1 //===-- LegalizeDAG.cpp - Implement SelectionDAG::Legalize ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SelectionDAG::Legalize method.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/CodeGen/SelectionDAG.h"
15 #include "llvm/ADT/SetVector.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallSet.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/ADT/Triple.h"
20 #include "llvm/CodeGen/Analysis.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineJumpTableInfo.h"
23 #include "llvm/IR/CallingConv.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/DataLayout.h"
26 #include "llvm/IR/DebugInfo.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetFrameLowering.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetSubtargetInfo.h"
40 #define DEBUG_TYPE "legalizedag"
42 //===----------------------------------------------------------------------===//
43 /// This takes an arbitrary SelectionDAG as input and
44 /// hacks on it until the target machine can handle it. This involves
45 /// eliminating value sizes the machine cannot handle (promoting small sizes to
46 /// large sizes or splitting up large values into small values) as well as
47 /// eliminating operations the machine cannot handle.
49 /// This code also does a small amount of optimization and recognition of idioms
50 /// as part of its processing. For example, if a target does not support a
51 /// 'setcc' instruction efficiently, but does support 'brcc' instruction, this
52 /// will attempt merge setcc and brc instructions into brcc's.
55 class SelectionDAGLegalize {
56 const TargetMachine &TM;
57 const TargetLowering &TLI;
60 /// \brief The set of nodes which have already been legalized. We hold a
61 /// reference to it in order to update as necessary on node deletion.
62 SmallPtrSetImpl<SDNode *> &LegalizedNodes;
64 /// \brief A set of all the nodes updated during legalization.
65 SmallSetVector<SDNode *, 16> *UpdatedNodes;
67 EVT getSetCCResultType(EVT VT) const {
68 return TLI.getSetCCResultType(*DAG.getContext(), VT);
71 // Libcall insertion helpers.
74 SelectionDAGLegalize(SelectionDAG &DAG,
75 SmallPtrSetImpl<SDNode *> &LegalizedNodes,
76 SmallSetVector<SDNode *, 16> *UpdatedNodes = nullptr)
77 : TM(DAG.getTarget()), TLI(DAG.getTargetLoweringInfo()), DAG(DAG),
78 LegalizedNodes(LegalizedNodes), UpdatedNodes(UpdatedNodes) {}
80 /// \brief Legalizes the given operation.
81 void LegalizeOp(SDNode *Node);
84 SDValue OptimizeFloatStore(StoreSDNode *ST);
86 void LegalizeLoadOps(SDNode *Node);
87 void LegalizeStoreOps(SDNode *Node);
89 /// Some targets cannot handle a variable
90 /// insertion index for the INSERT_VECTOR_ELT instruction. In this case, it
91 /// is necessary to spill the vector being inserted into to memory, perform
92 /// the insert there, and then read the result back.
93 SDValue PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val,
94 SDValue Idx, SDLoc dl);
95 SDValue ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val,
96 SDValue Idx, SDLoc dl);
98 /// Return a vector shuffle operation which
99 /// performs the same shuffe in terms of order or result bytes, but on a type
100 /// whose vector element type is narrower than the original shuffle type.
101 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
102 SDValue ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
103 SDValue N1, SDValue N2,
104 ArrayRef<int> Mask) const;
106 bool LegalizeSetCCCondCode(EVT VT, SDValue &LHS, SDValue &RHS, SDValue &CC,
107 bool &NeedInvert, SDLoc dl);
109 SDValue ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, bool isSigned);
110 SDValue ExpandLibCall(RTLIB::Libcall LC, EVT RetVT, const SDValue *Ops,
111 unsigned NumOps, bool isSigned, SDLoc dl);
113 std::pair<SDValue, SDValue> ExpandChainLibCall(RTLIB::Libcall LC,
114 SDNode *Node, bool isSigned);
115 SDValue ExpandFPLibCall(SDNode *Node, RTLIB::Libcall Call_F32,
116 RTLIB::Libcall Call_F64, RTLIB::Libcall Call_F80,
117 RTLIB::Libcall Call_F128,
118 RTLIB::Libcall Call_PPCF128);
119 SDValue ExpandIntLibCall(SDNode *Node, bool isSigned,
120 RTLIB::Libcall Call_I8,
121 RTLIB::Libcall Call_I16,
122 RTLIB::Libcall Call_I32,
123 RTLIB::Libcall Call_I64,
124 RTLIB::Libcall Call_I128);
125 void ExpandDivRemLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
126 void ExpandSinCosLibCall(SDNode *Node, SmallVectorImpl<SDValue> &Results);
128 SDValue EmitStackConvert(SDValue SrcOp, EVT SlotVT, EVT DestVT, SDLoc dl);
129 SDValue ExpandBUILD_VECTOR(SDNode *Node);
130 SDValue ExpandSCALAR_TO_VECTOR(SDNode *Node);
131 void ExpandDYNAMIC_STACKALLOC(SDNode *Node,
132 SmallVectorImpl<SDValue> &Results);
133 SDValue ExpandFCOPYSIGN(SDNode *Node);
134 SDValue ExpandLegalINT_TO_FP(bool isSigned, SDValue LegalOp, EVT DestVT,
136 SDValue PromoteLegalINT_TO_FP(SDValue LegalOp, EVT DestVT, bool isSigned,
138 SDValue PromoteLegalFP_TO_INT(SDValue LegalOp, EVT DestVT, bool isSigned,
141 SDValue ExpandBSWAP(SDValue Op, SDLoc dl);
142 SDValue ExpandBitCount(unsigned Opc, SDValue Op, SDLoc dl);
144 SDValue ExpandExtractFromVectorThroughStack(SDValue Op);
145 SDValue ExpandInsertToVectorThroughStack(SDValue Op);
146 SDValue ExpandVectorBuildThroughStack(SDNode* Node);
148 SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP);
150 std::pair<SDValue, SDValue> ExpandAtomic(SDNode *Node);
152 void ExpandNode(SDNode *Node);
153 void PromoteNode(SDNode *Node);
156 // Node replacement helpers
157 void ReplacedNode(SDNode *N) {
158 LegalizedNodes.erase(N);
160 UpdatedNodes->insert(N);
162 void ReplaceNode(SDNode *Old, SDNode *New) {
163 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
164 dbgs() << " with: "; New->dump(&DAG));
166 assert(Old->getNumValues() == New->getNumValues() &&
167 "Replacing one node with another that produces a different number "
169 DAG.ReplaceAllUsesWith(Old, New);
170 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i)
171 DAG.TransferDbgValues(SDValue(Old, i), SDValue(New, i));
173 UpdatedNodes->insert(New);
176 void ReplaceNode(SDValue Old, SDValue New) {
177 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG);
178 dbgs() << " with: "; New->dump(&DAG));
180 DAG.ReplaceAllUsesWith(Old, New);
181 DAG.TransferDbgValues(Old, New);
183 UpdatedNodes->insert(New.getNode());
184 ReplacedNode(Old.getNode());
186 void ReplaceNode(SDNode *Old, const SDValue *New) {
187 DEBUG(dbgs() << " ... replacing: "; Old->dump(&DAG));
189 DAG.ReplaceAllUsesWith(Old, New);
190 for (unsigned i = 0, e = Old->getNumValues(); i != e; ++i) {
191 DEBUG(dbgs() << (i == 0 ? " with: "
194 DAG.TransferDbgValues(SDValue(Old, i), New[i]);
196 UpdatedNodes->insert(New[i].getNode());
203 /// Return a vector shuffle operation which
204 /// performs the same shuffe in terms of order or result bytes, but on a type
205 /// whose vector element type is narrower than the original shuffle type.
206 /// e.g. <v4i32> <0, 1, 0, 1> -> v8i16 <0, 1, 2, 3, 0, 1, 2, 3>
208 SelectionDAGLegalize::ShuffleWithNarrowerEltType(EVT NVT, EVT VT, SDLoc dl,
209 SDValue N1, SDValue N2,
210 ArrayRef<int> Mask) const {
211 unsigned NumMaskElts = VT.getVectorNumElements();
212 unsigned NumDestElts = NVT.getVectorNumElements();
213 unsigned NumEltsGrowth = NumDestElts / NumMaskElts;
215 assert(NumEltsGrowth && "Cannot promote to vector type with fewer elts!");
217 if (NumEltsGrowth == 1)
218 return DAG.getVectorShuffle(NVT, dl, N1, N2, &Mask[0]);
220 SmallVector<int, 8> NewMask;
221 for (unsigned i = 0; i != NumMaskElts; ++i) {
223 for (unsigned j = 0; j != NumEltsGrowth; ++j) {
225 NewMask.push_back(-1);
227 NewMask.push_back(Idx * NumEltsGrowth + j);
230 assert(NewMask.size() == NumDestElts && "Non-integer NumEltsGrowth?");
231 assert(TLI.isShuffleMaskLegal(NewMask, NVT) && "Shuffle not legal?");
232 return DAG.getVectorShuffle(NVT, dl, N1, N2, &NewMask[0]);
235 /// Expands the ConstantFP node to an integer constant or
236 /// a load from the constant pool.
238 SelectionDAGLegalize::ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP) {
242 // If a FP immediate is precise when represented as a float and if the
243 // target can do an extending load from float to double, we put it into
244 // the constant pool as a float, even if it's is statically typed as a
245 // double. This shrinks FP constants and canonicalizes them for targets where
246 // an FP extending load is the same cost as a normal load (such as on the x87
247 // fp stack or PPC FP unit).
248 EVT VT = CFP->getValueType(0);
249 ConstantFP *LLVMC = const_cast<ConstantFP*>(CFP->getConstantFPValue());
251 assert((VT == MVT::f64 || VT == MVT::f32) && "Invalid type expansion");
252 return DAG.getConstant(LLVMC->getValueAPF().bitcastToAPInt(),
253 (VT == MVT::f64) ? MVT::i64 : MVT::i32);
258 while (SVT != MVT::f32 && SVT != MVT::f16) {
259 SVT = (MVT::SimpleValueType)(SVT.getSimpleVT().SimpleTy - 1);
260 if (ConstantFPSDNode::isValueValidForType(SVT, CFP->getValueAPF()) &&
261 // Only do this if the target has a native EXTLOAD instruction from
263 TLI.isLoadExtLegal(ISD::EXTLOAD, OrigVT, SVT) &&
264 TLI.ShouldShrinkFPConstant(OrigVT)) {
265 Type *SType = SVT.getTypeForEVT(*DAG.getContext());
266 LLVMC = cast<ConstantFP>(ConstantExpr::getFPTrunc(LLVMC, SType));
272 SDValue CPIdx = DAG.getConstantPool(LLVMC, TLI.getPointerTy());
273 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
276 DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT,
278 CPIdx, MachinePointerInfo::getConstantPool(),
279 VT, false, false, false, Alignment);
283 DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx,
284 MachinePointerInfo::getConstantPool(), false, false, false,
289 /// Expands an unaligned store to 2 half-size stores.
290 static void ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG,
291 const TargetLowering &TLI,
292 SelectionDAGLegalize *DAGLegalize) {
293 assert(ST->getAddressingMode() == ISD::UNINDEXED &&
294 "unaligned indexed stores not implemented!");
295 SDValue Chain = ST->getChain();
296 SDValue Ptr = ST->getBasePtr();
297 SDValue Val = ST->getValue();
298 EVT VT = Val.getValueType();
299 int Alignment = ST->getAlignment();
300 unsigned AS = ST->getAddressSpace();
303 if (ST->getMemoryVT().isFloatingPoint() ||
304 ST->getMemoryVT().isVector()) {
305 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
306 if (TLI.isTypeLegal(intVT)) {
307 // Expand to a bitconvert of the value to the integer type of the
308 // same size, then a (misaligned) int store.
309 // FIXME: Does not handle truncating floating point stores!
310 SDValue Result = DAG.getNode(ISD::BITCAST, dl, intVT, Val);
311 Result = DAG.getStore(Chain, dl, Result, Ptr, ST->getPointerInfo(),
312 ST->isVolatile(), ST->isNonTemporal(), Alignment);
313 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
316 // Do a (aligned) store to a stack slot, then copy from the stack slot
317 // to the final destination using (unaligned) integer loads and stores.
318 EVT StoredVT = ST->getMemoryVT();
320 TLI.getRegisterType(*DAG.getContext(),
321 EVT::getIntegerVT(*DAG.getContext(),
322 StoredVT.getSizeInBits()));
323 unsigned StoredBytes = StoredVT.getSizeInBits() / 8;
324 unsigned RegBytes = RegVT.getSizeInBits() / 8;
325 unsigned NumRegs = (StoredBytes + RegBytes - 1) / RegBytes;
327 // Make sure the stack slot is also aligned for the register type.
328 SDValue StackPtr = DAG.CreateStackTemporary(StoredVT, RegVT);
330 // Perform the original store, only redirected to the stack slot.
331 SDValue Store = DAG.getTruncStore(Chain, dl,
332 Val, StackPtr, MachinePointerInfo(),
333 StoredVT, false, false, 0);
334 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy(AS));
335 SmallVector<SDValue, 8> Stores;
338 // Do all but one copies using the full register width.
339 for (unsigned i = 1; i < NumRegs; i++) {
340 // Load one integer register's worth from the stack slot.
341 SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr,
342 MachinePointerInfo(),
343 false, false, false, 0);
344 // Store it to the final location. Remember the store.
345 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr,
346 ST->getPointerInfo().getWithOffset(Offset),
347 ST->isVolatile(), ST->isNonTemporal(),
348 MinAlign(ST->getAlignment(), Offset)));
349 // Increment the pointers.
