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, 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, MVT::i1) == TargetLowering::Promote)) {
948 // Promote to a byte-sized load if not loading an integral number of
949 // bytes. For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
950 unsigned NewWidth = SrcVT.getStoreSizeInBits();
951 EVT NVT = EVT::getIntegerVT(*DAG.getContext(), NewWidth);
954 // The extra bits are guaranteed to be zero, since we stored them that
955 // way. A zext load from NVT thus automatically gives zext from SrcVT.
957 ISD::LoadExtType NewExtType =
958 ExtType == ISD::ZEXTLOAD ? ISD::ZEXTLOAD : ISD::EXTLOAD;
961 DAG.getExtLoad(NewExtType, dl, Node->getValueType(0),
962 Chain, Ptr, LD->getPointerInfo(),
963 NVT, isVolatile, isNonTemporal, isInvariant, Alignment,
966 Ch = Result.getValue(1); // The chain.
968 if (ExtType == ISD::SEXTLOAD)
969 // Having the top bits zero doesn't help when sign extending.
970 Result = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
971 Result.getValueType(),
972 Result, DAG.getValueType(SrcVT));
973 else if (ExtType == ISD::ZEXTLOAD || NVT == Result.getValueType())
974 // All the top bits are guaranteed to be zero - inform the optimizers.
975 Result = DAG.getNode(ISD::AssertZext, dl,
976 Result.getValueType(), Result,
977 DAG.getValueType(SrcVT));
981 } else if (SrcWidth & (SrcWidth - 1)) {
982 // If not loading a power-of-2 number of bits, expand as two loads.
983 assert(!SrcVT.isVector() && "Unsupported extload!");
984 unsigned RoundWidth = 1 << Log2_32(SrcWidth);
985 assert(RoundWidth < SrcWidth);
986 unsigned ExtraWidth = SrcWidth - RoundWidth;
987 assert(ExtraWidth < RoundWidth);
988 assert(!(RoundWidth % 8) && !(ExtraWidth % 8) &&
989 "Load size not an integral number of bytes!");
990 EVT RoundVT = EVT::getIntegerVT(*DAG.getContext(), RoundWidth);
991 EVT ExtraVT = EVT::getIntegerVT(*DAG.getContext(), ExtraWidth);
993 unsigned IncrementSize;
995 if (TLI.isLittleEndian()) {
996 // EXTLOAD:i24 -> ZEXTLOAD:i16 | (shl EXTLOAD@+2:i8, 16)
997 // Load the bottom RoundWidth bits.
998 Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0),
1000 LD->getPointerInfo(), RoundVT, isVolatile,
1001 isNonTemporal, isInvariant, Alignment, AAInfo);
1003 // Load the remaining ExtraWidth bits.
1004 IncrementSize = RoundWidth / 8;
1005 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1006 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1007 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1008 LD->getPointerInfo().getWithOffset(IncrementSize),
1009 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1010 MinAlign(Alignment, IncrementSize), AAInfo);
1012 // Build a factor node to remember that this load is independent of
1014 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1017 // Move the top bits to the right place.
1018 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1019 DAG.getConstant(RoundWidth,
1020 TLI.getShiftAmountTy(Hi.getValueType())));
1022 // Join the hi and lo parts.
1023 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1025 // Big endian - avoid unaligned loads.
1026 // EXTLOAD:i24 -> (shl EXTLOAD:i16, 8) | ZEXTLOAD@+2:i8
1027 // Load the top RoundWidth bits.
1028 Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Chain, Ptr,
1029 LD->getPointerInfo(), RoundVT, isVolatile,
1030 isNonTemporal, isInvariant, Alignment, AAInfo);
1032 // Load the remaining ExtraWidth bits.
1033 IncrementSize = RoundWidth / 8;
1034 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr,
1035 DAG.getConstant(IncrementSize, Ptr.getValueType()));
1036 Lo = DAG.getExtLoad(ISD::ZEXTLOAD,
1037 dl, Node->getValueType(0), Chain, Ptr,
1038 LD->getPointerInfo().getWithOffset(IncrementSize),
1039 ExtraVT, isVolatile, isNonTemporal, isInvariant,
1040 MinAlign(Alignment, IncrementSize), AAInfo);
1042 // Build a factor node to remember that this load is independent of
1044 Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
1047 // Move the top bits to the right place.
1048 Hi = DAG.getNode(ISD::SHL, dl, Hi.getValueType(), Hi,
1049 DAG.getConstant(ExtraWidth,
1050 TLI.getShiftAmountTy(Hi.getValueType())));
1052 // Join the hi and lo parts.
1053 Value = DAG.getNode(ISD::OR, dl, Node->getValueType(0), Lo, Hi);
1058 bool isCustom = false;
1059 switch (TLI.getLoadExtAction(ExtType, SrcVT.getSimpleVT())) {
1060 default: llvm_unreachable("This action is not supported yet!");
1061 case TargetLowering::Custom:
1064 case TargetLowering::Legal: {
1065 Value = SDValue(Node, 0);
1066 Chain = SDValue(Node, 1);
1069 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1070 if (Res.getNode()) {
1072 Chain = Res.getValue(1);
1075 // If this is an unaligned load and the target doesn't support
1077 EVT MemVT = LD->getMemoryVT();
1078 unsigned AS = LD->getAddressSpace();
1079 unsigned Align = LD->getAlignment();
1080 if (!TLI.allowsMisalignedMemoryAccesses(MemVT, AS, Align)) {
1081 Type *Ty = LD->getMemoryVT().getTypeForEVT(*DAG.getContext());
1082 unsigned ABIAlignment = TLI.getDataLayout()->getABITypeAlignment(Ty);
1083 if (Align < ABIAlignment){
1084 ExpandUnalignedLoad(cast<LoadSDNode>(Node), DAG, TLI, Value, Chain);
1090 case TargetLowering::Expand:
1091 if (!TLI.isLoadExtLegal(ISD::EXTLOAD, SrcVT) && TLI.isTypeLegal(SrcVT)) {
1092 SDValue Load = DAG.getLoad(SrcVT, dl, Chain, Ptr, LD->getMemOperand());
1096 ExtendOp = (SrcVT.isFloatingPoint() ?
1097 ISD::FP_EXTEND : ISD::ANY_EXTEND);
1099 case ISD::SEXTLOAD: ExtendOp = ISD::SIGN_EXTEND; break;
1100 case ISD::ZEXTLOAD: ExtendOp = ISD::ZERO_EXTEND; break;
1101 default: llvm_unreachable("Unexpected extend load type!");
1103 Value = DAG.getNode(ExtendOp, dl, Node->getValueType(0), Load);
1104 Chain = Load.getValue(1);
1108 assert(!SrcVT.isVector() &&
1109 "Vector Loads are handled in LegalizeVectorOps");
1111 // FIXME: This does not work for vectors on most targets. Sign-
1112 // and zero-extend operations are currently folded into extending
1113 // loads, whether they are legal or not, and then we end up here
1114 // without any support for legalizing them.
1115 assert(ExtType != ISD::EXTLOAD &&
1116 "EXTLOAD should always be supported!");
1117 // Turn the unsupported load into an EXTLOAD followed by an
1118 // explicit zero/sign extend inreg.
1119 SDValue Result = DAG.getExtLoad(ISD::EXTLOAD, dl,
1120 Node->getValueType(0),
1122 LD->getMemOperand());
1124 if (ExtType == ISD::SEXTLOAD)
1125 ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl,
1126 Result.getValueType(),
1127 Result, DAG.getValueType(SrcVT));
1129 ValRes = DAG.getZeroExtendInReg(Result, dl, SrcVT.getScalarType());
1131 Chain = Result.getValue(1);
1136 // Since loads produce two values, make sure to remember that we legalized
1138 if (Chain.getNode() != Node) {
1139 assert(Value.getNode() != Node && "Load must be completely replaced");
1140 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Value);
1141 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
1143 UpdatedNodes->insert(Value.getNode());
1144 UpdatedNodes->insert(Chain.getNode());
1150 /// Return a legal replacement for the given operation, with all legal operands.
1151 void SelectionDAGLegalize::LegalizeOp(SDNode *Node) {
1152 DEBUG(dbgs() << "\nLegalizing: "; Node->dump(&DAG));
1154 if (Node->getOpcode() == ISD::TargetConstant) // Allow illegal target nodes.
1157 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1158 assert(TLI.getTypeAction(*DAG.getContext(), Node->getValueType(i)) ==
1159 TargetLowering::TypeLegal &&
1160 "Unexpected illegal type!");
1162 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i)
1163 assert((TLI.getTypeAction(*DAG.getContext(),
1164 Node->getOperand(i).getValueType()) ==
1165 TargetLowering::TypeLegal ||
1166 Node->getOperand(i).getOpcode() == ISD::TargetConstant) &&
1167 "Unexpected illegal type!");
1169 // Figure out the correct action; the way to query this varies by opcode
1170 TargetLowering::LegalizeAction Action = TargetLowering::Legal;
1171 bool SimpleFinishLegalizing = true;
1172 switch (Node->getOpcode()) {
1173 case ISD::INTRINSIC_W_CHAIN:
1174 case ISD::INTRINSIC_WO_CHAIN:
1175 case ISD::INTRINSIC_VOID:
1176 case ISD::STACKSAVE:
1177 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1180 Action = TLI.getOperationAction(Node->getOpcode(),
1181 Node->getValueType(0));
1182 if (Action != TargetLowering::Promote)
1183 Action = TLI.getOperationAction(Node->getOpcode(), MVT::Other);
1185 case ISD::FP_TO_FP16:
1186 case ISD::SINT_TO_FP:
1187 case ISD::UINT_TO_FP:
1188 case ISD::EXTRACT_VECTOR_ELT:
1189 Action = TLI.getOperationAction(Node->getOpcode(),
1190 Node->getOperand(0).getValueType());
1192 case ISD::FP_ROUND_INREG:
1193 case ISD::SIGN_EXTEND_INREG: {
1194 EVT InnerType = cast<VTSDNode>(Node->getOperand(1))->getVT();
1195 Action = TLI.getOperationAction(Node->getOpcode(), InnerType);
1198 case ISD::ATOMIC_STORE: {
1199 Action = TLI.getOperationAction(Node->getOpcode(),
1200 Node->getOperand(2).getValueType());
1203 case ISD::SELECT_CC:
1206 unsigned CCOperand = Node->getOpcode() == ISD::SELECT_CC ? 4 :
1207 Node->getOpcode() == ISD::SETCC ? 2 : 1;
1208 unsigned CompareOperand = Node->getOpcode() == ISD::BR_CC ? 2 : 0;
1209 MVT OpVT = Node->getOperand(CompareOperand).getSimpleValueType();
1210 ISD::CondCode CCCode =
1211 cast<CondCodeSDNode>(Node->getOperand(CCOperand))->get();
1212 Action = TLI.getCondCodeAction(CCCode, OpVT);
1213 if (Action == TargetLowering::Legal) {
1214 if (Node->getOpcode() == ISD::SELECT_CC)
1215 Action = TLI.getOperationAction(Node->getOpcode(),
1216 Node->getValueType(0));
1218 Action = TLI.getOperationAction(Node->getOpcode(), OpVT);
1224 // FIXME: Model these properly. LOAD and STORE are complicated, and
1225 // STORE expects the unlegalized operand in some cases.
1226 SimpleFinishLegalizing = false;
1228 case ISD::CALLSEQ_START:
1229 case ISD::CALLSEQ_END:
1230 // FIXME: This shouldn't be necessary. These nodes have special properties
1231 // dealing with the recursive nature of legalization. Removing this
1232 // special case should be done as part of making LegalizeDAG non-recursive.
1233 SimpleFinishLegalizing = false;
1235 case ISD::EXTRACT_ELEMENT:
1236 case ISD::FLT_ROUNDS_:
1244 case ISD::MERGE_VALUES:
1245 case ISD::EH_RETURN:
1246 case ISD::FRAME_TO_ARGS_OFFSET:
1247 case ISD::EH_SJLJ_SETJMP:
1248 case ISD::EH_SJLJ_LONGJMP:
1249 // These operations lie about being legal: when they claim to be legal,
1250 // they should actually be expanded.
1251 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1252 if (Action == TargetLowering::Legal)
1253 Action = TargetLowering::Expand;
1255 case ISD::INIT_TRAMPOLINE:
1256 case ISD::ADJUST_TRAMPOLINE:
1257 case ISD::FRAMEADDR:
1258 case ISD::RETURNADDR:
1259 // These operations lie about being legal: when they claim to be legal,
1260 // they should actually be custom-lowered.
1261 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1262 if (Action == TargetLowering::Legal)
1263 Action = TargetLowering::Custom;
1265 case ISD::READ_REGISTER:
1266 case ISD::WRITE_REGISTER:
1267 // Named register is legal in the DAG, but blocked by register name
1268 // selection if not implemented by target (to chose the correct register)
1269 // They'll be converted to Copy(To/From)Reg.
1270 Action = TargetLowering::Legal;
1272 case ISD::DEBUGTRAP:
1273 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1274 if (Action == TargetLowering::Expand) {
1275 // replace ISD::DEBUGTRAP with ISD::TRAP
1277 NewVal = DAG.getNode(ISD::TRAP, SDLoc(Node), Node->getVTList(),
1278 Node->getOperand(0));
1279 ReplaceNode(Node, NewVal.getNode());
1280 LegalizeOp(NewVal.getNode());
1286 if (Node->getOpcode() >= ISD::BUILTIN_OP_END) {
1287 Action = TargetLowering::Legal;
1289 Action = TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0));
1294 if (SimpleFinishLegalizing) {
1295 SDNode *NewNode = Node;
1296 switch (Node->getOpcode()) {
1303 // Legalizing shifts/rotates requires adjusting the shift amount
1304 // to the appropriate width.
1305 if (!Node->getOperand(1).getValueType().isVector()) {
1307 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1308 Node->getOperand(1));
1309 HandleSDNode Handle(SAO);
1310 LegalizeOp(SAO.getNode());
1311 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1315 case ISD::SRL_PARTS:
1316 case ISD::SRA_PARTS:
1317 case ISD::SHL_PARTS:
1318 // Legalizing shifts/rotates requires adjusting the shift amount
1319 // to the appropriate width.
1320 if (!Node->getOperand(2).getValueType().isVector()) {
1322 DAG.getShiftAmountOperand(Node->getOperand(0).getValueType(),
1323 Node->getOperand(2));
1324 HandleSDNode Handle(SAO);
1325 LegalizeOp(SAO.getNode());
1326 NewNode = DAG.UpdateNodeOperands(Node, Node->getOperand(0),
1327 Node->getOperand(1),
1333 if (NewNode != Node) {
1334 ReplaceNode(Node, NewNode);
1338 case TargetLowering::Legal:
1340 case TargetLowering::Custom: {
1341 // FIXME: The handling for custom lowering with multiple results is
1343 SDValue Res = TLI.LowerOperation(SDValue(Node, 0), DAG);
1344 if (Res.getNode()) {
1345 if (!(Res.getNode() != Node || Res.getResNo() != 0))
1348 if (Node->getNumValues() == 1) {
1349 // We can just directly replace this node with the lowered value.
1350 ReplaceNode(SDValue(Node, 0), Res);
1354 SmallVector<SDValue, 8> ResultVals;
1355 for (unsigned i = 0, e = Node->getNumValues(); i != e; ++i)
1356 ResultVals.push_back(Res.getValue(i));
1357 ReplaceNode(Node, ResultVals.data());
1362 case TargetLowering::Expand:
1365 case TargetLowering::Promote:
1371 switch (Node->getOpcode()) {
1378 llvm_unreachable("Do not know how to legalize this operator!");
1380 case ISD::CALLSEQ_START:
1381 case ISD::CALLSEQ_END:
1384 return LegalizeLoadOps(Node);
1387 return LegalizeStoreOps(Node);
1392 SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) {
1393 SDValue Vec = Op.getOperand(0);
1394 SDValue Idx = Op.getOperand(1);
1397 // Before we generate a new store to a temporary stack slot, see if there is
1398 // already one that we can use. There often is because when we scalarize
1399 // vector operations (using SelectionDAG::UnrollVectorOp for example) a whole
1400 // series of EXTRACT_VECTOR_ELT nodes are generated, one for each element in
1401 // the vector. If all are expanded here, we don't want one store per vector
1403 SDValue StackPtr, Ch;
1404 for (SDNode::use_iterator UI = Vec.getNode()->use_begin(),
1405 UE = Vec.getNode()->use_end(); UI != UE; ++UI) {
1407 if (StoreSDNode *ST = dyn_cast<StoreSDNode>(User)) {
1408 if (ST->isIndexed() || ST->isTruncatingStore() ||
1409 ST->getValue() != Vec)
1412 // Make sure that nothing else could have stored into the destination of
1414 if (!ST->getChain().reachesChainWithoutSideEffects(DAG.getEntryNode()))
1417 StackPtr = ST->getBasePtr();
1418 Ch = SDValue(ST, 0);
1423 if (!Ch.getNode()) {
1424 // Store the value to a temporary stack slot, then LOAD the returned part.
