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"
42 //===----------------------------------------------------------------------===//
43 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
44 /// hacks on it until the target machine can handle it. This involves
45 /// eliminating value sizes the machine cannot handle (promoting small sizes to
46 /// large sizes or splitting up large values into small values) as well as
47 /// eliminating operations the machine cannot handle.
49 /// This code also does a small amount of optimization and recognition of idioms
50 /// as part of its processing. For example, if a target does not support a
51 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
52 /// will attempt merge setcc and brc instructions into brcc's.
55 class SelectionDAGLegalize {
56 const TargetMachine &TM;
57 const TargetLowering &TLI;
60 /// \brief The set of nodes which have already been legalized. We hold a
61 /// reference to it in order to update as necessary on node deletion.
62 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
64 /// \brief A set of all the nodes updated during legalization.
65 SmallSetVector<SDNode *, 16> *UpdatedNodes;
67 EVT getSetCCResultType(EVT VT) const {
68 return TLI.getSetCCResultType(*DAG.getContext(), VT);
71 // Libcall insertion helpers.
74 SelectionDAGLegalize(SelectionDAG &DAG,
75 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
76 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
77 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
78 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {}
80 /// \brief Legalizes the given operation.
81 void LegalizeOp(SDNode *Node);
84 SDValue OptimizeFloatStore(StoreSDNode *ST);
86 void LegalizeLoadOps(SDNode *Node);
87 void LegalizeStoreOps(SDNode *Node);
89 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
90 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
91 /// is necessary to spill the vector being inserted into to memory, perform
92 /// the insert there, and then read the result back.
93 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
94 SDValue Idx, SDLoc dl);
95 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
96 SDValue Idx, SDLoc dl);
98 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
99 /// performs the same shuffe in terms of order or result bytes, but on a type
100 /// whose vector element type is narrower than the original shuffle type.
101 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
102 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
103 SDValue N1, SDValue N2,
104 ArrayRef<int> Mask) const;
106 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
107 bool &NeedInvert, SDLoc dl);
109 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
110 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
111 unsigned NumOps, bool isSigned, SDLoc dl);
113 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
114 SDNode *Node, bool isSigned);
115 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
116 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
117 RTLIB::Libcall Call_F128,
118 RTLIB::Libcall Call_PPCF128);
119 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
120 RTLIB::Libcall Call_I8,
121 RTLIB::Libcall Call_I16,
122 RTLIB::Libcall Call_I32,
123 RTLIB::Libcall Call_I64,
124 RTLIB::Libcall Call_I128);
125 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
126 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
128 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
129 SDValue ExpandBUILD_VECTOR(SDNode *Node);
130 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
131 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
132 SmallVectorImpl<SDValue> &Results);
133 SDValue ExpandFCOPYSIGN(SDNode *Node);
134 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
136 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
138 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
141 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
142 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
144 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
145 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
146 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
148 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
150 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
152 void ExpandNode(SDNode *Node);
153 void PromoteNode(SDNode *Node);
156 // Node replacement helpers
157 void ReplacedNode(SDNode *N) {
158 LegalizedNodes.erase(N);
160 UpdatedNodes->insert(N);
162 void ReplaceNode(SDNode *Old, SDNode *New) {
163 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
164 dbgs() << " with: "; New->dump(&DAG));
166 assert(Old->getNumValues() == New->getNumValues() &&
167 "Replacing one node with another that produces a different number "
169 DAG.ReplaceAllUsesWith(Old, New);
170 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
171 DAG.TransferDbgValues(SDValue(Old, i), SDValue(New, i));
173 UpdatedNodes->insert(New);
176 void ReplaceNode(SDValue Old, SDValue New) {
177 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
178 dbgs() << " with: "; New->dump(&DAG));
180 DAG.ReplaceAllUsesWith(Old, New);
181 DAG.TransferDbgValues(Old, New);
183 UpdatedNodes->insert(New.getNode());
184 ReplacedNode(Old.getNode());
186 void ReplaceNode(SDNode *Old, const SDValue *New) {
187 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
189 DAG.ReplaceAllUsesWith(Old, New);
190 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
191 DEBUG(dbgs() << (i == 0 ? " with: "
194 DAG.TransferDbgValues(SDValue(Old, i), New[i]);
196 UpdatedNodes->insert(New[i].getNode());
203 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
204 /// performs the same shuffe in terms of order or result bytes, but on a type
205 /// whose vector element type is narrower than the original shuffle type.
206 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
208 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
209 SDValue N1, SDValue N2,
210 ArrayRef<int> Mask) const {
211 unsigned NumMaskElts = VT.getVectorNumElements();
212 unsigned NumDestElts = NVT.getVectorNumElements();
213 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
215 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
217 if (NumEltsGrowth == 1)
218 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
220 SmallVector<int, 8> NewMask;
221 for (unsigned i = 0; i != NumMaskElts; ++i) {
223 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
225 NewMask.push_back(-1);
227 NewMask.push_back(Idx * NumEltsGrowth + j);
230 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
231 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
232 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
235 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
236 /// a load from the constant pool.
238 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
242 // If a FP immediate is precise when represented as a float and if the
243 // target can do an extending load from float to double, we put it into
244 // the constant pool as a float, even if it's is statically typed as a
245 // double. This shrinks FP constants and canonicalizes them for targets where
246 // an FP extending load is the same cost as a normal load (such as on the x87
247 // fp stack or PPC FP unit).
248 EVT VT = CFP->getValueType(0);
249 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
251 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
252 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
253 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
258 while (SVT != MVT::f32 && SVT != MVT::f16) {
259 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
260 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
261 // Only do this if the target has a native EXTLOAD instruction from
263 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
264 TLI.ShouldShrinkFPConstant(OrigVT)) {
265 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
266 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
272 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
273 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
276 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
278 CPIdx, MachinePointerInfo::getConstantPool(),
279 VT, false, false, false, Alignment);
283 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
284 MachinePointerInfo::getConstantPool(), false, false, false,
289 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
290 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
291 const TargetLowering &TLI,
292 SelectionDAGLegalize *DAGLegalize) {
293 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
294 "unaligned indexed stores not implemented!");
295 SDValue Chain = ST->getChain();
296 SDValue Ptr = ST->getBasePtr();
297 SDValue Val = ST->getValue();
298 EVT VT = Val.getValueType();
299 int Alignment = ST->getAlignment();
300 unsigned AS = ST->getAddressSpace();
303 if (ST->getMemoryVT().isFloatingPoint() ||
304 ST->getMemoryVT().isVector()) {
305 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
306 if (TLI.isTypeLegal(intVT)) {
307 // Expand to a bitconvert of the value to the integer type of the
308 // same size, then a (misaligned) int store.
309 // FIXME: Does not handle truncating floating point stores!
310 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
311 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
312 ST->isVolatile(), ST->isNonTemporal(), Alignment);
313 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
316 // Do a (aligned) store to a stack slot, then copy from the stack slot
317 // to the final destination using (unaligned) integer loads and stores.
318 EVT StoredVT = ST->getMemoryVT();
320 TLI.getRegisterType(*DAG.getContext(),
321 EVT::getIntegerVT(*DAG.getContext(),
322 StoredVT.getSizeInBits()));
323 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
324 unsigned RegBytes = RegVT.getSizeInBits() / 8;
325 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
327 // Make sure the stack slot is also aligned for the register type.
328 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
330 // Perform the original store, only redirected to the stack slot.
331 SDValue Store = DAG.getTruncStore(Chain, dl,
332 Val, StackPtr, MachinePointerInfo(),
333 StoredVT, false, false, 0);
334 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy(AS));
335 SmallVector<SDValue, 8> Stores;
338 // Do all but one copies using the full register width.
339 for (unsigned i = 1; i < NumRegs; i++) {
340 // Load one integer register's worth from the stack slot.
341 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
342 MachinePointerInfo(),
343 false, false, false, 0);
344 // Store it to the final location. Remember the store.
345 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
346 ST->getPointerInfo().getWithOffset(Offset),
347 ST->isVolatile(), ST->isNonTemporal(),
348 MinAlign(ST->getAlignment(), Offset)));
349 // Increment the pointers.
351 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
353 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
356 // The last store may be partial. Do a truncating store. On big-endian
357 // machines this requires an extending load from the stack slot to ensure
358 // that the bits are in the right place.
359 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
360 8 * (StoredBytes - Offset));
362 // Load from the stack slot.
363 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
364 MachinePointerInfo(),
365 MemVT, false, false, false, 0);
367 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
369 .getWithOffset(Offset),
370 MemVT, ST->isVolatile(),
372 MinAlign(ST->getAlignment(), Offset),
374 // The order of the stores doesn't matter - say it with a TokenFactor.
375 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
376 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
379 assert(ST->getMemoryVT().isInteger() &&
380 !ST->getMemoryVT().isVector() &&
381 "Unaligned store of unknown type.");
382 // Get the half-size VT
383 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
384 int NumBits = NewStoredVT.getSizeInBits();
385 int IncrementSize = NumBits / 8;
387 // Divide the stored value in two parts.
388 SDValue ShiftAmount = DAG.getConstant(NumBits,
389 TLI.getShiftAmountTy(Val.getValueType()));
391 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
393 // Store the two parts
394 SDValue Store1, Store2;
395 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
396 ST->getPointerInfo(), NewStoredVT,
397 ST->isVolatile(), ST->isNonTemporal(), Alignment);
399 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
400 DAG.getConstant(IncrementSize, TLI.getPointerTy(AS)));
401 Alignment = MinAlign(Alignment, IncrementSize);
402 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
403 ST->getPointerInfo().getWithOffset(IncrementSize),
404 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
405 Alignment, ST->getAAInfo());
408 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
409 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
412 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
414 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
415 const TargetLowering &TLI,
416 SDValue &ValResult, SDValue &ChainResult) {
417 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
418 "unaligned indexed loads not implemented!");
419 SDValue Chain = LD->getChain();
420 SDValue Ptr = LD->getBasePtr();
421 EVT VT = LD->getValueType(0);
422 EVT LoadedVT = LD->getMemoryVT();
424 if (VT.isFloatingPoint() || VT.isVector()) {
425 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
426 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
427 // Expand to a (misaligned) integer load of the same size,
428 // then bitconvert to floating point or vector.
429 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
430 LD->getMemOperand());
431 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
433 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
434 ISD::ANY_EXTEND, dl, VT, Result);
441 // Copy the value to a (aligned) stack slot using (unaligned) integer
442 // loads and stores, then do a (aligned) load from the stack slot.
443 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
444 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
445 unsigned RegBytes = RegVT.getSizeInBits() / 8;
446 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
448 // Make sure the stack slot is also aligned for the register type.
449 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
451 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
452 SmallVector<SDValue, 8> Stores;
453 SDValue StackPtr = StackBase;
456 // Do all but one copies using the full register width.
457 for (unsigned i = 1; i < NumRegs; i++) {
458 // Load one integer register's worth from the original location.
459 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
460 LD->getPointerInfo().getWithOffset(Offset),
461 LD->isVolatile(), LD->isNonTemporal(),
463 MinAlign(LD->getAlignment(), Offset),
465 // Follow the load with a store to the stack slot. Remember the store.
466 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
467 MachinePointerInfo(), false, false, 0));
468 // Increment the pointers.
470 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
471 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
475 // The last copy may be partial. Do an extending load.
476 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
477 8 * (LoadedBytes - Offset));
478 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
479 LD->getPointerInfo().getWithOffset(Offset),
480 MemVT, LD->isVolatile(),
483 MinAlign(LD->getAlignment(), Offset),
485 // Follow the load with a store to the stack slot. Remember the store.
