Tmp = getValue(I.getOperand(1));
3737 DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
3740 case Intrinsic::readcyclecounter: {
3741 SDValue Op = getRoot();
3742 SDValue Tmp = DAG.getNode(ISD::READCYCLECOUNTER,
3743 DAG.getNodeValueTypes(MVT::i64, MVT::Other), 2,
3746 DAG.setRoot(Tmp.getValue(1));
3749 case Intrinsic::part_select: {
3750 // Currently not implemented: just abort
3751 assert(0 && "part_select intrinsic not implemented");
3754 case Intrinsic::part_set: {
3755 // Currently not implemented: just abort
3756 assert(0 && "part_set intrinsic not implemented");
3759 case Intrinsic::bswap:
3760 setValue(&I, DAG.getNode(ISD::BSWAP,
3761 getValue(I.getOperand(1)).getValueType(),
3762 getValue(I.getOperand(1))));
3764 case Intrinsic::cttz: {
3765 SDValue Arg = getValue(I.getOperand(1));
3766 MVT Ty = Arg.getValueType();
3767 SDValue result = DAG.getNode(ISD::CTTZ, Ty, Arg);
3768 setValue(&I, result);
3771 case Intrinsic::ctlz: {
3772 SDValue Arg = getValue(I.getOperand(1));
3773 MVT Ty = Arg.getValueType();
3774 SDValue result = DAG.getNode(ISD::CTLZ, Ty, Arg);
3775 setValue(&I, result);
3778 case Intrinsic::ctpop: {
3779 SDValue Arg = getValue(I.getOperand(1));
3780 MVT Ty = Arg.getValueType();
3781 SDValue result = DAG.getNode(ISD::CTPOP, Ty, Arg);
3782 setValue(&I, result);
3785 case Intrinsic::stacksave: {
3786 SDValue Op = getRoot();
3787 SDValue Tmp = DAG.getNode(ISD::STACKSAVE,
3788 DAG.getNodeValueTypes(TLI.getPointerTy(), MVT::Other), 2, &Op, 1);
3790 DAG.setRoot(Tmp.getValue(1));
3793 case Intrinsic::stackrestore: {
3794 SDValue Tmp = getValue(I.getOperand(1));
3795 DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
3798 case Intrinsic::stackprotector_create: {
3799 // Emit code into the DAG to store the stack guard onto the stack.
3800 MachineFunction &MF = DAG.getMachineFunction();
3801 MachineFrameInfo *MFI = MF.getFrameInfo();
3802 MVT PtrTy = TLI.getPointerTy();
3804 SDValue Src = getValue(I.getOperand(1)); // The guard's value.
3805 AllocaInst *Slot = cast<AllocaInst>(I.getOperand(2));
3807 int FI = FuncInfo.StaticAllocaMap[Slot];
3808 MFI->setStackProtectorIndex(FI);
3810 SDValue FIN = DAG.getFrameIndex(FI, PtrTy);
3812 // Store the stack protector onto the stack.
3813 SDValue Result = DAG.getStore(getRoot(), Src, FIN,
3814 PseudoSourceValue::getFixedStack(FI),
3816 setValue(&I, Result);
3817 DAG.setRoot(Result);
3820 case Intrinsic::stackprotector_check: {
3821 // Emit code into the DAG to retrieve the stack guard off of the stack.
3822 MachineFunction &MF = DAG.getMachineFunction();
3823 MachineFrameInfo *MFI = MF.getFrameInfo();
3824 MVT PtrTy = TLI.getPointerTy();
3826 // Load the value stored on the stack.
3827 int FI = MFI->getStackProtectorIndex();
3828 SDValue FIN = DAG.getFrameIndex(MFI->getStackProtectorIndex(), PtrTy);
3829 setValue(&I, DAG.getLoad(PtrTy, getRoot(), FIN,
3830 PseudoSourceValue::getFixedStack(FI), 0, true));
3833 case Intrinsic::var_annotation:
3834 // Discard annotate attributes
3837 case Intrinsic::init_trampoline: {
3838 const Function *F = cast<Function>(I.getOperand(2)->stripPointerCasts());
3842 Ops[1] = getValue(I.getOperand(1));
3843 Ops[2] = getValue(I.getOperand(2));
3844 Ops[3] = getValue(I.getOperand(3));
3845 Ops[4] = DAG.getSrcValue(I.getOperand(1));
3846 Ops[5] = DAG.getSrcValue(F);
3848 SDValue Tmp = DAG.getNode(ISD::TRAMPOLINE,
3849 DAG.getNodeValueTypes(TLI.getPointerTy(),
3854 DAG.setRoot(Tmp.getValue(1));
3858 case Intrinsic::gcroot:
3860 Value *Alloca = I.getOperand(1);
3861 Constant *TypeMap = cast<Constant>(I.getOperand(2));
3863 FrameIndexSDNode *FI = cast<FrameIndexSDNode>(getValue(Alloca).getNode());
3864 GFI->addStackRoot(FI->getIndex(), TypeMap);
3868 case Intrinsic::gcread:
3869 case Intrinsic::gcwrite:
3870 assert(0 && "GC failed to lower gcread/gcwrite intrinsics!");
3873 case Intrinsic::flt_rounds: {
3874 setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, MVT::i32));
3878 case Intrinsic::trap: {
3879 DAG.setRoot(DAG.getNode(ISD::TRAP, MVT::Other, getRoot()));
3882 case Intrinsic::prefetch: {
3885 Ops[1] = getValue(I.getOperand(1));
3886 Ops[2] = getValue(I.getOperand(2));
3887 Ops[3] = getValue(I.getOperand(3));
3888 DAG.setRoot(DAG.getNode(ISD::PREFETCH, MVT::Other, &Ops[0], 4));
3892 case Intrinsic::memory_barrier: {
3895 for (int x = 1; x < 6; ++x)
3896 Ops[x] = getValue(I.getOperand(x));
3898 DAG.setRoot(DAG.getNode(ISD::MEMBARRIER, MVT::Other, &Ops[0], 6));
3901 case Intrinsic::atomic_cmp_swap: {
3902 SDValue Root = getRoot();
3904 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
3906 L = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP_8, Root,
3907 getValue(I.getOperand(1)),
3908 getValue(I.getOperand(2)),
3909 getValue(I.getOperand(3)),
3913 L = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP_16, Root,
3914 getValue(I.getOperand(1)),
3915 getValue(I.getOperand(2)),
3916 getValue(I.getOperand(3)),
3920 L = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP_32, Root,
3921 getValue(I.getOperand(1)),
3922 getValue(I.getOperand(2)),
3923 getValue(I.getOperand(3)),
3927 L = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP_64, Root,
3928 getValue(I.getOperand(1)),
3929 getValue(I.getOperand(2)),
3930 getValue(I.getOperand(3)),
3934 assert(0 && "Invalid atomic type");
3938 DAG.setRoot(L.getValue(1));
3941 case Intrinsic::atomic_load_add:
3942 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
3944 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD_8);
3946 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD_16);
3948 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD_32);
3950 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_ADD_64);
3952 assert(0 && "Invalid atomic type");
3955 case Intrinsic::atomic_load_sub:
3956 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
3958 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB_8);
3960 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB_16);
3962 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB_32);
3964 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_SUB_64);
3966 assert(0 && "Invalid atomic type");
3969 case Intrinsic::atomic_load_or:
3970 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
3972 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR_8);
3974 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR_16);
3976 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR_32);
3978 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_OR_64);
3980 assert(0 && "Invalid atomic type");
3983 case Intrinsic::atomic_load_xor:
3984 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
3986 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR_8);
3988 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR_16);
3990 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR_32);
3992 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_XOR_64);
3994 assert(0 && "Invalid atomic type");
3997 case Intrinsic::atomic_load_and:
3998 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4000 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND_8);
4002 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND_16);
4004 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND_32);
4006 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_AND_64);
4008 assert(0 && "Invalid atomic type");
4011 case Intrinsic::atomic_load_nand:
4012 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4014 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND_8);
4016 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND_16);
4018 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND_32);
4020 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_NAND_64);
4022 assert(0 && "Invalid atomic type");
4025 case Intrinsic::atomic_load_max:
4026 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4028 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX_8);
4030 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX_16);
4032 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX_32);
4034 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MAX_64);
4036 assert(0 && "Invalid atomic type");
4039 case Intrinsic::atomic_load_min:
4040 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4042 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN_8);
4044 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN_16);
4046 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN_32);
4048 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_MIN_64);
4050 assert(0 && "Invalid atomic type");
4053 case Intrinsic::atomic_load_umin:
4054 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4056 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN_8);
4058 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN_16);
4060 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN_32);
4062 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMIN_64);
4064 assert(0 && "Invalid atomic type");
4067 case Intrinsic::atomic_load_umax:
4068 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4070 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX_8);
4072 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX_16);
4074 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX_32);
4076 return implVisitBinaryAtomic(I, ISD::ATOMIC_LOAD_UMAX_64);
4078 assert(0 && "Invalid atomic type");
4081 case Intrinsic::atomic_swap:
4082 switch (getValue(I.getOperand(2)).getValueType().getSimpleVT()) {
4084 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP_8);
4086 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP_16);
4088 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP_32);
4090 return implVisitBinaryAtomic(I, ISD::ATOMIC_SWAP_64);
4092 assert(0 && "Invalid atomic type");
4099 void SelectionDAGLowering::LowerCallTo(CallSite CS, SDValue Callee,
4101 MachineBasicBlock *LandingPad) {
4102 const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
4103 const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
4104 MachineModuleInfo *MMI = DAG.getMachineModuleInfo();
4105 unsigned BeginLabel = 0, EndLabel = 0;
4107 TargetLowering::ArgListTy Args;
4108 TargetLowering::ArgListEntry Entry;
4109 Args.reserve(CS.arg_size());
4110 for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
4112 SDValue ArgNode = getValue(*i);
4113 Entry.Node = ArgNode; Entry.Ty = (*i)->getType();
4115 unsigned attrInd = i - CS.arg_begin() + 1;
4116 Entry.isSExt = CS.paramHasAttr(attrInd, Attribute::SExt);
4117 Entry.isZExt = CS.paramHasAttr(attrInd, Attribute::ZExt);
4118 Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
4119 Entry.isSRet = CS.paramHasAttr(attrInd, Attribute::StructRet);
4120 Entry.isNest = CS.paramHasAttr(attrInd, Attribute::Nest);
4121 Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
4122 Entry.Alignment = CS.getParamAlignment(attrInd);
4123 Args.push_back(Entry);
4126 if (LandingPad && MMI) {
4127 // Insert a label before the invoke call to mark the try range. This can be
4128 // used to detect deletion of the invoke via the MachineModuleInfo.
