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"
39 #define DEBUG_TYPE "legalizedag"
41 //===----------------------------------------------------------------------===//
42 /// SelectionDAGLegalize - This takes an arbitrary SelectionDAG as input and
43 /// hacks on it until the target machine can handle it. This involves
44 /// eliminating value sizes the machine cannot handle (promoting small sizes to
45 /// large sizes or splitting up large values into small values) as well as
46 /// eliminating operations the machine cannot handle.
48 /// This code also does a small amount of optimization and recognition of idioms
49 /// as part of its processing. For example, if a target does not support a
50 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
51 /// will attempt merge setcc and brc instructions into brcc's.
54 class SelectionDAGLegalize : public SelectionDAG::DAGUpdateListener {
55 const TargetMachine &TM;
56 const TargetLowering &TLI;
59 /// \brief The iterator being used to walk the DAG. We hold a reference to it
60 /// in order to update it as necessary on node deletion.
61 SelectionDAG::allnodes_iterator &LegalizePosition;
63 /// \brief The set of nodes which have already been legalized. We hold a
64 /// reference to it in order to update as necessary on node deletion.
65 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
67 /// \brief A set of all the nodes updated during legalization.
68 SmallSetVector<SDNode *, 16> *UpdatedNodes;
70 EVT getSetCCResultType(EVT VT) const {
71 return TLI.getSetCCResultType(*DAG.getContext(), VT);
74 // Libcall insertion helpers.
77 SelectionDAGLegalize(SelectionDAG &DAG,
78 SelectionDAG::allnodes_iterator &LegalizePosition,
79 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
80 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
81 : SelectionDAG::DAGUpdateListener(DAG), TM(DAG.getTarget()),
82 TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
83 LegalizePosition(LegalizePosition), LegalizedNodes(LegalizedNodes),
84 UpdatedNodes(UpdatedNodes) {}
86 /// \brief Legalizes the given operation.
87 void LegalizeOp(SDNode *Node);
90 SDValue OptimizeFloatStore(StoreSDNode *ST);
92 void LegalizeLoadOps(SDNode *Node);
93 void LegalizeStoreOps(SDNode *Node);
95 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
96 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
97 /// is necessary to spill the vector being inserted into to memory, perform
98 /// the insert there, and then read the result back.
99 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
100 SDValue Idx, SDLoc dl);
101 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
102 SDValue Idx, SDLoc dl);
104 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
105 /// performs the same shuffe in terms of order or result bytes, but on a type
106 /// whose vector element type is narrower than the original shuffle type.
107 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
108 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
109 SDValue N1, SDValue N2,
110 ArrayRef<int> Mask) const;
112 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
113 bool &NeedInvert, SDLoc dl);
115 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
116 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
117 unsigned NumOps, bool isSigned, SDLoc dl);
119 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
120 SDNode *Node, bool isSigned);
121 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
122 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
123 RTLIB::Libcall Call_F128,
124 RTLIB::Libcall Call_PPCF128);
125 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
126 RTLIB::Libcall Call_I8,
127 RTLIB::Libcall Call_I16,
128 RTLIB::Libcall Call_I32,
129 RTLIB::Libcall Call_I64,
130 RTLIB::Libcall Call_I128);
131 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
132 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
134 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
135 SDValue ExpandBUILD_VECTOR(SDNode *Node);
136 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
137 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
138 SmallVectorImpl<SDValue> &Results);
139 SDValue ExpandFCOPYSIGN(SDNode *Node);
140 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
142 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
144 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
147 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
148 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
150 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
151 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
152 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
154 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
156 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
158 void ExpandNode(SDNode *Node);
159 void PromoteNode(SDNode *Node);
161 void ForgetNode(SDNode *N) {
162 LegalizedNodes.erase(N);
163 if (LegalizePosition == SelectionDAG::allnodes_iterator(N))
166 UpdatedNodes->remove(N);
170 // DAGUpdateListener implementation.
171 void NodeDeleted(SDNode *N, SDNode *E) override {
174 void NodeUpdated(SDNode *N) override {}
176 // Node replacement helpers
177 void ReplacedNode(SDNode *N) {
178 if (N->use_empty()) {
179 DAG.RemoveDeadNode(N);
184 void ReplaceNode(SDNode *Old, SDNode *New) {
185 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
186 dbgs() << " with: "; New->dump(&DAG));
188 assert(Old->getNumValues() == New->getNumValues() &&
189 "Replacing one node with another that produces a different number "
191 DAG.ReplaceAllUsesWith(Old, New);
192 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
193 DAG.TransferDbgValues(SDValue(Old, i), SDValue(New, i));
195 UpdatedNodes->insert(New);
198 void ReplaceNode(SDValue Old, SDValue New) {
199 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
200 dbgs() << " with: "; New->dump(&DAG));
202 DAG.ReplaceAllUsesWith(Old, New);
203 DAG.TransferDbgValues(Old, New);
205 UpdatedNodes->insert(New.getNode());
206 ReplacedNode(Old.getNode());
208 void ReplaceNode(SDNode *Old, const SDValue *New) {
209 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
211 DAG.ReplaceAllUsesWith(Old, New);
212 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
213 DEBUG(dbgs() << (i == 0 ? " with: "
216 DAG.TransferDbgValues(SDValue(Old, i), New[i]);
218 UpdatedNodes->insert(New[i].getNode());
225 /// ShuffleWithNarrowerEltType - Return a vector shuffle operation which
226 /// performs the same shuffe in terms of order or result bytes, but on a type
227 /// whose vector element type is narrower than the original shuffle type.
228 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
230 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
231 SDValue N1, SDValue N2,
232 ArrayRef<int> Mask) const {
233 unsigned NumMaskElts = VT.getVectorNumElements();
234 unsigned NumDestElts = NVT.getVectorNumElements();
235 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
237 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
239 if (NumEltsGrowth == 1)
240 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
242 SmallVector<int, 8> NewMask;
243 for (unsigned i = 0; i != NumMaskElts; ++i) {
245 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
247 NewMask.push_back(-1);
249 NewMask.push_back(Idx * NumEltsGrowth + j);
252 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
253 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
254 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
257 /// ExpandConstantFP - Expands the ConstantFP node to an integer constant or
258 /// a load from the constant pool.
260 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
264 // If a FP immediate is precise when represented as a float and if the
265 // target can do an extending load from float to double, we put it into
266 // the constant pool as a float, even if it's is statically typed as a
267 // double. This shrinks FP constants and canonicalizes them for targets where
268 // an FP extending load is the same cost as a normal load (such as on the x87
269 // fp stack or PPC FP unit).
270 EVT VT = CFP->getValueType(0);
271 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
273 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
274 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
275 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
280 while (SVT != MVT::f32 && SVT != MVT::f16) {
281 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
282 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
283 // Only do this if the target has a native EXTLOAD instruction from
285 TLI.isLoadExtLegal(ISD::EXTLOAD, SVT) &&
286 TLI.ShouldShrinkFPConstant(OrigVT)) {
287 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
288 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
294 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
295 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
298 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
300 CPIdx, MachinePointerInfo::getConstantPool(),
301 VT, false, false, Alignment);
305 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
306 MachinePointerInfo::getConstantPool(), false, false, false,
311 /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores.
312 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
313 const TargetLowering &TLI,
314 SelectionDAGLegalize *DAGLegalize) {
315 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
316 "unaligned indexed stores not implemented!");
317 SDValue Chain = ST->getChain();
318 SDValue Ptr = ST->getBasePtr();
319 SDValue Val = ST->getValue();
320 EVT VT = Val.getValueType();
321 int Alignment = ST->getAlignment();
322 unsigned AS = ST->getAddressSpace();
325 if (ST->getMemoryVT().isFloatingPoint() ||
326 ST->getMemoryVT().isVector()) {
327 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
328 if (TLI.isTypeLegal(intVT)) {
329 // Expand to a bitconvert of the value to the integer type of the
330 // same size, then a (misaligned) int store.
331 // FIXME: Does not handle truncating floating point stores!
332 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
333 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
334 ST->isVolatile(), ST->isNonTemporal(), Alignment);
335 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
338 // Do a (aligned) store to a stack slot, then copy from the stack slot
339 // to the final destination using (unaligned) integer loads and stores.
340 EVT StoredVT = ST->getMemoryVT();
342 TLI.getRegisterType(*DAG.getContext(),
343 EVT::getIntegerVT(*DAG.getContext(),
344 StoredVT.getSizeInBits()));
345 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
346 unsigned RegBytes = RegVT.getSizeInBits() / 8;
347 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
349 // Make sure the stack slot is also aligned for the register type.
350 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
352 // Perform the original store, only redirected to the stack slot.
353 SDValue Store = DAG.getTruncStore(Chain, dl,
354 Val, StackPtr, MachinePointerInfo(),
355 StoredVT, false, false, 0);
356 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy(AS));
357 SmallVector<SDValue, 8> Stores;
360 // Do all but one copies using the full register width.
361 for (unsigned i = 1; i < NumRegs; i++) {
362 // Load one integer register's worth from the stack slot.
363 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
364 MachinePointerInfo(),
365 false, false, false, 0);
366 // Store it to the final location. Remember the store.
367 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
368 ST->getPointerInfo().getWithOffset(Offset),
369 ST->isVolatile(), ST->isNonTemporal(),
370 MinAlign(ST->getAlignment(), Offset)));
371 // Increment the pointers.
373 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
375 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
378 // The last store may be partial. Do a truncating store. On big-endian
379 // machines this requires an extending load from the stack slot to ensure
380 // that the bits are in the right place.
381 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
382 8 * (StoredBytes - Offset));
384 // Load from the stack slot.
385 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
386 MachinePointerInfo(),
387 MemVT, false, false, 0);
389 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
391 .getWithOffset(Offset),
392 MemVT, ST->isVolatile(),
394 MinAlign(ST->getAlignment(), Offset),
396 // The order of the stores doesn't matter - say it with a TokenFactor.
397 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
398 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
401 assert(ST->getMemoryVT().isInteger() &&
402 !ST->getMemoryVT().isVector() &&
403 "Unaligned store of unknown type.");
404 // Get the half-size VT
405 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
406 int NumBits = NewStoredVT.getSizeInBits();
407 int IncrementSize = NumBits / 8;
409 // Divide the stored value in two parts.
410 SDValue ShiftAmount = DAG.getConstant(NumBits,
411 TLI.getShiftAmountTy(Val.getValueType()));
413 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
415 // Store the two parts
416 SDValue Store1, Store2;
417 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
418 ST->getPointerInfo(), NewStoredVT,
419 ST->isVolatile(), ST->isNonTemporal(), Alignment);
421 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
422 DAG.getConstant(IncrementSize, TLI.getPointerTy(AS)));
423 Alignment = MinAlign(Alignment, IncrementSize);
424 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
425 ST->getPointerInfo().getWithOffset(IncrementSize),
426 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
427 Alignment, ST->getAAInfo());
430 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
431 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
434 /// ExpandUnalignedLoad - Expands an unaligned load to 2 half-size loads.
436 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
437 const TargetLowering &TLI,
438 SDValue &ValResult, SDValue &ChainResult) {
439 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
440 "unaligned indexed loads not implemented!");
441 SDValue Chain = LD->getChain();
442 SDValue Ptr = LD->getBasePtr();
443 EVT VT = LD->getValueType(0);
444 EVT LoadedVT = LD->getMemoryVT();
446 if (VT.isFloatingPoint() || VT.isVector()) {
447 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
448 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
449 // Expand to a (misaligned) integer load of the same size,
450 // then bitconvert to floating point or vector.
451 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
452 LD->getMemOperand());
453 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
455 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
456 ISD::ANY_EXTEND, dl, VT, Result);
463 // Copy the value to a (aligned) stack slot using (unaligned) integer
464 // loads and stores, then do a (aligned) load from the stack slot.
465 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
466 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
467 unsigned RegBytes = RegVT.getSizeInBits() / 8;
468 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
470 // Make sure the stack slot is also aligned for the register type.
471 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
473 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
474 SmallVector<SDValue, 8> Stores;
475 SDValue StackPtr = StackBase;
478 // Do all but one copies using the full register width.
479 for (unsigned i = 1; i < NumRegs; i++) {
480 // Load one integer register's worth from the original location.
481 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
482 LD->getPointerInfo().getWithOffset(Offset),
483 LD->isVolatile(), LD->isNonTemporal(),
485 MinAlign(LD->getAlignment(), Offset),
487 // Follow the load with a store to the stack slot. Remember the store.
488 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
489 MachinePointerInfo(), false, false, 0));
490 // Increment the pointers.
492 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
493 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
497 // The last copy may be partial. Do an extending load.
498 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
499 8 * (LoadedBytes - Offset));
500 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
501 LD->getPointerInfo().getWithOffset(Offset),
502 MemVT, LD->isVolatile(),
504 MinAlign(LD->getAlignment(), Offset),
506 // Follow the load with a store to the stack slot. Remember the store.
507 // On big-endian machines this requires a truncating store to ensure
508 // that the bits end up in the right place.
509 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
510 MachinePointerInfo(), MemVT,
513 // The order of the stores doesn't matter - say it with a TokenFactor.
514 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
516 // Finally, perform the original load only redirected to the stack slot.
517 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
518 MachinePointerInfo(), LoadedVT, false, false, 0);
520 // Callers expect a MERGE_VALUES node.