486 // On big-endian machines this requires a truncating store to ensure
487 // that the bits end up in the right place.
488 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
489 MachinePointerInfo(), MemVT,
492 // The order of the stores doesn't matter - say it with a TokenFactor.
493 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
495 // Finally, perform the original load only redirected to the stack slot.
496 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
497 MachinePointerInfo(), LoadedVT, false,false, false,
500 // Callers expect a MERGE_VALUES node.
505 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
506 "Unaligned load of unsupported type.");
508 // Compute the new VT that is half the size of the old one. This is an
510 unsigned NumBits = LoadedVT.getSizeInBits();
512 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
515 unsigned Alignment = LD->getAlignment();
516 unsigned IncrementSize = NumBits / 8;
517 ISD::LoadExtType HiExtType = LD->getExtensionType();
519 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
520 if (HiExtType == ISD::NON_EXTLOAD)
521 HiExtType = ISD::ZEXTLOAD;
523 // Load the value in two parts
525 if (TLI.isLittleEndian()) {
526 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
527 NewLoadedVT, LD->isVolatile(),
528 LD->isNonTemporal(), LD->isInvariant(), Alignment,
530 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
531 DAG.getConstant(IncrementSize, Ptr.getValueType()));
532 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
533 LD->getPointerInfo().getWithOffset(IncrementSize),
534 NewLoadedVT, LD->isVolatile(),
535 LD->isNonTemporal(),LD->isInvariant(),
536 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
538 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
539 NewLoadedVT, LD->isVolatile(),
540 LD->isNonTemporal(), LD->isInvariant(), Alignment,
542 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
543 DAG.getConstant(IncrementSize, Ptr.getValueType()));
544 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
545 LD->getPointerInfo().getWithOffset(IncrementSize),
546 NewLoadedVT, LD->isVolatile(),
547 LD->isNonTemporal(), LD->isInvariant(),
548 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
551 // aggregate the two parts
552 SDValue ShiftAmount = DAG.getConstant(NumBits,
553 TLI.getShiftAmountTy(Hi.getValueType()));
554 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
555 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
557 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
564 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
565 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
566 /// is necessary to spill the vector being inserted into to memory, perform
567 /// the insert there, and then read the result back.
568 SDValue SelectionDAGLegalize::
569 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
575 // If the target doesn't support this, we have to spill the input vector
576 // to a temporary stack slot, update the element, then reload it. This is
577 // badness. We could also load the value into a vector register (either
578 // with a "move to register" or "extload into register" instruction, then
579 // permute it into place, if the idx is a constant and if the idx is
580 // supported by the target.
581 EVT VT = Tmp1.getValueType();
582 EVT EltVT = VT.getVectorElementType();
583 EVT IdxVT = Tmp3.getValueType();
584 EVT PtrVT = TLI.getPointerTy();
585 SDValue StackPtr = DAG.CreateStackTemporary(VT);
587 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
590 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
591 MachinePointerInfo::getFixedStack(SPFI),
594 // Truncate or zero extend offset to target pointer type.
595 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
596 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
597 // Add the offset to the index.
598 unsigned EltSize = EltVT.getSizeInBits()/8;
599 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
600 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
601 // Store the scalar value.
602 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
604 // Load the updated vector.
605 return DAG.getLoad(VT, dl, Ch, StackPtr,
606 MachinePointerInfo::getFixedStack(SPFI), false, false,
611 SDValue SelectionDAGLegalize::
612 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
613 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
614 // SCALAR_TO_VECTOR requires that the type of the value being inserted
615 // match the element type of the vector being created, except for
616 // integers in which case the inserted value can be over width.
617 EVT EltVT = Vec.getValueType().getVectorElementType();
618 if (Val.getValueType() == EltVT ||
619 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
620 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
621 Vec.getValueType(), Val);
623 unsigned NumElts = Vec.getValueType().getVectorNumElements();
624 // We generate a shuffle of InVec and ScVec, so the shuffle mask
625 // should be 0,1,2,3,4,5... with the appropriate element replaced with
627 SmallVector<int, 8> ShufOps;
628 for (unsigned i = 0; i != NumElts; ++i)
629 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
631 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
635 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
638 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
639 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
640 // FIXME: We shouldn't do this for TargetConstantFP's.
641 // FIXME: move this to the DAG Combiner! Note that we can't regress due
642 // to phase ordering between legalized code and the dag combiner. This
643 // probably means that we need to integrate dag combiner and legalizer
645 // We generally can't do this one for long doubles.
646 SDValue Chain = ST->getChain();
647 SDValue Ptr = ST->getBasePtr();
648 unsigned Alignment = ST->getAlignment();
649 bool isVolatile = ST->isVolatile();
650 bool isNonTemporal = ST->isNonTemporal();
651 AAMDNodes AAInfo = ST->getAAInfo();
653 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
654 if (CFP->getValueType(0) == MVT::f32 &&
655 TLI.isTypeLegal(MVT::i32)) {
656 SDValue Con = DAG.getConstant(CFP->getValueAPF().
657 bitcastToAPInt().zextOrTrunc(32),
659 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
660 isVolatile, isNonTemporal, Alignment, AAInfo);
663 if (CFP->getValueType(0) == MVT::f64) {
664 // If this target supports 64-bit registers, do a single 64-bit store.
665 if (TLI.isTypeLegal(MVT::i64)) {
666 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
667 zextOrTrunc(64), MVT::i64);
668 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
669 isVolatile, isNonTemporal, Alignment, AAInfo);
672 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
673 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
674 // stores. If the target supports neither 32- nor 64-bits, this
675 // xform is certainly not worth it.
676 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
677 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
678 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
679 if (TLI.isBigEndian()) std::swap(Lo, Hi);
681 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
682 isNonTemporal, Alignment, AAInfo);
683 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
684 DAG.getConstant(4, Ptr.getValueType()));
685 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
686 ST->getPointerInfo().getWithOffset(4),
687 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
690 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
694 return SDValue(nullptr, 0);
697 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
698 StoreSDNode *ST = cast<StoreSDNode>(Node);
699 SDValue Chain = ST->getChain();
700 SDValue Ptr = ST->getBasePtr();
703 unsigned Alignment = ST->getAlignment();
704 bool isVolatile = ST->isVolatile();
705 bool isNonTemporal = ST->isNonTemporal();
706 AAMDNodes AAInfo = ST->getAAInfo();
708 if (!ST->isTruncatingStore()) {
709 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
710 ReplaceNode(ST, OptStore);
715 SDValue Value = ST->getValue();
716 MVT VT = Value.getSimpleValueType();
717 switch (TLI.getOperationAction(ISD::STORE, VT)) {
718 default: llvm_unreachable("This action is not supported yet!");
719 case TargetLowering::Legal: {
720 // If this is an unaligned store and the target doesn't support it,
722 unsigned AS = ST->getAddressSpace();
723 unsigned Align = ST->getAlignment();
724 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
725 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
726 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
727 if (Align < ABIAlignment)
728 ExpandUnalignedStore(cast<StoreSDNode>(Node),
733 case TargetLowering::Custom: {
734 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
736 ReplaceNode(SDValue(Node, 0), Res);
739 case TargetLowering::Promote: {
740 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
741 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
742 "Can only promote stores to same size type");
743 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
745 DAG.getStore(Chain, dl, Value, Ptr,
746 ST->getPointerInfo(), isVolatile,
747 isNonTemporal, Alignment, AAInfo);
748 ReplaceNode(SDValue(Node, 0), Result);
755 SDValue Value = ST->getValue();
757 EVT StVT = ST->getMemoryVT();
758 unsigned StWidth = StVT.getSizeInBits();
760 if (StWidth != StVT.getStoreSizeInBits()) {
761 // Promote to a byte-sized store with upper bits zero if not
762 // storing an integral number of bytes. For example, promote
763 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
764 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
765 StVT.getStoreSizeInBits());
766 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
768 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
769 NVT, isVolatile, isNonTemporal, Alignment,
771 ReplaceNode(SDValue(Node, 0), Result);
772 } else if (StWidth & (StWidth - 1)) {
773 // If not storing a power-of-2 number of bits, expand as two stores.
774 assert(!StVT.isVector() && "Unsupported truncstore!");
775 unsigned RoundWidth = 1 << Log2_32(StWidth);
776 assert(RoundWidth < StWidth);
777 unsigned ExtraWidth = StWidth - RoundWidth;
778 assert(ExtraWidth < RoundWidth);
779 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
780 "Store size not an integral number of bytes!");
781 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
782 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
784 unsigned IncrementSize;
786 if (TLI.isLittleEndian()) {
787 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
788 // Store the bottom RoundWidth bits.
789 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
791 isVolatile, isNonTemporal, Alignment,
794 // Store the remaining ExtraWidth bits.
795 IncrementSize = RoundWidth / 8;
796 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
797 DAG.getConstant(IncrementSize, Ptr.getValueType()));
798 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
799 DAG.getConstant(RoundWidth,
800 TLI.getShiftAmountTy(Value.getValueType())));
801 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
802 ST->getPointerInfo().getWithOffset(IncrementSize),
803 ExtraVT, isVolatile, isNonTemporal,
804 MinAlign(Alignment, IncrementSize), AAInfo);
806 // Big endian - avoid unaligned stores.
807 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
808 // Store the top RoundWidth bits.
809 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
810 DAG.getConstant(ExtraWidth,
811 TLI.getShiftAmountTy(Value.getValueType())));
812 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
813 RoundVT, isVolatile, isNonTemporal, Alignment,
816 // Store the remaining ExtraWidth bits.
817 IncrementSize = RoundWidth / 8;
818 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
819 DAG.getConstant(IncrementSize, Ptr.getValueType()));
820 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
821 ST->getPointerInfo().getWithOffset(IncrementSize),
822 ExtraVT, isVolatile, isNonTemporal,
823 MinAlign(Alignment, IncrementSize), AAInfo);
826 // The order of the stores doesn't matter.