4129 BeginLabel = MMI->NextLabelID();
4130 // Both PendingLoads and PendingExports must be flushed here;
4131 // this call might not return.
4133 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getControlRoot(), BeginLabel));
4136 std::pair<SDValue,SDValue> Result =
4137 TLI.LowerCallTo(getRoot(), CS.getType(),
4138 CS.paramHasAttr(0, Attribute::SExt),
4139 CS.paramHasAttr(0, Attribute::ZExt), FTy->isVarArg(),
4140 CS.paramHasAttr(0, Attribute::InReg),
4141 CS.getCallingConv(),
4142 IsTailCall && PerformTailCallOpt,
4144 if (CS.getType() != Type::VoidTy)
4145 setValue(CS.getInstruction(), Result.first);
4146 DAG.setRoot(Result.second);
4148 if (LandingPad && MMI) {
4149 // Insert a label at the end of the invoke call to mark the try range. This
4150 // can be used to detect deletion of the invoke via the MachineModuleInfo.
4151 EndLabel = MMI->NextLabelID();
4152 DAG.setRoot(DAG.getLabel(ISD::EH_LABEL, getRoot(), EndLabel));
4154 // Inform MachineModuleInfo of range.
4155 MMI->addInvoke(LandingPad, BeginLabel, EndLabel);
4160 void SelectionDAGLowering::visitCall(CallInst &I) {
4161 const char *RenameFn = 0;
4162 if (Function *F = I.getCalledFunction()) {
4163 if (F->isDeclaration()) {
4164 if (unsigned IID = F->getIntrinsicID()) {
4165 RenameFn = visitIntrinsicCall(I, IID);
4171 // Check for well-known libc/libm calls. If the function is internal, it
4172 // can't be a library call.
4173 unsigned NameLen = F->getNameLen();
4174 if (!F->hasInternalLinkage() && NameLen) {
4175 const char *NameStr = F->getNameStart();
4176 if (NameStr[0] == 'c' &&
4177 ((NameLen == 8 && !strcmp(NameStr, "copysign")) ||
4178 (NameLen == 9 && !strcmp(NameStr, "copysignf")))) {
4179 if (I.getNumOperands() == 3 && // Basic sanity checks.
4180 I.getOperand(1)->getType()->isFloatingPoint() &&
4181 I.getType() == I.getOperand(1)->getType() &&
4182 I.getType() == I.getOperand(2)->getType()) {
4183 SDValue LHS = getValue(I.getOperand(1));
4184 SDValue RHS = getValue(I.getOperand(2));
4185 setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
4189 } else if (NameStr[0] == 'f' &&
4190 ((NameLen == 4 && !strcmp(NameStr, "fabs")) ||
4191 (NameLen == 5 && !strcmp(NameStr, "fabsf")) ||
4192 (NameLen == 5 && !strcmp(NameStr, "fabsl")))) {
4193 if (I.getNumOperands() == 2 && // Basic sanity checks.
4194 I.getOperand(1)->getType()->isFloatingPoint() &&
4195 I.getType() == I.getOperand(1)->getType()) {
4196 SDValue Tmp = getValue(I.getOperand(1));
4197 setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
4200 } else if (NameStr[0] == 's' &&
4201 ((NameLen == 3 && !strcmp(NameStr, "sin")) ||
4202 (NameLen == 4 && !strcmp(NameStr, "sinf")) ||
4203 (NameLen == 4 && !strcmp(NameStr, "sinl")))) {
4204 if (I.getNumOperands() == 2 && // Basic sanity checks.
4205 I.getOperand(1)->getType()->isFloatingPoint() &&
4206 I.getType() == I.getOperand(1)->getType()) {
4207 SDValue Tmp = getValue(I.getOperand(1));
4208 setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
4211 } else if (NameStr[0] == 'c' &&
4212 ((NameLen == 3 && !strcmp(NameStr, "cos")) ||
4213 (NameLen == 4 && !strcmp(NameStr, "cosf")) ||
4214 (NameLen == 4 && !strcmp(NameStr, "cosl")))) {
4215 if (I.getNumOperands() == 2 && // Basic sanity checks.
4216 I.getOperand(1)->getType()->isFloatingPoint() &&
4217 I.getType() == I.getOperand(1)->getType()) {
4218 SDValue Tmp = getValue(I.getOperand(1));
4219 setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
4224 } else if (isa<InlineAsm>(I.getOperand(0))) {
4231 Callee = getValue(I.getOperand(0));
4233 Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
4235 LowerCallTo(&I, Callee, I.isTailCall());
4239 /// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
4240 /// this value and returns the result as a ValueVT value. This uses
4241 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4242 /// If the Flag pointer is NULL, no flag is used.
4243 SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
4245 SDValue *Flag) const {
4246 // Assemble the legal parts into the final values.
4247 SmallVector<SDValue, 4> Values(ValueVTs.size());
4248 SmallVector<SDValue, 8> Parts;
4249 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4250 // Copy the legal parts from the registers.
4251 MVT ValueVT = ValueVTs[Value];
4252 unsigned NumRegs = TLI->getNumRegisters(ValueVT);
4253 MVT RegisterVT = RegVTs[Value];
4255 Parts.resize(NumRegs);
4256 for (unsigned i = 0; i != NumRegs; ++i) {
4259 P = DAG.getCopyFromReg(Chain, Regs[Part+i], RegisterVT);
4261 P = DAG.getCopyFromReg(Chain, Regs[Part+i], RegisterVT, *Flag);
4262 *Flag = P.getValue(2);
4264 Chain = P.getValue(1);
4266 // If the source register was virtual and if we know something about it,
4267 // add an assert node.
4268 if (TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) &&
4269 RegisterVT.isInteger() && !RegisterVT.isVector()) {
4270 unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
4271 FunctionLoweringInfo &FLI = DAG.getFunctionLoweringInfo();
4272 if (FLI.LiveOutRegInfo.size() > SlotNo) {
4273 FunctionLoweringInfo::LiveOutInfo &LOI = FLI.LiveOutRegInfo[SlotNo];
4275 unsigned RegSize = RegisterVT.getSizeInBits();
4276 unsigned NumSignBits = LOI.NumSignBits;
4277 unsigned NumZeroBits = LOI.KnownZero.countLeadingOnes();
4279 // FIXME: We capture more information than the dag can represent. For
4280 // now, just use the tightest assertzext/assertsext possible.
