1 //===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
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 file implements the SelectionDAG::Legalize method.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DebugInfo.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetFrameLowering.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetSubtargetInfo.h"
40 #define DEBUG_TYPE "legalizedag"
44 struct FloatSignAsInt;
46 //===----------------------------------------------------------------------===//
47 /// This takes an arbitrary SelectionDAG as input and
48 /// hacks on it until the target machine can handle it. This involves
49 /// eliminating value sizes the machine cannot handle (promoting small sizes to
50 /// large sizes or splitting up large values into small values) as well as
51 /// eliminating operations the machine cannot handle.
53 /// This code also does a small amount of optimization and recognition of idioms
54 /// as part of its processing. For example, if a target does not support a
55 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
56 /// will attempt merge setcc and brc instructions into brcc's.
58 class SelectionDAGLegalize {
59 const TargetMachine &TM;
60 const TargetLowering &TLI;
63 /// \brief The set of nodes which have already been legalized. We hold a
64 /// reference to it in order to update as necessary on node deletion.
65 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
67 /// \brief A set of all the nodes updated during legalization.
68 SmallSetVector<SDNode *, 16> *UpdatedNodes;
70 EVT getSetCCResultType(EVT VT) const {
71 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
74 // Libcall insertion helpers.
77 SelectionDAGLegalize(SelectionDAG &DAG,
78 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
79 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
80 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
81 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {}
83 /// \brief Legalizes the given operation.
84 void LegalizeOp(SDNode *Node);
87 SDValue OptimizeFloatStore(StoreSDNode *ST);
89 void LegalizeLoadOps(SDNode *Node);
90 void LegalizeStoreOps(SDNode *Node);
92 /// Some targets cannot handle a variable
93 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
94 /// is necessary to spill the vector being inserted into to memory, perform
95 /// the insert there, and then read the result back.
96 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
97 SDValue Idx, SDLoc dl);
98 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
99 SDValue Idx, SDLoc dl);
101 /// Return a vector shuffle operation which
102 /// performs the same shuffe in terms of order or result bytes, but on a type
103 /// whose vector element type is narrower than the original shuffle type.
104 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
105 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
106 SDValue N1, SDValue N2,
107 ArrayRef<int> Mask) const;
109 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
110 bool &NeedInvert, SDLoc dl);
112 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
113 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
114 unsigned NumOps, bool isSigned, SDLoc dl);
116 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
117 SDNode *Node, bool isSigned);
118 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
119 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
120 RTLIB::Libcall Call_F128,
121 RTLIB::Libcall Call_PPCF128);
122 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
123 RTLIB::Libcall Call_I8,
124 RTLIB::Libcall Call_I16,
125 RTLIB::Libcall Call_I32,
126 RTLIB::Libcall Call_I64,
127 RTLIB::Libcall Call_I128);
128 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
129 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
131 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
132 SDValue ExpandBUILD_VECTOR(SDNode *Node);
133 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
134 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
135 SmallVectorImpl<SDValue> &Results);
136 void getSignAsIntValue(FloatSignAsInt &State, SDLoc DL, SDValue Value) const;
137 SDValue modifySignAsInt(const FloatSignAsInt &State, SDLoc DL,
138 SDValue NewIntValue) const;
139 SDValue ExpandFCOPYSIGN(SDNode *Node) const;
140 SDValue ExpandFABS(SDNode *Node) const;
141 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
143 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
145 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
148 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
149 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
151 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
152 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
153 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
155 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
157 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
159 void ExpandNode(SDNode *Node);
160 void PromoteNode(SDNode *Node);
163 // Node replacement helpers
164 void ReplacedNode(SDNode *N) {
165 LegalizedNodes.erase(N);
167 UpdatedNodes->insert(N);
169 void ReplaceNode(SDNode *Old, SDNode *New) {
170 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
171 dbgs() << " with: "; New->dump(&DAG));
173 assert(Old->getNumValues() == New->getNumValues() &&
174 "Replacing one node with another that produces a different number "
176 DAG.ReplaceAllUsesWith(Old, New);
177 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
178 DAG.TransferDbgValues(SDValue(Old, i), SDValue(New, i));
180 UpdatedNodes->insert(New);
183 void ReplaceNode(SDValue Old, SDValue New) {
184 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
185 dbgs() << " with: "; New->dump(&DAG));
187 DAG.ReplaceAllUsesWith(Old, New);
188 DAG.TransferDbgValues(Old, New);
190 UpdatedNodes->insert(New.getNode());
191 ReplacedNode(Old.getNode());
193 void ReplaceNode(SDNode *Old, const SDValue *New) {
194 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
196 DAG.ReplaceAllUsesWith(Old, New);
197 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
198 DEBUG(dbgs() << (i == 0 ? " with: "
201 DAG.TransferDbgValues(SDValue(Old, i), New[i]);
203 UpdatedNodes->insert(New[i].getNode());
210 /// Return a vector shuffle operation which
211 /// performs the same shuffe in terms of order or result bytes, but on a type
212 /// whose vector element type is narrower than the original shuffle type.
213 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
215 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
216 SDValue N1, SDValue N2,
217 ArrayRef<int> Mask) const {
218 unsigned NumMaskElts = VT.getVectorNumElements();
219 unsigned NumDestElts = NVT.getVectorNumElements();
220 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
222 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
224 if (NumEltsGrowth == 1)
225 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
227 SmallVector<int, 8> NewMask;
228 for (unsigned i = 0; i != NumMaskElts; ++i) {
230 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
232 NewMask.push_back(-1);
234 NewMask.push_back(Idx * NumEltsGrowth + j);
237 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
238 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
239 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
242 /// Expands the ConstantFP node to an integer constant or
243 /// a load from the constant pool.
245 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
249 // If a FP immediate is precise when represented as a float and if the
250 // target can do an extending load from float to double, we put it into
251 // the constant pool as a float, even if it's is statically typed as a
252 // double. This shrinks FP constants and canonicalizes them for targets where
253 // an FP extending load is the same cost as a normal load (such as on the x87
254 // fp stack or PPC FP unit).
255 EVT VT = CFP->getValueType(0);
256 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
258 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
259 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(), dl,
260 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
265 while (SVT != MVT::f32 && SVT != MVT::f16) {
266 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
267 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
268 // Only do this if the target has a native EXTLOAD instruction from
270 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) &&
271 TLI.ShouldShrinkFPConstant(OrigVT)) {
272 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
273 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
280 DAG.getConstantPool(LLVMC, TLI.getPointerTy(DAG.getDataLayout()));
281 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
284 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
286 CPIdx, MachinePointerInfo::getConstantPool(),
287 VT, false, false, false, Alignment);
291 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
292 MachinePointerInfo::getConstantPool(), false, false, false,
297 /// Expands an unaligned store to 2 half-size stores.
298 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
299 const TargetLowering &TLI,
300 SelectionDAGLegalize *DAGLegalize) {
301 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
302 "unaligned indexed stores not implemented!");
303 SDValue Chain = ST->getChain();
304 SDValue Ptr = ST->getBasePtr();
305 SDValue Val = ST->getValue();
306 EVT VT = Val.getValueType();
307 int Alignment = ST->getAlignment();
308 unsigned AS = ST->getAddressSpace();
311 if (ST->getMemoryVT().isFloatingPoint() ||
312 ST->getMemoryVT().isVector()) {
313 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
314 if (TLI.isTypeLegal(intVT)) {
315 // Expand to a bitconvert of the value to the integer type of the
316 // same size, then a (misaligned) int store.
317 // FIXME: Does not handle truncating floating point stores!
318 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
319 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
320 ST->isVolatile(), ST->isNonTemporal(), Alignment);
321 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
324 // Do a (aligned) store to a stack slot, then copy from the stack slot
325 // to the final destination using (unaligned) integer loads and stores.
326 EVT StoredVT = ST->getMemoryVT();
328 TLI.getRegisterType(*DAG.getContext(),
329 EVT::getIntegerVT(*DAG.getContext(),
330 StoredVT.getSizeInBits()));
331 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
332 unsigned RegBytes = RegVT.getSizeInBits() / 8;
333 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
335 // Make sure the stack slot is also aligned for the register type.
336 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
338 // Perform the original store, only redirected to the stack slot.
339 SDValue Store = DAG.getTruncStore(Chain, dl,
340 Val, StackPtr, MachinePointerInfo(),
341 StoredVT, false, false, 0);
342 SDValue Increment = DAG.getConstant(
343 RegBytes, dl, TLI.getPointerTy(DAG.getDataLayout(), AS));
344 SmallVector<SDValue, 8> Stores;
347 // Do all but one copies using the full register width.
348 for (unsigned i = 1; i < NumRegs; i++) {
349 // Load one integer register's worth from the stack slot.
350 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
351 MachinePointerInfo(),
352 false, false, false, 0);
353 // Store it to the final location. Remember the store.
354 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
355 ST->getPointerInfo().getWithOffset(Offset),
356 ST->isVolatile(), ST->isNonTemporal(),
357 MinAlign(ST->getAlignment(), Offset)));
358 // Increment the pointers.
360 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
362 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
365 // The last store may be partial. Do a truncating store. On big-endian
366 // machines this requires an extending load from the stack slot to ensure
367 // that the bits are in the right place.
368 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
369 8 * (StoredBytes - Offset));
371 // Load from the stack slot.
372 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
373 MachinePointerInfo(),
374 MemVT, false, false, false, 0);
376 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
378 .getWithOffset(Offset),
379 MemVT, ST->isVolatile(),
381 MinAlign(ST->getAlignment(), Offset),
383 // The order of the stores doesn't matter - say it with a TokenFactor.
384 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
385 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
388 assert(ST->getMemoryVT().isInteger() &&
389 !ST->getMemoryVT().isVector() &&
390 "Unaligned store of unknown type.");
391 // Get the half-size VT
392 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
393 int NumBits = NewStoredVT.getSizeInBits();
394 int IncrementSize = NumBits / 8;
396 // Divide the stored value in two parts.
397 SDValue ShiftAmount =
398 DAG.getConstant(NumBits, dl, TLI.getShiftAmountTy(Val.getValueType(),
399 DAG.getDataLayout()));
401 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
403 // Store the two parts
404 SDValue Store1, Store2;
405 Store1 = DAG.getTruncStore(Chain, dl,
406 DAG.getDataLayout().isLittleEndian() ? Lo : Hi,
407 Ptr, ST->getPointerInfo(), NewStoredVT,
408 ST->isVolatile(), ST->isNonTemporal(), Alignment);
410 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
411 DAG.getConstant(IncrementSize, dl,
412 TLI.getPointerTy(DAG.getDataLayout(), AS)));
413 Alignment = MinAlign(Alignment, IncrementSize);
414 Store2 = DAG.getTruncStore(
415 Chain, dl, DAG.getDataLayout().isLittleEndian() ? Hi : Lo, Ptr,
416 ST->getPointerInfo().getWithOffset(IncrementSize), NewStoredVT,
417 ST->isVolatile(), ST->isNonTemporal(), Alignment, ST->getAAInfo());
420 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
421 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
424 /// Expands an unaligned load to 2 half-size loads.
426 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
427 const TargetLowering &TLI,
428 SDValue &ValResult, SDValue &ChainResult) {
429 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
430 "unaligned indexed loads not implemented!");
431 SDValue Chain = LD->getChain();
432 SDValue Ptr = LD->getBasePtr();
433 EVT VT = LD->getValueType(0);
434 EVT LoadedVT = LD->getMemoryVT();
436 if (VT.isFloatingPoint() || VT.isVector()) {
437 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
438 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
439 // Expand to a (misaligned) integer load of the same size,
440 // then bitconvert to floating point or vector.
441 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
442 LD->getMemOperand());
443 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
445 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
446 ISD::ANY_EXTEND, dl, VT, Result);
453 // Copy the value to a (aligned) stack slot using (unaligned) integer
454 // loads and stores, then do a (aligned) load from the stack slot.
455 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
456 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
457 unsigned RegBytes = RegVT.getSizeInBits() / 8;
458 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
460 // Make sure the stack slot is also aligned for the register type.
461 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
464 DAG.getConstant(RegBytes, dl, TLI.getPointerTy(DAG.getDataLayout()));
465 SmallVector<SDValue, 8> Stores;
466 SDValue StackPtr = StackBase;
469 // Do all but one copies using the full register width.
470 for (unsigned i = 1; i < NumRegs; i++) {
471 // Load one integer register's worth from the original location.
472 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
473 LD->getPointerInfo().getWithOffset(Offset),
474 LD->isVolatile(), LD->isNonTemporal(),
476 MinAlign(LD->getAlignment(), Offset),
478 // Follow the load with a store to the stack slot. Remember the store.
479 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
480 MachinePointerInfo(), false, false, 0));
481 // Increment the pointers.
483 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
484 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
488 // The last copy may be partial. Do an extending load.
489 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
490 8 * (LoadedBytes - Offset));
491 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
492 LD->getPointerInfo().getWithOffset(Offset),
493 MemVT, LD->isVolatile(),
496 MinAlign(LD->getAlignment(), Offset),
498 // Follow the load with a store to the stack slot. Remember the store.
499 // On big-endian machines this requires a truncating store to ensure
500 // that the bits end up in the right place.
501 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
502 MachinePointerInfo(), MemVT,
505 // The order of the stores doesn't matter - say it with a TokenFactor.
506 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
508 // Finally, perform the original load only redirected to the stack slot.
509 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
510 MachinePointerInfo(), LoadedVT, false,false, false,
513 // Callers expect a MERGE_VALUES node.
518 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
519 "Unaligned load of unsupported type.");
521 // Compute the new VT that is half the size of the old one. This is an
523 unsigned NumBits = LoadedVT.getSizeInBits();
525 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
528 unsigned Alignment = LD->getAlignment();
529 unsigned IncrementSize = NumBits / 8;
530 ISD::LoadExtType HiExtType = LD->getExtensionType();
532 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
533 if (HiExtType == ISD::NON_EXTLOAD)
534 HiExtType = ISD::ZEXTLOAD;
536 // Load the value in two parts
538 if (DAG.getDataLayout().isLittleEndian()) {
539 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
540 NewLoadedVT, LD->isVolatile(),
541 LD->isNonTemporal(), LD->isInvariant(), Alignment,
543 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
544 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
545 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
546 LD->getPointerInfo().getWithOffset(IncrementSize),
547 NewLoadedVT, LD->isVolatile(),
548 LD->isNonTemporal(),LD->isInvariant(),
549 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
551 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
552 NewLoadedVT, LD->isVolatile(),
553 LD->isNonTemporal(), LD->isInvariant(), Alignment,
555 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
556 DAG.getConstant(IncrementSize, dl, Ptr.getValueType()));
557 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
558 LD->getPointerInfo().getWithOffset(IncrementSize),
559 NewLoadedVT, LD->isVolatile(),
560 LD->isNonTemporal(), LD->isInvariant(),
561 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
564 // aggregate the two parts
565 SDValue ShiftAmount =
566 DAG.getConstant(NumBits, dl, TLI.getShiftAmountTy(Hi.getValueType(),
567 DAG.getDataLayout()));
568 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
569 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
571 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
578 /// Some target cannot handle a variable insertion index for the
579 /// INSERT_VECTOR_ELT instruction. In this case, it
580 /// is necessary to spill the vector being inserted into to memory, perform
581 /// the insert there, and then read the result back.