1425 StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1426 Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr,
1427 MachinePointerInfo(), false, false, 0);
1430 // Add the offset to the index.
1432 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1433 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1434 DAG.getConstant(EltSize, Idx.getValueType()));
1436 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1437 StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr);
1439 if (Op.getValueType().isVector())
1440 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr,MachinePointerInfo(),
1441 false, false, false, 0);
1442 return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr,
1443 MachinePointerInfo(),
1444 Vec.getValueType().getVectorElementType(),
1445 false, false, false, 0);
1448 SDValue SelectionDAGLegalize::ExpandInsertToVectorThroughStack(SDValue Op) {
1449 assert(Op.getValueType().isVector() && "Non-vector insert subvector!");
1451 SDValue Vec = Op.getOperand(0);
1452 SDValue Part = Op.getOperand(1);
1453 SDValue Idx = Op.getOperand(2);
1456 // Store the value to a temporary stack slot, then LOAD the returned part.
1458 SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType());
1459 int FI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex();
1460 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1462 // First store the whole vector.
1463 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, PtrInfo,
1466 // Then store the inserted part.
1468 // Add the offset to the index.
1470 Vec.getValueType().getVectorElementType().getSizeInBits()/8;
1472 Idx = DAG.getNode(ISD::MUL, dl, Idx.getValueType(), Idx,
1473 DAG.getConstant(EltSize, Idx.getValueType()));
1474 Idx = DAG.getZExtOrTrunc(Idx, dl, TLI.getPointerTy());
1476 SDValue SubStackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx,
1479 // Store the subvector.
1480 Ch = DAG.getStore(Ch, dl, Part, SubStackPtr,
1481 MachinePointerInfo(), false, false, 0);
1483 // Finally, load the updated vector.
1484 return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, PtrInfo,
1485 false, false, false, 0);
1488 SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) {
1489 // We can't handle this case efficiently. Allocate a sufficiently
1490 // aligned object on the stack, store each element into it, then load
1491 // the result as a vector.
1492 // Create the stack frame object.
1493 EVT VT = Node->getValueType(0);
1494 EVT EltVT = VT.getVectorElementType();
1496 SDValue FIPtr = DAG.CreateStackTemporary(VT);
1497 int FI = cast<FrameIndexSDNode>(FIPtr.getNode())->getIndex();
1498 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(FI);
1500 // Emit a store of each element to the stack slot.
1501 SmallVector<SDValue, 8> Stores;
1502 unsigned TypeByteSize = EltVT.getSizeInBits() / 8;
1503 // Store (in the right endianness) the elements to memory.
1504 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
1505 // Ignore undef elements.
1506 if (Node->getOperand(i).getOpcode() == ISD::UNDEF) continue;
1508 unsigned Offset = TypeByteSize*i;
1510 SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType());
1511 Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx);
1513 // If the destination vector element type is narrower than the source
1514 // element type, only store the bits necessary.
1515 if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) {
1516 Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl,
1517 Node->getOperand(i), Idx,
1518 PtrInfo.getWithOffset(Offset),
1519 EltVT, false, false, 0));
1521 Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl,
1522 Node->getOperand(i), Idx,
1523 PtrInfo.getWithOffset(Offset),
1528 if (!Stores.empty()) // Not all undef elements?
1529 StoreChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Stores);
1531 StoreChain = DAG.getEntryNode();
1533 // Result is a load from the stack slot.
1534 return DAG.getLoad(VT, dl, StoreChain, FIPtr, PtrInfo,
1535 false, false, false, 0);
1538 SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) {
1540 SDValue Tmp1 = Node->getOperand(0);
1541 SDValue Tmp2 = Node->getOperand(1);
1543 // Get the sign bit of the RHS. First obtain a value that has the same
1544 // sign as the sign bit, i.e. negative if and only if the sign bit is 1.
1546 EVT FloatVT = Tmp2.getValueType();
1547 EVT IVT = EVT::getIntegerVT(*DAG.getContext(), FloatVT.getSizeInBits());
1548 if (TLI.isTypeLegal(IVT)) {
1549 // Convert to an integer with the same sign bit.
1550 SignBit = DAG.getNode(ISD::BITCAST, dl, IVT, Tmp2);
1552 // Store the float to memory, then load the sign part out as an integer.
1553 MVT LoadTy = TLI.getPointerTy();
1554 // First create a temporary that is aligned for both the load and store.
1555 SDValue StackPtr = DAG.CreateStackTemporary(FloatVT, LoadTy);
1556 // Then store the float to it.
1558 DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StackPtr, MachinePointerInfo(),
1560 if (TLI.isBigEndian()) {
1561 assert(FloatVT.isByteSized() && "Unsupported floating point type!");
1562 // Load out a legal integer with the same sign bit as the float.
1563 SignBit = DAG.getLoad(LoadTy, dl, Ch, StackPtr, MachinePointerInfo(),
1564 false, false, false, 0);
1565 } else { // Little endian
1566 SDValue LoadPtr = StackPtr;
1567 // The float may be wider than the integer we are going to load. Advance
1568 // the pointer so that the loaded integer will contain the sign bit.
1569 unsigned Strides = (FloatVT.getSizeInBits()-1)/LoadTy.getSizeInBits();
1570 unsigned ByteOffset = (Strides * LoadTy.getSizeInBits()) / 8;
1571 LoadPtr = DAG.getNode(ISD::ADD, dl, LoadPtr.getValueType(), LoadPtr,
1572 DAG.getConstant(ByteOffset, LoadPtr.getValueType()));
1573 // Load a legal integer containing the sign bit.
1574 SignBit = DAG.getLoad(LoadTy, dl, Ch, LoadPtr, MachinePointerInfo(),
1575 false, false, false, 0);
1576 // Move the sign bit to the top bit of the loaded integer.
1577 unsigned BitShift = LoadTy.getSizeInBits() -
1578 (FloatVT.getSizeInBits() - 8 * ByteOffset);
1579 assert(BitShift < LoadTy.getSizeInBits() && "Pointer advanced wrong?");
1581 SignBit = DAG.getNode(ISD::SHL, dl, LoadTy, SignBit,
1582 DAG.getConstant(BitShift,
1583 TLI.getShiftAmountTy(SignBit.getValueType())));
1586 // Now get the sign bit proper, by seeing whether the value is negative.
1587 SignBit = DAG.getSetCC(dl, getSetCCResultType(SignBit.getValueType()),
1588 SignBit, DAG.getConstant(0, SignBit.getValueType()),
1590 // Get the absolute value of the result.
1591 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, Tmp1.getValueType(), Tmp1);
1592 // Select between the nabs and abs value based on the sign bit of
1594 return DAG.getSelect(dl, AbsVal.getValueType(), SignBit,
1595 DAG.getNode(ISD::FNEG, dl, AbsVal.getValueType(), AbsVal),
1599 void SelectionDAGLegalize::ExpandDYNAMIC_STACKALLOC(SDNode* Node,
1600 SmallVectorImpl<SDValue> &Results) {
1601 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
1602 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
1603 " not tell us which reg is the stack pointer!");
1605 EVT VT = Node->getValueType(0);
1606 SDValue Tmp1 = SDValue(Node, 0);
1607 SDValue Tmp2 = SDValue(Node, 1);
1608 SDValue Tmp3 = Node->getOperand(2);
1609 SDValue Chain = Tmp1.getOperand(0);
1611 // Chain the dynamic stack allocation so that it doesn't modify the stack
1612 // pointer when other instructions are using the stack.
1613 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
1616 SDValue Size = Tmp2.getOperand(1);
1617 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
1618 Chain = SP.getValue(1);
1619 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
1620 unsigned StackAlign =
1621 DAG.getSubtarget().getFrameLowering()->getStackAlignment();
1622 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
1623 if (Align > StackAlign)
1624 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
1625 DAG.getConstant(-(uint64_t)Align, VT));
1626 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
1628 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
1629 DAG.getIntPtrConstant(0, true), SDValue(),
1632 Results.push_back(Tmp1);
1633 Results.push_back(Tmp2);
1636 /// Legalize a SETCC with given LHS and RHS and condition code CC on the current
1639 /// If the SETCC has been legalized using AND / OR, then the legalized node
1640 /// will be stored in LHS. RHS and CC will be set to SDValue(). NeedInvert
1641 /// will be set to false.
1643 /// If the SETCC has been legalized by using getSetCCSwappedOperands(),
1644 /// then the values of LHS and RHS will be swapped, CC will be set to the
1645 /// new condition, and NeedInvert will be set to false.
1647 /// If the SETCC has been legalized using the inverse condcode, then LHS and
1648 /// RHS will be unchanged, CC will set to the inverted condcode, and NeedInvert
1649 /// will be set to true. The caller must invert the result of the SETCC with
1650 /// SelectionDAG::getLogicalNOT() or take equivalent action to swap the effect
1651 /// of a true/false result.
1653 /// \returns true if the SetCC has been legalized, false if it hasn't.
1654 bool SelectionDAGLegalize::LegalizeSetCCCondCode(EVT VT,
1655 SDValue &LHS, SDValue &RHS,
1659 MVT OpVT = LHS.getSimpleValueType();
1660 ISD::CondCode CCCode = cast<CondCodeSDNode>(CC)->get();
1662 switch (TLI.getCondCodeAction(CCCode, OpVT)) {
1663 default: llvm_unreachable("Unknown condition code action!");
1664 case TargetLowering::Legal:
1667 case TargetLowering::Expand: {
1668 ISD::CondCode InvCC = ISD::getSetCCSwappedOperands(CCCode);
1669 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1670 std::swap(LHS, RHS);
1671 CC = DAG.getCondCode(InvCC);
1674 ISD::CondCode CC1 = ISD::SETCC_INVALID, CC2 = ISD::SETCC_INVALID;
1677 default: llvm_unreachable("Don't know how to expand this condition!");
1679 assert(TLI.getCondCodeAction(ISD::SETOEQ, OpVT)
1680 == TargetLowering::Legal
1681 && "If SETO is expanded, SETOEQ must be legal!");
1682 CC1 = ISD::SETOEQ; CC2 = ISD::SETOEQ; Opc = ISD::AND; break;
1684 assert(TLI.getCondCodeAction(ISD::SETUNE, OpVT)
1685 == TargetLowering::Legal
1686 && "If SETUO is expanded, SETUNE must be legal!");
1687 CC1 = ISD::SETUNE; CC2 = ISD::SETUNE; Opc = ISD::OR; break;
1700 // If we are floating point, assign and break, otherwise fall through.
1701 if (!OpVT.isInteger()) {
1702 // We can use the 4th bit to tell if we are the unordered
1703 // or ordered version of the opcode.
1704 CC2 = ((unsigned)CCCode & 0x8U) ? ISD::SETUO : ISD::SETO;
1705 Opc = ((unsigned)CCCode & 0x8U) ? ISD::OR : ISD::AND;
1706 CC1 = (ISD::CondCode)(((int)CCCode & 0x7) | 0x10);
1709 // Fallthrough if we are unsigned integer.
1714 // We only support using the inverted operation, which is computed above
1715 // and not a different manner of supporting expanding these cases.
1716 llvm_unreachable("Don't know how to expand this condition!");
1719 // Try inverting the result of the inverse condition.
1720 InvCC = CCCode == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ;
1721 if (TLI.isCondCodeLegal(InvCC, OpVT)) {
1722 CC = DAG.getCondCode(InvCC);
1726 // If inverting the condition didn't work then we have no means to expand
1728 llvm_unreachable("Don't know how to expand this condition!");
1731 SDValue SetCC1, SetCC2;
1732 if (CCCode != ISD::SETO && CCCode != ISD::SETUO) {
1733 // If we aren't the ordered or unorder operation,
1734 // then the pattern is (LHS CC1 RHS) Opc (LHS CC2 RHS).
1735 SetCC1 = DAG.getSetCC(dl, VT, LHS, RHS, CC1);
1736 SetCC2 = DAG.getSetCC(dl, VT, LHS, RHS, CC2);
1738 // Otherwise, the pattern is (LHS CC1 LHS) Opc (RHS CC2 RHS)
1739 SetCC1 = DAG.getSetCC(dl, VT, LHS, LHS, CC1);
1740 SetCC2 = DAG.getSetCC(dl, VT, RHS, RHS, CC2);
1742 LHS = DAG.getNode(Opc, dl, VT, SetCC1, SetCC2);
1751 /// Emit a store/load combination to the stack. This stores
1752 /// SrcOp to a stack slot of type SlotVT, truncating it if needed. It then does
1753 /// a load from the stack slot to DestVT, extending it if needed.
1754 /// The resultant code need not be legal.
1755 SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp,
1759 // Create the stack frame object.
1761 TLI.getDataLayout()->getPrefTypeAlignment(SrcOp.getValueType().
1762 getTypeForEVT(*DAG.getContext()));
1763 SDValue FIPtr = DAG.CreateStackTemporary(SlotVT, SrcAlign);
1765 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(FIPtr);
1766 int SPFI = StackPtrFI->getIndex();
1767 MachinePointerInfo PtrInfo = MachinePointerInfo::getFixedStack(SPFI);
1769 unsigned SrcSize = SrcOp.getValueType().getSizeInBits();
1770 unsigned SlotSize = SlotVT.getSizeInBits();
1771 unsigned DestSize = DestVT.getSizeInBits();
1772 Type *DestType = DestVT.getTypeForEVT(*DAG.getContext());
1773 unsigned DestAlign = TLI.getDataLayout()->getPrefTypeAlignment(DestType);
1775 // Emit a store to the stack slot. Use a truncstore if the input value is
1776 // later than DestVT.
1779 if (SrcSize > SlotSize)
1780 Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1781 PtrInfo, SlotVT, false, false, SrcAlign);
1783 assert(SrcSize == SlotSize && "Invalid store");
1784 Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr,
1785 PtrInfo, false, false, SrcAlign);
1788 // Result is a load from the stack slot.
1789 if (SlotSize == DestSize)
1790 return DAG.getLoad(DestVT, dl, Store, FIPtr, PtrInfo,
1791 false, false, false, DestAlign);
1793 assert(SlotSize < DestSize && "Unknown extension!");
1794 return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr,
1795 PtrInfo, SlotVT, false, false, false, DestAlign);
1798 SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) {
1800 // Create a vector sized/aligned stack slot, store the value to element #0,
1801 // then load the whole vector back out.
1802 SDValue StackPtr = DAG.CreateStackTemporary(Node->getValueType(0));
1804 FrameIndexSDNode *StackPtrFI = cast<FrameIndexSDNode>(StackPtr);
1805 int SPFI = StackPtrFI->getIndex();
1807 SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0),
1809 MachinePointerInfo::getFixedStack(SPFI),
1810 Node->getValueType(0).getVectorElementType(),
1812 return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr,
1813 MachinePointerInfo::getFixedStack(SPFI),
1814 false, false, false, 0);
1818 ExpandBVWithShuffles(SDNode *Node, SelectionDAG &DAG,
1819 const TargetLowering &TLI, SDValue &Res) {
1820 unsigned NumElems = Node->getNumOperands();
1822 EVT VT = Node->getValueType(0);
1824 // Try to group the scalars into pairs, shuffle the pairs together, then
1825 // shuffle the pairs of pairs together, etc. until the vector has
1826 // been built. This will work only if all of the necessary shuffle masks
1829 // We do this in two phases; first to check the legality of the shuffles,
1830 // and next, assuming that all shuffles are legal, to create the new nodes.
1831 for (int Phase = 0; Phase < 2; ++Phase) {
1832 SmallVector<std::pair<SDValue, SmallVector<int, 16> >, 16> IntermedVals,
1834 for (unsigned i = 0; i < NumElems; ++i) {
1835 SDValue V = Node->getOperand(i);
1836 if (V.getOpcode() == ISD::UNDEF)
1841 Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, V);
1842 IntermedVals.push_back(std::make_pair(Vec, SmallVector<int, 16>(1, i)));
1845 while (IntermedVals.size() > 2) {
1846 NewIntermedVals.clear();
1847 for (unsigned i = 0, e = (IntermedVals.size() & ~1u); i < e; i += 2) {
1848 // This vector and the next vector are shuffled together (simply to
1849 // append the one to the other).
1850 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1852 SmallVector<int, 16> FinalIndices;
1853 FinalIndices.reserve(IntermedVals[i].second.size() +
1854 IntermedVals[i+1].second.size());
1857 for (unsigned j = 0, f = IntermedVals[i].second.size(); j != f;
1860 FinalIndices.push_back(IntermedVals[i].second[j]);
1862 for (unsigned j = 0, f = IntermedVals[i+1].second.size(); j != f;
1864 ShuffleVec[k] = NumElems + j;
1865 FinalIndices.push_back(IntermedVals[i+1].second[j]);
1870 Shuffle = DAG.getVectorShuffle(VT, dl, IntermedVals[i].first,
1871 IntermedVals[i+1].first,
1873 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1875 NewIntermedVals.push_back(
1876 std::make_pair(Shuffle, std::move(FinalIndices)));
1879 // If we had an odd number of defined values, then append the last
1880 // element to the array of new vectors.