525 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
526 "Unaligned load of unsupported type.");
528 // Compute the new VT that is half the size of the old one. This is an
530 unsigned NumBits = LoadedVT.getSizeInBits();
532 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
535 unsigned Alignment = LD->getAlignment();
536 unsigned IncrementSize = NumBits / 8;
537 ISD::LoadExtType HiExtType = LD->getExtensionType();
539 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
540 if (HiExtType == ISD::NON_EXTLOAD)
541 HiExtType = ISD::ZEXTLOAD;
543 // Load the value in two parts
545 if (TLI.isLittleEndian()) {
546 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
547 NewLoadedVT, LD->isVolatile(),
548 LD->isNonTemporal(), Alignment, LD->getAAInfo());
549 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
550 DAG.getConstant(IncrementSize, Ptr.getValueType()));
551 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
552 LD->getPointerInfo().getWithOffset(IncrementSize),
553 NewLoadedVT, LD->isVolatile(),
554 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
557 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
558 NewLoadedVT, LD->isVolatile(),
559 LD->isNonTemporal(), Alignment, LD->getAAInfo());
560 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
561 DAG.getConstant(IncrementSize, Ptr.getValueType()));
562 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
563 LD->getPointerInfo().getWithOffset(IncrementSize),
564 NewLoadedVT, LD->isVolatile(),
565 LD->isNonTemporal(), MinAlign(Alignment, IncrementSize),
569 // aggregate the two parts
570 SDValue ShiftAmount = DAG.getConstant(NumBits,
571 TLI.getShiftAmountTy(Hi.getValueType()));
572 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
573 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
575 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
582 /// PerformInsertVectorEltInMemory - Some target cannot handle a variable
583 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
584 /// is necessary to spill the vector being inserted into to memory, perform
585 /// the insert there, and then read the result back.
586 SDValue SelectionDAGLegalize::
587 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
593 // If the target doesn't support this, we have to spill the input vector
594 // to a temporary stack slot, update the element, then reload it. This is
595 // badness. We could also load the value into a vector register (either
596 // with a "move to register" or "extload into register" instruction, then
597 // permute it into place, if the idx is a constant and if the idx is
598 // supported by the target.
599 EVT VT = Tmp1.getValueType();
600 EVT EltVT = VT.getVectorElementType();
601 EVT IdxVT = Tmp3.getValueType();
602 EVT PtrVT = TLI.getPointerTy();
603 SDValue StackPtr = DAG.CreateStackTemporary(VT);
605 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
608 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
609 MachinePointerInfo::getFixedStack(SPFI),
612 // Truncate or zero extend offset to target pointer type.
613 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
614 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
615 // Add the offset to the index.
616 unsigned EltSize = EltVT.getSizeInBits()/8;
617 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
618 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
619 // Store the scalar value.
620 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
622 // Load the updated vector.
623 return DAG.getLoad(VT, dl, Ch, StackPtr,
624 MachinePointerInfo::getFixedStack(SPFI), false, false,
629 SDValue SelectionDAGLegalize::
630 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
631 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
632 // SCALAR_TO_VECTOR requires that the type of the value being inserted
633 // match the element type of the vector being created, except for
634 // integers in which case the inserted value can be over width.
635 EVT EltVT = Vec.getValueType().getVectorElementType();
636 if (Val.getValueType() == EltVT ||
637 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
638 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
639 Vec.getValueType(), Val);
641 unsigned NumElts = Vec.getValueType().getVectorNumElements();
642 // We generate a shuffle of InVec and ScVec, so the shuffle mask
643 // should be 0,1,2,3,4,5... with the appropriate element replaced with
645 SmallVector<int, 8> ShufOps;
646 for (unsigned i = 0; i != NumElts; ++i)
647 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
649 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
653 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
656 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
657 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
658 // FIXME: We shouldn't do this for TargetConstantFP's.
659 // FIXME: move this to the DAG Combiner! Note that we can't regress due
660 // to phase ordering between legalized code and the dag combiner. This
661 // probably means that we need to integrate dag combiner and legalizer
663 // We generally can't do this one for long doubles.
664 SDValue Chain = ST->getChain();
665 SDValue Ptr = ST->getBasePtr();
666 unsigned Alignment = ST->getAlignment();
667 bool isVolatile = ST->isVolatile();
668 bool isNonTemporal = ST->isNonTemporal();
669 AAMDNodes AAInfo = ST->getAAInfo();
671 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
672 if (CFP->getValueType(0) == MVT::f32 &&
673 TLI.isTypeLegal(MVT::i32)) {
674 SDValue Con = DAG.getConstant(CFP->getValueAPF().
675 bitcastToAPInt().zextOrTrunc(32),
677 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
678 isVolatile, isNonTemporal, Alignment, AAInfo);
681 if (CFP->getValueType(0) == MVT::f64) {
682 // If this target supports 64-bit registers, do a single 64-bit store.
683 if (TLI.isTypeLegal(MVT::i64)) {
684 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
685 zextOrTrunc(64), MVT::i64);
686 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
687 isVolatile, isNonTemporal, Alignment, AAInfo);
690 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
691 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
692 // stores. If the target supports neither 32- nor 64-bits, this
693 // xform is certainly not worth it.
694 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
695 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
696 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
697 if (TLI.isBigEndian()) std::swap(Lo, Hi);
699 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
700 isNonTemporal, Alignment, AAInfo);
701 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
702 DAG.getConstant(4, Ptr.getValueType()));
703 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
704 ST->getPointerInfo().getWithOffset(4),
705 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
708 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
712 return SDValue(nullptr, 0);
715 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
716 StoreSDNode *ST = cast<StoreSDNode>(Node);
717 SDValue Chain = ST->getChain();
718 SDValue Ptr = ST->getBasePtr();
721 unsigned Alignment = ST->getAlignment();
722 bool isVolatile = ST->isVolatile();
723 bool isNonTemporal = ST->isNonTemporal();
724 AAMDNodes AAInfo = ST->getAAInfo();
726 if (!ST->isTruncatingStore()) {
727 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
728 ReplaceNode(ST, OptStore);
733 SDValue Value = ST->getValue();
734 MVT VT = Value.getSimpleValueType();
735 switch (TLI.getOperationAction(ISD::STORE, VT)) {
736 default: llvm_unreachable("This action is not supported yet!");
737 case TargetLowering::Legal: {
738 // If this is an unaligned store and the target doesn't support it,
740 unsigned AS = ST->getAddressSpace();
741 unsigned Align = ST->getAlignment();
742 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
743 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
744 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
745 if (Align < ABIAlignment)
746 ExpandUnalignedStore(cast<StoreSDNode>(Node),
751 case TargetLowering::Custom: {
752 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
754 ReplaceNode(SDValue(Node, 0), Res);
757 case TargetLowering::Promote: {
758 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
759 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
760 "Can only promote stores to same size type");
761 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
763 DAG.getStore(Chain, dl, Value, Ptr,
764 ST->getPointerInfo(), isVolatile,
765 isNonTemporal, Alignment, AAInfo);
766 ReplaceNode(SDValue(Node, 0), Result);
773 SDValue Value = ST->getValue();
775 EVT StVT = ST->getMemoryVT();
776 unsigned StWidth = StVT.getSizeInBits();
778 if (StWidth != StVT.getStoreSizeInBits()) {
779 // Promote to a byte-sized store with upper bits zero if not
780 // storing an integral number of bytes. For example, promote
781 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
782 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
783 StVT.getStoreSizeInBits());
784 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
786 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
787 NVT, isVolatile, isNonTemporal, Alignment,
789 ReplaceNode(SDValue(Node, 0), Result);
790 } else if (StWidth & (StWidth - 1)) {
791 // If not storing a power-of-2 number of bits, expand as two stores.
792 assert(!StVT.isVector() && "Unsupported truncstore!");
793 unsigned RoundWidth = 1 << Log2_32(StWidth);
794 assert(RoundWidth < StWidth);
795 unsigned ExtraWidth = StWidth - RoundWidth;
796 assert(ExtraWidth < RoundWidth);
797 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
798 "Store size not an integral number of bytes!");
799 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
800 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
802 unsigned IncrementSize;
804 if (TLI.isLittleEndian()) {
805 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
806 // Store the bottom RoundWidth bits.
807 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
809 isVolatile, isNonTemporal, Alignment,
812 // Store the remaining ExtraWidth bits.
813 IncrementSize = RoundWidth / 8;
814 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
815 DAG.getConstant(IncrementSize, Ptr.getValueType()));
816 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
817 DAG.getConstant(RoundWidth,
818 TLI.getShiftAmountTy(Value.getValueType())));
819 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
820 ST->getPointerInfo().getWithOffset(IncrementSize),
821 ExtraVT, isVolatile, isNonTemporal,
822 MinAlign(Alignment, IncrementSize), AAInfo);
824 // Big endian - avoid unaligned stores.
825 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
826 // Store the top RoundWidth bits.
827 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
828 DAG.getConstant(ExtraWidth,
829 TLI.getShiftAmountTy(Value.getValueType())));
830 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
831 RoundVT, isVolatile, isNonTemporal, Alignment,
834 // Store the remaining ExtraWidth bits.
835 IncrementSize = RoundWidth / 8;
836 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
837 DAG.getConstant(IncrementSize, Ptr.getValueType()));
838 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
839 ST->getPointerInfo().getWithOffset(IncrementSize),
840 ExtraVT, isVolatile, isNonTemporal,
841 MinAlign(Alignment, IncrementSize), AAInfo);
844 // The order of the stores doesn't matter.
845 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
846 ReplaceNode(SDValue(Node, 0), Result);
848 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
849 StVT.getSimpleVT())) {
850 default: llvm_unreachable("This action is not supported yet!");
851 case TargetLowering::Legal: {
852 unsigned AS = ST->getAddressSpace();
853 unsigned Align = ST->getAlignment();
854 // If this is an unaligned store and the target doesn't support it,
856 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
857 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
858 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
859 if (Align < ABIAlignment)
860 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
864 case TargetLowering::Custom: {
865 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
867 ReplaceNode(SDValue(Node, 0), Res);
870 case TargetLowering::Expand:
871 assert(!StVT.isVector() &&
872 "Vector Stores are handled in LegalizeVectorOps");
874 // TRUNCSTORE:i16 i32 -> STORE i16
875 assert(TLI.isTypeLegal(StVT) &&
876 "Do not know how to expand this store!");
877 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
879 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
880 isVolatile, isNonTemporal, Alignment, AAInfo);
881 ReplaceNode(SDValue(Node, 0), Result);
888 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
889 LoadSDNode *LD = cast<LoadSDNode>(Node);
890 SDValue Chain = LD->getChain(); // The chain.
891 SDValue Ptr = LD->getBasePtr(); // The base pointer.
892 SDValue Value; // The value returned by the load op.
895 ISD::LoadExtType ExtType = LD->getExtensionType();
896 if (ExtType == ISD::NON_EXTLOAD) {
897 MVT VT = Node->getSimpleValueType(0);
898 SDValue RVal = SDValue(Node, 0);
899 SDValue RChain = SDValue(Node, 1);
901 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
902 default: llvm_unreachable("This action is not supported yet!");
903 case TargetLowering::Legal: {
904 unsigned AS = LD->getAddressSpace();
905 unsigned Align = LD->getAlignment();
906 // If this is an unaligned load and the target doesn't support it,
908 if (!TLI.allowsMisalignedMemoryAccesses(LD->getMemoryVT(), AS, Align)) {
909 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
910 unsigned ABIAlignment =
911 TLI.getDataLayout()->getABITypeAlignment(Ty);
912 if (Align < ABIAlignment){
913 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
918 case TargetLowering::Custom: {
919 SDValue Res = TLI.LowerOperation(RVal, DAG);
922 RChain = Res.getValue(1);
926 case TargetLowering::Promote: {
927 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
928 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
929 "Can only promote loads to same size type");
931 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
932 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
933 RChain = Res.getValue(1);
937 if (RChain.getNode() != Node) {
938 assert(RVal.getNode() != Node && "Load must be completely replaced");
939 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
940 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
942 UpdatedNodes->insert(RVal.getNode());
943 UpdatedNodes->insert(RChain.getNode());
950 EVT SrcVT = LD->getMemoryVT();
951 unsigned SrcWidth = SrcVT.getSizeInBits();
952 unsigned Alignment = LD->getAlignment();
953 bool isVolatile = LD->isVolatile();
954 bool isNonTemporal = LD->isNonTemporal();
955 AAMDNodes AAInfo = LD->getAAInfo();
957 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
958 // Some targets pretend to have an i1 loading operation, and actually
959 // load an i8. This trick is correct for ZEXTLOAD because the top 7
960 // bits are guaranteed to be zero; it helps the optimizers understand
961 // that these bits are zero. It is also useful for EXTLOAD, since it
962 // tells the optimizers that those bits are undefined. It would be
963 // nice to have an effective generic way of getting these benefits...
964 // Until such a way is found, don't insist on promoting i1 here.
966 TLI.getLoadExtAction(ExtType, MVT::i1) == TargetLowering::Promote)) {
967 // Promote to a byte-sized load if not loading an integral number of
968 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
969 unsigned NewWidth = SrcVT.getStoreSizeInBits();
970 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
973 // The extra bits are guaranteed to be zero, since we stored them that
974 // way. A zext load from NVT thus automatically gives zext from SrcVT.
976 ISD::LoadExtType NewExtType =
977 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
980 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
981 Chain, Ptr, LD->getPointerInfo(),
982 NVT, isVolatile, isNonTemporal, Alignment, AAInfo);
984 Ch = Result.getValue(1); // The chain.
986 if (ExtType == ISD::SEXTLOAD)
987 // Having the top bits zero doesn't help when sign extending.
988 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
989 Result.getValueType(),
990 Result, DAG.getValueType(SrcVT));
991 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
992 // All the top bits are guaranteed to be zero - inform the optimizers.