827 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
828 ReplaceNode(SDValue(Node, 0), Result);
830 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
831 StVT.getSimpleVT())) {
832 default: llvm_unreachable("This action is not supported yet!");
833 case TargetLowering::Legal: {
834 unsigned AS = ST->getAddressSpace();
835 unsigned Align = ST->getAlignment();
836 // If this is an unaligned store and the target doesn't support it,
838 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
839 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
840 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
841 if (Align < ABIAlignment)
842 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
846 case TargetLowering::Custom: {
847 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
849 ReplaceNode(SDValue(Node, 0), Res);
852 case TargetLowering::Expand:
853 assert(!StVT.isVector() &&
854 "Vector Stores are handled in LegalizeVectorOps");
856 // TRUNCSTORE:i16 i32 -> STORE i16
857 assert(TLI.isTypeLegal(StVT) &&
858 "Do not know how to expand this store!");
859 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
861 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
862 isVolatile, isNonTemporal, Alignment, AAInfo);
863 ReplaceNode(SDValue(Node, 0), Result);
870 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
871 LoadSDNode *LD = cast<LoadSDNode>(Node);
872 SDValue Chain = LD->getChain(); // The chain.
873 SDValue Ptr = LD->getBasePtr(); // The base pointer.
874 SDValue Value; // The value returned by the load op.
877 ISD::LoadExtType ExtType = LD->getExtensionType();
878 if (ExtType == ISD::NON_EXTLOAD) {
879 MVT VT = Node->getSimpleValueType(0);
880 SDValue RVal = SDValue(Node, 0);
881 SDValue RChain = SDValue(Node, 1);
883 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
884 default: llvm_unreachable("This action is not supported yet!");
885 case TargetLowering::Legal: {
886 unsigned AS = LD->getAddressSpace();
887 unsigned Align = LD->getAlignment();
888 // If this is an unaligned load and the target doesn't support it,
890 if (!TLI.allowsMisalignedMemoryAccesses(LD->getMemoryVT(), AS, Align)) {
891 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
892 unsigned ABIAlignment =
893 TLI.getDataLayout()->getABITypeAlignment(Ty);
894 if (Align < ABIAlignment){
895 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
900 case TargetLowering::Custom: {
901 SDValue Res = TLI.LowerOperation(RVal, DAG);
904 RChain = Res.getValue(1);
908 case TargetLowering::Promote: {
909 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
910 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
911 "Can only promote loads to same size type");
913 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
914 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
915 RChain = Res.getValue(1);
919 if (RChain.getNode() != Node) {
920 assert(RVal.getNode() != Node && "Load must be completely replaced");
921 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
922 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
924 UpdatedNodes->insert(RVal.getNode());
925 UpdatedNodes->insert(RChain.getNode());
932 EVT SrcVT = LD->getMemoryVT();
933 unsigned SrcWidth = SrcVT.getSizeInBits();
934 unsigned Alignment = LD->getAlignment();
935 bool isVolatile = LD->isVolatile();
936 bool isNonTemporal = LD->isNonTemporal();
937 bool isInvariant = LD->isInvariant();
938 AAMDNodes AAInfo = LD->getAAInfo();
940 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
941 // Some targets pretend to have an i1 loading operation, and actually
942 // load an i8. This trick is correct for ZEXTLOAD because the top 7
943 // bits are guaranteed to be zero; it helps the optimizers understand
944 // that these bits are zero. It is also useful for EXTLOAD, since it
945 // tells the optimizers that those bits are undefined. It would be
946 // nice to have an effective generic way of getting these benefits...
947 // Until such a way is found, don't insist on promoting i1 here.
949 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
950 // Promote to a byte-sized load if not loading an integral number of
951 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
952 unsigned NewWidth = SrcVT.getStoreSizeInBits();
953 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
956 // The extra bits are guaranteed to be zero, since we stored them that
957 // way. A zext load from NVT thus automatically gives zext from SrcVT.
959 ISD::LoadExtType NewExtType =
960 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
963 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
964 Chain, Ptr, LD->getPointerInfo(),
965 NVT, isVolatile, isNonTemporal, isInvariant, Alignment,
968 Ch = Result.getValue(1); // The chain.
970 if (ExtType == ISD::SEXTLOAD)
971 // Having the top bits zero doesn't help when sign extending.
972 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
973 Result.getValueType(),
974 Result, DAG.getValueType(SrcVT));
975 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
976 // All the top bits are guaranteed to be zero - inform the optimizers.
977 Result = DAG.getNode(ISD::AssertZext, dl,
978 Result.getValueType(), Result,
979 DAG.getValueType(SrcVT));
983 } else if (SrcWidth & (SrcWidth - 1)) {
984 // If not loading a power-of-2 number of bits, expand as two loads.
985 assert(!SrcVT.isVector() && "Unsupported extload!");
986 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
987 assert(RoundWidth < SrcWidth);
988 unsigned ExtraWidth = SrcWidth - RoundWidth;
989 assert(ExtraWidth < RoundWidth);
990 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
991 "Load size not an integral number of bytes!");
992 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
993 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
995 unsigned IncrementSize;
997 if (TLI.isLittleEndian()) {
998 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
999 // Load the bottom RoundWidth bits.
1000 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1002 LD->getPointerInfo(), RoundVT, isVolatile,
1003 isNonTemporal, isInvariant, Alignment, AAInfo);
1005 // Load the remaining ExtraWidth bits.
1006 IncrementSize = RoundWidth / 8;
1007 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1008 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1009 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1010 LD->getPointerInfo().getWithOffset(IncrementSize),
1011 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1012 MinAlign(Alignment, IncrementSize), AAInfo);
1014 // Build a factor node to remember that this load is independent of
1016 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1019 // Move the top bits to the right place.
1020 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1021 DAG.getConstant(RoundWidth,
1022 TLI.getShiftAmountTy(Hi.getValueType())));
1024 // Join the hi and lo parts.
1025 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1027 // Big endian - avoid unaligned loads.
1028 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1029 // Load the top RoundWidth bits.
1030 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1031 LD->getPointerInfo(), RoundVT, isVolatile,
1032 isNonTemporal, isInvariant, Alignment, AAInfo);
1034 // Load the remaining ExtraWidth bits.
1035 IncrementSize = RoundWidth / 8;
1036 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1037 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1038 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1039 dl, Node->getValueType(0), Chain, Ptr,
1040 LD->getPointerInfo().getWithOffset(IncrementSize),
1041 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1042 MinAlign(Alignment, IncrementSize), AAInfo);
1044 // Build a factor node to remember that this load is independent of
1046 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1049 // Move the top bits to the right place.
1050 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1051 DAG.getConstant(ExtraWidth,
1052 TLI.getShiftAmountTy(Hi.getValueType())));
1054 // Join the hi and lo parts.
1055 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1060 bool isCustom = false;
1061 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1062 default: llvm_unreachable("This action is not supported yet!");
1063 case TargetLowering::Custom:
1066 case TargetLowering::Legal: {
1067 Value = SDValue(Node, 0);
1068 Chain = SDValue(Node, 1);
1071 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1072 if (Res.getNode()) {
1074 Chain = Res.getValue(1);
1077 // If this is an unaligned load and the target doesn't support
1079 EVT MemVT = LD->getMemoryVT();
1080 unsigned AS = LD->getAddressSpace();
1081 unsigned Align = LD->getAlignment();
1082 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1084 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1085 unsigned ABIAlignment =
1086 TLI.getDataLayout()->getABITypeAlignment(Ty);
1087 if (Align < ABIAlignment){
1088 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1089 DAG, TLI, Value, Chain);
1095 case TargetLowering::Expand:
1096 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) &&
1097 TLI.isTypeLegal(SrcVT)) {
1098 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1099 LD->getMemOperand());
1103 ExtendOp = (SrcVT.isFloatingPoint() ?
1104 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1106 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1107 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1108 default: llvm_unreachable("Unexpected extend load type!");
1110 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1111 Chain = Load.getValue(1);
1115 assert(!SrcVT.isVector() &&
1116 "Vector Loads are handled in LegalizeVectorOps");
1118 // FIXME: This does not work for vectors on most targets. Sign-
1119 // and zero-extend operations are currently folded into extending
1120 // loads, whether they are legal or not, and then we end up here
1121 // without any support for legalizing them.
1122 assert(ExtType != ISD::EXTLOAD &&
1123 "EXTLOAD should always be supported!");
1124 // Turn the unsupported load into an EXTLOAD followed by an
1125 // explicit zero/sign extend inreg.
1126 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1127 Node->getValueType(0),
1129 LD->getMemOperand());
1131 if (ExtType == ISD::SEXTLOAD)
1132 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1133 Result.getValueType(),
1134 Result, DAG.getValueType(SrcVT));
1136 ValRes = DAG.getZeroExtendInReg(Result, dl,
1137 SrcVT.getScalarType());
1139 Chain = Result.getValue(1);
1144 // Since loads produce two values, make sure to remember that we legalized
1146 if (Chain.getNode() != Node) {
1147 assert(Value.getNode() != Node && "Load must be completely replaced");
1148 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1149 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1151 UpdatedNodes->insert(Value.getNode());
1152 UpdatedNodes->insert(Chain.getNode());
1158 /// LegalizeOp - Return a legal replacement for the given operation, with
1159 /// all legal operands.
1160 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1161 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1163 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1166 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1167 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1168 TargetLowering::TypeLegal &&
1169 "Unexpected illegal type!");
1171 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1172 assert((TLI.getTypeAction(*DAG.getContext(),
1173 Node->getOperand(i).getValueType()) ==
1174 TargetLowering::TypeLegal ||
1175 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1176 "Unexpected illegal type!");
1178 // Figure out the correct action; the way to query this varies by opcode
1179 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1180 bool SimpleFinishLegalizing = true;
1181 switch (Node->getOpcode()) {
1182 case ISD::INTRINSIC_W_CHAIN:
1183 case ISD::INTRINSIC_WO_CHAIN:
1184 case ISD::INTRINSIC_VOID:
1185 case ISD::STACKSAVE:
1186 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1189 Action = TLI.getOperationAction(Node->getOpcode(),
1190 Node->getValueType(0));
1191 if (Action != TargetLowering::Promote)
1192 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1194 case ISD::FP_TO_FP16:
1195 case ISD::SINT_TO_FP:
1196 case ISD::UINT_TO_FP:
1197 case ISD::EXTRACT_VECTOR_ELT:
1198 Action = TLI.getOperationAction(Node->getOpcode(),
1199 Node->getOperand(0).getValueType());
1201 case ISD::FP_ROUND_INREG:
1202 case ISD::SIGN_EXTEND_INREG: {
1203 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1204 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1207 case ISD::ATOMIC_STORE: {
1208 Action = TLI.getOperationAction(Node->getOpcode(),
1209 Node->getOperand(2).getValueType());
1212 case ISD::SELECT_CC:
1215 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1216 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1217 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1218 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1219 ISD::CondCode CCCode =
1220 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1221 Action = TLI.getCondCodeAction(CCCode, OpVT);
1222 if (Action == TargetLowering::Legal) {
1223 if (Node->getOpcode() == ISD::SELECT_CC)
1224 Action = TLI.getOperationAction(Node->getOpcode(),
1225 Node->getValueType(0));
1227 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1233 // FIXME: Model these properly. LOAD and STORE are complicated, and
1234 // STORE expects the unlegalized operand in some cases.
1235 SimpleFinishLegalizing = false;
1237 case ISD::CALLSEQ_START:
1238 case ISD::CALLSEQ_END:
1239 // FIXME: This shouldn't be necessary. These nodes have special properties
1240 // dealing with the recursive nature of legalization. Removing this
1241 // special case should be done as part of making LegalizeDAG non-recursive.
1242 SimpleFinishLegalizing = false;
1244 case ISD::EXTRACT_ELEMENT:
1245 case ISD::FLT_ROUNDS_:
1253 case ISD::MERGE_VALUES:
1254 case ISD::EH_RETURN:
1255 case ISD::FRAME_TO_ARGS_OFFSET:
1256 case ISD::EH_SJLJ_SETJMP:
1257 case ISD::EH_SJLJ_LONGJMP:
1258 // These operations lie about being legal: when they claim to be legal,
1259 // they should actually be expanded.
1260 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1261 if (Action == TargetLowering::Legal)
1262 Action = TargetLowering::Expand;
1264 case ISD::INIT_TRAMPOLINE:
1265 case ISD::ADJUST_TRAMPOLINE:
1266 case ISD::FRAMEADDR:
1267 case ISD::RETURNADDR:
1268 // These operations lie about being legal: when they claim to be legal,
1269 // they should actually be custom-lowered.
1270 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1271 if (Action == TargetLowering::Legal)
1272 Action = TargetLowering::Custom;
1274 case ISD::READ_REGISTER:
1275 case ISD::WRITE_REGISTER:
1276 // Named register is legal in the DAG, but blocked by register name
1277 // selection if not implemented by target (to chose the correct register)
1278 // They'll be converted to Copy(To/From)Reg.