582 SDValue SelectionDAGLegalize::
583 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
589 // If the target doesn't support this, we have to spill the input vector
590 // to a temporary stack slot, update the element, then reload it. This is
591 // badness. We could also load the value into a vector register (either
592 // with a "move to register" or "extload into register" instruction, then
593 // permute it into place, if the idx is a constant and if the idx is
594 // supported by the target.
595 EVT VT = Tmp1.getValueType();
596 EVT EltVT = VT.getVectorElementType();
597 EVT IdxVT = Tmp3.getValueType();
598 EVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
599 SDValue StackPtr = DAG.CreateStackTemporary(VT);
601 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
604 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
605 MachinePointerInfo::getFixedStack(SPFI),
608 // Truncate or zero extend offset to target pointer type.
609 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
610 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
611 // Add the offset to the index.
612 unsigned EltSize = EltVT.getSizeInBits()/8;
613 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,
614 DAG.getConstant(EltSize, dl, IdxVT));
615 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
616 // Store the scalar value.
617 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
619 // Load the updated vector.
620 return DAG.getLoad(VT, dl, Ch, StackPtr,
621 MachinePointerInfo::getFixedStack(SPFI), false, false,
626 SDValue SelectionDAGLegalize::
627 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
628 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
629 // SCALAR_TO_VECTOR requires that the type of the value being inserted
630 // match the element type of the vector being created, except for
631 // integers in which case the inserted value can be over width.
632 EVT EltVT = Vec.getValueType().getVectorElementType();
633 if (Val.getValueType() == EltVT ||
634 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
635 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
636 Vec.getValueType(), Val);
638 unsigned NumElts = Vec.getValueType().getVectorNumElements();
639 // We generate a shuffle of InVec and ScVec, so the shuffle mask
640 // should be 0,1,2,3,4,5... with the appropriate element replaced with
642 SmallVector<int, 8> ShufOps;
643 for (unsigned i = 0; i != NumElts; ++i)
644 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
646 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
650 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
653 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
654 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
655 // FIXME: We shouldn't do this for TargetConstantFP's.
656 // FIXME: move this to the DAG Combiner! Note that we can't regress due
657 // to phase ordering between legalized code and the dag combiner. This
658 // probably means that we need to integrate dag combiner and legalizer
660 // We generally can't do this one for long doubles.
661 SDValue Chain = ST->getChain();
662 SDValue Ptr = ST->getBasePtr();
663 unsigned Alignment = ST->getAlignment();
664 bool isVolatile = ST->isVolatile();
665 bool isNonTemporal = ST->isNonTemporal();
666 AAMDNodes AAInfo = ST->getAAInfo();
668 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
669 if (CFP->getValueType(0) == MVT::f32 &&
670 TLI.isTypeLegal(MVT::i32)) {
671 SDValue Con = DAG.getConstant(CFP->getValueAPF().
672 bitcastToAPInt().zextOrTrunc(32),
673 SDLoc(CFP), MVT::i32);
674 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
675 isVolatile, isNonTemporal, Alignment, AAInfo);
678 if (CFP->getValueType(0) == MVT::f64) {
679 // If this target supports 64-bit registers, do a single 64-bit store.
680 if (TLI.isTypeLegal(MVT::i64)) {
681 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
682 zextOrTrunc(64), SDLoc(CFP), MVT::i64);
683 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
684 isVolatile, isNonTemporal, Alignment, AAInfo);
687 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
688 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
689 // stores. If the target supports neither 32- nor 64-bits, this
690 // xform is certainly not worth it.
691 const APInt &IntVal = CFP->getValueAPF().bitcastToAPInt();
692 SDValue Lo = DAG.getConstant(IntVal.trunc(32), dl, MVT::i32);
693 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), dl, MVT::i32);
694 if (DAG.getDataLayout().isBigEndian())
697 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
698 isNonTemporal, Alignment, AAInfo);
699 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
700 DAG.getConstant(4, dl, Ptr.getValueType()));
701 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
702 ST->getPointerInfo().getWithOffset(4),
703 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
706 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
710 return SDValue(nullptr, 0);
713 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
714 StoreSDNode *ST = cast<StoreSDNode>(Node);
715 SDValue Chain = ST->getChain();
716 SDValue Ptr = ST->getBasePtr();
719 unsigned Alignment = ST->getAlignment();
720 bool isVolatile = ST->isVolatile();
721 bool isNonTemporal = ST->isNonTemporal();
722 AAMDNodes AAInfo = ST->getAAInfo();
724 if (!ST->isTruncatingStore()) {
725 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
726 ReplaceNode(ST, OptStore);
731 SDValue Value = ST->getValue();
732 MVT VT = Value.getSimpleValueType();
733 switch (TLI.getOperationAction(ISD::STORE, VT)) {
734 default: llvm_unreachable("This action is not supported yet!");
735 case TargetLowering::Legal: {
736 // If this is an unaligned store and the target doesn't support it,
738 unsigned AS = ST->getAddressSpace();
739 unsigned Align = ST->getAlignment();
740 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
741 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
742 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
743 if (Align < ABIAlignment)
744 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
748 case TargetLowering::Custom: {
749 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
750 if (Res && Res != SDValue(Node, 0))
751 ReplaceNode(SDValue(Node, 0), Res);
754 case TargetLowering::Promote: {
755 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
756 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
757 "Can only promote stores to same size type");
758 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
760 DAG.getStore(Chain, dl, Value, Ptr,
761 ST->getPointerInfo(), isVolatile,
762 isNonTemporal, Alignment, AAInfo);
763 ReplaceNode(SDValue(Node, 0), Result);
770 SDValue Value = ST->getValue();
772 EVT StVT = ST->getMemoryVT();
773 unsigned StWidth = StVT.getSizeInBits();
774 auto &DL = DAG.getDataLayout();
776 if (StWidth != StVT.getStoreSizeInBits()) {
777 // Promote to a byte-sized store with upper bits zero if not
778 // storing an integral number of bytes. For example, promote
779 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
780 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
781 StVT.getStoreSizeInBits());
782 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
784 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
785 NVT, isVolatile, isNonTemporal, Alignment, AAInfo);
786 ReplaceNode(SDValue(Node, 0), Result);
787 } else if (StWidth & (StWidth - 1)) {
788 // If not storing a power-of-2 number of bits, expand as two stores.
789 assert(!StVT.isVector() && "Unsupported truncstore!");
790 unsigned RoundWidth = 1 << Log2_32(StWidth);
791 assert(RoundWidth < StWidth);
792 unsigned ExtraWidth = StWidth - RoundWidth;
793 assert(ExtraWidth < RoundWidth);
794 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
795 "Store size not an integral number of bytes!");
796 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
797 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
799 unsigned IncrementSize;
801 if (DL.isLittleEndian()) {
802 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
803 // Store the bottom RoundWidth bits.
804 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
806 isVolatile, isNonTemporal, Alignment,
809 // Store the remaining ExtraWidth bits.
810 IncrementSize = RoundWidth / 8;
811 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
812 DAG.getConstant(IncrementSize, dl,
813 Ptr.getValueType()));
815 ISD::SRL, dl, Value.getValueType(), Value,
816 DAG.getConstant(RoundWidth, dl,
817 TLI.getShiftAmountTy(Value.getValueType(), DL)));
818 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
819 ST->getPointerInfo().getWithOffset(IncrementSize),
820 ExtraVT, isVolatile, isNonTemporal,
821 MinAlign(Alignment, IncrementSize), AAInfo);
823 // Big endian - avoid unaligned stores.
824 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
825 // Store the top RoundWidth bits.
827 ISD::SRL, dl, Value.getValueType(), Value,
828 DAG.getConstant(ExtraWidth, dl,
829 TLI.getShiftAmountTy(Value.getValueType(), DL)));
830 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
831 RoundVT, isVolatile, isNonTemporal, Alignment,
834 // Store the remaining ExtraWidth bits.
835 IncrementSize = RoundWidth / 8;
836 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
837 DAG.getConstant(IncrementSize, dl,
838 Ptr.getValueType()));
839 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
840 ST->getPointerInfo().getWithOffset(IncrementSize),
841 ExtraVT, isVolatile, isNonTemporal,
842 MinAlign(Alignment, IncrementSize), AAInfo);
845 // The order of the stores doesn't matter.
846 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
847 ReplaceNode(SDValue(Node, 0), Result);
849 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
850 StVT.getSimpleVT())) {
851 default: llvm_unreachable("This action is not supported yet!");
852 case TargetLowering::Legal: {
853 unsigned AS = ST->getAddressSpace();
854 unsigned Align = ST->getAlignment();
855 // If this is an unaligned store and the target doesn't support it,
857 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
858 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
859 unsigned ABIAlignment = DL.getABITypeAlignment(Ty);
860 if (Align < ABIAlignment)
861 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
865 case TargetLowering::Custom: {
866 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
867 if (Res && Res != SDValue(Node, 0))
868 ReplaceNode(SDValue(Node, 0), Res);
871 case TargetLowering::Expand:
872 assert(!StVT.isVector() &&
873 "Vector Stores are handled in LegalizeVectorOps");
875 // TRUNCSTORE:i16 i32 -> STORE i16
876 assert(TLI.isTypeLegal(StVT) &&
877 "Do not know how to expand this store!");
878 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
880 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
881 isVolatile, isNonTemporal, Alignment, AAInfo);
882 ReplaceNode(SDValue(Node, 0), Result);
889 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
890 LoadSDNode *LD = cast<LoadSDNode>(Node);
891 SDValue Chain = LD->getChain(); // The chain.
892 SDValue Ptr = LD->getBasePtr(); // The base pointer.
893 SDValue Value; // The value returned by the load op.
896 ISD::LoadExtType ExtType = LD->getExtensionType();
897 if (ExtType == ISD::NON_EXTLOAD) {
898 MVT VT = Node->getSimpleValueType(0);
899 SDValue RVal = SDValue(Node, 0);
900 SDValue RChain = SDValue(Node, 1);
902 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
903 default: llvm_unreachable("This action is not supported yet!");
904 case TargetLowering::Legal: {
905 unsigned AS = LD->getAddressSpace();
906 unsigned Align = LD->getAlignment();
907 // If this is an unaligned load and the target doesn't support it,
909 if (!TLI.allowsMisalignedMemoryAccesses(LD->getMemoryVT(), AS, Align)) {
910 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
911 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
912 if (Align < ABIAlignment){
913 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
918 case TargetLowering::Custom: {
919 SDValue Res = TLI.LowerOperation(RVal, DAG);
922 RChain = Res.getValue(1);
926 case TargetLowering::Promote: {
927 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
928 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
929 "Can only promote loads to same size type");
931 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
932 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
933 RChain = Res.getValue(1);
937 if (RChain.getNode() != Node) {
938 assert(RVal.getNode() != Node && "Load must be completely replaced");
939 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
940 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
942 UpdatedNodes->insert(RVal.getNode());
943 UpdatedNodes->insert(RChain.getNode());
950 EVT SrcVT = LD->getMemoryVT();
951 unsigned SrcWidth = SrcVT.getSizeInBits();
952 unsigned Alignment = LD->getAlignment();
953 bool isVolatile = LD->isVolatile();
954 bool isNonTemporal = LD->isNonTemporal();
955 bool isInvariant = LD->isInvariant();
956 AAMDNodes AAInfo = LD->getAAInfo();
958 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
959 // Some targets pretend to have an i1 loading operation, and actually
960 // load an i8. This trick is correct for ZEXTLOAD because the top 7
961 // bits are guaranteed to be zero; it helps the optimizers understand
962 // that these bits are zero. It is also useful for EXTLOAD, since it
963 // tells the optimizers that those bits are undefined. It would be
964 // nice to have an effective generic way of getting these benefits...
965 // Until such a way is found, don't insist on promoting i1 here.
967 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) ==
968 TargetLowering::Promote)) {
969 // Promote to a byte-sized load if not loading an integral number of
970 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
971 unsigned NewWidth = SrcVT.getStoreSizeInBits();
972 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
975 // The extra bits are guaranteed to be zero, since we stored them that
976 // way. A zext load from NVT thus automatically gives zext from SrcVT.
978 ISD::LoadExtType NewExtType =
979 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
982 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
983 Chain, Ptr, LD->getPointerInfo(),
984 NVT, isVolatile, isNonTemporal, isInvariant, Alignment,
987 Ch = Result.getValue(1); // The chain.
989 if (ExtType == ISD::SEXTLOAD)
990 // Having the top bits zero doesn't help when sign extending.
991 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
992 Result.getValueType(),
993 Result, DAG.getValueType(SrcVT));
994 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
995 // All the top bits are guaranteed to be zero - inform the optimizers.
996 Result = DAG.getNode(ISD::AssertZext, dl,
997 Result.getValueType(), Result,
998 DAG.getValueType(SrcVT));
1002 } else if (SrcWidth & (SrcWidth - 1)) {
1003 // If not loading a power-of-2 number of bits, expand as two loads.
1004 assert(!SrcVT.isVector() && "Unsupported extload!");
1005 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
1006 assert(RoundWidth < SrcWidth);
1007 unsigned ExtraWidth = SrcWidth - RoundWidth;
1008 assert(ExtraWidth < RoundWidth);
1009 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
1010 "Load size not an integral number of bytes!");
1011 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
1012 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1014 unsigned IncrementSize;
1015 auto &DL = DAG.getDataLayout();
1017 if (DL.isLittleEndian()) {
1018 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
1019 // Load the bottom RoundWidth bits.
1020 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1022 LD->getPointerInfo(), RoundVT, isVolatile,
1023 isNonTemporal, isInvariant, Alignment, AAInfo);
1025 // Load the remaining ExtraWidth bits.
1026 IncrementSize = RoundWidth / 8;
1027 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1028 DAG.getConstant(IncrementSize, dl,
1029 Ptr.getValueType()));
1030 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1031 LD->getPointerInfo().getWithOffset(IncrementSize),
1032 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1033 MinAlign(Alignment, IncrementSize), AAInfo);
1035 // Build a factor node to remember that this load is independent of
1037 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1040 // Move the top bits to the right place.
1042 ISD::SHL, dl, Hi.getValueType(), Hi,
1043 DAG.getConstant(RoundWidth, dl,
1044 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
1046 // Join the hi and lo parts.
1047 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1049 // Big endian - avoid unaligned loads.
1050 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1051 // Load the top RoundWidth bits.
1052 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1053 LD->getPointerInfo(), RoundVT, isVolatile,
1054 isNonTemporal, isInvariant, Alignment, AAInfo);
1056 // Load the remaining ExtraWidth bits.
1057 IncrementSize = RoundWidth / 8;
1058 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1059 DAG.getConstant(IncrementSize, dl,
1060 Ptr.getValueType()));
1061 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1062 dl, Node->getValueType(0), Chain, Ptr,
1063 LD->getPointerInfo().getWithOffset(IncrementSize),
1064 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1065 MinAlign(Alignment, IncrementSize), AAInfo);
1067 // Build a factor node to remember that this load is independent of
1069 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1072 // Move the top bits to the right place.
1074 ISD::SHL, dl, Hi.getValueType(), Hi,
1075 DAG.getConstant(ExtraWidth, dl,
1076 TLI.getShiftAmountTy(Hi.getValueType(), DL)));
1078 // Join the hi and lo parts.