993 Result = DAG.getNode(ISD::AssertZext, dl,
994 Result.getValueType(), Result,
995 DAG.getValueType(SrcVT));
999 } else if (SrcWidth & (SrcWidth - 1)) {
1000 // If not loading a power-of-2 number of bits, expand as two loads.
1001 assert(!SrcVT.isVector() && "Unsupported extload!");
1002 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
1003 assert(RoundWidth < SrcWidth);
1004 unsigned ExtraWidth = SrcWidth - RoundWidth;
1005 assert(ExtraWidth < RoundWidth);
1006 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
1007 "Load size not an integral number of bytes!");
1008 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
1009 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
1011 unsigned IncrementSize;
1013 if (TLI.isLittleEndian()) {
1014 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
1015 // Load the bottom RoundWidth bits.
1016 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1018 LD->getPointerInfo(), RoundVT, isVolatile,
1019 isNonTemporal, Alignment, AAInfo);
1021 // Load the remaining ExtraWidth bits.
1022 IncrementSize = RoundWidth / 8;
1023 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1024 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1025 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1026 LD->getPointerInfo().getWithOffset(IncrementSize),
1027 ExtraVT, isVolatile, isNonTemporal,
1028 MinAlign(Alignment, IncrementSize), AAInfo);
1030 // Build a factor node to remember that this load is independent of
1032 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1035 // Move the top bits to the right place.
1036 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1037 DAG.getConstant(RoundWidth,
1038 TLI.getShiftAmountTy(Hi.getValueType())));
1040 // Join the hi and lo parts.
1041 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1043 // Big endian - avoid unaligned loads.
1044 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1045 // Load the top RoundWidth bits.
1046 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1047 LD->getPointerInfo(), RoundVT, isVolatile,
1048 isNonTemporal, Alignment, AAInfo);
1050 // Load the remaining ExtraWidth bits.
1051 IncrementSize = RoundWidth / 8;
1052 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1053 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1054 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1055 dl, Node->getValueType(0), Chain, Ptr,
1056 LD->getPointerInfo().getWithOffset(IncrementSize),
1057 ExtraVT, isVolatile, isNonTemporal,
1058 MinAlign(Alignment, IncrementSize), AAInfo);
1060 // Build a factor node to remember that this load is independent of
1062 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1065 // Move the top bits to the right place.
1066 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1067 DAG.getConstant(ExtraWidth,
1068 TLI.getShiftAmountTy(Hi.getValueType())));
1070 // Join the hi and lo parts.
1071 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1076 bool isCustom = false;
1077 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1078 default: llvm_unreachable("This action is not supported yet!");
1079 case TargetLowering::Custom:
1082 case TargetLowering::Legal: {
1083 Value = SDValue(Node, 0);
1084 Chain = SDValue(Node, 1);
1087 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1088 if (Res.getNode()) {
1090 Chain = Res.getValue(1);
1093 // If this is an unaligned load and the target doesn't support
1095 EVT MemVT = LD->getMemoryVT();
1096 unsigned AS = LD->getAddressSpace();
1097 unsigned Align = LD->getAlignment();
1098 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1100 LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1101 unsigned ABIAlignment =
1102 TLI.getDataLayout()->getABITypeAlignment(Ty);
1103 if (Align < ABIAlignment){
1104 ExpandUnalignedLoad(cast<LoadSDNode>(Node),
1105 DAG, TLI, Value, Chain);
1111 case TargetLowering::Expand:
1112 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) &&
1113 TLI.isTypeLegal(SrcVT)) {
1114 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr,
1115 LD->getMemOperand());
1119 ExtendOp = (SrcVT.isFloatingPoint() ?
1120 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1122 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1123 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1124 default: llvm_unreachable("Unexpected extend load type!");
1126 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1127 Chain = Load.getValue(1);
1131 assert(!SrcVT.isVector() &&
1132 "Vector Loads are handled in LegalizeVectorOps");
1134 // FIXME: This does not work for vectors on most targets. Sign-
1135 // and zero-extend operations are currently folded into extending
1136 // loads, whether they are legal or not, and then we end up here
1137 // without any support for legalizing them.
1138 assert(ExtType != ISD::EXTLOAD &&
1139 "EXTLOAD should always be supported!");
1140 // Turn the unsupported load into an EXTLOAD followed by an
1141 // explicit zero/sign extend inreg.
1142 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1143 Node->getValueType(0),
1145 LD->getMemOperand());
1147 if (ExtType == ISD::SEXTLOAD)
1148 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1149 Result.getValueType(),
1150 Result, DAG.getValueType(SrcVT));
1152 ValRes = DAG.getZeroExtendInReg(Result, dl,
1153 SrcVT.getScalarType());
1155 Chain = Result.getValue(1);
1160 // Since loads produce two values, make sure to remember that we legalized
1162 if (Chain.getNode() != Node) {
1163 assert(Value.getNode() != Node && "Load must be completely replaced");
1164 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1165 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1167 UpdatedNodes->insert(Value.getNode());
1168 UpdatedNodes->insert(Chain.getNode());
1174 /// LegalizeOp - Return a legal replacement for the given operation, with
1175 /// all legal operands.
1176 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1177 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1179 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1182 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1183 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1184 TargetLowering::TypeLegal &&
1185 "Unexpected illegal type!");
1187 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1188 assert((TLI.getTypeAction(*DAG.getContext(),
1189 Node->getOperand(i).getValueType()) ==
1190 TargetLowering::TypeLegal ||
1191 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1192 "Unexpected illegal type!");
1194 // Figure out the correct action; the way to query this varies by opcode
1195 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1196 bool SimpleFinishLegalizing = true;
1197 switch (Node->getOpcode()) {
1198 case ISD::INTRINSIC_W_CHAIN:
1199 case ISD::INTRINSIC_WO_CHAIN:
1200 case ISD::INTRINSIC_VOID:
1201 case ISD::STACKSAVE:
1202 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1205 Action = TLI.getOperationAction(Node->getOpcode(),
1206 Node->getValueType(0));
1207 if (Action != TargetLowering::Promote)
1208 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1210 case ISD::FP_TO_FP16:
1211 case ISD::SINT_TO_FP:
1212 case ISD::UINT_TO_FP:
1213 case ISD::EXTRACT_VECTOR_ELT:
1214 Action = TLI.getOperationAction(Node->getOpcode(),
1215 Node->getOperand(0).getValueType());
1217 case ISD::FP_ROUND_INREG:
1218 case ISD::SIGN_EXTEND_INREG: {
1219 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1220 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1223 case ISD::ATOMIC_STORE: {
1224 Action = TLI.getOperationAction(Node->getOpcode(),
1225 Node->getOperand(2).getValueType());
1228 case ISD::SELECT_CC:
1231 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1232 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1233 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1234 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1235 ISD::CondCode CCCode =
1236 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1237 Action = TLI.getCondCodeAction(CCCode, OpVT);
1238 if (Action == TargetLowering::Legal) {
1239 if (Node->getOpcode() == ISD::SELECT_CC)
1240 Action = TLI.getOperationAction(Node->getOpcode(),
1241 Node->getValueType(0));
1243 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1249 // FIXME: Model these properly. LOAD and STORE are complicated, and
1250 // STORE expects the unlegalized operand in some cases.
1251 SimpleFinishLegalizing = false;
1253 case ISD::CALLSEQ_START:
1254 case ISD::CALLSEQ_END:
1255 // FIXME: This shouldn't be necessary. These nodes have special properties
1256 // dealing with the recursive nature of legalization. Removing this
1257 // special case should be done as part of making LegalizeDAG non-recursive.
1258 SimpleFinishLegalizing = false;
1260 case ISD::EXTRACT_ELEMENT:
1261 case ISD::FLT_ROUNDS_:
1269 case ISD::MERGE_VALUES:
1270 case ISD::EH_RETURN:
1271 case ISD::FRAME_TO_ARGS_OFFSET:
1272 case ISD::EH_SJLJ_SETJMP:
1273 case ISD::EH_SJLJ_LONGJMP:
1274 // These operations lie about being legal: when they claim to be legal,
1275 // they should actually be expanded.
1276 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1277 if (Action == TargetLowering::Legal)
1278 Action = TargetLowering::Expand;
1280 case ISD::INIT_TRAMPOLINE:
1281 case ISD::ADJUST_TRAMPOLINE:
1282 case ISD::FRAMEADDR:
1283 case ISD::RETURNADDR:
1284 // These operations lie about being legal: when they claim to be legal,
1285 // they should actually be custom-lowered.
1286 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1287 if (Action == TargetLowering::Legal)
1288 Action = TargetLowering::Custom;
1290 case ISD::READ_REGISTER:
1291 case ISD::WRITE_REGISTER:
1292 // Named register is legal in the DAG, but blocked by register name
1293 // selection if not implemented by target (to chose the correct register)
1294 // They'll be converted to Copy(To/From)Reg.
1295 Action = TargetLowering::Legal;
1297 case ISD::DEBUGTRAP:
1298 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1299 if (Action == TargetLowering::Expand) {
1300 // replace ISD::DEBUGTRAP with ISD::TRAP
1302 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1303 Node->getOperand(0));
1304 ReplaceNode(Node, NewVal.getNode());
1305 LegalizeOp(NewVal.getNode());
1311 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1312 Action = TargetLowering::Legal;
1314 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1319 if (SimpleFinishLegalizing) {
1320 SDNode *NewNode = Node;
1321 switch (Node->getOpcode()) {
1328 // Legalizing shifts/rotates requires adjusting the shift amount
1329 // to the appropriate width.
1330 if (!Node->getOperand(1).getValueType().isVector()) {
1332 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1333 Node->getOperand(1));
1334 HandleSDNode Handle(SAO);
1335 LegalizeOp(SAO.getNode());
1336 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1340 case ISD::SRL_PARTS:
1341 case ISD::SRA_PARTS:
1342 case ISD::SHL_PARTS:
1343 // Legalizing shifts/rotates requires adjusting the shift amount
1344 // to the appropriate width.
1345 if (!Node->getOperand(2).getValueType().isVector()) {
1347 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1348 Node->getOperand(2));
1349 HandleSDNode Handle(SAO);
1350 LegalizeOp(SAO.getNode());
1351 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1352 Node->getOperand(1),
1358 if (NewNode != Node) {
1359 ReplaceNode(Node, NewNode);
1363 case TargetLowering::Legal:
1365 case TargetLowering::Custom: {
1366 // FIXME: The handling for custom lowering with multiple results is
1368 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1369 if (Res.getNode()) {
1370 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1373 if (Node->getNumValues() == 1) {
1374 // We can just directly replace this node with the lowered value.
1375 ReplaceNode(SDValue(Node, 0), Res);
1379 SmallVector<SDValue, 8> ResultVals;
1380 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1381 ResultVals.push_back(Res.getValue(i));
1382 ReplaceNode(Node, ResultVals.data());
1387 case TargetLowering::Expand:
1390 case TargetLowering::Promote:
1396 switch (Node->getOpcode()) {
1403 llvm_unreachable("Do not know how to legalize this operator!");
1405 case ISD::CALLSEQ_START:
1406 case ISD::CALLSEQ_END:
1409 return LegalizeLoadOps(Node);
1412 return LegalizeStoreOps(Node);
1417 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1418 SDValue Vec = Op.getOperand(0);
1419 SDValue Idx = Op.getOperand(1);
1422 // Before we generate a new store to a temporary stack slot, see if there is
1423 // already one that we can use. There often is because when we scalarize
1424 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1425 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1426 // the vector. If all are expanded here, we don't want one store per vector
1428 SDValue StackPtr, Ch;
1429 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1430 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1432 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1433 if (ST->isIndexed() || ST->isTruncatingStore() ||
1434 ST->getValue() != Vec)
1437 // Make sure that nothing else could have stored into the destination of
1439 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1442 StackPtr = ST->getBasePtr();
1443 Ch = SDValue(ST, 0);
1448 if (!Ch.getNode()) {
1449 // Store the value to a temporary stack slot, then LOAD the returned part.
1450 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1451 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1452 MachinePointerInfo(), false, false, 0);
1455 // Add the offset to the index.
1457 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1458 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1459 DAG.getConstant(EltSize, Idx.getValueType()));
1461 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1462 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1464 if (Op.getValueType().isVector())
1465 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1466 false, false, false, 0);
1467 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1468 MachinePointerInfo(),
1469 Vec.getValueType().getVectorElementType(),
1473 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1474 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1476 SDValue Vec = Op.getOperand(0);
1477 SDValue Part = Op.getOperand(1);
1478 SDValue Idx = Op.getOperand(2);
1481 // Store the value to a temporary stack slot, then LOAD the returned part.
1483 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1484 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1485 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1487 // First store the whole vector.
1488 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1491 // Then store the inserted part.
1493 // Add the offset to the index.
1495 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1497 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1498 DAG.getConstant(EltSize, Idx.getValueType()));
1499 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1501 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1504 // Store the subvector.
1505 Ch = DAG.getStore(DAG.getEntryNode(), dl, Part, SubStackPtr,
1506 MachinePointerInfo(), false, false, 0);
1508 // Finally, load the updated vector.
1509 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1510 false, false, false, 0);
1513 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1514 // We can't handle this case efficiently. Allocate a sufficiently
1515 // aligned object on the stack, store each element into it, then load
1516 // the result as a vector.
1517 // Create the stack frame object.
1518 EVT VT = Node->getValueType(0);
1519 EVT EltVT = VT.getVectorElementType();
1521 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1522 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1523 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1525 // Emit a store of each element to the stack slot.