1881 if ((IntermedVals.size() & 1) != 0)
1882 NewIntermedVals.push_back(IntermedVals.back());
1884 IntermedVals.swap(NewIntermedVals);
1887 assert(IntermedVals.size() <= 2 && IntermedVals.size() > 0 &&
1888 "Invalid number of intermediate vectors");
1889 SDValue Vec1 = IntermedVals[0].first;
1891 if (IntermedVals.size() > 1)
1892 Vec2 = IntermedVals[1].first;
1894 Vec2 = DAG.getUNDEF(VT);
1896 SmallVector<int, 16> ShuffleVec(NumElems, -1);
1897 for (unsigned i = 0, e = IntermedVals[0].second.size(); i != e; ++i)
1898 ShuffleVec[IntermedVals[0].second[i]] = i;
1899 for (unsigned i = 0, e = IntermedVals[1].second.size(); i != e; ++i)
1900 ShuffleVec[IntermedVals[1].second[i]] = NumElems + i;
1903 Res = DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
1904 else if (!TLI.isShuffleMaskLegal(ShuffleVec, VT))
1911 /// Expand a BUILD_VECTOR node on targets that don't
1912 /// support the operation, but do support the resultant vector type.
1913 SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) {
1914 unsigned NumElems = Node->getNumOperands();
1915 SDValue Value1, Value2;
1917 EVT VT = Node->getValueType(0);
1918 EVT OpVT = Node->getOperand(0).getValueType();
1919 EVT EltVT = VT.getVectorElementType();
1921 // If the only non-undef value is the low element, turn this into a
1922 // SCALAR_TO_VECTOR node. If this is { X, X, X, X }, determine X.
1923 bool isOnlyLowElement = true;
1924 bool MoreThanTwoValues = false;
1925 bool isConstant = true;
1926 for (unsigned i = 0; i < NumElems; ++i) {
1927 SDValue V = Node->getOperand(i);
1928 if (V.getOpcode() == ISD::UNDEF)
1931 isOnlyLowElement = false;
1932 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
1935 if (!Value1.getNode()) {
1937 } else if (!Value2.getNode()) {
1940 } else if (V != Value1 && V != Value2) {
1941 MoreThanTwoValues = true;
1945 if (!Value1.getNode())
1946 return DAG.getUNDEF(VT);
1948 if (isOnlyLowElement)
1949 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Node->getOperand(0));
1951 // If all elements are constants, create a load from the constant pool.
1953 SmallVector<Constant*, 16> CV;
1954 for (unsigned i = 0, e = NumElems; i != e; ++i) {
1955 if (ConstantFPSDNode *V =
1956 dyn_cast<ConstantFPSDNode>(Node->getOperand(i))) {
1957 CV.push_back(const_cast<ConstantFP *>(V->getConstantFPValue()));
1958 } else if (ConstantSDNode *V =
1959 dyn_cast<ConstantSDNode>(Node->getOperand(i))) {
1961 CV.push_back(const_cast<ConstantInt *>(V->getConstantIntValue()));
1963 // If OpVT and EltVT don't match, EltVT is not legal and the
1964 // element values have been promoted/truncated earlier. Undo this;
1965 // we don't want a v16i8 to become a v16i32 for example.
1966 const ConstantInt *CI = V->getConstantIntValue();
1967 CV.push_back(ConstantInt::get(EltVT.getTypeForEVT(*DAG.getContext()),
1968 CI->getZExtValue()));
1971 assert(Node->getOperand(i).getOpcode() == ISD::UNDEF);
1972 Type *OpNTy = EltVT.getTypeForEVT(*DAG.getContext());
1973 CV.push_back(UndefValue::get(OpNTy));
1976 Constant *CP = ConstantVector::get(CV);
1977 SDValue CPIdx = DAG.getConstantPool(CP, TLI.getPointerTy());
1978 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
1979 return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
1980 MachinePointerInfo::getConstantPool(),
1981 false, false, false, Alignment);
1984 SmallSet<SDValue, 16> DefinedValues;
1985 for (unsigned i = 0; i < NumElems; ++i) {
1986 if (Node->getOperand(i).getOpcode() == ISD::UNDEF)
1988 DefinedValues.insert(Node->getOperand(i));
1991 if (TLI.shouldExpandBuildVectorWithShuffles(VT, DefinedValues.size())) {
1992 if (!MoreThanTwoValues) {
1993 SmallVector<int, 8> ShuffleVec(NumElems, -1);
1994 for (unsigned i = 0; i < NumElems; ++i) {
1995 SDValue V = Node->getOperand(i);
1996 if (V.getOpcode() == ISD::UNDEF)
1998 ShuffleVec[i] = V == Value1 ? 0 : NumElems;
2000 if (TLI.isShuffleMaskLegal(ShuffleVec, Node->getValueType(0))) {
2001 // Get the splatted value into the low element of a vector register.
2002 SDValue Vec1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value1);
2004 if (Value2.getNode())
2005 Vec2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value2);
2007 Vec2 = DAG.getUNDEF(VT);
2009 // Return shuffle(LowValVec, undef, <0,0,0,0>)
2010 return DAG.getVectorShuffle(VT, dl, Vec1, Vec2, ShuffleVec.data());
2014 if (ExpandBVWithShuffles(Node, DAG, TLI, Res))
2019 // Otherwise, we can't handle this case efficiently.
2020 return ExpandVectorBuildThroughStack(Node);
2023 // Expand a node into a call to a libcall. If the result value
2024 // does not fit into a register, return the lo part and set the hi part to the
2025 // by-reg argument. If it does fit into a single register, return the result
2026 // and leave the Hi part unset.
2027 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node,
2029 TargetLowering::ArgListTy Args;
2030 TargetLowering::ArgListEntry Entry;
2031 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2032 EVT ArgVT = Node->getOperand(i).getValueType();
2033 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2034 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2035 Entry.isSExt = isSigned;
2036 Entry.isZExt = !isSigned;
2037 Args.push_back(Entry);
2039 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2040 TLI.getPointerTy());
2042 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2044 // By default, the input chain to this libcall is the entry node of the
2045 // function. If the libcall is going to be emitted as a tail call then
2046 // TLI.isUsedByReturnOnly will change it to the right chain if the return
2047 // node which is being folded has a non-entry input chain.
2048 SDValue InChain = DAG.getEntryNode();
2050 // isTailCall may be true since the callee does not reference caller stack
2051 // frame. Check if it's in the right position.
2052 SDValue TCChain = InChain;
2053 bool isTailCall = TLI.isInTailCallPosition(DAG, Node, TCChain);
2057 TargetLowering::CallLoweringInfo CLI(DAG);
2058 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2059 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2060 .setTailCall(isTailCall).setSExtResult(isSigned).setZExtResult(!isSigned);
2062 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2064 if (!CallInfo.second.getNode())
2065 // It's a tailcall, return the chain (which is the DAG root).
2066 return DAG.getRoot();
2068 return CallInfo.first;
2071 /// Generate a libcall taking the given operands as arguments
2072 /// and returning a result of type RetVT.
2073 SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, EVT RetVT,
2074 const SDValue *Ops, unsigned NumOps,
2075 bool isSigned, SDLoc dl) {
2076 TargetLowering::ArgListTy Args;
2077 Args.reserve(NumOps);
2079 TargetLowering::ArgListEntry Entry;
2080 for (unsigned i = 0; i != NumOps; ++i) {
2081 Entry.Node = Ops[i];
2082 Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext());
2083 Entry.isSExt = isSigned;
2084 Entry.isZExt = !isSigned;
2085 Args.push_back(Entry);
2087 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2088 TLI.getPointerTy());
2090 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2092 TargetLowering::CallLoweringInfo CLI(DAG);
2093 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
2094 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2095 .setSExtResult(isSigned).setZExtResult(!isSigned);
2097 std::pair<SDValue,SDValue> CallInfo = TLI.LowerCallTo(CLI);
2099 return CallInfo.first;
2102 // Expand a node into a call to a libcall. Similar to
2103 // ExpandLibCall except that the first operand is the in-chain.
2104 std::pair<SDValue, SDValue>
2105 SelectionDAGLegalize::ExpandChainLibCall(RTLIB::Libcall LC,
2108 SDValue InChain = Node->getOperand(0);
2110 TargetLowering::ArgListTy Args;
2111 TargetLowering::ArgListEntry Entry;
2112 for (unsigned i = 1, e = Node->getNumOperands(); i != e; ++i) {
2113 EVT ArgVT = Node->getOperand(i).getValueType();
2114 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2115 Entry.Node = Node->getOperand(i);
2117 Entry.isSExt = isSigned;
2118 Entry.isZExt = !isSigned;
2119 Args.push_back(Entry);
2121 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2122 TLI.getPointerTy());
2124 Type *RetTy = Node->getValueType(0).getTypeForEVT(*DAG.getContext());
2126 TargetLowering::CallLoweringInfo CLI(DAG);
2127 CLI.setDebugLoc(SDLoc(Node)).setChain(InChain)
2128 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2129 .setSExtResult(isSigned).setZExtResult(!isSigned);
2131 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2136 SDValue SelectionDAGLegalize::ExpandFPLibCall(SDNode* Node,
2137 RTLIB::Libcall Call_F32,
2138 RTLIB::Libcall Call_F64,
2139 RTLIB::Libcall Call_F80,
2140 RTLIB::Libcall Call_F128,
2141 RTLIB::Libcall Call_PPCF128) {
2143 switch (Node->getSimpleValueType(0).SimpleTy) {
2144 default: llvm_unreachable("Unexpected request for libcall!");
2145 case MVT::f32: LC = Call_F32; break;
2146 case MVT::f64: LC = Call_F64; break;
2147 case MVT::f80: LC = Call_F80; break;
2148 case MVT::f128: LC = Call_F128; break;
2149 case MVT::ppcf128: LC = Call_PPCF128; break;
2151 return ExpandLibCall(LC, Node, false);
2154 SDValue SelectionDAGLegalize::ExpandIntLibCall(SDNode* Node, bool isSigned,
2155 RTLIB::Libcall Call_I8,
2156 RTLIB::Libcall Call_I16,
2157 RTLIB::Libcall Call_I32,
2158 RTLIB::Libcall Call_I64,
2159 RTLIB::Libcall Call_I128) {
2161 switch (Node->getSimpleValueType(0).SimpleTy) {
2162 default: llvm_unreachable("Unexpected request for libcall!");
2163 case MVT::i8: LC = Call_I8; break;
2164 case MVT::i16: LC = Call_I16; break;
2165 case MVT::i32: LC = Call_I32; break;
2166 case MVT::i64: LC = Call_I64; break;
2167 case MVT::i128: LC = Call_I128; break;
2169 return ExpandLibCall(LC, Node, isSigned);
2172 /// Return true if divmod libcall is available.
2173 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
2174 const TargetLowering &TLI) {
2176 switch (Node->getSimpleValueType(0).SimpleTy) {
2177 default: llvm_unreachable("Unexpected request for libcall!");
2178 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2179 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2180 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2181 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2182 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2185 return TLI.getLibcallName(LC) != nullptr;
2188 /// Only issue divrem libcall if both quotient and remainder are needed.
2189 static bool useDivRem(SDNode *Node, bool isSigned, bool isDIV) {
2190 // The other use might have been replaced with a divrem already.
2191 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
2192 unsigned OtherOpcode = 0;
2194 OtherOpcode = isDIV ? ISD::SREM : ISD::SDIV;
2196 OtherOpcode = isDIV ? ISD::UREM : ISD::UDIV;
2198 SDValue Op0 = Node->getOperand(0);
2199 SDValue Op1 = Node->getOperand(1);
2200 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2201 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2205 if ((User->getOpcode() == OtherOpcode || User->getOpcode() == DivRemOpc) &&
2206 User->getOperand(0) == Op0 &&
2207 User->getOperand(1) == Op1)
2213 /// Issue libcalls to __{u}divmod to compute div / rem pairs.
2215 SelectionDAGLegalize::ExpandDivRemLibCall(SDNode *Node,
2216 SmallVectorImpl<SDValue> &Results) {
2217 unsigned Opcode = Node->getOpcode();
2218 bool isSigned = Opcode == ISD::SDIVREM;
2221 switch (Node->getSimpleValueType(0).SimpleTy) {
2222 default: llvm_unreachable("Unexpected request for libcall!");
2223 case MVT::i8: LC= isSigned ? RTLIB::SDIVREM_I8 : RTLIB::UDIVREM_I8; break;
2224 case MVT::i16: LC= isSigned ? RTLIB::SDIVREM_I16 : RTLIB::UDIVREM_I16; break;
2225 case MVT::i32: LC= isSigned ? RTLIB::SDIVREM_I32 : RTLIB::UDIVREM_I32; break;
2226 case MVT::i64: LC= isSigned ? RTLIB::SDIVREM_I64 : RTLIB::UDIVREM_I64; break;
2227 case MVT::i128: LC= isSigned ? RTLIB::SDIVREM_I128:RTLIB::UDIVREM_I128; break;
2230 // The input chain to this libcall is the entry node of the function.
2231 // Legalizing the call will automatically add the previous call to the
2233 SDValue InChain = DAG.getEntryNode();
2235 EVT RetVT = Node->getValueType(0);
2236 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2238 TargetLowering::ArgListTy Args;
2239 TargetLowering::ArgListEntry Entry;
2240 for (unsigned i = 0, e = Node->getNumOperands(); i != e; ++i) {
2241 EVT ArgVT = Node->getOperand(i).getValueType();
2242 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2243 Entry.Node = Node->getOperand(i); Entry.Ty = ArgTy;
2244 Entry.isSExt = isSigned;
2245 Entry.isZExt = !isSigned;
2246 Args.push_back(Entry);
2249 // Also pass the return address of the remainder.
2250 SDValue FIPtr = DAG.CreateStackTemporary(RetVT);
2252 Entry.Ty = RetTy->getPointerTo();
2253 Entry.isSExt = isSigned;
2254 Entry.isZExt = !isSigned;
2255 Args.push_back(Entry);
2257 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2258 TLI.getPointerTy());
2261 TargetLowering::CallLoweringInfo CLI(DAG);
2262 CLI.setDebugLoc(dl).setChain(InChain)
2263 .setCallee(TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args), 0)
2264 .setSExtResult(isSigned).setZExtResult(!isSigned);
2266 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2268 // Remainder is loaded back from the stack frame.
2269 SDValue Rem = DAG.getLoad(RetVT, dl, CallInfo.second, FIPtr,
2270 MachinePointerInfo(), false, false, false, 0);
2271 Results.push_back(CallInfo.first);
2272 Results.push_back(Rem);
2275 /// Return true if sincos libcall is available.
2276 static bool isSinCosLibcallAvailable(SDNode *Node, const TargetLowering &TLI) {
2278 switch (Node->getSimpleValueType(0).SimpleTy) {
2279 default: llvm_unreachable("Unexpected request for libcall!");
2280 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2281 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2282 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2283 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2284 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2286 return TLI.getLibcallName(LC) != nullptr;
2289 /// Return true if sincos libcall is available and can be used to combine sin
2291 static bool canCombineSinCosLibcall(SDNode *Node, const TargetLowering &TLI,
2292 const TargetMachine &TM) {
2293 if (!isSinCosLibcallAvailable(Node, TLI))
2295 // GNU sin/cos functions set errno while sincos does not. Therefore
2296 // combining sin and cos is only safe if unsafe-fpmath is enabled.
2297 bool isGNU = Triple(TM.getTargetTriple()).getEnvironment() == Triple::GNU;
2298 if (isGNU && !TM.Options.UnsafeFPMath)
2303 /// Only issue sincos libcall if both sin and cos are needed.
2304 static bool useSinCos(SDNode *Node) {
2305 unsigned OtherOpcode = Node->getOpcode() == ISD::FSIN
2306 ? ISD::FCOS : ISD::FSIN;
2308 SDValue Op0 = Node->getOperand(0);
2309 for (SDNode::use_iterator UI = Op0.getNode()->use_begin(),
2310 UE = Op0.getNode()->use_end(); UI != UE; ++UI) {
2314 // The other user might have been turned into sincos already.
2315 if (User->getOpcode() == OtherOpcode || User->getOpcode() == ISD::FSINCOS)
2321 /// Issue libcalls to sincos to compute sin / cos pairs.
2323 SelectionDAGLegalize::ExpandSinCosLibCall(SDNode *Node,
2324 SmallVectorImpl<SDValue> &Results) {
2326 switch (Node->getSimpleValueType(0).SimpleTy) {
2327 default: llvm_unreachable("Unexpected request for libcall!");
2328 case MVT::f32: LC = RTLIB::SINCOS_F32; break;
2329 case MVT::f64: LC = RTLIB::SINCOS_F64; break;
2330 case MVT::f80: LC = RTLIB::SINCOS_F80; break;
2331 case MVT::f128: LC = RTLIB::SINCOS_F128; break;
2332 case MVT::ppcf128: LC = RTLIB::SINCOS_PPCF128; break;
2335 // The input chain to this libcall is the entry node of the function.