1079 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1084 bool isCustom = false;
1085 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0),
1086 SrcVT.getSimpleVT())) {
1087 default: llvm_unreachable("This action is not supported yet!");
1088 case TargetLowering::Custom:
1091 case TargetLowering::Legal: {
1092 Value = SDValue(Node, 0);
1093 Chain = SDValue(Node, 1);
1096 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1097 if (Res.getNode()) {
1099 Chain = Res.getValue(1);
1102 // If this is an unaligned load and the target doesn't support
1104 EVT MemVT = LD->getMemoryVT();
1105 unsigned AS = LD->getAddressSpace();
1106 unsigned Align = LD->getAlignment();
1107 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1108 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1109 unsigned ABIAlignment = DAG.getDataLayout().getABITypeAlignment(Ty);
1110 if (Align < ABIAlignment){
1111 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, Value, Chain);
1117 case TargetLowering::Expand:
1118 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, Node->getValueType(0), SrcVT)) {
1119 // If the source type is not legal, see if there is a legal extload to
1120 // an intermediate type that we can then extend further.
1121 EVT LoadVT = TLI.getRegisterType(SrcVT.getSimpleVT());
1122 if (TLI.isTypeLegal(SrcVT) || // Same as SrcVT == LoadVT?
1123 TLI.isLoadExtLegal(ExtType, LoadVT, SrcVT)) {
1124 // If we are loading a legal type, this is a non-extload followed by a
1126 ISD::LoadExtType MidExtType =
1127 (LoadVT == SrcVT) ? ISD::NON_EXTLOAD : ExtType;
1129 SDValue Load = DAG.getExtLoad(MidExtType, dl, LoadVT, Chain, Ptr,
1130 SrcVT, LD->getMemOperand());
1132 ISD::getExtForLoadExtType(SrcVT.isFloatingPoint(), ExtType);
1133 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1134 Chain = Load.getValue(1);
1139 assert(!SrcVT.isVector() &&
1140 "Vector Loads are handled in LegalizeVectorOps");
1142 // FIXME: This does not work for vectors on most targets. Sign-
1143 // and zero-extend operations are currently folded into extending
1144 // loads, whether they are legal or not, and then we end up here
1145 // without any support for legalizing them.
1146 assert(ExtType != ISD::EXTLOAD &&
1147 "EXTLOAD should always be supported!");
1148 // Turn the unsupported load into an EXTLOAD followed by an
1149 // explicit zero/sign extend inreg.
1150 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1151 Node->getValueType(0),
1153 LD->getMemOperand());
1155 if (ExtType == ISD::SEXTLOAD)
1156 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1157 Result.getValueType(),
1158 Result, DAG.getValueType(SrcVT));
1160 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1162 Chain = Result.getValue(1);
1167 // Since loads produce two values, make sure to remember that we legalized
1169 if (Chain.getNode() != Node) {
1170 assert(Value.getNode() != Node && "Load must be completely replaced");
1171 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1172 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1174 UpdatedNodes->insert(Value.getNode());
1175 UpdatedNodes->insert(Chain.getNode());
1181 /// Return a legal replacement for the given operation, with all legal operands.
1182 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1183 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1185 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1189 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1190 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1191 TargetLowering::TypeLegal &&
1192 "Unexpected illegal type!");
1194 for (const SDValue &Op : Node->op_values())
1195 assert((TLI.getTypeAction(*DAG.getContext(),
1196 Op.getValueType()) == TargetLowering::TypeLegal ||
1197 Op.getOpcode() == ISD::TargetConstant) &&
1198 "Unexpected illegal type!");
1201 // Figure out the correct action; the way to query this varies by opcode
1202 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1203 bool SimpleFinishLegalizing = true;
1204 switch (Node->getOpcode()) {
1205 case ISD::INTRINSIC_W_CHAIN:
1206 case ISD::INTRINSIC_WO_CHAIN:
1207 case ISD::INTRINSIC_VOID:
1208 case ISD::STACKSAVE:
1209 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1212 Action = TLI.getOperationAction(Node->getOpcode(),
1213 Node->getValueType(0));
1214 if (Action != TargetLowering::Promote)
1215 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1217 case ISD::FP_TO_FP16:
1218 case ISD::SINT_TO_FP:
1219 case ISD::UINT_TO_FP:
1220 case ISD::EXTRACT_VECTOR_ELT:
1221 Action = TLI.getOperationAction(Node->getOpcode(),
1222 Node->getOperand(0).getValueType());
1224 case ISD::FP_ROUND_INREG:
1225 case ISD::SIGN_EXTEND_INREG: {
1226 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1227 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1230 case ISD::ATOMIC_STORE: {
1231 Action = TLI.getOperationAction(Node->getOpcode(),
1232 Node->getOperand(2).getValueType());
1235 case ISD::SELECT_CC:
1238 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1239 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1240 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1241 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1242 ISD::CondCode CCCode =
1243 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1244 Action = TLI.getCondCodeAction(CCCode, OpVT);
1245 if (Action == TargetLowering::Legal) {
1246 if (Node->getOpcode() == ISD::SELECT_CC)
1247 Action = TLI.getOperationAction(Node->getOpcode(),
1248 Node->getValueType(0));
1250 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1256 // FIXME: Model these properly. LOAD and STORE are complicated, and
1257 // STORE expects the unlegalized operand in some cases.
1258 SimpleFinishLegalizing = false;
1260 case ISD::CALLSEQ_START:
1261 case ISD::CALLSEQ_END:
1262 // FIXME: This shouldn't be necessary. These nodes have special properties
1263 // dealing with the recursive nature of legalization. Removing this
1264 // special case should be done as part of making LegalizeDAG non-recursive.
1265 SimpleFinishLegalizing = false;
1267 case ISD::EXTRACT_ELEMENT:
1268 case ISD::FLT_ROUNDS_:
1270 case ISD::MERGE_VALUES:
1271 case ISD::EH_RETURN:
1272 case ISD::FRAME_TO_ARGS_OFFSET:
1273 case ISD::EH_SJLJ_SETJMP:
1274 case ISD::EH_SJLJ_LONGJMP:
1275 // These operations lie about being legal: when they claim to be legal,
1276 // they should actually be expanded.
1277 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1278 if (Action == TargetLowering::Legal)
1279 Action = TargetLowering::Expand;
1281 case ISD::INIT_TRAMPOLINE:
1282 case ISD::ADJUST_TRAMPOLINE:
1283 case ISD::FRAMEADDR:
1284 case ISD::RETURNADDR:
1285 // These operations lie about being legal: when they claim to be legal,
1286 // they should actually be custom-lowered.
1287 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1288 if (Action == TargetLowering::Legal)
1289 Action = TargetLowering::Custom;
1291 case ISD::READ_REGISTER:
1292 case ISD::WRITE_REGISTER:
1293 // Named register is legal in the DAG, but blocked by register name
1294 // selection if not implemented by target (to chose the correct register)
1295 // They'll be converted to Copy(To/From)Reg.
1296 Action = TargetLowering::Legal;
1298 case ISD::DEBUGTRAP:
1299 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1300 if (Action == TargetLowering::Expand) {
1301 // replace ISD::DEBUGTRAP with ISD::TRAP
1303 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1304 Node->getOperand(0));
1305 ReplaceNode(Node, NewVal.getNode());
1306 LegalizeOp(NewVal.getNode());
1312 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1313 Action = TargetLowering::Legal;
1315 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1320 if (SimpleFinishLegalizing) {
1321 SDNode *NewNode = Node;
1322 switch (Node->getOpcode()) {
1329 // Legalizing shifts/rotates requires adjusting the shift amount
1330 // to the appropriate width.
1331 if (!Node->getOperand(1).getValueType().isVector()) {
1333 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1334 Node->getOperand(1));
1335 HandleSDNode Handle(SAO);
1336 LegalizeOp(SAO.getNode());
1337 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1341 case ISD::SRL_PARTS:
1342 case ISD::SRA_PARTS:
1343 case ISD::SHL_PARTS:
1344 // Legalizing shifts/rotates requires adjusting the shift amount
1345 // to the appropriate width.
1346 if (!Node->getOperand(2).getValueType().isVector()) {
1348 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1349 Node->getOperand(2));
1350 HandleSDNode Handle(SAO);
1351 LegalizeOp(SAO.getNode());
1352 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1353 Node->getOperand(1),
1359 if (NewNode != Node) {
1360 ReplaceNode(Node, NewNode);
1364 case TargetLowering::Legal:
1366 case TargetLowering::Custom: {
1367 // FIXME: The handling for custom lowering with multiple results is
1369 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1370 if (Res.getNode()) {
1371 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1374 if (Node->getNumValues() == 1) {
1375 // We can just directly replace this node with the lowered value.
1376 ReplaceNode(SDValue(Node, 0), Res);
1380 SmallVector<SDValue, 8> ResultVals;
1381 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1382 ResultVals.push_back(Res.getValue(i));
1383 ReplaceNode(Node, ResultVals.data());
1388 case TargetLowering::Expand:
1391 case TargetLowering::Promote:
1397 switch (Node->getOpcode()) {
1404 llvm_unreachable("Do not know how to legalize this operator!");
1406 case ISD::CALLSEQ_START:
1407 case ISD::CALLSEQ_END:
1410 return LegalizeLoadOps(Node);
1413 return LegalizeStoreOps(Node);
1418 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1419 SDValue Vec = Op.getOperand(0);
1420 SDValue Idx = Op.getOperand(1);
1423 // Before we generate a new store to a temporary stack slot, see if there is
1424 // already one that we can use. There often is because when we scalarize
1425 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1426 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1427 // the vector. If all are expanded here, we don't want one store per vector
1429 SDValue StackPtr, Ch;
1430 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1431 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1433 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1434 if (ST->isIndexed() || ST->isTruncatingStore() ||
1435 ST->getValue() != Vec)
1438 // Make sure that nothing else could have stored into the destination of
1440 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1443 StackPtr = ST->getBasePtr();
1444 Ch = SDValue(ST, 0);
1449 if (!Ch.getNode()) {
1450 // Store the value to a temporary stack slot, then LOAD the returned part.
1451 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1452 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1453 MachinePointerInfo(), false, false, 0);
1456 // Add the offset to the index.
1458 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1459 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1460 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType()));
1462 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout()));
1463 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1467 if (Op.getValueType().isVector())
1468 NewLoad = DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,
1469 MachinePointerInfo(), false, false, false, 0);
1471 NewLoad = DAG.getExtLoad(
1472 ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr, MachinePointerInfo(),
1473 Vec.getValueType().getVectorElementType(), false, false, false, 0);
1475 // Replace the chain going out of the store, by the one out of the load.
1476 DAG.ReplaceAllUsesOfValueWith(Ch, SDValue(NewLoad.getNode(), 1));
1478 // We introduced a cycle though, so update the loads operands, making sure
1479 // to use the original store's chain as an incoming chain.
1480 SmallVector<SDValue, 6> NewLoadOperands(NewLoad->op_begin(),
1482 NewLoadOperands[0] = Ch;
1484 SDValue(DAG.UpdateNodeOperands(NewLoad.getNode(), NewLoadOperands), 0);
1488 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1489 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1491 SDValue Vec = Op.getOperand(0);
1492 SDValue Part = Op.getOperand(1);
1493 SDValue Idx = Op.getOperand(2);
1496 // Store the value to a temporary stack slot, then LOAD the returned part.
1498 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1499 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1500 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1502 // First store the whole vector.
1503 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1506 // Then store the inserted part.
1508 // Add the offset to the index.
1510 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1512 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1513 DAG.getConstant(EltSize, SDLoc(Vec), Idx.getValueType()));
1514 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy(DAG.getDataLayout()));
1516 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1519 // Store the subvector.
1520 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr,
1521 MachinePointerInfo(), false, false, 0);
1523 // Finally, load the updated vector.
1524 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1525 false, false, false, 0);
1528 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1529 // We can't handle this case efficiently. Allocate a sufficiently
1530 // aligned object on the stack, store each element into it, then load
1531 // the result as a vector.
1532 // Create the stack frame object.
1533 EVT VT = Node->getValueType(0);
1534 EVT EltVT = VT.getVectorElementType();
1536 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1537 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1538 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1540 // Emit a store of each element to the stack slot.
1541 SmallVector<SDValue, 8> Stores;
1542 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1543 // Store (in the right endianness) the elements to memory.
1544 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1545 // Ignore undef elements.
1546 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1548 unsigned Offset = TypeByteSize*i;
1550 SDValue Idx = DAG.getConstant(Offset, dl, FIPtr.getValueType());
1551 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1553 // If the destination vector element type is narrower than the source
1554 // element type, only store the bits necessary.
1555 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1556 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1557 Node->getOperand(i), Idx,
1558 PtrInfo.getWithOffset(Offset),
1559 EltVT, false, false, 0));
1561 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1562 Node->getOperand(i), Idx,
1563 PtrInfo.getWithOffset(Offset),
1568 if (!Stores.empty()) // Not all undef elements?
1569 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1571 StoreChain = DAG.getEntryNode();
1573 // Result is a load from the stack slot.
1574 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1575 false, false, false, 0);
1579 /// Keeps track of state when getting the sign of a floatingpoint value as an
1581 struct FloatSignAsInt {
1586 MachinePointerInfo IntPointerInfo;
1587 MachinePointerInfo FloatPointerInfo;
1593 /// Bitcast a floatingpoint value to an integer value. Only bitcast the part
1594 /// containing the sign bit if the target has no integer value capable of
1595 /// holding all bits of the floatingpoint value.
1596 void SelectionDAGLegalize::getSignAsIntValue(FloatSignAsInt &State,
1597 SDLoc DL, SDValue Value) const {
1598 EVT FloatVT = Value.getValueType();
1599 unsigned NumBits = FloatVT.getSizeInBits();
1600 State.FloatVT = FloatVT;
1601 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), NumBits);
1602 // Convert to an integer of the same size.
1603 if (TLI.isTypeLegal(IVT)) {
1604 State.IntValue = DAG.getNode(ISD::BITCAST, DL, IVT, Value);
1605 State.SignMask = APInt::getSignBit(NumBits);
1609 auto &DataLayout = DAG.getDataLayout();
1610 // Store the float to memory, then load the sign part out as an integer.
1611 MVT LoadTy = TLI.getPointerTy(DataLayout);
1612 // First create a temporary that is aligned for both the load and store.
1613 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1614 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1615 // Then store the float to it.
1616 State.FloatPtr = StackPtr;
1617 State.FloatPointerInfo = MachinePointerInfo::getFixedStack(FI);
1618 State.Chain = DAG.getStore(DAG.getEntryNode(), DL, Value, State.FloatPtr,
1619 State.FloatPointerInfo, false, false, 0);
1621 if (DataLayout.isBigEndian()) {
1622 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1623 // Load out a legal integer with the same sign bit as the float.
1624 State.IntPtr = StackPtr;
1625 State.IntPointerInfo = State.FloatPointerInfo;
1626 State.IntValue = DAG.getLoad(LoadTy, DL, State.Chain, StackPtr,
1627 State.IntPointerInfo, false, false, false, 0);
1628 State.SignMask = APInt::getSignBit(LoadTy.getSizeInBits());
1629 } else { // Little endian
1630 // The float may be wider than the integer we are going to load. Advance
1631 // the pointer so that the loaded integer will contain the sign bit.
1632 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1633 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1635 = DAG.getNode(ISD::ADD, DL, StackPtr.getValueType(), StackPtr,
1636 DAG.getConstant(ByteOffset, DL, StackPtr.getValueType()));
1637 // Load a legal integer containing the sign bit.