1526 SmallVector<SDValue, 8> Stores;
1527 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1528 // Store (in the right endianness) the elements to memory.
1529 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1530 // Ignore undef elements.
1531 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1533 unsigned Offset = TypeByteSize*i;
1535 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1536 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1538 // If the destination vector element type is narrower than the source
1539 // element type, only store the bits necessary.
1540 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1541 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1542 Node->getOperand(i), Idx,
1543 PtrInfo.getWithOffset(Offset),
1544 EltVT, false, false, 0));
1546 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1547 Node->getOperand(i), Idx,
1548 PtrInfo.getWithOffset(Offset),
1553 if (!Stores.empty()) // Not all undef elements?
1554 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1556 StoreChain = DAG.getEntryNode();
1558 // Result is a load from the stack slot.
1559 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1560 false, false, false, 0);
1563 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1565 SDValue Tmp1 = Node->getOperand(0);
1566 SDValue Tmp2 = Node->getOperand(1);
1568 // Get the sign bit of the RHS. First obtain a value that has the same
1569 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1571 EVT FloatVT = Tmp2.getValueType();
1572 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1573 if (TLI.isTypeLegal(IVT)) {
1574 // Convert to an integer with the same sign bit.
1575 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1577 // Store the float to memory, then load the sign part out as an integer.
1578 MVT LoadTy = TLI.getPointerTy();
1579 // First create a temporary that is aligned for both the load and store.
1580 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1581 // Then store the float to it.
1583 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1585 if (TLI.isBigEndian()) {
1586 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1587 // Load out a legal integer with the same sign bit as the float.
1588 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1589 false, false, false, 0);
1590 } else { // Little endian
1591 SDValue LoadPtr = StackPtr;
1592 // The float may be wider than the integer we are going to load. Advance
1593 // the pointer so that the loaded integer will contain the sign bit.
1594 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1595 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1596 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1597 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1598 // Load a legal integer containing the sign bit.
1599 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1600 false, false, false, 0);
1601 // Move the sign bit to the top bit of the loaded integer.
1602 unsigned BitShift = LoadTy.getSizeInBits() -
1603 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1604 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1606 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1607 DAG.getConstant(BitShift,
1608 TLI.getShiftAmountTy(SignBit.getValueType())));
1611 // Now get the sign bit proper, by seeing whether the value is negative.
1612 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1613 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1615 // Get the absolute value of the result.
1616 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1617 // Select between the nabs and abs value based on the sign bit of
1619 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1620 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1624 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1625 SmallVectorImpl<SDValue> &Results) {
1626 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1627 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1628 " not tell us which reg is the stack pointer!");
1630 EVT VT = Node->getValueType(0);
1631 SDValue Tmp1 = SDValue(Node, 0);
1632 SDValue Tmp2 = SDValue(Node, 1);
1633 SDValue Tmp3 = Node->getOperand(2);
1634 SDValue Chain = Tmp1.getOperand(0);
1636 // Chain the dynamic stack allocation so that it doesn't modify the stack
1637 // pointer when other instructions are using the stack.
1638 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1641 SDValue Size = Tmp2.getOperand(1);
1642 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1643 Chain = SP.getValue(1);
1644 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1645 unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
1646 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1647 if (Align > StackAlign)
1648 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1649 DAG.getConstant(-(uint64_t)Align, VT));
1650 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1652 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1653 DAG.getIntPtrConstant(0, true), SDValue(),
1656 Results.push_back(Tmp1);
1657 Results.push_back(Tmp2);
1660 /// LegalizeSetCCCondCode - Legalize a SETCC with given LHS and RHS and
1661 /// condition code CC on the current target.
1663 /// If the SETCC has been legalized using AND / OR, then the legalized node
1664 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1665 /// will be set to false.
1667 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1668 /// then the values of LHS and RHS will be swapped, CC will be set to the
1669 /// new condition, and NeedInvert will be set to false.
1671 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1672 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1673 /// will be set to true. The caller must invert the result of the SETCC with
1674 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1675 /// of a true/false result.
1677 /// \returns true if the SetCC has been legalized, false if it hasn't.
1678 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1679 SDValue &LHS, SDValue &RHS,
1683 MVT OpVT = LHS.getSimpleValueType();
1684 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1686 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1687 default: llvm_unreachable("Unknown condition code action!");
1688 case TargetLowering::Legal:
1691 case TargetLowering::Expand: {
1692 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1693 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1694 std::swap(LHS, RHS);
1695 CC = DAG.getCondCode(InvCC);
1698 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1701 default: llvm_unreachable("Don't know how to expand this condition!");
1703 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1704 == TargetLowering::Legal
1705 && "If SETO is expanded, SETOEQ must be legal!");
1706 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1708 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1709 == TargetLowering::Legal
1710 && "If SETUO is expanded, SETUNE must be legal!");
1711 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1724 // If we are floating point, assign and break, otherwise fall through.
1725 if (!OpVT.isInteger()) {
1726 // We can use the 4th bit to tell if we are the unordered
1727 // or ordered version of the opcode.
1728 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1729 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1730 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1733 // Fallthrough if we are unsigned integer.
1738 // We only support using the inverted operation, which is computed above
1739 // and not a different manner of supporting expanding these cases.
1740 llvm_unreachable("Don't know how to expand this condition!");
1743 // Try inverting the result of the inverse condition.
1744 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1745 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1746 CC = DAG.getCondCode(InvCC);
1750 // If inverting the condition didn't work then we have no means to expand
1752 llvm_unreachable("Don't know how to expand this condition!");
1755 SDValue SetCC1, SetCC2;
1756 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1757 // If we aren't the ordered or unorder operation,
1758 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1759 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1760 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1762 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1763 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1764 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1766 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1775 /// EmitStackConvert - Emit a store/load combination to the stack. This stores
1776 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1777 /// a load from the stack slot to DestVT, extending it if needed.
1778 /// The resultant code need not be legal.
1779 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1783 // Create the stack frame object.
1785 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1786 getTypeForEVT(*DAG.getContext()));
1787 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1789 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1790 int SPFI = StackPtrFI->getIndex();
1791 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1793 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1794 unsigned SlotSize = SlotVT.getSizeInBits();
1795 unsigned DestSize = DestVT.getSizeInBits();
1796 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1797 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1799 // Emit a store to the stack slot. Use a truncstore if the input value is
1800 // later than DestVT.
1803 if (SrcSize > SlotSize)
1804 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1805 PtrInfo, SlotVT, false, false, SrcAlign);
1807 assert(SrcSize == SlotSize && "Invalid store");
1808 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1809 PtrInfo, false, false, SrcAlign);
1812 // Result is a load from the stack slot.
1813 if (SlotSize == DestSize)
1814 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1815 false, false, false, DestAlign);
1817 assert(SlotSize < DestSize && "Unknown extension!");
1818 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1819 PtrInfo, SlotVT, false, false, DestAlign);
1822 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1824 // Create a vector sized/aligned stack slot, store the value to element #0,
1825 // then load the whole vector back out.
1826 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1828 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1829 int SPFI = StackPtrFI->getIndex();
1831 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1833 MachinePointerInfo::getFixedStack(SPFI),
1834 Node->getValueType(0).getVectorElementType(),
1836 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1837 MachinePointerInfo::getFixedStack(SPFI),
1838 false, false, false, 0);
1842 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1843 const TargetLowering &TLI, SDValue &Res) {
1844 unsigned NumElems = Node->getNumOperands();
1846 EVT VT = Node->getValueType(0);
1848 // Try to group the scalars into pairs, shuffle the pairs together, then
1849 // shuffle the pairs of pairs together, etc. until the vector has
1850 // been built. This will work only if all of the necessary shuffle masks
1853 // We do this in two phases; first to check the legality of the shuffles,
1854 // and next, assuming that all shuffles are legal, to create the new nodes.
1855 for (int Phase = 0; Phase < 2; ++Phase) {
1856 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1858 for (unsigned i = 0; i < NumElems; ++i) {
1859 SDValue V = Node->getOperand(i);
1860 if (V.getOpcode() == ISD::UNDEF)
1865 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1866 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1869 while (IntermedVals.size() > 2) {
1870 NewIntermedVals.clear();
1871 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1872 // This vector and the next vector are shuffled together (simply to
1873 // append the one to the other).
1874 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1876 SmallVector<int, 16> FinalIndices;
1877 FinalIndices.reserve(IntermedVals[i].second.size() +
1878 IntermedVals[i+1].second.size());
1881 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1884 FinalIndices.push_back(IntermedVals[i].second[j]);
1886 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1888 ShuffleVec[k] = NumElems + j;
1889 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1894 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1895 IntermedVals[i+1].first,
1897 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1899 NewIntermedVals.push_back(std::make_pair(Shuffle, FinalIndices));
1902 // If we had an odd number of defined values, then append the last
1903 // element to the array of new vectors.
1904 if ((IntermedVals.size() & 1) != 0)
1905 NewIntermedVals.push_back(IntermedVals.back());
1907 IntermedVals.swap(NewIntermedVals);
1910 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1911 "Invalid number of intermediate vectors");
1912 SDValue Vec1 = IntermedVals[0].first;
1914 if (IntermedVals.size() > 1)
1915 Vec2 = IntermedVals[1].first;
1917 Vec2 = DAG.getUNDEF(VT);
1919 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1920 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1921 ShuffleVec[IntermedVals[0].second[i]] = i;
1922 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1923 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1926 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1927 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1934 /// ExpandBUILD_VECTOR - Expand a BUILD_VECTOR node on targets that don't
1935 /// support the operation, but do support the resultant vector type.
1936 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1937 unsigned NumElems = Node->getNumOperands();
1938 SDValue Value1, Value2;
1940 EVT VT = Node->getValueType(0);
1941 EVT OpVT = Node->getOperand(0).getValueType();
1942 EVT EltVT = VT.getVectorElementType();
1944 // If the only non-undef value is the low element, turn this into a
1945 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1946 bool isOnlyLowElement = true;
1947 bool MoreThanTwoValues = false;
1948 bool isConstant = true;
1949 for (unsigned i = 0; i < NumElems; ++i) {
1950 SDValue V = Node->getOperand(i);
1951 if (V.getOpcode() == ISD::UNDEF)
1954 isOnlyLowElement = false;
1955 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1958 if (!Value1.getNode()) {
1960 } else if (!Value2.getNode()) {
1963 } else if (V != Value1 && V != Value2) {
1964 MoreThanTwoValues = true;
1968 if (!Value1.getNode())
1969 return DAG.getUNDEF(VT);
1971 if (isOnlyLowElement)
1972 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1974 // If all elements are constants, create a load from the constant pool.
1976 SmallVector<Constant*, 16> CV;
1977 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1978 if (ConstantFPSDNode *V =
1979 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1980 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1981 } else if (ConstantSDNode *V =
1982 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1984 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1986 // If OpVT and EltVT don't match, EltVT is not legal and the
1987 // element values have been promoted/truncated earlier. Undo this;
1988 // we don't want a v16i8 to become a v16i32 for example.
1989 const ConstantInt *CI = V->getConstantIntValue();
1990 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1991 CI->getZExtValue()));
1994 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1995 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1996 CV.push_back(UndefValue::get(OpNTy));
1999 Constant *CP = ConstantVector::get(CV);
2000 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
2001 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2002 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
2003 MachinePointerInfo::getConstantPool(),
2004 false, false, false, Alignment);
2007 SmallSet<SDValue, 16> DefinedValues;
2008 for (unsigned i = 0; i < NumElems; ++i) {
2009 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
2011 DefinedValues.insert(Node->getOperand(i));
2014 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
2015 if (!MoreThanTwoValues) {
2016 SmallVector<int, 8> ShuffleVec(NumElems, -1);
2017 for (unsigned i = 0; i < NumElems; ++i) {
2018 SDValue V = Node->getOperand(i);
2019 if (V.getOpcode() == ISD::UNDEF)
2021 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2023 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2024 // Get the splatted value into the low element of a vector register.
2025 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2027 if (Value2.getNode())
2028 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2030 Vec2 = DAG.getUNDEF(VT);
2032 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2033 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2037 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2042 // Otherwise, we can't handle this case efficiently.
2043 return ExpandVectorBuildThroughStack(Node);
2046 // ExpandLibCall - Expand a node into a call to a libcall. If the result value
2047 // does not fit into a register, return the lo part and set the hi part to the
2048 // by-reg argument. If it does fit into a single register, return the result
2049 // and leave the Hi part unset.
2050 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2052 TargetLowering::ArgListTy Args;
2053 TargetLowering::ArgListEntry Entry;
2054 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2055 EVT ArgVT = Node->getOperand(i).getValueType();
2056 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2057 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2058 Entry.isSExt = isSigned;
2059 Entry.isZExt = !isSigned;
2060 Args.push_back(Entry);
2062 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2063 TLI.getPointerTy());
2065 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2067 // By default, the input chain to this libcall is the entry node of the
2068 // function. If the libcall is going to be emitted as a tail call then
2069 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2070 // node which is being folded has a non-entry input chain.
2071 SDValue InChain = DAG.getEntryNode();
2073 // isTailCall may be true since the callee does not reference caller stack
2074 // frame. Check if it's in the right position.
2075 SDValue TCChain = InChain;
2076 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2080 TargetLowering::CallLoweringInfo CLI(DAG);
2081 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2082 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2083 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2085 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2087 if (!CallInfo.second.getNode())
2088 // It's a tailcall, return the chain (which is the DAG root).
2089 return DAG.getRoot();
2091 return CallInfo.first;
2094 /// ExpandLibCall - Generate a libcall taking the given operands as arguments
2095 /// and returning a result of type RetVT.