351 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
353 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
356 // The last store may be partial. Do a truncating store. On big-endian
357 // machines this requires an extending load from the stack slot to ensure
358 // that the bits are in the right place.
359 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
360 8 * (StoredBytes - Offset));
362 // Load from the stack slot.
363 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr,
364 MachinePointerInfo(),
365 MemVT, false, false, false, 0);
367 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr,
369 .getWithOffset(Offset),
370 MemVT, ST->isVolatile(),
372 MinAlign(ST->getAlignment(), Offset),
374 // The order of the stores doesn't matter - say it with a TokenFactor.
375 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
376 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
379 assert(ST->getMemoryVT().isInteger() &&
380 !ST->getMemoryVT().isVector() &&
381 "Unaligned store of unknown type.");
382 // Get the half-size VT
383 EVT NewStoredVT = ST->getMemoryVT().getHalfSizedIntegerVT(*DAG.getContext());
384 int NumBits = NewStoredVT.getSizeInBits();
385 int IncrementSize = NumBits / 8;
387 // Divide the stored value in two parts.
388 SDValue ShiftAmount = DAG.getConstant(NumBits,
389 TLI.getShiftAmountTy(Val.getValueType()));
391 SDValue Hi = DAG.getNode(ISD::SRL, dl, VT, Val, ShiftAmount);
393 // Store the two parts
394 SDValue Store1, Store2;
395 Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr,
396 ST->getPointerInfo(), NewStoredVT,
397 ST->isVolatile(), ST->isNonTemporal(), Alignment);
399 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
400 DAG.getConstant(IncrementSize, TLI.getPointerTy(AS)));
401 Alignment = MinAlign(Alignment, IncrementSize);
402 Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr,
403 ST->getPointerInfo().getWithOffset(IncrementSize),
404 NewStoredVT, ST->isVolatile(), ST->isNonTemporal(),
405 Alignment, ST->getAAInfo());
408 DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2);
409 DAGLegalize->ReplaceNode(SDValue(ST, 0), Result);
412 /// Expands an unaligned load to 2 half-size loads.
414 ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG,
415 const TargetLowering &TLI,
416 SDValue &ValResult, SDValue &ChainResult) {
417 assert(LD->getAddressingMode() == ISD::UNINDEXED &&
418 "unaligned indexed loads not implemented!");
419 SDValue Chain = LD->getChain();
420 SDValue Ptr = LD->getBasePtr();
421 EVT VT = LD->getValueType(0);
422 EVT LoadedVT = LD->getMemoryVT();
424 if (VT.isFloatingPoint() || VT.isVector()) {
425 EVT intVT = EVT::getIntegerVT(*DAG.getContext(), LoadedVT.getSizeInBits());
426 if (TLI.isTypeLegal(intVT) && TLI.isTypeLegal(LoadedVT)) {
427 // Expand to a (misaligned) integer load of the same size,
428 // then bitconvert to floating point or vector.
429 SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr,
430 LD->getMemOperand());
431 SDValue Result = DAG.getNode(ISD::BITCAST, dl, LoadedVT, newLoad);
433 Result = DAG.getNode(VT.isFloatingPoint() ? ISD::FP_EXTEND :
434 ISD::ANY_EXTEND, dl, VT, Result);
441 // Copy the value to a (aligned) stack slot using (unaligned) integer
442 // loads and stores, then do a (aligned) load from the stack slot.
443 MVT RegVT = TLI.getRegisterType(*DAG.getContext(), intVT);
444 unsigned LoadedBytes = LoadedVT.getSizeInBits() / 8;
445 unsigned RegBytes = RegVT.getSizeInBits() / 8;
446 unsigned NumRegs = (LoadedBytes + RegBytes - 1) / RegBytes;
448 // Make sure the stack slot is also aligned for the register type.
449 SDValue StackBase = DAG.CreateStackTemporary(LoadedVT, RegVT);
451 SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy());
452 SmallVector<SDValue, 8> Stores;
453 SDValue StackPtr = StackBase;
456 // Do all but one copies using the full register width.
457 for (unsigned i = 1; i < NumRegs; i++) {
458 // Load one integer register's worth from the original location.
459 SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr,
460 LD->getPointerInfo().getWithOffset(Offset),
461 LD->isVolatile(), LD->isNonTemporal(),
463 MinAlign(LD->getAlignment(), Offset),
465 // Follow the load with a store to the stack slot. Remember the store.
466 Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr,
467 MachinePointerInfo(), false, false, 0));
468 // Increment the pointers.
470 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
471 StackPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), StackPtr,
475 // The last copy may be partial. Do an extending load.
476 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(),
477 8 * (LoadedBytes - Offset));
478 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr,
479 LD->getPointerInfo().getWithOffset(Offset),
480 MemVT, LD->isVolatile(),
483 MinAlign(LD->getAlignment(), Offset),
485 // Follow the load with a store to the stack slot. Remember the store.
486 // On big-endian machines this requires a truncating store to ensure
487 // that the bits end up in the right place.
488 Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr,
489 MachinePointerInfo(), MemVT,
492 // The order of the stores doesn't matter - say it with a TokenFactor.
493 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
495 // Finally, perform the original load only redirected to the stack slot.
496 Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase,
497 MachinePointerInfo(), LoadedVT, false,false, false,
500 // Callers expect a MERGE_VALUES node.
505 assert(LoadedVT.isInteger() && !LoadedVT.isVector() &&
506 "Unaligned load of unsupported type.");
508 // Compute the new VT that is half the size of the old one. This is an
510 unsigned NumBits = LoadedVT.getSizeInBits();
512 NewLoadedVT = EVT::getIntegerVT(*DAG.getContext(), NumBits/2);
515 unsigned Alignment = LD->getAlignment();
516 unsigned IncrementSize = NumBits / 8;
517 ISD::LoadExtType HiExtType = LD->getExtensionType();
519 // If the original load is NON_EXTLOAD, the hi part load must be ZEXTLOAD.
520 if (HiExtType == ISD::NON_EXTLOAD)
521 HiExtType = ISD::ZEXTLOAD;
523 // Load the value in two parts
525 if (TLI.isLittleEndian()) {
526 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getPointerInfo(),
527 NewLoadedVT, LD->isVolatile(),
528 LD->isNonTemporal(), LD->isInvariant(), Alignment,
530 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
531 DAG.getConstant(IncrementSize, Ptr.getValueType()));
532 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr,
533 LD->getPointerInfo().getWithOffset(IncrementSize),
534 NewLoadedVT, LD->isVolatile(),
535 LD->isNonTemporal(),LD->isInvariant(),
536 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
538 Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getPointerInfo(),
539 NewLoadedVT, LD->isVolatile(),
540 LD->isNonTemporal(), LD->isInvariant(), Alignment,
542 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
543 DAG.getConstant(IncrementSize, Ptr.getValueType()));
544 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr,
545 LD->getPointerInfo().getWithOffset(IncrementSize),
546 NewLoadedVT, LD->isVolatile(),
547 LD->isNonTemporal(), LD->isInvariant(),
548 MinAlign(Alignment, IncrementSize), LD->getAAInfo());
551 // aggregate the two parts
552 SDValue ShiftAmount = DAG.getConstant(NumBits,
553 TLI.getShiftAmountTy(Hi.getValueType()));
554 SDValue Result = DAG.getNode(ISD::SHL, dl, VT, Hi, ShiftAmount);
555 Result = DAG.getNode(ISD::OR, dl, VT, Result, Lo);
557 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
564 /// Some target cannot handle a variable insertion index for the
565 /// INSERT_VECTOR_ELT instruction. In this case, it
566 /// is necessary to spill the vector being inserted into to memory, perform
567 /// the insert there, and then read the result back.
568 SDValue SelectionDAGLegalize::
569 PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx,
575 // If the target doesn't support this, we have to spill the input vector
576 // to a temporary stack slot, update the element, then reload it. This is
577 // badness. We could also load the value into a vector register (either
578 // with a "move to register" or "extload into register" instruction, then
579 // permute it into place, if the idx is a constant and if the idx is
580 // supported by the target.
581 EVT VT = Tmp1.getValueType();
582 EVT EltVT = VT.getVectorElementType();
583 EVT IdxVT = Tmp3.getValueType();
584 EVT PtrVT = TLI.getPointerTy();
585 SDValue StackPtr = DAG.CreateStackTemporary(VT);
587 int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
590 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr,
591 MachinePointerInfo::getFixedStack(SPFI),
594 // Truncate or zero extend offset to target pointer type.
595 unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND;
596 Tmp3 = DAG.getNode(CastOpc, dl, PtrVT, Tmp3);
597 // Add the offset to the index.
598 unsigned EltSize = EltVT.getSizeInBits()/8;
599 Tmp3 = DAG.getNode(ISD::MUL, dl, IdxVT, Tmp3,DAG.getConstant(EltSize, IdxVT));
600 SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr);
601 // Store the scalar value.
602 Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, MachinePointerInfo(), EltVT,
604 // Load the updated vector.
605 return DAG.getLoad(VT, dl, Ch, StackPtr,
606 MachinePointerInfo::getFixedStack(SPFI), false, false,
611 SDValue SelectionDAGLegalize::
612 ExpandINSERT_VECTOR_ELT(SDValue Vec, SDValue Val, SDValue Idx, SDLoc dl) {
613 if (ConstantSDNode *InsertPos = dyn_cast<ConstantSDNode>(Idx)) {
614 // SCALAR_TO_VECTOR requires that the type of the value being inserted
615 // match the element type of the vector being created, except for
616 // integers in which case the inserted value can be over width.
617 EVT EltVT = Vec.getValueType().getVectorElementType();
618 if (Val.getValueType() == EltVT ||
619 (EltVT.isInteger() && Val.getValueType().bitsGE(EltVT))) {
620 SDValue ScVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
621 Vec.getValueType(), Val);
623 unsigned NumElts = Vec.getValueType().getVectorNumElements();
624 // We generate a shuffle of InVec and ScVec, so the shuffle mask
625 // should be 0,1,2,3,4,5... with the appropriate element replaced with
627 SmallVector<int, 8> ShufOps;
628 for (unsigned i = 0; i != NumElts; ++i)
629 ShufOps.push_back(i != InsertPos->getZExtValue() ? i : NumElts);
631 return DAG.getVectorShuffle(Vec.getValueType(), dl, Vec, ScVec,
635 return PerformInsertVectorEltInMemory(Vec, Val, Idx, dl);
638 SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) {
639 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
640 // FIXME: We shouldn't do this for TargetConstantFP's.
641 // FIXME: move this to the DAG Combiner! Note that we can't regress due
642 // to phase ordering between legalized code and the dag combiner. This
643 // probably means that we need to integrate dag combiner and legalizer
645 // We generally can't do this one for long doubles.
646 SDValue Chain = ST->getChain();
647 SDValue Ptr = ST->getBasePtr();
648 unsigned Alignment = ST->getAlignment();
649 bool isVolatile = ST->isVolatile();
650 bool isNonTemporal = ST->isNonTemporal();
651 AAMDNodes AAInfo = ST->getAAInfo();
653 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) {
654 if (CFP->getValueType(0) == MVT::f32 &&
655 TLI.isTypeLegal(MVT::i32)) {
656 SDValue Con = DAG.getConstant(CFP->getValueAPF().
657 bitcastToAPInt().zextOrTrunc(32),
659 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
660 isVolatile, isNonTemporal, Alignment, AAInfo);
663 if (CFP->getValueType(0) == MVT::f64) {
664 // If this target supports 64-bit registers, do a single 64-bit store.
665 if (TLI.isTypeLegal(MVT::i64)) {
666 SDValue Con = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
667 zextOrTrunc(64), MVT::i64);
668 return DAG.getStore(Chain, dl, Con, Ptr, ST->getPointerInfo(),
669 isVolatile, isNonTemporal, Alignment, AAInfo);
672 if (TLI.isTypeLegal(MVT::i32) && !ST->isVolatile()) {
673 // Otherwise, if the target supports 32-bit registers, use 2 32-bit
674 // stores. If the target supports neither 32- nor 64-bits, this
675 // xform is certainly not worth it.