1638 State.IntPtr = LoadPtr;
1639 State.IntPointerInfo = MachinePointerInfo::getFixedStack(FI, ByteOffset);
1640 State.IntValue = DAG.getLoad(LoadTy, DL, State.Chain, LoadPtr,
1641 State.IntPointerInfo, false, false, false, 0);
1642 unsigned NumBits = LoadTy.getSizeInBits();
1643 unsigned SignBit = (FloatVT.getSizeInBits()-1) - 8*ByteOffset;
1644 State.SignMask = APInt::getOneBitSet(NumBits, SignBit);
1648 /// Replace the integer value produced by getSignAsIntValue() with a new value
1649 /// and cast the result back to a floatingpoint type.
1650 SDValue SelectionDAGLegalize::modifySignAsInt(const FloatSignAsInt &State,
1651 SDLoc DL, SDValue NewIntValue) const {
1653 return DAG.getNode(ISD::BITCAST, DL, State.FloatVT, NewIntValue);
1656 // Override the part containing the sign bit in the value stored on the stack.
1657 SDValue Chain = DAG.getStore(State.Chain, DL, NewIntValue, State.IntPtr,
1658 State.IntPointerInfo, false, false, 0);
1659 return DAG.getLoad(State.FloatVT, DL, Chain, State.FloatPtr,
1660 State.FloatPointerInfo, false, false, false, 0);
1663 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode *Node) const {
1665 SDValue Mag = Node->getOperand(0);
1666 SDValue Sign = Node->getOperand(1);
1668 // Get sign bit into an integer value.
1669 FloatSignAsInt SignAsInt;
1670 getSignAsIntValue(SignAsInt, DL, Sign);
1672 EVT IntVT = SignAsInt.IntValue.getValueType();
1673 SDValue SignMask = DAG.getConstant(SignAsInt.SignMask, DL, IntVT);
1674 SDValue SignBit = DAG.getNode(ISD::AND, DL, IntVT, SignAsInt.IntValue,
1677 // Transform FCOPYSIGN(x, y) => sign(x) ? -FABS(x) : FABS(X) if FABS is legal.
1678 EVT FloatVT = Mag.getValueType();
1679 if (TLI.isOperationLegalOrCustom(ISD::FABS, FloatVT) &&
1680 TLI.isOperationLegalOrCustom(ISD::FNEG, FloatVT)) {
1681 SDValue AbsValue = DAG.getNode(ISD::FABS, DL, FloatVT, Mag);
1682 SDValue NegValue = DAG.getNode(ISD::FNEG, DL, FloatVT, AbsValue);
1683 SDValue Cond = DAG.getSetCC(DL, getSetCCResultType(IntVT), SignBit,
1684 DAG.getConstant(0, DL, IntVT), ISD::SETNE);
1685 return DAG.getSelect(DL, FloatVT, Cond, NegValue, AbsValue);
1688 // Transform values to integer, copy the sign bit and transform back.
1689 FloatSignAsInt MagAsInt;
1690 getSignAsIntValue(MagAsInt, DL, Mag);
1691 assert(SignAsInt.SignMask == MagAsInt.SignMask);
1692 SDValue ClearSignMask = DAG.getConstant(~SignAsInt.SignMask, DL, IntVT);
1693 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, MagAsInt.IntValue,
1695 SDValue CopiedSign = DAG.getNode(ISD::OR, DL, IntVT, ClearedSign, SignBit);
1697 return modifySignAsInt(MagAsInt, DL, CopiedSign);
1700 SDValue SelectionDAGLegalize::ExpandFABS(SDNode *Node) const {
1702 SDValue Value = Node->getOperand(0);
1704 // Transform FABS(x) => FCOPYSIGN(x, 0.0) if FCOPYSIGN is legal.
1705 EVT FloatVT = Value.getValueType();
1706 if (TLI.isOperationLegalOrCustom(ISD::FCOPYSIGN, FloatVT)) {
1707 SDValue Zero = DAG.getConstantFP(0.0, DL, FloatVT);
1708 return DAG.getNode(ISD::FCOPYSIGN, DL, FloatVT, Value, Zero);
1711 // Transform value to integer, clear the sign bit and transform back.
1712 FloatSignAsInt ValueAsInt;
1713 getSignAsIntValue(ValueAsInt, DL, Value);
1714 EVT IntVT = ValueAsInt.IntValue.getValueType();
1715 SDValue ClearSignMask = DAG.getConstant(~ValueAsInt.SignMask, DL, IntVT);
1716 SDValue ClearedSign = DAG.getNode(ISD::AND, DL, IntVT, ValueAsInt.IntValue,
1718 return modifySignAsInt(ValueAsInt, DL, ClearedSign);
1721 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1722 SmallVectorImpl<SDValue> &Results) {
1723 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1724 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1725 " not tell us which reg is the stack pointer!");
1727 EVT VT = Node->getValueType(0);
1728 SDValue Tmp1 = SDValue(Node, 0);
1729 SDValue Tmp2 = SDValue(Node, 1);
1730 SDValue Tmp3 = Node->getOperand(2);
1731 SDValue Chain = Tmp1.getOperand(0);
1733 // Chain the dynamic stack allocation so that it doesn't modify the stack
1734 // pointer when other instructions are using the stack.
1735 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true), dl);
1737 SDValue Size = Tmp2.getOperand(1);
1738 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1739 Chain = SP.getValue(1);
1740 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1741 unsigned StackAlign =
1742 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1743 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1744 if (Align > StackAlign)
1745 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1746 DAG.getConstant(-(uint64_t)Align, dl, VT));
1747 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1749 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
1750 DAG.getIntPtrConstant(0, dl, true), SDValue(), dl);
1752 Results.push_back(Tmp1);
1753 Results.push_back(Tmp2);
1756 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1759 /// If the SETCC has been legalized using AND / OR, then the legalized node
1760 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1761 /// will be set to false.
1763 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1764 /// then the values of LHS and RHS will be swapped, CC will be set to the
1765 /// new condition, and NeedInvert will be set to false.
1767 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1768 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1769 /// will be set to true. The caller must invert the result of the SETCC with
1770 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1771 /// of a true/false result.
1773 /// \returns true if the SetCC has been legalized, false if it hasn't.
1774 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1775 SDValue &LHS, SDValue &RHS,
1779 MVT OpVT = LHS.getSimpleValueType();
1780 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1782 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1783 default: llvm_unreachable("Unknown condition code action!");
1784 case TargetLowering::Legal:
1787 case TargetLowering::Expand: {
1788 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1789 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1790 std::swap(LHS, RHS);
1791 CC = DAG.getCondCode(InvCC);
1794 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1797 default: llvm_unreachable("Don't know how to expand this condition!");
1799 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1800 == TargetLowering::Legal
1801 && "If SETO is expanded, SETOEQ must be legal!");
1802 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1804 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1805 == TargetLowering::Legal
1806 && "If SETUO is expanded, SETUNE must be legal!");
1807 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1820 // If we are floating point, assign and break, otherwise fall through.
1821 if (!OpVT.isInteger()) {
1822 // We can use the 4th bit to tell if we are the unordered
1823 // or ordered version of the opcode.
1824 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1825 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1826 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1829 // Fallthrough if we are unsigned integer.
1834 // We only support using the inverted operation, which is computed above
1835 // and not a different manner of supporting expanding these cases.
1836 llvm_unreachable("Don't know how to expand this condition!");
1839 // Try inverting the result of the inverse condition.
1840 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1841 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1842 CC = DAG.getCondCode(InvCC);
1846 // If inverting the condition didn't work then we have no means to expand
1848 llvm_unreachable("Don't know how to expand this condition!");
1851 SDValue SetCC1, SetCC2;
1852 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1853 // If we aren't the ordered or unorder operation,
1854 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1855 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1856 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1858 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1859 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1860 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1862 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1871 /// Emit a store/load combination to the stack. This stores
1872 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1873 /// a load from the stack slot to DestVT, extending it if needed.
1874 /// The resultant code need not be legal.
1875 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1879 // Create the stack frame object.
1880 unsigned SrcAlign = DAG.getDataLayout().getPrefTypeAlignment(
1881 SrcOp.getValueType().getTypeForEVT(*DAG.getContext()));
1882 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1884 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1885 int SPFI = StackPtrFI->getIndex();
1886 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1888 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1889 unsigned SlotSize = SlotVT.getSizeInBits();
1890 unsigned DestSize = DestVT.getSizeInBits();
1891 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1892 unsigned DestAlign = DAG.getDataLayout().getPrefTypeAlignment(DestType);
1894 // Emit a store to the stack slot. Use a truncstore if the input value is
1895 // later than DestVT.
1898 if (SrcSize > SlotSize)
1899 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1900 PtrInfo, SlotVT, false, false, SrcAlign);
1902 assert(SrcSize == SlotSize && "Invalid store");
1903 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1904 PtrInfo, false, false, SrcAlign);
1907 // Result is a load from the stack slot.
1908 if (SlotSize == DestSize)
1909 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1910 false, false, false, DestAlign);
1912 assert(SlotSize < DestSize && "Unknown extension!");
1913 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1914 PtrInfo, SlotVT, false, false, false, DestAlign);
1917 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1919 // Create a vector sized/aligned stack slot, store the value to element #0,
1920 // then load the whole vector back out.
1921 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1923 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1924 int SPFI = StackPtrFI->getIndex();
1926 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1928 MachinePointerInfo::getFixedStack(SPFI),
1929 Node->getValueType(0).getVectorElementType(),
1931 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1932 MachinePointerInfo::getFixedStack(SPFI),
1933 false, false, false, 0);
1937 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1938 const TargetLowering &TLI, SDValue &Res) {
1939 unsigned NumElems = Node->getNumOperands();
1941 EVT VT = Node->getValueType(0);
1943 // Try to group the scalars into pairs, shuffle the pairs together, then
1944 // shuffle the pairs of pairs together, etc. until the vector has
1945 // been built. This will work only if all of the necessary shuffle masks
1948 // We do this in two phases; first to check the legality of the shuffles,
1949 // and next, assuming that all shuffles are legal, to create the new nodes.
1950 for (int Phase = 0; Phase < 2; ++Phase) {
1951 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1953 for (unsigned i = 0; i < NumElems; ++i) {
1954 SDValue V = Node->getOperand(i);
1955 if (V.getOpcode() == ISD::UNDEF)
1960 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1961 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1964 while (IntermedVals.size() > 2) {
1965 NewIntermedVals.clear();
1966 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1967 // This vector and the next vector are shuffled together (simply to
1968 // append the one to the other).
1969 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1971 SmallVector<int, 16> FinalIndices;
1972 FinalIndices.reserve(IntermedVals[i].second.size() +
1973 IntermedVals[i+1].second.size());
1976 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1979 FinalIndices.push_back(IntermedVals[i].second[j]);
1981 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1983 ShuffleVec[k] = NumElems + j;
1984 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1989 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1990 IntermedVals[i+1].first,
1992 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1994 NewIntermedVals.push_back(
1995 std::make_pair(Shuffle, std::move(FinalIndices)));
1998 // If we had an odd number of defined values, then append the last
1999 // element to the array of new vectors.
2000 if ((IntermedVals.size() & 1) != 0)
2001 NewIntermedVals.push_back(IntermedVals.back());
2003 IntermedVals.swap(NewIntermedVals);
2006 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
2007 "Invalid number of intermediate vectors");
2008 SDValue Vec1 = IntermedVals[0].first;
2010 if (IntermedVals.size() > 1)
2011 Vec2 = IntermedVals[1].first;
2013 Vec2 = DAG.getUNDEF(VT);
2015 SmallVector<int, 16> ShuffleVec(NumElems, -1);
2016 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
2017 ShuffleVec[IntermedVals[0].second[i]] = i;
2018 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
2019 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
2022 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2023 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
2030 /// Expand a BUILD_VECTOR node on targets that don't
2031 /// support the operation, but do support the resultant vector type.
2032 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
2033 unsigned NumElems = Node->getNumOperands();
2034 SDValue Value1, Value2;
2036 EVT VT = Node->getValueType(0);
2037 EVT OpVT = Node->getOperand(0).getValueType();
2038 EVT EltVT = VT.getVectorElementType();
2040 // If the only non-undef value is the low element, turn this into a
2041 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
2042 bool isOnlyLowElement = true;
2043 bool MoreThanTwoValues = false;
2044 bool isConstant = true;
2045 for (unsigned i = 0; i < NumElems; ++i) {
2046 SDValue V = Node->getOperand(i);
2047 if (V.getOpcode() == ISD::UNDEF)
2050 isOnlyLowElement = false;
2051 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
2054 if (!Value1.getNode()) {
2056 } else if (!Value2.getNode()) {
2059 } else if (V != Value1 && V != Value2) {
2060 MoreThanTwoValues = true;
2064 if (!Value1.getNode())
2065 return DAG.getUNDEF(VT);
2067 if (isOnlyLowElement)
2068 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
2070 // If all elements are constants, create a load from the constant pool.
2072 SmallVector<Constant*, 16> CV;
2073 for (unsigned i = 0, e = NumElems; i != e; ++i) {
2074 if (ConstantFPSDNode *V =
2075 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
2076 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
2077 } else if (ConstantSDNode *V =
2078 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
2080 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
2082 // If OpVT and EltVT don't match, EltVT is not legal and the
2083 // element values have been promoted/truncated earlier. Undo this;
2084 // we don't want a v16i8 to become a v16i32 for example.
2085 const ConstantInt *CI = V->getConstantIntValue();
2086 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
2087 CI->getZExtValue()));
2090 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
2091 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
2092 CV.push_back(UndefValue::get(OpNTy));
2095 Constant *CP = ConstantVector::get(CV);
2097 DAG.getConstantPool(CP, TLI.getPointerTy(DAG.getDataLayout()));
2098 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2099 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
2100 MachinePointerInfo::getConstantPool(),
2101 false, false, false, Alignment);
2104 SmallSet<SDValue, 16> DefinedValues;
2105 for (unsigned i = 0; i < NumElems; ++i) {
2106 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
2108 DefinedValues.insert(Node->getOperand(i));
2111 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
2112 if (!MoreThanTwoValues) {
2113 SmallVector<int, 8> ShuffleVec(NumElems, -1);
2114 for (unsigned i = 0; i < NumElems; ++i) {
2115 SDValue V = Node->getOperand(i);
2116 if (V.getOpcode() == ISD::UNDEF)
2118 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2120 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2121 // Get the splatted value into the low element of a vector register.
2122 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2124 if (Value2.getNode())
2125 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2127 Vec2 = DAG.getUNDEF(VT);
2129 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2130 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2134 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2139 // Otherwise, we can't handle this case efficiently.
2140 return ExpandVectorBuildThroughStack(Node);
2143 // Expand a node into a call to a libcall. If the result value
2144 // does not fit into a register, return the lo part and set the hi part to the
2145 // by-reg argument. If it does fit into a single register, return the result
2146 // and leave the Hi part unset.
2147 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2149 TargetLowering::ArgListTy Args;
2150 TargetLowering::ArgListEntry Entry;
2151 for (const SDValue &Op : Node->op_values()) {
2152 EVT ArgVT = Op.getValueType();
2153 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2156 Entry.isSExt = isSigned;
2157 Entry.isZExt = !isSigned;
2158 Args.push_back(Entry);
2160 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2161 TLI.getPointerTy(DAG.getDataLayout()));
2163 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2165 // By default, the input chain to this libcall is the entry node of the
2166 // function. If the libcall is going to be emitted as a tail call then
2167 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2168 // node which is being folded has a non-entry input chain.
2169 SDValue InChain = DAG.getEntryNode();
2171 // isTailCall may be true since the callee does not reference caller stack
2172 // frame. Check if it's in the right position.