4282 MVT FromVT(MVT::Other);
4283 if (NumSignBits == RegSize)
4284 isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
4285 else if (NumZeroBits >= RegSize-1)
4286 isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
4287 else if (NumSignBits > RegSize-8)
4288 isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
4289 else if (NumZeroBits >= RegSize-9)
4290 isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
4291 else if (NumSignBits > RegSize-16)
4292 isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
4293 else if (NumZeroBits >= RegSize-17)
4294 isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
4295 else if (NumSignBits > RegSize-32)
4296 isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
4297 else if (NumZeroBits >= RegSize-33)
4298 isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
4300 if (FromVT != MVT::Other) {
4301 P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext,
4302 RegisterVT, P, DAG.getValueType(FromVT));
4311 Values[Value] = getCopyFromParts(DAG, Parts.begin(), NumRegs, RegisterVT,
4317 return DAG.getMergeValues(DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
4318 &Values[0], ValueVTs.size());
4321 /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
4322 /// specified value into the registers specified by this object. This uses
4323 /// Chain/Flag as the input and updates them for the output Chain/Flag.
4324 /// If the Flag pointer is NULL, no flag is used.
4325 void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG,
4326 SDValue &Chain, SDValue *Flag) const {
4327 // Get the list of the values's legal parts.
4328 unsigned NumRegs = Regs.size();
4329 SmallVector<SDValue, 8> Parts(NumRegs);
4330 for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
4331 MVT ValueVT = ValueVTs[Value];
4332 unsigned NumParts = TLI->getNumRegisters(ValueVT);
4333 MVT RegisterVT = RegVTs[Value];
4335 getCopyToParts(DAG, Val.getValue(Val.getResNo() + Value),
4336 &Parts[Part], NumParts, RegisterVT);
4340 // Copy the parts into the registers.
4341 SmallVector<SDValue, 8> Chains(NumRegs);
4342 for (unsigned i = 0; i != NumRegs; ++i) {
4345 Part = DAG.getCopyToReg(Chain, Regs[i], Parts[i]);
4347 Part = DAG.getCopyToReg(Chain, Regs[i], Parts[i], *Flag);
4348 *Flag = Part.getValue(1);
4350 Chains[i] = Part.getValue(0);
4353 if (NumRegs == 1 || Flag)
4354 // If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
4355 // flagged to it. That is the CopyToReg nodes and the user are considered
4356 // a single scheduling unit. If we create a TokenFactor and return it as
4357 // chain, then the TokenFactor is both a predecessor (operand) of the
4358 // user as well as a successor (the TF operands are flagged to the user).
4359 // c1, f1 = CopyToReg
4360 // c2, f2 = CopyToReg
4361 // c3 = TokenFactor c1, c2
4364 Chain = Chains[NumRegs-1];
4366 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, &Chains[0], NumRegs);
4369 /// AddInlineAsmOperands - Add this value to the specified inlineasm node
4370 /// operand list. This adds the code marker and includes the number of
4371 /// values added into it.
4372 void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
4373 std::vector<SDValue> &Ops) const {
4374 MVT IntPtrTy = DAG.getTargetLoweringInfo().getPointerTy();
4375 Ops.push_back(DAG.getTargetConstant(Code | (Regs.size() << 3), IntPtrTy));
4376 for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
4377 unsigned NumRegs = TLI->getNumRegisters(ValueVTs[Value]);
4378 MVT RegisterVT = RegVTs[Value];
4379 for (unsigned i = 0; i != NumRegs; ++i) {
4380 assert(Reg < Regs.size() && "Mismatch in # registers expected");
4381 Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
4386 /// isAllocatableRegister - If the specified register is safe to allocate,
4387 /// i.e. it isn't a stack pointer or some other special register, return the
4388 /// register class for the register. Otherwise, return null.
4389 static const TargetRegisterClass *
4390 isAllocatableRegister(unsigned Reg, MachineFunction &MF,
4391 const TargetLowering &TLI,
4392 const TargetRegisterInfo *TRI) {
4393 MVT FoundVT = MVT::Other;
4394 const TargetRegisterClass *FoundRC = 0;
4395 for (TargetRegisterInfo::regclass_iterator RCI = TRI->regclass_begin(),
4396 E = TRI->regclass_end(); RCI != E; ++RCI) {
4397 MVT ThisVT = MVT::Other;
4399 const TargetRegisterClass *RC = *RCI;
4400 // If none of the the value types for this register class are valid, we
4401 // can't use it. For example, 64-bit reg classes on 32-bit targets.
4402 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
4404 if (TLI.isTypeLegal(*I)) {
4405 // If we have already found this register in a different register class,
4406 // choose the one with the largest VT specified. For example, on
4407 // PowerPC, we favor f64 register classes over f32.
4408 if (FoundVT == MVT::Other || FoundVT.bitsLT(*I)) {
4415 if (ThisVT == MVT::Other) continue;
4417 // NOTE: This isn't ideal. In particular, this might allocate the
4418 // frame pointer in functions that need it (due to them not being taken
4419 // out of allocation, because a variable sized allocation hasn't been seen
4420 // yet). This is a slight code pessimization, but should still work.
4421 for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
4422 E = RC->allocation_order_end(MF); I != E; ++I)
4424 // We found a matching register class. Keep looking at others in case
4425 // we find one with larger registers that this physreg is also in.
4436 /// AsmOperandInfo - This contains information for each constraint that we are
4438 struct VISIBILITY_HIDDEN SDISelAsmOperandInfo :
4439 public TargetLowering::AsmOperandInfo {
4440 /// CallOperand - If this is the result output operand or a clobber
4441 /// this is null, otherwise it is the incoming operand to the CallInst.
4442 /// This gets modified as the asm is processed.
4443 SDValue CallOperand;
4445 /// AssignedRegs - If this is a register or register class operand, this
4446 /// contains the set of register corresponding to the operand.
4447 RegsForValue AssignedRegs;
4449 explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
4450 : TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
4453 /// MarkAllocatedRegs - Once AssignedRegs is set, mark the assigned registers
4454 /// busy in OutputRegs/InputRegs.
4455 void MarkAllocatedRegs(bool isOutReg, bool isInReg,
4456 std::set<unsigned> &OutputRegs,
4457 std::set<unsigned> &InputRegs,
4458 const TargetRegisterInfo &TRI) const {
4460 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4461 MarkRegAndAliases(AssignedRegs.Regs[i], OutputRegs, TRI);
4464 for (unsigned i = 0, e = AssignedRegs.Regs.size(); i != e; ++i)
4465 MarkRegAndAliases(AssignedRegs.Regs[i], InputRegs, TRI);
4469 /// getCallOperandValMVT - Return the MVT of the Value* that this operand
4470 /// corresponds to. If there is no Value* for this operand, it returns
4472 MVT getCallOperandValMVT(const TargetLowering &TLI,
4473 const TargetData *TD) const {
4474 if (CallOperandVal == 0) return MVT::Other;
4476 if (isa<BasicBlock>(CallOperandVal))
4477 return TLI.getPointerTy();
4479 const llvm::Type *OpTy = CallOperandVal->getType();
4481 // If this is an indirect operand, the operand is a pointer to the
4484 OpTy = cast<PointerType>(OpTy)->getElementType();
4486 // If OpTy is not a single value, it may be a struct/union that we
4487 // can tile with integers.
4488 if (!OpTy->isSingleValueType() && OpTy->isSized()) {
4489 unsigned BitSize = TD->getTypeSizeInBits(OpTy);
4498 OpTy = IntegerType::get(BitSize);
4503 return TLI.getValueType(OpTy, true);
4507 /// MarkRegAndAliases - Mark the specified register and all aliases in the
4509 static void MarkRegAndAliases(unsigned Reg, std::set<unsigned> &Regs,
4510 const TargetRegisterInfo &TRI) {
4511 assert(TargetRegisterInfo::isPhysicalRegister(Reg) && "Isn't a physreg");
4513 if (const unsigned *Aliases = TRI.getAliasSet(Reg))
4514 for (; *Aliases; ++Aliases)
4515 Regs.insert(*Aliases);
4518 } // end llvm namespace.