676 const APInt &IntVal =CFP->getValueAPF().bitcastToAPInt();
677 SDValue Lo = DAG.getConstant(IntVal.trunc(32), MVT::i32);
678 SDValue Hi = DAG.getConstant(IntVal.lshr(32).trunc(32), MVT::i32);
679 if (TLI.isBigEndian()) std::swap(Lo, Hi);
681 Lo = DAG.getStore(Chain, dl, Lo, Ptr, ST->getPointerInfo(), isVolatile,
682 isNonTemporal, Alignment, AAInfo);
683 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
684 DAG.getConstant(4, Ptr.getValueType()));
685 Hi = DAG.getStore(Chain, dl, Hi, Ptr,
686 ST->getPointerInfo().getWithOffset(4),
687 isVolatile, isNonTemporal, MinAlign(Alignment, 4U),
690 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
694 return SDValue(nullptr, 0);
697 void SelectionDAGLegalize::LegalizeStoreOps(SDNode *Node) {
698 StoreSDNode *ST = cast<StoreSDNode>(Node);
699 SDValue Chain = ST->getChain();
700 SDValue Ptr = ST->getBasePtr();
703 unsigned Alignment = ST->getAlignment();
704 bool isVolatile = ST->isVolatile();
705 bool isNonTemporal = ST->isNonTemporal();
706 AAMDNodes AAInfo = ST->getAAInfo();
708 if (!ST->isTruncatingStore()) {
709 if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) {
710 ReplaceNode(ST, OptStore);
715 SDValue Value = ST->getValue();
716 MVT VT = Value.getSimpleValueType();
717 switch (TLI.getOperationAction(ISD::STORE, VT)) {
718 default: llvm_unreachable("This action is not supported yet!");
719 case TargetLowering::Legal: {
720 // If this is an unaligned store and the target doesn't support it,
722 unsigned AS = ST->getAddressSpace();
723 unsigned Align = ST->getAlignment();
724 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
725 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
726 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
727 if (Align < ABIAlignment)
728 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
732 case TargetLowering::Custom: {
733 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
735 ReplaceNode(SDValue(Node, 0), Res);
738 case TargetLowering::Promote: {
739 MVT NVT = TLI.getTypeToPromoteTo(ISD::STORE, VT);
740 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
741 "Can only promote stores to same size type");
742 Value = DAG.getNode(ISD::BITCAST, dl, NVT, Value);
744 DAG.getStore(Chain, dl, Value, Ptr,
745 ST->getPointerInfo(), isVolatile,
746 isNonTemporal, Alignment, AAInfo);
747 ReplaceNode(SDValue(Node, 0), Result);
754 SDValue Value = ST->getValue();
756 EVT StVT = ST->getMemoryVT();
757 unsigned StWidth = StVT.getSizeInBits();
759 if (StWidth != StVT.getStoreSizeInBits()) {
760 // Promote to a byte-sized store with upper bits zero if not
761 // storing an integral number of bytes. For example, promote
762 // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
763 EVT NVT = EVT::getIntegerVT(*DAG.getContext(),
764 StVT.getStoreSizeInBits());
765 Value = DAG.getZeroExtendInReg(Value, dl, StVT);
767 DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
768 NVT, isVolatile, isNonTemporal, Alignment, AAInfo);
769 ReplaceNode(SDValue(Node, 0), Result);
770 } else if (StWidth & (StWidth - 1)) {
771 // If not storing a power-of-2 number of bits, expand as two stores.
772 assert(!StVT.isVector() && "Unsupported truncstore!");
773 unsigned RoundWidth = 1 << Log2_32(StWidth);
774 assert(RoundWidth < StWidth);
775 unsigned ExtraWidth = StWidth - RoundWidth;
776 assert(ExtraWidth < RoundWidth);
777 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
778 "Store size not an integral number of bytes!");
779 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
780 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
782 unsigned IncrementSize;
784 if (TLI.isLittleEndian()) {
785 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 X, TRUNCSTORE@+2:i8 (srl X, 16)
786 // Store the bottom RoundWidth bits.
787 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
789 isVolatile, isNonTemporal, Alignment,
792 // Store the remaining ExtraWidth bits.
793 IncrementSize = RoundWidth / 8;
794 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
795 DAG.getConstant(IncrementSize, Ptr.getValueType()));
796 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
797 DAG.getConstant(RoundWidth,
798 TLI.getShiftAmountTy(Value.getValueType())));
799 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr,
800 ST->getPointerInfo().getWithOffset(IncrementSize),
801 ExtraVT, isVolatile, isNonTemporal,
802 MinAlign(Alignment, IncrementSize), AAInfo);
804 // Big endian - avoid unaligned stores.
805 // TRUNCSTORE:i24 X -> TRUNCSTORE:i16 (srl X, 8), TRUNCSTORE@+2:i8 X
806 // Store the top RoundWidth bits.
807 Hi = DAG.getNode(ISD::SRL, dl, Value.getValueType(), Value,
808 DAG.getConstant(ExtraWidth,
809 TLI.getShiftAmountTy(Value.getValueType())));
810 Hi = DAG.getTruncStore(Chain, dl, Hi, Ptr, ST->getPointerInfo(),
811 RoundVT, isVolatile, isNonTemporal, Alignment,
814 // Store the remaining ExtraWidth bits.
815 IncrementSize = RoundWidth / 8;
816 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
817 DAG.getConstant(IncrementSize, Ptr.getValueType()));
818 Lo = DAG.getTruncStore(Chain, dl, Value, Ptr,
819 ST->getPointerInfo().getWithOffset(IncrementSize),
820 ExtraVT, isVolatile, isNonTemporal,
821 MinAlign(Alignment, IncrementSize), AAInfo);
824 // The order of the stores doesn't matter.
825 SDValue Result = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi);
826 ReplaceNode(SDValue(Node, 0), Result);
828 switch (TLI.getTruncStoreAction(ST->getValue().getSimpleValueType(),
829 StVT.getSimpleVT())) {
830 default: llvm_unreachable("This action is not supported yet!");
831 case TargetLowering::Legal: {
832 unsigned AS = ST->getAddressSpace();
833 unsigned Align = ST->getAlignment();
834 // If this is an unaligned store and the target doesn't support it,
836 if (!TLI.allowsMisalignedMemoryAccesses(ST->getMemoryVT(), AS, Align)) {
837 Type *Ty = ST->getMemoryVT().getTypeForEVT(*DAG.getContext());
838 unsigned ABIAlignment= TLI.getDataLayout()->getABITypeAlignment(Ty);
839 if (Align < ABIAlignment)
840 ExpandUnalignedStore(cast<StoreSDNode>(Node), DAG, TLI, this);
844 case TargetLowering::Custom: {
845 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
847 ReplaceNode(SDValue(Node, 0), Res);
850 case TargetLowering::Expand:
851 assert(!StVT.isVector() &&
852 "Vector Stores are handled in LegalizeVectorOps");
854 // TRUNCSTORE:i16 i32 -> STORE i16
855 assert(TLI.isTypeLegal(StVT) &&
856 "Do not know how to expand this store!");
857 Value = DAG.getNode(ISD::TRUNCATE, dl, StVT, Value);
859 DAG.getStore(Chain, dl, Value, Ptr, ST->getPointerInfo(),
860 isVolatile, isNonTemporal, Alignment, AAInfo);
861 ReplaceNode(SDValue(Node, 0), Result);
868 void SelectionDAGLegalize::LegalizeLoadOps(SDNode *Node) {
869 LoadSDNode *LD = cast<LoadSDNode>(Node);
870 SDValue Chain = LD->getChain(); // The chain.
871 SDValue Ptr = LD->getBasePtr(); // The base pointer.
872 SDValue Value; // The value returned by the load op.
875 ISD::LoadExtType ExtType = LD->getExtensionType();
876 if (ExtType == ISD::NON_EXTLOAD) {
877 MVT VT = Node->getSimpleValueType(0);
878 SDValue RVal = SDValue(Node, 0);
879 SDValue RChain = SDValue(Node, 1);
881 switch (TLI.getOperationAction(Node->getOpcode(), VT)) {
882 default: llvm_unreachable("This action is not supported yet!");
883 case TargetLowering::Legal: {
884 unsigned AS = LD->getAddressSpace();
885 unsigned Align = LD->getAlignment();
886 // If this is an unaligned load and the target doesn't support it,
888 if (!TLI.allowsMisalignedMemoryAccesses(LD->getMemoryVT(), AS, Align)) {
889 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
890 unsigned ABIAlignment =
891 TLI.getDataLayout()->getABITypeAlignment(Ty);
892 if (Align < ABIAlignment){
893 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, RVal, RChain);
898 case TargetLowering::Custom: {
899 SDValue Res = TLI.LowerOperation(RVal, DAG);
902 RChain = Res.getValue(1);
906 case TargetLowering::Promote: {
907 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), VT);
908 assert(NVT.getSizeInBits() == VT.getSizeInBits() &&
909 "Can only promote loads to same size type");
911 SDValue Res = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getMemOperand());
912 RVal = DAG.getNode(ISD::BITCAST, dl, VT, Res);
913 RChain = Res.getValue(1);
917 if (RChain.getNode() != Node) {
918 assert(RVal.getNode() != Node && "Load must be completely replaced");
919 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), RVal);
920 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), RChain);
922 UpdatedNodes->insert(RVal.getNode());
923 UpdatedNodes->insert(RChain.getNode());
930 EVT SrcVT = LD->getMemoryVT();
931 unsigned SrcWidth = SrcVT.getSizeInBits();
932 unsigned Alignment = LD->getAlignment();
933 bool isVolatile = LD->isVolatile();
934 bool isNonTemporal = LD->isNonTemporal();
935 bool isInvariant = LD->isInvariant();
936 AAMDNodes AAInfo = LD->getAAInfo();
938 if (SrcWidth != SrcVT.getStoreSizeInBits() &&
939 // Some targets pretend to have an i1 loading operation, and actually
940 // load an i8. This trick is correct for ZEXTLOAD because the top 7
941 // bits are guaranteed to be zero; it helps the optimizers understand
942 // that these bits are zero. It is also useful for EXTLOAD, since it
943 // tells the optimizers that those bits are undefined. It would be
944 // nice to have an effective generic way of getting these benefits...
945 // Until such a way is found, don't insist on promoting i1 here.
947 TLI.getLoadExtAction(ExtType, Node->getValueType(0), MVT::i1) ==
948 TargetLowering::Promote)) {
949 // Promote to a byte-sized load if not loading an integral number of
950 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
951 unsigned NewWidth = SrcVT.getStoreSizeInBits();
952 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
955 // The extra bits are guaranteed to be zero, since we stored them that
956 // way. A zext load from NVT thus automatically gives zext from SrcVT.
958 ISD::LoadExtType NewExtType =
959 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
962 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
963 Chain, Ptr, LD->getPointerInfo(),
964 NVT, isVolatile, isNonTemporal, isInvariant, Alignment,
967 Ch = Result.getValue(1); // The chain.
969 if (ExtType == ISD::SEXTLOAD)
970 // Having the top bits zero doesn't help when sign extending.
971 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
972 Result.getValueType(),
973 Result, DAG.getValueType(SrcVT));
974 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
975 // All the top bits are guaranteed to be zero - inform the optimizers.
976 Result = DAG.getNode(ISD::AssertZext, dl,
977 Result.getValueType(), Result,
978 DAG.getValueType(SrcVT));
982 } else if (SrcWidth & (SrcWidth - 1)) {
983 // If not loading a power-of-2 number of bits, expand as two loads.
984 assert(!SrcVT.isVector() && "Unsupported extload!");
985 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
986 assert(RoundWidth < SrcWidth);
987 unsigned ExtraWidth = SrcWidth - RoundWidth;
988 assert(ExtraWidth < RoundWidth);
989 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
990 "Load size not an integral number of bytes!");
991 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
992 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
994 unsigned IncrementSize;
996 if (TLI.isLittleEndian()) {
997 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
998 // Load the bottom RoundWidth bits.
999 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1001 LD->getPointerInfo(), RoundVT, isVolatile,
1002 isNonTemporal, isInvariant, Alignment, AAInfo);
1004 // Load the remaining ExtraWidth bits.
1005 IncrementSize = RoundWidth / 8;
1006 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1007 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1008 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1009 LD->getPointerInfo().getWithOffset(IncrementSize),
1010 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1011 MinAlign(Alignment, IncrementSize), AAInfo);
1013 // Build a factor node to remember that this load is independent of
1015 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1018 // Move the top bits to the right place.
1019 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1020 DAG.getConstant(RoundWidth,
1021 TLI.getShiftAmountTy(Hi.getValueType())));
1023 // Join the hi and lo parts.
1024 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1026 // Big endian - avoid unaligned loads.
1027 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1028 // Load the top RoundWidth bits.
1029 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1030 LD->getPointerInfo(), RoundVT, isVolatile,
1031 isNonTemporal, isInvariant, Alignment, AAInfo);
1033 // Load the remaining ExtraWidth bits.
1034 IncrementSize = RoundWidth / 8;
1035 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1036 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1037 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1038 dl, Node->getValueType(0), Chain, Ptr,
1039 LD->getPointerInfo().getWithOffset(IncrementSize),
1040 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1041 MinAlign(Alignment, IncrementSize), AAInfo);
1043 // Build a factor node to remember that this load is independent of
1045 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1048 // Move the top bits to the right place.
1049 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1050 DAG.getConstant(ExtraWidth,
1051 TLI.getShiftAmountTy(Hi.getValueType())));
1053 // Join the hi and lo parts.
1054 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1059 bool isCustom = false;
1060 switch (TLI.getLoadExtAction(ExtType, Node->getValueType(0),
1061 SrcVT.getSimpleVT())) {
1062 default: llvm_unreachable("This action is not supported yet!");
1063 case TargetLowering::Custom:
1066 case TargetLowering::Legal: {
1067 Value = SDValue(Node, 0);
1068 Chain = SDValue(Node, 1);
1071 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1072 if (Res.getNode()) {
1074 Chain = Res.getValue(1);
1077 // If this is an unaligned load and the target doesn't support
1079 EVT MemVT = LD->getMemoryVT();
1080 unsigned AS = LD->getAddressSpace();
1081 unsigned Align = LD->getAlignment();
1082 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1083 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1084 unsigned ABIAlignment = TLI.getDataLayout()->getABITypeAlignment(Ty);
1085 if (Align < ABIAlignment){
1086 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, Value, Chain);
1092 case TargetLowering::Expand:
1093 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, Node->getValueType(0),
1094 SrcVT) && TLI.isTypeLegal(SrcVT)) {
1095 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr, LD->getMemOperand());
1099 ExtendOp = (SrcVT.isFloatingPoint() ?
1100 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1102 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1103 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1104 default: llvm_unreachable("Unexpected extend load type!");
1106 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1107 Chain = Load.getValue(1);
1111 assert(!SrcVT.isVector() &&
1112 "Vector Loads are handled in LegalizeVectorOps");
1114 // FIXME: This does not work for vectors on most targets. Sign-
1115 // and zero-extend operations are currently folded into extending
1116 // loads, whether they are legal or not, and then we end up here
1117 // without any support for legalizing them.