2173 SDValue TCChain = InChain;
2174 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2178 TargetLowering::CallLoweringInfo CLI(DAG);
2179 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2180 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2181 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2183 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2185 if (!CallInfo.second.getNode())
2186 // It's a tailcall, return the chain (which is the DAG root).
2187 return DAG.getRoot();
2189 return CallInfo.first;
2192 /// Generate a libcall taking the given operands as arguments
2193 /// and returning a result of type RetVT.
2194 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2195 const SDValue *Ops, unsigned NumOps,
2196 bool isSigned, SDLoc dl) {
2197 TargetLowering::ArgListTy Args;
2198 Args.reserve(NumOps);
2200 TargetLowering::ArgListEntry Entry;
2201 for (unsigned i = 0; i != NumOps; ++i) {
2202 Entry.Node = Ops[i];
2203 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2204 Entry.isSExt = isSigned;
2205 Entry.isZExt = !isSigned;
2206 Args.push_back(Entry);
2208 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2209 TLI.getPointerTy(DAG.getDataLayout()));
2211 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2213 TargetLowering::CallLoweringInfo CLI(DAG);
2214 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2215 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2216 .setSExtResult(isSigned).setZExtResult(!isSigned);
2218 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2220 return CallInfo.first;
2223 // Expand a node into a call to a libcall. Similar to
2224 // ExpandLibCall except that the first operand is the in-chain.
2225 std::pair<SDValue, SDValue>
2226 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2229 SDValue InChain = Node->getOperand(0);
2231 TargetLowering::ArgListTy Args;
2232 TargetLowering::ArgListEntry Entry;
2233 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2234 EVT ArgVT = Node->getOperand(i).getValueType();
2235 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2236 Entry.Node = Node->getOperand(i);
2238 Entry.isSExt = isSigned;
2239 Entry.isZExt = !isSigned;
2240 Args.push_back(Entry);
2242 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2243 TLI.getPointerTy(DAG.getDataLayout()));
2245 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2247 TargetLowering::CallLoweringInfo CLI(DAG);
2248 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2249 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2250 .setSExtResult(isSigned).setZExtResult(!isSigned);
2252 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2257 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2258 RTLIB::Libcall Call_F32,
2259 RTLIB::Libcall Call_F64,
2260 RTLIB::Libcall Call_F80,
2261 RTLIB::Libcall Call_F128,
2262 RTLIB::Libcall Call_PPCF128) {
2264 switch (Node->getSimpleValueType(0).SimpleTy) {
2265 default: llvm_unreachable("Unexpected request for libcall!");
2266 case MVT::f32: LC = Call_F32; break;
2267 case MVT::f64: LC = Call_F64; break;
2268 case MVT::f80: LC = Call_F80; break;
2269 case MVT::f128: LC = Call_F128; break;
2270 case MVT::ppcf128: LC = Call_PPCF128; break;
2272 return ExpandLibCall(LC, Node, false);
2275 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2276 RTLIB::Libcall Call_I8,
2277 RTLIB::Libcall Call_I16,
2278 RTLIB::Libcall Call_I32,
2279 RTLIB::Libcall Call_I64,
2280 RTLIB::Libcall Call_I128) {
2282 switch (Node->getSimpleValueType(0).SimpleTy) {
2283 default: llvm_unreachable("Unexpected request for libcall!");
2284 case MVT::i8: LC = Call_I8; break;
2285 case MVT::i16: LC = Call_I16; break;
2286 case MVT::i32: LC = Call_I32; break;
2287 case MVT::i64: LC = Call_I64; break;
2288 case MVT::i128: LC = Call_I128; break;
2290 return ExpandLibCall(LC, Node, isSigned);
2293 /// Return true if divmod libcall is available.
2294 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2295 const TargetLowering &TLI) {
2297 switch (Node->getSimpleValueType(0).SimpleTy) {
2298 default: llvm_unreachable("Unexpected request for libcall!");
2299 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2300 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2301 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2302 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2303 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2306 return TLI.getLibcallName(LC) != nullptr;
2309 /// Only issue divrem libcall if both quotient and remainder are needed.
2310 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2311 // The other use might have been replaced with a divrem already.
2312 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2313 unsigned OtherOpcode = 0;
2315 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2317 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2319 SDValue Op0 = Node->getOperand(0);
2320 SDValue Op1 = Node->getOperand(1);
2321 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2322 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2326 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2327 User->getOperand(0) == Op0 &&
2328 User->getOperand(1) == Op1)
2334 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2336 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2337 SmallVectorImpl<SDValue> &Results) {
2338 unsigned Opcode = Node->getOpcode();
2339 bool isSigned = Opcode == ISD::SDIVREM;
2342 switch (Node->getSimpleValueType(0).SimpleTy) {
2343 default: llvm_unreachable("Unexpected request for libcall!");
2344 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2345 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2346 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2347 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2348 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2351 // The input chain to this libcall is the entry node of the function.
2352 // Legalizing the call will automatically add the previous call to the
2354 SDValue InChain = DAG.getEntryNode();
2356 EVT RetVT = Node->getValueType(0);
2357 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2359 TargetLowering::ArgListTy Args;
2360 TargetLowering::ArgListEntry Entry;
2361 for (const SDValue &Op : Node->op_values()) {
2362 EVT ArgVT = Op.getValueType();
2363 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2366 Entry.isSExt = isSigned;
2367 Entry.isZExt = !isSigned;
2368 Args.push_back(Entry);
2371 // Also pass the return address of the remainder.
2372 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2374 Entry.Ty = RetTy->getPointerTo();
2375 Entry.isSExt = isSigned;
2376 Entry.isZExt = !isSigned;
2377 Args.push_back(Entry);
2379 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2380 TLI.getPointerTy(DAG.getDataLayout()));
2383 TargetLowering::CallLoweringInfo CLI(DAG);
2384 CLI.setDebugLoc(dl).setChain(InChain)
2385 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2386 .setSExtResult(isSigned).setZExtResult(!isSigned);
2388 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2390 // Remainder is loaded back from the stack frame.
2391 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2392 MachinePointerInfo(), false, false, false, 0);
2393 Results.push_back(CallInfo.first);
2394 Results.push_back(Rem);
2397 /// Return true if sincos libcall is available.
2398 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2400 switch (Node->getSimpleValueType(0).SimpleTy) {
2401 default: llvm_unreachable("Unexpected request for libcall!");
2402 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2403 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2404 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2405 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2406 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2408 return TLI.getLibcallName(LC) != nullptr;
2411 /// Return true if sincos libcall is available and can be used to combine sin
2413 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2414 const TargetMachine &TM) {
2415 if (!isSinCosLibcallAvailable(Node, TLI))
2417 // GNU sin/cos functions set errno while sincos does not. Therefore
2418 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2419 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2420 if (isGNU && !TM.Options.UnsafeFPMath)
2425 /// Only issue sincos libcall if both sin and cos are needed.
2426 static bool useSinCos(SDNode *Node) {
2427 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2428 ? ISD::FCOS : ISD::FSIN;
2430 SDValue Op0 = Node->getOperand(0);
2431 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2432 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2436 // The other user might have been turned into sincos already.
2437 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2443 /// Issue libcalls to sincos to compute sin / cos pairs.
2445 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2446 SmallVectorImpl<SDValue> &Results) {
2448 switch (Node->getSimpleValueType(0).SimpleTy) {
2449 default: llvm_unreachable("Unexpected request for libcall!");
2450 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2451 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2452 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2453 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2454 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2457 // The input chain to this libcall is the entry node of the function.
2458 // Legalizing the call will automatically add the previous call to the
2460 SDValue InChain = DAG.getEntryNode();
2462 EVT RetVT = Node->getValueType(0);
2463 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2465 TargetLowering::ArgListTy Args;
2466 TargetLowering::ArgListEntry Entry;
2468 // Pass the argument.
2469 Entry.Node = Node->getOperand(0);
2471 Entry.isSExt = false;
2472 Entry.isZExt = false;
2473 Args.push_back(Entry);
2475 // Pass the return address of sin.
2476 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2477 Entry.Node = SinPtr;
2478 Entry.Ty = RetTy->getPointerTo();
2479 Entry.isSExt = false;
2480 Entry.isZExt = false;
2481 Args.push_back(Entry);
2483 // Also pass the return address of the cos.
2484 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2485 Entry.Node = CosPtr;
2486 Entry.Ty = RetTy->getPointerTo();
2487 Entry.isSExt = false;
2488 Entry.isZExt = false;
2489 Args.push_back(Entry);
2491 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2492 TLI.getPointerTy(DAG.getDataLayout()));
2495 TargetLowering::CallLoweringInfo CLI(DAG);
2496 CLI.setDebugLoc(dl).setChain(InChain)
2497 .setCallee(TLI.getLibcallCallingConv(LC),
2498 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2500 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2502 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2503 MachinePointerInfo(), false, false, false, 0));
2504 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2505 MachinePointerInfo(), false, false, false, 0));
2508 /// This function is responsible for legalizing a
2509 /// INT_TO_FP operation of the specified operand when the target requests that
2510 /// we expand it. At this point, we know that the result and operand types are
2511 /// legal for the target.
2512 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2516 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2517 // simple 32-bit [signed|unsigned] integer to float/double expansion
2519 // Get the stack frame index of a 8 byte buffer.
2520 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2522 // word offset constant for Hi/Lo address computation
2523 SDValue WordOff = DAG.getConstant(sizeof(int), dl,
2524 StackSlot.getValueType());
2525 // set up Hi and Lo (into buffer) address based on endian
2526 SDValue Hi = StackSlot;
2527 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2528 StackSlot, WordOff);
2529 if (DAG.getDataLayout().isLittleEndian())
2532 // if signed map to unsigned space
2535 // constant used to invert sign bit (signed to unsigned mapping)
2536 SDValue SignBit = DAG.getConstant(0x80000000u, dl, MVT::i32);
2537 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2541 // store the lo of the constructed double - based on integer input
2542 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2543 Op0Mapped, Lo, MachinePointerInfo(),
2545 // initial hi portion of constructed double
2546 SDValue InitialHi = DAG.getConstant(0x43300000u, dl, MVT::i32);
2547 // store the hi of the constructed double - biased exponent
2548 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2549 MachinePointerInfo(),
2551 // load the constructed double
2552 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2553 MachinePointerInfo(), false, false, false, 0);
2554 // FP constant to bias correct the final result
2555 SDValue Bias = DAG.getConstantFP(isSigned ?
2556 BitsToDouble(0x4330000080000000ULL) :
2557 BitsToDouble(0x4330000000000000ULL),
2559 // subtract the bias
2560 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2563 // handle final rounding
2564 if (DestVT == MVT::f64) {
2567 } else if (DestVT.bitsLT(MVT::f64)) {
2568 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2569 DAG.getIntPtrConstant(0, dl));
2570 } else if (DestVT.bitsGT(MVT::f64)) {
2571 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2575 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2576 // Code below here assumes !isSigned without checking again.
2578 // Implementation of unsigned i64 to f64 following the algorithm in
2579 // __floatundidf in compiler_rt. This implementation has the advantage
2580 // of performing rounding correctly, both in the default rounding mode
2581 // and in all alternate rounding modes.
2582 // TODO: Generalize this for use with other types.
2583 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2585 DAG.getConstant(UINT64_C(0x4330000000000000), dl, MVT::i64);
2586 SDValue TwoP84PlusTwoP52 =
2587 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), dl,
2590 DAG.getConstant(UINT64_C(0x4530000000000000), dl, MVT::i64);
2592 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2593 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2594 DAG.getConstant(32, dl, MVT::i64));
2595 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2596 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2597 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2598 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2599 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2601 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2604 // Implementation of unsigned i64 to f32.
2605 // TODO: Generalize this for use with other types.
2606 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2607 // For unsigned conversions, convert them to signed conversions using the
2608 // algorithm from the x86_64 __floatundidf in compiler_rt.
2610 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2612 SDValue ShiftConst = DAG.getConstant(
2613 1, dl, TLI.getShiftAmountTy(Op0.getValueType(), DAG.getDataLayout()));
2614 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2615 SDValue AndConst = DAG.getConstant(1, dl, MVT::i64);
2616 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2617 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2619 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2620 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2622 // TODO: This really should be implemented using a branch rather than a
2623 // select. We happen to get lucky and machinesink does the right
2624 // thing most of the time. This would be a good candidate for a
2625 //pseudo-op, or, even better, for whole-function isel.
2626 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2627 Op0, DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
2628 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2631 // Otherwise, implement the fully general conversion.
2633 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2634 DAG.getConstant(UINT64_C(0xfffffffffffff800), dl, MVT::i64));
2635 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2636 DAG.getConstant(UINT64_C(0x800), dl, MVT::i64));
2637 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2638 DAG.getConstant(UINT64_C(0x7ff), dl, MVT::i64));
2639 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), And2,
2640 DAG.getConstant(UINT64_C(0), dl, MVT::i64),
2642 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2643 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64), Op0,
2644 DAG.getConstant(UINT64_C(0x0020000000000000), dl,
2647 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2648 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType(), DAG.getDataLayout());
2650 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2651 DAG.getConstant(32, dl, SHVT));
2652 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2653 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2655 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), dl,
2657 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2658 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2659 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2660 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2661 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2662 DAG.getIntPtrConstant(0, dl));
2665 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2667 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2669 DAG.getConstant(0, dl, Op0.getValueType()),
2671 SDValue Zero = DAG.getIntPtrConstant(0, dl),
2672 Four = DAG.getIntPtrConstant(4, dl);
2673 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2674 SignSet, Four, Zero);
2676 // If the sign bit of the integer is set, the large number will be treated
2677 // as a negative number. To counteract this, the dynamic code adds an
2678 // offset depending on the data type.
2680 switch (Op0.getSimpleValueType().SimpleTy) {
2681 default: llvm_unreachable("Unsupported integer type!");
2682 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2683 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2684 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2685 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2687 if (DAG.getDataLayout().isLittleEndian())
2689 Constant *FudgeFactor = ConstantInt::get(
2690 Type::getInt64Ty(*DAG.getContext()), FF);
2693 DAG.getConstantPool(FudgeFactor, TLI.getPointerTy(DAG.getDataLayout()));
2694 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2695 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2696 Alignment = std::min(Alignment, 4u);
2698 if (DestVT == MVT::f32)
2699 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2700 MachinePointerInfo::getConstantPool(),
2701 false, false, false, Alignment);
2703 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2704 DAG.getEntryNode(), CPIdx,
2705 MachinePointerInfo::getConstantPool(),
2706 MVT::f32, false, false, false, Alignment);
2707 HandleSDNode Handle(Load);
2708 LegalizeOp(Load.getNode());
2709 FudgeInReg = Handle.getValue();
2712 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2715 /// This function is responsible for legalizing a
2716 /// *INT_TO_FP operation of the specified operand when the target requests that
2717 /// we promote it. At this point, we know that the result and operand types are
2718 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2719 /// operation that takes a larger input.
2720 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2724 // First step, figure out the appropriate *INT_TO_FP operation to use.
2725 EVT NewInTy = LegalOp.getValueType();
2727 unsigned OpToUse = 0;
2729 // Scan for the appropriate larger type to use.
2731 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2732 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2734 // If the target supports SINT_TO_FP of this type, use it.
2735 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2736 OpToUse = ISD::SINT_TO_FP;
2739 if (isSigned) continue;
2741 // If the target supports UINT_TO_FP of this type, use it.
2742 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2743 OpToUse = ISD::UINT_TO_FP;
2747 // Otherwise, try a larger type.
2750 // Okay, we found the operation and type to use. Zero extend our input to the
2751 // desired type then run the operation on it.
2752 return DAG.getNode(OpToUse, dl, DestVT,
2753 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2754 dl, NewInTy, LegalOp));
2757 /// This function is responsible for legalizing a
2758 /// FP_TO_*INT operation of the specified operand when the target requests that
2759 /// we promote it. At this point, we know that the result and operand types are
2760 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2761 /// operation that returns a larger result.
2762 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2766 // First step, figure out the appropriate FP_TO*INT operation to use.
2767 EVT NewOutTy = DestVT;
2769 unsigned OpToUse = 0;
2771 // Scan for the appropriate larger type to use.
2773 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2774 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2776 // A larger signed type can hold all unsigned values of the requested type,
2777 // so using FP_TO_SINT is valid
2778 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2779 OpToUse = ISD::FP_TO_SINT;
2783 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2784 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2785 OpToUse = ISD::FP_TO_UINT;
2789 // Otherwise, try a larger type.
2793 // Okay, we found the operation and type to use.
2794 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2796 // Truncate the result of the extended FP_TO_*INT operation to the desired
2798 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2801 /// Open code the operations for BSWAP of the specified operation.
2802 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2803 EVT VT = Op.getValueType();
2804 EVT SHVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2805 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2806 switch (VT.getSimpleVT().SimpleTy) {
2807 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2809 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2810 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2811 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2813 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2814 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2815 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2816 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2817 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2818 DAG.getConstant(0xFF0000, dl, VT));
2819 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, dl, VT));
2820 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2821 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2822 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2824 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2825 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2826 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2827 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2828 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, dl, SHVT));
2829 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, dl, SHVT));
2830 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, dl, SHVT));
2831 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, dl, SHVT));
2832 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7,
2833 DAG.getConstant(255ULL<<48, dl, VT));
2834 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6,
2835 DAG.getConstant(255ULL<<40, dl, VT));
2836 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5,
2837 DAG.getConstant(255ULL<<32, dl, VT));
2838 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4,
2839 DAG.getConstant(255ULL<<24, dl, VT));
2840 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3,
2841 DAG.getConstant(255ULL<<16, dl, VT));
2842 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2,
2843 DAG.getConstant(255ULL<<8 , dl, VT));
2844 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2845 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2846 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2847 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2848 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2849 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2850 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2854 /// Expand the specified bitcount instruction into operations.
2855 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2858 default: llvm_unreachable("Cannot expand this yet!");
2860 EVT VT = Op.getValueType();
2861 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2862 unsigned Len = VT.getSizeInBits();
2864 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2865 "CTPOP not implemented for this type.");
2867 // This is the "best" algorithm from
2868 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2870 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)),
2872 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)),
2874 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)),
2876 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)),
2879 // v = v - ((v >> 1) & 0x55555555...)
2880 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2881 DAG.getNode(ISD::AND, dl, VT,
2882 DAG.getNode(ISD::SRL, dl, VT, Op,
2883 DAG.getConstant(1, dl, ShVT)),
2885 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2886 Op = DAG.getNode(ISD::ADD, dl, VT,
2887 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2888 DAG.getNode(ISD::AND, dl, VT,
2889 DAG.getNode(ISD::SRL, dl, VT, Op,
2890 DAG.getConstant(2, dl, ShVT)),
2892 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2893 Op = DAG.getNode(ISD::AND, dl, VT,
2894 DAG.getNode(ISD::ADD, dl, VT, Op,
2895 DAG.getNode(ISD::SRL, dl, VT, Op,
2896 DAG.getConstant(4, dl, ShVT))),
2898 // v = (v * 0x01010101...) >> (Len - 8)
2899 Op = DAG.getNode(ISD::SRL, dl, VT,
2900 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2901 DAG.getConstant(Len - 8, dl, ShVT));
2905 case ISD::CTLZ_ZERO_UNDEF:
2906 // This trivially expands to CTLZ.
2907 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2909 // for now, we do this:
2910 // x = x | (x >> 1);
2911 // x = x | (x >> 2);
2913 // x = x | (x >>16);
2914 // x = x | (x >>32); // for 64-bit input
2915 // return popcount(~x);
2917 // Ref: "Hacker's Delight" by Henry Warren
2918 EVT VT = Op.getValueType();
2919 EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
2920 unsigned len = VT.getSizeInBits();
2921 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2922 SDValue Tmp3 = DAG.getConstant(1ULL << i, dl, ShVT);
2923 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2924 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2926 Op = DAG.getNOT(dl, Op, VT);
2927 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2929 case ISD::CTTZ_ZERO_UNDEF:
2930 // This trivially expands to CTTZ.
2931 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2933 // for now, we use: { return popcount(~x & (x - 1)); }
2934 // unless the target has ctlz but not ctpop, in which case we use:
2935 // { return 32 - nlz(~x & (x-1)); }
2936 // Ref: "Hacker's Delight" by Henry Warren
2937 EVT VT = Op.getValueType();
2938 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2939 DAG.getNOT(dl, Op, VT),
2940 DAG.getNode(ISD::SUB, dl, VT, Op,
2941 DAG.getConstant(1, dl, VT)));
2942 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2943 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2944 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2945 return DAG.getNode(ISD::SUB, dl, VT,
2946 DAG.getConstant(VT.getSizeInBits(), dl, VT),
2947 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2948 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2953 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2954 unsigned Opc = Node->getOpcode();
2955 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2956 RTLIB::Libcall LC = RTLIB::getATOMIC(Opc, VT);
2957 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unexpected atomic op or value type!");
2959 return ExpandChainLibCall(LC, Node, false);
2962 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2963 SmallVector<SDValue, 8> Results;
2965 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2967 switch (Node->getOpcode()) {
2970 case ISD::CTLZ_ZERO_UNDEF:
2972 case ISD::CTTZ_ZERO_UNDEF:
2973 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2974 Results.push_back(Tmp1);
2977 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2979 case ISD::FRAMEADDR:
2980 case ISD::RETURNADDR:
2981 case ISD::FRAME_TO_ARGS_OFFSET:
2982 Results.push_back(DAG.getConstant(0, dl, Node->getValueType(0)));
2984 case ISD::FLT_ROUNDS_:
2985 Results.push_back(DAG.getConstant(1, dl, Node->getValueType(0)));
2987 case ISD::EH_RETURN:
2991 case ISD::EH_SJLJ_LONGJMP:
2992 // If the target didn't expand these, there's nothing to do, so just
2993 // preserve the chain and be done.
2994 Results.push_back(Node->getOperand(0));
2996 case ISD::EH_SJLJ_SETJMP:
2997 // If the target didn't expand this, just return 'zero' and preserve the
2999 Results.push_back(DAG.getConstant(0, dl, MVT::i32));
3000 Results.push_back(Node->getOperand(0));
3002 case ISD::ATOMIC_FENCE: {
3003 // If the target didn't lower this, lower it to '__sync_synchronize()' call
3004 // FIXME: handle "fence singlethread" more efficiently.
3005 TargetLowering::ArgListTy Args;
3007 TargetLowering::CallLoweringInfo CLI(DAG);
3009 .setChain(Node->getOperand(0))
3010 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3011 DAG.getExternalSymbol("__sync_synchronize",
3012 TLI.getPointerTy(DAG.getDataLayout())),
3013 std::move(Args), 0);
3015 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3017 Results.push_back(CallResult.second);
3020 case ISD::ATOMIC_LOAD: {
3021 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3022 SDValue Zero = DAG.getConstant(0, dl, Node->getValueType(0));
3023 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3024 SDValue Swap = DAG.getAtomicCmpSwap(
3025 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3026 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3027 cast<AtomicSDNode>(Node)->getMemOperand(),
3028 cast<AtomicSDNode>(Node)->getOrdering(),
3029 cast<AtomicSDNode>(Node)->getOrdering(),
3030 cast<AtomicSDNode>(Node)->getSynchScope());
3031 Results.push_back(Swap.getValue(0));
3032 Results.push_back(Swap.getValue(1));
3035 case ISD::ATOMIC_STORE: {
3036 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3037 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3038 cast<AtomicSDNode>(Node)->getMemoryVT(),
3039 Node->getOperand(0),
3040 Node->getOperand(1), Node->getOperand(2),
3041 cast<AtomicSDNode>(Node)->getMemOperand(),
3042 cast<AtomicSDNode>(Node)->getOrdering(),
3043 cast<AtomicSDNode>(Node)->getSynchScope());
3044 Results.push_back(Swap.getValue(1));
3047 // By default, atomic intrinsics are marked Legal and lowered. Targets
3048 // which don't support them directly, however, may want libcalls, in which
3049 // case they mark them Expand, and we get here.
3050 case ISD::ATOMIC_SWAP:
3051 case ISD::ATOMIC_LOAD_ADD:
3052 case ISD::ATOMIC_LOAD_SUB:
3053 case ISD::ATOMIC_LOAD_AND:
3054 case ISD::ATOMIC_LOAD_OR:
3055 case ISD::ATOMIC_LOAD_XOR:
3056 case ISD::ATOMIC_LOAD_NAND:
3057 case ISD::ATOMIC_LOAD_MIN:
3058 case ISD::ATOMIC_LOAD_MAX:
3059 case ISD::ATOMIC_LOAD_UMIN:
3060 case ISD::ATOMIC_LOAD_UMAX:
3061 case ISD::ATOMIC_CMP_SWAP: {
3062 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3063 Results.push_back(Tmp.first);
3064 Results.push_back(Tmp.second);
3067 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3068 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3069 // splits out the success value as a comparison. Expanding the resulting
3070 // ATOMIC_CMP_SWAP will produce a libcall.
3071 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3072 SDValue Res = DAG.getAtomicCmpSwap(
3073 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3074 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3075 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3076 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3077 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3078 cast<AtomicSDNode>(Node)->getSynchScope());
3080 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3081 Res, Node->getOperand(2), ISD::SETEQ);
3083 Results.push_back(Res.getValue(0));
3084 Results.push_back(Success);
3085 Results.push_back(Res.getValue(1));
3088 case ISD::DYNAMIC_STACKALLOC:
3089 ExpandDYNAMIC_STACKALLOC(Node, Results);
3091 case ISD::MERGE_VALUES:
3092 for (unsigned i = 0; i < Node->getNumValues(); i++)
3093 Results.push_back(Node->getOperand(i));
3096 EVT VT = Node->getValueType(0);
3098 Results.push_back(DAG.getConstant(0, dl, VT));
3100 assert(VT.isFloatingPoint() && "Unknown value type!");
3101 Results.push_back(DAG.getConstantFP(0, dl, VT));
3106 // If this operation is not supported, lower it to 'abort()' call
3107 TargetLowering::ArgListTy Args;
3108 TargetLowering::CallLoweringInfo CLI(DAG);
3110 .setChain(Node->getOperand(0))
3111 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3112 DAG.getExternalSymbol("abort",
3113 TLI.getPointerTy(DAG.getDataLayout())),
3114 std::move(Args), 0);
3115 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3117 Results.push_back(CallResult.second);
3122 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3123 Node->getValueType(0), dl);
3124 Results.push_back(Tmp1);
3126 case ISD::FP_EXTEND:
3127 Tmp1 = EmitStackConvert(Node->getOperand(0),
3128 Node->getOperand(0).getValueType(),
3129 Node->getValueType(0), dl);
3130 Results.push_back(Tmp1);
3132 case ISD::SIGN_EXTEND_INREG: {
3133 // NOTE: we could fall back on load/store here too for targets without
3134 // SAR. However, it is doubtful that any exist.
3135 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3136 EVT VT = Node->getValueType(0);
3137 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT, DAG.getDataLayout());
3140 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3141 ExtraVT.getScalarType().getSizeInBits();
3142 SDValue ShiftCst = DAG.getConstant(BitsDiff, dl, ShiftAmountTy);
3143 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3144 Node->getOperand(0), ShiftCst);
3145 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3146 Results.push_back(Tmp1);
3149 case ISD::FP_ROUND_INREG: {
3150 // The only way we can lower this is to turn it into a TRUNCSTORE,
3151 // EXTLOAD pair, targeting a temporary location (a stack slot).
3153 // NOTE: there is a choice here between constantly creating new stack
3154 // slots and always reusing the same one. We currently always create
3155 // new ones, as reuse may inhibit scheduling.
3156 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3157 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3158 Node->getValueType(0), dl);
3159 Results.push_back(Tmp1);
3162 case ISD::SINT_TO_FP:
3163 case ISD::UINT_TO_FP:
3164 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3165 Node->getOperand(0), Node->getValueType(0), dl);
3166 Results.push_back(Tmp1);
3168 case ISD::FP_TO_SINT:
3169 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3170 Results.push_back(Tmp1);
3172 case ISD::FP_TO_UINT: {
3173 SDValue True, False;
3174 EVT VT = Node->getOperand(0).getValueType();
3175 EVT NVT = Node->getValueType(0);
3176 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3177 APInt::getNullValue(VT.getSizeInBits()));
3178 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3179 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3180 Tmp1 = DAG.getConstantFP(apf, dl, VT);
3181 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3182 Node->getOperand(0),
3184 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3185 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3186 DAG.getNode(ISD::FSUB, dl, VT,
3187 Node->getOperand(0), Tmp1));
3188 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3189 DAG.getConstant(x, dl, NVT));
3190 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3191 Results.push_back(Tmp1);
3195 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3196 EVT VT = Node->getValueType(0);
3197 Tmp1 = Node->getOperand(0);
3198 Tmp2 = Node->getOperand(1);
3199 unsigned Align = Node->getConstantOperandVal(3);
3201 SDValue VAListLoad =
3202 DAG.getLoad(TLI.getPointerTy(DAG.getDataLayout()), dl, Tmp1, Tmp2,
3203 MachinePointerInfo(V), false, false, false, 0);
3204 SDValue VAList = VAListLoad;
3206 if (Align > TLI.getMinStackArgumentAlignment()) {
3207 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3209 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3210 DAG.getConstant(Align - 1, dl,
3211 VAList.getValueType()));
3213 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3214 DAG.getConstant(-(int64_t)Align, dl,
3215 VAList.getValueType()));
3218 // Increment the pointer, VAList, to the next vaarg
3219 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3220 DAG.getConstant(DAG.getDataLayout().getTypeAllocSize(
3221 VT.getTypeForEVT(*DAG.getContext())),
3222 dl, VAList.getValueType()));
3223 // Store the incremented VAList to the legalized pointer
3224 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3225 MachinePointerInfo(V), false, false, 0);
3226 // Load the actual argument out of the pointer VAList
3227 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3228 false, false, false, 0));
3229 Results.push_back(Results[0].getValue(1));
3233 // This defaults to loading a pointer from the input and storing it to the
3234 // output, returning the chain.
3235 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3236 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3237 Tmp1 = DAG.getLoad(TLI.getPointerTy(DAG.getDataLayout()), dl,
3238 Node->getOperand(0), Node->getOperand(2),
3239 MachinePointerInfo(VS), false, false, false, 0);
3240 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3241 MachinePointerInfo(VD), false, false, 0);
3242 Results.push_back(Tmp1);
3245 case ISD::EXTRACT_VECTOR_ELT:
3246 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3247 // This must be an access of the only element. Return it.
3248 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3249 Node->getOperand(0));
3251 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3252 Results.push_back(Tmp1);
3254 case ISD::EXTRACT_SUBVECTOR:
3255 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3257 case ISD::INSERT_SUBVECTOR:
3258 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3260 case ISD::CONCAT_VECTORS: {
3261 Results.push_back(ExpandVectorBuildThroughStack(Node));
3264 case ISD::SCALAR_TO_VECTOR:
3265 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3267 case ISD::INSERT_VECTOR_ELT:
3268 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3269 Node->getOperand(1),
3270 Node->getOperand(2), dl));
3272 case ISD::VECTOR_SHUFFLE: {
3273 SmallVector<int, 32> NewMask;
3274 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3276 EVT VT = Node->getValueType(0);
3277 EVT EltVT = VT.getVectorElementType();
3278 SDValue Op0 = Node->getOperand(0);
3279 SDValue Op1 = Node->getOperand(1);
3280 if (!TLI.isTypeLegal(EltVT)) {
3282 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3284 // BUILD_VECTOR operands are allowed to be wider than the element type.
3285 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3287 if (NewEltVT.bitsLT(EltVT)) {
3289 // Convert shuffle node.
3290 // If original node was v4i64 and the new EltVT is i32,
3291 // cast operands to v8i32 and re-build the mask.
3293 // Calculate new VT, the size of the new VT should be equal to original.
3295 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3296 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3297 assert(NewVT.bitsEq(VT));
3299 // cast operands to new VT
3300 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3301 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3303 // Convert the shuffle mask
3304 unsigned int factor =
3305 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3307 // EltVT gets smaller
3310 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3312 for (unsigned fi = 0; fi < factor; ++fi)
3313 NewMask.push_back(Mask[i]);
3316 for (unsigned fi = 0; fi < factor; ++fi)
3317 NewMask.push_back(Mask[i]*factor+fi);
3325 unsigned NumElems = VT.getVectorNumElements();
3326 SmallVector<SDValue, 16> Ops;
3327 for (unsigned i = 0; i != NumElems; ++i) {
3329 Ops.push_back(DAG.getUNDEF(EltVT));
3332 unsigned Idx = Mask[i];
3334 Ops.push_back(DAG.getNode(
3335 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op0,
3336 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout()))));
3338 Ops.push_back(DAG.getNode(
3339 ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Op1,
3340 DAG.getConstant(Idx - NumElems, dl,
3341 TLI.getVectorIdxTy(DAG.getDataLayout()))));
3344 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3345 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3346 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3347 Results.push_back(Tmp1);
3350 case ISD::EXTRACT_ELEMENT: {
3351 EVT OpTy = Node->getOperand(0).getValueType();
3352 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3354 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3355 DAG.getConstant(OpTy.getSizeInBits() / 2, dl,
3356 TLI.getShiftAmountTy(
3357 Node->getOperand(0).getValueType(),
3358 DAG.getDataLayout())));
3359 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3362 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3363 Node->getOperand(0));
3365 Results.push_back(Tmp1);
3368 case ISD::STACKSAVE:
3369 // Expand to CopyFromReg if the target set
3370 // StackPointerRegisterToSaveRestore.
3371 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3372 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3373 Node->getValueType(0)));
3374 Results.push_back(Results[0].getValue(1));
3376 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3377 Results.push_back(Node->getOperand(0));
3380 case ISD::STACKRESTORE:
3381 // Expand to CopyToReg if the target set
3382 // StackPointerRegisterToSaveRestore.
3383 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3384 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3385 Node->getOperand(1)));
3387 Results.push_back(Node->getOperand(0));
3390 case ISD::FCOPYSIGN:
3391 Results.push_back(ExpandFCOPYSIGN(Node));
3394 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3395 Tmp1 = DAG.getConstantFP(-0.0, dl, Node->getValueType(0));
3396 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3397 Node->getOperand(0));
3398 Results.push_back(Tmp1);
3401 Results.push_back(ExpandFABS(Node));
3407 // Expand Y = MAX(A, B) -> Y = (A > B) ? A : B
3409 switch (Node->getOpcode()) {
3410 default: llvm_unreachable("How did we get here?");
3411 case ISD::SMAX: Pred = ISD::SETGT; break;
3412 case ISD::SMIN: Pred = ISD::SETLT; break;
3413 case ISD::UMAX: Pred = ISD::SETUGT; break;
3414 case ISD::UMIN: Pred = ISD::SETULT; break;
3416 Tmp1 = Node->getOperand(0);
3417 Tmp2 = Node->getOperand(1);
3418 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp1, Tmp2, Pred);
3419 Results.push_back(Tmp1);
3424 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
3425 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
3426 RTLIB::FMIN_PPCF128));
3429 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
3430 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
3431 RTLIB::FMAX_PPCF128));
3434 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3435 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3436 RTLIB::SQRT_PPCF128));
3440 EVT VT = Node->getValueType(0);
3441 bool isSIN = Node->getOpcode() == ISD::FSIN;
3442 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3443 // fcos which share the same operand and both are used.
3444 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3445 canCombineSinCosLibcall(Node, TLI, TM))
3446 && useSinCos(Node)) {
3447 SDVTList VTs = DAG.getVTList(VT, VT);
3448 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3450 Tmp1 = Tmp1.getValue(1);
3451 Results.push_back(Tmp1);
3453 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3454 RTLIB::SIN_F80, RTLIB::SIN_F128,
3455 RTLIB::SIN_PPCF128));
3457 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3458 RTLIB::COS_F80, RTLIB::COS_F128,
3459 RTLIB::COS_PPCF128));
3464 // Expand into sincos libcall.
3465 ExpandSinCosLibCall(Node, Results);
3468 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3469 RTLIB::LOG_F80, RTLIB::LOG_F128,
3470 RTLIB::LOG_PPCF128));
3473 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3474 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3475 RTLIB::LOG2_PPCF128));
3478 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3479 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3480 RTLIB::LOG10_PPCF128));
3483 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3484 RTLIB::EXP_F80, RTLIB::EXP_F128,
3485 RTLIB::EXP_PPCF128));
3488 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3489 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3490 RTLIB::EXP2_PPCF128));
3493 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3494 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3495 RTLIB::TRUNC_PPCF128));
3498 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3499 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3500 RTLIB::FLOOR_PPCF128));
3503 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3504 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3505 RTLIB::CEIL_PPCF128));
3508 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3509 RTLIB::RINT_F80, RTLIB::RINT_F128,
3510 RTLIB::RINT_PPCF128));
3512 case ISD::FNEARBYINT:
3513 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3514 RTLIB::NEARBYINT_F64,
3515 RTLIB::NEARBYINT_F80,
3516 RTLIB::NEARBYINT_F128,
3517 RTLIB::NEARBYINT_PPCF128));
3520 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3524 RTLIB::ROUND_PPCF128));
3527 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3528 RTLIB::POWI_F80, RTLIB::POWI_F128,
3529 RTLIB::POWI_PPCF128));
3532 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3533 RTLIB::POW_F80, RTLIB::POW_F128,
3534 RTLIB::POW_PPCF128));
3537 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3538 RTLIB::DIV_F80, RTLIB::DIV_F128,
3539 RTLIB::DIV_PPCF128));
3542 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3543 RTLIB::REM_F80, RTLIB::REM_F128,
3544 RTLIB::REM_PPCF128));
3547 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3548 RTLIB::FMA_F80, RTLIB::FMA_F128,
3549 RTLIB::FMA_PPCF128));
3552 llvm_unreachable("Illegal fmad should never be formed");
3555 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
3556 RTLIB::ADD_F80, RTLIB::ADD_F128,
3557 RTLIB::ADD_PPCF128));
3560 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
3561 RTLIB::MUL_F80, RTLIB::MUL_F128,
3562 RTLIB::MUL_PPCF128));
3564 case ISD::FP16_TO_FP: {
3565 if (Node->getValueType(0) == MVT::f32) {
3566 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3570 // We can extend to types bigger than f32 in two steps without changing the
3571 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3572 // the option of emitting that before resorting to a libcall.
3574 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3576 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3579 case ISD::FP_TO_FP16: {
3580 if (!TLI.useSoftFloat() && TM.Options.UnsafeFPMath) {
3581 SDValue Op = Node->getOperand(0);
3582 MVT SVT = Op.getSimpleValueType();
3583 if ((SVT == MVT::f64 || SVT == MVT::f80) &&
3584 TLI.isOperationLegalOrCustom(ISD::FP_TO_FP16, MVT::f32)) {
3585 // Under fastmath, we can expand this node into a fround followed by
3586 // a float-half conversion.
3587 SDValue FloatVal = DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Op,
3588 DAG.getIntPtrConstant(0, dl));
3590 DAG.getNode(ISD::FP_TO_FP16, dl, MVT::i16, FloatVal));
3596 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3597 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3598 Results.push_back(ExpandLibCall(LC, Node, false));
3601 case ISD::ConstantFP: {
3602 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3603 // Check to see if this FP immediate is already legal.
3604 // If this is a legal constant, turn it into a TargetConstantFP node.
3605 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3606 Results.push_back(ExpandConstantFP(CFP, true));
3610 EVT VT = Node->getValueType(0);
3611 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3612 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3613 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3614 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3615 Results.push_back(Tmp1);
3617 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
3618 RTLIB::SUB_F80, RTLIB::SUB_F128,
3619 RTLIB::SUB_PPCF128));
3624 EVT VT = Node->getValueType(0);
3625 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3626 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3627 "Don't know how to expand this subtraction!");
3628 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3629 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), dl,
3631 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, dl, VT));
3632 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3637 EVT VT = Node->getValueType(0);
3638 bool isSigned = Node->getOpcode() == ISD::SREM;
3639 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3640 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3641 Tmp2 = Node->getOperand(0);
3642 Tmp3 = Node->getOperand(1);
3643 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3644 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3645 // If div is legal, it's better to do the normal expansion
3646 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3647 useDivRem(Node, isSigned, false))) {
3648 SDVTList VTs = DAG.getVTList(VT, VT);
3649 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3650 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3652 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3653 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3654 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3655 } else if (isSigned)
3656 Tmp1 = ExpandIntLibCall(Node, true,
3658 RTLIB::SREM_I16, RTLIB::SREM_I32,
3659 RTLIB::SREM_I64, RTLIB::SREM_I128);
3661 Tmp1 = ExpandIntLibCall(Node, false,
3663 RTLIB::UREM_I16, RTLIB::UREM_I32,
3664 RTLIB::UREM_I64, RTLIB::UREM_I128);
3665 Results.push_back(Tmp1);
3670 bool isSigned = Node->getOpcode() == ISD::SDIV;
3671 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3672 EVT VT = Node->getValueType(0);
3673 SDVTList VTs = DAG.getVTList(VT, VT);
3674 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3675 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3676 useDivRem(Node, isSigned, true)))
3677 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3678 Node->getOperand(1));
3680 Tmp1 = ExpandIntLibCall(Node, true,
3682 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3683 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3685 Tmp1 = ExpandIntLibCall(Node, false,
3687 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3688 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3689 Results.push_back(Tmp1);
3694 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3696 EVT VT = Node->getValueType(0);
3697 SDVTList VTs = DAG.getVTList(VT, VT);
3698 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3699 "If this wasn't legal, it shouldn't have been created!");
3700 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3701 Node->getOperand(1));
3702 Results.push_back(Tmp1.getValue(1));
3707 // Expand into divrem libcall
3708 ExpandDivRemLibCall(Node, Results);
3711 EVT VT = Node->getValueType(0);
3712 SDVTList VTs = DAG.getVTList(VT, VT);
3713 // See if multiply or divide can be lowered using two-result operations.
3714 // We just need the low half of the multiply; try both the signed
3715 // and unsigned forms. If the target supports both SMUL_LOHI and
3716 // UMUL_LOHI, form a preference by checking which forms of plain
3717 // MULH it supports.
3718 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3719 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3720 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3721 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3722 unsigned OpToUse = 0;
3723 if (HasSMUL_LOHI && !HasMULHS) {
3724 OpToUse = ISD::SMUL_LOHI;
3725 } else if (HasUMUL_LOHI && !HasMULHU) {
3726 OpToUse = ISD::UMUL_LOHI;
3727 } else if (HasSMUL_LOHI) {
3728 OpToUse = ISD::SMUL_LOHI;
3729 } else if (HasUMUL_LOHI) {
3730 OpToUse = ISD::UMUL_LOHI;
3733 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3734 Node->getOperand(1)));
3739 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3740 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3741 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3742 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3743 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3744 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3745 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3746 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3748 DAG.getConstant(HalfType.getSizeInBits(), dl,
3749 TLI.getShiftAmountTy(HalfType, DAG.getDataLayout()));
3750 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3751 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3755 Tmp1 = ExpandIntLibCall(Node, false,
3757 RTLIB::MUL_I16, RTLIB::MUL_I32,
3758 RTLIB::MUL_I64, RTLIB::MUL_I128);
3759 Results.push_back(Tmp1);
3764 SDValue LHS = Node->getOperand(0);
3765 SDValue RHS = Node->getOperand(1);
3766 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3767 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3769 Results.push_back(Sum);
3770 EVT ResultType = Node->getValueType(1);
3771 EVT OType = getSetCCResultType(Node->getValueType(0));
3773 SDValue Zero = DAG.getConstant(0, dl, LHS.getValueType());
3775 // LHSSign -> LHS >= 0
3776 // RHSSign -> RHS >= 0
3777 // SumSign -> Sum >= 0
3780 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3782 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3784 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3785 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3786 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3787 Node->getOpcode() == ISD::SADDO ?
3788 ISD::SETEQ : ISD::SETNE);
3790 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3791 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3793 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3794 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3799 SDValue LHS = Node->getOperand(0);
3800 SDValue RHS = Node->getOperand(1);
3801 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3802 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3804 Results.push_back(Sum);
3806 EVT ResultType = Node->getValueType(1);
3807 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3809 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3810 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3812 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3817 EVT VT = Node->getValueType(0);
3818 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3819 SDValue LHS = Node->getOperand(0);
3820 SDValue RHS = Node->getOperand(1);
3823 static const unsigned Ops[2][3] =
3824 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3825 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3826 bool isSigned = Node->getOpcode() == ISD::SMULO;
3827 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3828 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3829 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3830 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3831 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3833 TopHalf = BottomHalf.getValue(1);
3834 } else if (TLI.isTypeLegal(WideVT)) {
3835 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3836 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3837 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3838 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3839 DAG.getIntPtrConstant(0, dl));
3840 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3841 DAG.getIntPtrConstant(1, dl));
3843 // We can fall back to a libcall with an illegal type for the MUL if we
3844 // have a libcall big enough.
3845 // Also, we can fall back to a division in some cases, but that's a big
3846 // performance hit in the general case.
3847 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3848 if (WideVT == MVT::i16)
3849 LC = RTLIB::MUL_I16;
3850 else if (WideVT == MVT::i32)
3851 LC = RTLIB::MUL_I32;
3852 else if (WideVT == MVT::i64)
3853 LC = RTLIB::MUL_I64;
3854 else if (WideVT == MVT::i128)
3855 LC = RTLIB::MUL_I128;
3856 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3858 // The high part is obtained by SRA'ing all but one of the bits of low
3860 unsigned LoSize = VT.getSizeInBits();
3862 DAG.getNode(ISD::SRA, dl, VT, RHS,
3863 DAG.getConstant(LoSize - 1, dl,
3864 TLI.getPointerTy(DAG.getDataLayout())));
3866 DAG.getNode(ISD::SRA, dl, VT, LHS,
3867 DAG.getConstant(LoSize - 1, dl,
3868 TLI.getPointerTy(DAG.getDataLayout())));
3870 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3871 // pre-lowered to the correct types. This all depends upon WideVT not
3872 // being a legal type for the architecture and thus has to be split to
3874 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3875 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3876 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3877 DAG.getIntPtrConstant(0, dl));
3878 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3879 DAG.getIntPtrConstant(1, dl));
3880 // Ret is a node with an illegal type. Because such things are not
3881 // generally permitted during this phase of legalization, make sure the
3882 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been
3884 assert(Ret->use_empty() &&
3885 "Unexpected uses of illegally type from expanded lib call.");
3889 Tmp1 = DAG.getConstant(
3890 VT.getSizeInBits() - 1, dl,
3891 TLI.getShiftAmountTy(BottomHalf.getValueType(), DAG.getDataLayout()));
3892 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3893 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3896 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3897 DAG.getConstant(0, dl, VT), ISD::SETNE);
3899 Results.push_back(BottomHalf);
3900 Results.push_back(TopHalf);
3903 case ISD::BUILD_PAIR: {
3904 EVT PairTy = Node->getValueType(0);
3905 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3906 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3908 ISD::SHL, dl, PairTy, Tmp2,
3909 DAG.getConstant(PairTy.getSizeInBits() / 2, dl,
3910 TLI.getShiftAmountTy(PairTy, DAG.getDataLayout())));
3911 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3915 Tmp1 = Node->getOperand(0);
3916 Tmp2 = Node->getOperand(1);
3917 Tmp3 = Node->getOperand(2);
3918 if (Tmp1.getOpcode() == ISD::SETCC) {
3919 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3921 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3923 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3924 DAG.getConstant(0, dl, Tmp1.getValueType()),
3925 Tmp2, Tmp3, ISD::SETNE);
3927 Results.push_back(Tmp1);
3930 SDValue Chain = Node->getOperand(0);
3931 SDValue Table = Node->getOperand(1);
3932 SDValue Index = Node->getOperand(2);
3934 EVT PTy = TLI.getPointerTy(DAG.getDataLayout());
3936 const DataLayout &TD = DAG.getDataLayout();
3937 unsigned EntrySize =
3938 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3940 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(), Index,
3941 DAG.getConstant(EntrySize, dl, Index.getValueType()));
3942 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3945 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3946 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3947 MachinePointerInfo::getJumpTable(), MemVT,
3948 false, false, false, 0);
3950 if (TM.getRelocationModel() == Reloc::PIC_) {
3951 // For PIC, the sequence is:
3952 // BRIND(load(Jumptable + index) + RelocBase)
3953 // RelocBase can be JumpTable, GOT or some sort of global base.
3954 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3955 TLI.getPICJumpTableRelocBase(Table, DAG));
3957 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3958 Results.push_back(Tmp1);
3962 // Expand brcond's setcc into its constituent parts and create a BR_CC
3964 Tmp1 = Node->getOperand(0);
3965 Tmp2 = Node->getOperand(1);
3966 if (Tmp2.getOpcode() == ISD::SETCC) {
3967 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3968 Tmp1, Tmp2.getOperand(2),
3969 Tmp2.getOperand(0), Tmp2.getOperand(1),
3970 Node->getOperand(2));
3972 // We test only the i1 bit. Skip the AND if UNDEF.
3973 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3974 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3975 DAG.getConstant(1, dl, Tmp2.getValueType()));
3976 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3977 DAG.getCondCode(ISD::SETNE), Tmp3,
3978 DAG.getConstant(0, dl, Tmp3.getValueType()),
3979 Node->getOperand(2));
3981 Results.push_back(Tmp1);
3984 Tmp1 = Node->getOperand(0);
3985 Tmp2 = Node->getOperand(1);
3986 Tmp3 = Node->getOperand(2);
3987 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3988 Tmp3, NeedInvert, dl);
3991 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3992 // condition code, create a new SETCC node.
3994 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3997 // If we expanded the SETCC by inverting the condition code, then wrap
3998 // the existing SETCC in a NOT to restore the intended condition.
4000 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
4002 Results.push_back(Tmp1);
4006 // Otherwise, SETCC for the given comparison type must be completely
4007 // illegal; expand it into a SELECT_CC.
4008 EVT VT = Node->getValueType(0);
4010 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
4011 case TargetLowering::ZeroOrOneBooleanContent:
4012 case TargetLowering::UndefinedBooleanContent:
4015 case TargetLowering::ZeroOrNegativeOneBooleanContent:
4019 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
4020 DAG.getConstant(TrueValue, dl, VT),
4021 DAG.getConstant(0, dl, VT),
4023 Results.push_back(Tmp1);
4026 case ISD::SELECT_CC: {
4027 Tmp1 = Node->getOperand(0); // LHS
4028 Tmp2 = Node->getOperand(1); // RHS
4029 Tmp3 = Node->getOperand(2); // True
4030 Tmp4 = Node->getOperand(3); // False
4031 EVT VT = Node->getValueType(0);
4032 SDValue CC = Node->getOperand(4);
4033 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
4035 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
4036 // If the condition code is legal, then we need to expand this
4037 // node using SETCC and SELECT.
4038 EVT CmpVT = Tmp1.getValueType();
4039 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
4040 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
4043 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), CmpVT);
4044 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
4045 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
4049 // SELECT_CC is legal, so the condition code must not be.
4050 bool Legalized = false;
4051 // Try to legalize by inverting the condition. This is for targets that
4052 // might support an ordered version of a condition, but not the unordered
4053 // version (or vice versa).
4054 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
4055 Tmp1.getValueType().isInteger());
4056 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
4057 // Use the new condition code and swap true and false
4059 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
4061 // If The inverse is not legal, then try to swap the arguments using
4062 // the inverse condition code.
4063 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
4064 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
4065 // The swapped inverse condition is legal, so swap true and false,
4068 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
4073 Legalized = LegalizeSetCCCondCode(
4074 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
4077 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
4079 // If we expanded the SETCC by inverting the condition code, then swap
4080 // the True/False operands to match.
4082 std::swap(Tmp3, Tmp4);
4084 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
4085 // condition code, create a new SELECT_CC node.
4087 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
4088 Tmp1, Tmp2, Tmp3, Tmp4, CC);
4090 Tmp2 = DAG.getConstant(0, dl, Tmp1.getValueType());
4091 CC = DAG.getCondCode(ISD::SETNE);
4092 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
4093 Tmp2, Tmp3, Tmp4, CC);
4096 Results.push_back(Tmp1);
4100 Tmp1 = Node->getOperand(0); // Chain
4101 Tmp2 = Node->getOperand(2); // LHS
4102 Tmp3 = Node->getOperand(3); // RHS
4103 Tmp4 = Node->getOperand(1); // CC
4105 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
4106 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
4108 assert(Legalized && "Can't legalize BR_CC with legal condition!");
4110 // If we expanded the SETCC by inverting the condition code, then wrap
4111 // the existing SETCC in a NOT to restore the intended condition.
4113 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
4115 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
4117 if (Tmp4.getNode()) {
4118 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
4119 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
4121 Tmp3 = DAG.getConstant(0, dl, Tmp2.getValueType());
4122 Tmp4 = DAG.getCondCode(ISD::SETNE);
4123 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
4124 Tmp2, Tmp3, Node->getOperand(4));
4126 Results.push_back(Tmp1);
4129 case ISD::BUILD_VECTOR:
4130 Results.push_back(ExpandBUILD_VECTOR(Node));
4135 // Scalarize vector SRA/SRL/SHL.
4136 EVT VT = Node->getValueType(0);
4137 assert(VT.isVector() && "Unable to legalize non-vector shift");
4138 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
4139 unsigned NumElem = VT.getVectorNumElements();
4141 SmallVector<SDValue, 8> Scalars;
4142 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
4143 SDValue Ex = DAG.getNode(
4144 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(0),
4145 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
4146 SDValue Sh = DAG.getNode(
4147 ISD::EXTRACT_VECTOR_ELT, dl, VT.getScalarType(), Node->getOperand(1),
4148 DAG.getConstant(Idx, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
4149 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
4150 VT.getScalarType(), Ex, Sh));
4153 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars);
4154 ReplaceNode(SDValue(Node, 0), Result);
4157 case ISD::GLOBAL_OFFSET_TABLE:
4158 case ISD::GlobalAddress:
4159 case ISD::GlobalTLSAddress:
4160 case ISD::ExternalSymbol:
4161 case ISD::ConstantPool:
4162 case ISD::JumpTable:
4163 case ISD::INTRINSIC_W_CHAIN:
4164 case ISD::INTRINSIC_WO_CHAIN:
4165 case ISD::INTRINSIC_VOID:
4166 // FIXME: Custom lowering for these operations shouldn't return null!
4170 // Replace the original node with the legalized result.
4171 if (!Results.empty())
4172 ReplaceNode(Node, Results.data());
4175 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4176 SmallVector<SDValue, 8> Results;
4177 MVT OVT = Node->getSimpleValueType(0);
4178 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4179 Node->getOpcode() == ISD::SINT_TO_FP ||
4180 Node->getOpcode() == ISD::SETCC) {
4181 OVT = Node->getOperand(0).getSimpleValueType();
4183 if (Node->getOpcode() == ISD::BR_CC)
4184 OVT = Node->getOperand(2).getSimpleValueType();
4185 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4187 SDValue Tmp1, Tmp2, Tmp3;
4188 switch (Node->getOpcode()) {
4190 case ISD::CTTZ_ZERO_UNDEF:
4192 case ISD::CTLZ_ZERO_UNDEF:
4194 // Zero extend the argument.
4195 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4196 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4197 // already the correct result.
4198 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4199 if (Node->getOpcode() == ISD::CTTZ) {
4200 // FIXME: This should set a bit in the zero extended value instead.
4201 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
4202 Tmp1, DAG.getConstant(NVT.getSizeInBits(), dl, NVT),
4204 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
4205 DAG.getConstant(OVT.getSizeInBits(), dl, NVT), Tmp1);
4206 } else if (Node->getOpcode() == ISD::CTLZ ||
4207 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4208 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4209 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4210 DAG.getConstant(NVT.getSizeInBits() -
4211 OVT.getSizeInBits(), dl, NVT));
4213 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4216 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4217 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4218 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
4220 ISD::SRL, dl, NVT, Tmp1,
4221 DAG.getConstant(DiffBits, dl,
4222 TLI.getShiftAmountTy(NVT, DAG.getDataLayout())));
4223 Results.push_back(Tmp1);
4226 case ISD::FP_TO_UINT:
4227 case ISD::FP_TO_SINT:
4228 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4229 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4230 Results.push_back(Tmp1);
4232 case ISD::UINT_TO_FP:
4233 case ISD::SINT_TO_FP:
4234 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4235 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4236 Results.push_back(Tmp1);
4239 SDValue Chain = Node->getOperand(0); // Get the chain.
4240 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4243 if (OVT.isVector()) {
4244 TruncOp = ISD::BITCAST;
4246 assert(OVT.isInteger()
4247 && "VAARG promotion is supported only for vectors or integer types");
4248 TruncOp = ISD::TRUNCATE;
4251 // Perform the larger operation, then convert back
4252 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4253 Node->getConstantOperandVal(3));
4254 Chain = Tmp1.getValue(1);
4256 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4258 // Modified the chain result - switch anything that used the old chain to
4260 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4261 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4263 UpdatedNodes->insert(Tmp2.getNode());
4264 UpdatedNodes->insert(Chain.getNode());
4272 unsigned ExtOp, TruncOp;
4273 if (OVT.isVector()) {
4274 ExtOp = ISD::BITCAST;
4275 TruncOp = ISD::BITCAST;
4277 assert(OVT.isInteger() && "Cannot promote logic operation");
4278 ExtOp = ISD::ANY_EXTEND;
4279 TruncOp = ISD::TRUNCATE;
4281 // Promote each of the values to the new type.
4282 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4283 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4284 // Perform the larger operation, then convert back
4285 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4286 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4290 unsigned ExtOp, TruncOp;
4291 if (Node->getValueType(0).isVector() ||
4292 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4293 ExtOp = ISD::BITCAST;
4294 TruncOp = ISD::BITCAST;
4295 } else if (Node->getValueType(0).isInteger()) {
4296 ExtOp = ISD::ANY_EXTEND;
4297 TruncOp = ISD::TRUNCATE;
4299 ExtOp = ISD::FP_EXTEND;
4300 TruncOp = ISD::FP_ROUND;
4302 Tmp1 = Node->getOperand(0);
4303 // Promote each of the values to the new type.
4304 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4305 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4306 // Perform the larger operation, then round down.
4307 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4308 if (TruncOp != ISD::FP_ROUND)
4309 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4311 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4312 DAG.getIntPtrConstant(0, dl));
4313 Results.push_back(Tmp1);
4316 case ISD::VECTOR_SHUFFLE: {
4317 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4319 // Cast the two input vectors.
4320 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4321 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4323 // Convert the shuffle mask to the right # elements.
4324 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4325 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4326 Results.push_back(Tmp1);
4330 unsigned ExtOp = ISD::FP_EXTEND;
4331 if (NVT.isInteger()) {
4332 ISD::CondCode CCCode =
4333 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4334 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4336 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4337 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4338 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4339 Tmp1, Tmp2, Node->getOperand(2)));
4343 unsigned ExtOp = ISD::FP_EXTEND;
4344 if (NVT.isInteger()) {
4345 ISD::CondCode CCCode =
4346 cast<CondCodeSDNode>(Node->getOperand(1))->get();
4347 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4349 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4350 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(3));
4351 Results.push_back(DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0),
4352 Node->getOperand(0), Node->getOperand(1),
4353 Tmp1, Tmp2, Node->getOperand(4)));
4363 case ISD::FCOPYSIGN:
4365 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4366 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4367 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4368 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4369 Tmp3, DAG.getIntPtrConstant(0, dl)));
4373 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4374 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4375 Tmp3 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(2));
4377 DAG.getNode(ISD::FP_ROUND, dl, OVT,
4378 DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2, Tmp3),
4379 DAG.getIntPtrConstant(0, dl)));
4383 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4384 Tmp2 = Node->getOperand(1);
4385 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4386 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4387 Tmp3, DAG.getIntPtrConstant(0, dl)));
4393 case ISD::FNEARBYINT:
4406 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4407 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4408 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4409 Tmp2, DAG.getIntPtrConstant(0, dl)));
4414 // Replace the original node with the legalized result.
4415 if (!Results.empty())
4416 ReplaceNode(Node, Results.data());
4419 /// This is the entry point for the file.
4420 void SelectionDAG::Legalize() {
4421 AssignTopologicalOrder();
4423 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4424 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4426 // Visit all the nodes. We start in topological order, so that we see
4427 // nodes with their original operands intact. Legalization can produce
4428 // new nodes which may themselves need to be legalized. Iterate until all
4429 // nodes have been legalized.
4431 bool AnyLegalized = false;
4432 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4436 if (N->use_empty() && N != getRoot().getNode()) {
4442 if (LegalizedNodes.insert(N).second) {
4443 AnyLegalized = true;
4444 Legalizer.LegalizeOp(N);
4446 if (N->use_empty() && N != getRoot().getNode()) {
4457 // Remove dead nodes now.
4461 bool SelectionDAG::LegalizeOp(SDNode *N,
4462 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4463 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4464 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4466 // Directly insert the node in question, and legalize it. This will recurse
4467 // as needed through operands.
4468 LegalizedNodes.insert(N);
4469 Legalizer.LegalizeOp(N);
4471 return LegalizedNodes.count(N);