4521 /// GetRegistersForValue - Assign registers (virtual or physical) for the
4522 /// specified operand. We prefer to assign virtual registers, to allow the
4523 /// register allocator handle the assignment process. However, if the asm uses
4524 /// features that we can't model on machineinstrs, we have SDISel do the
4525 /// allocation. This produces generally horrible, but correct, code.
4527 /// OpInfo describes the operand.
4528 /// Input and OutputRegs are the set of already allocated physical registers.
4530 void SelectionDAGLowering::
4531 GetRegistersForValue(SDISelAsmOperandInfo &OpInfo,
4532 std::set<unsigned> &OutputRegs,
4533 std::set<unsigned> &InputRegs) {
4534 // Compute whether this value requires an input register, an output register,
4536 bool isOutReg = false;
4537 bool isInReg = false;
4538 switch (OpInfo.Type) {
4539 case InlineAsm::isOutput:
4542 // If there is an input constraint that matches this, we need to reserve
4543 // the input register so no other inputs allocate to it.
4544 isInReg = OpInfo.hasMatchingInput();
4546 case InlineAsm::isInput:
4550 case InlineAsm::isClobber:
4557 MachineFunction &MF = DAG.getMachineFunction();
4558 SmallVector<unsigned, 4> Regs;
4560 // If this is a constraint for a single physreg, or a constraint for a
4561 // register class, find it.
4562 std::pair<unsigned, const TargetRegisterClass*> PhysReg =
4563 TLI.getRegForInlineAsmConstraint(OpInfo.ConstraintCode,
4564 OpInfo.ConstraintVT);
4566 unsigned NumRegs = 1;
4567 if (OpInfo.ConstraintVT != MVT::Other) {
4568 // If this is a FP input in an integer register (or visa versa) insert a bit
4569 // cast of the input value. More generally, handle any case where the input
4570 // value disagrees with the register class we plan to stick this in.
4571 if (OpInfo.Type == InlineAsm::isInput &&
4572 PhysReg.second && !PhysReg.second->hasType(OpInfo.ConstraintVT)) {
4573 // Try to convert to the first MVT that the reg class contains. If the
4574 // types are identical size, use a bitcast to convert (e.g. two differing
4576 MVT RegVT = *PhysReg.second->vt_begin();
4577 if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
4578 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, RegVT,
4579 OpInfo.CallOperand);
4580 OpInfo.ConstraintVT = RegVT;
4581 } else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
4582 // If the input is a FP value and we want it in FP registers, do a
4583 // bitcast to the corresponding integer type. This turns an f64 value
4584 // into i64, which can be passed with two i32 values on a 32-bit
4586 RegVT = MVT::getIntegerVT(OpInfo.ConstraintVT.getSizeInBits());
4587 OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, RegVT,
4588 OpInfo.CallOperand);
4589 OpInfo.ConstraintVT = RegVT;
4593 NumRegs = TLI.getNumRegisters(OpInfo.ConstraintVT);
4597 MVT ValueVT = OpInfo.ConstraintVT;
4599 // If this is a constraint for a specific physical register, like {r17},
4601 if (PhysReg.first) {
4602 if (OpInfo.ConstraintVT == MVT::Other)
4603 ValueVT = *PhysReg.second->vt_begin();
4605 // Get the actual register value type. This is important, because the user
4606 // may have asked for (e.g.) the AX register in i32 type. We need to
4607 // remember that AX is actually i16 to get the right extension.
4608 RegVT = *PhysReg.second->vt_begin();
4610 // This is a explicit reference to a physical register.
4611 Regs.push_back(PhysReg.first);
4613 // If this is an expanded reference, add the rest of the regs to Regs.
4615 TargetRegisterClass::iterator I = PhysReg.second->begin();
4616 for (; *I != PhysReg.first; ++I)
4617 assert(I != PhysReg.second->end() && "Didn't find reg!");
4619 // Already added the first reg.
4621 for (; NumRegs; --NumRegs, ++I) {
4622 assert(I != PhysReg.second->end() && "Ran out of registers to allocate!");
4626 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4627 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4628 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
4632 // Otherwise, if this was a reference to an LLVM register class, create vregs
4633 // for this reference.
4634 std::vector<unsigned> RegClassRegs;
4635 const TargetRegisterClass *RC = PhysReg.second;
4637 // If this is a tied register, our regalloc doesn't know how to maintain
4638 // the constraint, so we have to pick a register to pin the input/output to.
4639 // If it isn't a matched constraint, go ahead and create vreg and let the
4640 // regalloc do its thing.
4641 if (!OpInfo.hasMatchingInput()) {
4642 RegVT = *PhysReg.second->vt_begin();
4643 if (OpInfo.ConstraintVT == MVT::Other)
4646 // Create the appropriate number of virtual registers.
4647 MachineRegisterInfo &RegInfo = MF.getRegInfo();
4648 for (; NumRegs; --NumRegs)
4649 Regs.push_back(RegInfo.createVirtualRegister(PhysReg.second));
4651 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, RegVT, ValueVT);
4655 // Otherwise, we can't allocate it. Let the code below figure out how to
4656 // maintain these constraints.
4657 RegClassRegs.assign(PhysReg.second->begin(), PhysReg.second->end());
4660 // This is a reference to a register class that doesn't directly correspond
4661 // to an LLVM register class. Allocate NumRegs consecutive, available,
4662 // registers from the class.
4663 RegClassRegs = TLI.getRegClassForInlineAsmConstraint(OpInfo.ConstraintCode,
4664 OpInfo.ConstraintVT);
4667 const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
4668 unsigned NumAllocated = 0;
4669 for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
4670 unsigned Reg = RegClassRegs[i];
4671 // See if this register is available.
4672 if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
4673 (isInReg && InputRegs.count(Reg))) { // Already used.
4674 // Make sure we find consecutive registers.
4679 // Check to see if this register is allocatable (i.e. don't give out the
4682 RC = isAllocatableRegister(Reg, MF, TLI, TRI);
4683 if (!RC) { // Couldn't allocate this register.
4684 // Reset NumAllocated to make sure we return consecutive registers.
4690 // Okay, this register is good, we can use it.
4693 // If we allocated enough consecutive registers, succeed.
4694 if (NumAllocated == NumRegs) {
4695 unsigned RegStart = (i-NumAllocated)+1;
4696 unsigned RegEnd = i+1;
4697 // Mark all of the allocated registers used.
4698 for (unsigned i = RegStart; i != RegEnd; ++i)
4699 Regs.push_back(RegClassRegs[i]);
4701 OpInfo.AssignedRegs = RegsForValue(TLI, Regs, *RC->vt_begin(),
4702 OpInfo.ConstraintVT);
4703 OpInfo.MarkAllocatedRegs(isOutReg, isInReg, OutputRegs, InputRegs, *TRI);
4708 // Otherwise, we couldn't allocate enough registers for this.
4711 /// hasInlineAsmMemConstraint - Return true if the inline asm instruction being
4712 /// processed uses a memory 'm' constraint.
4714 hasInlineAsmMemConstraint(std::vector<InlineAsm::ConstraintInfo> &CInfos,
4715 TargetLowering &TLI) {
4716 for (unsigned i = 0, e = CInfos.size(); i != e; ++i) {
4717 InlineAsm::ConstraintInfo &CI = CInfos[i];
4718 for (unsigned j = 0, ee = CI.Codes.size(); j != ee; ++j) {
4719 TargetLowering::ConstraintType CType = TLI.getConstraintType(CI.Codes[j]);
4720 if (CType == TargetLowering::C_Memory)
4728 /// visitInlineAsm - Handle a call to an InlineAsm object.
4730 void SelectionDAGLowering::visitInlineAsm(CallSite CS) {
4731 InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
4733 /// ConstraintOperands - Information about all of the constraints.
4734 std::vector<SDISelAsmOperandInfo> ConstraintOperands;
4736 SDValue Chain = getRoot();
4739 std::set<unsigned> OutputRegs, InputRegs;
4741 // Do a prepass over the constraints, canonicalizing them, and building up the
4742 // ConstraintOperands list.
4743 std::vector<InlineAsm::ConstraintInfo>
4744 ConstraintInfos = IA->ParseConstraints();
4746 bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
4748 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
4749 unsigned ResNo = 0; // ResNo - The result number of the next output.
4750 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
4751 ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
4752 SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
4754 MVT OpVT = MVT::Other;
4756 // Compute the value type for each operand.