1118 assert(ExtType != ISD::EXTLOAD &&
1119 "EXTLOAD should always be supported!");
1120 // Turn the unsupported load into an EXTLOAD followed by an
1121 // explicit zero/sign extend inreg.
1122 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1123 Node->getValueType(0),
1125 LD->getMemOperand());
1127 if (ExtType == ISD::SEXTLOAD)
1128 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1129 Result.getValueType(),
1130 Result, DAG.getValueType(SrcVT));
1132 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1134 Chain = Result.getValue(1);
1139 // Since loads produce two values, make sure to remember that we legalized
1141 if (Chain.getNode() != Node) {
1142 assert(Value.getNode() != Node && "Load must be completely replaced");
1143 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1144 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1146 UpdatedNodes->insert(Value.getNode());
1147 UpdatedNodes->insert(Chain.getNode());
1153 /// Return a legal replacement for the given operation, with all legal operands.
1154 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1155 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1157 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1160 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1161 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1162 TargetLowering::TypeLegal &&
1163 "Unexpected illegal type!");
1165 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1166 assert((TLI.getTypeAction(*DAG.getContext(),
1167 Node->getOperand(i).getValueType()) ==
1168 TargetLowering::TypeLegal ||
1169 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1170 "Unexpected illegal type!");
1172 // Figure out the correct action; the way to query this varies by opcode
1173 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1174 bool SimpleFinishLegalizing = true;
1175 switch (Node->getOpcode()) {
1176 case ISD::INTRINSIC_W_CHAIN:
1177 case ISD::INTRINSIC_WO_CHAIN:
1178 case ISD::INTRINSIC_VOID:
1179 case ISD::STACKSAVE:
1180 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1183 Action = TLI.getOperationAction(Node->getOpcode(),
1184 Node->getValueType(0));
1185 if (Action != TargetLowering::Promote)
1186 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1188 case ISD::FP_TO_FP16:
1189 case ISD::SINT_TO_FP:
1190 case ISD::UINT_TO_FP:
1191 case ISD::EXTRACT_VECTOR_ELT:
1192 Action = TLI.getOperationAction(Node->getOpcode(),
1193 Node->getOperand(0).getValueType());
1195 case ISD::FP_ROUND_INREG:
1196 case ISD::SIGN_EXTEND_INREG: {
1197 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1198 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1201 case ISD::ATOMIC_STORE: {
1202 Action = TLI.getOperationAction(Node->getOpcode(),
1203 Node->getOperand(2).getValueType());
1206 case ISD::SELECT_CC:
1209 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1210 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1211 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1212 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1213 ISD::CondCode CCCode =
1214 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1215 Action = TLI.getCondCodeAction(CCCode, OpVT);
1216 if (Action == TargetLowering::Legal) {
1217 if (Node->getOpcode() == ISD::SELECT_CC)
1218 Action = TLI.getOperationAction(Node->getOpcode(),
1219 Node->getValueType(0));
1221 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1227 // FIXME: Model these properly. LOAD and STORE are complicated, and
1228 // STORE expects the unlegalized operand in some cases.
1229 SimpleFinishLegalizing = false;
1231 case ISD::CALLSEQ_START:
1232 case ISD::CALLSEQ_END:
1233 // FIXME: This shouldn't be necessary. These nodes have special properties
1234 // dealing with the recursive nature of legalization. Removing this
1235 // special case should be done as part of making LegalizeDAG non-recursive.
1236 SimpleFinishLegalizing = false;
1238 case ISD::EXTRACT_ELEMENT:
1239 case ISD::FLT_ROUNDS_:
1247 case ISD::MERGE_VALUES:
1248 case ISD::EH_RETURN:
1249 case ISD::FRAME_TO_ARGS_OFFSET:
1250 case ISD::EH_SJLJ_SETJMP:
1251 case ISD::EH_SJLJ_LONGJMP:
1252 // These operations lie about being legal: when they claim to be legal,
1253 // they should actually be expanded.
1254 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1255 if (Action == TargetLowering::Legal)
1256 Action = TargetLowering::Expand;
1258 case ISD::INIT_TRAMPOLINE:
1259 case ISD::ADJUST_TRAMPOLINE:
1260 case ISD::FRAMEADDR:
1261 case ISD::RETURNADDR:
1262 // These operations lie about being legal: when they claim to be legal,
1263 // they should actually be custom-lowered.
1264 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1265 if (Action == TargetLowering::Legal)
1266 Action = TargetLowering::Custom;
1268 case ISD::READ_REGISTER:
1269 case ISD::WRITE_REGISTER:
1270 // Named register is legal in the DAG, but blocked by register name
1271 // selection if not implemented by target (to chose the correct register)
1272 // They'll be converted to Copy(To/From)Reg.
1273 Action = TargetLowering::Legal;
1275 case ISD::DEBUGTRAP:
1276 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1277 if (Action == TargetLowering::Expand) {
1278 // replace ISD::DEBUGTRAP with ISD::TRAP
1280 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1281 Node->getOperand(0));
1282 ReplaceNode(Node, NewVal.getNode());
1283 LegalizeOp(NewVal.getNode());
1289 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1290 Action = TargetLowering::Legal;
1292 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1297 if (SimpleFinishLegalizing) {
1298 SDNode *NewNode = Node;
1299 switch (Node->getOpcode()) {
1306 // Legalizing shifts/rotates requires adjusting the shift amount
1307 // to the appropriate width.
1308 if (!Node->getOperand(1).getValueType().isVector()) {
1310 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1311 Node->getOperand(1));
1312 HandleSDNode Handle(SAO);
1313 LegalizeOp(SAO.getNode());
1314 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1318 case ISD::SRL_PARTS:
1319 case ISD::SRA_PARTS:
1320 case ISD::SHL_PARTS:
1321 // Legalizing shifts/rotates requires adjusting the shift amount
1322 // to the appropriate width.
1323 if (!Node->getOperand(2).getValueType().isVector()) {
1325 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1326 Node->getOperand(2));
1327 HandleSDNode Handle(SAO);
1328 LegalizeOp(SAO.getNode());
1329 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1330 Node->getOperand(1),
1336 if (NewNode != Node) {
1337 ReplaceNode(Node, NewNode);
1341 case TargetLowering::Legal:
1343 case TargetLowering::Custom: {
1344 // FIXME: The handling for custom lowering with multiple results is
1346 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1347 if (Res.getNode()) {
1348 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1351 if (Node->getNumValues() == 1) {
1352 // We can just directly replace this node with the lowered value.
1353 ReplaceNode(SDValue(Node, 0), Res);
1357 SmallVector<SDValue, 8> ResultVals;
1358 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1359 ResultVals.push_back(Res.getValue(i));
1360 ReplaceNode(Node, ResultVals.data());
1365 case TargetLowering::Expand:
1368 case TargetLowering::Promote:
1374 switch (Node->getOpcode()) {
1381 llvm_unreachable("Do not know how to legalize this operator!");
1383 case ISD::CALLSEQ_START:
1384 case ISD::CALLSEQ_END:
1387 return LegalizeLoadOps(Node);
1390 return LegalizeStoreOps(Node);
1395 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1396 SDValue Vec = Op.getOperand(0);
1397 SDValue Idx = Op.getOperand(1);
1400 // Before we generate a new store to a temporary stack slot, see if there is
1401 // already one that we can use. There often is because when we scalarize
1402 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1403 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1404 // the vector. If all are expanded here, we don't want one store per vector
1406 SDValue StackPtr, Ch;
1407 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1408 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1410 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1411 if (ST->isIndexed() || ST->isTruncatingStore() ||
1412 ST->getValue() != Vec)
1415 // Make sure that nothing else could have stored into the destination of
1417 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1420 StackPtr = ST->getBasePtr();
1421 Ch = SDValue(ST, 0);
1426 if (!Ch.getNode()) {
1427 // Store the value to a temporary stack slot, then LOAD the returned part.
1428 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1429 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1430 MachinePointerInfo(), false, false, 0);
1433 // Add the offset to the index.
1435 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1436 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1437 DAG.getConstant(EltSize, Idx.getValueType()));
1439 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1440 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1442 if (Op.getValueType().isVector())
1443 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1444 false, false, false, 0);
1445 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1446 MachinePointerInfo(),
1447 Vec.getValueType().getVectorElementType(),
1448 false, false, false, 0);
1451 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1452 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1454 SDValue Vec = Op.getOperand(0);
1455 SDValue Part = Op.getOperand(1);
1456 SDValue Idx = Op.getOperand(2);
1459 // Store the value to a temporary stack slot, then LOAD the returned part.
1461 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1462 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1463 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1465 // First store the whole vector.
1466 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1469 // Then store the inserted part.
1471 // Add the offset to the index.
1473 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1475 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1476 DAG.getConstant(EltSize, Idx.getValueType()));
1477 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1479 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1482 // Store the subvector.
1483 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr,
1484 MachinePointerInfo(), false, false, 0);
1486 // Finally, load the updated vector.
1487 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1488 false, false, false, 0);
1491 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1492 // We can't handle this case efficiently. Allocate a sufficiently
1493 // aligned object on the stack, store each element into it, then load
1494 // the result as a vector.
1495 // Create the stack frame object.
1496 EVT VT = Node->getValueType(0);
1497 EVT EltVT = VT.getVectorElementType();
1499 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1500 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1501 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1503 // Emit a store of each element to the stack slot.
1504 SmallVector<SDValue, 8> Stores;
1505 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1506 // Store (in the right endianness) the elements to memory.
1507 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1508 // Ignore undef elements.
1509 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1511 unsigned Offset = TypeByteSize*i;
1513 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1514 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1516 // If the destination vector element type is narrower than the source
1517 // element type, only store the bits necessary.
1518 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1519 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1520 Node->getOperand(i), Idx,
1521 PtrInfo.getWithOffset(Offset),
1522 EltVT, false, false, 0));
1524 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1525 Node->getOperand(i), Idx,
1526 PtrInfo.getWithOffset(Offset),
1531 if (!Stores.empty()) // Not all undef elements?
1532 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1534 StoreChain = DAG.getEntryNode();
1536 // Result is a load from the stack slot.
1537 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1538 false, false, false, 0);
1541 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1543 SDValue Tmp1 = Node->getOperand(0);
1544 SDValue Tmp2 = Node->getOperand(1);
1546 // Get the sign bit of the RHS. First obtain a value that has the same
1547 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1549 EVT FloatVT = Tmp2.getValueType();
1550 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1551 if (TLI.isTypeLegal(IVT)) {
1552 // Convert to an integer with the same sign bit.
1553 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1555 // Store the float to memory, then load the sign part out as an integer.
1556 MVT LoadTy = TLI.getPointerTy();
1557 // First create a temporary that is aligned for both the load and store.
1558 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1559 // Then store the float to it.
1561 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1563 if (TLI.isBigEndian()) {
1564 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1565 // Load out a legal integer with the same sign bit as the float.
1566 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1567 false, false, false, 0);
1568 } else { // Little endian
1569 SDValue LoadPtr = StackPtr;
1570 // The float may be wider than the integer we are going to load. Advance
1571 // the pointer so that the loaded integer will contain the sign bit.
1572 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1573 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1574 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1575 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1576 // Load a legal integer containing the sign bit.
1577 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1578 false, false, false, 0);
1579 // Move the sign bit to the top bit of the loaded integer.
1580 unsigned BitShift = LoadTy.getSizeInBits() -
1581 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1582 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1584 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1585 DAG.getConstant(BitShift,
1586 TLI.getShiftAmountTy(SignBit.getValueType())));
1589 // Now get the sign bit proper, by seeing whether the value is negative.
1590 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1591 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1593 // Get the absolute value of the result.
1594 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1595 // Select between the nabs and abs value based on the sign bit of
1597 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1598 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1602 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1603 SmallVectorImpl<SDValue> &Results) {
1604 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1605 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1606 " not tell us which reg is the stack pointer!");
1608 EVT VT = Node->getValueType(0);
1609 SDValue Tmp1 = SDValue(Node, 0);
1610 SDValue Tmp2 = SDValue(Node, 1);
1611 SDValue Tmp3 = Node->getOperand(2);
1612 SDValue Chain = Tmp1.getOperand(0);
1614 // Chain the dynamic stack allocation so that it doesn't modify the stack
1615 // pointer when other instructions are using the stack.
1616 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1619 SDValue Size = Tmp2.getOperand(1);
1620 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1621 Chain = SP.getValue(1);
1622 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1623 unsigned StackAlign =
1624 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1625 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1626 if (Align > StackAlign)
1627 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1628 DAG.getConstant(-(uint64_t)Align, VT));
1629 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1631 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1632 DAG.getIntPtrConstant(0, true), SDValue(),
1635 Results.push_back(Tmp1);
1636 Results.push_back(Tmp2);
1639 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1642 /// If the SETCC has been legalized using AND / OR, then the legalized node
1643 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1644 /// will be set to false.
1646 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1647 /// then the values of LHS and RHS will be swapped, CC will be set to the
1648 /// new condition, and NeedInvert will be set to false.
1650 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1651 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1652 /// will be set to true. The caller must invert the result of the SETCC with
1653 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1654 /// of a true/false result.
1656 /// \returns true if the SetCC has been legalized, false if it hasn't.
1657 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1658 SDValue &LHS, SDValue &RHS,
1662 MVT OpVT = LHS.getSimpleValueType();
1663 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1665 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1666 default: llvm_unreachable("Unknown condition code action!");
1667 case TargetLowering::Legal:
1670 case TargetLowering::Expand: {
1671 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1672 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1673 std::swap(LHS, RHS);
1674 CC = DAG.getCondCode(InvCC);
1677 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1680 default: llvm_unreachable("Don't know how to expand this condition!");
1682 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1683 == TargetLowering::Legal
1684 && "If SETO is expanded, SETOEQ must be legal!");
1685 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1687 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1688 == TargetLowering::Legal
1689 && "If SETUO is expanded, SETUNE must be legal!");
1690 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1703 // If we are floating point, assign and break, otherwise fall through.