1279 Action = TargetLowering::Legal;
1281 case ISD::DEBUGTRAP:
1282 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1283 if (Action == TargetLowering::Expand) {
1284 // replace ISD::DEBUGTRAP with ISD::TRAP
1286 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1287 Node->getOperand(0));
1288 ReplaceNode(Node, NewVal.getNode());
1289 LegalizeOp(NewVal.getNode());
1295 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1296 Action = TargetLowering::Legal;
1298 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1303 if (SimpleFinishLegalizing) {
1304 SDNode *NewNode = Node;
1305 switch (Node->getOpcode()) {
1312 // Legalizing shifts/rotates requires adjusting the shift amount
1313 // to the appropriate width.
1314 if (!Node->getOperand(1).getValueType().isVector()) {
1316 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1317 Node->getOperand(1));
1318 HandleSDNode Handle(SAO);
1319 LegalizeOp(SAO.getNode());
1320 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1324 case ISD::SRL_PARTS:
1325 case ISD::SRA_PARTS:
1326 case ISD::SHL_PARTS:
1327 // Legalizing shifts/rotates requires adjusting the shift amount
1328 // to the appropriate width.
1329 if (!Node->getOperand(2).getValueType().isVector()) {
1331 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1332 Node->getOperand(2));
1333 HandleSDNode Handle(SAO);
1334 LegalizeOp(SAO.getNode());
1335 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1336 Node->getOperand(1),
1342 if (NewNode != Node) {
1343 ReplaceNode(Node, NewNode);
1347 case TargetLowering::Legal:
1349 case TargetLowering::Custom: {
1350 // FIXME: The handling for custom lowering with multiple results is
1352 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1353 if (Res.getNode()) {
1354 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1357 if (Node->getNumValues() == 1) {
1358 // We can just directly replace this node with the lowered value.
1359 ReplaceNode(SDValue(Node, 0), Res);
1363 SmallVector<SDValue, 8> ResultVals;
1364 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1365 ResultVals.push_back(Res.getValue(i));
1366 ReplaceNode(Node, ResultVals.data());
1371 case TargetLowering::Expand:
1374 case TargetLowering::Promote:
1380 switch (Node->getOpcode()) {
1387 llvm_unreachable("Do not know how to legalize this operator!");
1389 case ISD::CALLSEQ_START:
1390 case ISD::CALLSEQ_END:
1393 return LegalizeLoadOps(Node);
1396 return LegalizeStoreOps(Node);
1401 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1402 SDValue Vec = Op.getOperand(0);
1403 SDValue Idx = Op.getOperand(1);
1406 // Before we generate a new store to a temporary stack slot, see if there is
1407 // already one that we can use. There often is because when we scalarize
1408 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1409 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1410 // the vector. If all are expanded here, we don't want one store per vector
1412 SDValue StackPtr, Ch;
1413 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1414 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1416 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1417 if (ST->isIndexed() || ST->isTruncatingStore() ||
1418 ST->getValue() != Vec)
1421 // Make sure that nothing else could have stored into the destination of
1423 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1426 StackPtr = ST->getBasePtr();
1427 Ch = SDValue(ST, 0);
1432 if (!Ch.getNode()) {
1433 // Store the value to a temporary stack slot, then LOAD the returned part.
1434 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1435 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1436 MachinePointerInfo(), false, false, 0);
1439 // Add the offset to the index.
1441 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1442 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1443 DAG.getConstant(EltSize, Idx.getValueType()));
1445 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1446 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1448 if (Op.getValueType().isVector())
1449 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1450 false, false, false, 0);
1451 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1452 MachinePointerInfo(),
1453 Vec.getValueType().getVectorElementType(),
1454 false, false, false, 0);
1457 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1458 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1460 SDValue Vec = Op.getOperand(0);
1461 SDValue Part = Op.getOperand(1);
1462 SDValue Idx = Op.getOperand(2);
1465 // Store the value to a temporary stack slot, then LOAD the returned part.
1467 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1468 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1469 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1471 // First store the whole vector.
1472 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1475 // Then store the inserted part.
1477 // Add the offset to the index.
1479 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1481 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1482 DAG.getConstant(EltSize, Idx.getValueType()));
1483 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1485 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1488 // Store the subvector.
1489 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1490 MachinePointerInfo(), false, false, 0);
1492 // Finally, load the updated vector.
1493 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1494 false, false, false, 0);
1497 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1498 // We can't handle this case efficiently. Allocate a sufficiently
1499 // aligned object on the stack, store each element into it, then load
1500 // the result as a vector.
1501 // Create the stack frame object.
1502 EVT VT = Node->getValueType(0);
1503 EVT EltVT = VT.getVectorElementType();
1505 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1506 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1507 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1509 // Emit a store of each element to the stack slot.
1510 SmallVector<SDValue, 8> Stores;
1511 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1512 // Store (in the right endianness) the elements to memory.
1513 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1514 // Ignore undef elements.
1515 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1517 unsigned Offset = TypeByteSize*i;
1519 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1520 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1522 // If the destination vector element type is narrower than the source
1523 // element type, only store the bits necessary.
1524 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1525 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1526 Node->getOperand(i), Idx,
1527 PtrInfo.getWithOffset(Offset),
1528 EltVT, false, false, 0));
1530 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1531 Node->getOperand(i), Idx,
1532 PtrInfo.getWithOffset(Offset),
1537 if (!Stores.empty()) // Not all undef elements?
1538 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1540 StoreChain = DAG.getEntryNode();
1542 // Result is a load from the stack slot.
1543 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1544 false, false, false, 0);
1547 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1549 SDValue Tmp1 = Node->getOperand(0);
1550 SDValue Tmp2 = Node->getOperand(1);
1552 // Get the sign bit of the RHS. First obtain a value that has the same
1553 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1555 EVT FloatVT = Tmp2.getValueType();
1556 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1557 if (TLI.isTypeLegal(IVT)) {
1558 // Convert to an integer with the same sign bit.
1559 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1561 // Store the float to memory, then load the sign part out as an integer.
1562 MVT LoadTy = TLI.getPointerTy();
1563 // First create a temporary that is aligned for both the load and store.
1564 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1565 // Then store the float to it.
1567 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1569 if (TLI.isBigEndian()) {
1570 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1571 // Load out a legal integer with the same sign bit as the float.
1572 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1573 false, false, false, 0);
1574 } else { // Little endian
1575 SDValue LoadPtr = StackPtr;
1576 // The float may be wider than the integer we are going to load. Advance
1577 // the pointer so that the loaded integer will contain the sign bit.
1578 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1579 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1580 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1581 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1582 // Load a legal integer containing the sign bit.
1583 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1584 false, false, false, 0);
1585 // Move the sign bit to the top bit of the loaded integer.
1586 unsigned BitShift = LoadTy.getSizeInBits() -
1587 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1588 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1590 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1591 DAG.getConstant(BitShift,
1592 TLI.getShiftAmountTy(SignBit.getValueType())));
1595 // Now get the sign bit proper, by seeing whether the value is negative.
1596 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1597 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1599 // Get the absolute value of the result.
1600 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1601 // Select between the nabs and abs value based on the sign bit of
1603 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1604 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1608 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1609 SmallVectorImpl<SDValue> &Results) {
1610 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1611 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1612 " not tell us which reg is the stack pointer!");
1614 EVT VT = Node->getValueType(0);
1615 SDValue Tmp1 = SDValue(Node, 0);
1616 SDValue Tmp2 = SDValue(Node, 1);
1617 SDValue Tmp3 = Node->getOperand(2);
1618 SDValue Chain = Tmp1.getOperand(0);
1620 // Chain the dynamic stack allocation so that it doesn't modify the stack
1621 // pointer when other instructions are using the stack.
1622 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1625 SDValue Size = Tmp2.getOperand(1);
1626 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1627 Chain = SP.getValue(1);
1628 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1629 unsigned StackAlign =
1630 TM.getSubtargetImpl()->getFrameLowering()->getStackAlignment();
1631 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1632 if (Align > StackAlign)
1633 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1634 DAG.getConstant(-(uint64_t)Align, VT));
1635 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1637 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1638 DAG.getIntPtrConstant(0, true), SDValue(),
1641 Results.push_back(Tmp1);
1642 Results.push_back(Tmp2);
1645 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1646 /// condition code CC on the current target.
1648 /// If the SETCC has been legalized using AND / OR, then the legalized node
1649 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1650 /// will be set to false.
1652 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1653 /// then the values of LHS and RHS will be swapped, CC will be set to the
1654 /// new condition, and NeedInvert will be set to false.
1656 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1657 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1658 /// will be set to true. The caller must invert the result of the SETCC with
1659 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1660 /// of a true/false result.
1662 /// \returns true if the SetCC has been legalized, false if it hasn't.
1663 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1664 SDValue &LHS, SDValue &RHS,
1668 MVT OpVT = LHS.getSimpleValueType();
1669 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1671 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1672 default: llvm_unreachable("Unknown condition code action!");
1673 case TargetLowering::Legal:
1676 case TargetLowering::Expand: {
1677 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1678 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1679 std::swap(LHS, RHS);
1680 CC = DAG.getCondCode(InvCC);
1683 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1686 default: llvm_unreachable("Don't know how to expand this condition!");
1688 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1689 == TargetLowering::Legal
1690 && "If SETO is expanded, SETOEQ must be legal!");
1691 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1693 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1694 == TargetLowering::Legal
1695 && "If SETUO is expanded, SETUNE must be legal!");
1696 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1709 // If we are floating point, assign and break, otherwise fall through.
1710 if (!OpVT.isInteger()) {
1711 // We can use the 4th bit to tell if we are the unordered
1712 // or ordered version of the opcode.
1713 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1714 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1715 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1718 // Fallthrough if we are unsigned integer.
1723 // We only support using the inverted operation, which is computed above
1724 // and not a different manner of supporting expanding these cases.
1725 llvm_unreachable("Don't know how to expand this condition!");
1728 // Try inverting the result of the inverse condition.
1729 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1730 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1731 CC = DAG.getCondCode(InvCC);
1735 // If inverting the condition didn't work then we have no means to expand
1737 llvm_unreachable("Don't know how to expand this condition!");
1740 SDValue SetCC1, SetCC2;
1741 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1742 // If we aren't the ordered or unorder operation,
1743 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1744 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1745 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1747 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1748 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1749 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1751 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1760 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1761 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1762 /// a load from the stack slot to DestVT, extending it if needed.
1763 /// The resultant code need not be legal.
1764 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1768 // Create the stack frame object.
1770 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1771 getTypeForEVT(*DAG.getContext()));
1772 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1774 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1775 int SPFI = StackPtrFI->getIndex();
1776 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1778 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1779 unsigned SlotSize = SlotVT.getSizeInBits();
1780 unsigned DestSize = DestVT.getSizeInBits();
1781 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1782 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1784 // Emit a store to the stack slot. Use a truncstore if the input value is
1785 // later than DestVT.
1788 if (SrcSize > SlotSize)
1789 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1790 PtrInfo, SlotVT, false, false, SrcAlign);
1792 assert(SrcSize == SlotSize && "Invalid store");
1793 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1794 PtrInfo, false, false, SrcAlign);
1797 // Result is a load from the stack slot.
1798 if (SlotSize == DestSize)
1799 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1800 false, false, false, DestAlign);
1802 assert(SlotSize < DestSize && "Unknown extension!");
1803 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1804 PtrInfo, SlotVT, false, false, false, DestAlign);
1807 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1809 // Create a vector sized/aligned stack slot, store the value to element #0,
1810 // then load the whole vector back out.