4757 switch (OpInfo.Type) {
4758 case InlineAsm::isOutput:
4759 // Indirect outputs just consume an argument.
4760 if (OpInfo.isIndirect) {
4761 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
4765 // The return value of the call is this value. As such, there is no
4766 // corresponding argument.
4767 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
4768 if (const StructType *STy = dyn_cast<StructType>(CS.getType())) {
4769 OpVT = TLI.getValueType(STy->getElementType(ResNo));
4771 assert(ResNo == 0 && "Asm only has one result!");
4772 OpVT = TLI.getValueType(CS.getType());
4776 case InlineAsm::isInput:
4777 OpInfo.CallOperandVal = CS.getArgument(ArgNo++);
4779 case InlineAsm::isClobber:
4784 // If this is an input or an indirect output, process the call argument.
4785 // BasicBlocks are labels, currently appearing only in asm's.
4786 if (OpInfo.CallOperandVal) {
4787 if (BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
4788 OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
4790 OpInfo.CallOperand = getValue(OpInfo.CallOperandVal);
4793 OpVT = OpInfo.getCallOperandValMVT(TLI, TD);
4796 OpInfo.ConstraintVT = OpVT;
4799 // Second pass over the constraints: compute which constraint option to use
4800 // and assign registers to constraints that want a specific physreg.
4801 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
4802 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
4804 // If this is an output operand with a matching input operand, look up the
4805 // matching input. It might have a different type (e.g. the output might be
4806 // i32 and the input i64) and we need to pick the larger width to ensure we
4807 // reserve the right number of registers.
4808 if (OpInfo.hasMatchingInput()) {
4809 SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
4810 if (OpInfo.ConstraintVT != Input.ConstraintVT) {
4811 assert(OpInfo.ConstraintVT.isInteger() &&
4812 Input.ConstraintVT.isInteger() &&
4813 "Asm constraints must be the same or different sized integers");
4814 if (OpInfo.ConstraintVT.getSizeInBits() <
4815 Input.ConstraintVT.getSizeInBits())
4816 OpInfo.ConstraintVT = Input.ConstraintVT;
4818 Input.ConstraintVT = OpInfo.ConstraintVT;
4822 // Compute the constraint code and ConstraintType to use.
4823 TLI.ComputeConstraintToUse(OpInfo, OpInfo.CallOperand, hasMemory, &DAG);
4825 // If this is a memory input, and if the operand is not indirect, do what we
4826 // need to to provide an address for the memory input.
4827 if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
4828 !OpInfo.isIndirect) {
4829 assert(OpInfo.Type == InlineAsm::isInput &&
4830 "Can only indirectify direct input operands!");
4832 // Memory operands really want the address of the value. If we don't have
4833 // an indirect input, put it in the constpool if we can, otherwise spill
4834 // it to a stack slot.
4836 // If the operand is a float, integer, or vector constant, spill to a
4837 // constant pool entry to get its address.
4838 Value *OpVal = OpInfo.CallOperandVal;
4839 if (isa<ConstantFP>(OpVal) || isa<ConstantInt>(OpVal) ||
4840 isa<ConstantVector>(OpVal)) {
4841 OpInfo.CallOperand = DAG.getConstantPool(cast<Constant>(OpVal),
4842 TLI.getPointerTy());
4844 // Otherwise, create a stack slot and emit a store to it before the
4846 const Type *Ty = OpVal->getType();
4847 uint64_t TySize = TLI.getTargetData()->getABITypeSize(Ty);
4848 unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(Ty);
4849 MachineFunction &MF = DAG.getMachineFunction();
4850 int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align);
4851 SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
4852 Chain = DAG.getStore(Chain, OpInfo.CallOperand, StackSlot, NULL, 0);
4853 OpInfo.CallOperand = StackSlot;
4856 // There is no longer a Value* corresponding to this operand.
4857 OpInfo.CallOperandVal = 0;
4858 // It is now an indirect operand.
4859 OpInfo.isIndirect = true;
4862 // If this constraint is for a specific register, allocate it before
4864 if (OpInfo.ConstraintType == TargetLowering::C_Register)
4865 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
4867 ConstraintInfos.clear();
4870 // Second pass - Loop over all of the operands, assigning virtual or physregs
4871 // to register class operands.
4872 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
4873 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
4875 // C_Register operands have already been allocated, Other/Memory don't need
4877 if (OpInfo.ConstraintType == TargetLowering::C_RegisterClass)
4878 GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
4881 // AsmNodeOperands - The operands for the ISD::INLINEASM node.
4882 std::vector<SDValue> AsmNodeOperands;
4883 AsmNodeOperands.push_back(SDValue()); // reserve space for input chain
4884 AsmNodeOperands.push_back(
4885 DAG.getTargetExternalSymbol(IA->getAsmString().c_str(), MVT::Other));
4888 // Loop over all of the inputs, copying the operand values into the
4889 // appropriate registers and processing the output regs.
4890 RegsForValue RetValRegs;
4892 // IndirectStoresToEmit - The set of stores to emit after the inline asm node.
4893 std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
4895 for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
4896 SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
4898 switch (OpInfo.Type) {
4899 case InlineAsm::isOutput: {
4900 if (OpInfo.ConstraintType != TargetLowering::C_RegisterClass &&
4901 OpInfo.ConstraintType != TargetLowering::C_Register) {
4902 // Memory output, or 'other' output (e.g. 'X' constraint).
4903 assert(OpInfo.isIndirect && "Memory output must be indirect operand");
4905 // Add information to the INLINEASM node to know about this output.
4906 unsigned ResOpType = 4/*MEM*/ | (1<<3);
4907 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
4908 TLI.getPointerTy()));
4909 AsmNodeOperands.push_back(OpInfo.CallOperand);
4913 // Otherwise, this is a register or register class output.
4915 // Copy the output from the appropriate register. Find a register that
4917 if (OpInfo.AssignedRegs.Regs.empty()) {
4918 cerr << "Couldn't allocate output reg for constraint '"
4919 << OpInfo.ConstraintCode << "'!\n";
4923 // If this is an indirect operand, store through the pointer after the
4925 if (OpInfo.isIndirect) {
4926 IndirectStoresToEmit.push_back(std::make_pair(OpInfo.AssignedRegs,
4927 OpInfo.CallOperandVal));
4929 // This is the result value of the call.
4930 assert(CS.getType() != Type::VoidTy && "Bad inline asm!");
4931 // Concatenate this output onto the outputs list.
4932 RetValRegs.append(OpInfo.AssignedRegs);
4935 // Add information to the INLINEASM node to know that this register is
4937 OpInfo.AssignedRegs.AddInlineAsmOperands(OpInfo.isEarlyClobber ?
4938 6 /* EARLYCLOBBER REGDEF */ :
4940 DAG, AsmNodeOperands);
4943 case InlineAsm::isInput: {
4944 SDValue InOperandVal = OpInfo.CallOperand;
4946 if (OpInfo.isMatchingInputConstraint()) { // Matching constraint?
4947 // If this is required to match an output register we have already set,
4948 // just use its register.
4949 unsigned OperandNo = OpInfo.getMatchedOperand();
4951 // Scan until we find the definition we already emitted of this operand.
4952 // When we find it, create a RegsForValue operand.
4953 unsigned CurOp = 2; // The first operand.
4954 for (; OperandNo; --OperandNo) {
4955 // Advance to the next operand.
4957 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
4958 assert(((NumOps & 7) == 2 /*REGDEF*/ ||
4959 (NumOps & 7) == 6 /*EARLYCLOBBER REGDEF*/ ||
4960 (NumOps & 7) == 4 /*MEM*/) &&
4961 "Skipped past definitions?");
4962 CurOp += (NumOps>>3)+1;
4966 cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getZExtValue();
4967 if ((NumOps & 7) == 2 /*REGDEF*/
4968 || (NumOps & 7) == 6 /* EARLYCLOBBER REGDEF */) {
4969 // Add NumOps>>3 registers to MatchedRegs.
4970 RegsForValue MatchedRegs;
4971 MatchedRegs.TLI = &TLI;
4972 MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
4973 MatchedRegs.RegVTs.push_back(AsmNodeOperands[CurOp+1].getValueType());
4974 for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
4976 cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
4977 MatchedRegs.Regs.push_back(Reg);
4980 // Use the produced MatchedRegs object to
4981 MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
4982 MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
4985 assert(((NumOps & 7) == 4) && "Unknown matching constraint!");
4986 assert((NumOps >> 3) == 1 && "Unexpected number of operands");
4987 // Add information to the INLINEASM node to know about this input.