1704 if (!OpVT.isInteger()) {
1705 // We can use the 4th bit to tell if we are the unordered
1706 // or ordered version of the opcode.
1707 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1708 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1709 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1712 // Fallthrough if we are unsigned integer.
1717 // We only support using the inverted operation, which is computed above
1718 // and not a different manner of supporting expanding these cases.
1719 llvm_unreachable("Don't know how to expand this condition!");
1722 // Try inverting the result of the inverse condition.
1723 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1724 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1725 CC = DAG.getCondCode(InvCC);
1729 // If inverting the condition didn't work then we have no means to expand
1731 llvm_unreachable("Don't know how to expand this condition!");
1734 SDValue SetCC1, SetCC2;
1735 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1736 // If we aren't the ordered or unorder operation,
1737 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1738 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1739 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1741 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1742 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1743 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1745 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1754 /// Emit a store/load combination to the stack. This stores
1755 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1756 /// a load from the stack slot to DestVT, extending it if needed.
1757 /// The resultant code need not be legal.
1758 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1762 // Create the stack frame object.
1764 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1765 getTypeForEVT(*DAG.getContext()));
1766 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1768 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1769 int SPFI = StackPtrFI->getIndex();
1770 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1772 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1773 unsigned SlotSize = SlotVT.getSizeInBits();
1774 unsigned DestSize = DestVT.getSizeInBits();
1775 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1776 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1778 // Emit a store to the stack slot. Use a truncstore if the input value is
1779 // later than DestVT.
1782 if (SrcSize > SlotSize)
1783 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1784 PtrInfo, SlotVT, false, false, SrcAlign);
1786 assert(SrcSize == SlotSize && "Invalid store");
1787 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1788 PtrInfo, false, false, SrcAlign);
1791 // Result is a load from the stack slot.
1792 if (SlotSize == DestSize)
1793 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1794 false, false, false, DestAlign);
1796 assert(SlotSize < DestSize && "Unknown extension!");
1797 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1798 PtrInfo, SlotVT, false, false, false, DestAlign);
1801 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1803 // Create a vector sized/aligned stack slot, store the value to element #0,
1804 // then load the whole vector back out.
1805 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1807 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1808 int SPFI = StackPtrFI->getIndex();
1810 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1812 MachinePointerInfo::getFixedStack(SPFI),
1813 Node->getValueType(0).getVectorElementType(),
1815 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1816 MachinePointerInfo::getFixedStack(SPFI),
1817 false, false, false, 0);
1821 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1822 const TargetLowering &TLI, SDValue &Res) {
1823 unsigned NumElems = Node->getNumOperands();
1825 EVT VT = Node->getValueType(0);
1827 // Try to group the scalars into pairs, shuffle the pairs together, then
1828 // shuffle the pairs of pairs together, etc. until the vector has
1829 // been built. This will work only if all of the necessary shuffle masks
1832 // We do this in two phases; first to check the legality of the shuffles,
1833 // and next, assuming that all shuffles are legal, to create the new nodes.
1834 for (int Phase = 0; Phase < 2; ++Phase) {
1835 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1837 for (unsigned i = 0; i < NumElems; ++i) {
1838 SDValue V = Node->getOperand(i);
1839 if (V.getOpcode() == ISD::UNDEF)
1844 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1845 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1848 while (IntermedVals.size() > 2) {
1849 NewIntermedVals.clear();
1850 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1851 // This vector and the next vector are shuffled together (simply to
1852 // append the one to the other).
1853 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1855 SmallVector<int, 16> FinalIndices;
1856 FinalIndices.reserve(IntermedVals[i].second.size() +
1857 IntermedVals[i+1].second.size());
1860 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1863 FinalIndices.push_back(IntermedVals[i].second[j]);
1865 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1867 ShuffleVec[k] = NumElems + j;
1868 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1873 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1874 IntermedVals[i+1].first,
1876 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1878 NewIntermedVals.push_back(
1879 std::make_pair(Shuffle, std::move(FinalIndices)));
1882 // If we had an odd number of defined values, then append the last
1883 // element to the array of new vectors.
1884 if ((IntermedVals.size() & 1) != 0)
1885 NewIntermedVals.push_back(IntermedVals.back());
1887 IntermedVals.swap(NewIntermedVals);
1890 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1891 "Invalid number of intermediate vectors");
1892 SDValue Vec1 = IntermedVals[0].first;
1894 if (IntermedVals.size() > 1)
1895 Vec2 = IntermedVals[1].first;
1897 Vec2 = DAG.getUNDEF(VT);
1899 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1900 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1901 ShuffleVec[IntermedVals[0].second[i]] = i;
1902 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1903 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1906 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1907 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1914 /// Expand a BUILD_VECTOR node on targets that don't
1915 /// support the operation, but do support the resultant vector type.
1916 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1917 unsigned NumElems = Node->getNumOperands();
1918 SDValue Value1, Value2;
1920 EVT VT = Node->getValueType(0);
1921 EVT OpVT = Node->getOperand(0).getValueType();
1922 EVT EltVT = VT.getVectorElementType();
1924 // If the only non-undef value is the low element, turn this into a
1925 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1926 bool isOnlyLowElement = true;
1927 bool MoreThanTwoValues = false;
1928 bool isConstant = true;
1929 for (unsigned i = 0; i < NumElems; ++i) {
1930 SDValue V = Node->getOperand(i);
1931 if (V.getOpcode() == ISD::UNDEF)
1934 isOnlyLowElement = false;
1935 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1938 if (!Value1.getNode()) {
1940 } else if (!Value2.getNode()) {
1943 } else if (V != Value1 && V != Value2) {
1944 MoreThanTwoValues = true;
1948 if (!Value1.getNode())
1949 return DAG.getUNDEF(VT);
1951 if (isOnlyLowElement)
1952 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1954 // If all elements are constants, create a load from the constant pool.
1956 SmallVector<Constant*, 16> CV;
1957 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1958 if (ConstantFPSDNode *V =
1959 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1960 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1961 } else if (ConstantSDNode *V =
1962 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1964 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1966 // If OpVT and EltVT don't match, EltVT is not legal and the
1967 // element values have been promoted/truncated earlier. Undo this;
1968 // we don't want a v16i8 to become a v16i32 for example.
1969 const ConstantInt *CI = V->getConstantIntValue();
1970 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1971 CI->getZExtValue()));
1974 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1975 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1976 CV.push_back(UndefValue::get(OpNTy));
1979 Constant *CP = ConstantVector::get(CV);
1980 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1981 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1982 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1983 MachinePointerInfo::getConstantPool(),
1984 false, false, false, Alignment);
1987 SmallSet<SDValue, 16> DefinedValues;
1988 for (unsigned i = 0; i < NumElems; ++i) {
1989 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
1991 DefinedValues.insert(Node->getOperand(i));
1994 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1995 if (!MoreThanTwoValues) {
1996 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1997 for (unsigned i = 0; i < NumElems; ++i) {
1998 SDValue V = Node->getOperand(i);
1999 if (V.getOpcode() == ISD::UNDEF)
2001 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2003 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2004 // Get the splatted value into the low element of a vector register.
2005 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2007 if (Value2.getNode())
2008 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2010 Vec2 = DAG.getUNDEF(VT);
2012 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2013 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2017 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2022 // Otherwise, we can't handle this case efficiently.
2023 return ExpandVectorBuildThroughStack(Node);
2026 // Expand a node into a call to a libcall. If the result value
2027 // does not fit into a register, return the lo part and set the hi part to the
2028 // by-reg argument. If it does fit into a single register, return the result
2029 // and leave the Hi part unset.
2030 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2032 TargetLowering::ArgListTy Args;
2033 TargetLowering::ArgListEntry Entry;
2034 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2035 EVT ArgVT = Node->getOperand(i).getValueType();
2036 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2037 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2038 Entry.isSExt = isSigned;
2039 Entry.isZExt = !isSigned;
2040 Args.push_back(Entry);
2042 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2043 TLI.getPointerTy());
2045 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2047 // By default, the input chain to this libcall is the entry node of the
2048 // function. If the libcall is going to be emitted as a tail call then
2049 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2050 // node which is being folded has a non-entry input chain.
2051 SDValue InChain = DAG.getEntryNode();
2053 // isTailCall may be true since the callee does not reference caller stack
2054 // frame. Check if it's in the right position.
2055 SDValue TCChain = InChain;
2056 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2060 TargetLowering::CallLoweringInfo CLI(DAG);
2061 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2062 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2063 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2065 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2067 if (!CallInfo.second.getNode())
2068 // It's a tailcall, return the chain (which is the DAG root).
2069 return DAG.getRoot();
2071 return CallInfo.first;
2074 /// Generate a libcall taking the given operands as arguments
2075 /// and returning a result of type RetVT.
2076 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2077 const SDValue *Ops, unsigned NumOps,
2078 bool isSigned, SDLoc dl) {
2079 TargetLowering::ArgListTy Args;
2080 Args.reserve(NumOps);
2082 TargetLowering::ArgListEntry Entry;
2083 for (unsigned i = 0; i != NumOps; ++i) {
2084 Entry.Node = Ops[i];
2085 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2086 Entry.isSExt = isSigned;
2087 Entry.isZExt = !isSigned;
2088 Args.push_back(Entry);
2090 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2091 TLI.getPointerTy());
2093 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2095 TargetLowering::CallLoweringInfo CLI(DAG);
2096 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2097 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2098 .setSExtResult(isSigned).setZExtResult(!isSigned);
2100 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2102 return CallInfo.first;
2105 // Expand a node into a call to a libcall. Similar to
2106 // ExpandLibCall except that the first operand is the in-chain.
2107 std::pair<SDValue, SDValue>
2108 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2111 SDValue InChain = Node->getOperand(0);
2113 TargetLowering::ArgListTy Args;
2114 TargetLowering::ArgListEntry Entry;
2115 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2116 EVT ArgVT = Node->getOperand(i).getValueType();
2117 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2118 Entry.Node = Node->getOperand(i);
2120 Entry.isSExt = isSigned;
2121 Entry.isZExt = !isSigned;
2122 Args.push_back(Entry);
2124 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2125 TLI.getPointerTy());
2127 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2129 TargetLowering::CallLoweringInfo CLI(DAG);
2130 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2131 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2132 .setSExtResult(isSigned).setZExtResult(!isSigned);
2134 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2139 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2140 RTLIB::Libcall Call_F32,
2141 RTLIB::Libcall Call_F64,
2142 RTLIB::Libcall Call_F80,
2143 RTLIB::Libcall Call_F128,
2144 RTLIB::Libcall Call_PPCF128) {
2146 switch (Node->getSimpleValueType(0).SimpleTy) {
2147 default: llvm_unreachable("Unexpected request for libcall!");
2148 case MVT::f32: LC = Call_F32; break;
2149 case MVT::f64: LC = Call_F64; break;
2150 case MVT::f80: LC = Call_F80; break;
2151 case MVT::f128: LC = Call_F128; break;
2152 case MVT::ppcf128: LC = Call_PPCF128; break;
2154 return ExpandLibCall(LC, Node, false);
2157 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2158 RTLIB::Libcall Call_I8,
2159 RTLIB::Libcall Call_I16,
2160 RTLIB::Libcall Call_I32,
2161 RTLIB::Libcall Call_I64,
2162 RTLIB::Libcall Call_I128) {
2164 switch (Node->getSimpleValueType(0).SimpleTy) {
2165 default: llvm_unreachable("Unexpected request for libcall!");
2166 case MVT::i8: LC = Call_I8; break;
2167 case MVT::i16: LC = Call_I16; break;
2168 case MVT::i32: LC = Call_I32; break;
2169 case MVT::i64: LC = Call_I64; break;
2170 case MVT::i128: LC = Call_I128; break;
2172 return ExpandLibCall(LC, Node, isSigned);
2175 /// Return true if divmod libcall is available.
2176 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2177 const TargetLowering &TLI) {
2179 switch (Node->getSimpleValueType(0).SimpleTy) {
2180 default: llvm_unreachable("Unexpected request for libcall!");
2181 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2182 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2183 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2184 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2185 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2188 return TLI.getLibcallName(LC) != nullptr;
2191 /// Only issue divrem libcall if both quotient and remainder are needed.
2192 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2193 // The other use might have been replaced with a divrem already.
2194 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2195 unsigned OtherOpcode = 0;
2197 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2199 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2201 SDValue Op0 = Node->getOperand(0);
2202 SDValue Op1 = Node->getOperand(1);
2203 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2204 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2208 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2209 User->getOperand(0) == Op0 &&
2210 User->getOperand(1) == Op1)
2216 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2218 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2219 SmallVectorImpl<SDValue> &Results) {
2220 unsigned Opcode = Node->getOpcode();
2221 bool isSigned = Opcode == ISD::SDIVREM;
2224 switch (Node->getSimpleValueType(0).SimpleTy) {
2225 default: llvm_unreachable("Unexpected request for libcall!");
2226 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2227 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2228 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2229 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2230 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2233 // The input chain to this libcall is the entry node of the function.