2096 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2097 const SDValue *Ops, unsigned NumOps,
2098 bool isSigned, SDLoc dl) {
2099 TargetLowering::ArgListTy Args;
2100 Args.reserve(NumOps);
2102 TargetLowering::ArgListEntry Entry;
2103 for (unsigned i = 0; i != NumOps; ++i) {
2104 Entry.Node = Ops[i];
2105 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2106 Entry.isSExt = isSigned;
2107 Entry.isZExt = !isSigned;
2108 Args.push_back(Entry);
2110 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2111 TLI.getPointerTy());
2113 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2115 TargetLowering::CallLoweringInfo CLI(DAG);
2116 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2117 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2118 .setSExtResult(isSigned).setZExtResult(!isSigned);
2120 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2122 return CallInfo.first;
2125 // ExpandChainLibCall - Expand a node into a call to a libcall. Similar to
2126 // ExpandLibCall except that the first operand is the in-chain.
2127 std::pair<SDValue, SDValue>
2128 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2131 SDValue InChain = Node->getOperand(0);
2133 TargetLowering::ArgListTy Args;
2134 TargetLowering::ArgListEntry Entry;
2135 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2136 EVT ArgVT = Node->getOperand(i).getValueType();
2137 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2138 Entry.Node = Node->getOperand(i);
2140 Entry.isSExt = isSigned;
2141 Entry.isZExt = !isSigned;
2142 Args.push_back(Entry);
2144 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2145 TLI.getPointerTy());
2147 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2149 TargetLowering::CallLoweringInfo CLI(DAG);
2150 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2151 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2152 .setSExtResult(isSigned).setZExtResult(!isSigned);
2154 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2159 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2160 RTLIB::Libcall Call_F32,
2161 RTLIB::Libcall Call_F64,
2162 RTLIB::Libcall Call_F80,
2163 RTLIB::Libcall Call_F128,
2164 RTLIB::Libcall Call_PPCF128) {
2166 switch (Node->getSimpleValueType(0).SimpleTy) {
2167 default: llvm_unreachable("Unexpected request for libcall!");
2168 case MVT::f32: LC = Call_F32; break;
2169 case MVT::f64: LC = Call_F64; break;
2170 case MVT::f80: LC = Call_F80; break;
2171 case MVT::f128: LC = Call_F128; break;
2172 case MVT::ppcf128: LC = Call_PPCF128; break;
2174 return ExpandLibCall(LC, Node, false);
2177 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2178 RTLIB::Libcall Call_I8,
2179 RTLIB::Libcall Call_I16,
2180 RTLIB::Libcall Call_I32,
2181 RTLIB::Libcall Call_I64,
2182 RTLIB::Libcall Call_I128) {
2184 switch (Node->getSimpleValueType(0).SimpleTy) {
2185 default: llvm_unreachable("Unexpected request for libcall!");
2186 case MVT::i8: LC = Call_I8; break;
2187 case MVT::i16: LC = Call_I16; break;
2188 case MVT::i32: LC = Call_I32; break;
2189 case MVT::i64: LC = Call_I64; break;
2190 case MVT::i128: LC = Call_I128; break;
2192 return ExpandLibCall(LC, Node, isSigned);
2195 /// isDivRemLibcallAvailable - Return true if divmod libcall is available.
2196 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2197 const TargetLowering &TLI) {
2199 switch (Node->getSimpleValueType(0).SimpleTy) {
2200 default: llvm_unreachable("Unexpected request for libcall!");
2201 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2202 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2203 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2204 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2205 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2208 return TLI.getLibcallName(LC) != nullptr;
2211 /// useDivRem - Only issue divrem libcall if both quotient and remainder are
2213 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2214 // The other use might have been replaced with a divrem already.
2215 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2216 unsigned OtherOpcode = 0;
2218 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2220 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2222 SDValue Op0 = Node->getOperand(0);
2223 SDValue Op1 = Node->getOperand(1);
2224 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2225 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2229 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2230 User->getOperand(0) == Op0 &&
2231 User->getOperand(1) == Op1)
2237 /// ExpandDivRemLibCall - Issue libcalls to __{u}divmod to compute div / rem
2240 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2241 SmallVectorImpl<SDValue> &Results) {
2242 unsigned Opcode = Node->getOpcode();
2243 bool isSigned = Opcode == ISD::SDIVREM;
2246 switch (Node->getSimpleValueType(0).SimpleTy) {
2247 default: llvm_unreachable("Unexpected request for libcall!");
2248 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2249 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2250 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2251 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2252 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2255 // The input chain to this libcall is the entry node of the function.
2256 // Legalizing the call will automatically add the previous call to the
2258 SDValue InChain = DAG.getEntryNode();
2260 EVT RetVT = Node->getValueType(0);
2261 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2263 TargetLowering::ArgListTy Args;
2264 TargetLowering::ArgListEntry Entry;
2265 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2266 EVT ArgVT = Node->getOperand(i).getValueType();
2267 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2268 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2269 Entry.isSExt = isSigned;
2270 Entry.isZExt = !isSigned;
2271 Args.push_back(Entry);
2274 // Also pass the return address of the remainder.
2275 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2277 Entry.Ty = RetTy->getPointerTo();
2278 Entry.isSExt = isSigned;
2279 Entry.isZExt = !isSigned;
2280 Args.push_back(Entry);
2282 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2283 TLI.getPointerTy());
2286 TargetLowering::CallLoweringInfo CLI(DAG);
2287 CLI.setDebugLoc(dl).setChain(InChain)
2288 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2289 .setSExtResult(isSigned).setZExtResult(!isSigned);
2291 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2293 // Remainder is loaded back from the stack frame.
2294 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2295 MachinePointerInfo(), false, false, false, 0);
2296 Results.push_back(CallInfo.first);
2297 Results.push_back(Rem);
2300 /// isSinCosLibcallAvailable - Return true if sincos libcall is available.
2301 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2303 switch (Node->getSimpleValueType(0).SimpleTy) {
2304 default: llvm_unreachable("Unexpected request for libcall!");
2305 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2306 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2307 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2308 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2309 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2311 return TLI.getLibcallName(LC) != nullptr;
2314 /// canCombineSinCosLibcall - Return true if sincos libcall is available and
2315 /// can be used to combine sin and cos.
2316 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2317 const TargetMachine &TM) {
2318 if (!isSinCosLibcallAvailable(Node, TLI))
2320 // GNU sin/cos functions set errno while sincos does not. Therefore
2321 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2322 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2323 if (isGNU && !TM.Options.UnsafeFPMath)
2328 /// useSinCos - Only issue sincos libcall if both sin and cos are
2330 static bool useSinCos(SDNode *Node) {
2331 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2332 ? ISD::FCOS : ISD::FSIN;
2334 SDValue Op0 = Node->getOperand(0);
2335 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2336 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2340 // The other user might have been turned into sincos already.
2341 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2347 /// ExpandSinCosLibCall - Issue libcalls to sincos to compute sin / cos
2350 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2351 SmallVectorImpl<SDValue> &Results) {
2353 switch (Node->getSimpleValueType(0).SimpleTy) {
2354 default: llvm_unreachable("Unexpected request for libcall!");
2355 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2356 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2357 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2358 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2359 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2362 // The input chain to this libcall is the entry node of the function.
2363 // Legalizing the call will automatically add the previous call to the
2365 SDValue InChain = DAG.getEntryNode();
2367 EVT RetVT = Node->getValueType(0);
2368 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2370 TargetLowering::ArgListTy Args;
2371 TargetLowering::ArgListEntry Entry;
2373 // Pass the argument.
2374 Entry.Node = Node->getOperand(0);
2376 Entry.isSExt = false;
2377 Entry.isZExt = false;
2378 Args.push_back(Entry);
2380 // Pass the return address of sin.
2381 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2382 Entry.Node = SinPtr;
2383 Entry.Ty = RetTy->getPointerTo();
2384 Entry.isSExt = false;
2385 Entry.isZExt = false;
2386 Args.push_back(Entry);
2388 // Also pass the return address of the cos.
2389 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2390 Entry.Node = CosPtr;
2391 Entry.Ty = RetTy->getPointerTo();
2392 Entry.isSExt = false;
2393 Entry.isZExt = false;
2394 Args.push_back(Entry);
2396 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2397 TLI.getPointerTy());
2400 TargetLowering::CallLoweringInfo CLI(DAG);
2401 CLI.setDebugLoc(dl).setChain(InChain)
2402 .setCallee(TLI.getLibcallCallingConv(LC),
2403 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2405 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2407 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2408 MachinePointerInfo(), false, false, false, 0));
2409 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2410 MachinePointerInfo(), false, false, false, 0));
2413 /// ExpandLegalINT_TO_FP - This function is responsible for legalizing a
2414 /// INT_TO_FP operation of the specified operand when the target requests that
2415 /// we expand it. At this point, we know that the result and operand types are
2416 /// legal for the target.
2417 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2421 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2422 // simple 32-bit [signed|unsigned] integer to float/double expansion
2424 // Get the stack frame index of a 8 byte buffer.
2425 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2427 // word offset constant for Hi/Lo address computation
2428 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2429 // set up Hi and Lo (into buffer) address based on endian
2430 SDValue Hi = StackSlot;
2431 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2432 StackSlot, WordOff);
2433 if (TLI.isLittleEndian())
2436 // if signed map to unsigned space
2439 // constant used to invert sign bit (signed to unsigned mapping)
2440 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2441 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2445 // store the lo of the constructed double - based on integer input
2446 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2447 Op0Mapped, Lo, MachinePointerInfo(),
2449 // initial hi portion of constructed double
2450 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2451 // store the hi of the constructed double - biased exponent
2452 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2453 MachinePointerInfo(),
2455 // load the constructed double
2456 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2457 MachinePointerInfo(), false, false, false, 0);
2458 // FP constant to bias correct the final result
2459 SDValue Bias = DAG.getConstantFP(isSigned ?
2460 BitsToDouble(0x4330000080000000ULL) :
2461 BitsToDouble(0x4330000000000000ULL),
2463 // subtract the bias
2464 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2467 // handle final rounding
2468 if (DestVT == MVT::f64) {
2471 } else if (DestVT.bitsLT(MVT::f64)) {
2472 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2473 DAG.getIntPtrConstant(0));
2474 } else if (DestVT.bitsGT(MVT::f64)) {
2475 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2479 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2480 // Code below here assumes !isSigned without checking again.
2482 // Implementation of unsigned i64 to f64 following the algorithm in
2483 // __floatundidf in compiler_rt. This implementation has the advantage
2484 // of performing rounding correctly, both in the default rounding mode
2485 // and in all alternate rounding modes.
2486 // TODO: Generalize this for use with other types.
2487 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2489 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2490 SDValue TwoP84PlusTwoP52 =
2491 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2493 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2495 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2496 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2497 DAG.getConstant(32, MVT::i64));
2498 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2499 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2500 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2501 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2502 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2504 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2507 // Implementation of unsigned i64 to f32.
2508 // TODO: Generalize this for use with other types.
2509 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2510 // For unsigned conversions, convert them to signed conversions using the
2511 // algorithm from the x86_64 __floatundidf in compiler_rt.
2513 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2515 SDValue ShiftConst =
2516 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2517 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2518 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2519 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2520 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2522 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2523 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2525 // TODO: This really should be implemented using a branch rather than a
2526 // select. We happen to get lucky and machinesink does the right
2527 // thing most of the time. This would be a good candidate for a
2528 //pseudo-op, or, even better, for whole-function isel.
2529 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2530 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2531 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2534 // Otherwise, implement the fully general conversion.
2536 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2537 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2538 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2539 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2540 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2541 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2542 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2543 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2544 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2545 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2546 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2548 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2549 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2551 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2552 DAG.getConstant(32, SHVT));
2553 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2554 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2556 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2557 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2558 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2559 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2560 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2561 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2562 DAG.getIntPtrConstant(0));
2565 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2567 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2568 Op0, DAG.getConstant(0, Op0.getValueType()),
2570 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2571 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2572 SignSet, Four, Zero);
2574 // If the sign bit of the integer is set, the large number will be treated
2575 // as a negative number. To counteract this, the dynamic code adds an
2576 // offset depending on the data type.
2578 switch (Op0.getSimpleValueType().SimpleTy) {
2579 default: llvm_unreachable("Unsupported integer type!");
2580 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2581 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2582 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2583 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2585 if (TLI.isLittleEndian()) FF <<= 32;
2586 Constant *FudgeFactor = ConstantInt::get(
2587 Type::getInt64Ty(*DAG.getContext()), FF);
2589 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2590 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2591 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2592 Alignment = std::min(Alignment, 4u);
2594 if (DestVT == MVT::f32)
2595 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2596 MachinePointerInfo::getConstantPool(),
2597 false, false, false, Alignment);
2599 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2600 DAG.getEntryNode(), CPIdx,
2601 MachinePointerInfo::getConstantPool(),
2602 MVT::f32, false, false, Alignment);
2603 HandleSDNode Handle(Load);
2604 LegalizeOp(Load.getNode());
2605 FudgeInReg = Handle.getValue();
2608 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2611 /// PromoteLegalINT_TO_FP - This function is responsible for legalizing a
2612 /// *INT_TO_FP operation of the specified operand when the target requests that
2613 /// we promote it. At this point, we know that the result and operand types are
2614 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2615 /// operation that takes a larger input.
2616 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2620 // First step, figure out the appropriate *INT_TO_FP operation to use.
2621 EVT NewInTy = LegalOp.getValueType();
2623 unsigned OpToUse = 0;
2625 // Scan for the appropriate larger type to use.
2627 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2628 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2630 // If the target supports SINT_TO_FP of this type, use it.
2631 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2632 OpToUse = ISD::SINT_TO_FP;
2635 if (isSigned) continue;
2637 // If the target supports UINT_TO_FP of this type, use it.
2638 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2639 OpToUse = ISD::UINT_TO_FP;
2643 // Otherwise, try a larger type.
2646 // Okay, we found the operation and type to use. Zero extend our input to the
2647 // desired type then run the operation on it.
2648 return DAG.getNode(OpToUse, dl, DestVT,
2649 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2650 dl, NewInTy, LegalOp));
2653 /// PromoteLegalFP_TO_INT - This function is responsible for legalizing a
2654 /// FP_TO_*INT operation of the specified operand when the target requests that
2655 /// we promote it. At this point, we know that the result and operand types are
2656 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2657 /// operation that returns a larger result.
2658 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2662 // First step, figure out the appropriate FP_TO*INT operation to use.
2663 EVT NewOutTy = DestVT;
2665 unsigned OpToUse = 0;
2667 // Scan for the appropriate larger type to use.
2669 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2670 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2672 // A larger signed type can hold all unsigned values of the requested type,
2673 // so using FP_TO_SINT is valid
2674 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2675 OpToUse = ISD::FP_TO_SINT;
2679 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2680 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2681 OpToUse = ISD::FP_TO_UINT;
2685 // Otherwise, try a larger type.
2689 // Okay, we found the operation and type to use.
2690 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2692 // Truncate the result of the extended FP_TO_*INT operation to the desired
2694 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2697 /// ExpandBSWAP - Open code the operations for BSWAP of the specified operation.
2699 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2700 EVT VT = Op.getValueType();
2701 EVT SHVT = TLI.getShiftAmountTy(VT);
2702 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2703 switch (VT.getSimpleVT().SimpleTy) {
2704 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2706 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2707 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2708 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2710 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2711 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2712 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2713 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2714 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2715 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2716 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2717 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2718 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2720 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2721 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2722 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2723 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2724 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2725 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2726 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2727 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2728 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2729 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2730 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2731 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2732 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2733 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2734 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2735 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2736 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2737 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2738 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2739 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2740 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2744 /// ExpandBitCount - Expand the specified bitcount instruction into operations.
2746 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2749 default: llvm_unreachable("Cannot expand this yet!");
2751 EVT VT = Op.getValueType();
2752 EVT ShVT = TLI.getShiftAmountTy(VT);
2753 unsigned Len = VT.getSizeInBits();
2755 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2756 "CTPOP not implemented for this type.");
2758 // This is the "best" algorithm from
2759 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2761 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2762 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2763 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2764 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2766 // v = v - ((v >> 1) & 0x55555555...)
2767 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2768 DAG.getNode(ISD::AND, dl, VT,
2769 DAG.getNode(ISD::SRL, dl, VT, Op,
2770 DAG.getConstant(1, ShVT)),
2772 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2773 Op = DAG.getNode(ISD::ADD, dl, VT,
2774 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2775 DAG.getNode(ISD::AND, dl, VT,
2776 DAG.getNode(ISD::SRL, dl, VT, Op,
2777 DAG.getConstant(2, ShVT)),
2779 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2780 Op = DAG.getNode(ISD::AND, dl, VT,
2781 DAG.getNode(ISD::ADD, dl, VT, Op,
2782 DAG.getNode(ISD::SRL, dl, VT, Op,
2783 DAG.getConstant(4, ShVT))),
2785 // v = (v * 0x01010101...) >> (Len - 8)
2786 Op = DAG.getNode(ISD::SRL, dl, VT,
2787 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2788 DAG.getConstant(Len - 8, ShVT));
2792 case ISD::CTLZ_ZERO_UNDEF:
2793 // This trivially expands to CTLZ.
2794 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2796 // for now, we do this:
2797 // x = x | (x >> 1);
2798 // x = x | (x >> 2);
2800 // x = x | (x >>16);
2801 // x = x | (x >>32); // for 64-bit input
2802 // return popcount(~x);
2804 // but see also: http://www.hackersdelight.org/HDcode/nlz.cc
2805 EVT VT = Op.getValueType();
2806 EVT ShVT = TLI.getShiftAmountTy(VT);
2807 unsigned len = VT.getSizeInBits();
2808 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2809 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2810 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2811 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2813 Op = DAG.getNOT(dl, Op, VT);
2814 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2816 case ISD::CTTZ_ZERO_UNDEF:
2817 // This trivially expands to CTTZ.
2818 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2820 // for now, we use: { return popcount(~x & (x - 1)); }
2821 // unless the target has ctlz but not ctpop, in which case we use:
2822 // { return 32 - nlz(~x & (x-1)); }
2823 // see also http://www.hackersdelight.org/HDcode/ntz.cc
2824 EVT VT = Op.getValueType();
2825 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2826 DAG.getNOT(dl, Op, VT),
2827 DAG.getNode(ISD::SUB, dl, VT, Op,
2828 DAG.getConstant(1, VT)));
2829 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2830 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2831 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2832 return DAG.getNode(ISD::SUB, dl, VT,
2833 DAG.getConstant(VT.getSizeInBits(), VT),
2834 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2835 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2840 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2841 unsigned Opc = Node->getOpcode();
2842 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2847 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2848 case ISD::ATOMIC_SWAP:
2849 switch (VT.SimpleTy) {
2850 default: llvm_unreachable("Unexpected value type for atomic!");
2851 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2852 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2853 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2854 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2855 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2858 case ISD::ATOMIC_CMP_SWAP:
2859 switch (VT.SimpleTy) {
2860 default: llvm_unreachable("Unexpected value type for atomic!");
2861 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2862 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2863 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2864 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2865 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2868 case ISD::ATOMIC_LOAD_ADD:
2869 switch (VT.SimpleTy) {
2870 default: llvm_unreachable("Unexpected value type for atomic!");
2871 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2872 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2873 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2874 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2875 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2878 case ISD::ATOMIC_LOAD_SUB:
2879 switch (VT.SimpleTy) {
2880 default: llvm_unreachable("Unexpected value type for atomic!");
2881 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2882 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2883 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2884 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2885 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2888 case ISD::ATOMIC_LOAD_AND:
2889 switch (VT.SimpleTy) {
2890 default: llvm_unreachable("Unexpected value type for atomic!");
2891 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2892 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2893 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2894 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2895 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2898 case ISD::ATOMIC_LOAD_OR:
2899 switch (VT.SimpleTy) {
2900 default: llvm_unreachable("Unexpected value type for atomic!");
2901 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2902 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2903 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2904 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2905 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2908 case ISD::ATOMIC_LOAD_XOR:
2909 switch (VT.SimpleTy) {
2910 default: llvm_unreachable("Unexpected value type for atomic!");
2911 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2912 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2913 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2914 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2915 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2918 case ISD::ATOMIC_LOAD_NAND:
2919 switch (VT.SimpleTy) {
2920 default: llvm_unreachable("Unexpected value type for atomic!");
2921 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2922 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2923 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2924 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2925 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2928 case ISD::ATOMIC_LOAD_MAX:
2929 switch (VT.SimpleTy) {
2930 default: llvm_unreachable("Unexpected value type for atomic!");
2931 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2932 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2933 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2934 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2935 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2938 case ISD::ATOMIC_LOAD_UMAX:
2939 switch (VT.SimpleTy) {
2940 default: llvm_unreachable("Unexpected value type for atomic!");
2941 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2942 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2943 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2944 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2945 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2948 case ISD::ATOMIC_LOAD_MIN:
2949 switch (VT.SimpleTy) {
2950 default: llvm_unreachable("Unexpected value type for atomic!");
2951 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2952 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2953 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2954 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2955 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2958 case ISD::ATOMIC_LOAD_UMIN:
2959 switch (VT.SimpleTy) {
2960 default: llvm_unreachable("Unexpected value type for atomic!");
2961 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2962 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2963 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2964 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2965 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2970 return ExpandChainLibCall(LC, Node, false);
2973 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2974 SmallVector<SDValue, 8> Results;
2976 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2978 switch (Node->getOpcode()) {
2981 case ISD::CTLZ_ZERO_UNDEF:
2983 case ISD::CTTZ_ZERO_UNDEF:
2984 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2985 Results.push_back(Tmp1);
2988 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2990 case ISD::FRAMEADDR:
2991 case ISD::RETURNADDR:
2992 case ISD::FRAME_TO_ARGS_OFFSET:
2993 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2995 case ISD::FLT_ROUNDS_:
2996 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2998 case ISD::EH_RETURN:
3002 case ISD::EH_SJLJ_LONGJMP:
3003 // If the target didn't expand these, there's nothing to do, so just
3004 // preserve the chain and be done.
3005 Results.push_back(Node->getOperand(0));
3007 case ISD::EH_SJLJ_SETJMP:
3008 // If the target didn't expand this, just return 'zero' and preserve the
3010 Results.push_back(DAG.getConstant(0, MVT::i32));
3011 Results.push_back(Node->getOperand(0));
3013 case ISD::ATOMIC_FENCE: {
3014 // If the target didn't lower this, lower it to '__sync_synchronize()' call
3015 // FIXME: handle "fence singlethread" more efficiently.
3016 TargetLowering::ArgListTy Args;
3018 TargetLowering::CallLoweringInfo CLI(DAG);
3019 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3020 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3021 DAG.getExternalSymbol("__sync_synchronize",
3022 TLI.getPointerTy()), std::move(Args), 0);
3024 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3026 Results.push_back(CallResult.second);
3029 case ISD::ATOMIC_LOAD: {
3030 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3031 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
3032 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3033 SDValue Swap = DAG.getAtomicCmpSwap(
3034 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3035 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3036 cast<AtomicSDNode>(Node)->getMemOperand(),
3037 cast<AtomicSDNode>(Node)->getOrdering(),
3038 cast<AtomicSDNode>(Node)->getOrdering(),
3039 cast<AtomicSDNode>(Node)->getSynchScope());
3040 Results.push_back(Swap.getValue(0));
3041 Results.push_back(Swap.getValue(1));
3044 case ISD::ATOMIC_STORE: {
3045 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3046 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3047 cast<AtomicSDNode>(Node)->getMemoryVT(),
3048 Node->getOperand(0),
3049 Node->getOperand(1), Node->getOperand(2),
3050 cast<AtomicSDNode>(Node)->getMemOperand(),
3051 cast<AtomicSDNode>(Node)->getOrdering(),
3052 cast<AtomicSDNode>(Node)->getSynchScope());
3053 Results.push_back(Swap.getValue(1));
3056 // By default, atomic intrinsics are marked Legal and lowered. Targets
3057 // which don't support them directly, however, may want libcalls, in which
3058 // case they mark them Expand, and we get here.
3059 case ISD::ATOMIC_SWAP:
3060 case ISD::ATOMIC_LOAD_ADD:
3061 case ISD::ATOMIC_LOAD_SUB:
3062 case ISD::ATOMIC_LOAD_AND:
3063 case ISD::ATOMIC_LOAD_OR:
3064 case ISD::ATOMIC_LOAD_XOR:
3065 case ISD::ATOMIC_LOAD_NAND:
3066 case ISD::ATOMIC_LOAD_MIN:
3067 case ISD::ATOMIC_LOAD_MAX:
3068 case ISD::ATOMIC_LOAD_UMIN:
3069 case ISD::ATOMIC_LOAD_UMAX:
3070 case ISD::ATOMIC_CMP_SWAP: {
3071 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3072 Results.push_back(Tmp.first);
3073 Results.push_back(Tmp.second);
3076 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3077 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3078 // splits out the success value as a comparison. Expanding the resulting
3079 // ATOMIC_CMP_SWAP will produce a libcall.
3080 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3081 SDValue Res = DAG.getAtomicCmpSwap(
3082 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3083 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3084 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3085 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3086 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3087 cast<AtomicSDNode>(Node)->getSynchScope());
3089 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3090 Res, Node->getOperand(2), ISD::SETEQ);
3092 Results.push_back(Res.getValue(0));
3093 Results.push_back(Success);
3094 Results.push_back(Res.getValue(1));
3097 case ISD::DYNAMIC_STACKALLOC:
3098 ExpandDYNAMIC_STACKALLOC(Node, Results);
3100 case ISD::MERGE_VALUES:
3101 for (unsigned i = 0; i < Node->getNumValues(); i++)
3102 Results.push_back(Node->getOperand(i));
3105 EVT VT = Node->getValueType(0);
3107 Results.push_back(DAG.getConstant(0, VT));
3109 assert(VT.isFloatingPoint() && "Unknown value type!");
3110 Results.push_back(DAG.getConstantFP(0, VT));
3115 // If this operation is not supported, lower it to 'abort()' call
3116 TargetLowering::ArgListTy Args;
3117 TargetLowering::CallLoweringInfo CLI(DAG);
3118 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3119 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3120 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
3121 std::move(Args), 0);
3122 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3124 Results.push_back(CallResult.second);
3129 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3130 Node->getValueType(0), dl);
3131 Results.push_back(Tmp1);
3133 case ISD::FP_EXTEND:
3134 Tmp1 = EmitStackConvert(Node->getOperand(0),
3135 Node->getOperand(0).getValueType(),
3136 Node->getValueType(0), dl);
3137 Results.push_back(Tmp1);
3139 case ISD::SIGN_EXTEND_INREG: {
3140 // NOTE: we could fall back on load/store here too for targets without
3141 // SAR. However, it is doubtful that any exist.
3142 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3143 EVT VT = Node->getValueType(0);
3144 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3147 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3148 ExtraVT.getScalarType().getSizeInBits();
3149 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3150 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3151 Node->getOperand(0), ShiftCst);
3152 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3153 Results.push_back(Tmp1);
3156 case ISD::FP_ROUND_INREG: {
3157 // The only way we can lower this is to turn it into a TRUNCSTORE,
3158 // EXTLOAD pair, targeting a temporary location (a stack slot).
3160 // NOTE: there is a choice here between constantly creating new stack
3161 // slots and always reusing the same one. We currently always create
3162 // new ones, as reuse may inhibit scheduling.
3163 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3164 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3165 Node->getValueType(0), dl);
3166 Results.push_back(Tmp1);
3169 case ISD::SINT_TO_FP:
3170 case ISD::UINT_TO_FP:
3171 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3172 Node->getOperand(0), Node->getValueType(0), dl);
3173 Results.push_back(Tmp1);
3175 case ISD::FP_TO_SINT:
3176 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3177 Results.push_back(Tmp1);
3179 case ISD::FP_TO_UINT: {
3180 SDValue True, False;
3181 EVT VT = Node->getOperand(0).getValueType();
3182 EVT NVT = Node->getValueType(0);
3183 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3184 APInt::getNullValue(VT.getSizeInBits()));
3185 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3186 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3187 Tmp1 = DAG.getConstantFP(apf, VT);
3188 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3189 Node->getOperand(0),
3191 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3192 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3193 DAG.getNode(ISD::FSUB, dl, VT,
3194 Node->getOperand(0), Tmp1));
3195 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3196 DAG.getConstant(x, NVT));
3197 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3198 Results.push_back(Tmp1);
3202 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3203 EVT VT = Node->getValueType(0);
3204 Tmp1 = Node->getOperand(0);
3205 Tmp2 = Node->getOperand(1);
3206 unsigned Align = Node->getConstantOperandVal(3);
3208 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3209 MachinePointerInfo(V),
3210 false, false, false, 0);
3211 SDValue VAList = VAListLoad;
3213 if (Align > TLI.getMinStackArgumentAlignment()) {
3214 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3216 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3217 DAG.getConstant(Align - 1,
3218 VAList.getValueType()));
3220 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3221 DAG.getConstant(-(int64_t)Align,
3222 VAList.getValueType()));
3225 // Increment the pointer, VAList, to the next vaarg
3226 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3227 DAG.getConstant(TLI.getDataLayout()->
3228 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3229 VAList.getValueType()));
3230 // Store the incremented VAList to the legalized pointer
3231 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3232 MachinePointerInfo(V), false, false, 0);
3233 // Load the actual argument out of the pointer VAList
3234 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3235 false, false, false, 0));
3236 Results.push_back(Results[0].getValue(1));
3240 // This defaults to loading a pointer from the input and storing it to the
3241 // output, returning the chain.
3242 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3243 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3244 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3245 Node->getOperand(2), MachinePointerInfo(VS),
3246 false, false, false, 0);
3247 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3248 MachinePointerInfo(VD), false, false, 0);
3249 Results.push_back(Tmp1);
3252 case ISD::EXTRACT_VECTOR_ELT:
3253 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3254 // This must be an access of the only element. Return it.
3255 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3256 Node->getOperand(0));
3258 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3259 Results.push_back(Tmp1);
3261 case ISD::EXTRACT_SUBVECTOR:
3262 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3264 case ISD::INSERT_SUBVECTOR:
3265 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3267 case ISD::CONCAT_VECTORS: {
3268 Results.push_back(ExpandVectorBuildThroughStack(Node));
3271 case ISD::SCALAR_TO_VECTOR:
3272 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3274 case ISD::INSERT_VECTOR_ELT:
3275 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3276 Node->getOperand(1),
3277 Node->getOperand(2), dl));
3279 case ISD::VECTOR_SHUFFLE: {
3280 SmallVector<int, 32> NewMask;
3281 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3283 EVT VT = Node->getValueType(0);
3284 EVT EltVT = VT.getVectorElementType();
3285 SDValue Op0 = Node->getOperand(0);
3286 SDValue Op1 = Node->getOperand(1);
3287 if (!TLI.isTypeLegal(EltVT)) {
3289 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3291 // BUILD_VECTOR operands are allowed to be wider than the element type.
3292 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3294 if (NewEltVT.bitsLT(EltVT)) {
3296 // Convert shuffle node.
3297 // If original node was v4i64 and the new EltVT is i32,
3298 // cast operands to v8i32 and re-build the mask.
3300 // Calculate new VT, the size of the new VT should be equal to original.
3302 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3303 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3304 assert(NewVT.bitsEq(VT));
3306 // cast operands to new VT
3307 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3308 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3310 // Convert the shuffle mask
3311 unsigned int factor =
3312 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3314 // EltVT gets smaller
3317 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3319 for (unsigned fi = 0; fi < factor; ++fi)
3320 NewMask.push_back(Mask[i]);
3323 for (unsigned fi = 0; fi < factor; ++fi)
3324 NewMask.push_back(Mask[i]*factor+fi);
3332 unsigned NumElems = VT.getVectorNumElements();
3333 SmallVector<SDValue, 16> Ops;
3334 for (unsigned i = 0; i != NumElems; ++i) {
3336 Ops.push_back(DAG.getUNDEF(EltVT));
3339 unsigned Idx = Mask[i];
3341 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3343 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3345 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3347 DAG.getConstant(Idx - NumElems,
3348 TLI.getVectorIdxTy())));
3351 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3352 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3353 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3354 Results.push_back(Tmp1);
3357 case ISD::EXTRACT_ELEMENT: {
3358 EVT OpTy = Node->getOperand(0).getValueType();
3359 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3361 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3362 DAG.getConstant(OpTy.getSizeInBits()/2,
3363 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3364 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3367 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3368 Node->getOperand(0));
3370 Results.push_back(Tmp1);
3373 case ISD::STACKSAVE:
3374 // Expand to CopyFromReg if the target set
3375 // StackPointerRegisterToSaveRestore.
3376 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3377 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3378 Node->getValueType(0)));
3379 Results.push_back(Results[0].getValue(1));
3381 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3382 Results.push_back(Node->getOperand(0));
3385 case ISD::STACKRESTORE:
3386 // Expand to CopyToReg if the target set
3387 // StackPointerRegisterToSaveRestore.
3388 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3389 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3390 Node->getOperand(1)));
3392 Results.push_back(Node->getOperand(0));
3395 case ISD::FCOPYSIGN:
3396 Results.push_back(ExpandFCOPYSIGN(Node));
3399 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3400 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3401 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3402 Node->getOperand(0));
3403 Results.push_back(Tmp1);
3406 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3407 EVT VT = Node->getValueType(0);
3408 Tmp1 = Node->getOperand(0);
3409 Tmp2 = DAG.getConstantFP(0.0, VT);
3410 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3411 Tmp1, Tmp2, ISD::SETUGT);
3412 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3413 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3414 Results.push_back(Tmp1);
3418 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3419 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3420 RTLIB::SQRT_PPCF128));
3424 EVT VT = Node->getValueType(0);
3425 bool isSIN = Node->getOpcode() == ISD::FSIN;
3426 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3427 // fcos which share the same operand and both are used.
3428 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3429 canCombineSinCosLibcall(Node, TLI, TM))
3430 && useSinCos(Node)) {
3431 SDVTList VTs = DAG.getVTList(VT, VT);
3432 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3434 Tmp1 = Tmp1.getValue(1);
3435 Results.push_back(Tmp1);
3437 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3438 RTLIB::SIN_F80, RTLIB::SIN_F128,
3439 RTLIB::SIN_PPCF128));
3441 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3442 RTLIB::COS_F80, RTLIB::COS_F128,
3443 RTLIB::COS_PPCF128));
3448 // Expand into sincos libcall.
3449 ExpandSinCosLibCall(Node, Results);
3452 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3453 RTLIB::LOG_F80, RTLIB::LOG_F128,
3454 RTLIB::LOG_PPCF128));
3457 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3458 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3459 RTLIB::LOG2_PPCF128));
3462 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3463 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3464 RTLIB::LOG10_PPCF128));
3467 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3468 RTLIB::EXP_F80, RTLIB::EXP_F128,
3469 RTLIB::EXP_PPCF128));
3472 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3473 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3474 RTLIB::EXP2_PPCF128));
3477 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3478 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3479 RTLIB::TRUNC_PPCF128));
3482 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3483 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3484 RTLIB::FLOOR_PPCF128));
3487 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3488 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3489 RTLIB::CEIL_PPCF128));
3492 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3493 RTLIB::RINT_F80, RTLIB::RINT_F128,
3494 RTLIB::RINT_PPCF128));
3496 case ISD::FNEARBYINT:
3497 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3498 RTLIB::NEARBYINT_F64,
3499 RTLIB::NEARBYINT_F80,
3500 RTLIB::NEARBYINT_F128,
3501 RTLIB::NEARBYINT_PPCF128));
3504 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3508 RTLIB::ROUND_PPCF128));
3511 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3512 RTLIB::POWI_F80, RTLIB::POWI_F128,
3513 RTLIB::POWI_PPCF128));
3516 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3517 RTLIB::POW_F80, RTLIB::POW_F128,
3518 RTLIB::POW_PPCF128));
3521 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3522 RTLIB::DIV_F80, RTLIB::DIV_F128,
3523 RTLIB::DIV_PPCF128));
3526 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3527 RTLIB::REM_F80, RTLIB::REM_F128,
3528 RTLIB::REM_PPCF128));
3531 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3532 RTLIB::FMA_F80, RTLIB::FMA_F128,
3533 RTLIB::FMA_PPCF128));
3535 case ISD::FP16_TO_FP: {
3536 if (Node->getValueType(0) == MVT::f32) {
3537 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3541 // We can extend to types bigger than f32 in two steps without changing the
3542 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3543 // the option of emitting that before resorting to a libcall.
3545 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3547 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3550 case ISD::FP_TO_FP16: {
3552 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3553 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3554 Results.push_back(ExpandLibCall(LC, Node, false));
3557 case ISD::ConstantFP: {
3558 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3559 // Check to see if this FP immediate is already legal.
3560 // If this is a legal constant, turn it into a TargetConstantFP node.
3561 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3562 Results.push_back(ExpandConstantFP(CFP, true));
3566 EVT VT = Node->getValueType(0);
3567 assert(TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3568 TLI.isOperationLegalOrCustom(ISD::FNEG, VT) &&
3569 "Don't know how to expand this FP subtraction!");
3570 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3571 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3572 Results.push_back(Tmp1);
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)));
3690 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3691 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3692 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3693 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3694 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3695 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3696 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3697 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3698 SDValue Shift = DAG.getConstant(HalfType.getSizeInBits(),
3699 TLI.getShiftAmountTy(HalfType));
3700 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3701 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3705 Tmp1 = ExpandIntLibCall(Node, false,
3707 RTLIB::MUL_I16, RTLIB::MUL_I32,
3708 RTLIB::MUL_I64, RTLIB::MUL_I128);
3709 Results.push_back(Tmp1);
3714 SDValue LHS = Node->getOperand(0);
3715 SDValue RHS = Node->getOperand(1);
3716 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3717 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3719 Results.push_back(Sum);
3720 EVT ResultType = Node->getValueType(1);
3721 EVT OType = getSetCCResultType(Node->getValueType(0));
3723 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3725 // LHSSign -> LHS >= 0
3726 // RHSSign -> RHS >= 0
3727 // SumSign -> Sum >= 0
3730 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3732 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3734 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3735 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3736 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3737 Node->getOpcode() == ISD::SADDO ?
3738 ISD::SETEQ : ISD::SETNE);
3740 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3741 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3743 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3744 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3749 SDValue LHS = Node->getOperand(0);
3750 SDValue RHS = Node->getOperand(1);
3751 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3752 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3754 Results.push_back(Sum);
3756 EVT ResultType = Node->getValueType(1);
3757 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3759 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3760 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3762 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3767 EVT VT = Node->getValueType(0);
3768 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3769 SDValue LHS = Node->getOperand(0);
3770 SDValue RHS = Node->getOperand(1);
3773 static const unsigned Ops[2][3] =
3774 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3775 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3776 bool isSigned = Node->getOpcode() == ISD::SMULO;
3777 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3778 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3779 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3780 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3781 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3783 TopHalf = BottomHalf.getValue(1);
3784 } else if (TLI.isTypeLegal(WideVT)) {
3785 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3786 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3787 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3788 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3789 DAG.getIntPtrConstant(0));
3790 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3791 DAG.getIntPtrConstant(1));
3793 // We can fall back to a libcall with an illegal type for the MUL if we
3794 // have a libcall big enough.
3795 // Also, we can fall back to a division in some cases, but that's a big
3796 // performance hit in the general case.
3797 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3798 if (WideVT == MVT::i16)
3799 LC = RTLIB::MUL_I16;
3800 else if (WideVT == MVT::i32)
3801 LC = RTLIB::MUL_I32;
3802 else if (WideVT == MVT::i64)
3803 LC = RTLIB::MUL_I64;
3804 else if (WideVT == MVT::i128)
3805 LC = RTLIB::MUL_I128;
3806 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3808 // The high part is obtained by SRA'ing all but one of the bits of low
3810 unsigned LoSize = VT.getSizeInBits();
3811 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3812 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3813 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3814 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3816 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3817 // pre-lowered to the correct types. This all depends upon WideVT not
3818 // being a legal type for the architecture and thus has to be split to
3820 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3821 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3822 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3823 DAG.getIntPtrConstant(0));
3824 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3825 DAG.getIntPtrConstant(1));
3826 // Ret is a node with an illegal type. Because such things are not
3827 // generally permitted during this phase of legalization, delete the
3828 // node. The above EXTRACT_ELEMENT nodes should have been folded.
3829 DAG.DeleteNode(Ret.getNode());
3833 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3834 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3835 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3836 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3839 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3840 DAG.getConstant(0, VT), ISD::SETNE);
3842 Results.push_back(BottomHalf);
3843 Results.push_back(TopHalf);
3846 case ISD::BUILD_PAIR: {
3847 EVT PairTy = Node->getValueType(0);
3848 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3849 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3850 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3851 DAG.getConstant(PairTy.getSizeInBits()/2,
3852 TLI.getShiftAmountTy(PairTy)));
3853 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3857 Tmp1 = Node->getOperand(0);
3858 Tmp2 = Node->getOperand(1);
3859 Tmp3 = Node->getOperand(2);
3860 if (Tmp1.getOpcode() == ISD::SETCC) {
3861 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3863 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3865 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3866 DAG.getConstant(0, Tmp1.getValueType()),
3867 Tmp2, Tmp3, ISD::SETNE);
3869 Results.push_back(Tmp1);
3872 SDValue Chain = Node->getOperand(0);
3873 SDValue Table = Node->getOperand(1);
3874 SDValue Index = Node->getOperand(2);
3876 EVT PTy = TLI.getPointerTy();
3878 const DataLayout &TD = *TLI.getDataLayout();
3879 unsigned EntrySize =
3880 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3882 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(),
3883 Index, DAG.getConstant(EntrySize, Index.getValueType()));
3884 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3887 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3888 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3889 MachinePointerInfo::getJumpTable(), MemVT,
3892 if (TM.getRelocationModel() == Reloc::PIC_) {
3893 // For PIC, the sequence is:
3894 // BRIND(load(Jumptable + index) + RelocBase)
3895 // RelocBase can be JumpTable, GOT or some sort of global base.
3896 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3897 TLI.getPICJumpTableRelocBase(Table, DAG));
3899 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3900 Results.push_back(Tmp1);
3904 // Expand brcond's setcc into its constituent parts and create a BR_CC
3906 Tmp1 = Node->getOperand(0);
3907 Tmp2 = Node->getOperand(1);
3908 if (Tmp2.getOpcode() == ISD::SETCC) {
3909 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3910 Tmp1, Tmp2.getOperand(2),
3911 Tmp2.getOperand(0), Tmp2.getOperand(1),
3912 Node->getOperand(2));
3914 // We test only the i1 bit. Skip the AND if UNDEF.
3915 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3916 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3917 DAG.getConstant(1, Tmp2.getValueType()));
3918 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3919 DAG.getCondCode(ISD::SETNE), Tmp3,
3920 DAG.getConstant(0, Tmp3.getValueType()),
3921 Node->getOperand(2));
3923 Results.push_back(Tmp1);
3926 Tmp1 = Node->getOperand(0);
3927 Tmp2 = Node->getOperand(1);
3928 Tmp3 = Node->getOperand(2);
3929 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3930 Tmp3, NeedInvert, dl);
3933 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3934 // condition code, create a new SETCC node.
3936 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3939 // If we expanded the SETCC by inverting the condition code, then wrap
3940 // the existing SETCC in a NOT to restore the intended condition.
3942 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3944 Results.push_back(Tmp1);
3948 // Otherwise, SETCC for the given comparison type must be completely
3949 // illegal; expand it into a SELECT_CC.
3950 EVT VT = Node->getValueType(0);
3952 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
3953 case TargetLowering::ZeroOrOneBooleanContent:
3954 case TargetLowering::UndefinedBooleanContent:
3957 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3961 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3962 DAG.getConstant(TrueValue, VT), DAG.getConstant(0, VT),
3964 Results.push_back(Tmp1);
3967 case ISD::SELECT_CC: {
3968 Tmp1 = Node->getOperand(0); // LHS
3969 Tmp2 = Node->getOperand(1); // RHS
3970 Tmp3 = Node->getOperand(2); // True
3971 Tmp4 = Node->getOperand(3); // False
3972 EVT VT = Node->getValueType(0);
3973 SDValue CC = Node->getOperand(4);
3974 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3976 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
3977 // If the condition code is legal, then we need to expand this
3978 // node using SETCC and SELECT.
3979 EVT CmpVT = Tmp1.getValueType();
3980 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3981 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3983 EVT CCVT = TLI.getSetCCResultType(*DAG.getContext(), CmpVT);
3984 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3985 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3989 // SELECT_CC is legal, so the condition code must not be.
3990 bool Legalized = false;
3991 // Try to legalize by inverting the condition. This is for targets that
3992 // might support an ordered version of a condition, but not the unordered
3993 // version (or vice versa).
3994 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3995 Tmp1.getValueType().isInteger());
3996 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3997 // Use the new condition code and swap true and false
3999 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
4001 // If The inverse is not legal, then try to swap the arguments using
4002 // the inverse condition code.
4003 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
4004 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
4005 // The swapped inverse condition is legal, so swap true and false,
4008 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
4013 Legalized = LegalizeSetCCCondCode(
4014 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
4017 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
4019 // If we expanded the SETCC by inverting the condition code, then swap
4020 // the True/False operands to match.
4022 std::swap(Tmp3, Tmp4);
4024 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
4025 // condition code, create a new SELECT_CC node.
4027 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
4028 Tmp1, Tmp2, Tmp3, Tmp4, CC);
4030 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
4031 CC = DAG.getCondCode(ISD::SETNE);
4032 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
4033 Tmp2, Tmp3, Tmp4, CC);
4036 Results.push_back(Tmp1);
4040 Tmp1 = Node->getOperand(0); // Chain
4041 Tmp2 = Node->getOperand(2); // LHS
4042 Tmp3 = Node->getOperand(3); // RHS
4043 Tmp4 = Node->getOperand(1); // CC
4045 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
4046 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
4048 assert(Legalized && "Can't legalize BR_CC with legal condition!");
4050 // If we expanded the SETCC by inverting the condition code, then wrap
4051 // the existing SETCC in a NOT to restore the intended condition.
4053 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
4055 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
4057 if (Tmp4.getNode()) {
4058 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
4059 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
4061 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
4062 Tmp4 = DAG.getCondCode(ISD::SETNE);
4063 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
4064 Tmp2, Tmp3, Node->getOperand(4));
4066 Results.push_back(Tmp1);
4069 case ISD::BUILD_VECTOR:
4070 Results.push_back(ExpandBUILD_VECTOR(Node));
4075 // Scalarize vector SRA/SRL/SHL.
4076 EVT VT = Node->getValueType(0);
4077 assert(VT.isVector() && "Unable to legalize non-vector shift");
4078 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
4079 unsigned NumElem = VT.getVectorNumElements();
4081 SmallVector<SDValue, 8> Scalars;
4082 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
4083 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4085 Node->getOperand(0), DAG.getConstant(Idx,
4086 TLI.getVectorIdxTy()));
4087 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4089 Node->getOperand(1), DAG.getConstant(Idx,
4090 TLI.getVectorIdxTy()));
4091 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
4092 VT.getScalarType(), Ex, Sh));
4095 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars);
4096 ReplaceNode(SDValue(Node, 0), Result);
4099 case ISD::GLOBAL_OFFSET_TABLE:
4100 case ISD::GlobalAddress:
4101 case ISD::GlobalTLSAddress:
4102 case ISD::ExternalSymbol:
4103 case ISD::ConstantPool:
4104 case ISD::JumpTable:
4105 case ISD::INTRINSIC_W_CHAIN:
4106 case ISD::INTRINSIC_WO_CHAIN:
4107 case ISD::INTRINSIC_VOID:
4108 // FIXME: Custom lowering for these operations shouldn't return null!
4112 // Replace the original node with the legalized result.
4113 if (!Results.empty())
4114 ReplaceNode(Node, Results.data());
4117 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4118 SmallVector<SDValue, 8> Results;
4119 MVT OVT = Node->getSimpleValueType(0);
4120 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4121 Node->getOpcode() == ISD::SINT_TO_FP ||
4122 Node->getOpcode() == ISD::SETCC) {
4123 OVT = Node->getOperand(0).getSimpleValueType();
4125 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4127 SDValue Tmp1, Tmp2, Tmp3;
4128 switch (Node->getOpcode()) {
4130 case ISD::CTTZ_ZERO_UNDEF:
4132 case ISD::CTLZ_ZERO_UNDEF:
4134 // Zero extend the argument.
4135 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4136 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4137 // already the correct result.
4138 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4139 if (Node->getOpcode() == ISD::CTTZ) {
4140 // FIXME: This should set a bit in the zero extended value instead.
4141 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
4142 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
4144 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
4145 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
4146 } else if (Node->getOpcode() == ISD::CTLZ ||
4147 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4148 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4149 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4150 DAG.getConstant(NVT.getSizeInBits() -
4151 OVT.getSizeInBits(), NVT));
4153 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4156 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4157 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4158 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
4159 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
4160 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
4161 Results.push_back(Tmp1);
4164 case ISD::FP_TO_UINT:
4165 case ISD::FP_TO_SINT:
4166 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4167 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4168 Results.push_back(Tmp1);
4170 case ISD::UINT_TO_FP:
4171 case ISD::SINT_TO_FP:
4172 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4173 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4174 Results.push_back(Tmp1);
4177 SDValue Chain = Node->getOperand(0); // Get the chain.
4178 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4181 if (OVT.isVector()) {
4182 TruncOp = ISD::BITCAST;
4184 assert(OVT.isInteger()
4185 && "VAARG promotion is supported only for vectors or integer types");
4186 TruncOp = ISD::TRUNCATE;
4189 // Perform the larger operation, then convert back
4190 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4191 Node->getConstantOperandVal(3));
4192 Chain = Tmp1.getValue(1);
4194 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4196 // Modified the chain result - switch anything that used the old chain to
4198 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4199 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4201 UpdatedNodes->insert(Tmp2.getNode());
4202 UpdatedNodes->insert(Chain.getNode());
4210 unsigned ExtOp, TruncOp;
4211 if (OVT.isVector()) {
4212 ExtOp = ISD::BITCAST;
4213 TruncOp = ISD::BITCAST;
4215 assert(OVT.isInteger() && "Cannot promote logic operation");
4216 ExtOp = ISD::ANY_EXTEND;
4217 TruncOp = ISD::TRUNCATE;
4219 // Promote each of the values to the new type.
4220 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4221 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4222 // Perform the larger operation, then convert back
4223 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4224 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4228 unsigned ExtOp, TruncOp;
4229 if (Node->getValueType(0).isVector() ||
4230 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4231 ExtOp = ISD::BITCAST;
4232 TruncOp = ISD::BITCAST;
4233 } else if (Node->getValueType(0).isInteger()) {
4234 ExtOp = ISD::ANY_EXTEND;
4235 TruncOp = ISD::TRUNCATE;
4237 ExtOp = ISD::FP_EXTEND;
4238 TruncOp = ISD::FP_ROUND;
4240 Tmp1 = Node->getOperand(0);
4241 // Promote each of the values to the new type.
4242 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4243 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4244 // Perform the larger operation, then round down.
4245 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4246 if (TruncOp != ISD::FP_ROUND)
4247 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4249 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4250 DAG.getIntPtrConstant(0));
4251 Results.push_back(Tmp1);
4254 case ISD::VECTOR_SHUFFLE: {
4255 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4257 // Cast the two input vectors.
4258 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4259 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4261 // Convert the shuffle mask to the right # elements.
4262 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4263 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4264 Results.push_back(Tmp1);
4268 unsigned ExtOp = ISD::FP_EXTEND;
4269 if (NVT.isInteger()) {
4270 ISD::CondCode CCCode =
4271 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4272 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4274 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4275 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4276 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4277 Tmp1, Tmp2, Node->getOperand(2)));
4283 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4284 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4285 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4286 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4287 Tmp3, DAG.getIntPtrConstant(0)));
4294 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4295 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4296 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4297 Tmp2, DAG.getIntPtrConstant(0)));
4302 // Replace the original node with the legalized result.
4303 if (!Results.empty())
4304 ReplaceNode(Node, Results.data());
4307 // SelectionDAG::Legalize - This is the entry point for the file.
4309 void SelectionDAG::Legalize() {
4310 AssignTopologicalOrder();
4312 allnodes_iterator LegalizePosition;
4313 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4314 SelectionDAGLegalize Legalizer(*this, LegalizePosition, LegalizedNodes);
4316 // Visit all the nodes. We start in topological order, so that we see
4317 // nodes with their original operands intact. Legalization can produce
4318 // new nodes which may themselves need to be legalized. Iterate until all
4319 // nodes have been legalized.
4321 bool AnyLegalized = false;
4322 for (LegalizePosition = allnodes_end();
4323 LegalizePosition != allnodes_begin(); ) {
4326 SDNode *N = LegalizePosition;
4327 if (LegalizedNodes.insert(N)) {
4328 AnyLegalized = true;
4329 Legalizer.LegalizeOp(N);
4337 // Remove dead nodes now.
4341 bool SelectionDAG::LegalizeOp(SDNode *N,
4342 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4343 allnodes_iterator LegalizePosition(N);
4344 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4345 SelectionDAGLegalize Legalizer(*this, LegalizePosition, LegalizedNodes,
4348 // Directly insert the node in question, and legalize it. This will recurse
4349 // as needed through operands.
4350 LegalizedNodes.insert(N);
4351 Legalizer.LegalizeOp(N);
4353 return LegalizedNodes.count(N);