4988 AsmNodeOperands.push_back(DAG.getTargetConstant(NumOps,
4989 TLI.getPointerTy()));
4990 AsmNodeOperands.push_back(AsmNodeOperands[CurOp+1]);
4995 if (OpInfo.ConstraintType == TargetLowering::C_Other) {
4996 assert(!OpInfo.isIndirect &&
4997 "Don't know how to handle indirect other inputs yet!");
4999 std::vector<SDValue> Ops;
5000 TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
5001 hasMemory, Ops, DAG);
5003 cerr << "Invalid operand for inline asm constraint '"
5004 << OpInfo.ConstraintCode << "'!\n";
5008 // Add information to the INLINEASM node to know about this input.
5009 unsigned ResOpType = 3 /*IMM*/ | (Ops.size() << 3);
5010 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5011 TLI.getPointerTy()));
5012 AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
5014 } else if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
5015 assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
5016 assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
5017 "Memory operands expect pointer values");
5019 // Add information to the INLINEASM node to know about this input.
5020 unsigned ResOpType = 4/*MEM*/ | (1<<3);
5021 AsmNodeOperands.push_back(DAG.getTargetConstant(ResOpType,
5022 TLI.getPointerTy()));
5023 AsmNodeOperands.push_back(InOperandVal);
5027 assert((OpInfo.ConstraintType == TargetLowering::C_RegisterClass ||
5028 OpInfo.ConstraintType == TargetLowering::C_Register) &&
5029 "Unknown constraint type!");
5030 assert(!OpInfo.isIndirect &&
5031 "Don't know how to handle indirect register inputs yet!");
5033 // Copy the input into the appropriate registers.
5034 if (OpInfo.AssignedRegs.Regs.empty()) {
5035 cerr << "Couldn't allocate output reg for constraint '"
5036 << OpInfo.ConstraintCode << "'!\n";
5040 OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, Chain, &Flag);
5042 OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/,
5043 DAG, AsmNodeOperands);
5046 case InlineAsm::isClobber: {
5047 // Add the clobbered value to the operand list, so that the register
5048 // allocator is aware that the physreg got clobbered.
5049 if (!OpInfo.AssignedRegs.Regs.empty())
5050 OpInfo.AssignedRegs.AddInlineAsmOperands(6 /* EARLYCLOBBER REGDEF */,
5051 DAG, AsmNodeOperands);
5057 // Finish up input operands.
5058 AsmNodeOperands[0] = Chain;
5059 if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
5061 Chain = DAG.getNode(ISD::INLINEASM,
5062 DAG.getNodeValueTypes(MVT::Other, MVT::Flag), 2,
5063 &AsmNodeOperands[0], AsmNodeOperands.size());
5064 Flag = Chain.getValue(1);
5066 // If this asm returns a register value, copy the result from that register
5067 // and set it as the value of the call.
5068 if (!RetValRegs.Regs.empty()) {
5069 SDValue Val = RetValRegs.getCopyFromRegs(DAG, Chain, &Flag);
5071 // FIXME: Why don't we do this for inline asms with MRVs?
5072 if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) {
5073 MVT ResultType = TLI.getValueType(CS.getType());
5075 // If any of the results of the inline asm is a vector, it may have the
5076 // wrong width/num elts. This can happen for register classes that can
5077 // contain multiple different value types. The preg or vreg allocated may
5078 // not have the same VT as was expected. Convert it to the right type
5079 // with bit_convert.
5080 if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
5081 Val = DAG.getNode(ISD::BIT_CONVERT, ResultType, Val);
5083 } else if (ResultType != Val.getValueType() &&
5084 ResultType.isInteger() && Val.getValueType().isInteger()) {
5085 // If a result value was tied to an input value, the computed result may
5086 // have a wider width than the expected result. Extract the relevant
5088 Val = DAG.getNode(ISD::TRUNCATE, ResultType, Val);
5091 assert(ResultType == Val.getValueType() && "Asm result value mismatch!");
5094 setValue(CS.getInstruction(), Val);
5097 std::vector<std::pair<SDValue, Value*> > StoresToEmit;
5099 // Process indirect outputs, first output all of the flagged copies out of
5101 for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
5102 RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
5103 Value *Ptr = IndirectStoresToEmit[i].second;
5104 SDValue OutVal = OutRegs.getCopyFromRegs(DAG, Chain, &Flag);
5105 StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
5108 // Emit the non-flagged stores from the physregs.
5109 SmallVector<SDValue, 8> OutChains;
5110 for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
5111 OutChains.push_back(DAG.getStore(Chain, StoresToEmit[i].first,
5112 getValue(StoresToEmit[i].second),
5113 StoresToEmit[i].second, 0));
5114 if (!OutChains.empty())
5115 Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
5116 &OutChains[0], OutChains.size());
5121 void SelectionDAGLowering::visitMalloc(MallocInst &I) {
5122 SDValue Src = getValue(I.getOperand(0));
5124 MVT IntPtr = TLI.getPointerTy();
5126 if (IntPtr.bitsLT(Src.getValueType()))
5127 Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
5128 else if (IntPtr.bitsGT(Src.getValueType()))
5129 Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
5131 // Scale the source by the type size.
5132 uint64_t ElementSize = TD->getABITypeSize(I.getType()->getElementType());
5133 Src = DAG.getNode(ISD::MUL, Src.getValueType(),
5134 Src, DAG.getIntPtrConstant(ElementSize));
5136 TargetLowering::ArgListTy Args;
5137 TargetLowering::ArgListEntry Entry;
5139 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5140 Args.push_back(Entry);
5142 std::pair<SDValue,SDValue> Result =
5143 TLI.LowerCallTo(getRoot(), I.getType(), false, false, false, false,
5144 CallingConv::C, PerformTailCallOpt,
5145 DAG.getExternalSymbol("malloc", IntPtr),
5147 setValue(&I, Result.first); // Pointers always fit in registers
5148 DAG.setRoot(Result.second);
5151 void SelectionDAGLowering::visitFree(FreeInst &I) {
5152 TargetLowering::ArgListTy Args;
5153 TargetLowering::ArgListEntry Entry;
5154 Entry.Node = getValue(I.getOperand(0));
5155 Entry.Ty = TLI.getTargetData()->getIntPtrType();
5156 Args.push_back(Entry);
5157 MVT IntPtr = TLI.getPointerTy();
5158 std::pair<SDValue,SDValue> Result =
5159 TLI.LowerCallTo(getRoot(), Type::VoidTy, false, false, false, false,
5160 CallingConv::C, PerformTailCallOpt,
5161 DAG.getExternalSymbol("free", IntPtr), Args, DAG);
5162 DAG.setRoot(Result.second);
5165 void SelectionDAGLowering::visitVAStart(CallInst &I) {
5166 DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
5167 getValue(I.getOperand(1)),
5168 DAG.getSrcValue(I.getOperand(1))));
5171 void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
5172 SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
5173 getValue(I.getOperand(0)),
5174 DAG.getSrcValue(I.getOperand(0)));
5176 DAG.setRoot(V.getValue(1));
5179 void SelectionDAGLowering::visitVAEnd(CallInst &I) {
5180 DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
5181 getValue(I.getOperand(1)),
5182 DAG.getSrcValue(I.getOperand(1))));
5185 void SelectionDAGLowering::visitVACopy(CallInst &I) {
5186 DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
5187 getValue(I.getOperand(1)),
5188 getValue(I.getOperand(2)),
5189 DAG.getSrcValue(I.getOperand(1)),
5190 DAG.getSrcValue(I.getOperand(2))));
5193 /// TargetLowering::LowerArguments - This is the default LowerArguments
5194 /// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
5195 /// targets are migrated to using FORMAL_ARGUMENTS, this hook should be
5196 /// integrated into SDISel.
5197 void TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG,
5198 SmallVectorImpl<SDValue> &ArgValues) {
5199 // Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
5200 SmallVector<SDValue, 3+16> Ops;
5201 Ops.push_back(DAG.getRoot());
5202 Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
5203 Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
5205 // Add one result value for each formal argument.