1811 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1813 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1814 int SPFI = StackPtrFI->getIndex();
1816 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1818 MachinePointerInfo::getFixedStack(SPFI),
1819 Node->getValueType(0).getVectorElementType(),
1821 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1822 MachinePointerInfo::getFixedStack(SPFI),
1823 false, false, false, 0);
1827 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1828 const TargetLowering &TLI, SDValue &Res) {
1829 unsigned NumElems = Node->getNumOperands();
1831 EVT VT = Node->getValueType(0);
1833 // Try to group the scalars into pairs, shuffle the pairs together, then
1834 // shuffle the pairs of pairs together, etc. until the vector has
1835 // been built. This will work only if all of the necessary shuffle masks
1838 // We do this in two phases; first to check the legality of the shuffles,
1839 // and next, assuming that all shuffles are legal, to create the new nodes.
1840 for (int Phase = 0; Phase < 2; ++Phase) {
1841 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1843 for (unsigned i = 0; i < NumElems; ++i) {
1844 SDValue V = Node->getOperand(i);
1845 if (V.getOpcode() == ISD::UNDEF)
1850 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1851 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1854 while (IntermedVals.size() > 2) {
1855 NewIntermedVals.clear();
1856 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1857 // This vector and the next vector are shuffled together (simply to
1858 // append the one to the other).
1859 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1861 SmallVector<int, 16> FinalIndices;
1862 FinalIndices.reserve(IntermedVals[i].second.size() +
1863 IntermedVals[i+1].second.size());
1866 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1869 FinalIndices.push_back(IntermedVals[i].second[j]);
1871 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1873 ShuffleVec[k] = NumElems + j;
1874 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1879 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1880 IntermedVals[i+1].first,
1882 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1884 NewIntermedVals.push_back(
1885 std::make_pair(Shuffle, std::move(FinalIndices)));
1888 // If we had an odd number of defined values, then append the last
1889 // element to the array of new vectors.
1890 if ((IntermedVals.size() & 1) != 0)
1891 NewIntermedVals.push_back(IntermedVals.back());
1893 IntermedVals.swap(NewIntermedVals);
1896 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1897 "Invalid number of intermediate vectors");
1898 SDValue Vec1 = IntermedVals[0].first;
1900 if (IntermedVals.size() > 1)
1901 Vec2 = IntermedVals[1].first;
1903 Vec2 = DAG.getUNDEF(VT);
1905 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1906 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1907 ShuffleVec[IntermedVals[0].second[i]] = i;
1908 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1909 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1912 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1913 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1920 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1921 /// support the operation, but do support the resultant vector type.
1922 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1923 unsigned NumElems = Node->getNumOperands();
1924 SDValue Value1, Value2;
1926 EVT VT = Node->getValueType(0);
1927 EVT OpVT = Node->getOperand(0).getValueType();
1928 EVT EltVT = VT.getVectorElementType();
1930 // If the only non-undef value is the low element, turn this into a
1931 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1932 bool isOnlyLowElement = true;
1933 bool MoreThanTwoValues = false;
1934 bool isConstant = true;
1935 for (unsigned i = 0; i < NumElems; ++i) {
1936 SDValue V = Node->getOperand(i);
1937 if (V.getOpcode() == ISD::UNDEF)
1940 isOnlyLowElement = false;
1941 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1944 if (!Value1.getNode()) {
1946 } else if (!Value2.getNode()) {
1949 } else if (V != Value1 && V != Value2) {
1950 MoreThanTwoValues = true;
1954 if (!Value1.getNode())
1955 return DAG.getUNDEF(VT);
1957 if (isOnlyLowElement)
1958 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1960 // If all elements are constants, create a load from the constant pool.
1962 SmallVector<Constant*, 16> CV;
1963 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1964 if (ConstantFPSDNode *V =
1965 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1966 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1967 } else if (ConstantSDNode *V =
1968 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1970 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1972 // If OpVT and EltVT don't match, EltVT is not legal and the
1973 // element values have been promoted/truncated earlier. Undo this;
1974 // we don't want a v16i8 to become a v16i32 for example.
1975 const ConstantInt *CI = V->getConstantIntValue();
1976 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1977 CI->getZExtValue()));
1980 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1981 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1982 CV.push_back(UndefValue::get(OpNTy));
1985 Constant *CP = ConstantVector::get(CV);
1986 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1987 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1988 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1989 MachinePointerInfo::getConstantPool(),
1990 false, false, false, Alignment);
1993 SmallSet<SDValue, 16> DefinedValues;
1994 for (unsigned i = 0; i < NumElems; ++i) {
1995 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
1997 DefinedValues.insert(Node->getOperand(i));
2000 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
2001 if (!MoreThanTwoValues) {
2002 SmallVector<int, 8> ShuffleVec(NumElems, -1);
2003 for (unsigned i = 0; i < NumElems; ++i) {
2004 SDValue V = Node->getOperand(i);
2005 if (V.getOpcode() == ISD::UNDEF)
2007 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2009 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2010 // Get the splatted value into the low element of a vector register.
2011 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2013 if (Value2.getNode())
2014 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2016 Vec2 = DAG.getUNDEF(VT);
2018 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2019 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2023 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2028 // Otherwise, we can't handle this case efficiently.
2029 return ExpandVectorBuildThroughStack(Node);
2032 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
2033 // does not fit into a register, return the lo part and set the hi part to the
2034 // by-reg argument. If it does fit into a single register, return the result
2035 // and leave the Hi part unset.
2036 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2038 TargetLowering::ArgListTy Args;
2039 TargetLowering::ArgListEntry Entry;
2040 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2041 EVT ArgVT = Node->getOperand(i).getValueType();
2042 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2043 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2044 Entry.isSExt = isSigned;
2045 Entry.isZExt = !isSigned;
2046 Args.push_back(Entry);
2048 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2049 TLI.getPointerTy());
2051 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2053 // By default, the input chain to this libcall is the entry node of the
2054 // function. If the libcall is going to be emitted as a tail call then
2055 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2056 // node which is being folded has a non-entry input chain.
2057 SDValue InChain = DAG.getEntryNode();
2059 // isTailCall may be true since the callee does not reference caller stack
2060 // frame. Check if it's in the right position.
2061 SDValue TCChain = InChain;
2062 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2066 TargetLowering::CallLoweringInfo CLI(DAG);
2067 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2068 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2069 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2071 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2073 if (!CallInfo.second.getNode())
2074 // It's a tailcall, return the chain (which is the DAG root).
2075 return DAG.getRoot();
2077 return CallInfo.first;
2080 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
2081 /// and returning a result of type RetVT.
2082 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2083 const SDValue *Ops, unsigned NumOps,
2084 bool isSigned, SDLoc dl) {
2085 TargetLowering::ArgListTy Args;
2086 Args.reserve(NumOps);
2088 TargetLowering::ArgListEntry Entry;
2089 for (unsigned i = 0; i != NumOps; ++i) {
2090 Entry.Node = Ops[i];
2091 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2092 Entry.isSExt = isSigned;
2093 Entry.isZExt = !isSigned;
2094 Args.push_back(Entry);
2096 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2097 TLI.getPointerTy());
2099 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2101 TargetLowering::CallLoweringInfo CLI(DAG);
2102 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2103 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2104 .setSExtResult(isSigned).setZExtResult(!isSigned);
2106 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2108 return CallInfo.first;
2111 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
2112 // ExpandLibCall except that the first operand is the in-chain.
2113 std::pair<SDValue, SDValue>
2114 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2117 SDValue InChain = Node->getOperand(0);
2119 TargetLowering::ArgListTy Args;
2120 TargetLowering::ArgListEntry Entry;
2121 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2122 EVT ArgVT = Node->getOperand(i).getValueType();
2123 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2124 Entry.Node = Node->getOperand(i);
2126 Entry.isSExt = isSigned;
2127 Entry.isZExt = !isSigned;
2128 Args.push_back(Entry);
2130 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2131 TLI.getPointerTy());
2133 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2135 TargetLowering::CallLoweringInfo CLI(DAG);
2136 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2137 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2138 .setSExtResult(isSigned).setZExtResult(!isSigned);
2140 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2145 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2146 RTLIB::Libcall Call_F32,
2147 RTLIB::Libcall Call_F64,
2148 RTLIB::Libcall Call_F80,
2149 RTLIB::Libcall Call_F128,
2150 RTLIB::Libcall Call_PPCF128) {
2152 switch (Node->getSimpleValueType(0).SimpleTy) {
2153 default: llvm_unreachable("Unexpected request for libcall!");
2154 case MVT::f32: LC = Call_F32; break;
2155 case MVT::f64: LC = Call_F64; break;
2156 case MVT::f80: LC = Call_F80; break;
2157 case MVT::f128: LC = Call_F128; break;
2158 case MVT::ppcf128: LC = Call_PPCF128; break;
2160 return ExpandLibCall(LC, Node, false);
2163 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2164 RTLIB::Libcall Call_I8,
2165 RTLIB::Libcall Call_I16,
2166 RTLIB::Libcall Call_I32,
2167 RTLIB::Libcall Call_I64,
2168 RTLIB::Libcall Call_I128) {
2170 switch (Node->getSimpleValueType(0).SimpleTy) {
2171 default: llvm_unreachable("Unexpected request for libcall!");
2172 case MVT::i8: LC = Call_I8; break;
2173 case MVT::i16: LC = Call_I16; break;
2174 case MVT::i32: LC = Call_I32; break;
2175 case MVT::i64: LC = Call_I64; break;
2176 case MVT::i128: LC = Call_I128; break;
2178 return ExpandLibCall(LC, Node, isSigned);
2181 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2182 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2183 const TargetLowering &TLI) {
2185 switch (Node->getSimpleValueType(0).SimpleTy) {
2186 default: llvm_unreachable("Unexpected request for libcall!");
2187 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2188 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2189 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2190 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2191 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2194 return TLI.getLibcallName(LC) != nullptr;
2197 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2199 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2200 // The other use might have been replaced with a divrem already.
2201 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2202 unsigned OtherOpcode = 0;
2204 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2206 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2208 SDValue Op0 = Node->getOperand(0);
2209 SDValue Op1 = Node->getOperand(1);
2210 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2211 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2215 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2216 User->getOperand(0) == Op0 &&
2217 User->getOperand(1) == Op1)
2223 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2226 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2227 SmallVectorImpl<SDValue> &Results) {
2228 unsigned Opcode = Node->getOpcode();
2229 bool isSigned = Opcode == ISD::SDIVREM;
2232 switch (Node->getSimpleValueType(0).SimpleTy) {
2233 default: llvm_unreachable("Unexpected request for libcall!");
2234 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2235 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2236 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2237 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2238 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2241 // The input chain to this libcall is the entry node of the function.
2242 // Legalizing the call will automatically add the previous call to the
2244 SDValue InChain = DAG.getEntryNode();
2246 EVT RetVT = Node->getValueType(0);
2247 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2249 TargetLowering::ArgListTy Args;
2250 TargetLowering::ArgListEntry Entry;
2251 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2252 EVT ArgVT = Node->getOperand(i).getValueType();
2253 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2254 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2255 Entry.isSExt = isSigned;
2256 Entry.isZExt = !isSigned;
2257 Args.push_back(Entry);
2260 // Also pass the return address of the remainder.
2261 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2263 Entry.Ty = RetTy->getPointerTo();
2264 Entry.isSExt = isSigned;
2265 Entry.isZExt = !isSigned;
2266 Args.push_back(Entry);
2268 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2269 TLI.getPointerTy());
2272 TargetLowering::CallLoweringInfo CLI(DAG);
2273 CLI.setDebugLoc(dl).setChain(InChain)
2274 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2275 .setSExtResult(isSigned).setZExtResult(!isSigned);
2277 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2279 // Remainder is loaded back from the stack frame.
2280 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2281 MachinePointerInfo(), false, false, false, 0);
2282 Results.push_back(CallInfo.first);
2283 Results.push_back(Rem);
2286 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2287 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2289 switch (Node->getSimpleValueType(0).SimpleTy) {
2290 default: llvm_unreachable("Unexpected request for libcall!");
2291 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2292 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2293 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2294 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2295 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2297 return TLI.getLibcallName(LC) != nullptr;
2300 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2301 /// can be used to combine sin and cos.
2302 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2303 const TargetMachine &TM) {
2304 if (!isSinCosLibcallAvailable(Node, TLI))
2306 // GNU sin/cos functions set errno while sincos does not. Therefore
2307 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2308 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2309 if (isGNU && !TM.Options.UnsafeFPMath)
2314 /// useSinCos - Only issue sincos libcall if both sin and cos are
2316 static bool useSinCos(SDNode *Node) {
2317 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2318 ? ISD::FCOS : ISD::FSIN;
2320 SDValue Op0 = Node->getOperand(0);
2321 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2322 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2326 // The other user might have been turned into sincos already.
2327 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2333 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2336 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2337 SmallVectorImpl<SDValue> &Results) {
2339 switch (Node->getSimpleValueType(0).SimpleTy) {
2340 default: llvm_unreachable("Unexpected request for libcall!");
2341 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2342 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2343 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2344 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2345 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2348 // The input chain to this libcall is the entry node of the function.
2349 // Legalizing the call will automatically add the previous call to the
2351 SDValue InChain = DAG.getEntryNode();
2353 EVT RetVT = Node->getValueType(0);
2354 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2356 TargetLowering::ArgListTy Args;
2357 TargetLowering::ArgListEntry Entry;
2359 // Pass the argument.
2360 Entry.Node = Node->getOperand(0);
2362 Entry.isSExt = false;
2363 Entry.isZExt = false;
2364 Args.push_back(Entry);
2366 // Pass the return address of sin.
2367 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2368 Entry.Node = SinPtr;
2369 Entry.Ty = RetTy->getPointerTo();
2370 Entry.isSExt = false;
2371 Entry.isZExt = false;
2372 Args.push_back(Entry);
2374 // Also pass the return address of the cos.
2375 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2376 Entry.Node = CosPtr;
2377 Entry.Ty = RetTy->getPointerTo();
2378 Entry.isSExt = false;
2379 Entry.isZExt = false;
2380 Args.push_back(Entry);
2382 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2383 TLI.getPointerTy());
2386 TargetLowering::CallLoweringInfo CLI(DAG);
2387 CLI.setDebugLoc(dl).setChain(InChain)
2388 .setCallee(TLI.getLibcallCallingConv(LC),
2389 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2391 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2393 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2394 MachinePointerInfo(), false, false, false, 0));
2395 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2396 MachinePointerInfo(), false, false, false, 0));
2399 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2400 /// INT_TO_FP operation of the specified operand when the target requests that
2401 /// we expand it. At this point, we know that the result and operand types are
2402 /// legal for the target.
2403 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2407 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2408 // simple 32-bit [signed|unsigned] integer to float/double expansion
2410 // Get the stack frame index of a 8 byte buffer.
2411 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2413 // word offset constant for Hi/Lo address computation
2414 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2415 // set up Hi and Lo (into buffer) address based on endian
2416 SDValue Hi = StackSlot;
2417 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2418 StackSlot, WordOff);
2419 if (TLI.isLittleEndian())
2422 // if signed map to unsigned space
2425 // constant used to invert sign bit (signed to unsigned mapping)
2426 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2427 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2431 // store the lo of the constructed double - based on integer input
2432 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2433 Op0Mapped, Lo, MachinePointerInfo(),
2435 // initial hi portion of constructed double
2436 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2437 // store the hi of the constructed double - biased exponent
2438 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2439 MachinePointerInfo(),
2441 // load the constructed double
2442 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2443 MachinePointerInfo(), false, false, false, 0);
2444 // FP constant to bias correct the final result
2445 SDValue Bias = DAG.getConstantFP(isSigned ?
2446 BitsToDouble(0x4330000080000000ULL) :
2447 BitsToDouble(0x4330000000000000ULL),
2449 // subtract the bias
2450 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2453 // handle final rounding
2454 if (DestVT == MVT::f64) {
2457 } else if (DestVT.bitsLT(MVT::f64)) {
2458 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2459 DAG.getIntPtrConstant(0));
2460 } else if (DestVT.bitsGT(MVT::f64)) {
2461 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2465 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2466 // Code below here assumes !isSigned without checking again.
2468 // Implementation of unsigned i64 to f64 following the algorithm in
2469 // __floatundidf in compiler_rt. This implementation has the advantage
2470 // of performing rounding correctly, both in the default rounding mode
2471 // and in all alternate rounding modes.
2472 // TODO: Generalize this for use with other types.
2473 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2475 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2476 SDValue TwoP84PlusTwoP52 =
2477 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2479 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2481 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2482 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2483 DAG.getConstant(32, MVT::i64));
2484 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2485 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2486 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2487 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2488 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2490 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2493 // Implementation of unsigned i64 to f32.
2494 // TODO: Generalize this for use with other types.
2495 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2496 // For unsigned conversions, convert them to signed conversions using the
2497 // algorithm from the x86_64 __floatundidf in compiler_rt.
2499 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2501 SDValue ShiftConst =
2502 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2503 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2504 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2505 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2506 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2508 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2509 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2511 // TODO: This really should be implemented using a branch rather than a
2512 // select. We happen to get lucky and machinesink does the right
2513 // thing most of the time. This would be a good candidate for a
2514 //pseudo-op, or, even better, for whole-function isel.
2515 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2516 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2517 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2520 // Otherwise, implement the fully general conversion.
2522 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2523 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2524 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2525 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2526 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2527 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2528 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2529 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2530 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2531 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2532 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2534 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2535 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2537 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2538 DAG.getConstant(32, SHVT));
2539 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2540 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2542 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2543 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2544 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2545 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2546 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2547 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2548 DAG.getIntPtrConstant(0));
2551 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2553 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2554 Op0, DAG.getConstant(0, Op0.getValueType()),
2556 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2557 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2558 SignSet, Four, Zero);
2560 // If the sign bit of the integer is set, the large number will be treated
2561 // as a negative number. To counteract this, the dynamic code adds an
2562 // offset depending on the data type.
2564 switch (Op0.getSimpleValueType().SimpleTy) {
2565 default: llvm_unreachable("Unsupported integer type!");
2566 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2567 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2568 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2569 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2571 if (TLI.isLittleEndian()) FF <<= 32;
2572 Constant *FudgeFactor = ConstantInt::get(
2573 Type::getInt64Ty(*DAG.getContext()), FF);
2575 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2576 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2577 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2578 Alignment = std::min(Alignment, 4u);
2580 if (DestVT == MVT::f32)
2581 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2582 MachinePointerInfo::getConstantPool(),
2583 false, false, false, Alignment);
2585 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2586 DAG.getEntryNode(), CPIdx,
2587 MachinePointerInfo::getConstantPool(),
2588 MVT::f32, false, false, false, Alignment);
2589 HandleSDNode Handle(Load);
2590 LegalizeOp(Load.getNode());
2591 FudgeInReg = Handle.getValue();
2594 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2597 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2598 /// *INT_TO_FP operation of the specified operand when the target requests that
2599 /// we promote it. At this point, we know that the result and operand types are
2600 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2601 /// operation that takes a larger input.
2602 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2606 // First step, figure out the appropriate *INT_TO_FP operation to use.
2607 EVT NewInTy = LegalOp.getValueType();
2609 unsigned OpToUse = 0;
2611 // Scan for the appropriate larger type to use.
2613 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2614 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2616 // If the target supports SINT_TO_FP of this type, use it.
2617 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2618 OpToUse = ISD::SINT_TO_FP;
2621 if (isSigned) continue;
2623 // If the target supports UINT_TO_FP of this type, use it.
2624 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2625 OpToUse = ISD::UINT_TO_FP;
2629 // Otherwise, try a larger type.
2632 // Okay, we found the operation and type to use. Zero extend our input to the
2633 // desired type then run the operation on it.
2634 return DAG.getNode(OpToUse, dl, DestVT,
2635 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2636 dl, NewInTy, LegalOp));
2639 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2640 /// FP_TO_*INT operation of the specified operand when the target requests that
2641 /// we promote it. At this point, we know that the result and operand types are
2642 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2643 /// operation that returns a larger result.
2644 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2648 // First step, figure out the appropriate FP_TO*INT operation to use.
2649 EVT NewOutTy = DestVT;
2651 unsigned OpToUse = 0;
2653 // Scan for the appropriate larger type to use.
2655 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2656 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2658 // A larger signed type can hold all unsigned values of the requested type,
2659 // so using FP_TO_SINT is valid
2660 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2661 OpToUse = ISD::FP_TO_SINT;
2665 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2666 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2667 OpToUse = ISD::FP_TO_UINT;
2671 // Otherwise, try a larger type.
2675 // Okay, we found the operation and type to use.
2676 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2678 // Truncate the result of the extended FP_TO_*INT operation to the desired
2680 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2683 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2685 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2686 EVT VT = Op.getValueType();
2687 EVT SHVT = TLI.getShiftAmountTy(VT);
2688 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2689 switch (VT.getSimpleVT().SimpleTy) {
2690 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2692 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2693 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2694 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2696 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2697 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2698 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2699 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2700 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2701 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2702 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2703 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2704 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2706 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2707 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2708 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2709 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2710 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2711 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2712 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2713 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2714 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2715 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2716 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2717 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2718 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2719 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2720 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2721 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2722 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2723 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2724 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2725 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2726 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2730 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2732 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2735 default: llvm_unreachable("Cannot expand this yet!");
2737 EVT VT = Op.getValueType();
2738 EVT ShVT = TLI.getShiftAmountTy(VT);
2739 unsigned Len = VT.getSizeInBits();
2741 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2742 "CTPOP not implemented for this type.");
2744 // This is the "best" algorithm from
2745 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2747 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2748 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2749 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2750 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2752 // v = v - ((v >> 1) & 0x55555555...)
2753 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2754 DAG.getNode(ISD::AND, dl, VT,
2755 DAG.getNode(ISD::SRL, dl, VT, Op,
2756 DAG.getConstant(1, ShVT)),
2758 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2759 Op = DAG.getNode(ISD::ADD, dl, VT,
2760 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2761 DAG.getNode(ISD::AND, dl, VT,
2762 DAG.getNode(ISD::SRL, dl, VT, Op,
2763 DAG.getConstant(2, ShVT)),
2765 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2766 Op = DAG.getNode(ISD::AND, dl, VT,
2767 DAG.getNode(ISD::ADD, dl, VT, Op,
2768 DAG.getNode(ISD::SRL, dl, VT, Op,
2769 DAG.getConstant(4, ShVT))),
2771 // v = (v * 0x01010101...) >> (Len - 8)
2772 Op = DAG.getNode(ISD::SRL, dl, VT,
2773 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2774 DAG.getConstant(Len - 8, ShVT));
2778 case ISD::CTLZ_ZERO_UNDEF:
2779 // This trivially expands to CTLZ.
2780 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2782 // for now, we do this:
2783 // x = x | (x >> 1);
2784 // x = x | (x >> 2);
2786 // x = x | (x >>16);
2787 // x = x | (x >>32); // for 64-bit input
2788 // return popcount(~x);
2790 // Ref: "Hacker's Delight" by Henry Warren
2791 EVT VT = Op.getValueType();
2792 EVT ShVT = TLI.getShiftAmountTy(VT);
2793 unsigned len = VT.getSizeInBits();
2794 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2795 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2796 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2797 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2799 Op = DAG.getNOT(dl, Op, VT);
2800 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2802 case ISD::CTTZ_ZERO_UNDEF:
2803 // This trivially expands to CTTZ.
2804 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2806 // for now, we use: { return popcount(~x & (x - 1)); }
2807 // unless the target has ctlz but not ctpop, in which case we use:
2808 // { return 32 - nlz(~x & (x-1)); }
2809 // Ref: "Hacker's Delight" by Henry Warren
2810 EVT VT = Op.getValueType();
2811 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2812 DAG.getNOT(dl, Op, VT),
2813 DAG.getNode(ISD::SUB, dl, VT, Op,
2814 DAG.getConstant(1, VT)));
2815 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2816 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2817 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2818 return DAG.getNode(ISD::SUB, dl, VT,
2819 DAG.getConstant(VT.getSizeInBits(), VT),
2820 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2821 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2826 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2827 unsigned Opc = Node->getOpcode();
2828 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2833 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2834 case ISD::ATOMIC_SWAP:
2835 switch (VT.SimpleTy) {
2836 default: llvm_unreachable("Unexpected value type for atomic!");
2837 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2838 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2839 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2840 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2841 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2844 case ISD::ATOMIC_CMP_SWAP:
2845 switch (VT.SimpleTy) {
2846 default: llvm_unreachable("Unexpected value type for atomic!");
2847 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2848 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2849 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2850 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2851 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2854 case ISD::ATOMIC_LOAD_ADD:
2855 switch (VT.SimpleTy) {
2856 default: llvm_unreachable("Unexpected value type for atomic!");
2857 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2858 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2859 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2860 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2861 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2864 case ISD::ATOMIC_LOAD_SUB:
2865 switch (VT.SimpleTy) {
2866 default: llvm_unreachable("Unexpected value type for atomic!");
2867 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2868 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2869 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2870 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2871 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2874 case ISD::ATOMIC_LOAD_AND:
2875 switch (VT.SimpleTy) {
2876 default: llvm_unreachable("Unexpected value type for atomic!");
2877 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2878 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2879 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2880 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2881 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2884 case ISD::ATOMIC_LOAD_OR:
2885 switch (VT.SimpleTy) {
2886 default: llvm_unreachable("Unexpected value type for atomic!");
2887 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2888 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2889 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2890 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2891 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2894 case ISD::ATOMIC_LOAD_XOR:
2895 switch (VT.SimpleTy) {
2896 default: llvm_unreachable("Unexpected value type for atomic!");
2897 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2898 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2899 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2900 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2901 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2904 case ISD::ATOMIC_LOAD_NAND:
2905 switch (VT.SimpleTy) {
2906 default: llvm_unreachable("Unexpected value type for atomic!");
2907 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2908 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2909 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2910 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2911 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2914 case ISD::ATOMIC_LOAD_MAX:
2915 switch (VT.SimpleTy) {
2916 default: llvm_unreachable("Unexpected value type for atomic!");
2917 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2918 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2919 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2920 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2921 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2924 case ISD::ATOMIC_LOAD_UMAX:
2925 switch (VT.SimpleTy) {
2926 default: llvm_unreachable("Unexpected value type for atomic!");
2927 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2928 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2929 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2930 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2931 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2934 case ISD::ATOMIC_LOAD_MIN:
2935 switch (VT.SimpleTy) {
2936 default: llvm_unreachable("Unexpected value type for atomic!");
2937 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2938 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2939 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2940 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2941 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2944 case ISD::ATOMIC_LOAD_UMIN:
2945 switch (VT.SimpleTy) {
2946 default: llvm_unreachable("Unexpected value type for atomic!");
2947 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2948 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2949 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2950 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2951 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2956 return ExpandChainLibCall(LC, Node, false);
2959 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2960 SmallVector<SDValue, 8> Results;
2962 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2964 switch (Node->getOpcode()) {
2967 case ISD::CTLZ_ZERO_UNDEF:
2969 case ISD::CTTZ_ZERO_UNDEF:
2970 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2971 Results.push_back(Tmp1);
2974 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2976 case ISD::FRAMEADDR:
2977 case ISD::RETURNADDR:
2978 case ISD::FRAME_TO_ARGS_OFFSET:
2979 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2981 case ISD::FLT_ROUNDS_:
2982 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2984 case ISD::EH_RETURN:
2988 case ISD::EH_SJLJ_LONGJMP:
2989 // If the target didn't expand these, there's nothing to do, so just
2990 // preserve the chain and be done.
2991 Results.push_back(Node->getOperand(0));
2993 case ISD::EH_SJLJ_SETJMP:
2994 // If the target didn't expand this, just return 'zero' and preserve the
2996 Results.push_back(DAG.getConstant(0, MVT::i32));
2997 Results.push_back(Node->getOperand(0));
2999 case ISD::ATOMIC_FENCE: {
3000 // If the target didn't lower this, lower it to '__sync_synchronize()' call
3001 // FIXME: handle "fence singlethread" more efficiently.
3002 TargetLowering::ArgListTy Args;
3004 TargetLowering::CallLoweringInfo CLI(DAG);
3005 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3006 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3007 DAG.getExternalSymbol("__sync_synchronize",
3008 TLI.getPointerTy()), std::move(Args), 0);
3010 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3012 Results.push_back(CallResult.second);
3015 case ISD::ATOMIC_LOAD: {
3016 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3017 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
3018 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3019 SDValue Swap = DAG.getAtomicCmpSwap(
3020 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3021 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3022 cast<AtomicSDNode>(Node)->getMemOperand(),
3023 cast<AtomicSDNode>(Node)->getOrdering(),
3024 cast<AtomicSDNode>(Node)->getOrdering(),
3025 cast<AtomicSDNode>(Node)->getSynchScope());
3026 Results.push_back(Swap.getValue(0));
3027 Results.push_back(Swap.getValue(1));
3030 case ISD::ATOMIC_STORE: {
3031 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3032 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3033 cast<AtomicSDNode>(Node)->getMemoryVT(),
3034 Node->getOperand(0),
3035 Node->getOperand(1), Node->getOperand(2),
3036 cast<AtomicSDNode>(Node)->getMemOperand(),
3037 cast<AtomicSDNode>(Node)->getOrdering(),
3038 cast<AtomicSDNode>(Node)->getSynchScope());
3039 Results.push_back(Swap.getValue(1));
3042 // By default, atomic intrinsics are marked Legal and lowered. Targets
3043 // which don't support them directly, however, may want libcalls, in which
3044 // case they mark them Expand, and we get here.
3045 case ISD::ATOMIC_SWAP:
3046 case ISD::ATOMIC_LOAD_ADD:
3047 case ISD::ATOMIC_LOAD_SUB:
3048 case ISD::ATOMIC_LOAD_AND:
3049 case ISD::ATOMIC_LOAD_OR:
3050 case ISD::ATOMIC_LOAD_XOR:
3051 case ISD::ATOMIC_LOAD_NAND:
3052 case ISD::ATOMIC_LOAD_MIN:
3053 case ISD::ATOMIC_LOAD_MAX:
3054 case ISD::ATOMIC_LOAD_UMIN:
3055 case ISD::ATOMIC_LOAD_UMAX:
3056 case ISD::ATOMIC_CMP_SWAP: {
3057 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3058 Results.push_back(Tmp.first);
3059 Results.push_back(Tmp.second);
3062 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3063 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3064 // splits out the success value as a comparison. Expanding the resulting
3065 // ATOMIC_CMP_SWAP will produce a libcall.
3066 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3067 SDValue Res = DAG.getAtomicCmpSwap(
3068 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3069 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3070 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3071 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3072 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3073 cast<AtomicSDNode>(Node)->getSynchScope());
3075 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3076 Res, Node->getOperand(2), ISD::SETEQ);
3078 Results.push_back(Res.getValue(0));
3079 Results.push_back(Success);
3080 Results.push_back(Res.getValue(1));
3083 case ISD::DYNAMIC_STACKALLOC:
3084 ExpandDYNAMIC_STACKALLOC(Node, Results);
3086 case ISD::MERGE_VALUES:
3087 for (unsigned i = 0; i < Node->getNumValues(); i++)
3088 Results.push_back(Node->getOperand(i));
3091 EVT VT = Node->getValueType(0);
3093 Results.push_back(DAG.getConstant(0, VT));
3095 assert(VT.isFloatingPoint() && "Unknown value type!");
3096 Results.push_back(DAG.getConstantFP(0, VT));
3101 // If this operation is not supported, lower it to 'abort()' call
3102 TargetLowering::ArgListTy Args;
3103 TargetLowering::CallLoweringInfo CLI(DAG);
3104 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3105 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3106 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
3107 std::move(Args), 0);
3108 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3110 Results.push_back(CallResult.second);
3115 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3116 Node->getValueType(0), dl);
3117 Results.push_back(Tmp1);
3119 case ISD::FP_EXTEND:
3120 Tmp1 = EmitStackConvert(Node->getOperand(0),
3121 Node->getOperand(0).getValueType(),
3122 Node->getValueType(0), dl);
3123 Results.push_back(Tmp1);
3125 case ISD::SIGN_EXTEND_INREG: {
3126 // NOTE: we could fall back on load/store here too for targets without
3127 // SAR. However, it is doubtful that any exist.
3128 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3129 EVT VT = Node->getValueType(0);
3130 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3133 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3134 ExtraVT.getScalarType().getSizeInBits();
3135 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3136 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3137 Node->getOperand(0), ShiftCst);
3138 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3139 Results.push_back(Tmp1);
3142 case ISD::FP_ROUND_INREG: {
3143 // The only way we can lower this is to turn it into a TRUNCSTORE,
3144 // EXTLOAD pair, targeting a temporary location (a stack slot).
3146 // NOTE: there is a choice here between constantly creating new stack
3147 // slots and always reusing the same one. We currently always create
3148 // new ones, as reuse may inhibit scheduling.
3149 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3150 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3151 Node->getValueType(0), dl);
3152 Results.push_back(Tmp1);
3155 case ISD::SINT_TO_FP:
3156 case ISD::UINT_TO_FP:
3157 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3158 Node->getOperand(0), Node->getValueType(0), dl);
3159 Results.push_back(Tmp1);
3161 case ISD::FP_TO_SINT:
3162 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3163 Results.push_back(Tmp1);
3165 case ISD::FP_TO_UINT: {
3166 SDValue True, False;
3167 EVT VT = Node->getOperand(0).getValueType();
3168 EVT NVT = Node->getValueType(0);
3169 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3170 APInt::getNullValue(VT.getSizeInBits()));
3171 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3172 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3173 Tmp1 = DAG.getConstantFP(apf, VT);
3174 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3175 Node->getOperand(0),
3177 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3178 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3179 DAG.getNode(ISD::FSUB, dl, VT,
3180 Node->getOperand(0), Tmp1));
3181 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3182 DAG.getConstant(x, NVT));
3183 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3184 Results.push_back(Tmp1);
3188 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3189 EVT VT = Node->getValueType(0);
3190 Tmp1 = Node->getOperand(0);
3191 Tmp2 = Node->getOperand(1);
3192 unsigned Align = Node->getConstantOperandVal(3);
3194 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3195 MachinePointerInfo(V),
3196 false, false, false, 0);
3197 SDValue VAList = VAListLoad;
3199 if (Align > TLI.getMinStackArgumentAlignment()) {
3200 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3202 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3203 DAG.getConstant(Align - 1,
3204 VAList.getValueType()));
3206 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3207 DAG.getConstant(-(int64_t)Align,
3208 VAList.getValueType()));
3211 // Increment the pointer, VAList, to the next vaarg
3212 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3213 DAG.getConstant(TLI.getDataLayout()->
3214 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3215 VAList.getValueType()));
3216 // Store the incremented VAList to the legalized pointer
3217 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3218 MachinePointerInfo(V), false, false, 0);
3219 // Load the actual argument out of the pointer VAList
3220 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3221 false, false, false, 0));
3222 Results.push_back(Results[0].getValue(1));
3226 // This defaults to loading a pointer from the input and storing it to the
3227 // output, returning the chain.
3228 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3229 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3230 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3231 Node->getOperand(2), MachinePointerInfo(VS),
3232 false, false, false, 0);
3233 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3234 MachinePointerInfo(VD), false, false, 0);
3235 Results.push_back(Tmp1);
3238 case ISD::EXTRACT_VECTOR_ELT:
3239 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3240 // This must be an access of the only element. Return it.
3241 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3242 Node->getOperand(0));
3244 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3245 Results.push_back(Tmp1);
3247 case ISD::EXTRACT_SUBVECTOR:
3248 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3250 case ISD::INSERT_SUBVECTOR:
3251 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3253 case ISD::CONCAT_VECTORS: {
3254 Results.push_back(ExpandVectorBuildThroughStack(Node));
3257 case ISD::SCALAR_TO_VECTOR:
3258 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3260 case ISD::INSERT_VECTOR_ELT:
3261 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3262 Node->getOperand(1),
3263 Node->getOperand(2), dl));
3265 case ISD::VECTOR_SHUFFLE: {
3266 SmallVector<int, 32> NewMask;
3267 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3269 EVT VT = Node->getValueType(0);
3270 EVT EltVT = VT.getVectorElementType();
3271 SDValue Op0 = Node->getOperand(0);
3272 SDValue Op1 = Node->getOperand(1);
3273 if (!TLI.isTypeLegal(EltVT)) {
3275 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3277 // BUILD_VECTOR operands are allowed to be wider than the element type.
3278 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3280 if (NewEltVT.bitsLT(EltVT)) {
3282 // Convert shuffle node.
3283 // If original node was v4i64 and the new EltVT is i32,
3284 // cast operands to v8i32 and re-build the mask.
3286 // Calculate new VT, the size of the new VT should be equal to original.
3288 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3289 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3290 assert(NewVT.bitsEq(VT));
3292 // cast operands to new VT
3293 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3294 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3296 // Convert the shuffle mask
3297 unsigned int factor =
3298 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3300 // EltVT gets smaller
3303 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3305 for (unsigned fi = 0; fi < factor; ++fi)
3306 NewMask.push_back(Mask[i]);
3309 for (unsigned fi = 0; fi < factor; ++fi)
3310 NewMask.push_back(Mask[i]*factor+fi);
3318 unsigned NumElems = VT.getVectorNumElements();
3319 SmallVector<SDValue, 16> Ops;
3320 for (unsigned i = 0; i != NumElems; ++i) {
3322 Ops.push_back(DAG.getUNDEF(EltVT));
3325 unsigned Idx = Mask[i];
3327 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3329 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3331 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3333 DAG.getConstant(Idx - NumElems,
3334 TLI.getVectorIdxTy())));
3337 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3338 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3339 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3340 Results.push_back(Tmp1);
3343 case ISD::EXTRACT_ELEMENT: {
3344 EVT OpTy = Node->getOperand(0).getValueType();
3345 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3347 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3348 DAG.getConstant(OpTy.getSizeInBits()/2,
3349 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3350 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3353 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3354 Node->getOperand(0));
3356 Results.push_back(Tmp1);
3359 case ISD::STACKSAVE:
3360 // Expand to CopyFromReg if the target set
3361 // StackPointerRegisterToSaveRestore.
3362 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3363 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3364 Node->getValueType(0)));
3365 Results.push_back(Results[0].getValue(1));
3367 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3368 Results.push_back(Node->getOperand(0));
3371 case ISD::STACKRESTORE:
3372 // Expand to CopyToReg if the target set
3373 // StackPointerRegisterToSaveRestore.
3374 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3375 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3376 Node->getOperand(1)));
3378 Results.push_back(Node->getOperand(0));
3381 case ISD::FCOPYSIGN:
3382 Results.push_back(ExpandFCOPYSIGN(Node));
3385 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3386 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3387 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3388 Node->getOperand(0));
3389 Results.push_back(Tmp1);
3392 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3393 EVT VT = Node->getValueType(0);
3394 Tmp1 = Node->getOperand(0);
3395 Tmp2 = DAG.getConstantFP(0.0, VT);
3396 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3397 Tmp1, Tmp2, ISD::SETUGT);
3398 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3399 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3400 Results.push_back(Tmp1);
3404 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3405 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3406 RTLIB::SQRT_PPCF128));
3410 EVT VT = Node->getValueType(0);
3411 bool isSIN = Node->getOpcode() == ISD::FSIN;
3412 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3413 // fcos which share the same operand and both are used.
3414 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3415 canCombineSinCosLibcall(Node, TLI, TM))
3416 && useSinCos(Node)) {
3417 SDVTList VTs = DAG.getVTList(VT, VT);
3418 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3420 Tmp1 = Tmp1.getValue(1);
3421 Results.push_back(Tmp1);
3423 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3424 RTLIB::SIN_F80, RTLIB::SIN_F128,
3425 RTLIB::SIN_PPCF128));
3427 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3428 RTLIB::COS_F80, RTLIB::COS_F128,
3429 RTLIB::COS_PPCF128));
3434 // Expand into sincos libcall.
3435 ExpandSinCosLibCall(Node, Results);
3438 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3439 RTLIB::LOG_F80, RTLIB::LOG_F128,
3440 RTLIB::LOG_PPCF128));
3443 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3444 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3445 RTLIB::LOG2_PPCF128));
3448 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3449 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3450 RTLIB::LOG10_PPCF128));
3453 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3454 RTLIB::EXP_F80, RTLIB::EXP_F128,
3455 RTLIB::EXP_PPCF128));
3458 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3459 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3460 RTLIB::EXP2_PPCF128));
3463 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3464 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3465 RTLIB::TRUNC_PPCF128));
3468 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3469 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3470 RTLIB::FLOOR_PPCF128));
3473 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3474 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3475 RTLIB::CEIL_PPCF128));
3478 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3479 RTLIB::RINT_F80, RTLIB::RINT_F128,
3480 RTLIB::RINT_PPCF128));
3482 case ISD::FNEARBYINT:
3483 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3484 RTLIB::NEARBYINT_F64,
3485 RTLIB::NEARBYINT_F80,
3486 RTLIB::NEARBYINT_F128,
3487 RTLIB::NEARBYINT_PPCF128));
3490 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3494 RTLIB::ROUND_PPCF128));
3497 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3498 RTLIB::POWI_F80, RTLIB::POWI_F128,
3499 RTLIB::POWI_PPCF128));
3502 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3503 RTLIB::POW_F80, RTLIB::POW_F128,
3504 RTLIB::POW_PPCF128));
3507 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3508 RTLIB::DIV_F80, RTLIB::DIV_F128,
3509 RTLIB::DIV_PPCF128));
3512 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3513 RTLIB::REM_F80, RTLIB::REM_F128,
3514 RTLIB::REM_PPCF128));
3517 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3518 RTLIB::FMA_F80, RTLIB::FMA_F128,
3519 RTLIB::FMA_PPCF128));
3522 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
3523 RTLIB::ADD_F80, RTLIB::ADD_F128,
3524 RTLIB::ADD_PPCF128));
3527 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
3528 RTLIB::MUL_F80, RTLIB::MUL_F128,
3529 RTLIB::MUL_PPCF128));
3531 case ISD::FP16_TO_FP: {
3532 if (Node->getValueType(0) == MVT::f32) {
3533 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3537 // We can extend to types bigger than f32 in two steps without changing the
3538 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3539 // the option of emitting that before resorting to a libcall.
3541 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3543 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3546 case ISD::FP_TO_FP16: {
3548 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3549 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3550 Results.push_back(ExpandLibCall(LC, Node, false));
3553 case ISD::ConstantFP: {
3554 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3555 // Check to see if this FP immediate is already legal.
3556 // If this is a legal constant, turn it into a TargetConstantFP node.
3557 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3558 Results.push_back(ExpandConstantFP(CFP, true));
3562 EVT VT = Node->getValueType(0);
3563 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3564 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3565 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3566 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3567 Results.push_back(Tmp1);
3569 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
3570 RTLIB::SUB_F80, RTLIB::SUB_F128,
3571 RTLIB::SUB_PPCF128));
3576 EVT VT = Node->getValueType(0);
3577 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3578 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3579 "Don't know how to expand this subtraction!");
3580 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3581 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3582 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3583 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3588 EVT VT = Node->getValueType(0);
3589 bool isSigned = Node->getOpcode() == ISD::SREM;
3590 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3591 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3592 Tmp2 = Node->getOperand(0);
3593 Tmp3 = Node->getOperand(1);
3594 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3595 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3596 // If div is legal, it's better to do the normal expansion
3597 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3598 useDivRem(Node, isSigned, false))) {
3599 SDVTList VTs = DAG.getVTList(VT, VT);
3600 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3601 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3603 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3604 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3605 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3606 } else if (isSigned)
3607 Tmp1 = ExpandIntLibCall(Node, true,
3609 RTLIB::SREM_I16, RTLIB::SREM_I32,
3610 RTLIB::SREM_I64, RTLIB::SREM_I128);
3612 Tmp1 = ExpandIntLibCall(Node, false,
3614 RTLIB::UREM_I16, RTLIB::UREM_I32,
3615 RTLIB::UREM_I64, RTLIB::UREM_I128);
3616 Results.push_back(Tmp1);
3621 bool isSigned = Node->getOpcode() == ISD::SDIV;
3622 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3623 EVT VT = Node->getValueType(0);
3624 SDVTList VTs = DAG.getVTList(VT, VT);
3625 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3626 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3627 useDivRem(Node, isSigned, true)))
3628 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3629 Node->getOperand(1));
3631 Tmp1 = ExpandIntLibCall(Node, true,
3633 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3634 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3636 Tmp1 = ExpandIntLibCall(Node, false,
3638 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3639 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3640 Results.push_back(Tmp1);
3645 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3647 EVT VT = Node->getValueType(0);
3648 SDVTList VTs = DAG.getVTList(VT, VT);
3649 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3650 "If this wasn't legal, it shouldn't have been created!");
3651 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3652 Node->getOperand(1));
3653 Results.push_back(Tmp1.getValue(1));
3658 // Expand into divrem libcall
3659 ExpandDivRemLibCall(Node, Results);
3662 EVT VT = Node->getValueType(0);
3663 SDVTList VTs = DAG.getVTList(VT, VT);
3664 // See if multiply or divide can be lowered using two-result operations.
3665 // We just need the low half of the multiply; try both the signed
3666 // and unsigned forms. If the target supports both SMUL_LOHI and
3667 // UMUL_LOHI, form a preference by checking which forms of plain
3668 // MULH it supports.
3669 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3670 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3671 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3672 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3673 unsigned OpToUse = 0;
3674 if (HasSMUL_LOHI && !HasMULHS) {
3675 OpToUse = ISD::SMUL_LOHI;
3676 } else if (HasUMUL_LOHI && !HasMULHU) {
3677 OpToUse = ISD::UMUL_LOHI;
3678 } else if (HasSMUL_LOHI) {
3679 OpToUse = ISD::SMUL_LOHI;
3680 } else if (HasUMUL_LOHI) {
3681 OpToUse = ISD::UMUL_LOHI;
3684 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3685 Node->getOperand(1)));