2336 // Legalizing the call will automatically add the previous call to the
2338 SDValue InChain = DAG.getEntryNode();
2340 EVT RetVT = Node->getValueType(0);
2341 Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext());
2343 TargetLowering::ArgListTy Args;
2344 TargetLowering::ArgListEntry Entry;
2346 // Pass the argument.
2347 Entry.Node = Node->getOperand(0);
2349 Entry.isSExt = false;
2350 Entry.isZExt = false;
2351 Args.push_back(Entry);
2353 // Pass the return address of sin.
2354 SDValue SinPtr = DAG.CreateStackTemporary(RetVT);
2355 Entry.Node = SinPtr;
2356 Entry.Ty = RetTy->getPointerTo();
2357 Entry.isSExt = false;
2358 Entry.isZExt = false;
2359 Args.push_back(Entry);
2361 // Also pass the return address of the cos.
2362 SDValue CosPtr = DAG.CreateStackTemporary(RetVT);
2363 Entry.Node = CosPtr;
2364 Entry.Ty = RetTy->getPointerTo();
2365 Entry.isSExt = false;
2366 Entry.isZExt = false;
2367 Args.push_back(Entry);
2369 SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
2370 TLI.getPointerTy());
2373 TargetLowering::CallLoweringInfo CLI(DAG);
2374 CLI.setDebugLoc(dl).setChain(InChain)
2375 .setCallee(TLI.getLibcallCallingConv(LC),
2376 Type::getVoidTy(*DAG.getContext()), Callee, std::move(Args), 0);
2378 std::pair<SDValue, SDValue> CallInfo = TLI.LowerCallTo(CLI);
2380 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, SinPtr,
2381 MachinePointerInfo(), false, false, false, 0));
2382 Results.push_back(DAG.getLoad(RetVT, dl, CallInfo.second, CosPtr,
2383 MachinePointerInfo(), false, false, false, 0));
2386 /// This function is responsible for legalizing a
2387 /// INT_TO_FP operation of the specified operand when the target requests that
2388 /// we expand it. At this point, we know that the result and operand types are
2389 /// legal for the target.
2390 SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned,
2394 if (Op0.getValueType() == MVT::i32 && TLI.isTypeLegal(MVT::f64)) {
2395 // simple 32-bit [signed|unsigned] integer to float/double expansion
2397 // Get the stack frame index of a 8 byte buffer.
2398 SDValue StackSlot = DAG.CreateStackTemporary(MVT::f64);
2400 // word offset constant for Hi/Lo address computation
2401 SDValue WordOff = DAG.getConstant(sizeof(int), StackSlot.getValueType());
2402 // set up Hi and Lo (into buffer) address based on endian
2403 SDValue Hi = StackSlot;
2404 SDValue Lo = DAG.getNode(ISD::ADD, dl, StackSlot.getValueType(),
2405 StackSlot, WordOff);
2406 if (TLI.isLittleEndian())
2409 // if signed map to unsigned space
2412 // constant used to invert sign bit (signed to unsigned mapping)
2413 SDValue SignBit = DAG.getConstant(0x80000000u, MVT::i32);
2414 Op0Mapped = DAG.getNode(ISD::XOR, dl, MVT::i32, Op0, SignBit);
2418 // store the lo of the constructed double - based on integer input
2419 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl,
2420 Op0Mapped, Lo, MachinePointerInfo(),
2422 // initial hi portion of constructed double
2423 SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32);
2424 // store the hi of the constructed double - biased exponent
2425 SDValue Store2 = DAG.getStore(Store1, dl, InitialHi, Hi,
2426 MachinePointerInfo(),
2428 // load the constructed double
2429 SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot,
2430 MachinePointerInfo(), false, false, false, 0);
2431 // FP constant to bias correct the final result
2432 SDValue Bias = DAG.getConstantFP(isSigned ?
2433 BitsToDouble(0x4330000080000000ULL) :
2434 BitsToDouble(0x4330000000000000ULL),
2436 // subtract the bias
2437 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Load, Bias);
2440 // handle final rounding
2441 if (DestVT == MVT::f64) {
2444 } else if (DestVT.bitsLT(MVT::f64)) {
2445 Result = DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
2446 DAG.getIntPtrConstant(0));
2447 } else if (DestVT.bitsGT(MVT::f64)) {
2448 Result = DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
2452 assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet");
2453 // Code below here assumes !isSigned without checking again.
2455 // Implementation of unsigned i64 to f64 following the algorithm in
2456 // __floatundidf in compiler_rt. This implementation has the advantage
2457 // of performing rounding correctly, both in the default rounding mode
2458 // and in all alternate rounding modes.
2459 // TODO: Generalize this for use with other types.
2460 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) {
2462 DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64);
2463 SDValue TwoP84PlusTwoP52 =
2464 DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64);
2466 DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64);
2468 SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32);
2469 SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0,
2470 DAG.getConstant(32, MVT::i64));
2471 SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52);
2472 SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84);
2473 SDValue LoFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, LoOr);
2474 SDValue HiFlt = DAG.getNode(ISD::BITCAST, dl, MVT::f64, HiOr);
2475 SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt,
2477 return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub);
2480 // Implementation of unsigned i64 to f32.
2481 // TODO: Generalize this for use with other types.
2482 if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f32) {
2483 // For unsigned conversions, convert them to signed conversions using the
2484 // algorithm from the x86_64 __floatundidf in compiler_rt.
2486 SDValue Fast = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Op0);
2488 SDValue ShiftConst =
2489 DAG.getConstant(1, TLI.getShiftAmountTy(Op0.getValueType()));
2490 SDValue Shr = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, ShiftConst);
2491 SDValue AndConst = DAG.getConstant(1, MVT::i64);
2492 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0, AndConst);
2493 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And, Shr);
2495 SDValue SignCvt = DAG.getNode(ISD::SINT_TO_FP, dl, MVT::f32, Or);
2496 SDValue Slow = DAG.getNode(ISD::FADD, dl, MVT::f32, SignCvt, SignCvt);
2498 // TODO: This really should be implemented using a branch rather than a
2499 // select. We happen to get lucky and machinesink does the right
2500 // thing most of the time. This would be a good candidate for a
2501 //pseudo-op, or, even better, for whole-function isel.
2502 SDValue SignBitTest = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2503 Op0, DAG.getConstant(0, MVT::i64), ISD::SETLT);
2504 return DAG.getSelect(dl, MVT::f32, SignBitTest, Slow, Fast);
2507 // Otherwise, implement the fully general conversion.
2509 SDValue And = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2510 DAG.getConstant(UINT64_C(0xfffffffffffff800), MVT::i64));
2511 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::i64, And,
2512 DAG.getConstant(UINT64_C(0x800), MVT::i64));
2513 SDValue And2 = DAG.getNode(ISD::AND, dl, MVT::i64, Op0,
2514 DAG.getConstant(UINT64_C(0x7ff), MVT::i64));
2515 SDValue Ne = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2516 And2, DAG.getConstant(UINT64_C(0), MVT::i64), ISD::SETNE);
2517 SDValue Sel = DAG.getSelect(dl, MVT::i64, Ne, Or, Op0);
2518 SDValue Ge = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
2519 Op0, DAG.getConstant(UINT64_C(0x0020000000000000), MVT::i64),
2521 SDValue Sel2 = DAG.getSelect(dl, MVT::i64, Ge, Sel, Op0);
2522 EVT SHVT = TLI.getShiftAmountTy(Sel2.getValueType());
2524 SDValue Sh = DAG.getNode(ISD::SRL, dl, MVT::i64, Sel2,
2525 DAG.getConstant(32, SHVT));
2526 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sh);
2527 SDValue Fcvt = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Trunc);
2529 DAG.getConstantFP(BitsToDouble(UINT64_C(0x41f0000000000000)), MVT::f64);
2530 SDValue Fmul = DAG.getNode(ISD::FMUL, dl, MVT::f64, TwoP32, Fcvt);
2531 SDValue Lo = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Sel2);
2532 SDValue Fcvt2 = DAG.getNode(ISD::UINT_TO_FP, dl, MVT::f64, Lo);
2533 SDValue Fadd = DAG.getNode(ISD::FADD, dl, MVT::f64, Fmul, Fcvt2);
2534 return DAG.getNode(ISD::FP_ROUND, dl, MVT::f32, Fadd,
2535 DAG.getIntPtrConstant(0));
2538 SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0);
2540 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(Op0.getValueType()),
2541 Op0, DAG.getConstant(0, Op0.getValueType()),
2543 SDValue Zero = DAG.getIntPtrConstant(0), Four = DAG.getIntPtrConstant(4);
2544 SDValue CstOffset = DAG.getSelect(dl, Zero.getValueType(),
2545 SignSet, Four, Zero);
2547 // If the sign bit of the integer is set, the large number will be treated
2548 // as a negative number. To counteract this, the dynamic code adds an
2549 // offset depending on the data type.
2551 switch (Op0.getSimpleValueType().SimpleTy) {
2552 default: llvm_unreachable("Unsupported integer type!");
2553 case MVT::i8 : FF = 0x43800000ULL; break; // 2^8 (as a float)
2554 case MVT::i16: FF = 0x47800000ULL; break; // 2^16 (as a float)
2555 case MVT::i32: FF = 0x4F800000ULL; break; // 2^32 (as a float)
2556 case MVT::i64: FF = 0x5F800000ULL; break; // 2^64 (as a float)
2558 if (TLI.isLittleEndian()) FF <<= 32;
2559 Constant *FudgeFactor = ConstantInt::get(
2560 Type::getInt64Ty(*DAG.getContext()), FF);
2562 SDValue CPIdx = DAG.getConstantPool(FudgeFactor, TLI.getPointerTy());
2563 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
2564 CPIdx = DAG.getNode(ISD::ADD, dl, CPIdx.getValueType(), CPIdx, CstOffset);
2565 Alignment = std::min(Alignment, 4u);
2567 if (DestVT == MVT::f32)
2568 FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx,
2569 MachinePointerInfo::getConstantPool(),
2570 false, false, false, Alignment);
2572 SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT,
2573 DAG.getEntryNode(), CPIdx,
2574 MachinePointerInfo::getConstantPool(),
2575 MVT::f32, false, false, false, Alignment);
2576 HandleSDNode Handle(Load);
2577 LegalizeOp(Load.getNode());
2578 FudgeInReg = Handle.getValue();
2581 return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg);
2584 /// This function is responsible for legalizing a
2585 /// *INT_TO_FP operation of the specified operand when the target requests that
2586 /// we promote it. At this point, we know that the result and operand types are
2587 /// legal for the target, and that there is a legal UINT_TO_FP or SINT_TO_FP
2588 /// operation that takes a larger input.
2589 SDValue SelectionDAGLegalize::PromoteLegalINT_TO_FP(SDValue LegalOp,
2593 // First step, figure out the appropriate *INT_TO_FP operation to use.
2594 EVT NewInTy = LegalOp.getValueType();
2596 unsigned OpToUse = 0;
2598 // Scan for the appropriate larger type to use.
2600 NewInTy = (MVT::SimpleValueType)(NewInTy.getSimpleVT().SimpleTy+1);
2601 assert(NewInTy.isInteger() && "Ran out of possibilities!");
2603 // If the target supports SINT_TO_FP of this type, use it.
2604 if (TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, NewInTy)) {
2605 OpToUse = ISD::SINT_TO_FP;
2608 if (isSigned) continue;
2610 // If the target supports UINT_TO_FP of this type, use it.
2611 if (TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, NewInTy)) {
2612 OpToUse = ISD::UINT_TO_FP;
2616 // Otherwise, try a larger type.
2619 // Okay, we found the operation and type to use. Zero extend our input to the
2620 // desired type then run the operation on it.
2621 return DAG.getNode(OpToUse, dl, DestVT,
2622 DAG.getNode(isSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND,
2623 dl, NewInTy, LegalOp));
2626 /// This function is responsible for legalizing a
2627 /// FP_TO_*INT operation of the specified operand when the target requests that
2628 /// we promote it. At this point, we know that the result and operand types are
2629 /// legal for the target, and that there is a legal FP_TO_UINT or FP_TO_SINT
2630 /// operation that returns a larger result.
2631 SDValue SelectionDAGLegalize::PromoteLegalFP_TO_INT(SDValue LegalOp,
2635 // First step, figure out the appropriate FP_TO*INT operation to use.
2636 EVT NewOutTy = DestVT;
2638 unsigned OpToUse = 0;
2640 // Scan for the appropriate larger type to use.
2642 NewOutTy = (MVT::SimpleValueType)(NewOutTy.getSimpleVT().SimpleTy+1);
2643 assert(NewOutTy.isInteger() && "Ran out of possibilities!");
2645 // A larger signed type can hold all unsigned values of the requested type,
2646 // so using FP_TO_SINT is valid
2647 if (TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NewOutTy)) {
2648 OpToUse = ISD::FP_TO_SINT;
2652 // However, if the value may be < 0.0, we *must* use some FP_TO_SINT.
2653 if (!isSigned && TLI.isOperationLegalOrCustom(ISD::FP_TO_UINT, NewOutTy)) {
2654 OpToUse = ISD::FP_TO_UINT;
2658 // Otherwise, try a larger type.
2662 // Okay, we found the operation and type to use.
2663 SDValue Operation = DAG.getNode(OpToUse, dl, NewOutTy, LegalOp);
2665 // Truncate the result of the extended FP_TO_*INT operation to the desired
2667 return DAG.getNode(ISD::TRUNCATE, dl, DestVT, Operation);
2670 /// Open code the operations for BSWAP of the specified operation.
2671 SDValue SelectionDAGLegalize::ExpandBSWAP(SDValue Op, SDLoc dl) {
2672 EVT VT = Op.getValueType();
2673 EVT SHVT = TLI.getShiftAmountTy(VT);
2674 SDValue Tmp1, Tmp2, Tmp3, Tmp4, Tmp5, Tmp6, Tmp7, Tmp8;
2675 switch (VT.getSimpleVT().SimpleTy) {
2676 default: llvm_unreachable("Unhandled Expand type in BSWAP!");
2678 Tmp2 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2679 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2680 return DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
2682 Tmp4 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2683 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2684 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2685 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2686 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(0xFF0000, VT));
2687 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(0xFF00, VT));
2688 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2689 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2690 return DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2692 Tmp8 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(56, SHVT));
2693 Tmp7 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(40, SHVT));
2694 Tmp6 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(24, SHVT));
2695 Tmp5 = DAG.getNode(ISD::SHL, dl, VT, Op, DAG.getConstant(8, SHVT));
2696 Tmp4 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(8, SHVT));
2697 Tmp3 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(24, SHVT));
2698 Tmp2 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(40, SHVT));
2699 Tmp1 = DAG.getNode(ISD::SRL, dl, VT, Op, DAG.getConstant(56, SHVT));
2700 Tmp7 = DAG.getNode(ISD::AND, dl, VT, Tmp7, DAG.getConstant(255ULL<<48, VT));
2701 Tmp6 = DAG.getNode(ISD::AND, dl, VT, Tmp6, DAG.getConstant(255ULL<<40, VT));
2702 Tmp5 = DAG.getNode(ISD::AND, dl, VT, Tmp5, DAG.getConstant(255ULL<<32, VT));
2703 Tmp4 = DAG.getNode(ISD::AND, dl, VT, Tmp4, DAG.getConstant(255ULL<<24, VT));
2704 Tmp3 = DAG.getNode(ISD::AND, dl, VT, Tmp3, DAG.getConstant(255ULL<<16, VT));
2705 Tmp2 = DAG.getNode(ISD::AND, dl, VT, Tmp2, DAG.getConstant(255ULL<<8 , VT));
2706 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp7);
2707 Tmp6 = DAG.getNode(ISD::OR, dl, VT, Tmp6, Tmp5);
2708 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp3);
2709 Tmp2 = DAG.getNode(ISD::OR, dl, VT, Tmp2, Tmp1);
2710 Tmp8 = DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp6);
2711 Tmp4 = DAG.getNode(ISD::OR, dl, VT, Tmp4, Tmp2);
2712 return DAG.getNode(ISD::OR, dl, VT, Tmp8, Tmp4);
2716 /// Expand the specified bitcount instruction into operations.
2717 SDValue SelectionDAGLegalize::ExpandBitCount(unsigned Opc, SDValue Op,
2720 default: llvm_unreachable("Cannot expand this yet!");
2722 EVT VT = Op.getValueType();
2723 EVT ShVT = TLI.getShiftAmountTy(VT);
2724 unsigned Len = VT.getSizeInBits();
2726 assert(VT.isInteger() && Len <= 128 && Len % 8 == 0 &&
2727 "CTPOP not implemented for this type.");
2729 // This is the "best" algorithm from
2730 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
2732 SDValue Mask55 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x55)), VT);
2733 SDValue Mask33 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x33)), VT);
2734 SDValue Mask0F = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x0F)), VT);
2735 SDValue Mask01 = DAG.getConstant(APInt::getSplat(Len, APInt(8, 0x01)), VT);
2737 // v = v - ((v >> 1) & 0x55555555...)