5206 SmallVector<MVT, 16> RetVals;
5208 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end();
5210 SmallVector<MVT, 4> ValueVTs;
5211 ComputeValueVTs(*this, I->getType(), ValueVTs);
5212 for (unsigned Value = 0, NumValues = ValueVTs.size();
5213 Value != NumValues; ++Value) {
5214 MVT VT = ValueVTs[Value];
5215 const Type *ArgTy = VT.getTypeForMVT();
5216 ISD::ArgFlagsTy Flags;
5217 unsigned OriginalAlignment =
5218 getTargetData()->getABITypeAlignment(ArgTy);
5220 if (F.paramHasAttr(j, Attribute::ZExt))
5222 if (F.paramHasAttr(j, Attribute::SExt))
5224 if (F.paramHasAttr(j, Attribute::InReg))
5226 if (F.paramHasAttr(j, Attribute::StructRet))
5228 if (F.paramHasAttr(j, Attribute::ByVal)) {
5230 const PointerType *Ty = cast<PointerType>(I->getType());
5231 const Type *ElementTy = Ty->getElementType();
5232 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5233 unsigned FrameSize = getTargetData()->getABITypeSize(ElementTy);
5234 // For ByVal, alignment should be passed from FE. BE will guess if
5235 // this info is not there but there are cases it cannot get right.
5236 if (F.getParamAlignment(j))
5237 FrameAlign = F.getParamAlignment(j);
5238 Flags.setByValAlign(FrameAlign);
5239 Flags.setByValSize(FrameSize);
5241 if (F.paramHasAttr(j, Attribute::Nest))
5243 Flags.setOrigAlign(OriginalAlignment);
5245 MVT RegisterVT = getRegisterType(VT);
5246 unsigned NumRegs = getNumRegisters(VT);
5247 for (unsigned i = 0; i != NumRegs; ++i) {
5248 RetVals.push_back(RegisterVT);
5249 ISD::ArgFlagsTy MyFlags = Flags;
5250 if (NumRegs > 1 && i == 0)
5252 // if it isn't first piece, alignment must be 1
5254 MyFlags.setOrigAlign(1);
5255 Ops.push_back(DAG.getArgFlags(MyFlags));
5260 RetVals.push_back(MVT::Other);
5263 SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS,
5264 DAG.getVTList(&RetVals[0], RetVals.size()),
5265 &Ops[0], Ops.size()).getNode();
5267 // Prelower FORMAL_ARGUMENTS. This isn't required for functionality, but
5268 // allows exposing the loads that may be part of the argument access to the
5269 // first DAGCombiner pass.
5270 SDValue TmpRes = LowerOperation(SDValue(Result, 0), DAG);
5272 // The number of results should match up, except that the lowered one may have
5273 // an extra flag result.
5274 assert((Result->getNumValues() == TmpRes.getNode()->getNumValues() ||
5275 (Result->getNumValues()+1 == TmpRes.getNode()->getNumValues() &&
5276 TmpRes.getValue(Result->getNumValues()).getValueType() == MVT::Flag))
5277 && "Lowering produced unexpected number of results!");
5279 // The FORMAL_ARGUMENTS node itself is likely no longer needed.
5280 if (Result != TmpRes.getNode() && Result->use_empty()) {
5281 HandleSDNode Dummy(DAG.getRoot());
5282 DAG.RemoveDeadNode(Result);
5285 Result = TmpRes.getNode();
5287 unsigned NumArgRegs = Result->getNumValues() - 1;
5288 DAG.setRoot(SDValue(Result, NumArgRegs));
5290 // Set up the return result vector.
5293 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E;
5295 SmallVector<MVT, 4> ValueVTs;
5296 ComputeValueVTs(*this, I->getType(), ValueVTs);
5297 for (unsigned Value = 0, NumValues = ValueVTs.size();
5298 Value != NumValues; ++Value) {
5299 MVT VT = ValueVTs[Value];
5300 MVT PartVT = getRegisterType(VT);
5302 unsigned NumParts = getNumRegisters(VT);
5303 SmallVector<SDValue, 4> Parts(NumParts);
5304 for (unsigned j = 0; j != NumParts; ++j)
5305 Parts[j] = SDValue(Result, i++);
5307 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5308 if (F.paramHasAttr(Idx, Attribute::SExt))
5309 AssertOp = ISD::AssertSext;
5310 else if (F.paramHasAttr(Idx, Attribute::ZExt))
5311 AssertOp = ISD::AssertZext;
5313 ArgValues.push_back(getCopyFromParts(DAG, &Parts[0], NumParts, PartVT, VT,
5317 assert(i == NumArgRegs && "Argument register count mismatch!");
5321 /// TargetLowering::LowerCallTo - This is the default LowerCallTo
5322 /// implementation, which just inserts an ISD::CALL node, which is later custom
5323 /// lowered by the target to something concrete. FIXME: When all targets are
5324 /// migrated to using ISD::CALL, this hook should be integrated into SDISel.
5325 std::pair<SDValue, SDValue>
5326 TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy,
5327 bool RetSExt, bool RetZExt, bool isVarArg,
5329 unsigned CallingConv, bool isTailCall,
5331 ArgListTy &Args, SelectionDAG &DAG) {
5332 assert((!isTailCall || PerformTailCallOpt) &&
5333 "isTailCall set when tail-call optimizations are disabled!");
5335 SmallVector<SDValue, 32> Ops;
5336 Ops.push_back(Chain); // Op#0 - Chain
5337 Ops.push_back(Callee);
5339 // Handle all of the outgoing arguments.
5340 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
5341 SmallVector<MVT, 4> ValueVTs;
5342 ComputeValueVTs(*this, Args[i].Ty, ValueVTs);
5343 for (unsigned Value = 0, NumValues = ValueVTs.size();
5344 Value != NumValues; ++Value) {
5345 MVT VT = ValueVTs[Value];
5346 const Type *ArgTy = VT.getTypeForMVT();
5347 SDValue Op = SDValue(Args[i].Node.getNode(),
5348 Args[i].Node.getResNo() + Value);
5349 ISD::ArgFlagsTy Flags;
5350 unsigned OriginalAlignment =
5351 getTargetData()->getABITypeAlignment(ArgTy);
5357 if (Args[i].isInReg)
5361 if (Args[i].isByVal) {
5363 const PointerType *Ty = cast<PointerType>(Args[i].Ty);
5364 const Type *ElementTy = Ty->getElementType();
5365 unsigned FrameAlign = getByValTypeAlignment(ElementTy);
5366 unsigned FrameSize = getTargetData()->getABITypeSize(ElementTy);
5367 // For ByVal, alignment should come from FE. BE will guess if this
5368 // info is not there but there are cases it cannot get right.
5369 if (Args[i].Alignment)
5370 FrameAlign = Args[i].Alignment;
5371 Flags.setByValAlign(FrameAlign);
5372 Flags.setByValSize(FrameSize);
5376 Flags.setOrigAlign(OriginalAlignment);
5378 MVT PartVT = getRegisterType(VT);
5379 unsigned NumParts = getNumRegisters(VT);
5380 SmallVector<SDValue, 4> Parts(NumParts);
5381 ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
5384 ExtendKind = ISD::SIGN_EXTEND;
5385 else if (Args[i].isZExt)
5386 ExtendKind = ISD::ZERO_EXTEND;
5388 getCopyToParts(DAG, Op, &Parts[0], NumParts, PartVT, ExtendKind);
5390 for (unsigned i = 0; i != NumParts; ++i) {
5391 // if it isn't first piece, alignment must be 1
5392 ISD::ArgFlagsTy MyFlags = Flags;
5393 if (NumParts > 1 && i == 0)
5396 MyFlags.setOrigAlign(1);
5398 Ops.push_back(Parts[i]);
5399 Ops.push_back(DAG.getArgFlags(MyFlags));
5404 // Figure out the result value types. We start by making a list of
5405 // the potentially illegal return value types.
5406 SmallVector<MVT, 4> LoweredRetTys;
5407 SmallVector<MVT, 4> RetTys;
5408 ComputeValueVTs(*this, RetTy, RetTys);
5410 // Then we translate that to a list of legal types.
5411 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5413 MVT RegisterVT = getRegisterType(VT);
5414 unsigned NumRegs = getNumRegisters(VT);
5415 for (unsigned i = 0; i != NumRegs; ++i)
5416 LoweredRetTys.push_back(RegisterVT);
5419 LoweredRetTys.push_back(MVT::Other); // Always has a chain.
5421 // Create the CALL node.