2234 // Legalizing the call will automatically add the previous call to the
2236 SDValue InChain = DAG.getEntryNode();
2238 EVT RetVT = Node->getValueType(0);
2239 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2241 TargetLowering::ArgListTy Args;
2242 TargetLowering::ArgListEntry Entry;
2243 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2244 EVT ArgVT = Node->getOperand(i).getValueType();
2245 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2246 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2247 Entry.isSExt = isSigned;
2248 Entry.isZExt = !isSigned;
2249 Args.push_back(Entry);
2252 // Also pass the return address of the remainder.
2253 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2255 Entry.Ty = RetTy->getPointerTo();
2256 Entry.isSExt = isSigned;
2257 Entry.isZExt = !isSigned;
2258 Args.push_back(Entry);
2260 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2261 TLI.getPointerTy());
2264 TargetLowering::CallLoweringInfo CLI(DAG);
2265 CLI.setDebugLoc(dl).setChain(InChain)
2266 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2267 .setSExtResult(isSigned).setZExtResult(!isSigned);
2269 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2271 // Remainder is loaded back from the stack frame.
2272 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2273 MachinePointerInfo(), false, false, false, 0);
2274 Results.push_back(CallInfo.first);
2275 Results.push_back(Rem);
2278 /// Return true if sincos libcall is available.
2279 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2281 switch (Node->getSimpleValueType(0).SimpleTy) {
2282 default: llvm_unreachable("Unexpected request for libcall!");
2283 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2284 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2285 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2286 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2287 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2289 return TLI.getLibcallName(LC) != nullptr;
2292 /// Return true if sincos libcall is available and can be used to combine sin
2294 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2295 const TargetMachine &TM) {
2296 if (!isSinCosLibcallAvailable(Node, TLI))
2298 // GNU sin/cos functions set errno while sincos does not. Therefore
2299 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2300 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2301 if (isGNU && !TM.Options.UnsafeFPMath)
2306 /// Only issue sincos libcall if both sin and cos are needed.
2307 static bool useSinCos(SDNode *Node) {
2308 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2309 ? ISD::FCOS : ISD::FSIN;
2311 SDValue Op0 = Node->getOperand(0);
2312 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2313 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2317 // The other user might have been turned into sincos already.
2318 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2324 /// Issue libcalls to sincos to compute sin / cos pairs.
2326 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2327 SmallVectorImpl<SDValue> &Results) {
2329 switch (Node->getSimpleValueType(0).SimpleTy) {
2330 default: llvm_unreachable("Unexpected request for libcall!");
2331 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2332 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2333 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2334 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2335 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2338 // The input chain to this libcall is the entry node of the function.
2339 // Legalizing the call will automatically add the previous call to the
2341 SDValue InChain = DAG.getEntryNode();
2343 EVT RetVT = Node->getValueType(0);
2344 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2346 TargetLowering::ArgListTy Args;
2347 TargetLowering::ArgListEntry Entry;
2349 // Pass the argument.
2350 Entry.Node = Node->getOperand(0);
2352 Entry.isSExt = false;
2353 Entry.isZExt = false;
2354 Args.push_back(Entry);
2356 // Pass the return address of sin.
2357 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2358 Entry.Node = SinPtr;
2359 Entry.Ty = RetTy->getPointerTo();
2360 Entry.isSExt = false;
2361 Entry.isZExt = false;
2362 Args.push_back(Entry);
2364 // Also pass the return address of the cos.
2365 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2366 Entry.Node = CosPtr;
2367 Entry.Ty = RetTy->getPointerTo();
2368 Entry.isSExt = false;
2369 Entry.isZExt = false;
2370 Args.push_back(Entry);
2372 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2373 TLI.getPointerTy());
2376 TargetLowering::CallLoweringInfo CLI(DAG);
2377 CLI.setDebugLoc(dl).setChain(InChain)
2378 .setCallee(TLI.getLibcallCallingConv(LC),
2379 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2381 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2383 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2384 MachinePointerInfo(), false, false, false, 0));
2385 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2386 MachinePointerInfo(), false, false, false, 0));
2389 /// This function is responsible for legalizing a
2390 /// INT_TO_FP operation of the specified operand when the target requests that
2391 /// we expand it. At this point, we know that the result and operand types are
2392 /// legal for the target.
2393 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2397 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2398 // simple 32-bit [signed|unsigned] integer to float/double expansion
2400 // Get the stack frame index of a 8 byte buffer.
2401 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2403 // word offset constant for Hi/Lo address computation
2404 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2405 // set up Hi and Lo (into buffer) address based on endian
2406 SDValue Hi = StackSlot;
2407 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2408 StackSlot, WordOff);
2409 if (TLI.isLittleEndian())
2412 // if signed map to unsigned space
2415 // constant used to invert sign bit (signed to unsigned mapping)
2416 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2417 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2421 // store the lo of the constructed double - based on integer input
2422 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2423 Op0Mapped, Lo, MachinePointerInfo(),
2425 // initial hi portion of constructed double
2426 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2427 // store the hi of the constructed double - biased exponent
2428 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2429 MachinePointerInfo(),
2431 // load the constructed double
2432 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2433 MachinePointerInfo(), false, false, false, 0);
2434 // FP constant to bias correct the final result
2435 SDValue Bias = DAG.getConstantFP(isSigned ?
2436 BitsToDouble(0x4330000080000000ULL) :
2437 BitsToDouble(0x4330000000000000ULL),
2439 // subtract the bias
2440 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2443 // handle final rounding
2444 if (DestVT == MVT::f64) {
2447 } else if (DestVT.bitsLT(MVT::f64)) {
2448 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2449 DAG.getIntPtrConstant(0));
2450 } else if (DestVT.bitsGT(MVT::f64)) {
2451 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2455 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2456 // Code below here assumes !isSigned without checking again.
2458 // Implementation of unsigned i64 to f64 following the algorithm in
2459 // __floatundidf in compiler_rt. This implementation has the advantage
2460 // of performing rounding correctly, both in the default rounding mode
2461 // and in all alternate rounding modes.
2462 // TODO: Generalize this for use with other types.
2463 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2465 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2466 SDValue TwoP84PlusTwoP52 =
2467 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2469 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2471 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2472 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2473 DAG.getConstant(32, MVT::i64));
2474 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2475 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2476 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2477 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2478 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2480 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2483 // Implementation of unsigned i64 to f32.
2484 // TODO: Generalize this for use with other types.
2485 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2486 // For unsigned conversions, convert them to signed conversions using the
2487 // algorithm from the x86_64 __floatundidf in compiler_rt.
2489 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2491 SDValue ShiftConst =
2492 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2493 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2494 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2495 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2496 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2498 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2499 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2501 // TODO: This really should be implemented using a branch rather than a
2502 // select. We happen to get lucky and machinesink does the right
2503 // thing most of the time. This would be a good candidate for a
2504 //pseudo-op, or, even better, for whole-function isel.
2505 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2506 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2507 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2510 // Otherwise, implement the fully general conversion.
2512 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2513 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2514 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2515 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2516 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2517 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2518 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2519 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2520 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2521 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2522 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2524 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2525 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2527 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2528 DAG.getConstant(32, SHVT));
2529 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2530 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2532 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2533 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2534 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2535 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2536 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2537 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2538 DAG.getIntPtrConstant(0));
2541 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2543 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2544 Op0, DAG.getConstant(0, Op0.getValueType()),
2546 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2547 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2548 SignSet, Four, Zero);
2550 // If the sign bit of the integer is set, the large number will be treated
2551 // as a negative number. To counteract this, the dynamic code adds an
2552 // offset depending on the data type.
2554 switch (Op0.getSimpleValueType().SimpleTy) {
2555 default: llvm_unreachable("Unsupported integer type!");
2556 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2557 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2558 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2559 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2561 if (TLI.isLittleEndian()) FF <<= 32;
2562 Constant *FudgeFactor = ConstantInt::get(
2563 Type::getInt64Ty(*DAG.getContext()), FF);
2565 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2566 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2567 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2568 Alignment = std::min(Alignment, 4u);
2570 if (DestVT == MVT::f32)
2571 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2572 MachinePointerInfo::getConstantPool(),
2573 false, false, false, Alignment);
2575 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2576 DAG.getEntryNode(), CPIdx,
2577 MachinePointerInfo::getConstantPool(),
2578 MVT::f32, false, false, false, Alignment);
2579 HandleSDNode Handle(Load);
2580 LegalizeOp(Load.getNode());
2581 FudgeInReg = Handle.getValue();
2584 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2587 /// This function is responsible for legalizing a
2588 /// *INT_TO_FP operation of the specified operand when the target requests that
2589 /// we promote it. At this point, we know that the result and operand types are
2590 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2591 /// operation that takes a larger input.
2592 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2596 // First step, figure out the appropriate *INT_TO_FP operation to use.
2597 EVT NewInTy = LegalOp.getValueType();
2599 unsigned OpToUse = 0;
2601 // Scan for the appropriate larger type to use.
2603 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2604 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2606 // If the target supports SINT_TO_FP of this type, use it.
2607 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2608 OpToUse = ISD::SINT_TO_FP;
2611 if (isSigned) continue;
2613 // If the target supports UINT_TO_FP of this type, use it.
2614 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2615 OpToUse = ISD::UINT_TO_FP;
2619 // Otherwise, try a larger type.
2622 // Okay, we found the operation and type to use. Zero extend our input to the
2623 // desired type then run the operation on it.
2624 return DAG.getNode(OpToUse, dl, DestVT,
2625 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2626 dl, NewInTy, LegalOp));
2629 /// This function is responsible for legalizing a
2630 /// FP_TO_*INT operation of the specified operand when the target requests that
2631 /// we promote it. At this point, we know that the result and operand types are
2632 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2633 /// operation that returns a larger result.
2634 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2638 // First step, figure out the appropriate FP_TO*INT operation to use.
2639 EVT NewOutTy = DestVT;
2641 unsigned OpToUse = 0;
2643 // Scan for the appropriate larger type to use.
2645 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2646 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2648 // A larger signed type can hold all unsigned values of the requested type,
2649 // so using FP_TO_SINT is valid
2650 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2651 OpToUse = ISD::FP_TO_SINT;
2655 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2656 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2657 OpToUse = ISD::FP_TO_UINT;
2661 // Otherwise, try a larger type.
2665 // Okay, we found the operation and type to use.
2666 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2668 // Truncate the result of the extended FP_TO_*INT operation to the desired
2670 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2673 /// Open code the operations for BSWAP of the specified operation.
2674 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2675 EVT VT = Op.getValueType();
2676 EVT SHVT = TLI.getShiftAmountTy(VT);
2677 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2678 switch (VT.getSimpleVT().SimpleTy) {
2679 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2681 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2682 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2683 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2685 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2686 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2687 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2688 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2689 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2690 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2691 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2692 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2693 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2695 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2696 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2697 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2698 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2699 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2700 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2701 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2702 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2703 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2704 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2705 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2706 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2707 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2708 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2709 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2710 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2711 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2712 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2713 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2714 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2715 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2719 /// Expand the specified bitcount instruction into operations.
2720 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2723 default: llvm_unreachable("Cannot expand this yet!");
2725 EVT VT = Op.getValueType();
2726 EVT ShVT = TLI.getShiftAmountTy(VT);
2727 unsigned Len = VT.getSizeInBits();
2729 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2730 "CTPOP not implemented for this type.");
2732 // This is the "best" algorithm from
2733 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2735 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2736 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2737 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2738 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2740 // v = v - ((v >> 1) & 0x55555555...)
2741 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2742 DAG.getNode(ISD::AND, dl, VT,
2743 DAG.getNode(ISD::SRL, dl, VT, Op,
2744 DAG.getConstant(1, ShVT)),
2746 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2747 Op = DAG.getNode(ISD::ADD, dl, VT,
2748 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2749 DAG.getNode(ISD::AND, dl, VT,
2750 DAG.getNode(ISD::SRL, dl, VT, Op,
2751 DAG.getConstant(2, ShVT)),
2753 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2754 Op = DAG.getNode(ISD::AND, dl, VT,
2755 DAG.getNode(ISD::ADD, dl, VT, Op,
2756 DAG.getNode(ISD::SRL, dl, VT, Op,
2757 DAG.getConstant(4, ShVT))),
2759 // v = (v * 0x01010101...) >> (Len - 8)
2760 Op = DAG.getNode(ISD::SRL, dl, VT,
2761 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2762 DAG.getConstant(Len - 8, ShVT));
2766 case ISD::CTLZ_ZERO_UNDEF:
2767 // This trivially expands to CTLZ.
2768 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2770 // for now, we do this:
2771 // x = x | (x >> 1);
2772 // x = x | (x >> 2);
2774 // x = x | (x >>16);
2775 // x = x | (x >>32); // for 64-bit input
2776 // return popcount(~x);
2778 // Ref: "Hacker's Delight" by Henry Warren
2779 EVT VT = Op.getValueType();
2780 EVT ShVT = TLI.getShiftAmountTy(VT);
2781 unsigned len = VT.getSizeInBits();
2782 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2783 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2784 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2785 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2787 Op = DAG.getNOT(dl, Op, VT);
2788 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2790 case ISD::CTTZ_ZERO_UNDEF:
2791 // This trivially expands to CTTZ.
2792 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2794 // for now, we use: { return popcount(~x & (x - 1)); }
2795 // unless the target has ctlz but not ctpop, in which case we use:
2796 // { return 32 - nlz(~x & (x-1)); }
2797 // Ref: "Hacker's Delight" by Henry Warren
2798 EVT VT = Op.getValueType();
2799 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2800 DAG.getNOT(dl, Op, VT),
2801 DAG.getNode(ISD::SUB, dl, VT, Op,
2802 DAG.getConstant(1, VT)));
2803 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2804 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2805 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2806 return DAG.getNode(ISD::SUB, dl, VT,
2807 DAG.getConstant(VT.getSizeInBits(), VT),
2808 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2809 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2814 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2815 unsigned Opc = Node->getOpcode();
2816 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2821 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2822 case ISD::ATOMIC_SWAP:
2823 switch (VT.SimpleTy) {
2824 default: llvm_unreachable("Unexpected value type for atomic!");
2825 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2826 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2827 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2828 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2829 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2832 case ISD::ATOMIC_CMP_SWAP:
2833 switch (VT.SimpleTy) {
2834 default: llvm_unreachable("Unexpected value type for atomic!");
2835 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2836 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2837 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2838 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2839 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2842 case ISD::ATOMIC_LOAD_ADD:
2843 switch (VT.SimpleTy) {
2844 default: llvm_unreachable("Unexpected value type for atomic!");
2845 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2846 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2847 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2848 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2849 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2852 case ISD::ATOMIC_LOAD_SUB:
2853 switch (VT.SimpleTy) {
2854 default: llvm_unreachable("Unexpected value type for atomic!");
2855 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2856 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2857 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2858 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2859 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2862 case ISD::ATOMIC_LOAD_AND:
2863 switch (VT.SimpleTy) {
2864 default: llvm_unreachable("Unexpected value type for atomic!");
2865 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2866 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2867 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2868 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2869 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2872 case ISD::ATOMIC_LOAD_OR:
2873 switch (VT.SimpleTy) {
2874 default: llvm_unreachable("Unexpected value type for atomic!");
2875 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2876 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2877 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2878 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2879 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2882 case ISD::ATOMIC_LOAD_XOR:
2883 switch (VT.SimpleTy) {
2884 default: llvm_unreachable("Unexpected value type for atomic!");
2885 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2886 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2887 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2888 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2889 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2892 case ISD::ATOMIC_LOAD_NAND:
2893 switch (VT.SimpleTy) {
2894 default: llvm_unreachable("Unexpected value type for atomic!");
2895 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2896 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2897 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2898 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2899 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2902 case ISD::ATOMIC_LOAD_MAX:
2903 switch (VT.SimpleTy) {
2904 default: llvm_unreachable("Unexpected value type for atomic!");
2905 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2906 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2907 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2908 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2909 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2912 case ISD::ATOMIC_LOAD_UMAX:
2913 switch (VT.SimpleTy) {
2914 default: llvm_unreachable("Unexpected value type for atomic!");
2915 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2916 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2917 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2918 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2919 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2922 case ISD::ATOMIC_LOAD_MIN:
2923 switch (VT.SimpleTy) {
2924 default: llvm_unreachable("Unexpected value type for atomic!");
2925 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2926 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2927 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2928 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2929 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2932 case ISD::ATOMIC_LOAD_UMIN:
2933 switch (VT.SimpleTy) {
2934 default: llvm_unreachable("Unexpected value type for atomic!");
2935 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2936 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2937 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2938 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2939 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2944 return ExpandChainLibCall(LC, Node, false);
2947 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2948 SmallVector<SDValue, 8> Results;
2950 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2952 switch (Node->getOpcode()) {
2955 case ISD::CTLZ_ZERO_UNDEF:
2957 case ISD::CTTZ_ZERO_UNDEF:
2958 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2959 Results.push_back(Tmp1);
2962 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2964 case ISD::FRAMEADDR:
2965 case ISD::RETURNADDR:
2966 case ISD::FRAME_TO_ARGS_OFFSET:
2967 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2969 case ISD::FLT_ROUNDS_:
2970 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2972 case ISD::EH_RETURN:
2976 case ISD::EH_SJLJ_LONGJMP:
2977 // If the target didn't expand these, there's nothing to do, so just
2978 // preserve the chain and be done.
2979 Results.push_back(Node->getOperand(0));
2981 case ISD::EH_SJLJ_SETJMP:
2982 // If the target didn't expand this, just return 'zero' and preserve the
2984 Results.push_back(DAG.getConstant(0, MVT::i32));
2985 Results.push_back(Node->getOperand(0));
2987 case ISD::ATOMIC_FENCE: {
2988 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2989 // FIXME: handle "fence singlethread" more efficiently.
2990 TargetLowering::ArgListTy Args;
2992 TargetLowering::CallLoweringInfo CLI(DAG);
2993 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
2994 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2995 DAG.getExternalSymbol("__sync_synchronize",
2996 TLI.getPointerTy()), std::move(Args), 0);
2998 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3000 Results.push_back(CallResult.second);
3003 case ISD::ATOMIC_LOAD: {
3004 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3005 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
3006 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3007 SDValue Swap = DAG.getAtomicCmpSwap(
3008 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3009 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3010 cast<AtomicSDNode>(Node)->getMemOperand(),
3011 cast<AtomicSDNode>(Node)->getOrdering(),
3012 cast<AtomicSDNode>(Node)->getOrdering(),
3013 cast<AtomicSDNode>(Node)->getSynchScope());
3014 Results.push_back(Swap.getValue(0));
3015 Results.push_back(Swap.getValue(1));
3018 case ISD::ATOMIC_STORE: {
3019 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3020 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3021 cast<AtomicSDNode>(Node)->getMemoryVT(),
3022 Node->getOperand(0),
3023 Node->getOperand(1), Node->getOperand(2),
3024 cast<AtomicSDNode>(Node)->getMemOperand(),
3025 cast<AtomicSDNode>(Node)->getOrdering(),
3026 cast<AtomicSDNode>(Node)->getSynchScope());
3027 Results.push_back(Swap.getValue(1));
3030 // By default, atomic intrinsics are marked Legal and lowered. Targets
3031 // which don't support them directly, however, may want libcalls, in which
3032 // case they mark them Expand, and we get here.
3033 case ISD::ATOMIC_SWAP:
3034 case ISD::ATOMIC_LOAD_ADD:
3035 case ISD::ATOMIC_LOAD_SUB:
3036 case ISD::ATOMIC_LOAD_AND:
3037 case ISD::ATOMIC_LOAD_OR:
3038 case ISD::ATOMIC_LOAD_XOR:
3039 case ISD::ATOMIC_LOAD_NAND:
3040 case ISD::ATOMIC_LOAD_MIN:
3041 case ISD::ATOMIC_LOAD_MAX:
3042 case ISD::ATOMIC_LOAD_UMIN:
3043 case ISD::ATOMIC_LOAD_UMAX:
3044 case ISD::ATOMIC_CMP_SWAP: {
3045 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3046 Results.push_back(Tmp.first);
3047 Results.push_back(Tmp.second);
3050 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3051 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3052 // splits out the success value as a comparison. Expanding the resulting
3053 // ATOMIC_CMP_SWAP will produce a libcall.
3054 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3055 SDValue Res = DAG.getAtomicCmpSwap(
3056 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3057 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3058 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3059 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3060 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3061 cast<AtomicSDNode>(Node)->getSynchScope());
3063 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3064 Res, Node->getOperand(2), ISD::SETEQ);
3066 Results.push_back(Res.getValue(0));
3067 Results.push_back(Success);
3068 Results.push_back(Res.getValue(1));
3071 case ISD::DYNAMIC_STACKALLOC:
3072 ExpandDYNAMIC_STACKALLOC(Node, Results);
3074 case ISD::MERGE_VALUES:
3075 for (unsigned i = 0; i < Node->getNumValues(); i++)
3076 Results.push_back(Node->getOperand(i));
3079 EVT VT = Node->getValueType(0);
3081 Results.push_back(DAG.getConstant(0, VT));
3083 assert(VT.isFloatingPoint() && "Unknown value type!");
3084 Results.push_back(DAG.getConstantFP(0, VT));
3089 // If this operation is not supported, lower it to 'abort()' call
3090 TargetLowering::ArgListTy Args;
3091 TargetLowering::CallLoweringInfo CLI(DAG);
3092 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3093 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3094 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
3095 std::move(Args), 0);
3096 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3098 Results.push_back(CallResult.second);
3103 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3104 Node->getValueType(0), dl);
3105 Results.push_back(Tmp1);
3107 case ISD::FP_EXTEND:
3108 Tmp1 = EmitStackConvert(Node->getOperand(0),
3109 Node->getOperand(0).getValueType(),
3110 Node->getValueType(0), dl);
3111 Results.push_back(Tmp1);
3113 case ISD::SIGN_EXTEND_INREG: {
3114 // NOTE: we could fall back on load/store here too for targets without
3115 // SAR. However, it is doubtful that any exist.
3116 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3117 EVT VT = Node->getValueType(0);
3118 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3121 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3122 ExtraVT.getScalarType().getSizeInBits();
3123 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3124 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3125 Node->getOperand(0), ShiftCst);
3126 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3127 Results.push_back(Tmp1);
3130 case ISD::FP_ROUND_INREG: {
3131 // The only way we can lower this is to turn it into a TRUNCSTORE,
3132 // EXTLOAD pair, targeting a temporary location (a stack slot).
3134 // NOTE: there is a choice here between constantly creating new stack
3135 // slots and always reusing the same one. We currently always create
3136 // new ones, as reuse may inhibit scheduling.
3137 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3138 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3139 Node->getValueType(0), dl);
3140 Results.push_back(Tmp1);
3143 case ISD::SINT_TO_FP:
3144 case ISD::UINT_TO_FP:
3145 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3146 Node->getOperand(0), Node->getValueType(0), dl);
3147 Results.push_back(Tmp1);
3149 case ISD::FP_TO_SINT:
3150 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3151 Results.push_back(Tmp1);
3153 case ISD::FP_TO_UINT: {
3154 SDValue True, False;
3155 EVT VT = Node->getOperand(0).getValueType();
3156 EVT NVT = Node->getValueType(0);
3157 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3158 APInt::getNullValue(VT.getSizeInBits()));
3159 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3160 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3161 Tmp1 = DAG.getConstantFP(apf, VT);
3162 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3163 Node->getOperand(0),
3165 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3166 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3167 DAG.getNode(ISD::FSUB, dl, VT,
3168 Node->getOperand(0), Tmp1));
3169 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3170 DAG.getConstant(x, NVT));
3171 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3172 Results.push_back(Tmp1);
3176 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3177 EVT VT = Node->getValueType(0);
3178 Tmp1 = Node->getOperand(0);
3179 Tmp2 = Node->getOperand(1);
3180 unsigned Align = Node->getConstantOperandVal(3);
3182 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3183 MachinePointerInfo(V),
3184 false, false, false, 0);
3185 SDValue VAList = VAListLoad;
3187 if (Align > TLI.getMinStackArgumentAlignment()) {
3188 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3190 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3191 DAG.getConstant(Align - 1,
3192 VAList.getValueType()));
3194 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3195 DAG.getConstant(-(int64_t)Align,
3196 VAList.getValueType()));
3199 // Increment the pointer, VAList, to the next vaarg
3200 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3201 DAG.getConstant(TLI.getDataLayout()->
3202 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3203 VAList.getValueType()));
3204 // Store the incremented VAList to the legalized pointer
3205 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3206 MachinePointerInfo(V), false, false, 0);
3207 // Load the actual argument out of the pointer VAList
3208 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3209 false, false, false, 0));
3210 Results.push_back(Results[0].getValue(1));
3214 // This defaults to loading a pointer from the input and storing it to the
3215 // output, returning the chain.
3216 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3217 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3218 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3219 Node->getOperand(2), MachinePointerInfo(VS),
3220 false, false, false, 0);
3221 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3222 MachinePointerInfo(VD), false, false, 0);
3223 Results.push_back(Tmp1);
3226 case ISD::EXTRACT_VECTOR_ELT:
3227 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3228 // This must be an access of the only element. Return it.
3229 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3230 Node->getOperand(0));
3232 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3233 Results.push_back(Tmp1);
3235 case ISD::EXTRACT_SUBVECTOR:
3236 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3238 case ISD::INSERT_SUBVECTOR:
3239 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3241 case ISD::CONCAT_VECTORS: {
3242 Results.push_back(ExpandVectorBuildThroughStack(Node));
3245 case ISD::SCALAR_TO_VECTOR:
3246 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3248 case ISD::INSERT_VECTOR_ELT:
3249 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3250 Node->getOperand(1),
3251 Node->getOperand(2), dl));
3253 case ISD::VECTOR_SHUFFLE: {
3254 SmallVector<int, 32> NewMask;
3255 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3257 EVT VT = Node->getValueType(0);
3258 EVT EltVT = VT.getVectorElementType();
3259 SDValue Op0 = Node->getOperand(0);
3260 SDValue Op1 = Node->getOperand(1);
3261 if (!TLI.isTypeLegal(EltVT)) {
3263 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3265 // BUILD_VECTOR operands are allowed to be wider than the element type.
3266 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3268 if (NewEltVT.bitsLT(EltVT)) {
3270 // Convert shuffle node.
3271 // If original node was v4i64 and the new EltVT is i32,
3272 // cast operands to v8i32 and re-build the mask.
3274 // Calculate new VT, the size of the new VT should be equal to original.
3276 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3277 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3278 assert(NewVT.bitsEq(VT));
3280 // cast operands to new VT
3281 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3282 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3284 // Convert the shuffle mask
3285 unsigned int factor =
3286 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3288 // EltVT gets smaller
3291 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3293 for (unsigned fi = 0; fi < factor; ++fi)
3294 NewMask.push_back(Mask[i]);
3297 for (unsigned fi = 0; fi < factor; ++fi)
3298 NewMask.push_back(Mask[i]*factor+fi);
3306 unsigned NumElems = VT.getVectorNumElements();
3307 SmallVector<SDValue, 16> Ops;
3308 for (unsigned i = 0; i != NumElems; ++i) {
3310 Ops.push_back(DAG.getUNDEF(EltVT));
3313 unsigned Idx = Mask[i];
3315 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3317 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3319 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3321 DAG.getConstant(Idx - NumElems,
3322 TLI.getVectorIdxTy())));
3325 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3326 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3327 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3328 Results.push_back(Tmp1);
3331 case ISD::EXTRACT_ELEMENT: {
3332 EVT OpTy = Node->getOperand(0).getValueType();
3333 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3335 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3336 DAG.getConstant(OpTy.getSizeInBits()/2,
3337 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3338 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3341 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3342 Node->getOperand(0));
3344 Results.push_back(Tmp1);
3347 case ISD::STACKSAVE:
3348 // Expand to CopyFromReg if the target set
3349 // StackPointerRegisterToSaveRestore.
3350 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3351 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3352 Node->getValueType(0)));
3353 Results.push_back(Results[0].getValue(1));
3355 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3356 Results.push_back(Node->getOperand(0));
3359 case ISD::STACKRESTORE:
3360 // Expand to CopyToReg if the target set
3361 // StackPointerRegisterToSaveRestore.
3362 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3363 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3364 Node->getOperand(1)));
3366 Results.push_back(Node->getOperand(0));
3369 case ISD::FCOPYSIGN:
3370 Results.push_back(ExpandFCOPYSIGN(Node));
3373 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3374 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3375 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3376 Node->getOperand(0));
3377 Results.push_back(Tmp1);
3380 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3381 EVT VT = Node->getValueType(0);
3382 Tmp1 = Node->getOperand(0);
3383 Tmp2 = DAG.getConstantFP(0.0, VT);
3384 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3385 Tmp1, Tmp2, ISD::SETUGT);
3386 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3387 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3388 Results.push_back(Tmp1);
3392 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
3393 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
3394 RTLIB::FMIN_PPCF128));
3397 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
3398 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
3399 RTLIB::FMAX_PPCF128));
3402 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3403 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3404 RTLIB::SQRT_PPCF128));
3408 EVT VT = Node->getValueType(0);
3409 bool isSIN = Node->getOpcode() == ISD::FSIN;
3410 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3411 // fcos which share the same operand and both are used.
3412 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3413 canCombineSinCosLibcall(Node, TLI, TM))
3414 && useSinCos(Node)) {
3415 SDVTList VTs = DAG.getVTList(VT, VT);
3416 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3418 Tmp1 = Tmp1.getValue(1);
3419 Results.push_back(Tmp1);
3421 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3422 RTLIB::SIN_F80, RTLIB::SIN_F128,
3423 RTLIB::SIN_PPCF128));
3425 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3426 RTLIB::COS_F80, RTLIB::COS_F128,
3427 RTLIB::COS_PPCF128));
3432 // Expand into sincos libcall.
3433 ExpandSinCosLibCall(Node, Results);
3436 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3437 RTLIB::LOG_F80, RTLIB::LOG_F128,
3438 RTLIB::LOG_PPCF128));
3441 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3442 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3443 RTLIB::LOG2_PPCF128));
3446 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3447 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3448 RTLIB::LOG10_PPCF128));
3451 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3452 RTLIB::EXP_F80, RTLIB::EXP_F128,
3453 RTLIB::EXP_PPCF128));
3456 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3457 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3458 RTLIB::EXP2_PPCF128));
3461 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3462 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3463 RTLIB::TRUNC_PPCF128));
3466 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3467 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3468 RTLIB::FLOOR_PPCF128));
3471 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3472 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3473 RTLIB::CEIL_PPCF128));
3476 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3477 RTLIB::RINT_F80, RTLIB::RINT_F128,
3478 RTLIB::RINT_PPCF128));
3480 case ISD::FNEARBYINT:
3481 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3482 RTLIB::NEARBYINT_F64,
3483 RTLIB::NEARBYINT_F80,
3484 RTLIB::NEARBYINT_F128,
3485 RTLIB::NEARBYINT_PPCF128));
3488 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3492 RTLIB::ROUND_PPCF128));
3495 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3496 RTLIB::POWI_F80, RTLIB::POWI_F128,
3497 RTLIB::POWI_PPCF128));
3500 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3501 RTLIB::POW_F80, RTLIB::POW_F128,
3502 RTLIB::POW_PPCF128));
3505 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3506 RTLIB::DIV_F80, RTLIB::DIV_F128,
3507 RTLIB::DIV_PPCF128));
3510 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3511 RTLIB::REM_F80, RTLIB::REM_F128,
3512 RTLIB::REM_PPCF128));
3515 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3516 RTLIB::FMA_F80, RTLIB::FMA_F128,
3517 RTLIB::FMA_PPCF128));
3520 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
3521 RTLIB::ADD_F80, RTLIB::ADD_F128,
3522 RTLIB::ADD_PPCF128));
3525 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
3526 RTLIB::MUL_F80, RTLIB::MUL_F128,
3527 RTLIB::MUL_PPCF128));
3529 case ISD::FP16_TO_FP: {
3530 if (Node->getValueType(0) == MVT::f32) {
3531 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3535 // We can extend to types bigger than f32 in two steps without changing the
3536 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3537 // the option of emitting that before resorting to a libcall.
3539 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3541 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3544 case ISD::FP_TO_FP16: {
3546 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3547 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3548 Results.push_back(ExpandLibCall(LC, Node, false));
3551 case ISD::ConstantFP: {
3552 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3553 // Check to see if this FP immediate is already legal.
3554 // If this is a legal constant, turn it into a TargetConstantFP node.
3555 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3556 Results.push_back(ExpandConstantFP(CFP, true));
3560 EVT VT = Node->getValueType(0);
3561 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3562 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3563 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3564 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3565 Results.push_back(Tmp1);
3567 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
3568 RTLIB::SUB_F80, RTLIB::SUB_F128,
3569 RTLIB::SUB_PPCF128));
3574 EVT VT = Node->getValueType(0);
3575 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3576 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3577 "Don't know how to expand this subtraction!");
3578 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3579 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3580 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3581 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3586 EVT VT = Node->getValueType(0);
3587 bool isSigned = Node->getOpcode() == ISD::SREM;
3588 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3589 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3590 Tmp2 = Node->getOperand(0);
3591 Tmp3 = Node->getOperand(1);
3592 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3593 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3594 // If div is legal, it's better to do the normal expansion
3595 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3596 useDivRem(Node, isSigned, false))) {
3597 SDVTList VTs = DAG.getVTList(VT, VT);
3598 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3599 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3601 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3602 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3603 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3604 } else if (isSigned)
3605 Tmp1 = ExpandIntLibCall(Node, true,
3607 RTLIB::SREM_I16, RTLIB::SREM_I32,
3608 RTLIB::SREM_I64, RTLIB::SREM_I128);
3610 Tmp1 = ExpandIntLibCall(Node, false,
3612 RTLIB::UREM_I16, RTLIB::UREM_I32,
3613 RTLIB::UREM_I64, RTLIB::UREM_I128);
3614 Results.push_back(Tmp1);
3619 bool isSigned = Node->getOpcode() == ISD::SDIV;
3620 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3621 EVT VT = Node->getValueType(0);
3622 SDVTList VTs = DAG.getVTList(VT, VT);
3623 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3624 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3625 useDivRem(Node, isSigned, true)))
3626 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3627 Node->getOperand(1));
3629 Tmp1 = ExpandIntLibCall(Node, true,
3631 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3632 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3634 Tmp1 = ExpandIntLibCall(Node, false,
3636 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3637 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3638 Results.push_back(Tmp1);
3643 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3645 EVT VT = Node->getValueType(0);
3646 SDVTList VTs = DAG.getVTList(VT, VT);
3647 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3648 "If this wasn't legal, it shouldn't have been created!");
3649 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3650 Node->getOperand(1));
3651 Results.push_back(Tmp1.getValue(1));
3656 // Expand into divrem libcall
3657 ExpandDivRemLibCall(Node, Results);
3660 EVT VT = Node->getValueType(0);
3661 SDVTList VTs = DAG.getVTList(VT, VT);
3662 // See if multiply or divide can be lowered using two-result operations.
3663 // We just need the low half of the multiply; try both the signed
3664 // and unsigned forms. If the target supports both SMUL_LOHI and
3665 // UMUL_LOHI, form a preference by checking which forms of plain
3666 // MULH it supports.
3667 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3668 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3669 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3670 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3671 unsigned OpToUse = 0;
3672 if (HasSMUL_LOHI && !HasMULHS) {
3673 OpToUse = ISD::SMUL_LOHI;
3674 } else if (HasUMUL_LOHI && !HasMULHU) {
3675 OpToUse = ISD::UMUL_LOHI;
3676 } else if (HasSMUL_LOHI) {
3677 OpToUse = ISD::SMUL_LOHI;
3678 } else if (HasUMUL_LOHI) {
3679 OpToUse = ISD::UMUL_LOHI;
3682 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3683 Node->getOperand(1)));
3688 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3689 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3690 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3691 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3692 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3693 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3694 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3695 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3696 SDValue Shift = DAG.getConstant(HalfType.getSizeInBits(),
3697 TLI.getShiftAmountTy(HalfType));
3698 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3699 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3703 Tmp1 = ExpandIntLibCall(Node, false,
3705 RTLIB::MUL_I16, RTLIB::MUL_I32,
3706 RTLIB::MUL_I64, RTLIB::MUL_I128);
3707 Results.push_back(Tmp1);
3712 SDValue LHS = Node->getOperand(0);
3713 SDValue RHS = Node->getOperand(1);
3714 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3715 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3717 Results.push_back(Sum);
3718 EVT ResultType = Node->getValueType(1);
3719 EVT OType = getSetCCResultType(Node->getValueType(0));
3721 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3723 // LHSSign -> LHS >= 0
3724 // RHSSign -> RHS >= 0
3725 // SumSign -> Sum >= 0
3728 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3730 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3732 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3733 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3734 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3735 Node->getOpcode() == ISD::SADDO ?
3736 ISD::SETEQ : ISD::SETNE);
3738 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3739 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3741 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3742 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3747 SDValue LHS = Node->getOperand(0);
3748 SDValue RHS = Node->getOperand(1);
3749 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3750 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3752 Results.push_back(Sum);
3754 EVT ResultType = Node->getValueType(1);
3755 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3757 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3758 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3760 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3765 EVT VT = Node->getValueType(0);
3766 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3767 SDValue LHS = Node->getOperand(0);
3768 SDValue RHS = Node->getOperand(1);
3771 static const unsigned Ops[2][3] =
3772 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3773 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3774 bool isSigned = Node->getOpcode() == ISD::SMULO;
3775 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3776 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3777 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3778 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3779 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3781 TopHalf = BottomHalf.getValue(1);
3782 } else if (TLI.isTypeLegal(WideVT)) {
3783 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3784 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3785 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3786 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3787 DAG.getIntPtrConstant(0));
3788 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3789 DAG.getIntPtrConstant(1));
3791 // We can fall back to a libcall with an illegal type for the MUL if we
3792 // have a libcall big enough.
3793 // Also, we can fall back to a division in some cases, but that's a big
3794 // performance hit in the general case.
3795 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3796 if (WideVT == MVT::i16)
3797 LC = RTLIB::MUL_I16;
3798 else if (WideVT == MVT::i32)
3799 LC = RTLIB::MUL_I32;
3800 else if (WideVT == MVT::i64)
3801 LC = RTLIB::MUL_I64;
3802 else if (WideVT == MVT::i128)
3803 LC = RTLIB::MUL_I128;
3804 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3806 // The high part is obtained by SRA'ing all but one of the bits of low
3808 unsigned LoSize = VT.getSizeInBits();
3809 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3810 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3811 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3812 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3814 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3815 // pre-lowered to the correct types. This all depends upon WideVT not
3816 // being a legal type for the architecture and thus has to be split to
3818 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3819 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3820 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3821 DAG.getIntPtrConstant(0));
3822 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3823 DAG.getIntPtrConstant(1));
3824 // Ret is a node with an illegal type. Because such things are not
3825 // generally permitted during this phase of legalization, make sure the
3826 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been
3828 assert(Ret->use_empty() &&
3829 "Unexpected uses of illegally type from expanded lib call.");
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,
3890 false, false, false, 0);
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)));
4286 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4287 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4288 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4289 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4290 Tmp3, DAG.getIntPtrConstant(0)));
4297 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4298 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4299 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4300 Tmp2, DAG.getIntPtrConstant(0)));
4305 // Replace the original node with the legalized result.
4306 if (!Results.empty())
4307 ReplaceNode(Node, Results.data());
4310 /// This is the entry point for the file.
4311 void SelectionDAG::Legalize() {
4312 AssignTopologicalOrder();
4314 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4315 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4317 // Visit all the nodes. We start in topological order, so that we see
4318 // nodes with their original operands intact. Legalization can produce
4319 // new nodes which may themselves need to be legalized. Iterate until all
4320 // nodes have been legalized.
4322 bool AnyLegalized = false;
4323 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4327 if (N->use_empty() && N != getRoot().getNode()) {
4333 if (LegalizedNodes.insert(N).second) {
4334 AnyLegalized = true;
4335 Legalizer.LegalizeOp(N);
4337 if (N->use_empty() && N != getRoot().getNode()) {
4348 // Remove dead nodes now.
4352 bool SelectionDAG::LegalizeOp(SDNode *N,
4353 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4354 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4355 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4357 // Directly insert the node in question, and legalize it. This will recurse
4358 // as needed through operands.
4359 LegalizedNodes.insert(N);
4360 Legalizer.LegalizeOp(N);
4362 return LegalizedNodes.count(N);