2738 Op = DAG.getNode(ISD::SUB, dl, VT, Op,
2739 DAG.getNode(ISD::AND, dl, VT,
2740 DAG.getNode(ISD::SRL, dl, VT, Op,
2741 DAG.getConstant(1, ShVT)),
2743 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
2744 Op = DAG.getNode(ISD::ADD, dl, VT,
2745 DAG.getNode(ISD::AND, dl, VT, Op, Mask33),
2746 DAG.getNode(ISD::AND, dl, VT,
2747 DAG.getNode(ISD::SRL, dl, VT, Op,
2748 DAG.getConstant(2, ShVT)),
2750 // v = (v + (v >> 4)) & 0x0F0F0F0F...
2751 Op = DAG.getNode(ISD::AND, dl, VT,
2752 DAG.getNode(ISD::ADD, dl, VT, Op,
2753 DAG.getNode(ISD::SRL, dl, VT, Op,
2754 DAG.getConstant(4, ShVT))),
2756 // v = (v * 0x01010101...) >> (Len - 8)
2757 Op = DAG.getNode(ISD::SRL, dl, VT,
2758 DAG.getNode(ISD::MUL, dl, VT, Op, Mask01),
2759 DAG.getConstant(Len - 8, ShVT));
2763 case ISD::CTLZ_ZERO_UNDEF:
2764 // This trivially expands to CTLZ.
2765 return DAG.getNode(ISD::CTLZ, dl, Op.getValueType(), Op);
2767 // for now, we do this:
2768 // x = x | (x >> 1);
2769 // x = x | (x >> 2);
2771 // x = x | (x >>16);
2772 // x = x | (x >>32); // for 64-bit input
2773 // return popcount(~x);
2775 // Ref: "Hacker's Delight" by Henry Warren
2776 EVT VT = Op.getValueType();
2777 EVT ShVT = TLI.getShiftAmountTy(VT);
2778 unsigned len = VT.getSizeInBits();
2779 for (unsigned i = 0; (1U << i) <= (len / 2); ++i) {
2780 SDValue Tmp3 = DAG.getConstant(1ULL << i, ShVT);
2781 Op = DAG.getNode(ISD::OR, dl, VT, Op,
2782 DAG.getNode(ISD::SRL, dl, VT, Op, Tmp3));
2784 Op = DAG.getNOT(dl, Op, VT);
2785 return DAG.getNode(ISD::CTPOP, dl, VT, Op);
2787 case ISD::CTTZ_ZERO_UNDEF:
2788 // This trivially expands to CTTZ.
2789 return DAG.getNode(ISD::CTTZ, dl, Op.getValueType(), Op);
2791 // for now, we use: { return popcount(~x & (x - 1)); }
2792 // unless the target has ctlz but not ctpop, in which case we use:
2793 // { return 32 - nlz(~x & (x-1)); }
2794 // Ref: "Hacker's Delight" by Henry Warren
2795 EVT VT = Op.getValueType();
2796 SDValue Tmp3 = DAG.getNode(ISD::AND, dl, VT,
2797 DAG.getNOT(dl, Op, VT),
2798 DAG.getNode(ISD::SUB, dl, VT, Op,
2799 DAG.getConstant(1, VT)));
2800 // If ISD::CTLZ is legal and CTPOP isn't, then do that instead.
2801 if (!TLI.isOperationLegalOrCustom(ISD::CTPOP, VT) &&
2802 TLI.isOperationLegalOrCustom(ISD::CTLZ, VT))
2803 return DAG.getNode(ISD::SUB, dl, VT,
2804 DAG.getConstant(VT.getSizeInBits(), VT),
2805 DAG.getNode(ISD::CTLZ, dl, VT, Tmp3));
2806 return DAG.getNode(ISD::CTPOP, dl, VT, Tmp3);
2811 std::pair <SDValue, SDValue> SelectionDAGLegalize::ExpandAtomic(SDNode *Node) {
2812 unsigned Opc = Node->getOpcode();
2813 MVT VT = cast<AtomicSDNode>(Node)->getMemoryVT().getSimpleVT();
2818 llvm_unreachable("Unhandled atomic intrinsic Expand!");
2819 case ISD::ATOMIC_SWAP:
2820 switch (VT.SimpleTy) {
2821 default: llvm_unreachable("Unexpected value type for atomic!");
2822 case MVT::i8: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_1; break;
2823 case MVT::i16: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_2; break;
2824 case MVT::i32: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_4; break;
2825 case MVT::i64: LC = RTLIB::SYNC_LOCK_TEST_AND_SET_8; break;
2826 case MVT::i128:LC = RTLIB::SYNC_LOCK_TEST_AND_SET_16;break;
2829 case ISD::ATOMIC_CMP_SWAP:
2830 switch (VT.SimpleTy) {
2831 default: llvm_unreachable("Unexpected value type for atomic!");
2832 case MVT::i8: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1; break;
2833 case MVT::i16: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2; break;
2834 case MVT::i32: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4; break;
2835 case MVT::i64: LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8; break;
2836 case MVT::i128:LC = RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16;break;
2839 case ISD::ATOMIC_LOAD_ADD:
2840 switch (VT.SimpleTy) {
2841 default: llvm_unreachable("Unexpected value type for atomic!");
2842 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_ADD_1; break;
2843 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_ADD_2; break;
2844 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_ADD_4; break;
2845 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_ADD_8; break;
2846 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_ADD_16;break;
2849 case ISD::ATOMIC_LOAD_SUB:
2850 switch (VT.SimpleTy) {
2851 default: llvm_unreachable("Unexpected value type for atomic!");
2852 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_SUB_1; break;
2853 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_SUB_2; break;
2854 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_SUB_4; break;
2855 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_SUB_8; break;
2856 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_SUB_16;break;
2859 case ISD::ATOMIC_LOAD_AND:
2860 switch (VT.SimpleTy) {
2861 default: llvm_unreachable("Unexpected value type for atomic!");
2862 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_AND_1; break;
2863 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_AND_2; break;
2864 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_AND_4; break;
2865 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_AND_8; break;
2866 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_AND_16;break;
2869 case ISD::ATOMIC_LOAD_OR:
2870 switch (VT.SimpleTy) {
2871 default: llvm_unreachable("Unexpected value type for atomic!");
2872 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_OR_1; break;
2873 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_OR_2; break;
2874 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_OR_4; break;
2875 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_OR_8; break;
2876 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_OR_16;break;
2879 case ISD::ATOMIC_LOAD_XOR:
2880 switch (VT.SimpleTy) {
2881 default: llvm_unreachable("Unexpected value type for atomic!");
2882 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_XOR_1; break;
2883 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_XOR_2; break;
2884 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_XOR_4; break;
2885 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_XOR_8; break;
2886 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_XOR_16;break;
2889 case ISD::ATOMIC_LOAD_NAND:
2890 switch (VT.SimpleTy) {
2891 default: llvm_unreachable("Unexpected value type for atomic!");
2892 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_NAND_1; break;
2893 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_NAND_2; break;
2894 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_NAND_4; break;
2895 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_NAND_8; break;
2896 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_NAND_16;break;
2899 case ISD::ATOMIC_LOAD_MAX:
2900 switch (VT.SimpleTy) {
2901 default: llvm_unreachable("Unexpected value type for atomic!");
2902 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MAX_1; break;
2903 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MAX_2; break;
2904 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MAX_4; break;
2905 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MAX_8; break;
2906 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MAX_16;break;
2909 case ISD::ATOMIC_LOAD_UMAX:
2910 switch (VT.SimpleTy) {
2911 default: llvm_unreachable("Unexpected value type for atomic!");
2912 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMAX_1; break;
2913 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMAX_2; break;
2914 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMAX_4; break;
2915 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMAX_8; break;
2916 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMAX_16;break;
2919 case ISD::ATOMIC_LOAD_MIN:
2920 switch (VT.SimpleTy) {
2921 default: llvm_unreachable("Unexpected value type for atomic!");
2922 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_MIN_1; break;
2923 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_MIN_2; break;
2924 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_MIN_4; break;
2925 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_MIN_8; break;
2926 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_MIN_16;break;
2929 case ISD::ATOMIC_LOAD_UMIN:
2930 switch (VT.SimpleTy) {
2931 default: llvm_unreachable("Unexpected value type for atomic!");
2932 case MVT::i8: LC = RTLIB::SYNC_FETCH_AND_UMIN_1; break;
2933 case MVT::i16: LC = RTLIB::SYNC_FETCH_AND_UMIN_2; break;
2934 case MVT::i32: LC = RTLIB::SYNC_FETCH_AND_UMIN_4; break;
2935 case MVT::i64: LC = RTLIB::SYNC_FETCH_AND_UMIN_8; break;
2936 case MVT::i128:LC = RTLIB::SYNC_FETCH_AND_UMIN_16;break;
2941 return ExpandChainLibCall(LC, Node, false);
2944 void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
2945 SmallVector<SDValue, 8> Results;
2947 SDValue Tmp1, Tmp2, Tmp3, Tmp4;
2949 switch (Node->getOpcode()) {
2952 case ISD::CTLZ_ZERO_UNDEF:
2954 case ISD::CTTZ_ZERO_UNDEF:
2955 Tmp1 = ExpandBitCount(Node->getOpcode(), Node->getOperand(0), dl);
2956 Results.push_back(Tmp1);
2959 Results.push_back(ExpandBSWAP(Node->getOperand(0), dl));
2961 case ISD::FRAMEADDR:
2962 case ISD::RETURNADDR:
2963 case ISD::FRAME_TO_ARGS_OFFSET:
2964 Results.push_back(DAG.getConstant(0, Node->getValueType(0)));
2966 case ISD::FLT_ROUNDS_:
2967 Results.push_back(DAG.getConstant(1, Node->getValueType(0)));
2969 case ISD::EH_RETURN:
2973 case ISD::EH_SJLJ_LONGJMP:
2974 // If the target didn't expand these, there's nothing to do, so just
2975 // preserve the chain and be done.
2976 Results.push_back(Node->getOperand(0));
2978 case ISD::EH_SJLJ_SETJMP:
2979 // If the target didn't expand this, just return 'zero' and preserve the
2981 Results.push_back(DAG.getConstant(0, MVT::i32));
2982 Results.push_back(Node->getOperand(0));
2984 case ISD::ATOMIC_FENCE: {
2985 // If the target didn't lower this, lower it to '__sync_synchronize()' call
2986 // FIXME: handle "fence singlethread" more efficiently.
2987 TargetLowering::ArgListTy Args;
2989 TargetLowering::CallLoweringInfo CLI(DAG);
2990 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
2991 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
2992 DAG.getExternalSymbol("__sync_synchronize",
2993 TLI.getPointerTy()), std::move(Args), 0);
2995 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
2997 Results.push_back(CallResult.second);
3000 case ISD::ATOMIC_LOAD: {
3001 // There is no libcall for atomic load; fake it with ATOMIC_CMP_SWAP.
3002 SDValue Zero = DAG.getConstant(0, Node->getValueType(0));
3003 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3004 SDValue Swap = DAG.getAtomicCmpSwap(
3005 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3006 Node->getOperand(0), Node->getOperand(1), Zero, Zero,
3007 cast<AtomicSDNode>(Node)->getMemOperand(),
3008 cast<AtomicSDNode>(Node)->getOrdering(),
3009 cast<AtomicSDNode>(Node)->getOrdering(),
3010 cast<AtomicSDNode>(Node)->getSynchScope());
3011 Results.push_back(Swap.getValue(0));
3012 Results.push_back(Swap.getValue(1));
3015 case ISD::ATOMIC_STORE: {
3016 // There is no libcall for atomic store; fake it with ATOMIC_SWAP.
3017 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
3018 cast<AtomicSDNode>(Node)->getMemoryVT(),
3019 Node->getOperand(0),
3020 Node->getOperand(1), Node->getOperand(2),
3021 cast<AtomicSDNode>(Node)->getMemOperand(),
3022 cast<AtomicSDNode>(Node)->getOrdering(),
3023 cast<AtomicSDNode>(Node)->getSynchScope());
3024 Results.push_back(Swap.getValue(1));
3027 // By default, atomic intrinsics are marked Legal and lowered. Targets
3028 // which don't support them directly, however, may want libcalls, in which
3029 // case they mark them Expand, and we get here.
3030 case ISD::ATOMIC_SWAP:
3031 case ISD::ATOMIC_LOAD_ADD:
3032 case ISD::ATOMIC_LOAD_SUB:
3033 case ISD::ATOMIC_LOAD_AND:
3034 case ISD::ATOMIC_LOAD_OR:
3035 case ISD::ATOMIC_LOAD_XOR:
3036 case ISD::ATOMIC_LOAD_NAND:
3037 case ISD::ATOMIC_LOAD_MIN:
3038 case ISD::ATOMIC_LOAD_MAX:
3039 case ISD::ATOMIC_LOAD_UMIN:
3040 case ISD::ATOMIC_LOAD_UMAX:
3041 case ISD::ATOMIC_CMP_SWAP: {
3042 std::pair<SDValue, SDValue> Tmp = ExpandAtomic(Node);
3043 Results.push_back(Tmp.first);
3044 Results.push_back(Tmp.second);
3047 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
3048 // Expanding an ATOMIC_CMP_SWAP_WITH_SUCCESS produces an ATOMIC_CMP_SWAP and
3049 // splits out the success value as a comparison. Expanding the resulting
3050 // ATOMIC_CMP_SWAP will produce a libcall.
3051 SDVTList VTs = DAG.getVTList(Node->getValueType(0), MVT::Other);
3052 SDValue Res = DAG.getAtomicCmpSwap(
3053 ISD::ATOMIC_CMP_SWAP, dl, cast<AtomicSDNode>(Node)->getMemoryVT(), VTs,
3054 Node->getOperand(0), Node->getOperand(1), Node->getOperand(2),
3055 Node->getOperand(3), cast<MemSDNode>(Node)->getMemOperand(),
3056 cast<AtomicSDNode>(Node)->getSuccessOrdering(),
3057 cast<AtomicSDNode>(Node)->getFailureOrdering(),
3058 cast<AtomicSDNode>(Node)->getSynchScope());
3060 SDValue Success = DAG.getSetCC(SDLoc(Node), Node->getValueType(1),
3061 Res, Node->getOperand(2), ISD::SETEQ);
3063 Results.push_back(Res.getValue(0));
3064 Results.push_back(Success);
3065 Results.push_back(Res.getValue(1));
3068 case ISD::DYNAMIC_STACKALLOC:
3069 ExpandDYNAMIC_STACKALLOC(Node, Results);
3071 case ISD::MERGE_VALUES:
3072 for (unsigned i = 0; i < Node->getNumValues(); i++)
3073 Results.push_back(Node->getOperand(i));
3076 EVT VT = Node->getValueType(0);
3078 Results.push_back(DAG.getConstant(0, VT));
3080 assert(VT.isFloatingPoint() && "Unknown value type!");
3081 Results.push_back(DAG.getConstantFP(0, VT));
3086 // If this operation is not supported, lower it to 'abort()' call
3087 TargetLowering::ArgListTy Args;
3088 TargetLowering::CallLoweringInfo CLI(DAG);
3089 CLI.setDebugLoc(dl).setChain(Node->getOperand(0))
3090 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
3091 DAG.getExternalSymbol("abort", TLI.getPointerTy()),
3092 std::move(Args), 0);
3093 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
3095 Results.push_back(CallResult.second);
3100 Tmp1 = EmitStackConvert(Node->getOperand(0), Node->getValueType(0),
3101 Node->getValueType(0), dl);
3102 Results.push_back(Tmp1);
3104 case ISD::FP_EXTEND:
3105 Tmp1 = EmitStackConvert(Node->getOperand(0),
3106 Node->getOperand(0).getValueType(),
3107 Node->getValueType(0), dl);
3108 Results.push_back(Tmp1);
3110 case ISD::SIGN_EXTEND_INREG: {
3111 // NOTE: we could fall back on load/store here too for targets without
3112 // SAR. However, it is doubtful that any exist.
3113 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3114 EVT VT = Node->getValueType(0);
3115 EVT ShiftAmountTy = TLI.getShiftAmountTy(VT);
3118 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
3119 ExtraVT.getScalarType().getSizeInBits();
3120 SDValue ShiftCst = DAG.getConstant(BitsDiff, ShiftAmountTy);
3121 Tmp1 = DAG.getNode(ISD::SHL, dl, Node->getValueType(0),
3122 Node->getOperand(0), ShiftCst);
3123 Tmp1 = DAG.getNode(ISD::SRA, dl, Node->getValueType(0), Tmp1, ShiftCst);
3124 Results.push_back(Tmp1);
3127 case ISD::FP_ROUND_INREG: {
3128 // The only way we can lower this is to turn it into a TRUNCSTORE,
3129 // EXTLOAD pair, targeting a temporary location (a stack slot).
3131 // NOTE: there is a choice here between constantly creating new stack
3132 // slots and always reusing the same one. We currently always create
3133 // new ones, as reuse may inhibit scheduling.
3134 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
3135 Tmp1 = EmitStackConvert(Node->getOperand(0), ExtraVT,
3136 Node->getValueType(0), dl);
3137 Results.push_back(Tmp1);
3140 case ISD::SINT_TO_FP:
3141 case ISD::UINT_TO_FP:
3142 Tmp1 = ExpandLegalINT_TO_FP(Node->getOpcode() == ISD::SINT_TO_FP,
3143 Node->getOperand(0), Node->getValueType(0), dl);
3144 Results.push_back(Tmp1);
3146 case ISD::FP_TO_SINT:
3147 if (TLI.expandFP_TO_SINT(Node, Tmp1, DAG))
3148 Results.push_back(Tmp1);
3150 case ISD::FP_TO_UINT: {
3151 SDValue True, False;
3152 EVT VT = Node->getOperand(0).getValueType();
3153 EVT NVT = Node->getValueType(0);
3154 APFloat apf(DAG.EVTToAPFloatSemantics(VT),
3155 APInt::getNullValue(VT.getSizeInBits()));
3156 APInt x = APInt::getSignBit(NVT.getSizeInBits());
3157 (void)apf.convertFromAPInt(x, false, APFloat::rmNearestTiesToEven);
3158 Tmp1 = DAG.getConstantFP(apf, VT);
3159 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(VT),
3160 Node->getOperand(0),
3162 True = DAG.getNode(ISD::FP_TO_SINT, dl, NVT, Node->getOperand(0));
3163 False = DAG.getNode(ISD::FP_TO_SINT, dl, NVT,
3164 DAG.getNode(ISD::FSUB, dl, VT,
3165 Node->getOperand(0), Tmp1));
3166 False = DAG.getNode(ISD::XOR, dl, NVT, False,
3167 DAG.getConstant(x, NVT));
3168 Tmp1 = DAG.getSelect(dl, NVT, Tmp2, True, False);
3169 Results.push_back(Tmp1);
3173 const Value *V = cast<SrcValueSDNode>(Node->getOperand(2))->getValue();
3174 EVT VT = Node->getValueType(0);
3175 Tmp1 = Node->getOperand(0);
3176 Tmp2 = Node->getOperand(1);
3177 unsigned Align = Node->getConstantOperandVal(3);
3179 SDValue VAListLoad = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2,
3180 MachinePointerInfo(V),
3181 false, false, false, 0);
3182 SDValue VAList = VAListLoad;
3184 if (Align > TLI.getMinStackArgumentAlignment()) {
3185 assert(((Align & (Align-1)) == 0) && "Expected Align to be a power of 2");
3187 VAList = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3188 DAG.getConstant(Align - 1,
3189 VAList.getValueType()));
3191 VAList = DAG.getNode(ISD::AND, dl, VAList.getValueType(), VAList,
3192 DAG.getConstant(-(int64_t)Align,
3193 VAList.getValueType()));
3196 // Increment the pointer, VAList, to the next vaarg
3197 Tmp3 = DAG.getNode(ISD::ADD, dl, VAList.getValueType(), VAList,
3198 DAG.getConstant(TLI.getDataLayout()->
3199 getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())),
3200 VAList.getValueType()));
3201 // Store the incremented VAList to the legalized pointer
3202 Tmp3 = DAG.getStore(VAListLoad.getValue(1), dl, Tmp3, Tmp2,
3203 MachinePointerInfo(V), false, false, 0);
3204 // Load the actual argument out of the pointer VAList
3205 Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, MachinePointerInfo(),
3206 false, false, false, 0));
3207 Results.push_back(Results[0].getValue(1));
3211 // This defaults to loading a pointer from the input and storing it to the
3212 // output, returning the chain.
3213 const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue();
3214 const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue();
3215 Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0),
3216 Node->getOperand(2), MachinePointerInfo(VS),
3217 false, false, false, 0);
3218 Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1),
3219 MachinePointerInfo(VD), false, false, 0);
3220 Results.push_back(Tmp1);
3223 case ISD::EXTRACT_VECTOR_ELT:
3224 if (Node->getOperand(0).getValueType().getVectorNumElements() == 1)
3225 // This must be an access of the only element. Return it.
3226 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0),
3227 Node->getOperand(0));
3229 Tmp1 = ExpandExtractFromVectorThroughStack(SDValue(Node, 0));
3230 Results.push_back(Tmp1);
3232 case ISD::EXTRACT_SUBVECTOR:
3233 Results.push_back(ExpandExtractFromVectorThroughStack(SDValue(Node, 0)));
3235 case ISD::INSERT_SUBVECTOR:
3236 Results.push_back(ExpandInsertToVectorThroughStack(SDValue(Node, 0)));
3238 case ISD::CONCAT_VECTORS: {
3239 Results.push_back(ExpandVectorBuildThroughStack(Node));
3242 case ISD::SCALAR_TO_VECTOR:
3243 Results.push_back(ExpandSCALAR_TO_VECTOR(Node));
3245 case ISD::INSERT_VECTOR_ELT:
3246 Results.push_back(ExpandINSERT_VECTOR_ELT(Node->getOperand(0),
3247 Node->getOperand(1),
3248 Node->getOperand(2), dl));
3250 case ISD::VECTOR_SHUFFLE: {
3251 SmallVector<int, 32> NewMask;
3252 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
3254 EVT VT = Node->getValueType(0);
3255 EVT EltVT = VT.getVectorElementType();
3256 SDValue Op0 = Node->getOperand(0);
3257 SDValue Op1 = Node->getOperand(1);
3258 if (!TLI.isTypeLegal(EltVT)) {
3260 EVT NewEltVT = TLI.getTypeToTransformTo(*DAG.getContext(), EltVT);
3262 // BUILD_VECTOR operands are allowed to be wider than the element type.
3263 // But if NewEltVT is smaller that EltVT the BUILD_VECTOR does not accept
3265 if (NewEltVT.bitsLT(EltVT)) {
3267 // Convert shuffle node.
3268 // If original node was v4i64 and the new EltVT is i32,
3269 // cast operands to v8i32 and re-build the mask.
3271 // Calculate new VT, the size of the new VT should be equal to original.
3273 EVT::getVectorVT(*DAG.getContext(), NewEltVT,
3274 VT.getSizeInBits() / NewEltVT.getSizeInBits());
3275 assert(NewVT.bitsEq(VT));
3277 // cast operands to new VT
3278 Op0 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op0);
3279 Op1 = DAG.getNode(ISD::BITCAST, dl, NewVT, Op1);
3281 // Convert the shuffle mask
3282 unsigned int factor =
3283 NewVT.getVectorNumElements()/VT.getVectorNumElements();
3285 // EltVT gets smaller
3288 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
3290 for (unsigned fi = 0; fi < factor; ++fi)
3291 NewMask.push_back(Mask[i]);
3294 for (unsigned fi = 0; fi < factor; ++fi)
3295 NewMask.push_back(Mask[i]*factor+fi);
3303 unsigned NumElems = VT.getVectorNumElements();
3304 SmallVector<SDValue, 16> Ops;
3305 for (unsigned i = 0; i != NumElems; ++i) {
3307 Ops.push_back(DAG.getUNDEF(EltVT));
3310 unsigned Idx = Mask[i];
3312 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3314 DAG.getConstant(Idx, TLI.getVectorIdxTy())));
3316 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
3318 DAG.getConstant(Idx - NumElems,
3319 TLI.getVectorIdxTy())));
3322 Tmp1 = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
3323 // We may have changed the BUILD_VECTOR type. Cast it back to the Node type.
3324 Tmp1 = DAG.getNode(ISD::BITCAST, dl, Node->getValueType(0), Tmp1);
3325 Results.push_back(Tmp1);
3328 case ISD::EXTRACT_ELEMENT: {
3329 EVT OpTy = Node->getOperand(0).getValueType();
3330 if (cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue()) {
3332 Tmp1 = DAG.getNode(ISD::SRL, dl, OpTy, Node->getOperand(0),
3333 DAG.getConstant(OpTy.getSizeInBits()/2,
3334 TLI.getShiftAmountTy(Node->getOperand(0).getValueType())));
3335 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0), Tmp1);
3338 Tmp1 = DAG.getNode(ISD::TRUNCATE, dl, Node->getValueType(0),
3339 Node->getOperand(0));
3341 Results.push_back(Tmp1);
3344 case ISD::STACKSAVE:
3345 // Expand to CopyFromReg if the target set
3346 // StackPointerRegisterToSaveRestore.
3347 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3348 Results.push_back(DAG.getCopyFromReg(Node->getOperand(0), dl, SP,
3349 Node->getValueType(0)));
3350 Results.push_back(Results[0].getValue(1));
3352 Results.push_back(DAG.getUNDEF(Node->getValueType(0)));
3353 Results.push_back(Node->getOperand(0));
3356 case ISD::STACKRESTORE:
3357 // Expand to CopyToReg if the target set
3358 // StackPointerRegisterToSaveRestore.
3359 if (unsigned SP = TLI.getStackPointerRegisterToSaveRestore()) {
3360 Results.push_back(DAG.getCopyToReg(Node->getOperand(0), dl, SP,
3361 Node->getOperand(1)));
3363 Results.push_back(Node->getOperand(0));
3366 case ISD::FCOPYSIGN:
3367 Results.push_back(ExpandFCOPYSIGN(Node));
3370 // Expand Y = FNEG(X) -> Y = SUB -0.0, X
3371 Tmp1 = DAG.getConstantFP(-0.0, Node->getValueType(0));
3372 Tmp1 = DAG.getNode(ISD::FSUB, dl, Node->getValueType(0), Tmp1,
3373 Node->getOperand(0));
3374 Results.push_back(Tmp1);
3377 // Expand Y = FABS(X) -> Y = (X >u 0.0) ? X : fneg(X).
3378 EVT VT = Node->getValueType(0);
3379 Tmp1 = Node->getOperand(0);
3380 Tmp2 = DAG.getConstantFP(0.0, VT);
3381 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(Tmp1.getValueType()),
3382 Tmp1, Tmp2, ISD::SETUGT);
3383 Tmp3 = DAG.getNode(ISD::FNEG, dl, VT, Tmp1);
3384 Tmp1 = DAG.getSelect(dl, VT, Tmp2, Tmp1, Tmp3);
3385 Results.push_back(Tmp1);
3389 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMIN_F32, RTLIB::FMIN_F64,
3390 RTLIB::FMIN_F80, RTLIB::FMIN_F128,
3391 RTLIB::FMIN_PPCF128));
3394 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMAX_F32, RTLIB::FMAX_F64,
3395 RTLIB::FMAX_F80, RTLIB::FMAX_F128,
3396 RTLIB::FMAX_PPCF128));
3399 Results.push_back(ExpandFPLibCall(Node, RTLIB::SQRT_F32, RTLIB::SQRT_F64,
3400 RTLIB::SQRT_F80, RTLIB::SQRT_F128,
3401 RTLIB::SQRT_PPCF128));
3405 EVT VT = Node->getValueType(0);
3406 bool isSIN = Node->getOpcode() == ISD::FSIN;
3407 // Turn fsin / fcos into ISD::FSINCOS node if there are a pair of fsin /
3408 // fcos which share the same operand and both are used.
3409 if ((TLI.isOperationLegalOrCustom(ISD::FSINCOS, VT) ||
3410 canCombineSinCosLibcall(Node, TLI, TM))
3411 && useSinCos(Node)) {
3412 SDVTList VTs = DAG.getVTList(VT, VT);
3413 Tmp1 = DAG.getNode(ISD::FSINCOS, dl, VTs, Node->getOperand(0));
3415 Tmp1 = Tmp1.getValue(1);
3416 Results.push_back(Tmp1);
3418 Results.push_back(ExpandFPLibCall(Node, RTLIB::SIN_F32, RTLIB::SIN_F64,
3419 RTLIB::SIN_F80, RTLIB::SIN_F128,
3420 RTLIB::SIN_PPCF128));
3422 Results.push_back(ExpandFPLibCall(Node, RTLIB::COS_F32, RTLIB::COS_F64,
3423 RTLIB::COS_F80, RTLIB::COS_F128,
3424 RTLIB::COS_PPCF128));
3429 // Expand into sincos libcall.
3430 ExpandSinCosLibCall(Node, Results);
3433 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG_F32, RTLIB::LOG_F64,
3434 RTLIB::LOG_F80, RTLIB::LOG_F128,
3435 RTLIB::LOG_PPCF128));
3438 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG2_F32, RTLIB::LOG2_F64,
3439 RTLIB::LOG2_F80, RTLIB::LOG2_F128,
3440 RTLIB::LOG2_PPCF128));
3443 Results.push_back(ExpandFPLibCall(Node, RTLIB::LOG10_F32, RTLIB::LOG10_F64,
3444 RTLIB::LOG10_F80, RTLIB::LOG10_F128,
3445 RTLIB::LOG10_PPCF128));
3448 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP_F32, RTLIB::EXP_F64,
3449 RTLIB::EXP_F80, RTLIB::EXP_F128,
3450 RTLIB::EXP_PPCF128));
3453 Results.push_back(ExpandFPLibCall(Node, RTLIB::EXP2_F32, RTLIB::EXP2_F64,
3454 RTLIB::EXP2_F80, RTLIB::EXP2_F128,
3455 RTLIB::EXP2_PPCF128));
3458 Results.push_back(ExpandFPLibCall(Node, RTLIB::TRUNC_F32, RTLIB::TRUNC_F64,
3459 RTLIB::TRUNC_F80, RTLIB::TRUNC_F128,
3460 RTLIB::TRUNC_PPCF128));
3463 Results.push_back(ExpandFPLibCall(Node, RTLIB::FLOOR_F32, RTLIB::FLOOR_F64,
3464 RTLIB::FLOOR_F80, RTLIB::FLOOR_F128,
3465 RTLIB::FLOOR_PPCF128));
3468 Results.push_back(ExpandFPLibCall(Node, RTLIB::CEIL_F32, RTLIB::CEIL_F64,
3469 RTLIB::CEIL_F80, RTLIB::CEIL_F128,
3470 RTLIB::CEIL_PPCF128));
3473 Results.push_back(ExpandFPLibCall(Node, RTLIB::RINT_F32, RTLIB::RINT_F64,
3474 RTLIB::RINT_F80, RTLIB::RINT_F128,
3475 RTLIB::RINT_PPCF128));
3477 case ISD::FNEARBYINT:
3478 Results.push_back(ExpandFPLibCall(Node, RTLIB::NEARBYINT_F32,
3479 RTLIB::NEARBYINT_F64,
3480 RTLIB::NEARBYINT_F80,
3481 RTLIB::NEARBYINT_F128,
3482 RTLIB::NEARBYINT_PPCF128));
3485 Results.push_back(ExpandFPLibCall(Node, RTLIB::ROUND_F32,
3489 RTLIB::ROUND_PPCF128));
3492 Results.push_back(ExpandFPLibCall(Node, RTLIB::POWI_F32, RTLIB::POWI_F64,
3493 RTLIB::POWI_F80, RTLIB::POWI_F128,
3494 RTLIB::POWI_PPCF128));
3497 Results.push_back(ExpandFPLibCall(Node, RTLIB::POW_F32, RTLIB::POW_F64,
3498 RTLIB::POW_F80, RTLIB::POW_F128,
3499 RTLIB::POW_PPCF128));
3502 Results.push_back(ExpandFPLibCall(Node, RTLIB::DIV_F32, RTLIB::DIV_F64,
3503 RTLIB::DIV_F80, RTLIB::DIV_F128,
3504 RTLIB::DIV_PPCF128));
3507 Results.push_back(ExpandFPLibCall(Node, RTLIB::REM_F32, RTLIB::REM_F64,
3508 RTLIB::REM_F80, RTLIB::REM_F128,
3509 RTLIB::REM_PPCF128));
3512 Results.push_back(ExpandFPLibCall(Node, RTLIB::FMA_F32, RTLIB::FMA_F64,
3513 RTLIB::FMA_F80, RTLIB::FMA_F128,
3514 RTLIB::FMA_PPCF128));
3517 Results.push_back(ExpandFPLibCall(Node, RTLIB::ADD_F32, RTLIB::ADD_F64,
3518 RTLIB::ADD_F80, RTLIB::ADD_F128,
3519 RTLIB::ADD_PPCF128));
3522 Results.push_back(ExpandFPLibCall(Node, RTLIB::MUL_F32, RTLIB::MUL_F64,
3523 RTLIB::MUL_F80, RTLIB::MUL_F128,
3524 RTLIB::MUL_PPCF128));
3526 case ISD::FP16_TO_FP: {
3527 if (Node->getValueType(0) == MVT::f32) {
3528 Results.push_back(ExpandLibCall(RTLIB::FPEXT_F16_F32, Node, false));
3532 // We can extend to types bigger than f32 in two steps without changing the
3533 // result. Since "f16 -> f32" is much more commonly available, give CodeGen
3534 // the option of emitting that before resorting to a libcall.
3536 DAG.getNode(ISD::FP16_TO_FP, dl, MVT::f32, Node->getOperand(0));
3538 DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Res));
3541 case ISD::FP_TO_FP16: {
3543 RTLIB::getFPROUND(Node->getOperand(0).getValueType(), MVT::f16);
3544 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unable to expand fp_to_fp16");
3545 Results.push_back(ExpandLibCall(LC, Node, false));
3548 case ISD::ConstantFP: {
3549 ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Node);
3550 // Check to see if this FP immediate is already legal.
3551 // If this is a legal constant, turn it into a TargetConstantFP node.
3552 if (!TLI.isFPImmLegal(CFP->getValueAPF(), Node->getValueType(0)))
3553 Results.push_back(ExpandConstantFP(CFP, true));
3557 EVT VT = Node->getValueType(0);
3558 if (TLI.isOperationLegalOrCustom(ISD::FADD, VT) &&
3559 TLI.isOperationLegalOrCustom(ISD::FNEG, VT)) {
3560 Tmp1 = DAG.getNode(ISD::FNEG, dl, VT, Node->getOperand(1));
3561 Tmp1 = DAG.getNode(ISD::FADD, dl, VT, Node->getOperand(0), Tmp1);
3562 Results.push_back(Tmp1);
3564 Results.push_back(ExpandFPLibCall(Node, RTLIB::SUB_F32, RTLIB::SUB_F64,
3565 RTLIB::SUB_F80, RTLIB::SUB_F128,
3566 RTLIB::SUB_PPCF128));
3571 EVT VT = Node->getValueType(0);
3572 assert(TLI.isOperationLegalOrCustom(ISD::ADD, VT) &&
3573 TLI.isOperationLegalOrCustom(ISD::XOR, VT) &&
3574 "Don't know how to expand this subtraction!");
3575 Tmp1 = DAG.getNode(ISD::XOR, dl, VT, Node->getOperand(1),
3576 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), VT));
3577 Tmp1 = DAG.getNode(ISD::ADD, dl, VT, Tmp1, DAG.getConstant(1, VT));
3578 Results.push_back(DAG.getNode(ISD::ADD, dl, VT, Node->getOperand(0), Tmp1));
3583 EVT VT = Node->getValueType(0);
3584 bool isSigned = Node->getOpcode() == ISD::SREM;
3585 unsigned DivOpc = isSigned ? ISD::SDIV : ISD::UDIV;
3586 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3587 Tmp2 = Node->getOperand(0);
3588 Tmp3 = Node->getOperand(1);
3589 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3590 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3591 // If div is legal, it's better to do the normal expansion
3592 !TLI.isOperationLegalOrCustom(DivOpc, Node->getValueType(0)) &&
3593 useDivRem(Node, isSigned, false))) {
3594 SDVTList VTs = DAG.getVTList(VT, VT);
3595 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Tmp2, Tmp3).getValue(1);
3596 } else if (TLI.isOperationLegalOrCustom(DivOpc, VT)) {
3598 Tmp1 = DAG.getNode(DivOpc, dl, VT, Tmp2, Tmp3);
3599 Tmp1 = DAG.getNode(ISD::MUL, dl, VT, Tmp1, Tmp3);
3600 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, Tmp2, Tmp1);
3601 } else if (isSigned)
3602 Tmp1 = ExpandIntLibCall(Node, true,
3604 RTLIB::SREM_I16, RTLIB::SREM_I32,
3605 RTLIB::SREM_I64, RTLIB::SREM_I128);
3607 Tmp1 = ExpandIntLibCall(Node, false,
3609 RTLIB::UREM_I16, RTLIB::UREM_I32,
3610 RTLIB::UREM_I64, RTLIB::UREM_I128);
3611 Results.push_back(Tmp1);
3616 bool isSigned = Node->getOpcode() == ISD::SDIV;
3617 unsigned DivRemOpc = isSigned ? ISD::SDIVREM : ISD::UDIVREM;
3618 EVT VT = Node->getValueType(0);
3619 SDVTList VTs = DAG.getVTList(VT, VT);
3620 if (TLI.isOperationLegalOrCustom(DivRemOpc, VT) ||
3621 (isDivRemLibcallAvailable(Node, isSigned, TLI) &&
3622 useDivRem(Node, isSigned, true)))
3623 Tmp1 = DAG.getNode(DivRemOpc, dl, VTs, Node->getOperand(0),
3624 Node->getOperand(1));
3626 Tmp1 = ExpandIntLibCall(Node, true,
3628 RTLIB::SDIV_I16, RTLIB::SDIV_I32,
3629 RTLIB::SDIV_I64, RTLIB::SDIV_I128);
3631 Tmp1 = ExpandIntLibCall(Node, false,
3633 RTLIB::UDIV_I16, RTLIB::UDIV_I32,
3634 RTLIB::UDIV_I64, RTLIB::UDIV_I128);
3635 Results.push_back(Tmp1);
3640 unsigned ExpandOpcode = Node->getOpcode() == ISD::MULHU ? ISD::UMUL_LOHI :
3642 EVT VT = Node->getValueType(0);
3643 SDVTList VTs = DAG.getVTList(VT, VT);
3644 assert(TLI.isOperationLegalOrCustom(ExpandOpcode, VT) &&
3645 "If this wasn't legal, it shouldn't have been created!");
3646 Tmp1 = DAG.getNode(ExpandOpcode, dl, VTs, Node->getOperand(0),
3647 Node->getOperand(1));
3648 Results.push_back(Tmp1.getValue(1));
3653 // Expand into divrem libcall
3654 ExpandDivRemLibCall(Node, Results);
3657 EVT VT = Node->getValueType(0);
3658 SDVTList VTs = DAG.getVTList(VT, VT);
3659 // See if multiply or divide can be lowered using two-result operations.
3660 // We just need the low half of the multiply; try both the signed
3661 // and unsigned forms. If the target supports both SMUL_LOHI and
3662 // UMUL_LOHI, form a preference by checking which forms of plain
3663 // MULH it supports.
3664 bool HasSMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::SMUL_LOHI, VT);
3665 bool HasUMUL_LOHI = TLI.isOperationLegalOrCustom(ISD::UMUL_LOHI, VT);
3666 bool HasMULHS = TLI.isOperationLegalOrCustom(ISD::MULHS, VT);
3667 bool HasMULHU = TLI.isOperationLegalOrCustom(ISD::MULHU, VT);
3668 unsigned OpToUse = 0;
3669 if (HasSMUL_LOHI && !HasMULHS) {
3670 OpToUse = ISD::SMUL_LOHI;
3671 } else if (HasUMUL_LOHI && !HasMULHU) {
3672 OpToUse = ISD::UMUL_LOHI;
3673 } else if (HasSMUL_LOHI) {
3674 OpToUse = ISD::SMUL_LOHI;
3675 } else if (HasUMUL_LOHI) {
3676 OpToUse = ISD::UMUL_LOHI;
3679 Results.push_back(DAG.getNode(OpToUse, dl, VTs, Node->getOperand(0),
3680 Node->getOperand(1)));
3685 EVT HalfType = VT.getHalfSizedIntegerVT(*DAG.getContext());
3686 if (TLI.isOperationLegalOrCustom(ISD::ZERO_EXTEND, VT) &&
3687 TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND, VT) &&
3688 TLI.isOperationLegalOrCustom(ISD::SHL, VT) &&
3689 TLI.isOperationLegalOrCustom(ISD::OR, VT) &&
3690 TLI.expandMUL(Node, Lo, Hi, HalfType, DAG)) {
3691 Lo = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Lo);
3692 Hi = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Hi);
3693 SDValue Shift = DAG.getConstant(HalfType.getSizeInBits(),
3694 TLI.getShiftAmountTy(HalfType));
3695 Hi = DAG.getNode(ISD::SHL, dl, VT, Hi, Shift);
3696 Results.push_back(DAG.getNode(ISD::OR, dl, VT, Lo, Hi));
3700 Tmp1 = ExpandIntLibCall(Node, false,
3702 RTLIB::MUL_I16, RTLIB::MUL_I32,
3703 RTLIB::MUL_I64, RTLIB::MUL_I128);
3704 Results.push_back(Tmp1);
3709 SDValue LHS = Node->getOperand(0);
3710 SDValue RHS = Node->getOperand(1);
3711 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::SADDO ?
3712 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3714 Results.push_back(Sum);
3715 EVT ResultType = Node->getValueType(1);
3716 EVT OType = getSetCCResultType(Node->getValueType(0));
3718 SDValue Zero = DAG.getConstant(0, LHS.getValueType());
3720 // LHSSign -> LHS >= 0
3721 // RHSSign -> RHS >= 0
3722 // SumSign -> Sum >= 0
3725 // Overflow -> (LHSSign == RHSSign) && (LHSSign != SumSign)
3727 // Overflow -> (LHSSign != RHSSign) && (LHSSign != SumSign)
3729 SDValue LHSSign = DAG.getSetCC(dl, OType, LHS, Zero, ISD::SETGE);
3730 SDValue RHSSign = DAG.getSetCC(dl, OType, RHS, Zero, ISD::SETGE);
3731 SDValue SignsMatch = DAG.getSetCC(dl, OType, LHSSign, RHSSign,
3732 Node->getOpcode() == ISD::SADDO ?
3733 ISD::SETEQ : ISD::SETNE);
3735 SDValue SumSign = DAG.getSetCC(dl, OType, Sum, Zero, ISD::SETGE);
3736 SDValue SumSignNE = DAG.getSetCC(dl, OType, LHSSign, SumSign, ISD::SETNE);
3738 SDValue Cmp = DAG.getNode(ISD::AND, dl, OType, SignsMatch, SumSignNE);
3739 Results.push_back(DAG.getBoolExtOrTrunc(Cmp, dl, ResultType, ResultType));
3744 SDValue LHS = Node->getOperand(0);
3745 SDValue RHS = Node->getOperand(1);
3746 SDValue Sum = DAG.getNode(Node->getOpcode() == ISD::UADDO ?
3747 ISD::ADD : ISD::SUB, dl, LHS.getValueType(),
3749 Results.push_back(Sum);
3751 EVT ResultType = Node->getValueType(1);
3752 EVT SetCCType = getSetCCResultType(Node->getValueType(0));
3754 = Node->getOpcode() == ISD::UADDO ? ISD::SETULT : ISD::SETUGT;
3755 SDValue SetCC = DAG.getSetCC(dl, SetCCType, Sum, LHS, CC);
3757 Results.push_back(DAG.getBoolExtOrTrunc(SetCC, dl, ResultType, ResultType));
3762 EVT VT = Node->getValueType(0);
3763 EVT WideVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits() * 2);
3764 SDValue LHS = Node->getOperand(0);
3765 SDValue RHS = Node->getOperand(1);
3768 static const unsigned Ops[2][3] =
3769 { { ISD::MULHU, ISD::UMUL_LOHI, ISD::ZERO_EXTEND },
3770 { ISD::MULHS, ISD::SMUL_LOHI, ISD::SIGN_EXTEND }};
3771 bool isSigned = Node->getOpcode() == ISD::SMULO;
3772 if (TLI.isOperationLegalOrCustom(Ops[isSigned][0], VT)) {
3773 BottomHalf = DAG.getNode(ISD::MUL, dl, VT, LHS, RHS);
3774 TopHalf = DAG.getNode(Ops[isSigned][0], dl, VT, LHS, RHS);
3775 } else if (TLI.isOperationLegalOrCustom(Ops[isSigned][1], VT)) {
3776 BottomHalf = DAG.getNode(Ops[isSigned][1], dl, DAG.getVTList(VT, VT), LHS,
3778 TopHalf = BottomHalf.getValue(1);
3779 } else if (TLI.isTypeLegal(WideVT)) {
3780 LHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, LHS);
3781 RHS = DAG.getNode(Ops[isSigned][2], dl, WideVT, RHS);
3782 Tmp1 = DAG.getNode(ISD::MUL, dl, WideVT, LHS, RHS);
3783 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3784 DAG.getIntPtrConstant(0));
3785 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Tmp1,
3786 DAG.getIntPtrConstant(1));
3788 // We can fall back to a libcall with an illegal type for the MUL if we
3789 // have a libcall big enough.
3790 // Also, we can fall back to a division in some cases, but that's a big
3791 // performance hit in the general case.
3792 RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
3793 if (WideVT == MVT::i16)
3794 LC = RTLIB::MUL_I16;
3795 else if (WideVT == MVT::i32)
3796 LC = RTLIB::MUL_I32;
3797 else if (WideVT == MVT::i64)
3798 LC = RTLIB::MUL_I64;
3799 else if (WideVT == MVT::i128)
3800 LC = RTLIB::MUL_I128;
3801 assert(LC != RTLIB::UNKNOWN_LIBCALL && "Cannot expand this operation!");
3803 // The high part is obtained by SRA'ing all but one of the bits of low
3805 unsigned LoSize = VT.getSizeInBits();
3806 SDValue HiLHS = DAG.getNode(ISD::SRA, dl, VT, RHS,
3807 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3808 SDValue HiRHS = DAG.getNode(ISD::SRA, dl, VT, LHS,
3809 DAG.getConstant(LoSize-1, TLI.getPointerTy()));
3811 // Here we're passing the 2 arguments explicitly as 4 arguments that are
3812 // pre-lowered to the correct types. This all depends upon WideVT not
3813 // being a legal type for the architecture and thus has to be split to
3815 SDValue Args[] = { LHS, HiLHS, RHS, HiRHS };
3816 SDValue Ret = ExpandLibCall(LC, WideVT, Args, 4, isSigned, dl);
3817 BottomHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3818 DAG.getIntPtrConstant(0));
3819 TopHalf = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, VT, Ret,
3820 DAG.getIntPtrConstant(1));
3821 // Ret is a node with an illegal type. Because such things are not
3822 // generally permitted during this phase of legalization, make sure the
3823 // node has no more uses. The above EXTRACT_ELEMENT nodes should have been
3825 assert(Ret->use_empty() &&
3826 "Unexpected uses of illegally type from expanded lib call.");
3830 Tmp1 = DAG.getConstant(VT.getSizeInBits() - 1,
3831 TLI.getShiftAmountTy(BottomHalf.getValueType()));
3832 Tmp1 = DAG.getNode(ISD::SRA, dl, VT, BottomHalf, Tmp1);
3833 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf, Tmp1,
3836 TopHalf = DAG.getSetCC(dl, getSetCCResultType(VT), TopHalf,
3837 DAG.getConstant(0, VT), ISD::SETNE);
3839 Results.push_back(BottomHalf);
3840 Results.push_back(TopHalf);
3843 case ISD::BUILD_PAIR: {
3844 EVT PairTy = Node->getValueType(0);
3845 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, PairTy, Node->getOperand(0));
3846 Tmp2 = DAG.getNode(ISD::ANY_EXTEND, dl, PairTy, Node->getOperand(1));
3847 Tmp2 = DAG.getNode(ISD::SHL, dl, PairTy, Tmp2,
3848 DAG.getConstant(PairTy.getSizeInBits()/2,
3849 TLI.getShiftAmountTy(PairTy)));
3850 Results.push_back(DAG.getNode(ISD::OR, dl, PairTy, Tmp1, Tmp2));
3854 Tmp1 = Node->getOperand(0);
3855 Tmp2 = Node->getOperand(1);
3856 Tmp3 = Node->getOperand(2);
3857 if (Tmp1.getOpcode() == ISD::SETCC) {
3858 Tmp1 = DAG.getSelectCC(dl, Tmp1.getOperand(0), Tmp1.getOperand(1),
3860 cast<CondCodeSDNode>(Tmp1.getOperand(2))->get());
3862 Tmp1 = DAG.getSelectCC(dl, Tmp1,
3863 DAG.getConstant(0, Tmp1.getValueType()),
3864 Tmp2, Tmp3, ISD::SETNE);
3866 Results.push_back(Tmp1);
3869 SDValue Chain = Node->getOperand(0);
3870 SDValue Table = Node->getOperand(1);
3871 SDValue Index = Node->getOperand(2);
3873 EVT PTy = TLI.getPointerTy();
3875 const DataLayout &TD = *TLI.getDataLayout();
3876 unsigned EntrySize =
3877 DAG.getMachineFunction().getJumpTableInfo()->getEntrySize(TD);
3879 Index = DAG.getNode(ISD::MUL, dl, Index.getValueType(),
3880 Index, DAG.getConstant(EntrySize, Index.getValueType()));
3881 SDValue Addr = DAG.getNode(ISD::ADD, dl, Index.getValueType(),
3884 EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8);
3885 SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr,
3886 MachinePointerInfo::getJumpTable(), MemVT,
3887 false, false, false, 0);
3889 if (TM.getRelocationModel() == Reloc::PIC_) {
3890 // For PIC, the sequence is:
3891 // BRIND(load(Jumptable + index) + RelocBase)
3892 // RelocBase can be JumpTable, GOT or some sort of global base.
3893 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr,
3894 TLI.getPICJumpTableRelocBase(Table, DAG));
3896 Tmp1 = DAG.getNode(ISD::BRIND, dl, MVT::Other, LD.getValue(1), Addr);
3897 Results.push_back(Tmp1);
3901 // Expand brcond's setcc into its constituent parts and create a BR_CC
3903 Tmp1 = Node->getOperand(0);
3904 Tmp2 = Node->getOperand(1);
3905 if (Tmp2.getOpcode() == ISD::SETCC) {
3906 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other,
3907 Tmp1, Tmp2.getOperand(2),
3908 Tmp2.getOperand(0), Tmp2.getOperand(1),
3909 Node->getOperand(2));
3911 // We test only the i1 bit. Skip the AND if UNDEF.
3912 Tmp3 = (Tmp2.getOpcode() == ISD::UNDEF) ? Tmp2 :
3913 DAG.getNode(ISD::AND, dl, Tmp2.getValueType(), Tmp2,
3914 DAG.getConstant(1, Tmp2.getValueType()));
3915 Tmp1 = DAG.getNode(ISD::BR_CC, dl, MVT::Other, Tmp1,
3916 DAG.getCondCode(ISD::SETNE), Tmp3,
3917 DAG.getConstant(0, Tmp3.getValueType()),
3918 Node->getOperand(2));
3920 Results.push_back(Tmp1);
3923 Tmp1 = Node->getOperand(0);
3924 Tmp2 = Node->getOperand(1);
3925 Tmp3 = Node->getOperand(2);
3926 bool Legalized = LegalizeSetCCCondCode(Node->getValueType(0), Tmp1, Tmp2,
3927 Tmp3, NeedInvert, dl);
3930 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
3931 // condition code, create a new SETCC node.
3933 Tmp1 = DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
3936 // If we expanded the SETCC by inverting the condition code, then wrap
3937 // the existing SETCC in a NOT to restore the intended condition.
3939 Tmp1 = DAG.getLogicalNOT(dl, Tmp1, Tmp1->getValueType(0));
3941 Results.push_back(Tmp1);
3945 // Otherwise, SETCC for the given comparison type must be completely
3946 // illegal; expand it into a SELECT_CC.
3947 EVT VT = Node->getValueType(0);
3949 switch (TLI.getBooleanContents(Tmp1->getValueType(0))) {
3950 case TargetLowering::ZeroOrOneBooleanContent:
3951 case TargetLowering::UndefinedBooleanContent:
3954 case TargetLowering::ZeroOrNegativeOneBooleanContent:
3958 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, VT, Tmp1, Tmp2,
3959 DAG.getConstant(TrueValue, VT), DAG.getConstant(0, VT),
3961 Results.push_back(Tmp1);
3964 case ISD::SELECT_CC: {
3965 Tmp1 = Node->getOperand(0); // LHS
3966 Tmp2 = Node->getOperand(1); // RHS
3967 Tmp3 = Node->getOperand(2); // True
3968 Tmp4 = Node->getOperand(3); // False
3969 EVT VT = Node->getValueType(0);
3970 SDValue CC = Node->getOperand(4);
3971 ISD::CondCode CCOp = cast<CondCodeSDNode>(CC)->get();
3973 if (TLI.isCondCodeLegal(CCOp, Tmp1.getSimpleValueType())) {
3974 // If the condition code is legal, then we need to expand this
3975 // node using SETCC and SELECT.
3976 EVT CmpVT = Tmp1.getValueType();
3977 assert(!TLI.isOperationExpand(ISD::SELECT, VT) &&
3978 "Cannot expand ISD::SELECT_CC when ISD::SELECT also needs to be "
3980 EVT CCVT = TLI.getSetCCResultType(*DAG.getContext(), CmpVT);
3981 SDValue Cond = DAG.getNode(ISD::SETCC, dl, CCVT, Tmp1, Tmp2, CC);
3982 Results.push_back(DAG.getSelect(dl, VT, Cond, Tmp3, Tmp4));
3986 // SELECT_CC is legal, so the condition code must not be.
3987 bool Legalized = false;
3988 // Try to legalize by inverting the condition. This is for targets that
3989 // might support an ordered version of a condition, but not the unordered
3990 // version (or vice versa).
3991 ISD::CondCode InvCC = ISD::getSetCCInverse(CCOp,
3992 Tmp1.getValueType().isInteger());
3993 if (TLI.isCondCodeLegal(InvCC, Tmp1.getSimpleValueType())) {
3994 // Use the new condition code and swap true and false
3996 Tmp1 = DAG.getSelectCC(dl, Tmp1, Tmp2, Tmp4, Tmp3, InvCC);
3998 // If The inverse is not legal, then try to swap the arguments using
3999 // the inverse condition code.
4000 ISD::CondCode SwapInvCC = ISD::getSetCCSwappedOperands(InvCC);
4001 if (TLI.isCondCodeLegal(SwapInvCC, Tmp1.getSimpleValueType())) {
4002 // The swapped inverse condition is legal, so swap true and false,
4005 Tmp1 = DAG.getSelectCC(dl, Tmp2, Tmp1, Tmp4, Tmp3, SwapInvCC);
4010 Legalized = LegalizeSetCCCondCode(
4011 getSetCCResultType(Tmp1.getValueType()), Tmp1, Tmp2, CC, NeedInvert,
4014 assert(Legalized && "Can't legalize SELECT_CC with legal condition!");
4016 // If we expanded the SETCC by inverting the condition code, then swap
4017 // the True/False operands to match.
4019 std::swap(Tmp3, Tmp4);
4021 // If we expanded the SETCC by swapping LHS and RHS, or by inverting the
4022 // condition code, create a new SELECT_CC node.
4024 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0),
4025 Tmp1, Tmp2, Tmp3, Tmp4, CC);
4027 Tmp2 = DAG.getConstant(0, Tmp1.getValueType());
4028 CC = DAG.getCondCode(ISD::SETNE);
4029 Tmp1 = DAG.getNode(ISD::SELECT_CC, dl, Node->getValueType(0), Tmp1,
4030 Tmp2, Tmp3, Tmp4, CC);
4033 Results.push_back(Tmp1);
4037 Tmp1 = Node->getOperand(0); // Chain
4038 Tmp2 = Node->getOperand(2); // LHS
4039 Tmp3 = Node->getOperand(3); // RHS
4040 Tmp4 = Node->getOperand(1); // CC
4042 bool Legalized = LegalizeSetCCCondCode(getSetCCResultType(
4043 Tmp2.getValueType()), Tmp2, Tmp3, Tmp4, NeedInvert, dl);
4045 assert(Legalized && "Can't legalize BR_CC with legal condition!");
4047 // If we expanded the SETCC by inverting the condition code, then wrap
4048 // the existing SETCC in a NOT to restore the intended condition.
4050 Tmp4 = DAG.getNOT(dl, Tmp4, Tmp4->getValueType(0));
4052 // If we expanded the SETCC by swapping LHS and RHS, create a new BR_CC
4054 if (Tmp4.getNode()) {
4055 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1,
4056 Tmp4, Tmp2, Tmp3, Node->getOperand(4));
4058 Tmp3 = DAG.getConstant(0, Tmp2.getValueType());
4059 Tmp4 = DAG.getCondCode(ISD::SETNE);
4060 Tmp1 = DAG.getNode(ISD::BR_CC, dl, Node->getValueType(0), Tmp1, Tmp4,
4061 Tmp2, Tmp3, Node->getOperand(4));
4063 Results.push_back(Tmp1);
4066 case ISD::BUILD_VECTOR:
4067 Results.push_back(ExpandBUILD_VECTOR(Node));
4072 // Scalarize vector SRA/SRL/SHL.
4073 EVT VT = Node->getValueType(0);
4074 assert(VT.isVector() && "Unable to legalize non-vector shift");
4075 assert(TLI.isTypeLegal(VT.getScalarType())&& "Element type must be legal");
4076 unsigned NumElem = VT.getVectorNumElements();
4078 SmallVector<SDValue, 8> Scalars;
4079 for (unsigned Idx = 0; Idx < NumElem; Idx++) {
4080 SDValue Ex = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4082 Node->getOperand(0), DAG.getConstant(Idx,
4083 TLI.getVectorIdxTy()));
4084 SDValue Sh = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
4086 Node->getOperand(1), DAG.getConstant(Idx,
4087 TLI.getVectorIdxTy()));
4088 Scalars.push_back(DAG.getNode(Node->getOpcode(), dl,
4089 VT.getScalarType(), Ex, Sh));
4092 DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Scalars);
4093 ReplaceNode(SDValue(Node, 0), Result);
4096 case ISD::GLOBAL_OFFSET_TABLE:
4097 case ISD::GlobalAddress:
4098 case ISD::GlobalTLSAddress:
4099 case ISD::ExternalSymbol:
4100 case ISD::ConstantPool:
4101 case ISD::JumpTable:
4102 case ISD::INTRINSIC_W_CHAIN:
4103 case ISD::INTRINSIC_WO_CHAIN:
4104 case ISD::INTRINSIC_VOID:
4105 // FIXME: Custom lowering for these operations shouldn't return null!
4109 // Replace the original node with the legalized result.
4110 if (!Results.empty())
4111 ReplaceNode(Node, Results.data());
4114 void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
4115 SmallVector<SDValue, 8> Results;
4116 MVT OVT = Node->getSimpleValueType(0);
4117 if (Node->getOpcode() == ISD::UINT_TO_FP ||
4118 Node->getOpcode() == ISD::SINT_TO_FP ||
4119 Node->getOpcode() == ISD::SETCC) {
4120 OVT = Node->getOperand(0).getSimpleValueType();
4122 MVT NVT = TLI.getTypeToPromoteTo(Node->getOpcode(), OVT);
4124 SDValue Tmp1, Tmp2, Tmp3;
4125 switch (Node->getOpcode()) {
4127 case ISD::CTTZ_ZERO_UNDEF:
4129 case ISD::CTLZ_ZERO_UNDEF:
4131 // Zero extend the argument.
4132 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4133 // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
4134 // already the correct result.
4135 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4136 if (Node->getOpcode() == ISD::CTTZ) {
4137 // FIXME: This should set a bit in the zero extended value instead.
4138 Tmp2 = DAG.getSetCC(dl, getSetCCResultType(NVT),
4139 Tmp1, DAG.getConstant(NVT.getSizeInBits(), NVT),
4141 Tmp1 = DAG.getSelect(dl, NVT, Tmp2,
4142 DAG.getConstant(OVT.getSizeInBits(), NVT), Tmp1);
4143 } else if (Node->getOpcode() == ISD::CTLZ ||
4144 Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
4145 // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
4146 Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
4147 DAG.getConstant(NVT.getSizeInBits() -
4148 OVT.getSizeInBits(), NVT));
4150 Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
4153 unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
4154 Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
4155 Tmp1 = DAG.getNode(ISD::BSWAP, dl, NVT, Tmp1);
4156 Tmp1 = DAG.getNode(ISD::SRL, dl, NVT, Tmp1,
4157 DAG.getConstant(DiffBits, TLI.getShiftAmountTy(NVT)));
4158 Results.push_back(Tmp1);
4161 case ISD::FP_TO_UINT:
4162 case ISD::FP_TO_SINT:
4163 Tmp1 = PromoteLegalFP_TO_INT(Node->getOperand(0), Node->getValueType(0),
4164 Node->getOpcode() == ISD::FP_TO_SINT, dl);
4165 Results.push_back(Tmp1);
4167 case ISD::UINT_TO_FP:
4168 case ISD::SINT_TO_FP:
4169 Tmp1 = PromoteLegalINT_TO_FP(Node->getOperand(0), Node->getValueType(0),
4170 Node->getOpcode() == ISD::SINT_TO_FP, dl);
4171 Results.push_back(Tmp1);
4174 SDValue Chain = Node->getOperand(0); // Get the chain.
4175 SDValue Ptr = Node->getOperand(1); // Get the pointer.
4178 if (OVT.isVector()) {
4179 TruncOp = ISD::BITCAST;
4181 assert(OVT.isInteger()
4182 && "VAARG promotion is supported only for vectors or integer types");
4183 TruncOp = ISD::TRUNCATE;
4186 // Perform the larger operation, then convert back
4187 Tmp1 = DAG.getVAArg(NVT, dl, Chain, Ptr, Node->getOperand(2),
4188 Node->getConstantOperandVal(3));
4189 Chain = Tmp1.getValue(1);
4191 Tmp2 = DAG.getNode(TruncOp, dl, OVT, Tmp1);
4193 // Modified the chain result - switch anything that used the old chain to
4195 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 0), Tmp2);
4196 DAG.ReplaceAllUsesOfValueWith(SDValue(Node, 1), Chain);
4198 UpdatedNodes->insert(Tmp2.getNode());
4199 UpdatedNodes->insert(Chain.getNode());
4207 unsigned ExtOp, TruncOp;
4208 if (OVT.isVector()) {
4209 ExtOp = ISD::BITCAST;
4210 TruncOp = ISD::BITCAST;
4212 assert(OVT.isInteger() && "Cannot promote logic operation");
4213 ExtOp = ISD::ANY_EXTEND;
4214 TruncOp = ISD::TRUNCATE;
4216 // Promote each of the values to the new type.
4217 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4218 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4219 // Perform the larger operation, then convert back
4220 Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4221 Results.push_back(DAG.getNode(TruncOp, dl, OVT, Tmp1));
4225 unsigned ExtOp, TruncOp;
4226 if (Node->getValueType(0).isVector() ||
4227 Node->getValueType(0).getSizeInBits() == NVT.getSizeInBits()) {
4228 ExtOp = ISD::BITCAST;
4229 TruncOp = ISD::BITCAST;
4230 } else if (Node->getValueType(0).isInteger()) {
4231 ExtOp = ISD::ANY_EXTEND;
4232 TruncOp = ISD::TRUNCATE;
4234 ExtOp = ISD::FP_EXTEND;
4235 TruncOp = ISD::FP_ROUND;
4237 Tmp1 = Node->getOperand(0);
4238 // Promote each of the values to the new type.
4239 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4240 Tmp3 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(2));
4241 // Perform the larger operation, then round down.
4242 Tmp1 = DAG.getSelect(dl, NVT, Tmp1, Tmp2, Tmp3);
4243 if (TruncOp != ISD::FP_ROUND)
4244 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1);
4246 Tmp1 = DAG.getNode(TruncOp, dl, Node->getValueType(0), Tmp1,
4247 DAG.getIntPtrConstant(0));
4248 Results.push_back(Tmp1);
4251 case ISD::VECTOR_SHUFFLE: {
4252 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Node)->getMask();
4254 // Cast the two input vectors.
4255 Tmp1 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(0));
4256 Tmp2 = DAG.getNode(ISD::BITCAST, dl, NVT, Node->getOperand(1));
4258 // Convert the shuffle mask to the right # elements.
4259 Tmp1 = ShuffleWithNarrowerEltType(NVT, OVT, dl, Tmp1, Tmp2, Mask);
4260 Tmp1 = DAG.getNode(ISD::BITCAST, dl, OVT, Tmp1);
4261 Results.push_back(Tmp1);
4265 unsigned ExtOp = ISD::FP_EXTEND;
4266 if (NVT.isInteger()) {
4267 ISD::CondCode CCCode =
4268 cast<CondCodeSDNode>(Node->getOperand(2))->get();
4269 ExtOp = isSignedIntSetCC(CCCode) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
4271 Tmp1 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(0));
4272 Tmp2 = DAG.getNode(ExtOp, dl, NVT, Node->getOperand(1));
4273 Results.push_back(DAG.getNode(ISD::SETCC, dl, Node->getValueType(0),
4274 Tmp1, Tmp2, Node->getOperand(2)));
4283 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4284 Tmp2 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(1));
4285 Tmp3 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1, Tmp2);
4286 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4287 Tmp3, DAG.getIntPtrConstant(0)));
4294 Tmp1 = DAG.getNode(ISD::FP_EXTEND, dl, NVT, Node->getOperand(0));
4295 Tmp2 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
4296 Results.push_back(DAG.getNode(ISD::FP_ROUND, dl, OVT,
4297 Tmp2, DAG.getIntPtrConstant(0)));
4302 // Replace the original node with the legalized result.
4303 if (!Results.empty())
4304 ReplaceNode(Node, Results.data());
4307 /// This is the entry point for the file.
4308 void SelectionDAG::Legalize() {
4309 AssignTopologicalOrder();
4311 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4312 SelectionDAGLegalize Legalizer(*this, LegalizedNodes);
4314 // Visit all the nodes. We start in topological order, so that we see
4315 // nodes with their original operands intact. Legalization can produce
4316 // new nodes which may themselves need to be legalized. Iterate until all
4317 // nodes have been legalized.
4319 bool AnyLegalized = false;
4320 for (auto NI = allnodes_end(); NI != allnodes_begin();) {
4324 if (N->use_empty() && N != getRoot().getNode()) {
4330 if (LegalizedNodes.insert(N).second) {
4331 AnyLegalized = true;
4332 Legalizer.LegalizeOp(N);
4334 if (N->use_empty() && N != getRoot().getNode()) {
4345 // Remove dead nodes now.
4349 bool SelectionDAG::LegalizeOp(SDNode *N,
4350 SmallSetVector<SDNode *, 16> &UpdatedNodes) {
4351 SmallPtrSet<SDNode *, 16> LegalizedNodes;
4352 SelectionDAGLegalize Legalizer(*this, LegalizedNodes, &UpdatedNodes);
4354 // Directly insert the node in question, and legalize it. This will recurse
4355 // as needed through operands.
4356 LegalizedNodes.insert(N);
4357 Legalizer.LegalizeOp(N);
4359 return LegalizedNodes.count(N);