5422 SDValue Res = DAG.getCall(CallingConv, isVarArg, isTailCall, isInreg,
5423 DAG.getVTList(&LoweredRetTys[0],
5424 LoweredRetTys.size()),
5427 Chain = Res.getValue(LoweredRetTys.size() - 1);
5429 // Gather up the call result into a single value.
5430 if (RetTy != Type::VoidTy && !RetTys.empty()) {
5431 ISD::NodeType AssertOp = ISD::DELETED_NODE;
5434 AssertOp = ISD::AssertSext;
5436 AssertOp = ISD::AssertZext;
5438 SmallVector<SDValue, 4> ReturnValues;
5440 for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
5442 MVT RegisterVT = getRegisterType(VT);
5443 unsigned NumRegs = getNumRegisters(VT);
5444 unsigned RegNoEnd = NumRegs + RegNo;
5445 SmallVector<SDValue, 4> Results;
5446 for (; RegNo != RegNoEnd; ++RegNo)
5447 Results.push_back(Res.getValue(RegNo));
5448 SDValue ReturnValue =
5449 getCopyFromParts(DAG, &Results[0], NumRegs, RegisterVT, VT,
5451 ReturnValues.push_back(ReturnValue);
5453 Res = DAG.getMergeValues(DAG.getVTList(&RetTys[0], RetTys.size()),
5454 &ReturnValues[0], ReturnValues.size());
5457 return std::make_pair(Res, Chain);
5460 SDValue TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
5461 assert(0 && "LowerOperation not implemented for this target!");
5467 void SelectionDAGLowering::CopyValueToVirtualRegister(Value *V, unsigned Reg) {
5468 SDValue Op = getValue(V);
5469 assert((Op.getOpcode() != ISD::CopyFromReg ||
5470 cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
5471 "Copy from a reg to the same reg!");
5472 assert(!TargetRegisterInfo::isPhysicalRegister(Reg) && "Is a physreg");
5474 RegsForValue RFV(TLI, Reg, V->getType());
5475 SDValue Chain = DAG.getEntryNode();
5476 RFV.getCopyToRegs(Op, DAG, Chain, 0);
5477 PendingExports.push_back(Chain);
5480 #include "llvm/CodeGen/SelectionDAGISel.h"
5482 void SelectionDAGISel::
5483 LowerArguments(BasicBlock *LLVMBB) {
5484 // If this is the entry block, emit arguments.
5485 Function &F = *LLVMBB->getParent();
5486 SDValue OldRoot = SDL->DAG.getRoot();
5487 SmallVector<SDValue, 16> Args;
5488 TLI.LowerArguments(F, SDL->DAG, Args);
5491 for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
5493 SmallVector<MVT, 4> ValueVTs;
5494 ComputeValueVTs(TLI, AI->getType(), ValueVTs);
5495 unsigned NumValues = ValueVTs.size();
5496 if (!AI->use_empty()) {
5497 SDL->setValue(AI, SDL->DAG.getMergeValues(&Args[a], NumValues));
5498 // If this argument is live outside of the entry block, insert a copy from
5499 // whereever we got it to the vreg that other BB's will reference it as.
5500 DenseMap<const Value*, unsigned>::iterator VMI=FuncInfo->ValueMap.find(AI);
5501 if (VMI != FuncInfo->ValueMap.end()) {
5502 SDL->CopyValueToVirtualRegister(AI, VMI->second);
5508 // Finally, if the target has anything special to do, allow it to do so.
5509 // FIXME: this should insert code into the DAG!
5510 EmitFunctionEntryCode(F, SDL->DAG.getMachineFunction());
5513 /// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
5514 /// ensure constants are generated when needed. Remember the virtual registers
5515 /// that need to be added to the Machine PHI nodes as input. We cannot just
5516 /// directly add them, because expansion might result in multiple MBB's for one
5517 /// BB. As such, the start of the BB might correspond to a different MBB than
5521 SelectionDAGISel::HandlePHINodesInSuccessorBlocks(BasicBlock *LLVMBB) {
5522 TerminatorInst *TI = LLVMBB->getTerminator();
5524 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5526 // Check successor nodes' PHI nodes that expect a constant to be available
5528 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5529 BasicBlock *SuccBB = TI->getSuccessor(succ);
5530 if (!isa<PHINode>(SuccBB->begin())) continue;
5531 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5533 // If this terminator has multiple identical successors (common for
5534 // switches), only handle each succ once.
5535 if (!SuccsHandled.insert(SuccMBB)) continue;
5537 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5540 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5541 // nodes and Machine PHI nodes, but the incoming operands have not been
5543 for (BasicBlock::iterator I = SuccBB->begin();
5544 (PN = dyn_cast<PHINode>(I)); ++I) {
5545 // Ignore dead phi's.
5546 if (PN->use_empty()) continue;
5549 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5551 if (Constant *C = dyn_cast<Constant>(PHIOp)) {
5552 unsigned &RegOut = SDL->ConstantsOut[C];
5554 RegOut = FuncInfo->CreateRegForValue(C);
5555 SDL->CopyValueToVirtualRegister(C, RegOut);
5559 Reg = FuncInfo->ValueMap[PHIOp];
5561 assert(isa<AllocaInst>(PHIOp) &&
5562 FuncInfo->StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
5563 "Didn't codegen value into a register!??");
5564 Reg = FuncInfo->CreateRegForValue(PHIOp);
5565 SDL->CopyValueToVirtualRegister(PHIOp, Reg);
5569 // Remember that this register needs to added to the machine PHI node as
5570 // the input for this MBB.
5571 SmallVector<MVT, 4> ValueVTs;
5572 ComputeValueVTs(TLI, PN->getType(), ValueVTs);
5573 for (unsigned vti = 0, vte = ValueVTs.size(); vti != vte; ++vti) {
5574 MVT VT = ValueVTs[vti];
5575 unsigned NumRegisters = TLI.getNumRegisters(VT);
5576 for (unsigned i = 0, e = NumRegisters; i != e; ++i)
5577 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
5578 Reg += NumRegisters;
5582 SDL->ConstantsOut.clear();
5585 /// This is the Fast-ISel version of HandlePHINodesInSuccessorBlocks. It only
5586 /// supports legal types, and it emits MachineInstrs directly instead of
5587 /// creating SelectionDAG nodes.
5590 SelectionDAGISel::HandlePHINodesInSuccessorBlocksFast(BasicBlock *LLVMBB,
5592 TerminatorInst *TI = LLVMBB->getTerminator();
5594 SmallPtrSet<MachineBasicBlock *, 4> SuccsHandled;
5595 unsigned OrigNumPHINodesToUpdate = SDL->PHINodesToUpdate.size();
5597 // Check successor nodes' PHI nodes that expect a constant to be available
5599 for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
5600 BasicBlock *SuccBB = TI->getSuccessor(succ);
5601 if (!isa<PHINode>(SuccBB->begin())) continue;
5602 MachineBasicBlock *SuccMBB = FuncInfo->MBBMap[SuccBB];
5604 // If this terminator has multiple identical successors (common for
5605 // switches), only handle each succ once.
5606 if (!SuccsHandled.insert(SuccMBB)) continue;
5608 MachineBasicBlock::iterator MBBI = SuccMBB->begin();
5611 // At this point we know that there is a 1-1 correspondence between LLVM PHI
5612 // nodes and Machine PHI nodes, but the incoming operands have not been
5614 for (BasicBlock::iterator I = SuccBB->begin();
5615 (PN = dyn_cast<PHINode>(I)); ++I) {
5616 // Ignore dead phi's.
5617 if (PN->use_empty()) continue;
5619 // Only handle legal types. Two interesting things to note here. First,
5620 // by bailing out early, we may leave behind some dead instructions,
5621 // since SelectionDAG's HandlePHINodesInSuccessorBlocks will insert its
5622 // own moves. Second, this check is necessary becuase FastISel doesn't
5623 // use CreateRegForValue to create registers, so it always creates
5624 // exactly one register for each non-void instruction.
5625 MVT VT = TLI.getValueType(PN->getType(), /*AllowUnknown=*/true);
5626 if (VT == MVT::Other || !TLI.isTypeLegal(VT)) {
5629 VT = TLI.getTypeToTransformTo(VT);
5631 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5636 Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
5638 unsigned Reg = F->getRegForValue(PHIOp);
5640 SDL->PHINodesToUpdate.resize(OrigNumPHINodesToUpdate);
5643 SDL->PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg));